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>
28 #include <linux/pagewalk.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/psi.h>
61 #include <linux/seq_buf.h>
67 #include <linux/uaccess.h>
69 #include <trace/events/vmscan.h>
71 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
72 EXPORT_SYMBOL(memory_cgrp_subsys);
74 struct mem_cgroup *root_mem_cgroup __read_mostly;
76 /* Active memory cgroup to use from an interrupt context */
77 DEFINE_PER_CPU(struct mem_cgroup *, int_active_memcg);
79 /* Socket memory accounting disabled? */
80 static bool cgroup_memory_nosocket;
82 /* Kernel memory accounting disabled? */
83 static bool cgroup_memory_nokmem;
85 /* Whether the swap controller is active */
86 #ifdef CONFIG_MEMCG_SWAP
87 bool cgroup_memory_noswap __read_mostly;
89 #define cgroup_memory_noswap 1
92 #ifdef CONFIG_CGROUP_WRITEBACK
93 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
96 /* Whether legacy memory+swap accounting is active */
97 static bool do_memsw_account(void)
99 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_noswap;
102 #define THRESHOLDS_EVENTS_TARGET 128
103 #define SOFTLIMIT_EVENTS_TARGET 1024
106 * Cgroups above their limits are maintained in a RB-Tree, independent of
107 * their hierarchy representation
110 struct mem_cgroup_tree_per_node {
111 struct rb_root rb_root;
112 struct rb_node *rb_rightmost;
116 struct mem_cgroup_tree {
117 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
120 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
123 struct mem_cgroup_eventfd_list {
124 struct list_head list;
125 struct eventfd_ctx *eventfd;
129 * cgroup_event represents events which userspace want to receive.
131 struct mem_cgroup_event {
133 * memcg which the event belongs to.
135 struct mem_cgroup *memcg;
137 * eventfd to signal userspace about the event.
139 struct eventfd_ctx *eventfd;
141 * Each of these stored in a list by the cgroup.
143 struct list_head list;
145 * register_event() callback will be used to add new userspace
146 * waiter for changes related to this event. Use eventfd_signal()
147 * on eventfd to send notification to userspace.
149 int (*register_event)(struct mem_cgroup *memcg,
150 struct eventfd_ctx *eventfd, const char *args);
152 * unregister_event() callback will be called when userspace closes
153 * the eventfd or on cgroup removing. This callback must be set,
154 * if you want provide notification functionality.
156 void (*unregister_event)(struct mem_cgroup *memcg,
157 struct eventfd_ctx *eventfd);
159 * All fields below needed to unregister event when
160 * userspace closes eventfd.
163 wait_queue_head_t *wqh;
164 wait_queue_entry_t wait;
165 struct work_struct remove;
168 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
169 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
171 /* Stuffs for move charges at task migration. */
173 * Types of charges to be moved.
175 #define MOVE_ANON 0x1U
176 #define MOVE_FILE 0x2U
177 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
179 /* "mc" and its members are protected by cgroup_mutex */
180 static struct move_charge_struct {
181 spinlock_t lock; /* for from, to */
182 struct mm_struct *mm;
183 struct mem_cgroup *from;
184 struct mem_cgroup *to;
186 unsigned long precharge;
187 unsigned long moved_charge;
188 unsigned long moved_swap;
189 struct task_struct *moving_task; /* a task moving charges */
190 wait_queue_head_t waitq; /* a waitq for other context */
192 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
193 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
197 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
198 * limit reclaim to prevent infinite loops, if they ever occur.
200 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
201 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
203 /* for encoding cft->private value on file */
212 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
213 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
214 #define MEMFILE_ATTR(val) ((val) & 0xffff)
215 /* Used for OOM nofiier */
216 #define OOM_CONTROL (0)
219 * Iteration constructs for visiting all cgroups (under a tree). If
220 * loops are exited prematurely (break), mem_cgroup_iter_break() must
221 * be used for reference counting.
223 #define for_each_mem_cgroup_tree(iter, root) \
224 for (iter = mem_cgroup_iter(root, NULL, NULL); \
226 iter = mem_cgroup_iter(root, iter, NULL))
228 #define for_each_mem_cgroup(iter) \
229 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
231 iter = mem_cgroup_iter(NULL, iter, NULL))
233 static inline bool task_is_dying(void)
235 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
236 (current->flags & PF_EXITING);
239 /* Some nice accessors for the vmpressure. */
240 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
243 memcg = root_mem_cgroup;
244 return &memcg->vmpressure;
247 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
249 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
252 #ifdef CONFIG_MEMCG_KMEM
253 static DEFINE_SPINLOCK(objcg_lock);
255 static void obj_cgroup_release(struct percpu_ref *ref)
257 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
258 struct mem_cgroup *memcg;
259 unsigned int nr_bytes;
260 unsigned int nr_pages;
264 * At this point all allocated objects are freed, and
265 * objcg->nr_charged_bytes can't have an arbitrary byte value.
266 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
268 * The following sequence can lead to it:
269 * 1) CPU0: objcg == stock->cached_objcg
270 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
271 * PAGE_SIZE bytes are charged
272 * 3) CPU1: a process from another memcg is allocating something,
273 * the stock if flushed,
274 * objcg->nr_charged_bytes = PAGE_SIZE - 92
275 * 5) CPU0: we do release this object,
276 * 92 bytes are added to stock->nr_bytes
277 * 6) CPU0: stock is flushed,
278 * 92 bytes are added to objcg->nr_charged_bytes
280 * In the result, nr_charged_bytes == PAGE_SIZE.
281 * This page will be uncharged in obj_cgroup_release().
283 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
284 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
285 nr_pages = nr_bytes >> PAGE_SHIFT;
287 spin_lock_irqsave(&objcg_lock, flags);
288 memcg = obj_cgroup_memcg(objcg);
290 __memcg_kmem_uncharge(memcg, nr_pages);
291 list_del(&objcg->list);
292 mem_cgroup_put(memcg);
293 spin_unlock_irqrestore(&objcg_lock, flags);
295 percpu_ref_exit(ref);
296 kfree_rcu(objcg, rcu);
299 static struct obj_cgroup *obj_cgroup_alloc(void)
301 struct obj_cgroup *objcg;
304 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
308 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
314 INIT_LIST_HEAD(&objcg->list);
318 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
319 struct mem_cgroup *parent)
321 struct obj_cgroup *objcg, *iter;
323 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
325 spin_lock_irq(&objcg_lock);
327 /* Move active objcg to the parent's list */
328 xchg(&objcg->memcg, parent);
329 css_get(&parent->css);
330 list_add(&objcg->list, &parent->objcg_list);
332 /* Move already reparented objcgs to the parent's list */
333 list_for_each_entry(iter, &memcg->objcg_list, list) {
334 css_get(&parent->css);
335 xchg(&iter->memcg, parent);
336 css_put(&memcg->css);
338 list_splice(&memcg->objcg_list, &parent->objcg_list);
340 spin_unlock_irq(&objcg_lock);
342 percpu_ref_kill(&objcg->refcnt);
346 * This will be used as a shrinker list's index.
347 * The main reason for not using cgroup id for this:
348 * this works better in sparse environments, where we have a lot of memcgs,
349 * but only a few kmem-limited. Or also, if we have, for instance, 200
350 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
351 * 200 entry array for that.
353 * The current size of the caches array is stored in memcg_nr_cache_ids. It
354 * will double each time we have to increase it.
356 static DEFINE_IDA(memcg_cache_ida);
357 int memcg_nr_cache_ids;
359 /* Protects memcg_nr_cache_ids */
360 static DECLARE_RWSEM(memcg_cache_ids_sem);
362 void memcg_get_cache_ids(void)
364 down_read(&memcg_cache_ids_sem);
367 void memcg_put_cache_ids(void)
369 up_read(&memcg_cache_ids_sem);
373 * MIN_SIZE is different than 1, because we would like to avoid going through
374 * the alloc/free process all the time. In a small machine, 4 kmem-limited
375 * cgroups is a reasonable guess. In the future, it could be a parameter or
376 * tunable, but that is strictly not necessary.
378 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
379 * this constant directly from cgroup, but it is understandable that this is
380 * better kept as an internal representation in cgroup.c. In any case, the
381 * cgrp_id space is not getting any smaller, and we don't have to necessarily
382 * increase ours as well if it increases.
384 #define MEMCG_CACHES_MIN_SIZE 4
385 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
388 * A lot of the calls to the cache allocation functions are expected to be
389 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
390 * conditional to this static branch, we'll have to allow modules that does
391 * kmem_cache_alloc and the such to see this symbol as well
393 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
394 EXPORT_SYMBOL(memcg_kmem_enabled_key);
397 static int memcg_shrinker_map_size;
398 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
400 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
402 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
405 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
406 int size, int old_size)
408 struct memcg_shrinker_map *new, *old;
411 lockdep_assert_held(&memcg_shrinker_map_mutex);
414 old = rcu_dereference_protected(
415 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
416 /* Not yet online memcg */
420 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
424 /* Set all old bits, clear all new bits */
425 memset(new->map, (int)0xff, old_size);
426 memset((void *)new->map + old_size, 0, size - old_size);
428 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
429 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
435 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
437 struct mem_cgroup_per_node *pn;
438 struct memcg_shrinker_map *map;
441 if (mem_cgroup_is_root(memcg))
445 pn = mem_cgroup_nodeinfo(memcg, nid);
446 map = rcu_dereference_protected(pn->shrinker_map, true);
449 rcu_assign_pointer(pn->shrinker_map, NULL);
453 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
455 struct memcg_shrinker_map *map;
456 int nid, size, ret = 0;
458 if (mem_cgroup_is_root(memcg))
461 mutex_lock(&memcg_shrinker_map_mutex);
462 size = memcg_shrinker_map_size;
464 map = kvzalloc_node(sizeof(*map) + size, GFP_KERNEL, nid);
466 memcg_free_shrinker_maps(memcg);
470 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
472 mutex_unlock(&memcg_shrinker_map_mutex);
477 int memcg_expand_shrinker_maps(int new_id)
479 int size, old_size, ret = 0;
480 struct mem_cgroup *memcg;
482 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
483 old_size = memcg_shrinker_map_size;
484 if (size <= old_size)
487 mutex_lock(&memcg_shrinker_map_mutex);
488 if (!root_mem_cgroup)
491 for_each_mem_cgroup(memcg) {
492 if (mem_cgroup_is_root(memcg))
494 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
496 mem_cgroup_iter_break(NULL, memcg);
502 memcg_shrinker_map_size = size;
503 mutex_unlock(&memcg_shrinker_map_mutex);
507 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
509 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
510 struct memcg_shrinker_map *map;
513 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
514 /* Pairs with smp mb in shrink_slab() */
515 smp_mb__before_atomic();
516 set_bit(shrinker_id, map->map);
522 * mem_cgroup_css_from_page - css of the memcg associated with a page
523 * @page: page of interest
525 * If memcg is bound to the default hierarchy, css of the memcg associated
526 * with @page is returned. The returned css remains associated with @page
527 * until it is released.
529 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
532 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
534 struct mem_cgroup *memcg;
536 memcg = page->mem_cgroup;
538 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
539 memcg = root_mem_cgroup;
545 * page_cgroup_ino - return inode number of the memcg a page is charged to
548 * Look up the closest online ancestor of the memory cgroup @page is charged to
549 * and return its inode number or 0 if @page is not charged to any cgroup. It
550 * is safe to call this function without holding a reference to @page.
552 * Note, this function is inherently racy, because there is nothing to prevent
553 * the cgroup inode from getting torn down and potentially reallocated a moment
554 * after page_cgroup_ino() returns, so it only should be used by callers that
555 * do not care (such as procfs interfaces).
557 ino_t page_cgroup_ino(struct page *page)
559 struct mem_cgroup *memcg;
560 unsigned long ino = 0;
563 memcg = page->mem_cgroup;
566 * The lowest bit set means that memcg isn't a valid
567 * memcg pointer, but a obj_cgroups pointer.
568 * In this case the page is shared and doesn't belong
569 * to any specific memory cgroup.
571 if ((unsigned long) memcg & 0x1UL)
574 while (memcg && !(memcg->css.flags & CSS_ONLINE))
575 memcg = parent_mem_cgroup(memcg);
577 ino = cgroup_ino(memcg->css.cgroup);
582 static struct mem_cgroup_per_node *
583 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
585 int nid = page_to_nid(page);
587 return memcg->nodeinfo[nid];
590 static struct mem_cgroup_tree_per_node *
591 soft_limit_tree_node(int nid)
593 return soft_limit_tree.rb_tree_per_node[nid];
596 static struct mem_cgroup_tree_per_node *
597 soft_limit_tree_from_page(struct page *page)
599 int nid = page_to_nid(page);
601 return soft_limit_tree.rb_tree_per_node[nid];
604 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
605 struct mem_cgroup_tree_per_node *mctz,
606 unsigned long new_usage_in_excess)
608 struct rb_node **p = &mctz->rb_root.rb_node;
609 struct rb_node *parent = NULL;
610 struct mem_cgroup_per_node *mz_node;
611 bool rightmost = true;
616 mz->usage_in_excess = new_usage_in_excess;
617 if (!mz->usage_in_excess)
621 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
623 if (mz->usage_in_excess < mz_node->usage_in_excess) {
629 * We can't avoid mem cgroups that are over their soft
630 * limit by the same amount
632 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
637 mctz->rb_rightmost = &mz->tree_node;
639 rb_link_node(&mz->tree_node, parent, p);
640 rb_insert_color(&mz->tree_node, &mctz->rb_root);
644 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
645 struct mem_cgroup_tree_per_node *mctz)
650 if (&mz->tree_node == mctz->rb_rightmost)
651 mctz->rb_rightmost = rb_prev(&mz->tree_node);
653 rb_erase(&mz->tree_node, &mctz->rb_root);
657 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
658 struct mem_cgroup_tree_per_node *mctz)
662 spin_lock_irqsave(&mctz->lock, flags);
663 __mem_cgroup_remove_exceeded(mz, mctz);
664 spin_unlock_irqrestore(&mctz->lock, flags);
667 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
669 unsigned long nr_pages = page_counter_read(&memcg->memory);
670 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
671 unsigned long excess = 0;
673 if (nr_pages > soft_limit)
674 excess = nr_pages - soft_limit;
679 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
681 unsigned long excess;
682 struct mem_cgroup_per_node *mz;
683 struct mem_cgroup_tree_per_node *mctz;
685 mctz = soft_limit_tree_from_page(page);
689 * Necessary to update all ancestors when hierarchy is used.
690 * because their event counter is not touched.
692 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
693 mz = mem_cgroup_page_nodeinfo(memcg, page);
694 excess = soft_limit_excess(memcg);
696 * We have to update the tree if mz is on RB-tree or
697 * mem is over its softlimit.
699 if (excess || mz->on_tree) {
702 spin_lock_irqsave(&mctz->lock, flags);
703 /* if on-tree, remove it */
705 __mem_cgroup_remove_exceeded(mz, mctz);
707 * Insert again. mz->usage_in_excess will be updated.
708 * If excess is 0, no tree ops.
710 __mem_cgroup_insert_exceeded(mz, mctz, excess);
711 spin_unlock_irqrestore(&mctz->lock, flags);
716 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
718 struct mem_cgroup_tree_per_node *mctz;
719 struct mem_cgroup_per_node *mz;
723 mz = mem_cgroup_nodeinfo(memcg, nid);
724 mctz = soft_limit_tree_node(nid);
726 mem_cgroup_remove_exceeded(mz, mctz);
730 static struct mem_cgroup_per_node *
731 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
733 struct mem_cgroup_per_node *mz;
737 if (!mctz->rb_rightmost)
738 goto done; /* Nothing to reclaim from */
740 mz = rb_entry(mctz->rb_rightmost,
741 struct mem_cgroup_per_node, tree_node);
743 * Remove the node now but someone else can add it back,
744 * we will to add it back at the end of reclaim to its correct
745 * position in the tree.
747 __mem_cgroup_remove_exceeded(mz, mctz);
748 if (!soft_limit_excess(mz->memcg) ||
749 !css_tryget(&mz->memcg->css))
755 static struct mem_cgroup_per_node *
756 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
758 struct mem_cgroup_per_node *mz;
760 spin_lock_irq(&mctz->lock);
761 mz = __mem_cgroup_largest_soft_limit_node(mctz);
762 spin_unlock_irq(&mctz->lock);
767 * __mod_memcg_state - update cgroup memory statistics
768 * @memcg: the memory cgroup
769 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
770 * @val: delta to add to the counter, can be negative
772 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
774 long x, threshold = MEMCG_CHARGE_BATCH;
776 if (mem_cgroup_disabled())
779 if (memcg_stat_item_in_bytes(idx))
780 threshold <<= PAGE_SHIFT;
782 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
783 if (unlikely(abs(x) > threshold)) {
784 struct mem_cgroup *mi;
787 * Batch local counters to keep them in sync with
788 * the hierarchical ones.
790 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
791 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
792 atomic_long_add(x, &mi->vmstats[idx]);
795 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
798 static struct mem_cgroup_per_node *
799 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
801 struct mem_cgroup *parent;
803 parent = parent_mem_cgroup(pn->memcg);
806 return mem_cgroup_nodeinfo(parent, nid);
809 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
812 struct mem_cgroup_per_node *pn;
813 struct mem_cgroup *memcg;
814 long x, threshold = MEMCG_CHARGE_BATCH;
816 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
820 __mod_memcg_state(memcg, idx, val);
823 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
825 if (vmstat_item_in_bytes(idx))
826 threshold <<= PAGE_SHIFT;
828 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
829 if (unlikely(abs(x) > threshold)) {
830 pg_data_t *pgdat = lruvec_pgdat(lruvec);
831 struct mem_cgroup_per_node *pi;
833 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
834 atomic_long_add(x, &pi->lruvec_stat[idx]);
837 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
841 * __mod_lruvec_state - update lruvec memory statistics
842 * @lruvec: the lruvec
843 * @idx: the stat item
844 * @val: delta to add to the counter, can be negative
846 * The lruvec is the intersection of the NUMA node and a cgroup. This
847 * function updates the all three counters that are affected by a
848 * change of state at this level: per-node, per-cgroup, per-lruvec.
850 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
854 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
856 /* Update memcg and lruvec */
857 if (!mem_cgroup_disabled())
858 __mod_memcg_lruvec_state(lruvec, idx, val);
861 void __mod_lruvec_slab_state(void *p, enum node_stat_item idx, int val)
863 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
864 struct mem_cgroup *memcg;
865 struct lruvec *lruvec;
868 memcg = mem_cgroup_from_obj(p);
871 * Untracked pages have no memcg, no lruvec. Update only the
872 * node. If we reparent the slab objects to the root memcg,
873 * when we free the slab object, we need to update the per-memcg
874 * vmstats to keep it correct for the root memcg.
877 __mod_node_page_state(pgdat, idx, val);
879 lruvec = mem_cgroup_lruvec(memcg, pgdat);
880 __mod_lruvec_state(lruvec, idx, val);
885 void mod_memcg_obj_state(void *p, int idx, int val)
887 struct mem_cgroup *memcg;
890 memcg = mem_cgroup_from_obj(p);
892 mod_memcg_state(memcg, idx, val);
897 * __count_memcg_events - account VM events in a cgroup
898 * @memcg: the memory cgroup
899 * @idx: the event item
900 * @count: the number of events that occured
902 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
907 if (mem_cgroup_disabled())
910 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
911 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
912 struct mem_cgroup *mi;
915 * Batch local counters to keep them in sync with
916 * the hierarchical ones.
918 __this_cpu_add(memcg->vmstats_local->events[idx], x);
919 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
920 atomic_long_add(x, &mi->vmevents[idx]);
923 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
926 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
928 return atomic_long_read(&memcg->vmevents[event]);
931 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
936 for_each_possible_cpu(cpu)
937 x += per_cpu(memcg->vmstats_local->events[event], cpu);
941 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
945 /* pagein of a big page is an event. So, ignore page size */
947 __count_memcg_events(memcg, PGPGIN, 1);
949 __count_memcg_events(memcg, PGPGOUT, 1);
950 nr_pages = -nr_pages; /* for event */
953 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
956 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
957 enum mem_cgroup_events_target target)
959 unsigned long val, next;
961 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
962 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
963 /* from time_after() in jiffies.h */
964 if ((long)(next - val) < 0) {
966 case MEM_CGROUP_TARGET_THRESH:
967 next = val + THRESHOLDS_EVENTS_TARGET;
969 case MEM_CGROUP_TARGET_SOFTLIMIT:
970 next = val + SOFTLIMIT_EVENTS_TARGET;
975 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
982 * Check events in order.
