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 #define MEM_CGROUP_RECLAIM_RETRIES 5
78 /* Socket memory accounting disabled? */
79 static bool cgroup_memory_nosocket;
81 /* Kernel memory accounting disabled? */
82 static bool cgroup_memory_nokmem;
84 /* Whether the swap controller is active */
85 #ifdef CONFIG_MEMCG_SWAP
86 bool cgroup_memory_noswap __read_mostly;
88 #define cgroup_memory_noswap 1
91 #ifdef CONFIG_CGROUP_WRITEBACK
92 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
95 /* Whether legacy memory+swap accounting is active */
96 static bool do_memsw_account(void)
98 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_noswap;
101 #define THRESHOLDS_EVENTS_TARGET 128
102 #define SOFTLIMIT_EVENTS_TARGET 1024
105 * Cgroups above their limits are maintained in a RB-Tree, independent of
106 * their hierarchy representation
109 struct mem_cgroup_tree_per_node {
110 struct rb_root rb_root;
111 struct rb_node *rb_rightmost;
115 struct mem_cgroup_tree {
116 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
119 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
122 struct mem_cgroup_eventfd_list {
123 struct list_head list;
124 struct eventfd_ctx *eventfd;
128 * cgroup_event represents events which userspace want to receive.
130 struct mem_cgroup_event {
132 * memcg which the event belongs to.
134 struct mem_cgroup *memcg;
136 * eventfd to signal userspace about the event.
138 struct eventfd_ctx *eventfd;
140 * Each of these stored in a list by the cgroup.
142 struct list_head list;
144 * register_event() callback will be used to add new userspace
145 * waiter for changes related to this event. Use eventfd_signal()
146 * on eventfd to send notification to userspace.
148 int (*register_event)(struct mem_cgroup *memcg,
149 struct eventfd_ctx *eventfd, const char *args);
151 * unregister_event() callback will be called when userspace closes
152 * the eventfd or on cgroup removing. This callback must be set,
153 * if you want provide notification functionality.
155 void (*unregister_event)(struct mem_cgroup *memcg,
156 struct eventfd_ctx *eventfd);
158 * All fields below needed to unregister event when
159 * userspace closes eventfd.
162 wait_queue_head_t *wqh;
163 wait_queue_entry_t wait;
164 struct work_struct remove;
167 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
168 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
170 /* Stuffs for move charges at task migration. */
172 * Types of charges to be moved.
174 #define MOVE_ANON 0x1U
175 #define MOVE_FILE 0x2U
176 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
178 /* "mc" and its members are protected by cgroup_mutex */
179 static struct move_charge_struct {
180 spinlock_t lock; /* for from, to */
181 struct mm_struct *mm;
182 struct mem_cgroup *from;
183 struct mem_cgroup *to;
185 unsigned long precharge;
186 unsigned long moved_charge;
187 unsigned long moved_swap;
188 struct task_struct *moving_task; /* a task moving charges */
189 wait_queue_head_t waitq; /* a waitq for other context */
191 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
192 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
196 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
197 * limit reclaim to prevent infinite loops, if they ever occur.
199 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
200 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
203 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
204 MEM_CGROUP_CHARGE_TYPE_ANON,
205 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
206 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
210 /* for encoding cft->private value on file */
219 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
220 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
221 #define MEMFILE_ATTR(val) ((val) & 0xffff)
222 /* Used for OOM nofiier */
223 #define OOM_CONTROL (0)
226 * Iteration constructs for visiting all cgroups (under a tree). If
227 * loops are exited prematurely (break), mem_cgroup_iter_break() must
228 * be used for reference counting.
230 #define for_each_mem_cgroup_tree(iter, root) \
231 for (iter = mem_cgroup_iter(root, NULL, NULL); \
233 iter = mem_cgroup_iter(root, iter, NULL))
235 #define for_each_mem_cgroup(iter) \
236 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
238 iter = mem_cgroup_iter(NULL, iter, NULL))
240 static inline bool should_force_charge(void)
242 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
243 (current->flags & PF_EXITING);
246 /* Some nice accessors for the vmpressure. */
247 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
250 memcg = root_mem_cgroup;
251 return &memcg->vmpressure;
254 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
256 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
259 #ifdef CONFIG_MEMCG_KMEM
260 extern spinlock_t css_set_lock;
262 static void obj_cgroup_release(struct percpu_ref *ref)
264 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
265 struct mem_cgroup *memcg;
266 unsigned int nr_bytes;
267 unsigned int nr_pages;
271 * At this point all allocated objects are freed, and
272 * objcg->nr_charged_bytes can't have an arbitrary byte value.