985 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
987 /* threshold event is triggered in finer grain than soft limit */
988 if (unlikely(mem_cgroup_event_ratelimit(memcg,
989 MEM_CGROUP_TARGET_THRESH))) {
992 do_softlimit = mem_cgroup_event_ratelimit(memcg,
993 MEM_CGROUP_TARGET_SOFTLIMIT);
994 mem_cgroup_threshold(memcg);
995 if (unlikely(do_softlimit))
996 mem_cgroup_update_tree(memcg, page);
1000 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1003 * mm_update_next_owner() may clear mm->owner to NULL
1004 * if it races with swapoff, page migration, etc.
1005 * So this can be called with p == NULL.
1010 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1012 EXPORT_SYMBOL(mem_cgroup_from_task);
1015 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1016 * @mm: mm from which memcg should be extracted. It can be NULL.
1018 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
1019 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
1022 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1024 struct mem_cgroup *memcg;
1026 if (mem_cgroup_disabled())
1032 * Page cache insertions can happen withou an
1033 * actual mm context, e.g. during disk probing
1034 * on boot, loopback IO, acct() writes etc.
1037 memcg = root_mem_cgroup;
1039 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1040 if (unlikely(!memcg))
1041 memcg = root_mem_cgroup;
1043 } while (!css_tryget(&memcg->css));
1047 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1050 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
1051 * @page: page from which memcg should be extracted.
1053 * Obtain a reference on page->memcg and returns it if successful. Otherwise
1054 * root_mem_cgroup is returned.
1056 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
1058 struct mem_cgroup *memcg = page->mem_cgroup;
1060 if (mem_cgroup_disabled())
1064 /* Page should not get uncharged and freed memcg under us. */
1065 if (!memcg || WARN_ON_ONCE(!css_tryget(&memcg->css)))
1066 memcg = root_mem_cgroup;
1070 EXPORT_SYMBOL(get_mem_cgroup_from_page);
1072 static __always_inline struct mem_cgroup *active_memcg(void)
1075 return this_cpu_read(int_active_memcg);
1077 return current->active_memcg;
1080 static __always_inline struct mem_cgroup *get_active_memcg(void)
1082 struct mem_cgroup *memcg;
1085 memcg = active_memcg();
1086 /* remote memcg must hold a ref. */
1087 if (memcg && WARN_ON_ONCE(!css_tryget(&memcg->css)))
1088 memcg = root_mem_cgroup;
1094 static __always_inline bool memcg_kmem_bypass(void)
1096 /* Allow remote memcg charging from any context. */
1097 if (unlikely(active_memcg()))
1100 /* Memcg to charge can't be determined. */
1101 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
1108 * If active memcg is set, do not fallback to current->mm->memcg.
1110 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
1112 if (memcg_kmem_bypass())
1115 if (unlikely(active_memcg()))
1116 return get_active_memcg();
1118 return get_mem_cgroup_from_mm(current->mm);
1122 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1123 * @root: hierarchy root
1124 * @prev: previously returned memcg, NULL on first invocation
1125 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1127 * Returns references to children of the hierarchy below @root, or
1128 * @root itself, or %NULL after a full round-trip.
1130 * Caller must pass the return value in @prev on subsequent
1131 * invocations for reference counting, or use mem_cgroup_iter_break()
1132 * to cancel a hierarchy walk before the round-trip is complete.
1134 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1135 * in the hierarchy among all concurrent reclaimers operating on the
1138 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1139 struct mem_cgroup *prev,
1140 struct mem_cgroup_reclaim_cookie *reclaim)
1142 struct mem_cgroup_reclaim_iter *iter;
1143 struct cgroup_subsys_state *css = NULL;
1144 struct mem_cgroup *memcg = NULL;
1145 struct mem_cgroup *pos = NULL;
1147 if (mem_cgroup_disabled())
1151 root = root_mem_cgroup;
1153 if (prev && !reclaim)
1156 if (!root->use_hierarchy && root != root_mem_cgroup) {
1165 struct mem_cgroup_per_node *mz;
1167 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1170 if (prev && reclaim->generation != iter->generation)
1174 pos = READ_ONCE(iter->position);
1175 if (!pos || css_tryget(&pos->css))
1178 * css reference reached zero, so iter->position will
1179 * be cleared by ->css_released. However, we should not
1180 * rely on this happening soon, because ->css_released
1181 * is called from a work queue, and by busy-waiting we
1182 * might block it. So we clear iter->position right
1185 (void)cmpxchg(&iter->position, pos, NULL);
1193 css = css_next_descendant_pre(css, &root->css);
1196 * Reclaimers share the hierarchy walk, and a
1197 * new one might jump in right at the end of
1198 * the hierarchy - make sure they see at least
1199 * one group and restart from the beginning.
1207 * Verify the css and acquire a reference. The root
1208 * is provided by the caller, so we know it's alive
1209 * and kicking, and don't take an extra reference.
1211 memcg = mem_cgroup_from_css(css);
1213 if (css == &root->css)
1216 if (css_tryget(css))
1224 * The position could have already been updated by a competing
1225 * thread, so check that the value hasn't changed since we read
1226 * it to avoid reclaiming from the same cgroup twice.
1228 (void)cmpxchg(&iter->position, pos, memcg);
1236 reclaim->generation = iter->generation;
1242 if (prev && prev != root)
1243 css_put(&prev->css);
1249 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1250 * @root: hierarchy root
1251 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1253 void mem_cgroup_iter_break(struct mem_cgroup *root,
1254 struct mem_cgroup *prev)
1257 root = root_mem_cgroup;
1258 if (prev && prev != root)
1259 css_put(&prev->css);
1262 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1263 struct mem_cgroup *dead_memcg)
1265 struct mem_cgroup_reclaim_iter *iter;
1266 struct mem_cgroup_per_node *mz;
1269 for_each_node(nid) {
1270 mz = mem_cgroup_nodeinfo(from, nid);
1272 cmpxchg(&iter->position, dead_memcg, NULL);
1276 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1278 struct mem_cgroup *memcg = dead_memcg;
1279 struct mem_cgroup *last;
1282 __invalidate_reclaim_iterators(memcg, dead_memcg);
1284 } while ((memcg = parent_mem_cgroup(memcg)));
1287 * When cgruop1 non-hierarchy mode is used,
1288 * parent_mem_cgroup() does not walk all the way up to the
1289 * cgroup root (root_mem_cgroup). So we have to handle
1290 * dead_memcg from cgroup root separately.
1292 if (last != root_mem_cgroup)
1293 __invalidate_reclaim_iterators(root_mem_cgroup,
1298 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1299 * @memcg: hierarchy root
1300 * @fn: function to call for each task
1301 * @arg: argument passed to @fn
1303 * This function iterates over tasks attached to @memcg or to any of its
1304 * descendants and calls @fn for each task. If @fn returns a non-zero
1305 * value, the function breaks the iteration loop and returns the value.
1306 * Otherwise, it will iterate over all tasks and return 0.
1308 * This function must not be called for the root memory cgroup.
1310 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1311 int (*fn)(struct task_struct *, void *), void *arg)
1313 struct mem_cgroup *iter;
1316 BUG_ON(memcg == root_mem_cgroup);
1318 for_each_mem_cgroup_tree(iter, memcg) {
1319 struct css_task_iter it;
1320 struct task_struct *task;
1322 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1323 while (!ret && (task = css_task_iter_next(&it)))
1324 ret = fn(task, arg);
1325 css_task_iter_end(&it);
1327 mem_cgroup_iter_break(memcg, iter);
1335 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1337 * @pgdat: pgdat of the page
1339 * This function relies on page->mem_cgroup being stable - see the
1340 * access rules in commit_charge().
1342 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1344 struct mem_cgroup_per_node *mz;
1345 struct mem_cgroup *memcg;
1346 struct lruvec *lruvec;
1348 if (mem_cgroup_disabled()) {
1349 lruvec = &pgdat->__lruvec;
1353 memcg = page->mem_cgroup;
1355 * Swapcache readahead pages are added to the LRU - and
1356 * possibly migrated - before they are charged.
1359 memcg = root_mem_cgroup;
1361 mz = mem_cgroup_page_nodeinfo(memcg, page);
1362 lruvec = &mz->lruvec;
1365 * Since a node can be onlined after the mem_cgroup was created,
1366 * we have to be prepared to initialize lruvec->zone here;
1367 * and if offlined then reonlined, we need to reinitialize it.
1369 if (unlikely(lruvec->pgdat != pgdat))
1370 lruvec->pgdat = pgdat;
1375 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1376 * @lruvec: mem_cgroup per zone lru vector
1377 * @lru: index of lru list the page is sitting on
1378 * @zid: zone id of the accounted pages
1379 * @nr_pages: positive when adding or negative when removing
1381 * This function must be called under lru_lock, just before a page is added
1382 * to or just after a page is removed from an lru list (that ordering being
1383 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1385 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1386 int zid, int nr_pages)
1388 struct mem_cgroup_per_node *mz;
1389 unsigned long *lru_size;
1392 if (mem_cgroup_disabled())
1395 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1396 lru_size = &mz->lru_zone_size[zid][lru];
1399 *lru_size += nr_pages;
1402 if (WARN_ONCE(size < 0,
1403 "%s(%p, %d, %d): lru_size %ld\n",
1404 __func__, lruvec, lru, nr_pages, size)) {
1410 *lru_size += nr_pages;
1414 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1415 * @memcg: the memory cgroup
1417 * Returns the maximum amount of memory @mem can be charged with, in
1420 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1422 unsigned long margin = 0;
1423 unsigned long count;
1424 unsigned long limit;
1426 count = page_counter_read(&memcg->memory);
1427 limit = READ_ONCE(memcg->memory.max);
1429 margin = limit - count;
1431 if (do_memsw_account()) {
1432 count = page_counter_read(&memcg->memsw);
1433 limit = READ_ONCE(memcg->memsw.max);
1435 margin = min(margin, limit - count);
1444 * A routine for checking "mem" is under move_account() or not.
1446 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1447 * moving cgroups. This is for waiting at high-memory pressure
1450 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1452 struct mem_cgroup *from;
1453 struct mem_cgroup *to;
1456 * Unlike task_move routines, we access mc.to, mc.from not under
1457 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1459 spin_lock(&mc.lock);
1465 ret = mem_cgroup_is_descendant(from, memcg) ||
1466 mem_cgroup_is_descendant(to, memcg);
1468 spin_unlock(&mc.lock);
1472 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1474 if (mc.moving_task && current != mc.moving_task) {
1475 if (mem_cgroup_under_move(memcg)) {
1477 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1478 /* moving charge context might have finished. */
1481 finish_wait(&mc.waitq, &wait);
1488 struct memory_stat {
1494 static struct memory_stat memory_stats[] = {
1495 { "anon", PAGE_SIZE, NR_ANON_MAPPED },
1496 { "file", PAGE_SIZE, NR_FILE_PAGES },
1497 { "kernel_stack", 1024, NR_KERNEL_STACK_KB },
1498 { "percpu", 1, MEMCG_PERCPU_B },
1499 { "sock", PAGE_SIZE, MEMCG_SOCK },
1500 { "shmem", PAGE_SIZE, NR_SHMEM },
1501 { "file_mapped", PAGE_SIZE, NR_FILE_MAPPED },
1502 { "file_dirty", PAGE_SIZE, NR_FILE_DIRTY },
1503 { "file_writeback", PAGE_SIZE, NR_WRITEBACK },
1504 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1506 * The ratio will be initialized in memory_stats_init(). Because
1507 * on some architectures, the macro of HPAGE_PMD_SIZE is not
1508 * constant(e.g. powerpc).
1510 { "anon_thp", 0, NR_ANON_THPS },
1512 { "inactive_anon", PAGE_SIZE, NR_INACTIVE_ANON },
1513 { "active_anon", PAGE_SIZE, NR_ACTIVE_ANON },
1514 { "inactive_file", PAGE_SIZE, NR_INACTIVE_FILE },
1515 { "active_file", PAGE_SIZE, NR_ACTIVE_FILE },
1516 { "unevictable", PAGE_SIZE, NR_UNEVICTABLE },
1519 * Note: The slab_reclaimable and slab_unreclaimable must be
1520 * together and slab_reclaimable must be in front.
1522 { "slab_reclaimable", 1, NR_SLAB_RECLAIMABLE_B },
1523 { "slab_unreclaimable", 1, NR_SLAB_UNRECLAIMABLE_B },
1525 /* The memory events */
1526 { "workingset_refault_anon", 1, WORKINGSET_REFAULT_ANON },
1527 { "workingset_refault_file", 1, WORKINGSET_REFAULT_FILE },
1528 { "workingset_activate_anon", 1, WORKINGSET_ACTIVATE_ANON },
1529 { "workingset_activate_file", 1, WORKINGSET_ACTIVATE_FILE },
1530 { "workingset_restore_anon", 1, WORKINGSET_RESTORE_ANON },
1531 { "workingset_restore_file", 1, WORKINGSET_RESTORE_FILE },
1532 { "workingset_nodereclaim", 1, WORKINGSET_NODERECLAIM },
1535 static int __init memory_stats_init(void)
1539 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1540 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1541 if (memory_stats[i].idx == NR_ANON_THPS)
1542 memory_stats[i].ratio = HPAGE_PMD_SIZE;
1544 VM_BUG_ON(!memory_stats[i].ratio);
1545 VM_BUG_ON(memory_stats[i].idx >= MEMCG_NR_STAT);
1550 pure_initcall(memory_stats_init);
1552 static char *memory_stat_format(struct mem_cgroup *memcg)
1557 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1562 * Provide statistics on the state of the memory subsystem as
1563 * well as cumulative event counters that show past behavior.
1565 * This list is ordered following a combination of these gradients:
1566 * 1) generic big picture -> specifics and details
1567 * 2) reflecting userspace activity -> reflecting kernel heuristics
1569 * Current memory state:
1572 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1575 size = memcg_page_state(memcg, memory_stats[i].idx);
1576 size *= memory_stats[i].ratio;
1577 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1579 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1580 size = memcg_page_state(memcg, NR_SLAB_RECLAIMABLE_B) +
1581 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE_B);
1582 seq_buf_printf(&s, "slab %llu\n", size);
1586 /* Accumulated memory events */
1588 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1589 memcg_events(memcg, PGFAULT));
1590 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1591 memcg_events(memcg, PGMAJFAULT));
1592 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1593 memcg_events(memcg, PGREFILL));
1594 seq_buf_printf(&s, "pgscan %lu\n",
1595 memcg_events(memcg, PGSCAN_KSWAPD) +
1596 memcg_events(memcg, PGSCAN_DIRECT));
1597 seq_buf_printf(&s, "pgsteal %lu\n",
1598 memcg_events(memcg, PGSTEAL_KSWAPD) +
1599 memcg_events(memcg, PGSTEAL_DIRECT));
1600 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1601 memcg_events(memcg, PGACTIVATE));
1602 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1603 memcg_events(memcg, PGDEACTIVATE));
1604 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1605 memcg_events(memcg, PGLAZYFREE));
1606 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1607 memcg_events(memcg, PGLAZYFREED));
1609 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1610 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1611 memcg_events(memcg, THP_FAULT_ALLOC));
1612 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1613 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1614 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1616 /* The above should easily fit into one page */
1617 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1622 #define K(x) ((x) << (PAGE_SHIFT-10))
1624 * mem_cgroup_print_oom_context: Print OOM information relevant to
1625 * memory controller.
1626 * @memcg: The memory cgroup that went over limit
1627 * @p: Task that is going to be killed
1629 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1632 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1637 pr_cont(",oom_memcg=");
1638 pr_cont_cgroup_path(memcg->css.cgroup);
1640 pr_cont(",global_oom");
1642 pr_cont(",task_memcg=");
1643 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1649 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1650 * memory controller.
1651 * @memcg: The memory cgroup that went over limit
1653 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1657 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1658 K((u64)page_counter_read(&memcg->memory)),
1659 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1660 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1661 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1662 K((u64)page_counter_read(&memcg->swap)),
1663 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1665 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1666 K((u64)page_counter_read(&memcg->memsw)),
1667 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1668 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1669 K((u64)page_counter_read(&memcg->kmem)),
1670 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1673 pr_info("Memory cgroup stats for ");
1674 pr_cont_cgroup_path(memcg->css.cgroup);
1676 buf = memory_stat_format(memcg);
1684 * Return the memory (and swap, if configured) limit for a memcg.
1686 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1688 unsigned long max = READ_ONCE(memcg->memory.max);
1690 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1691 if (mem_cgroup_swappiness(memcg))
1692 max += min(READ_ONCE(memcg->swap.max),
1693 (unsigned long)total_swap_pages);
1695 if (mem_cgroup_swappiness(memcg)) {
1696 /* Calculate swap excess capacity from memsw limit */
1697 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1699 max += min(swap, (unsigned long)total_swap_pages);
1705 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1707 return page_counter_read(&memcg->memory);
1710 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1713 struct oom_control oc = {
1717 .gfp_mask = gfp_mask,
1722 if (mutex_lock_killable(&oom_lock))
1725 if (mem_cgroup_margin(memcg) >= (1 << order))
1729 * A few threads which were not waiting at mutex_lock_killable() can
1730 * fail to bail out. Therefore, check again after holding oom_lock.
1732 ret = task_is_dying() || out_of_memory(&oc);
1735 mutex_unlock(&oom_lock);
1739 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1742 unsigned long *total_scanned)
1744 struct mem_cgroup *victim = NULL;
1747 unsigned long excess;
1748 unsigned long nr_scanned;
1749 struct mem_cgroup_reclaim_cookie reclaim = {
1753 excess = soft_limit_excess(root_memcg);
1756 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1761 * If we have not been able to reclaim
1762 * anything, it might because there are
1763 * no reclaimable pages under this hierarchy
1768 * We want to do more targeted reclaim.
1769 * excess >> 2 is not to excessive so as to
1770 * reclaim too much, nor too less that we keep
1771 * coming back to reclaim from this cgroup
1773 if (total >= (excess >> 2) ||
1774 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1779 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1780 pgdat, &nr_scanned);
1781 *total_scanned += nr_scanned;
1782 if (!soft_limit_excess(root_memcg))
1785 mem_cgroup_iter_break(root_memcg, victim);
1789 #ifdef CONFIG_LOCKDEP
1790 static struct lockdep_map memcg_oom_lock_dep_map = {
1791 .name = "memcg_oom_lock",
1795 static DEFINE_SPINLOCK(memcg_oom_lock);
1798 * Check OOM-Killer is already running under our hierarchy.
1799 * If someone is running, return false.
1801 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1803 struct mem_cgroup *iter, *failed = NULL;
1805 spin_lock(&memcg_oom_lock);
1807 for_each_mem_cgroup_tree(iter, memcg) {
1808 if (iter->oom_lock) {
1810 * this subtree of our hierarchy is already locked
1811 * so we cannot give a lock.
1814 mem_cgroup_iter_break(memcg, iter);
1817 iter->oom_lock = true;
1822 * OK, we failed to lock the whole subtree so we have
1823 * to clean up what we set up to the failing subtree
1825 for_each_mem_cgroup_tree(iter, memcg) {
1826 if (iter == failed) {
1827 mem_cgroup_iter_break(memcg, iter);
1830 iter->oom_lock = false;
1833 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1835 spin_unlock(&memcg_oom_lock);
1840 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1842 struct mem_cgroup *iter;
1844 spin_lock(&memcg_oom_lock);
1845 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1846 for_each_mem_cgroup_tree(iter, memcg)
1847 iter->oom_lock = false;
1848 spin_unlock(&memcg_oom_lock);
1851 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1853 struct mem_cgroup *iter;
1855 spin_lock(&memcg_oom_lock);
1856 for_each_mem_cgroup_tree(iter, memcg)
1858 spin_unlock(&memcg_oom_lock);
1861 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1863 struct mem_cgroup *iter;
1866 * Be careful about under_oom underflows becase a child memcg
1867 * could have been added after mem_cgroup_mark_under_oom.
1869 spin_lock(&memcg_oom_lock);
1870 for_each_mem_cgroup_tree(iter, memcg)
1871 if (iter->under_oom > 0)
1873 spin_unlock(&memcg_oom_lock);
1876 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1878 struct oom_wait_info {
1879 struct mem_cgroup *memcg;
1880 wait_queue_entry_t wait;
1883 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1884 unsigned mode, int sync, void *arg)
1886 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1887 struct mem_cgroup *oom_wait_memcg;
1888 struct oom_wait_info *oom_wait_info;
1890 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1891 oom_wait_memcg = oom_wait_info->memcg;
1893 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1894 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1896 return autoremove_wake_function(wait, mode, sync, arg);
1899 static void memcg_oom_recover(struct mem_cgroup *memcg)
1902 * For the following lockless ->under_oom test, the only required
1903 * guarantee is that it must see the state asserted by an OOM when
1904 * this function is called as a result of userland actions
1905 * triggered by the notification of the OOM. This is trivially
1906 * achieved by invoking mem_cgroup_mark_under_oom() before
1907 * triggering notification.