273 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
275 * The following sequence can lead to it:
276 * 1) CPU0: objcg == stock->cached_objcg
277 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
278 * PAGE_SIZE bytes are charged
279 * 3) CPU1: a process from another memcg is allocating something,
280 * the stock if flushed,
281 * objcg->nr_charged_bytes = PAGE_SIZE - 92
282 * 5) CPU0: we do release this object,
283 * 92 bytes are added to stock->nr_bytes
284 * 6) CPU0: stock is flushed,
285 * 92 bytes are added to objcg->nr_charged_bytes
287 * In the result, nr_charged_bytes == PAGE_SIZE.
288 * This page will be uncharged in obj_cgroup_release().
290 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
291 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
292 nr_pages = nr_bytes >> PAGE_SHIFT;
294 spin_lock_irqsave(&css_set_lock, flags);
295 memcg = obj_cgroup_memcg(objcg);
297 __memcg_kmem_uncharge(memcg, nr_pages);
298 list_del(&objcg->list);
299 mem_cgroup_put(memcg);
300 spin_unlock_irqrestore(&css_set_lock, flags);
302 percpu_ref_exit(ref);
303 kfree_rcu(objcg, rcu);
306 static struct obj_cgroup *obj_cgroup_alloc(void)
308 struct obj_cgroup *objcg;
311 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
315 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
321 INIT_LIST_HEAD(&objcg->list);
325 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
326 struct mem_cgroup *parent)
328 struct obj_cgroup *objcg, *iter;
330 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
332 spin_lock_irq(&css_set_lock);
334 /* Move active objcg to the parent's list */
335 xchg(&objcg->memcg, parent);
336 css_get(&parent->css);
337 list_add(&objcg->list, &parent->objcg_list);
339 /* Move already reparented objcgs to the parent's list */
340 list_for_each_entry(iter, &memcg->objcg_list, list) {
341 css_get(&parent->css);
342 xchg(&iter->memcg, parent);
343 css_put(&memcg->css);
345 list_splice(&memcg->objcg_list, &parent->objcg_list);
347 spin_unlock_irq(&css_set_lock);
349 percpu_ref_kill(&objcg->refcnt);
353 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
354 * The main reason for not using cgroup id for this:
355 * this works better in sparse environments, where we have a lot of memcgs,
356 * but only a few kmem-limited. Or also, if we have, for instance, 200
357 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
358 * 200 entry array for that.
360 * The current size of the caches array is stored in memcg_nr_cache_ids. It
361 * will double each time we have to increase it.
363 static DEFINE_IDA(memcg_cache_ida);
364 int memcg_nr_cache_ids;
366 /* Protects memcg_nr_cache_ids */
367 static DECLARE_RWSEM(memcg_cache_ids_sem);
369 void memcg_get_cache_ids(void)
371 down_read(&memcg_cache_ids_sem);
374 void memcg_put_cache_ids(void)
376 up_read(&memcg_cache_ids_sem);
380 * MIN_SIZE is different than 1, because we would like to avoid going through
381 * the alloc/free process all the time. In a small machine, 4 kmem-limited
382 * cgroups is a reasonable guess. In the future, it could be a parameter or
383 * tunable, but that is strictly not necessary.
385 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
386 * this constant directly from cgroup, but it is understandable that this is
387 * better kept as an internal representation in cgroup.c. In any case, the
388 * cgrp_id space is not getting any smaller, and we don't have to necessarily
389 * increase ours as well if it increases.