1909 if (memcg && memcg->under_oom)
1910 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1920 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1922 enum oom_status ret;
1925 if (order > PAGE_ALLOC_COSTLY_ORDER)
1928 memcg_memory_event(memcg, MEMCG_OOM);
1931 * We are in the middle of the charge context here, so we
1932 * don't want to block when potentially sitting on a callstack
1933 * that holds all kinds of filesystem and mm locks.
1935 * cgroup1 allows disabling the OOM killer and waiting for outside
1936 * handling until the charge can succeed; remember the context and put
1937 * the task to sleep at the end of the page fault when all locks are
1940 * On the other hand, in-kernel OOM killer allows for an async victim
1941 * memory reclaim (oom_reaper) and that means that we are not solely
1942 * relying on the oom victim to make a forward progress and we can
1943 * invoke the oom killer here.
1945 * Please note that mem_cgroup_out_of_memory might fail to find a
1946 * victim and then we have to bail out from the charge path.
1948 if (memcg->oom_kill_disable) {
1949 if (!current->in_user_fault)
1951 css_get(&memcg->css);
1952 current->memcg_in_oom = memcg;
1953 current->memcg_oom_gfp_mask = mask;
1954 current->memcg_oom_order = order;
1959 mem_cgroup_mark_under_oom(memcg);
1961 locked = mem_cgroup_oom_trylock(memcg);
1964 mem_cgroup_oom_notify(memcg);
1966 mem_cgroup_unmark_under_oom(memcg);
1967 if (mem_cgroup_out_of_memory(memcg, mask, order))
1973 mem_cgroup_oom_unlock(memcg);
1979 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1980 * @handle: actually kill/wait or just clean up the OOM state
1982 * This has to be called at the end of a page fault if the memcg OOM
1983 * handler was enabled.
1985 * Memcg supports userspace OOM handling where failed allocations must
1986 * sleep on a waitqueue until the userspace task resolves the
1987 * situation. Sleeping directly in the charge context with all kinds
1988 * of locks held is not a good idea, instead we remember an OOM state
1989 * in the task and mem_cgroup_oom_synchronize() has to be called at
1990 * the end of the page fault to complete the OOM handling.
1992 * Returns %true if an ongoing memcg OOM situation was detected and
1993 * completed, %false otherwise.
1995 bool mem_cgroup_oom_synchronize(bool handle)
1997 struct mem_cgroup *memcg = current->memcg_in_oom;
1998 struct oom_wait_info owait;
2001 /* OOM is global, do not handle */
2008 owait.memcg = memcg;
2009 owait.wait.flags = 0;
2010 owait.wait.func = memcg_oom_wake_function;
2011 owait.wait.private = current;
2012 INIT_LIST_HEAD(&owait.wait.entry);
2014 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2015 mem_cgroup_mark_under_oom(memcg);
2017 locked = mem_cgroup_oom_trylock(memcg);
2020 mem_cgroup_oom_notify(memcg);
2022 if (locked && !memcg->oom_kill_disable) {
2023 mem_cgroup_unmark_under_oom(memcg);
2024 finish_wait(&memcg_oom_waitq, &owait.wait);
2025 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
2026 current->memcg_oom_order);
2029 mem_cgroup_unmark_under_oom(memcg);
2030 finish_wait(&memcg_oom_waitq, &owait.wait);
2034 mem_cgroup_oom_unlock(memcg);
2036 * There is no guarantee that an OOM-lock contender
2037 * sees the wakeups triggered by the OOM kill
2038 * uncharges. Wake any sleepers explicitely.
2040 memcg_oom_recover(memcg);
2043 current->memcg_in_oom = NULL;
2044 css_put(&memcg->css);
2049 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2050 * @victim: task to be killed by the OOM killer
2051 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2053 * Returns a pointer to a memory cgroup, which has to be cleaned up
2054 * by killing all belonging OOM-killable tasks.
2056 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2058 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2059 struct mem_cgroup *oom_domain)
2061 struct mem_cgroup *oom_group = NULL;
2062 struct mem_cgroup *memcg;
2064 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2068 oom_domain = root_mem_cgroup;
2072 memcg = mem_cgroup_from_task(victim);
2073 if (memcg == root_mem_cgroup)
2077 * If the victim task has been asynchronously moved to a different
2078 * memory cgroup, we might end up killing tasks outside oom_domain.
2079 * In this case it's better to ignore memory.group.oom.
2081 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2085 * Traverse the memory cgroup hierarchy from the victim task's
2086 * cgroup up to the OOMing cgroup (or root) to find the
2087 * highest-level memory cgroup with oom.group set.
2089 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2090 if (memcg->oom_group)
2093 if (memcg == oom_domain)
2098 css_get(&oom_group->css);
2105 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2107 pr_info("Tasks in ");
2108 pr_cont_cgroup_path(memcg->css.cgroup);
2109 pr_cont(" are going to be killed due to memory.oom.group set\n");
2113 * lock_page_memcg - lock a page->mem_cgroup binding
2116 * This function protects unlocked LRU pages from being moved to
2119 * It ensures lifetime of the returned memcg. Caller is responsible
2120 * for the lifetime of the page; __unlock_page_memcg() is available
2121 * when @page might get freed inside the locked section.
2123 struct mem_cgroup *lock_page_memcg(struct page *page)
2125 struct page *head = compound_head(page); /* rmap on tail pages */
2126 struct mem_cgroup *memcg;
2127 unsigned long flags;
2130 * The RCU lock is held throughout the transaction. The fast
2131 * path can get away without acquiring the memcg->move_lock
2132 * because page moving starts with an RCU grace period.
2134 * The RCU lock also protects the memcg from being freed when
2135 * the page state that is going to change is the only thing
2136 * preventing the page itself from being freed. E.g. writeback
2137 * doesn't hold a page reference and relies on PG_writeback to
2138 * keep off truncation, migration and so forth.
2142 if (mem_cgroup_disabled())
2145 memcg = head->mem_cgroup;
2146 if (unlikely(!memcg))
2149 if (atomic_read(&memcg->moving_account) <= 0)
2152 spin_lock_irqsave(&memcg->move_lock, flags);
2153 if (memcg != head->mem_cgroup) {
2154 spin_unlock_irqrestore(&memcg->move_lock, flags);
2159 * When charge migration first begins, we can have locked and
2160 * unlocked page stat updates happening concurrently. Track
2161 * the task who has the lock for unlock_page_memcg().
2163 memcg->move_lock_task = current;
2164 memcg->move_lock_flags = flags;
2168 EXPORT_SYMBOL(lock_page_memcg);
2171 * __unlock_page_memcg - unlock and unpin a memcg
2174 * Unlock and unpin a memcg returned by lock_page_memcg().
2176 void __unlock_page_memcg(struct mem_cgroup *memcg)
2178 if (memcg && memcg->move_lock_task == current) {
2179 unsigned long flags = memcg->move_lock_flags;
2181 memcg->move_lock_task = NULL;
2182 memcg->move_lock_flags = 0;
2184 spin_unlock_irqrestore(&memcg->move_lock, flags);
2191 * unlock_page_memcg - unlock a page->mem_cgroup binding
2194 void unlock_page_memcg(struct page *page)
2196 struct page *head = compound_head(page);
2198 __unlock_page_memcg(head->mem_cgroup);
2200 EXPORT_SYMBOL(unlock_page_memcg);
2202 struct memcg_stock_pcp {
2203 struct mem_cgroup *cached; /* this never be root cgroup */
2204 unsigned int nr_pages;
2206 #ifdef CONFIG_MEMCG_KMEM
2207 struct obj_cgroup *cached_objcg;
2208 unsigned int nr_bytes;
2211 struct work_struct work;
2212 unsigned long flags;
2213 #define FLUSHING_CACHED_CHARGE 0
2215 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2216 static DEFINE_MUTEX(percpu_charge_mutex);
2218 #ifdef CONFIG_MEMCG_KMEM
2219 static void drain_obj_stock(struct memcg_stock_pcp *stock);
2220 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2221 struct mem_cgroup *root_memcg);
2224 static inline void drain_obj_stock(struct memcg_stock_pcp *stock)
2227 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2228 struct mem_cgroup *root_memcg)
2235 * consume_stock: Try to consume stocked charge on this cpu.
2236 * @memcg: memcg to consume from.
2237 * @nr_pages: how many pages to charge.
2239 * The charges will only happen if @memcg matches the current cpu's memcg
2240 * stock, and at least @nr_pages are available in that stock. Failure to
2241 * service an allocation will refill the stock.
2243 * returns true if successful, false otherwise.
2245 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2247 struct memcg_stock_pcp *stock;
2248 unsigned long flags;
2251 if (nr_pages > MEMCG_CHARGE_BATCH)
2254 local_irq_save(flags);
2256 stock = this_cpu_ptr(&memcg_stock);
2257 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2258 stock->nr_pages -= nr_pages;
2262 local_irq_restore(flags);
2268 * Returns stocks cached in percpu and reset cached information.
2270 static void drain_stock(struct memcg_stock_pcp *stock)
2272 struct mem_cgroup *old = stock->cached;
2277 if (stock->nr_pages) {
2278 page_counter_uncharge(&old->memory, stock->nr_pages);
2279 if (do_memsw_account())
2280 page_counter_uncharge(&old->memsw, stock->nr_pages);
2281 stock->nr_pages = 0;
2285 stock->cached = NULL;
2288 static void drain_local_stock(struct work_struct *dummy)
2290 struct memcg_stock_pcp *stock;
2291 unsigned long flags;
2294 * The only protection from memory hotplug vs. drain_stock races is
2295 * that we always operate on local CPU stock here with IRQ disabled
2297 local_irq_save(flags);
2299 stock = this_cpu_ptr(&memcg_stock);
2300 drain_obj_stock(stock);
2302 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2304 local_irq_restore(flags);
2308 * Cache charges(val) to local per_cpu area.
2309 * This will be consumed by consume_stock() function, later.
2311 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2313 struct memcg_stock_pcp *stock;
2314 unsigned long flags;
2316 local_irq_save(flags);
2318 stock = this_cpu_ptr(&memcg_stock);
2319 if (stock->cached != memcg) { /* reset if necessary */
2321 css_get(&memcg->css);
2322 stock->cached = memcg;
2324 stock->nr_pages += nr_pages;
2326 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2329 local_irq_restore(flags);
2333 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2334 * of the hierarchy under it.
2336 static void drain_all_stock(struct mem_cgroup *root_memcg)
2340 /* If someone's already draining, avoid adding running more workers. */
2341 if (!mutex_trylock(&percpu_charge_mutex))
2344 * Notify other cpus that system-wide "drain" is running
2345 * We do not care about races with the cpu hotplug because cpu down
2346 * as well as workers from this path always operate on the local
2347 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2350 for_each_online_cpu(cpu) {
2351 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2352 struct mem_cgroup *memcg;
2356 memcg = stock->cached;
2357 if (memcg && stock->nr_pages &&
2358 mem_cgroup_is_descendant(memcg, root_memcg))
2360 if (obj_stock_flush_required(stock, root_memcg))
2365 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2367 drain_local_stock(&stock->work);
2369 schedule_work_on(cpu, &stock->work);
2373 mutex_unlock(&percpu_charge_mutex);
2376 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2378 struct memcg_stock_pcp *stock;
2379 struct mem_cgroup *memcg, *mi;
2381 stock = &per_cpu(memcg_stock, cpu);
2384 for_each_mem_cgroup(memcg) {
2387 for (i = 0; i < MEMCG_NR_STAT; i++) {
2391 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2393 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2394 atomic_long_add(x, &memcg->vmstats[i]);
2396 if (i >= NR_VM_NODE_STAT_ITEMS)
2399 for_each_node(nid) {
2400 struct mem_cgroup_per_node *pn;
2402 pn = mem_cgroup_nodeinfo(memcg, nid);
2403 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2406 atomic_long_add(x, &pn->lruvec_stat[i]);
2407 } while ((pn = parent_nodeinfo(pn, nid)));
2411 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2414 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2416 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2417 atomic_long_add(x, &memcg->vmevents[i]);
2424 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2425 unsigned int nr_pages,
2428 unsigned long nr_reclaimed = 0;
2431 unsigned long pflags;
2433 if (page_counter_read(&memcg->memory) <=
2434 READ_ONCE(memcg->memory.high))
2437 memcg_memory_event(memcg, MEMCG_HIGH);
2439 psi_memstall_enter(&pflags);
2440 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2442 psi_memstall_leave(&pflags);
2443 } while ((memcg = parent_mem_cgroup(memcg)) &&
2444 !mem_cgroup_is_root(memcg));
2446 return nr_reclaimed;
2449 static void high_work_func(struct work_struct *work)
2451 struct mem_cgroup *memcg;
2453 memcg = container_of(work, struct mem_cgroup, high_work);
2454 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2458 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2459 * enough to still cause a significant slowdown in most cases, while still
2460 * allowing diagnostics and tracing to proceed without becoming stuck.
2462 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2465 * When calculating the delay, we use these either side of the exponentiation to
2466 * maintain precision and scale to a reasonable number of jiffies (see the table
2469 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2470 * overage ratio to a delay.
2471 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2472 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2473 * to produce a reasonable delay curve.
2475 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2476 * reasonable delay curve compared to precision-adjusted overage, not
2477 * penalising heavily at first, but still making sure that growth beyond the
2478 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2479 * example, with a high of 100 megabytes:
2481 * +-------+------------------------+
2482 * | usage | time to allocate in ms |
2483 * +-------+------------------------+
2505 * +-------+------------------------+
2507 #define MEMCG_DELAY_PRECISION_SHIFT 20
2508 #define MEMCG_DELAY_SCALING_SHIFT 14
2510 static u64 calculate_overage(unsigned long usage, unsigned long high)
2518 * Prevent division by 0 in overage calculation by acting as if
2519 * it was a threshold of 1 page
2521 high = max(high, 1UL);
2523 overage = usage - high;
2524 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2525 return div64_u64(overage, high);
2528 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2530 u64 overage, max_overage = 0;
2533 overage = calculate_overage(page_counter_read(&memcg->memory),
2534 READ_ONCE(memcg->memory.high));
2535 max_overage = max(overage, max_overage);
2536 } while ((memcg = parent_mem_cgroup(memcg)) &&
2537 !mem_cgroup_is_root(memcg));
2542 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2544 u64 overage, max_overage = 0;
2547 overage = calculate_overage(page_counter_read(&memcg->swap),
2548 READ_ONCE(memcg->swap.high));
2550 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2551 max_overage = max(overage, max_overage);
2552 } while ((memcg = parent_mem_cgroup(memcg)) &&
2553 !mem_cgroup_is_root(memcg));
2559 * Get the number of jiffies that we should penalise a mischievous cgroup which
2560 * is exceeding its memory.high by checking both it and its ancestors.
2562 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2563 unsigned int nr_pages,
2566 unsigned long penalty_jiffies;
2572 * We use overage compared to memory.high to calculate the number of
2573 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2574 * fairly lenient on small overages, and increasingly harsh when the
2575 * memcg in question makes it clear that it has no intention of stopping
2576 * its crazy behaviour, so we exponentially increase the delay based on
2579 penalty_jiffies = max_overage * max_overage * HZ;
2580 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2581 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2584 * Factor in the task's own contribution to the overage, such that four
2585 * N-sized allocations are throttled approximately the same as one
2586 * 4N-sized allocation.
2588 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2589 * larger the current charge patch is than that.
2591 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2595 * Scheduled by try_charge() to be executed from the userland return path
2596 * and reclaims memory over the high limit.
2598 void mem_cgroup_handle_over_high(void)
2600 unsigned long penalty_jiffies;
2601 unsigned long pflags;
2602 unsigned long nr_reclaimed;
2603 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2604 int nr_retries = MAX_RECLAIM_RETRIES;
2605 struct mem_cgroup *memcg;
2606 bool in_retry = false;
2608 if (likely(!nr_pages))
2611 memcg = get_mem_cgroup_from_mm(current->mm);
2612 current->memcg_nr_pages_over_high = 0;
2616 * The allocating task should reclaim at least the batch size, but for
2617 * subsequent retries we only want to do what's necessary to prevent oom
2618 * or breaching resource isolation.
2620 * This is distinct from memory.max or page allocator behaviour because
2621 * memory.high is currently batched, whereas memory.max and the page
2622 * allocator run every time an allocation is made.
2624 nr_reclaimed = reclaim_high(memcg,
2625 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2629 * memory.high is breached and reclaim is unable to keep up. Throttle
2630 * allocators proactively to slow down excessive growth.
2632 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2633 mem_find_max_overage(memcg));
2635 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2636 swap_find_max_overage(memcg));
2639 * Clamp the max delay per usermode return so as to still keep the
2640 * application moving forwards and also permit diagnostics, albeit
2643 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2646 * Don't sleep if the amount of jiffies this memcg owes us is so low
2647 * that it's not even worth doing, in an attempt to be nice to those who
2648 * go only a small amount over their memory.high value and maybe haven't
2649 * been aggressively reclaimed enough yet.
2651 if (penalty_jiffies <= HZ / 100)
2655 * If reclaim is making forward progress but we're still over
2656 * memory.high, we want to encourage that rather than doing allocator
2659 if (nr_reclaimed || nr_retries--) {
2665 * If we exit early, we're guaranteed to die (since
2666 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2667 * need to account for any ill-begotten jiffies to pay them off later.
2669 psi_memstall_enter(&pflags);
2670 schedule_timeout_killable(penalty_jiffies);
2671 psi_memstall_leave(&pflags);
2674 css_put(&memcg->css);
2677 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2678 unsigned int nr_pages)
2680 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2681 int nr_retries = MAX_RECLAIM_RETRIES;
2682 struct mem_cgroup *mem_over_limit;
2683 struct page_counter *counter;
2684 enum oom_status oom_status;
2685 unsigned long nr_reclaimed;
2686 bool passed_oom = false;
2687 bool may_swap = true;
2688 bool drained = false;
2689 unsigned long pflags;
2691 if (mem_cgroup_is_root(memcg))
2694 if (consume_stock(memcg, nr_pages))
2697 if (!do_memsw_account() ||
2698 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2699 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2701 if (do_memsw_account())
2702 page_counter_uncharge(&memcg->memsw, batch);
2703 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2705 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2709 if (batch > nr_pages) {
2715 * Memcg doesn't have a dedicated reserve for atomic
2716 * allocations. But like the global atomic pool, we need to
2717 * put the burden of reclaim on regular allocation requests
2718 * and let these go through as privileged allocations.
2720 if (gfp_mask & __GFP_ATOMIC)
2724 * Prevent unbounded recursion when reclaim operations need to
2725 * allocate memory. This might exceed the limits temporarily,
2726 * but we prefer facilitating memory reclaim and getting back
2727 * under the limit over triggering OOM kills in these cases.
2729 if (unlikely(current->flags & PF_MEMALLOC))
2732 if (unlikely(task_in_memcg_oom(current)))
2735 if (!gfpflags_allow_blocking(gfp_mask))
2738 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2740 psi_memstall_enter(&pflags);
2741 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2742 gfp_mask, may_swap);
2743 psi_memstall_leave(&pflags);
2745 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2749 drain_all_stock(mem_over_limit);
2754 if (gfp_mask & __GFP_NORETRY)
2757 * Even though the limit is exceeded at this point, reclaim
2758 * may have been able to free some pages. Retry the charge
2759 * before killing the task.
2761 * Only for regular pages, though: huge pages are rather
2762 * unlikely to succeed so close to the limit, and we fall back
2763 * to regular pages anyway in case of failure.
2765 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2768 * At task move, charge accounts can be doubly counted. So, it's
2769 * better to wait until the end of task_move if something is going on.
2771 if (mem_cgroup_wait_acct_move(mem_over_limit))
2777 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2780 if (gfp_mask & __GFP_NOFAIL)
2783 /* Avoid endless loop for tasks bypassed by the oom killer */
2784 if (passed_oom && task_is_dying())
2788 * keep retrying as long as the memcg oom killer is able to make
2789 * a forward progress or bypass the charge if the oom killer
2790 * couldn't make any progress.
2792 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2793 get_order(nr_pages * PAGE_SIZE));
2794 if (oom_status == OOM_SUCCESS) {
2796 nr_retries = MAX_RECLAIM_RETRIES;
2800 if (!(gfp_mask & __GFP_NOFAIL))
2804 * The allocation either can't fail or will lead to more memory
2805 * being freed very soon. Allow memory usage go over the limit
2806 * temporarily by force charging it.
2808 page_counter_charge(&memcg->memory, nr_pages);
2809 if (do_memsw_account())
2810 page_counter_charge(&memcg->memsw, nr_pages);
2815 if (batch > nr_pages)
2816 refill_stock(memcg, batch - nr_pages);
2819 * If the hierarchy is above the normal consumption range, schedule
2820 * reclaim on returning to userland. We can perform reclaim here
2821 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2822 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2823 * not recorded as it most likely matches current's and won't
2824 * change in the meantime. As high limit is checked again before
2825 * reclaim, the cost of mismatch is negligible.
2828 bool mem_high, swap_high;
2830 mem_high = page_counter_read(&memcg->memory) >
2831 READ_ONCE(memcg->memory.high);
2832 swap_high = page_counter_read(&memcg->swap) >
2833 READ_ONCE(memcg->swap.high);
2835 /* Don't bother a random interrupted task */
2836 if (in_interrupt()) {
2838 schedule_work(&memcg->high_work);
2844 if (mem_high || swap_high) {
2846 * The allocating tasks in this cgroup will need to do
2847 * reclaim or be throttled to prevent further growth
2848 * of the memory or swap footprints.
2850 * Target some best-effort fairness between the tasks,
2851 * and distribute reclaim work and delay penalties
2852 * based on how much each task is actually allocating.
2854 current->memcg_nr_pages_over_high += batch;
2855 set_notify_resume(current);
2858 } while ((memcg = parent_mem_cgroup(memcg)));
2863 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2864 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2866 if (mem_cgroup_is_root(memcg))
2869 page_counter_uncharge(&memcg->memory, nr_pages);
2870 if (do_memsw_account())
2871 page_counter_uncharge(&memcg->memsw, nr_pages);
2875 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2877 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2879 * Any of the following ensures page->mem_cgroup stability:
2883 * - lock_page_memcg()
2884 * - exclusive reference
2886 page->mem_cgroup = memcg;
2889 #ifdef CONFIG_MEMCG_KMEM
2891 * The allocated objcg pointers array is not accounted directly.
2892 * Moreover, it should not come from DMA buffer and is not readily
2893 * reclaimable. So those GFP bits should be masked off.
2895 #define OBJCGS_CLEAR_MASK (__GFP_DMA | __GFP_RECLAIMABLE | \
2896 __GFP_ACCOUNT | __GFP_NOFAIL)
2898 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2901 unsigned int objects = objs_per_slab_page(s, page);
2904 gfp &= ~OBJCGS_CLEAR_MASK;
2905 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2910 if (cmpxchg(&page->obj_cgroups, NULL,
2911 (struct obj_cgroup **) ((unsigned long)vec | 0x1UL)))
2914 kmemleak_not_leak(vec);
2920 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2922 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2923 * cgroup_mutex, etc.
2925 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2929 if (mem_cgroup_disabled())
2932 page = virt_to_head_page(p);
2935 * If page->mem_cgroup is set, it's either a simple mem_cgroup pointer
2936 * or a pointer to obj_cgroup vector. In the latter case the lowest
2937 * bit of the pointer is set.
2938 * The page->mem_cgroup pointer can be asynchronously changed
2939 * from NULL to (obj_cgroup_vec | 0x1UL), but can't be changed
2940 * from a valid memcg pointer to objcg vector or back.
2942 if (!page->mem_cgroup)
2946 * Slab objects are accounted individually, not per-page.
2947 * Memcg membership data for each individual object is saved in
2948 * the page->obj_cgroups.
2950 if (page_has_obj_cgroups(page)) {
2951 struct obj_cgroup *objcg;
2954 off = obj_to_index(page->slab_cache, page, p);
2955 objcg = page_obj_cgroups(page)[off];
2957 return obj_cgroup_memcg(objcg);
2962 /* All other pages use page->mem_cgroup */
2963 return page->mem_cgroup;
2966 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2968 struct obj_cgroup *objcg = NULL;
2969 struct mem_cgroup *memcg;
2971 if (memcg_kmem_bypass())
2975 if (unlikely(active_memcg()))
2976 memcg = active_memcg();
2978 memcg = mem_cgroup_from_task(current);
2980 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2981 objcg = rcu_dereference(memcg->objcg);
2982 if (objcg && obj_cgroup_tryget(objcg))
2991 static int memcg_alloc_cache_id(void)
2996 id = ida_simple_get(&memcg_cache_ida,
2997 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3001 if (id < memcg_nr_cache_ids)
3005 * There's no space for the new id in memcg_caches arrays,
3006 * so we have to grow them.
3008 down_write(&memcg_cache_ids_sem);
3010 size = 2 * (id + 1);
3011 if (size < MEMCG_CACHES_MIN_SIZE)
3012 size = MEMCG_CACHES_MIN_SIZE;
3013 else if (size > MEMCG_CACHES_MAX_SIZE)
3014 size = MEMCG_CACHES_MAX_SIZE;
3016 err = memcg_update_all_list_lrus(size);
3018 memcg_nr_cache_ids = size;
3020 up_write(&memcg_cache_ids_sem);
3023 ida_simple_remove(&memcg_cache_ida, id);
3029 static void memcg_free_cache_id(int id)
3031 ida_simple_remove(&memcg_cache_ida, id);
3035 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
3036 * @memcg: memory cgroup to charge
3037 * @gfp: reclaim mode
3038 * @nr_pages: number of pages to charge
3040 * Returns 0 on success, an error code on failure.
3042 int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
3043 unsigned int nr_pages)
3045 struct page_counter *counter;
3048 ret = try_charge(memcg, gfp, nr_pages);
3052 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3053 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3056 * Enforce __GFP_NOFAIL allocation because callers are not
3057 * prepared to see failures and likely do not have any failure
3060 if (gfp & __GFP_NOFAIL) {
3061 page_counter_charge(&memcg->kmem, nr_pages);
3064 cancel_charge(memcg, nr_pages);
3071 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
3072 * @memcg: memcg to uncharge
3073 * @nr_pages: number of pages to uncharge
3075 void __memcg_kmem_uncharge(struct mem_cgroup *memcg, unsigned int nr_pages)
3077 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3078 page_counter_uncharge(&memcg->kmem, nr_pages);
3080 refill_stock(memcg, nr_pages);
3084 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3085 * @page: page to charge
3086 * @gfp: reclaim mode
3087 * @order: allocation order
3089 * Returns 0 on success, an error code on failure.
3091 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3093 struct mem_cgroup *memcg;
3096 memcg = get_mem_cgroup_from_current();
3097 if (memcg && !mem_cgroup_is_root(memcg)) {
3098 ret = __memcg_kmem_charge(memcg, gfp, 1 << order);
3100 page->mem_cgroup = memcg;
3101 __SetPageKmemcg(page);
3104 css_put(&memcg->css);
3110 * __memcg_kmem_uncharge_page: uncharge a kmem page
3111 * @page: page to uncharge
3112 * @order: allocation order
3114 void __memcg_kmem_uncharge_page(struct page *page, int order)
3116 struct mem_cgroup *memcg = page->mem_cgroup;
3117 unsigned int nr_pages = 1 << order;
3122 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3123 __memcg_kmem_uncharge(memcg, nr_pages);
3124 page->mem_cgroup = NULL;
3125 css_put(&memcg->css);
3127 /* slab pages do not have PageKmemcg flag set */
3128 if (PageKmemcg(page))
3129 __ClearPageKmemcg(page);
3132 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3134 struct memcg_stock_pcp *stock;
3135 unsigned long flags;
3138 local_irq_save(flags);
3140 stock = this_cpu_ptr(&memcg_stock);
3141 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3142 stock->nr_bytes -= nr_bytes;
3146 local_irq_restore(flags);
3151 static void drain_obj_stock(struct memcg_stock_pcp *stock)
3153 struct obj_cgroup *old = stock->cached_objcg;
3158 if (stock->nr_bytes) {
3159 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3160 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3163 struct mem_cgroup *memcg;
3167 memcg = obj_cgroup_memcg(old);
3168 if (unlikely(!css_tryget(&memcg->css)))
3172 __memcg_kmem_uncharge(memcg, nr_pages);
3173 css_put(&memcg->css);
3177 * The leftover is flushed to the centralized per-memcg value.
3178 * On the next attempt to refill obj stock it will be moved
3179 * to a per-cpu stock (probably, on an other CPU), see
3180 * refill_obj_stock().
3182 * How often it's flushed is a trade-off between the memory
3183 * limit enforcement accuracy and potential CPU contention,
3184 * so it might be changed in the future.
3186 atomic_add(nr_bytes, &old->nr_charged_bytes);
3187 stock->nr_bytes = 0;
3190 obj_cgroup_put(old);
3191 stock->cached_objcg = NULL;
3194 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3195 struct mem_cgroup *root_memcg)
3197 struct mem_cgroup *memcg;
3199 if (stock->cached_objcg) {
3200 memcg = obj_cgroup_memcg(stock->cached_objcg);
3201 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3208 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3210 struct memcg_stock_pcp *stock;
3211 unsigned long flags;
3213 local_irq_save(flags);
3215 stock = this_cpu_ptr(&memcg_stock);
3216 if (stock->cached_objcg != objcg) { /* reset if necessary */
3217 drain_obj_stock(stock);
3218 obj_cgroup_get(objcg);
3219 stock->cached_objcg = objcg;
3220 stock->nr_bytes = atomic_xchg(&objcg->nr_charged_bytes, 0);
3222 stock->nr_bytes += nr_bytes;
3224 if (stock->nr_bytes > PAGE_SIZE)
3225 drain_obj_stock(stock);
3227 local_irq_restore(flags);
3230 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3232 struct mem_cgroup *memcg;
3233 unsigned int nr_pages, nr_bytes;
3236 if (consume_obj_stock(objcg, size))
3240 * In theory, memcg->nr_charged_bytes can have enough
3241 * pre-charged bytes to satisfy the allocation. However,
3242 * flushing memcg->nr_charged_bytes requires two atomic
3243 * operations, and memcg->nr_charged_bytes can't be big,
3244 * so it's better to ignore it and try grab some new pages.
3245 * memcg->nr_charged_bytes will be flushed in
3246 * refill_obj_stock(), called from this function or
3247 * independently later.
3251 memcg = obj_cgroup_memcg(objcg);
3252 if (unlikely(!css_tryget(&memcg->css)))
3256 nr_pages = size >> PAGE_SHIFT;
3257 nr_bytes = size & (PAGE_SIZE - 1);
3262 ret = __memcg_kmem_charge(memcg, gfp, nr_pages);
3263 if (!ret && nr_bytes)
3264 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes);
3266 css_put(&memcg->css);
3270 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3272 refill_obj_stock(objcg, size);
3275 #endif /* CONFIG_MEMCG_KMEM */
3278 * Because head->mem_cgroup is not set on tails, set it now.
3280 void split_page_memcg(struct page *head, unsigned int nr)
3282 struct mem_cgroup *memcg = head->mem_cgroup;
3283 int kmemcg = PageKmemcg(head);
3286 if (mem_cgroup_disabled() || !memcg)
3289 for (i = 1; i < nr; i++) {
3290 head[i].mem_cgroup = memcg;
3292 __SetPageKmemcg(head + i);
3294 css_get_many(&memcg->css, nr - 1);
3297 #ifdef CONFIG_MEMCG_SWAP
3299 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3300 * @entry: swap entry to be moved
3301 * @from: mem_cgroup which the entry is moved from
3302 * @to: mem_cgroup which the entry is moved to
3304 * It succeeds only when the swap_cgroup's record for this entry is the same
3305 * as the mem_cgroup's id of @from.
3307 * Returns 0 on success, -EINVAL on failure.
3309 * The caller must have charged to @to, IOW, called page_counter_charge() about
3310 * both res and memsw, and called css_get().
3312 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3313 struct mem_cgroup *from, struct mem_cgroup *to)
3315 unsigned short old_id, new_id;
3317 old_id = mem_cgroup_id(from);
3318 new_id = mem_cgroup_id(to);
3320 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3321 mod_memcg_state(from, MEMCG_SWAP, -1);
3322 mod_memcg_state(to, MEMCG_SWAP, 1);
3328 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3329 struct mem_cgroup *from, struct mem_cgroup *to)
3335 static DEFINE_MUTEX(memcg_max_mutex);
3337 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3338 unsigned long max, bool memsw)
3340 bool enlarge = false;
3341 bool drained = false;
3343 bool limits_invariant;
3344 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3347 if (signal_pending(current)) {
3352 mutex_lock(&memcg_max_mutex);
3354 * Make sure that the new limit (memsw or memory limit) doesn't
3355 * break our basic invariant rule memory.max <= memsw.max.
3357 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3358 max <= memcg->memsw.max;
3359 if (!limits_invariant) {
3360 mutex_unlock(&memcg_max_mutex);
3364 if (max > counter->max)
3366 ret = page_counter_set_max(counter, max);
3367 mutex_unlock(&memcg_max_mutex);
3373 drain_all_stock(memcg);
3378 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3379 GFP_KERNEL, !memsw)) {
3385 if (!ret && enlarge)
3386 memcg_oom_recover(memcg);
3391 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3393 unsigned long *total_scanned)
3395 unsigned long nr_reclaimed = 0;
3396 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3397 unsigned long reclaimed;
3399 struct mem_cgroup_tree_per_node *mctz;
3400 unsigned long excess;
3401 unsigned long nr_scanned;
3406 mctz = soft_limit_tree_node(pgdat->node_id);
3409 * Do not even bother to check the largest node if the root
3410 * is empty. Do it lockless to prevent lock bouncing. Races
3411 * are acceptable as soft limit is best effort anyway.
3413 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3417 * This loop can run a while, specially if mem_cgroup's continuously
3418 * keep exceeding their soft limit and putting the system under
3425 mz = mem_cgroup_largest_soft_limit_node(mctz);
3430 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3431 gfp_mask, &nr_scanned);
3432 nr_reclaimed += reclaimed;
3433 *total_scanned += nr_scanned;
3434 spin_lock_irq(&mctz->lock);
3435 __mem_cgroup_remove_exceeded(mz, mctz);
3438 * If we failed to reclaim anything from this memory cgroup
3439 * it is time to move on to the next cgroup
3443 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3445 excess = soft_limit_excess(mz->memcg);
3447 * One school of thought says that we should not add
3448 * back the node to the tree if reclaim returns 0.
3449 * But our reclaim could return 0, simply because due
3450 * to priority we are exposing a smaller subset of
3451 * memory to reclaim from. Consider this as a longer
3454 /* If excess == 0, no tree ops */
3455 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3456 spin_unlock_irq(&mctz->lock);
3457 css_put(&mz->memcg->css);
3460 * Could not reclaim anything and there are no more
3461 * mem cgroups to try or we seem to be looping without
3462 * reclaiming anything.
3464 if (!nr_reclaimed &&
3466 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3468 } while (!nr_reclaimed);
3470 css_put(&next_mz->memcg->css);
3471 return nr_reclaimed;
3475 * Test whether @memcg has children, dead or alive. Note that this
3476 * function doesn't care whether @memcg has use_hierarchy enabled and
3477 * returns %true if there are child csses according to the cgroup
3478 * hierarchy. Testing use_hierarchy is the caller's responsibility.
3480 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3485 ret = css_next_child(NULL, &memcg->css);
3491 * Reclaims as many pages from the given memcg as possible.
3493 * Caller is responsible for holding css reference for memcg.
3495 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3497 int nr_retries = MAX_RECLAIM_RETRIES;
3499 /* we call try-to-free pages for make this cgroup empty */
3500 lru_add_drain_all();
3502 drain_all_stock(memcg);
3504 /* try to free all pages in this cgroup */
3505 while (nr_retries && page_counter_read(&memcg->memory)) {
3508 if (signal_pending(current))
3511 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3515 /* maybe some writeback is necessary */
3516 congestion_wait(BLK_RW_ASYNC, HZ/10);
3524 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3525 char *buf, size_t nbytes,
3528 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3530 if (mem_cgroup_is_root(memcg))
3532 return mem_cgroup_force_empty(memcg) ?: nbytes;
3535 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3538 return mem_cgroup_from_css(css)->use_hierarchy;
3541 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3542 struct cftype *cft, u64 val)
3545 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3546 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3548 if (memcg->use_hierarchy == val)
3552 * If parent's use_hierarchy is set, we can't make any modifications
3553 * in the child subtrees. If it is unset, then the change can
3554 * occur, provided the current cgroup has no children.
3556 * For the root cgroup, parent_mem is NULL, we allow value to be
3557 * set if there are no children.
3559 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3560 (val == 1 || val == 0)) {
3561 if (!memcg_has_children(memcg))
3562 memcg->use_hierarchy = val;
3571 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3575 if (mem_cgroup_is_root(memcg)) {
3576 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3577 memcg_page_state(memcg, NR_ANON_MAPPED);
3579 val += memcg_page_state(memcg, MEMCG_SWAP);
3582 val = page_counter_read(&memcg->memory);
3584 val = page_counter_read(&memcg->memsw);
3597 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3600 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3601 struct page_counter *counter;
3603 switch (MEMFILE_TYPE(cft->private)) {
3605 counter = &memcg->memory;
3608 counter = &memcg->memsw;
3611 counter = &memcg->kmem;
3614 counter = &memcg->tcpmem;
3620 switch (MEMFILE_ATTR(cft->private)) {
3622 if (counter == &memcg->memory)
3623 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3624 if (counter == &memcg->memsw)
3625 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3626 return (u64)page_counter_read(counter) * PAGE_SIZE;
3628 return (u64)counter->max * PAGE_SIZE;
3630 return (u64)counter->watermark * PAGE_SIZE;
3632 return counter->failcnt;
3633 case RES_SOFT_LIMIT:
3634 return (u64)memcg->soft_limit * PAGE_SIZE;
3640 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3642 unsigned long stat[MEMCG_NR_STAT] = {0};
3643 struct mem_cgroup *mi;
3646 for_each_online_cpu(cpu)
3647 for (i = 0; i < MEMCG_NR_STAT; i++)
3648 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3650 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3651 for (i = 0; i < MEMCG_NR_STAT; i++)
3652 atomic_long_add(stat[i], &mi->vmstats[i]);
3654 for_each_node(node) {
3655 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3656 struct mem_cgroup_per_node *pi;
3658 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3661 for_each_online_cpu(cpu)
3662 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3664 pn->lruvec_stat_cpu->count[i], cpu);
3666 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3667 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3668 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3672 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3674 unsigned long events[NR_VM_EVENT_ITEMS];
3675 struct mem_cgroup *mi;
3678 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3681 for_each_online_cpu(cpu)
3682 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3683 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3686 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3687 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3688 atomic_long_add(events[i], &mi->vmevents[i]);
3691 #ifdef CONFIG_MEMCG_KMEM
3692 static int memcg_online_kmem(struct mem_cgroup *memcg)
3694 struct obj_cgroup *objcg;
3697 if (cgroup_memory_nokmem)
3700 BUG_ON(memcg->kmemcg_id >= 0);
3701 BUG_ON(memcg->kmem_state);
3703 memcg_id = memcg_alloc_cache_id();
3707 objcg = obj_cgroup_alloc();
3709 memcg_free_cache_id(memcg_id);
3712 objcg->memcg = memcg;
3713 rcu_assign_pointer(memcg->objcg, objcg);
3715 static_branch_enable(&memcg_kmem_enabled_key);
3718 * A memory cgroup is considered kmem-online as soon as it gets
3719 * kmemcg_id. Setting the id after enabling static branching will
3720 * guarantee no one starts accounting before all call sites are
3723 memcg->kmemcg_id = memcg_id;
3724 memcg->kmem_state = KMEM_ONLINE;
3729 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3731 struct cgroup_subsys_state *css;
3732 struct mem_cgroup *parent, *child;
3735 if (memcg->kmem_state != KMEM_ONLINE)
3738 memcg->kmem_state = KMEM_ALLOCATED;
3740 parent = parent_mem_cgroup(memcg);
3742 parent = root_mem_cgroup;
3744 memcg_reparent_objcgs(memcg, parent);
3746 kmemcg_id = memcg->kmemcg_id;
3747 BUG_ON(kmemcg_id < 0);
3750 * Change kmemcg_id of this cgroup and all its descendants to the
3751 * parent's id, and then move all entries from this cgroup's list_lrus
3752 * to ones of the parent. After we have finished, all list_lrus
3753 * corresponding to this cgroup are guaranteed to remain empty. The
3754 * ordering is imposed by list_lru_node->lock taken by
3755 * memcg_drain_all_list_lrus().
3757 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3758 css_for_each_descendant_pre(css, &memcg->css) {
3759 child = mem_cgroup_from_css(css);
3760 BUG_ON(child->kmemcg_id != kmemcg_id);
3761 child->kmemcg_id = parent->kmemcg_id;
3762 if (!memcg->use_hierarchy)
3767 memcg_drain_all_list_lrus(kmemcg_id, parent);
3769 memcg_free_cache_id(kmemcg_id);
3772 static void memcg_free_kmem(struct mem_cgroup *memcg)
3774 /* css_alloc() failed, offlining didn't happen */
3775 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3776 memcg_offline_kmem(memcg);
3779 static int memcg_online_kmem(struct mem_cgroup *memcg)
3783 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3786 static void memcg_free_kmem(struct mem_cgroup *memcg)
3789 #endif /* CONFIG_MEMCG_KMEM */
3791 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3796 mutex_lock(&memcg_max_mutex);
3797 ret = page_counter_set_max(&memcg->kmem, max);
3798 mutex_unlock(&memcg_max_mutex);
3802 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3806 mutex_lock(&memcg_max_mutex);
3808 ret = page_counter_set_max(&memcg->tcpmem, max);
3812 if (!memcg->tcpmem_active) {
3814 * The active flag needs to be written after the static_key
3815 * update. This is what guarantees that the socket activation
3816 * function is the last one to run. See mem_cgroup_sk_alloc()
3817 * for details, and note that we don't mark any socket as
3818 * belonging to this memcg until that flag is up.
3820 * We need to do this, because static_keys will span multiple
3821 * sites, but we can't control their order. If we mark a socket
3822 * as accounted, but the accounting functions are not patched in
3823 * yet, we'll lose accounting.
3825 * We never race with the readers in mem_cgroup_sk_alloc(),
3826 * because when this value change, the code to process it is not
3829 static_branch_inc(&memcg_sockets_enabled_key);
3830 memcg->tcpmem_active = true;
3833 mutex_unlock(&memcg_max_mutex);
3838 * The user of this function is...
3841 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3842 char *buf, size_t nbytes, loff_t off)
3844 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3845 unsigned long nr_pages;
3848 buf = strstrip(buf);
3849 ret = page_counter_memparse(buf, "-1", &nr_pages);
3853 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3855 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3859 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3861 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3864 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3867 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3868 "Please report your usecase to linux-mm@kvack.org if you "
3869 "depend on this functionality.\n");
3870 ret = memcg_update_kmem_max(memcg, nr_pages);
3873 ret = memcg_update_tcp_max(memcg, nr_pages);
3877 case RES_SOFT_LIMIT:
3878 memcg->soft_limit = nr_pages;
3882 return ret ?: nbytes;
3885 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3886 size_t nbytes, loff_t off)
3888 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3889 struct page_counter *counter;
3891 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3893 counter = &memcg->memory;
3896 counter = &memcg->memsw;
3899 counter = &memcg->kmem;
3902 counter = &memcg->tcpmem;
3908 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3910 page_counter_reset_watermark(counter);
3913 counter->failcnt = 0;
3922 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3925 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3929 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3930 struct cftype *cft, u64 val)
3932 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3934 pr_warn_once("Cgroup memory moving (move_charge_at_immigrate) is deprecated. "
3935 "Please report your usecase to linux-mm@kvack.org if you "
3936 "depend on this functionality.\n");
3938 if (val & ~MOVE_MASK)
3942 * No kind of locking is needed in here, because ->can_attach() will
3943 * check this value once in the beginning of the process, and then carry
3944 * on with stale data. This means that changes to this value will only
3945 * affect task migrations starting after the change.
3947 memcg->move_charge_at_immigrate = val;
3951 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3952 struct cftype *cft, u64 val)
3960 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3961 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3962 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3964 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3965 int nid, unsigned int lru_mask, bool tree)
3967 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3968 unsigned long nr = 0;
3971 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3974 if (!(BIT(lru) & lru_mask))
3977 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3979 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3984 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3985 unsigned int lru_mask,
3988 unsigned long nr = 0;
3992 if (!(BIT(lru) & lru_mask))
3995 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3997 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
4002 static int memcg_numa_stat_show(struct seq_file *m, void *v)
4006 unsigned int lru_mask;
4009 static const struct numa_stat stats[] = {
4010 { "total", LRU_ALL },
4011 { "file", LRU_ALL_FILE },
4012 { "anon", LRU_ALL_ANON },
4013 { "unevictable", BIT(LRU_UNEVICTABLE) },
4015 const struct numa_stat *stat;
4017 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4019 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4020 seq_printf(m, "%s=%lu", stat->name,
4021 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4023 for_each_node_state(nid, N_MEMORY)
4024 seq_printf(m, " N%d=%lu", nid,
4025 mem_cgroup_node_nr_lru_pages(memcg, nid,
4026 stat->lru_mask, false));
4030 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4032 seq_printf(m, "hierarchical_%s=%lu", stat->name,
4033 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4035 for_each_node_state(nid, N_MEMORY)
4036 seq_printf(m, " N%d=%lu", nid,
4037 mem_cgroup_node_nr_lru_pages(memcg, nid,
4038 stat->lru_mask, true));
4044 #endif /* CONFIG_NUMA */
4046 static const unsigned int memcg1_stats[] = {
4049 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4059 static const char *const memcg1_stat_names[] = {
4062 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4072 /* Universal VM events cgroup1 shows, original sort order */
4073 static const unsigned int memcg1_events[] = {
4080 static int memcg_stat_show(struct seq_file *m, void *v)
4082 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4083 unsigned long memory, memsw;
4084 struct mem_cgroup *mi;
4087 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4089 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4092 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4094 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4095 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4096 if (memcg1_stats[i] == NR_ANON_THPS)
4099 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4102 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4103 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4104 memcg_events_local(memcg, memcg1_events[i]));
4106 for (i = 0; i < NR_LRU_LISTS; i++)
4107 seq_printf(m, "%s %lu\n", lru_list_name(i),
4108 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4111 /* Hierarchical information */
4112 memory = memsw = PAGE_COUNTER_MAX;
4113 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4114 memory = min(memory, READ_ONCE(mi->memory.max));
4115 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4117 seq_printf(m, "hierarchical_memory_limit %llu\n",
4118 (u64)memory * PAGE_SIZE);
4119 if (do_memsw_account())
4120 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4121 (u64)memsw * PAGE_SIZE);
4123 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4126 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4128 nr = memcg_page_state(memcg, memcg1_stats[i]);
4129 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4130 if (memcg1_stats[i] == NR_ANON_THPS)
4133 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4134 (u64)nr * PAGE_SIZE);
4137 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4138 seq_printf(m, "total_%s %llu\n",
4139 vm_event_name(memcg1_events[i]),
4140 (u64)memcg_events(memcg, memcg1_events[i]));
4142 for (i = 0; i < NR_LRU_LISTS; i++)
4143 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4144 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4147 #ifdef CONFIG_DEBUG_VM
4150 struct mem_cgroup_per_node *mz;
4151 unsigned long anon_cost = 0;
4152 unsigned long file_cost = 0;
4154 for_each_online_pgdat(pgdat) {
4155 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
4157 anon_cost += mz->lruvec.anon_cost;
4158 file_cost += mz->lruvec.file_cost;
4160 seq_printf(m, "anon_cost %lu\n", anon_cost);
4161 seq_printf(m, "file_cost %lu\n", file_cost);
4168 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4171 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4173 return mem_cgroup_swappiness(memcg);
4176 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4177 struct cftype *cft, u64 val)
4179 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4185 memcg->swappiness = val;
4187 vm_swappiness = val;
4192 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4194 struct mem_cgroup_threshold_ary *t;
4195 unsigned long usage;
4200 t = rcu_dereference(memcg->thresholds.primary);
4202 t = rcu_dereference(memcg->memsw_thresholds.primary);
4207 usage = mem_cgroup_usage(memcg, swap);
4210 * current_threshold points to threshold just below or equal to usage.
4211 * If it's not true, a threshold was crossed after last
4212 * call of __mem_cgroup_threshold().
4214 i = t->current_threshold;
4217 * Iterate backward over array of thresholds starting from
4218 * current_threshold and check if a threshold is crossed.
4219 * If none of thresholds below usage is crossed, we read
4220 * only one element of the array here.
4222 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4223 eventfd_signal(t->entries[i].eventfd, 1);
4225 /* i = current_threshold + 1 */
4229 * Iterate forward over array of thresholds starting from
4230 * current_threshold+1 and check if a threshold is crossed.
4231 * If none of thresholds above usage is crossed, we read
4232 * only one element of the array here.
4234 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4235 eventfd_signal(t->entries[i].eventfd, 1);
4237 /* Update current_threshold */
4238 t->current_threshold = i - 1;
4243 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4246 __mem_cgroup_threshold(memcg, false);
4247 if (do_memsw_account())
4248 __mem_cgroup_threshold(memcg, true);
4250 memcg = parent_mem_cgroup(memcg);
4254 static int compare_thresholds(const void *a, const void *b)
4256 const struct mem_cgroup_threshold *_a = a;
4257 const struct mem_cgroup_threshold *_b = b;
4259 if (_a->threshold > _b->threshold)
4262 if (_a->threshold < _b->threshold)
4268 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4270 struct mem_cgroup_eventfd_list *ev;
4272 spin_lock(&memcg_oom_lock);
4274 list_for_each_entry(ev, &memcg->oom_notify, list)
4275 eventfd_signal(ev->eventfd, 1);
4277 spin_unlock(&memcg_oom_lock);
4281 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4283 struct mem_cgroup *iter;
4285 for_each_mem_cgroup_tree(iter, memcg)
4286 mem_cgroup_oom_notify_cb(iter);
4289 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4290 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4292 struct mem_cgroup_thresholds *thresholds;
4293 struct mem_cgroup_threshold_ary *new;
4294 unsigned long threshold;
4295 unsigned long usage;
4298 ret = page_counter_memparse(args, "-1", &threshold);
4302 mutex_lock(&memcg->thresholds_lock);
4305 thresholds = &memcg->thresholds;
4306 usage = mem_cgroup_usage(memcg, false);
4307 } else if (type == _MEMSWAP) {
4308 thresholds = &memcg->memsw_thresholds;
4309 usage = mem_cgroup_usage(memcg, true);
4313 /* Check if a threshold crossed before adding a new one */
4314 if (thresholds->primary)
4315 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4317 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4319 /* Allocate memory for new array of thresholds */
4320 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4327 /* Copy thresholds (if any) to new array */
4328 if (thresholds->primary)
4329 memcpy(new->entries, thresholds->primary->entries,
4330 flex_array_size(new, entries, size - 1));
4332 /* Add new threshold */
4333 new->entries[size - 1].eventfd = eventfd;
4334 new->entries[size - 1].threshold = threshold;
4336 /* Sort thresholds. Registering of new threshold isn't time-critical */
4337 sort(new->entries, size, sizeof(*new->entries),
4338 compare_thresholds, NULL);
4340 /* Find current threshold */
4341 new->current_threshold = -1;
4342 for (i = 0; i < size; i++) {
4343 if (new->entries[i].threshold <= usage) {
4345 * new->current_threshold will not be used until
4346 * rcu_assign_pointer(), so it's safe to increment
4349 ++new->current_threshold;
4354 /* Free old spare buffer and save old primary buffer as spare */
4355 kfree(thresholds->spare);
4356 thresholds->spare = thresholds->primary;
4358 rcu_assign_pointer(thresholds->primary, new);
4360 /* To be sure that nobody uses thresholds */
4364 mutex_unlock(&memcg->thresholds_lock);
4369 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4370 struct eventfd_ctx *eventfd, const char *args)
4372 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4375 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4376 struct eventfd_ctx *eventfd, const char *args)
4378 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4381 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4382 struct eventfd_ctx *eventfd, enum res_type type)
4384 struct mem_cgroup_thresholds *thresholds;
4385 struct mem_cgroup_threshold_ary *new;
4386 unsigned long usage;
4387 int i, j, size, entries;
4389 mutex_lock(&memcg->thresholds_lock);
4392 thresholds = &memcg->thresholds;
4393 usage = mem_cgroup_usage(memcg, false);
4394 } else if (type == _MEMSWAP) {
4395 thresholds = &memcg->memsw_thresholds;
4396 usage = mem_cgroup_usage(memcg, true);
4400 if (!thresholds->primary)
4403 /* Check if a threshold crossed before removing */
4404 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4406 /* Calculate new number of threshold */
4408 for (i = 0; i < thresholds->primary->size; i++) {
4409 if (thresholds->primary->entries[i].eventfd != eventfd)
4415 new = thresholds->spare;
4417 /* If no items related to eventfd have been cleared, nothing to do */
4421 /* Set thresholds array to NULL if we don't have thresholds */
4430 /* Copy thresholds and find current threshold */
4431 new->current_threshold = -1;
4432 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4433 if (thresholds->primary->entries[i].eventfd == eventfd)
4436 new->entries[j] = thresholds->primary->entries[i];
4437 if (new->entries[j].threshold <= usage) {
4439 * new->current_threshold will not be used
4440 * until rcu_assign_pointer(), so it's safe to increment
4443 ++new->current_threshold;
4449 /* Swap primary and spare array */
4450 thresholds->spare = thresholds->primary;
4452 rcu_assign_pointer(thresholds->primary, new);
4454 /* To be sure that nobody uses thresholds */
4457 /* If all events are unregistered, free the spare array */
4459 kfree(thresholds->spare);
4460 thresholds->spare = NULL;
4463 mutex_unlock(&memcg->thresholds_lock);
4466 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4467 struct eventfd_ctx *eventfd)
4469 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4472 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4473 struct eventfd_ctx *eventfd)
4475 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4478 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4479 struct eventfd_ctx *eventfd, const char *args)
4481 struct mem_cgroup_eventfd_list *event;
4483 event = kmalloc(sizeof(*event), GFP_KERNEL);
4487 spin_lock(&memcg_oom_lock);
4489 event->eventfd = eventfd;
4490 list_add(&event->list, &memcg->oom_notify);
4492 /* already in OOM ? */
4493 if (memcg->under_oom)
4494 eventfd_signal(eventfd, 1);
4495 spin_unlock(&memcg_oom_lock);
4500 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4501 struct eventfd_ctx *eventfd)
4503 struct mem_cgroup_eventfd_list *ev, *tmp;
4505 spin_lock(&memcg_oom_lock);
4507 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4508 if (ev->eventfd == eventfd) {
4509 list_del(&ev->list);
4514 spin_unlock(&memcg_oom_lock);
4517 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4519 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4521 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4522 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4523 seq_printf(sf, "oom_kill %lu\n",
4524 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4528 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4529 struct cftype *cft, u64 val)
4531 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4533 /* cannot set to root cgroup and only 0 and 1 are allowed */
4534 if (!css->parent || !((val == 0) || (val == 1)))
4537 memcg->oom_kill_disable = val;
4539 memcg_oom_recover(memcg);
4544 #ifdef CONFIG_CGROUP_WRITEBACK
4546 #include <trace/events/writeback.h>
4548 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4550 return wb_domain_init(&memcg->cgwb_domain, gfp);
4553 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4555 wb_domain_exit(&memcg->cgwb_domain);
4558 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4560 wb_domain_size_changed(&memcg->cgwb_domain);
4563 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4565 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4567 if (!memcg->css.parent)
4570 return &memcg->cgwb_domain;
4574 * idx can be of type enum memcg_stat_item or node_stat_item.
4575 * Keep in sync with memcg_exact_page().
4577 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4579 long x = atomic_long_read(&memcg->vmstats[idx]);
4582 for_each_online_cpu(cpu)
4583 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4590 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4591 * @wb: bdi_writeback in question
4592 * @pfilepages: out parameter for number of file pages
4593 * @pheadroom: out parameter for number of allocatable pages according to memcg
4594 * @pdirty: out parameter for number of dirty pages
4595 * @pwriteback: out parameter for number of pages under writeback
4597 * Determine the numbers of file, headroom, dirty, and writeback pages in
4598 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4599 * is a bit more involved.
4601 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4602 * headroom is calculated as the lowest headroom of itself and the
4603 * ancestors. Note that this doesn't consider the actual amount of
4604 * available memory in the system. The caller should further cap
4605 * *@pheadroom accordingly.
4607 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4608 unsigned long *pheadroom, unsigned long *pdirty,
4609 unsigned long *pwriteback)
4611 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4612 struct mem_cgroup *parent;
4614 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4616 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4617 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4618 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4619 *pheadroom = PAGE_COUNTER_MAX;
4621 while ((parent = parent_mem_cgroup(memcg))) {
4622 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4623 READ_ONCE(memcg->memory.high));
4624 unsigned long used = page_counter_read(&memcg->memory);
4626 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4632 * Foreign dirty flushing
4634 * There's an inherent mismatch between memcg and writeback. The former
4635 * trackes ownership per-page while the latter per-inode. This was a
4636 * deliberate design decision because honoring per-page ownership in the
4637 * writeback path is complicated, may lead to higher CPU and IO overheads
4638 * and deemed unnecessary given that write-sharing an inode across
4639 * different cgroups isn't a common use-case.
4641 * Combined with inode majority-writer ownership switching, this works well
4642 * enough in most cases but there are some pathological cases. For
4643 * example, let's say there are two cgroups A and B which keep writing to
4644 * different but confined parts of the same inode. B owns the inode and
4645 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4646 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4647 * triggering background writeback. A will be slowed down without a way to
4648 * make writeback of the dirty pages happen.
4650 * Conditions like the above can lead to a cgroup getting repatedly and
4651 * severely throttled after making some progress after each
4652 * dirty_expire_interval while the underyling IO device is almost
4655 * Solving this problem completely requires matching the ownership tracking
4656 * granularities between memcg and writeback in either direction. However,
4657 * the more egregious behaviors can be avoided by simply remembering the
4658 * most recent foreign dirtying events and initiating remote flushes on
4659 * them when local writeback isn't enough to keep the memory clean enough.
4661 * The following two functions implement such mechanism. When a foreign
4662 * page - a page whose memcg and writeback ownerships don't match - is
4663 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4664 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4665 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4666 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4667 * foreign bdi_writebacks which haven't expired. Both the numbers of
4668 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4669 * limited to MEMCG_CGWB_FRN_CNT.
4671 * The mechanism only remembers IDs and doesn't hold any object references.
4672 * As being wrong occasionally doesn't matter, updates and accesses to the
4673 * records are lockless and racy.
4675 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4676 struct bdi_writeback *wb)
4678 struct mem_cgroup *memcg = page->mem_cgroup;
4679 struct memcg_cgwb_frn *frn;
4680 u64 now = get_jiffies_64();
4681 u64 oldest_at = now;
4685 trace_track_foreign_dirty(page, wb);
4688 * Pick the slot to use. If there is already a slot for @wb, keep
4689 * using it. If not replace the oldest one which isn't being
4692 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4693 frn = &memcg->cgwb_frn[i];
4694 if (frn->bdi_id == wb->bdi->id &&
4695 frn->memcg_id == wb->memcg_css->id)
4697 if (time_before64(frn->at, oldest_at) &&
4698 atomic_read(&frn->done.cnt) == 1) {
4700 oldest_at = frn->at;
4704 if (i < MEMCG_CGWB_FRN_CNT) {
4706 * Re-using an existing one. Update timestamp lazily to
4707 * avoid making the cacheline hot. We want them to be
4708 * reasonably up-to-date and significantly shorter than
4709 * dirty_expire_interval as that's what expires the record.
4710 * Use the shorter of 1s and dirty_expire_interval / 8.
4712 unsigned long update_intv =
4713 min_t(unsigned long, HZ,
4714 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4716 if (time_before64(frn->at, now - update_intv))
4718 } else if (oldest >= 0) {
4719 /* replace the oldest free one */
4720 frn = &memcg->cgwb_frn[oldest];
4721 frn->bdi_id = wb->bdi->id;
4722 frn->memcg_id = wb->memcg_css->id;
4727 /* issue foreign writeback flushes for recorded foreign dirtying events */
4728 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4730 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4731 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4732 u64 now = jiffies_64;
4735 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4736 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4739 * If the record is older than dirty_expire_interval,
4740 * writeback on it has already started. No need to kick it
4741 * off again. Also, don't start a new one if there's
4742 * already one in flight.
4744 if (time_after64(frn->at, now - intv) &&
4745 atomic_read(&frn->done.cnt) == 1) {
4747 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4748 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4749 WB_REASON_FOREIGN_FLUSH,
4755 #else /* CONFIG_CGROUP_WRITEBACK */
4757 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4762 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4766 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4770 #endif /* CONFIG_CGROUP_WRITEBACK */
4773 * DO NOT USE IN NEW FILES.
4775 * "cgroup.event_control" implementation.
4777 * This is way over-engineered. It tries to support fully configurable
4778 * events for each user. Such level of flexibility is completely
4779 * unnecessary especially in the light of the planned unified hierarchy.
4781 * Please deprecate this and replace with something simpler if at all
4786 * Unregister event and free resources.
4788 * Gets called from workqueue.
4790 static void memcg_event_remove(struct work_struct *work)
4792 struct mem_cgroup_event *event =
4793 container_of(work, struct mem_cgroup_event, remove);
4794 struct mem_cgroup *memcg = event->memcg;
4796 remove_wait_queue(event->wqh, &event->wait);
4798 event->unregister_event(memcg, event->eventfd);
4800 /* Notify userspace the event is going away. */
4801 eventfd_signal(event->eventfd, 1);
4803 eventfd_ctx_put(event->eventfd);
4805 css_put(&memcg->css);
4809 * Gets called on EPOLLHUP on eventfd when user closes it.
4811 * Called with wqh->lock held and interrupts disabled.
4813 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4814 int sync, void *key)
4816 struct mem_cgroup_event *event =
4817 container_of(wait, struct mem_cgroup_event, wait);
4818 struct mem_cgroup *memcg = event->memcg;
4819 __poll_t flags = key_to_poll(key);
4821 if (flags & EPOLLHUP) {
4823 * If the event has been detached at cgroup removal, we
4824 * can simply return knowing the other side will cleanup
4827 * We can't race against event freeing since the other
4828 * side will require wqh->lock via remove_wait_queue(),
4831 spin_lock(&memcg->event_list_lock);
4832 if (!list_empty(&event->list)) {
4833 list_del_init(&event->list);
4835 * We are in atomic context, but cgroup_event_remove()
4836 * may sleep, so we have to call it in workqueue.
4838 schedule_work(&event->remove);
4840 spin_unlock(&memcg->event_list_lock);
4846 static void memcg_event_ptable_queue_proc(struct file *file,
4847 wait_queue_head_t *wqh, poll_table *pt)
4849 struct mem_cgroup_event *event =
4850 container_of(pt, struct mem_cgroup_event, pt);
4853 add_wait_queue(wqh, &event->wait);
4857 * DO NOT USE IN NEW FILES.
4859 * Parse input and register new cgroup event handler.
4861 * Input must be in format '<event_fd> <control_fd> <args>'.
4862 * Interpretation of args is defined by control file implementation.
4864 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4865 char *buf, size_t nbytes, loff_t off)
4867 struct cgroup_subsys_state *css = of_css(of);
4868 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4869 struct mem_cgroup_event *event;
4870 struct cgroup_subsys_state *cfile_css;
4871 unsigned int efd, cfd;
4874 struct dentry *cdentry;
4879 buf = strstrip(buf);
4881 efd = simple_strtoul(buf, &endp, 10);
4886 cfd = simple_strtoul(buf, &endp, 10);
4887 if ((*endp != ' ') && (*endp != '\0'))
4891 event = kzalloc(sizeof(*event), GFP_KERNEL);
4895 event->memcg = memcg;
4896 INIT_LIST_HEAD(&event->list);
4897 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4898 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4899 INIT_WORK(&event->remove, memcg_event_remove);
4907 event->eventfd = eventfd_ctx_fileget(efile.file);
4908 if (IS_ERR(event->eventfd)) {
4909 ret = PTR_ERR(event->eventfd);
4916 goto out_put_eventfd;
4919 /* the process need read permission on control file */
4920 /* AV: shouldn't we check that it's been opened for read instead? */
4921 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4926 * The control file must be a regular cgroup1 file. As a regular cgroup
4927 * file can't be renamed, it's safe to access its name afterwards.
4929 cdentry = cfile.file->f_path.dentry;
4930 if (cdentry->d_sb->s_type != &cgroup_fs_type || !d_is_reg(cdentry)) {
4936 * Determine the event callbacks and set them in @event. This used
4937 * to be done via struct cftype but cgroup core no longer knows
4938 * about these events. The following is crude but the whole thing
4939 * is for compatibility anyway.
4941 * DO NOT ADD NEW FILES.
4943 name = cdentry->d_name.name;
4945 if (!strcmp(name, "memory.usage_in_bytes")) {
4946 event->register_event = mem_cgroup_usage_register_event;
4947 event->unregister_event = mem_cgroup_usage_unregister_event;
4948 } else if (!strcmp(name, "memory.oom_control")) {
4949 event->register_event = mem_cgroup_oom_register_event;
4950 event->unregister_event = mem_cgroup_oom_unregister_event;
4951 } else if (!strcmp(name, "memory.pressure_level")) {
4952 event->register_event = vmpressure_register_event;
4953 event->unregister_event = vmpressure_unregister_event;
4954 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4955 event->register_event = memsw_cgroup_usage_register_event;
4956 event->unregister_event = memsw_cgroup_usage_unregister_event;
4963 * Verify @cfile should belong to @css. Also, remaining events are
4964 * automatically removed on cgroup destruction but the removal is
4965 * asynchronous, so take an extra ref on @css.
4967 cfile_css = css_tryget_online_from_dir(cdentry->d_parent,
4968 &memory_cgrp_subsys);
4970 if (IS_ERR(cfile_css))
4972 if (cfile_css != css) {
4977 ret = event->register_event(memcg, event->eventfd, buf);
4981 vfs_poll(efile.file, &event->pt);
4983 spin_lock(&memcg->event_list_lock);
4984 list_add(&event->list, &memcg->event_list);
4985 spin_unlock(&memcg->event_list_lock);
4997 eventfd_ctx_put(event->eventfd);
5006 static struct cftype mem_cgroup_legacy_files[] = {
5008 .name = "usage_in_bytes",
5009 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5010 .read_u64 = mem_cgroup_read_u64,
5013 .name = "max_usage_in_bytes",
5014 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5015 .write = mem_cgroup_reset,
5016 .read_u64 = mem_cgroup_read_u64,
5019 .name = "limit_in_bytes",
5020 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5021 .write = mem_cgroup_write,
5022 .read_u64 = mem_cgroup_read_u64,
5025 .name = "soft_limit_in_bytes",
5026 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5027 .write = mem_cgroup_write,
5028 .read_u64 = mem_cgroup_read_u64,
5032 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5033 .write = mem_cgroup_reset,
5034 .read_u64 = mem_cgroup_read_u64,
5038 .seq_show = memcg_stat_show,
5041 .name = "force_empty",
5042 .write = mem_cgroup_force_empty_write,
5045 .name = "use_hierarchy",
5046 .write_u64 = mem_cgroup_hierarchy_write,
5047 .read_u64 = mem_cgroup_hierarchy_read,
5050 .name = "cgroup.event_control", /* XXX: for compat */
5051 .write = memcg_write_event_control,
5052 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
5055 .name = "swappiness",
5056 .read_u64 = mem_cgroup_swappiness_read,
5057 .write_u64 = mem_cgroup_swappiness_write,
5060 .name = "move_charge_at_immigrate",
5061 .read_u64 = mem_cgroup_move_charge_read,
5062 .write_u64 = mem_cgroup_move_charge_write,
5065 .name = "oom_control",
5066 .seq_show = mem_cgroup_oom_control_read,
5067 .write_u64 = mem_cgroup_oom_control_write,
5068 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5071 .name = "pressure_level",
5075 .name = "numa_stat",
5076 .seq_show = memcg_numa_stat_show,
5080 .name = "kmem.limit_in_bytes",
5081 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5082 .write = mem_cgroup_write,
5083 .read_u64 = mem_cgroup_read_u64,
5086 .name = "kmem.usage_in_bytes",
5087 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5088 .read_u64 = mem_cgroup_read_u64,
5091 .name = "kmem.failcnt",
5092 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5093 .write = mem_cgroup_reset,
5094 .read_u64 = mem_cgroup_read_u64,
5097 .name = "kmem.max_usage_in_bytes",
5098 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5099 .write = mem_cgroup_reset,
5100 .read_u64 = mem_cgroup_read_u64,
5102 #if defined(CONFIG_MEMCG_KMEM) && \
5103 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5105 .name = "kmem.slabinfo",
5106 .seq_show = memcg_slab_show,
5110 .name = "kmem.tcp.limit_in_bytes",
5111 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5112 .write = mem_cgroup_write,
5113 .read_u64 = mem_cgroup_read_u64,
5116 .name = "kmem.tcp.usage_in_bytes",
5117 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5118 .read_u64 = mem_cgroup_read_u64,
5121 .name = "kmem.tcp.failcnt",
5122 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5123 .write = mem_cgroup_reset,
5124 .read_u64 = mem_cgroup_read_u64,
5127 .name = "kmem.tcp.max_usage_in_bytes",
5128 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5129 .write = mem_cgroup_reset,
5130 .read_u64 = mem_cgroup_read_u64,
5132 { }, /* terminate */
5136 * Private memory cgroup IDR
5138 * Swap-out records and page cache shadow entries need to store memcg
5139 * references in constrained space, so we maintain an ID space that is
5140 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5141 * memory-controlled cgroups to 64k.
5143 * However, there usually are many references to the offline CSS after
5144 * the cgroup has been destroyed, such as page cache or reclaimable
5145 * slab objects, that don't need to hang on to the ID. We want to keep
5146 * those dead CSS from occupying IDs, or we might quickly exhaust the
5147 * relatively small ID space and prevent the creation of new cgroups
5148 * even when there are much fewer than 64k cgroups - possibly none.
5150 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5151 * be freed and recycled when it's no longer needed, which is usually
5152 * when the CSS is offlined.
5154 * The only exception to that are records of swapped out tmpfs/shmem
5155 * pages that need to be attributed to live ancestors on swapin. But
5156 * those references are manageable from userspace.
5159 static DEFINE_IDR(mem_cgroup_idr);
5161 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5163 if (memcg->id.id > 0) {
5164 idr_remove(&mem_cgroup_idr, memcg->id.id);
5169 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5172 refcount_add(n, &memcg->id.ref);
5175 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5177 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5178 mem_cgroup_id_remove(memcg);
5180 /* Memcg ID pins CSS */
5181 css_put(&memcg->css);
5185 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5187 mem_cgroup_id_put_many(memcg, 1);
5191 * mem_cgroup_from_id - look up a memcg from a memcg id
5192 * @id: the memcg id to look up
5194 * Caller must hold rcu_read_lock().
5196 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5198 WARN_ON_ONCE(!rcu_read_lock_held());
5199 return idr_find(&mem_cgroup_idr, id);
5202 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5204 struct mem_cgroup_per_node *pn;
5207 * This routine is called against possible nodes.
5208 * But it's BUG to call kmalloc() against offline node.
5210 * TODO: this routine can waste much memory for nodes which will
5211 * never be onlined. It's better to use memory hotplug callback
5214 if (!node_state(node, N_NORMAL_MEMORY))
5216 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5220 pn->lruvec_stat_local = alloc_percpu_gfp(struct lruvec_stat,
5221 GFP_KERNEL_ACCOUNT);
5222 if (!pn->lruvec_stat_local) {
5227 pn->lruvec_stat_cpu = alloc_percpu_gfp(struct lruvec_stat,
5228 GFP_KERNEL_ACCOUNT);
5229 if (!pn->lruvec_stat_cpu) {
5230 free_percpu(pn->lruvec_stat_local);
5235 lruvec_init(&pn->lruvec);
5236 pn->usage_in_excess = 0;
5237 pn->on_tree = false;
5240 memcg->nodeinfo[node] = pn;
5244 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5246 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5251 free_percpu(pn->lruvec_stat_cpu);
5252 free_percpu(pn->lruvec_stat_local);
5256 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5261 free_mem_cgroup_per_node_info(memcg, node);
5262 free_percpu(memcg->vmstats_percpu);
5263 free_percpu(memcg->vmstats_local);
5267 static void mem_cgroup_free(struct mem_cgroup *memcg)
5269 memcg_wb_domain_exit(memcg);
5271 * Flush percpu vmstats and vmevents to guarantee the value correctness
5272 * on parent's and all ancestor levels.
5274 memcg_flush_percpu_vmstats(memcg);
5275 memcg_flush_percpu_vmevents(memcg);
5276 __mem_cgroup_free(memcg);
5279 static struct mem_cgroup *mem_cgroup_alloc(void)
5281 struct mem_cgroup *memcg;
5284 int __maybe_unused i;
5285 long error = -ENOMEM;
5287 size = sizeof(struct mem_cgroup);
5288 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5290 memcg = kzalloc(size, GFP_KERNEL);
5292 return ERR_PTR(error);
5294 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5295 1, MEM_CGROUP_ID_MAX,
5297 if (memcg->id.id < 0) {
5298 error = memcg->id.id;
5302 memcg->vmstats_local = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5303 GFP_KERNEL_ACCOUNT);
5304 if (!memcg->vmstats_local)
5307 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5308 GFP_KERNEL_ACCOUNT);
5309 if (!memcg->vmstats_percpu)
5313 if (alloc_mem_cgroup_per_node_info(memcg, node))
5316 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5319 INIT_WORK(&memcg->high_work, high_work_func);
5320 INIT_LIST_HEAD(&memcg->oom_notify);
5321 mutex_init(&memcg->thresholds_lock);
5322 spin_lock_init(&memcg->move_lock);
5323 vmpressure_init(&memcg->vmpressure);
5324 INIT_LIST_HEAD(&memcg->event_list);
5325 spin_lock_init(&memcg->event_list_lock);
5326 memcg->socket_pressure = jiffies;
5327 #ifdef CONFIG_MEMCG_KMEM
5328 memcg->kmemcg_id = -1;
5329 INIT_LIST_HEAD(&memcg->objcg_list);
5331 #ifdef CONFIG_CGROUP_WRITEBACK
5332 INIT_LIST_HEAD(&memcg->cgwb_list);
5333 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5334 memcg->cgwb_frn[i].done =
5335 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5337 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5338 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5339 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5340 memcg->deferred_split_queue.split_queue_len = 0;
5342 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5345 mem_cgroup_id_remove(memcg);
5346 __mem_cgroup_free(memcg);
5347 return ERR_PTR(error);
5350 static struct cgroup_subsys_state * __ref
5351 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5353 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5354 struct mem_cgroup *memcg, *old_memcg;
5355 long error = -ENOMEM;
5357 old_memcg = set_active_memcg(parent);
5358 memcg = mem_cgroup_alloc();
5359 set_active_memcg(old_memcg);
5361 return ERR_CAST(memcg);
5363 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5364 memcg->soft_limit = PAGE_COUNTER_MAX;
5365 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5367 memcg->swappiness = mem_cgroup_swappiness(parent);
5368 memcg->oom_kill_disable = parent->oom_kill_disable;
5371 page_counter_init(&memcg->memory, NULL);
5372 page_counter_init(&memcg->swap, NULL);
5373 page_counter_init(&memcg->kmem, NULL);
5374 page_counter_init(&memcg->tcpmem, NULL);
5375 } else if (parent->use_hierarchy) {
5376 memcg->use_hierarchy = true;
5377 page_counter_init(&memcg->memory, &parent->memory);
5378 page_counter_init(&memcg->swap, &parent->swap);
5379 page_counter_init(&memcg->kmem, &parent->kmem);
5380 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5382 page_counter_init(&memcg->memory, &root_mem_cgroup->memory);
5383 page_counter_init(&memcg->swap, &root_mem_cgroup->swap);
5384 page_counter_init(&memcg->kmem, &root_mem_cgroup->kmem);
5385 page_counter_init(&memcg->tcpmem, &root_mem_cgroup->tcpmem);
5387 * Deeper hierachy with use_hierarchy == false doesn't make
5388 * much sense so let cgroup subsystem know about this
5389 * unfortunate state in our controller.
5391 if (parent != root_mem_cgroup)
5392 memory_cgrp_subsys.broken_hierarchy = true;
5395 /* The following stuff does not apply to the root */
5397 root_mem_cgroup = memcg;
5401 error = memcg_online_kmem(memcg);
5405 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5406 static_branch_inc(&memcg_sockets_enabled_key);
5410 mem_cgroup_id_remove(memcg);
5411 mem_cgroup_free(memcg);
5412 return ERR_PTR(error);
5415 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5417 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5420 * A memcg must be visible for memcg_expand_shrinker_maps()
5421 * by the time the maps are allocated. So, we allocate maps
5422 * here, when for_each_mem_cgroup() can't skip it.
5424 if (memcg_alloc_shrinker_maps(memcg)) {
5425 mem_cgroup_id_remove(memcg);
5429 /* Online state pins memcg ID, memcg ID pins CSS */
5430 refcount_set(&memcg->id.ref, 1);
5435 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5437 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5438 struct mem_cgroup_event *event, *tmp;
5441 * Unregister events and notify userspace.
5442 * Notify userspace about cgroup removing only after rmdir of cgroup
5443 * directory to avoid race between userspace and kernelspace.
5445 spin_lock(&memcg->event_list_lock);
5446 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5447 list_del_init(&event->list);
5448 schedule_work(&event->remove);
5450 spin_unlock(&memcg->event_list_lock);
5452 page_counter_set_min(&memcg->memory, 0);
5453 page_counter_set_low(&memcg->memory, 0);
5455 memcg_offline_kmem(memcg);
5456 wb_memcg_offline(memcg);
5458 drain_all_stock(memcg);
5460 mem_cgroup_id_put(memcg);
5463 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5465 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5467 invalidate_reclaim_iterators(memcg);
5470 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5472 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5473 int __maybe_unused i;
5475 #ifdef CONFIG_CGROUP_WRITEBACK
5476 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5477 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5479 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5480 static_branch_dec(&memcg_sockets_enabled_key);
5482 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5483 static_branch_dec(&memcg_sockets_enabled_key);
5485 vmpressure_cleanup(&memcg->vmpressure);
5486 cancel_work_sync(&memcg->high_work);
5487 mem_cgroup_remove_from_trees(memcg);
5488 memcg_free_shrinker_maps(memcg);
5489 memcg_free_kmem(memcg);
5490 mem_cgroup_free(memcg);
5494 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5495 * @css: the target css
5497 * Reset the states of the mem_cgroup associated with @css. This is
5498 * invoked when the userland requests disabling on the default hierarchy
5499 * but the memcg is pinned through dependency. The memcg should stop
5500 * applying policies and should revert to the vanilla state as it may be
5501 * made visible again.
5503 * The current implementation only resets the essential configurations.
5504 * This needs to be expanded to cover all the visible parts.
5506 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5508 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5510 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5511 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5512 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5513 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5514 page_counter_set_min(&memcg->memory, 0);
5515 page_counter_set_low(&memcg->memory, 0);
5516 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5517 memcg->soft_limit = PAGE_COUNTER_MAX;
5518 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5519 memcg_wb_domain_size_changed(memcg);
5523 /* Handlers for move charge at task migration. */
5524 static int mem_cgroup_do_precharge(unsigned long count)
5528 /* Try a single bulk charge without reclaim first, kswapd may wake */
5529 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5531 mc.precharge += count;
5535 /* Try charges one by one with reclaim, but do not retry */
5537 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5551 enum mc_target_type {
5558 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5559 unsigned long addr, pte_t ptent)
5561 struct page *page = vm_normal_page(vma, addr, ptent);
5563 if (!page || !page_mapped(page))
5565 if (PageAnon(page)) {
5566 if (!(mc.flags & MOVE_ANON))
5569 if (!(mc.flags & MOVE_FILE))
5572 if (!get_page_unless_zero(page))
5578 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5579 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5580 pte_t ptent, swp_entry_t *entry)
5582 struct page *page = NULL;
5583 swp_entry_t ent = pte_to_swp_entry(ptent);
5585 if (!(mc.flags & MOVE_ANON))
5589 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5590 * a device and because they are not accessible by CPU they are store
5591 * as special swap entry in the CPU page table.
5593 if (is_device_private_entry(ent)) {
5594 page = device_private_entry_to_page(ent);
5596 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5597 * a refcount of 1 when free (unlike normal page)
5599 if (!page_ref_add_unless(page, 1, 1))
5604 if (non_swap_entry(ent))
5608 * Because lookup_swap_cache() updates some statistics counter,
5609 * we call find_get_page() with swapper_space directly.
5611 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5612 entry->val = ent.val;
5617 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5618 pte_t ptent, swp_entry_t *entry)
5624 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5625 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5627 if (!vma->vm_file) /* anonymous vma */
5629 if (!(mc.flags & MOVE_FILE))
5632 /* page is moved even if it's not RSS of this task(page-faulted). */
5633 /* shmem/tmpfs may report page out on swap: account for that too. */
5634 return find_get_incore_page(vma->vm_file->f_mapping,
5635 linear_page_index(vma, addr));
5639 * mem_cgroup_move_account - move account of the page
5641 * @compound: charge the page as compound or small page
5642 * @from: mem_cgroup which the page is moved from.
5643 * @to: mem_cgroup which the page is moved to. @from != @to.
5645 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5647 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5650 static int mem_cgroup_move_account(struct page *page,
5652 struct mem_cgroup *from,
5653 struct mem_cgroup *to)
5655 struct lruvec *from_vec, *to_vec;
5656 struct pglist_data *pgdat;
5657 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5660 VM_BUG_ON(from == to);
5661 VM_BUG_ON_PAGE(PageLRU(page), page);
5662 VM_BUG_ON(compound && !PageTransHuge(page));
5665 * Prevent mem_cgroup_migrate() from looking at
5666 * page->mem_cgroup of its source page while we change it.
5669 if (!trylock_page(page))
5673 if (page->mem_cgroup != from)
5676 pgdat = page_pgdat(page);
5677 from_vec = mem_cgroup_lruvec(from, pgdat);
5678 to_vec = mem_cgroup_lruvec(to, pgdat);
5680 lock_page_memcg(page);
5682 if (PageAnon(page)) {
5683 if (page_mapped(page)) {
5684 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5685 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5686 if (PageTransHuge(page)) {
5687 __dec_lruvec_state(from_vec, NR_ANON_THPS);
5688 __inc_lruvec_state(to_vec, NR_ANON_THPS);
5693 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5694 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5696 if (PageSwapBacked(page)) {
5697 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5698 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5701 if (page_mapped(page)) {
5702 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5703 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5706 if (PageDirty(page)) {
5707 struct address_space *mapping = page_mapping(page);
5709 if (mapping_can_writeback(mapping)) {
5710 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5712 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5718 if (PageWriteback(page)) {
5719 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5720 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5724 * All state has been migrated, let's switch to the new memcg.
5726 * It is safe to change page->mem_cgroup here because the page
5727 * is referenced, charged, isolated, and locked: we can't race
5728 * with (un)charging, migration, LRU putback, or anything else
5729 * that would rely on a stable page->mem_cgroup.
5731 * Note that lock_page_memcg is a memcg lock, not a page lock,
5732 * to save space. As soon as we switch page->mem_cgroup to a
5733 * new memcg that isn't locked, the above state can change
5734 * concurrently again. Make sure we're truly done with it.
5739 css_put(&from->css);
5741 page->mem_cgroup = to;
5743 __unlock_page_memcg(from);
5747 local_irq_disable();
5748 mem_cgroup_charge_statistics(to, page, nr_pages);
5749 memcg_check_events(to, page);
5750 mem_cgroup_charge_statistics(from, page, -nr_pages);
5751 memcg_check_events(from, page);
5760 * get_mctgt_type - get target type of moving charge
5761 * @vma: the vma the pte to be checked belongs
5762 * @addr: the address corresponding to the pte to be checked
5763 * @ptent: the pte to be checked
5764 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5767 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5768 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5769 * move charge. if @target is not NULL, the page is stored in target->page
5770 * with extra refcnt got(Callers should handle it).
5771 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5772 * target for charge migration. if @target is not NULL, the entry is stored
5774 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5775 * (so ZONE_DEVICE page and thus not on the lru).
5776 * For now we such page is charge like a regular page would be as for all
5777 * intent and purposes it is just special memory taking the place of a
5780 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5782 * Called with pte lock held.
5785 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5786 unsigned long addr, pte_t ptent, union mc_target *target)
5788 struct page *page = NULL;
5789 enum mc_target_type ret = MC_TARGET_NONE;
5790 swp_entry_t ent = { .val = 0 };
5792 if (pte_present(ptent))
5793 page = mc_handle_present_pte(vma, addr, ptent);
5794 else if (is_swap_pte(ptent))
5795 page = mc_handle_swap_pte(vma, ptent, &ent);
5796 else if (pte_none(ptent))
5797 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5799 if (!page && !ent.val)
5803 * Do only loose check w/o serialization.
5804 * mem_cgroup_move_account() checks the page is valid or
5805 * not under LRU exclusion.
5807 if (page->mem_cgroup == mc.from) {
5808 ret = MC_TARGET_PAGE;
5809 if (is_device_private_page(page))
5810 ret = MC_TARGET_DEVICE;
5812 target->page = page;
5814 if (!ret || !target)
5818 * There is a swap entry and a page doesn't exist or isn't charged.
5819 * But we cannot move a tail-page in a THP.
5821 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5822 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5823 ret = MC_TARGET_SWAP;
5830 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5832 * We don't consider PMD mapped swapping or file mapped pages because THP does
5833 * not support them for now.
5834 * Caller should make sure that pmd_trans_huge(pmd) is true.
5836 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5837 unsigned long addr, pmd_t pmd, union mc_target *target)
5839 struct page *page = NULL;
5840 enum mc_target_type ret = MC_TARGET_NONE;
5842 if (unlikely(is_swap_pmd(pmd))) {
5843 VM_BUG_ON(thp_migration_supported() &&
5844 !is_pmd_migration_entry(pmd));
5847 page = pmd_page(pmd);
5848 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5849 if (!(mc.flags & MOVE_ANON))
5851 if (page->mem_cgroup == mc.from) {
5852 ret = MC_TARGET_PAGE;
5855 target->page = page;
5861 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5862 unsigned long addr, pmd_t pmd, union mc_target *target)
5864 return MC_TARGET_NONE;
5868 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5869 unsigned long addr, unsigned long end,
5870 struct mm_walk *walk)
5872 struct vm_area_struct *vma = walk->vma;
5876 ptl = pmd_trans_huge_lock(pmd, vma);
5879 * Note their can not be MC_TARGET_DEVICE for now as we do not
5880 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5881 * this might change.
5883 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5884 mc.precharge += HPAGE_PMD_NR;
5889 if (pmd_trans_unstable(pmd))
5891 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5892 for (; addr != end; pte++, addr += PAGE_SIZE)
5893 if (get_mctgt_type(vma, addr, *pte, NULL))
5894 mc.precharge++; /* increment precharge temporarily */
5895 pte_unmap_unlock(pte - 1, ptl);
5901 static const struct mm_walk_ops precharge_walk_ops = {
5902 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5905 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5907 unsigned long precharge;
5910 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5911 mmap_read_unlock(mm);
5913 precharge = mc.precharge;
5919 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5921 unsigned long precharge = mem_cgroup_count_precharge(mm);
5923 VM_BUG_ON(mc.moving_task);
5924 mc.moving_task = current;
5925 return mem_cgroup_do_precharge(precharge);
5928 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5929 static void __mem_cgroup_clear_mc(void)
5931 struct mem_cgroup *from = mc.from;
5932 struct mem_cgroup *to = mc.to;
5934 /* we must uncharge all the leftover precharges from mc.to */
5936 cancel_charge(mc.to, mc.precharge);
5940 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5941 * we must uncharge here.
5943 if (mc.moved_charge) {
5944 cancel_charge(mc.from, mc.moved_charge);
5945 mc.moved_charge = 0;
5947 /* we must fixup refcnts and charges */
5948 if (mc.moved_swap) {
5949 /* uncharge swap account from the old cgroup */
5950 if (!mem_cgroup_is_root(mc.from))
5951 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5953 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5956 * we charged both to->memory and to->memsw, so we
5957 * should uncharge to->memory.
5959 if (!mem_cgroup_is_root(mc.to))
5960 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5964 memcg_oom_recover(from);
5965 memcg_oom_recover(to);
5966 wake_up_all(&mc.waitq);
5969 static void mem_cgroup_clear_mc(void)
5971 struct mm_struct *mm = mc.mm;
5974 * we must clear moving_task before waking up waiters at the end of
5977 mc.moving_task = NULL;
5978 __mem_cgroup_clear_mc();
5979 spin_lock(&mc.lock);
5983 spin_unlock(&mc.lock);
5988 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5990 struct cgroup_subsys_state *css;
5991 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5992 struct mem_cgroup *from;
5993 struct task_struct *leader, *p;
5994 struct mm_struct *mm;
5995 unsigned long move_flags;
5998 /* charge immigration isn't supported on the default hierarchy */
5999 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6003 * Multi-process migrations only happen on the default hierarchy
6004 * where charge immigration is not used. Perform charge
6005 * immigration if @tset contains a leader and whine if there are
6009 cgroup_taskset_for_each_leader(leader, css, tset) {
6012 memcg = mem_cgroup_from_css(css);
6018 * We are now commited to this value whatever it is. Changes in this
6019 * tunable will only affect upcoming migrations, not the current one.
6020 * So we need to save it, and keep it going.
6022 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
6026 from = mem_cgroup_from_task(p);
6028 VM_BUG_ON(from == memcg);
6030 mm = get_task_mm(p);
6033 /* We move charges only when we move a owner of the mm */
6034 if (mm->owner == p) {
6037 VM_BUG_ON(mc.precharge);
6038 VM_BUG_ON(mc.moved_charge);
6039 VM_BUG_ON(mc.moved_swap);
6041 spin_lock(&mc.lock);
6045 mc.flags = move_flags;
6046 spin_unlock(&mc.lock);
6047 /* We set mc.moving_task later */
6049 ret = mem_cgroup_precharge_mc(mm);
6051 mem_cgroup_clear_mc();
6058 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6061 mem_cgroup_clear_mc();
6064 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6065 unsigned long addr, unsigned long end,
6066 struct mm_walk *walk)
6069 struct vm_area_struct *vma = walk->vma;
6072 enum mc_target_type target_type;
6073 union mc_target target;
6076 ptl = pmd_trans_huge_lock(pmd, vma);
6078 if (mc.precharge < HPAGE_PMD_NR) {
6082 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6083 if (target_type == MC_TARGET_PAGE) {
6085 if (!isolate_lru_page(page)) {
6086 if (!mem_cgroup_move_account(page, true,
6088 mc.precharge -= HPAGE_PMD_NR;
6089 mc.moved_charge += HPAGE_PMD_NR;
6091 putback_lru_page(page);
6094 } else if (target_type == MC_TARGET_DEVICE) {
6096 if (!mem_cgroup_move_account(page, true,
6098 mc.precharge -= HPAGE_PMD_NR;
6099 mc.moved_charge += HPAGE_PMD_NR;
6107 if (pmd_trans_unstable(pmd))
6110 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6111 for (; addr != end; addr += PAGE_SIZE) {
6112 pte_t ptent = *(pte++);
6113 bool device = false;
6119 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6120 case MC_TARGET_DEVICE:
6123 case MC_TARGET_PAGE:
6126 * We can have a part of the split pmd here. Moving it
6127 * can be done but it would be too convoluted so simply
6128 * ignore such a partial THP and keep it in original
6129 * memcg. There should be somebody mapping the head.
6131 if (PageTransCompound(page))
6133 if (!device && isolate_lru_page(page))
6135 if (!mem_cgroup_move_account(page, false,
6138 /* we uncharge from mc.from later. */
6142 putback_lru_page(page);
6143 put: /* get_mctgt_type() gets the page */
6146 case MC_TARGET_SWAP:
6148 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6150 mem_cgroup_id_get_many(mc.to, 1);
6151 /* we fixup other refcnts and charges later. */
6159 pte_unmap_unlock(pte - 1, ptl);
6164 * We have consumed all precharges we got in can_attach().
6165 * We try charge one by one, but don't do any additional
6166 * charges to mc.to if we have failed in charge once in attach()
6169 ret = mem_cgroup_do_precharge(1);
6177 static const struct mm_walk_ops charge_walk_ops = {
6178 .pmd_entry = mem_cgroup_move_charge_pte_range,
6181 static void mem_cgroup_move_charge(void)
6183 lru_add_drain_all();
6185 * Signal lock_page_memcg() to take the memcg's move_lock
6186 * while we're moving its pages to another memcg. Then wait
6187 * for already started RCU-only updates to finish.
6189 atomic_inc(&mc.from->moving_account);
6192 if (unlikely(!mmap_read_trylock(mc.mm))) {
6194 * Someone who are holding the mmap_lock might be waiting in
6195 * waitq. So we cancel all extra charges, wake up all waiters,
6196 * and retry. Because we cancel precharges, we might not be able
6197 * to move enough charges, but moving charge is a best-effort
6198 * feature anyway, so it wouldn't be a big problem.
6200 __mem_cgroup_clear_mc();
6205 * When we have consumed all precharges and failed in doing
6206 * additional charge, the page walk just aborts.
6208 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6211 mmap_read_unlock(mc.mm);
6212 atomic_dec(&mc.from->moving_account);
6215 static void mem_cgroup_move_task(void)
6218 mem_cgroup_move_charge();
6219 mem_cgroup_clear_mc();
6222 #else /* !CONFIG_MMU */
6223 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6227 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6230 static void mem_cgroup_move_task(void)
6236 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6237 * to verify whether we're attached to the default hierarchy on each mount
6240 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6243 * use_hierarchy is forced on the default hierarchy. cgroup core
6244 * guarantees that @root doesn't have any children, so turning it
6245 * on for the root memcg is enough.
6247 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6248 root_mem_cgroup->use_hierarchy = true;
6250 root_mem_cgroup->use_hierarchy = false;
6253 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6255 if (value == PAGE_COUNTER_MAX)
6256 seq_puts(m, "max\n");
6258 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6263 static u64 memory_current_read(struct cgroup_subsys_state *css,
6266 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6268 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6271 static int memory_min_show(struct seq_file *m, void *v)
6273 return seq_puts_memcg_tunable(m,
6274 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6277 static ssize_t memory_min_write(struct kernfs_open_file *of,
6278 char *buf, size_t nbytes, loff_t off)
6280 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6284 buf = strstrip(buf);
6285 err = page_counter_memparse(buf, "max", &min);
6289 page_counter_set_min(&memcg->memory, min);
6294 static int memory_low_show(struct seq_file *m, void *v)
6296 return seq_puts_memcg_tunable(m,
6297 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6300 static ssize_t memory_low_write(struct kernfs_open_file *of,
6301 char *buf, size_t nbytes, loff_t off)
6303 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6307 buf = strstrip(buf);
6308 err = page_counter_memparse(buf, "max", &low);
6312 page_counter_set_low(&memcg->memory, low);
6317 static int memory_high_show(struct seq_file *m, void *v)
6319 return seq_puts_memcg_tunable(m,
6320 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6323 static ssize_t memory_high_write(struct kernfs_open_file *of,
6324 char *buf, size_t nbytes, loff_t off)
6326 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6327 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6328 bool drained = false;
6332 buf = strstrip(buf);
6333 err = page_counter_memparse(buf, "max", &high);
6337 page_counter_set_high(&memcg->memory, high);
6340 unsigned long nr_pages = page_counter_read(&memcg->memory);
6341 unsigned long reclaimed;
6343 if (nr_pages <= high)
6346 if (signal_pending(current))
6350 drain_all_stock(memcg);
6355 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6358 if (!reclaimed && !nr_retries--)
6362 memcg_wb_domain_size_changed(memcg);
6366 static int memory_max_show(struct seq_file *m, void *v)
6368 return seq_puts_memcg_tunable(m,
6369 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6372 static ssize_t memory_max_write(struct kernfs_open_file *of,
6373 char *buf, size_t nbytes, loff_t off)
6375 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6376 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6377 bool drained = false;
6381 buf = strstrip(buf);
6382 err = page_counter_memparse(buf, "max", &max);
6386 xchg(&memcg->memory.max, max);
6389 unsigned long nr_pages = page_counter_read(&memcg->memory);
6391 if (nr_pages <= max)
6394 if (signal_pending(current))
6398 drain_all_stock(memcg);
6404 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6410 memcg_memory_event(memcg, MEMCG_OOM);
6411 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6415 memcg_wb_domain_size_changed(memcg);
6419 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6421 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6422 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6423 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6424 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6425 seq_printf(m, "oom_kill %lu\n",
6426 atomic_long_read(&events[MEMCG_OOM_KILL]));
6429 static int memory_events_show(struct seq_file *m, void *v)
6431 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6433 __memory_events_show(m, memcg->memory_events);
6437 static int memory_events_local_show(struct seq_file *m, void *v)
6439 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6441 __memory_events_show(m, memcg->memory_events_local);
6445 static int memory_stat_show(struct seq_file *m, void *v)
6447 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6450 buf = memory_stat_format(memcg);
6459 static int memory_numa_stat_show(struct seq_file *m, void *v)
6462 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6464 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6467 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6470 seq_printf(m, "%s", memory_stats[i].name);
6471 for_each_node_state(nid, N_MEMORY) {
6473 struct lruvec *lruvec;
6475 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6476 size = lruvec_page_state(lruvec, memory_stats[i].idx);
6477 size *= memory_stats[i].ratio;
6478 seq_printf(m, " N%d=%llu", nid, size);
6487 static int memory_oom_group_show(struct seq_file *m, void *v)
6489 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6491 seq_printf(m, "%d\n", memcg->oom_group);
6496 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6497 char *buf, size_t nbytes, loff_t off)
6499 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6502 buf = strstrip(buf);
6506 ret = kstrtoint(buf, 0, &oom_group);
6510 if (oom_group != 0 && oom_group != 1)
6513 memcg->oom_group = oom_group;
6518 static struct cftype memory_files[] = {
6521 .flags = CFTYPE_NOT_ON_ROOT,
6522 .read_u64 = memory_current_read,
6526 .flags = CFTYPE_NOT_ON_ROOT,
6527 .seq_show = memory_min_show,
6528 .write = memory_min_write,
6532 .flags = CFTYPE_NOT_ON_ROOT,
6533 .seq_show = memory_low_show,
6534 .write = memory_low_write,
6538 .flags = CFTYPE_NOT_ON_ROOT,
6539 .seq_show = memory_high_show,
6540 .write = memory_high_write,
6544 .flags = CFTYPE_NOT_ON_ROOT,
6545 .seq_show = memory_max_show,
6546 .write = memory_max_write,
6550 .flags = CFTYPE_NOT_ON_ROOT,
6551 .file_offset = offsetof(struct mem_cgroup, events_file),
6552 .seq_show = memory_events_show,
6555 .name = "events.local",
6556 .flags = CFTYPE_NOT_ON_ROOT,
6557 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6558 .seq_show = memory_events_local_show,
6562 .seq_show = memory_stat_show,
6566 .name = "numa_stat",
6567 .seq_show = memory_numa_stat_show,
6571 .name = "oom.group",
6572 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6573 .seq_show = memory_oom_group_show,
6574 .write = memory_oom_group_write,
6579 struct cgroup_subsys memory_cgrp_subsys = {
6580 .css_alloc = mem_cgroup_css_alloc,
6581 .css_online = mem_cgroup_css_online,
6582 .css_offline = mem_cgroup_css_offline,
6583 .css_released = mem_cgroup_css_released,
6584 .css_free = mem_cgroup_css_free,
6585 .css_reset = mem_cgroup_css_reset,
6586 .can_attach = mem_cgroup_can_attach,
6587 .cancel_attach = mem_cgroup_cancel_attach,
6588 .post_attach = mem_cgroup_move_task,
6589 .bind = mem_cgroup_bind,
6590 .dfl_cftypes = memory_files,
6591 .legacy_cftypes = mem_cgroup_legacy_files,
6596 * This function calculates an individual cgroup's effective
6597 * protection which is derived from its own memory.min/low, its
6598 * parent's and siblings' settings, as well as the actual memory
6599 * distribution in the tree.
6601 * The following rules apply to the effective protection values:
6603 * 1. At the first level of reclaim, effective protection is equal to
6604 * the declared protection in memory.min and memory.low.
6606 * 2. To enable safe delegation of the protection configuration, at
6607 * subsequent levels the effective protection is capped to the
6608 * parent's effective protection.
6610 * 3. To make complex and dynamic subtrees easier to configure, the
6611 * user is allowed to overcommit the declared protection at a given
6612 * level. If that is the case, the parent's effective protection is
6613 * distributed to the children in proportion to how much protection
6614 * they have declared and how much of it they are utilizing.
6616 * This makes distribution proportional, but also work-conserving:
6617 * if one cgroup claims much more protection than it uses memory,
6618 * the unused remainder is available to its siblings.
6620 * 4. Conversely, when the declared protection is undercommitted at a
6621 * given level, the distribution of the larger parental protection
6622 * budget is NOT proportional. A cgroup's protection from a sibling
6623 * is capped to its own memory.min/low setting.
6625 * 5. However, to allow protecting recursive subtrees from each other
6626 * without having to declare each individual cgroup's fixed share
6627 * of the ancestor's claim to protection, any unutilized -
6628 * "floating" - protection from up the tree is distributed in
6629 * proportion to each cgroup's *usage*. This makes the protection
6630 * neutral wrt sibling cgroups and lets them compete freely over
6631 * the shared parental protection budget, but it protects the
6632 * subtree as a whole from neighboring subtrees.
6634 * Note that 4. and 5. are not in conflict: 4. is about protecting
6635 * against immediate siblings whereas 5. is about protecting against
6636 * neighboring subtrees.
6638 static unsigned long effective_protection(unsigned long usage,
6639 unsigned long parent_usage,
6640 unsigned long setting,
6641 unsigned long parent_effective,
6642 unsigned long siblings_protected)
6644 unsigned long protected;
6647 protected = min(usage, setting);
6649 * If all cgroups at this level combined claim and use more
6650 * protection then what the parent affords them, distribute
6651 * shares in proportion to utilization.
6653 * We are using actual utilization rather than the statically
6654 * claimed protection in order to be work-conserving: claimed
6655 * but unused protection is available to siblings that would
6656 * otherwise get a smaller chunk than what they claimed.
6658 if (siblings_protected > parent_effective)
6659 return protected * parent_effective / siblings_protected;
6662 * Ok, utilized protection of all children is within what the
6663 * parent affords them, so we know whatever this child claims
6664 * and utilizes is effectively protected.
6666 * If there is unprotected usage beyond this value, reclaim
6667 * will apply pressure in proportion to that amount.
6669 * If there is unutilized protection, the cgroup will be fully
6670 * shielded from reclaim, but we do return a smaller value for
6671 * protection than what the group could enjoy in theory. This
6672 * is okay. With the overcommit distribution above, effective
6673 * protection is always dependent on how memory is actually
6674 * consumed among the siblings anyway.
6679 * If the children aren't claiming (all of) the protection
6680 * afforded to them by the parent, distribute the remainder in
6681 * proportion to the (unprotected) memory of each cgroup. That
6682 * way, cgroups that aren't explicitly prioritized wrt each
6683 * other compete freely over the allowance, but they are
6684 * collectively protected from neighboring trees.
6686 * We're using unprotected memory for the weight so that if
6687 * some cgroups DO claim explicit protection, we don't protect
6688 * the same bytes twice.
6690 * Check both usage and parent_usage against the respective
6691 * protected values. One should imply the other, but they
6692 * aren't read atomically - make sure the division is sane.
6694 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6696 if (parent_effective > siblings_protected &&
6697 parent_usage > siblings_protected &&
6698 usage > protected) {
6699 unsigned long unclaimed;
6701 unclaimed = parent_effective - siblings_protected;
6702 unclaimed *= usage - protected;
6703 unclaimed /= parent_usage - siblings_protected;
6712 * mem_cgroup_protected - check if memory consumption is in the normal range
6713 * @root: the top ancestor of the sub-tree being checked
6714 * @memcg: the memory cgroup to check
6716 * WARNING: This function is not stateless! It can only be used as part
6717 * of a top-down tree iteration, not for isolated queries.
6719 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6720 struct mem_cgroup *memcg)
6722 unsigned long usage, parent_usage;
6723 struct mem_cgroup *parent;
6725 if (mem_cgroup_disabled())
6729 root = root_mem_cgroup;
6732 * Effective values of the reclaim targets are ignored so they
6733 * can be stale. Have a look at mem_cgroup_protection for more
6735 * TODO: calculation should be more robust so that we do not need
6736 * that special casing.
6741 usage = page_counter_read(&memcg->memory);
6745 parent = parent_mem_cgroup(memcg);
6746 /* No parent means a non-hierarchical mode on v1 memcg */
6750 if (parent == root) {
6751 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6752 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6756 parent_usage = page_counter_read(&parent->memory);
6758 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6759 READ_ONCE(memcg->memory.min),
6760 READ_ONCE(parent->memory.emin),
6761 atomic_long_read(&parent->memory.children_min_usage)));
6763 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6764 READ_ONCE(memcg->memory.low),
6765 READ_ONCE(parent->memory.elow),
6766 atomic_long_read(&parent->memory.children_low_usage)));
6770 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6771 * @page: page to charge
6772 * @mm: mm context of the victim
6773 * @gfp_mask: reclaim mode
6775 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6776 * pages according to @gfp_mask if necessary.
6778 * Returns 0 on success. Otherwise, an error code is returned.
6780 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6782 unsigned int nr_pages = thp_nr_pages(page);
6783 struct mem_cgroup *memcg = NULL;
6786 if (mem_cgroup_disabled())
6789 if (PageSwapCache(page)) {
6790 swp_entry_t ent = { .val = page_private(page), };
6794 * Every swap fault against a single page tries to charge the
6795 * page, bail as early as possible. shmem_unuse() encounters
6796 * already charged pages, too. page->mem_cgroup is protected
6797 * by the page lock, which serializes swap cache removal, which
6798 * in turn serializes uncharging.
6800 VM_BUG_ON_PAGE(!PageLocked(page), page);
6801 if (compound_head(page)->mem_cgroup)
6804 id = lookup_swap_cgroup_id(ent);
6806 memcg = mem_cgroup_from_id(id);
6807 if (memcg && !css_tryget_online(&memcg->css))
6813 memcg = get_mem_cgroup_from_mm(mm);
6815 ret = try_charge(memcg, gfp_mask, nr_pages);
6819 css_get(&memcg->css);
6820 commit_charge(page, memcg);
6822 local_irq_disable();
6823 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6824 memcg_check_events(memcg, page);
6828 * Cgroup1's unified memory+swap counter has been charged with the
6829 * new swapcache page, finish the transfer by uncharging the swap
6830 * slot. The swap slot would also get uncharged when it dies, but
6831 * it can stick around indefinitely and we'd count the page twice
6834 * Cgroup2 has separate resource counters for memory and swap,
6835 * so this is a non-issue here. Memory and swap charge lifetimes
6836 * correspond 1:1 to page and swap slot lifetimes: we charge the
6837 * page to memory here, and uncharge swap when the slot is freed.
6839 if (do_memsw_account() && PageSwapCache(page)) {
6840 swp_entry_t entry = { .val = page_private(page) };
6842 * The swap entry might not get freed for a long time,
6843 * let's not wait for it. The page already received a
6844 * memory+swap charge, drop the swap entry duplicate.
6846 mem_cgroup_uncharge_swap(entry, nr_pages);
6850 css_put(&memcg->css);
6855 struct uncharge_gather {
6856 struct mem_cgroup *memcg;
6857 unsigned long nr_pages;
6858 unsigned long pgpgout;
6859 unsigned long nr_kmem;
6860 struct page *dummy_page;
6863 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6865 memset(ug, 0, sizeof(*ug));
6868 static void uncharge_batch(const struct uncharge_gather *ug)
6870 unsigned long flags;
6872 if (!mem_cgroup_is_root(ug->memcg)) {
6873 page_counter_uncharge(&ug->memcg->memory, ug->nr_pages);
6874 if (do_memsw_account())
6875 page_counter_uncharge(&ug->memcg->memsw, ug->nr_pages);
6876 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6877 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6878 memcg_oom_recover(ug->memcg);
6881 local_irq_save(flags);
6882 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6883 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_pages);
6884 memcg_check_events(ug->memcg, ug->dummy_page);
6885 local_irq_restore(flags);
6887 /* drop reference from uncharge_page */
6888 css_put(&ug->memcg->css);
6891 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6893 unsigned long nr_pages;
6895 VM_BUG_ON_PAGE(PageLRU(page), page);
6897 if (!page->mem_cgroup)
6901 * Nobody should be changing or seriously looking at
6902 * page->mem_cgroup at this point, we have fully
6903 * exclusive access to the page.
6906 if (ug->memcg != page->mem_cgroup) {
6909 uncharge_gather_clear(ug);
6911 ug->memcg = page->mem_cgroup;
6913 /* pairs with css_put in uncharge_batch */
6914 css_get(&ug->memcg->css);
6917 nr_pages = compound_nr(page);
6918 ug->nr_pages += nr_pages;
6920 if (!PageKmemcg(page)) {
6923 ug->nr_kmem += nr_pages;
6924 __ClearPageKmemcg(page);
6927 ug->dummy_page = page;
6928 page->mem_cgroup = NULL;
6929 css_put(&ug->memcg->css);
6932 static void uncharge_list(struct list_head *page_list)
6934 struct uncharge_gather ug;
6935 struct list_head *next;
6937 uncharge_gather_clear(&ug);
6940 * Note that the list can be a single page->lru; hence the
6941 * do-while loop instead of a simple list_for_each_entry().
6943 next = page_list->next;
6947 page = list_entry(next, struct page, lru);
6948 next = page->lru.next;
6950 uncharge_page(page, &ug);
6951 } while (next != page_list);
6954 uncharge_batch(&ug);
6958 * mem_cgroup_uncharge - uncharge a page
6959 * @page: page to uncharge
6961 * Uncharge a page previously charged with mem_cgroup_charge().
6963 void mem_cgroup_uncharge(struct page *page)
6965 struct uncharge_gather ug;
6967 if (mem_cgroup_disabled())
6970 /* Don't touch page->lru of any random page, pre-check: */
6971 if (!page->mem_cgroup)
6974 uncharge_gather_clear(&ug);
6975 uncharge_page(page, &ug);
6976 uncharge_batch(&ug);
6980 * mem_cgroup_uncharge_list - uncharge a list of page
6981 * @page_list: list of pages to uncharge
6983 * Uncharge a list of pages previously charged with
6984 * mem_cgroup_charge().
6986 void mem_cgroup_uncharge_list(struct list_head *page_list)
6988 if (mem_cgroup_disabled())
6991 if (!list_empty(page_list))
6992 uncharge_list(page_list);
6996 * mem_cgroup_migrate - charge a page's replacement
6997 * @oldpage: currently circulating page
6998 * @newpage: replacement page
7000 * Charge @newpage as a replacement page for @oldpage. @oldpage will
7001 * be uncharged upon free.
7003 * Both pages must be locked, @newpage->mapping must be set up.
7005 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
7007 struct mem_cgroup *memcg;
7008 unsigned int nr_pages;
7009 unsigned long flags;
7011 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
7012 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
7013 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
7014 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
7017 if (mem_cgroup_disabled())
7020 /* Page cache replacement: new page already charged? */
7021 if (newpage->mem_cgroup)
7024 /* Swapcache readahead pages can get replaced before being charged */
7025 memcg = oldpage->mem_cgroup;
7029 /* Force-charge the new page. The old one will be freed soon */
7030 nr_pages = thp_nr_pages(newpage);
7032 page_counter_charge(&memcg->memory, nr_pages);
7033 if (do_memsw_account())
7034 page_counter_charge(&memcg->memsw, nr_pages);
7036 css_get(&memcg->css);
7037 commit_charge(newpage, memcg);
7039 local_irq_save(flags);
7040 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
7041 memcg_check_events(memcg, newpage);
7042 local_irq_restore(flags);
7045 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
7046 EXPORT_SYMBOL(memcg_sockets_enabled_key);
7048 void mem_cgroup_sk_alloc(struct sock *sk)
7050 struct mem_cgroup *memcg;
7052 if (!mem_cgroup_sockets_enabled)
7055 /* Do not associate the sock with unrelated interrupted task's memcg. */
7060 memcg = mem_cgroup_from_task(current);
7061 if (memcg == root_mem_cgroup)
7063 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
7065 if (css_tryget(&memcg->css))
7066 sk->sk_memcg = memcg;
7071 void mem_cgroup_sk_free(struct sock *sk)
7074 css_put(&sk->sk_memcg->css);
7078 * mem_cgroup_charge_skmem - charge socket memory
7079 * @memcg: memcg to charge
7080 * @nr_pages: number of pages to charge
7082 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7083 * @memcg's configured limit, %false if the charge had to be forced.
7085 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7087 gfp_t gfp_mask = GFP_KERNEL;
7089 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7090 struct page_counter *fail;
7092 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7093 memcg->tcpmem_pressure = 0;
7096 page_counter_charge(&memcg->tcpmem, nr_pages);
7097 memcg->tcpmem_pressure = 1;
7101 /* Don't block in the packet receive path */
7103 gfp_mask = GFP_NOWAIT;
7105 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7107 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
7110 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
7115 * mem_cgroup_uncharge_skmem - uncharge socket memory
7116 * @memcg: memcg to uncharge
7117 * @nr_pages: number of pages to uncharge
7119 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7121 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7122 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7126 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7128 refill_stock(memcg, nr_pages);
7131 static int __init cgroup_memory(char *s)
7135 while ((token = strsep(&s, ",")) != NULL) {
7138 if (!strcmp(token, "nosocket"))
7139 cgroup_memory_nosocket = true;
7140 if (!strcmp(token, "nokmem"))
7141 cgroup_memory_nokmem = true;
7145 __setup("cgroup.memory=", cgroup_memory);
7148 * subsys_initcall() for memory controller.
7150 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7151 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7152 * basically everything that doesn't depend on a specific mem_cgroup structure
7153 * should be initialized from here.
7155 static int __init mem_cgroup_init(void)
7159 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7160 memcg_hotplug_cpu_dead);
7162 for_each_possible_cpu(cpu)
7163 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7166 for_each_node(node) {
7167 struct mem_cgroup_tree_per_node *rtpn;
7169 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7170 node_online(node) ? node : NUMA_NO_NODE);
7172 rtpn->rb_root = RB_ROOT;
7173 rtpn->rb_rightmost = NULL;
7174 spin_lock_init(&rtpn->lock);
7175 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7180 subsys_initcall(mem_cgroup_init);
7182 #ifdef CONFIG_MEMCG_SWAP
7183 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7185 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7187 * The root cgroup cannot be destroyed, so it's refcount must
7190 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7194 memcg = parent_mem_cgroup(memcg);
7196 memcg = root_mem_cgroup;
7202 * mem_cgroup_swapout - transfer a memsw charge to swap
7203 * @page: page whose memsw charge to transfer
7204 * @entry: swap entry to move the charge to
7206 * Transfer the memsw charge of @page to @entry.
7208 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7210 struct mem_cgroup *memcg, *swap_memcg;
7211 unsigned int nr_entries;
7212 unsigned short oldid;
7214 VM_BUG_ON_PAGE(PageLRU(page), page);
7215 VM_BUG_ON_PAGE(page_count(page), page);
7217 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7220 memcg = page->mem_cgroup;
7222 /* Readahead page, never charged */
7227 * In case the memcg owning these pages has been offlined and doesn't
7228 * have an ID allocated to it anymore, charge the closest online
7229 * ancestor for the swap instead and transfer the memory+swap charge.
7231 swap_memcg = mem_cgroup_id_get_online(memcg);
7232 nr_entries = thp_nr_pages(page);
7233 /* Get references for the tail pages, too */
7235 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7236 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7238 VM_BUG_ON_PAGE(oldid, page);
7239 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7241 page->mem_cgroup = NULL;
7243 if (!mem_cgroup_is_root(memcg))
7244 page_counter_uncharge(&memcg->memory, nr_entries);
7246 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7247 if (!mem_cgroup_is_root(swap_memcg))
7248 page_counter_charge(&swap_memcg->memsw, nr_entries);
7249 page_counter_uncharge(&memcg->memsw, nr_entries);
7253 * Interrupts should be disabled here because the caller holds the
7254 * i_pages lock which is taken with interrupts-off. It is
7255 * important here to have the interrupts disabled because it is the
7256 * only synchronisation we have for updating the per-CPU variables.
7258 VM_BUG_ON(!irqs_disabled());
7259 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7260 memcg_check_events(memcg, page);
7262 css_put(&memcg->css);
7266 * mem_cgroup_try_charge_swap - try charging swap space for a page
7267 * @page: page being added to swap
7268 * @entry: swap entry to charge
7270 * Try to charge @page's memcg for the swap space at @entry.
7272 * Returns 0 on success, -ENOMEM on failure.
7274 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7276 unsigned int nr_pages = thp_nr_pages(page);
7277 struct page_counter *counter;
7278 struct mem_cgroup *memcg;
7279 unsigned short oldid;
7281 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7284 memcg = page->mem_cgroup;
7286 /* Readahead page, never charged */
7291 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7295 memcg = mem_cgroup_id_get_online(memcg);
7297 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7298 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7299 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7300 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7301 mem_cgroup_id_put(memcg);
7305 /* Get references for the tail pages, too */
7307 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7308 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7309 VM_BUG_ON_PAGE(oldid, page);
7310 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7316 * mem_cgroup_uncharge_swap - uncharge swap space
7317 * @entry: swap entry to uncharge
7318 * @nr_pages: the amount of swap space to uncharge
7320 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7322 struct mem_cgroup *memcg;
7325 id = swap_cgroup_record(entry, 0, nr_pages);
7327 memcg = mem_cgroup_from_id(id);
7329 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7330 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7331 page_counter_uncharge(&memcg->swap, nr_pages);
7333 page_counter_uncharge(&memcg->memsw, nr_pages);
7335 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7336 mem_cgroup_id_put_many(memcg, nr_pages);
7341 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7343 long nr_swap_pages = get_nr_swap_pages();
7345 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7346 return nr_swap_pages;
7347 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7348 nr_swap_pages = min_t(long, nr_swap_pages,
7349 READ_ONCE(memcg->swap.max) -
7350 page_counter_read(&memcg->swap));
7351 return nr_swap_pages;
7354 bool mem_cgroup_swap_full(struct page *page)
7356 struct mem_cgroup *memcg;
7358 VM_BUG_ON_PAGE(!PageLocked(page), page);
7362 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7365 memcg = page->mem_cgroup;
7369 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7370 unsigned long usage = page_counter_read(&memcg->swap);
7372 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7373 usage * 2 >= READ_ONCE(memcg->swap.max))
7380 static int __init setup_swap_account(char *s)
7382 if (!strcmp(s, "1"))
7383 cgroup_memory_noswap = 0;
7384 else if (!strcmp(s, "0"))
7385 cgroup_memory_noswap = 1;
7388 __setup("swapaccount=", setup_swap_account);
7390 static u64 swap_current_read(struct cgroup_subsys_state *css,
7393 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7395 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7398 static int swap_high_show(struct seq_file *m, void *v)
7400 return seq_puts_memcg_tunable(m,
7401 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7404 static ssize_t swap_high_write(struct kernfs_open_file *of,
7405 char *buf, size_t nbytes, loff_t off)
7407 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7411 buf = strstrip(buf);
7412 err = page_counter_memparse(buf, "max", &high);
7416 page_counter_set_high(&memcg->swap, high);
7421 static int swap_max_show(struct seq_file *m, void *v)
7423 return seq_puts_memcg_tunable(m,
7424 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7427 static ssize_t swap_max_write(struct kernfs_open_file *of,
7428 char *buf, size_t nbytes, loff_t off)
7430 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7434 buf = strstrip(buf);
7435 err = page_counter_memparse(buf, "max", &max);
7439 xchg(&memcg->swap.max, max);
7444 static int swap_events_show(struct seq_file *m, void *v)
7446 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7448 seq_printf(m, "high %lu\n",
7449 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7450 seq_printf(m, "max %lu\n",
7451 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7452 seq_printf(m, "fail %lu\n",
7453 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7458 static struct cftype swap_files[] = {
7460 .name = "swap.current",
7461 .flags = CFTYPE_NOT_ON_ROOT,
7462 .read_u64 = swap_current_read,
7465 .name = "swap.high",
7466 .flags = CFTYPE_NOT_ON_ROOT,
7467 .seq_show = swap_high_show,
7468 .write = swap_high_write,
7472 .flags = CFTYPE_NOT_ON_ROOT,
7473 .seq_show = swap_max_show,
7474 .write = swap_max_write,
7477 .name = "swap.events",
7478 .flags = CFTYPE_NOT_ON_ROOT,
7479 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7480 .seq_show = swap_events_show,
7485 static struct cftype memsw_files[] = {
7487 .name = "memsw.usage_in_bytes",
7488 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7489 .read_u64 = mem_cgroup_read_u64,
7492 .name = "memsw.max_usage_in_bytes",
7493 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7494 .write = mem_cgroup_reset,
7495 .read_u64 = mem_cgroup_read_u64,
7498 .name = "memsw.limit_in_bytes",
7499 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7500 .write = mem_cgroup_write,
7501 .read_u64 = mem_cgroup_read_u64,
7504 .name = "memsw.failcnt",
7505 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7506 .write = mem_cgroup_reset,
7507 .read_u64 = mem_cgroup_read_u64,
7509 { }, /* terminate */
7513 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7514 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7515 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7516 * boot parameter. This may result in premature OOPS inside
7517 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7519 static int __init mem_cgroup_swap_init(void)
7521 /* No memory control -> no swap control */
7522 if (mem_cgroup_disabled())
7523 cgroup_memory_noswap = true;
7525 if (cgroup_memory_noswap)
7528 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7529 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7533 core_initcall(mem_cgroup_swap_init);
7535 #endif /* CONFIG_MEMCG_SWAP */