391 #define MEMCG_CACHES_MIN_SIZE 4
392 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
395 * A lot of the calls to the cache allocation functions are expected to be
396 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
397 * conditional to this static branch, we'll have to allow modules that does
398 * kmem_cache_alloc and the such to see this symbol as well
400 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
401 EXPORT_SYMBOL(memcg_kmem_enabled_key);
403 struct workqueue_struct *memcg_kmem_cache_wq;
406 static int memcg_shrinker_map_size;
407 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
409 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
411 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
414 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
415 int size, int old_size)
417 struct memcg_shrinker_map *new, *old;
420 lockdep_assert_held(&memcg_shrinker_map_mutex);
423 old = rcu_dereference_protected(
424 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
425 /* Not yet online memcg */
429 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
433 /* Set all old bits, clear all new bits */
434 memset(new->map, (int)0xff, old_size);
435 memset((void *)new->map + old_size, 0, size - old_size);
437 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
438 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
444 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
446 struct mem_cgroup_per_node *pn;
447 struct memcg_shrinker_map *map;
450 if (mem_cgroup_is_root(memcg))
454 pn = mem_cgroup_nodeinfo(memcg, nid);
455 map = rcu_dereference_protected(pn->shrinker_map, true);
458 rcu_assign_pointer(pn->shrinker_map, NULL);
462 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
464 struct memcg_shrinker_map *map;
465 int nid, size, ret = 0;
467 if (mem_cgroup_is_root(memcg))
470 mutex_lock(&memcg_shrinker_map_mutex);
471 size = memcg_shrinker_map_size;
473 map = kvzalloc_node(sizeof(*map) + size, GFP_KERNEL, nid);
475 memcg_free_shrinker_maps(memcg);
479 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
481 mutex_unlock(&memcg_shrinker_map_mutex);
486 int memcg_expand_shrinker_maps(int new_id)
488 int size, old_size, ret = 0;
489 struct mem_cgroup *memcg;
491 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
492 old_size = memcg_shrinker_map_size;
493 if (size <= old_size)
496 mutex_lock(&memcg_shrinker_map_mutex);
497 if (!root_mem_cgroup)
500 for_each_mem_cgroup(memcg) {
501 if (mem_cgroup_is_root(memcg))
503 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
505 mem_cgroup_iter_break(NULL, memcg);
511 memcg_shrinker_map_size = size;
512 mutex_unlock(&memcg_shrinker_map_mutex);
516 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
518 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
519 struct memcg_shrinker_map *map;
522 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
523 /* Pairs with smp mb in shrink_slab() */
524 smp_mb__before_atomic();
525 set_bit(shrinker_id, map->map);
531 * mem_cgroup_css_from_page - css of the memcg associated with a page
532 * @page: page of interest
534 * If memcg is bound to the default hierarchy, css of the memcg associated
535 * with @page is returned. The returned css remains associated with @page
536 * until it is released.
538 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
541 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
543 struct mem_cgroup *memcg;
545 memcg = page->mem_cgroup;
547 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
548 memcg = root_mem_cgroup;
554 * page_cgroup_ino - return inode number of the memcg a page is charged to
557 * Look up the closest online ancestor of the memory cgroup @page is charged to
558 * and return its inode number or 0 if @page is not charged to any cgroup. It
559 * is safe to call this function without holding a reference to @page.
561 * Note, this function is inherently racy, because there is nothing to prevent
562 * the cgroup inode from getting torn down and potentially reallocated a moment
563 * after page_cgroup_ino() returns, so it only should be used by callers that
564 * do not care (such as procfs interfaces).
566 ino_t page_cgroup_ino(struct page *page)
568 struct mem_cgroup *memcg;
569 unsigned long ino = 0;
572 if (PageSlab(page) && !PageTail(page)) {
573 memcg = memcg_from_slab_page(page);
575 memcg = page->mem_cgroup;
578 * The lowest bit set means that memcg isn't a valid
579 * memcg pointer, but a obj_cgroups pointer.
580 * In this case the page is shared and doesn't belong
581 * to any specific memory cgroup.
583 if ((unsigned long) memcg & 0x1UL)
587 while (memcg && !(memcg->css.flags & CSS_ONLINE))
588 memcg = parent_mem_cgroup(memcg);
590 ino = cgroup_ino(memcg->css.cgroup);
595 static struct mem_cgroup_per_node *
596 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
598 int nid = page_to_nid(page);
600 return memcg->nodeinfo[nid];
603 static struct mem_cgroup_tree_per_node *
604 soft_limit_tree_node(int nid)
606 return soft_limit_tree.rb_tree_per_node[nid];
609 static struct mem_cgroup_tree_per_node *
610 soft_limit_tree_from_page(struct page *page)
612 int nid = page_to_nid(page);
614 return soft_limit_tree.rb_tree_per_node[nid];
617 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
618 struct mem_cgroup_tree_per_node *mctz,
619 unsigned long new_usage_in_excess)
621 struct rb_node **p = &mctz->rb_root.rb_node;
622 struct rb_node *parent = NULL;
623 struct mem_cgroup_per_node *mz_node;
624 bool rightmost = true;
629 mz->usage_in_excess = new_usage_in_excess;
630 if (!mz->usage_in_excess)
634 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
636 if (mz->usage_in_excess < mz_node->usage_in_excess) {
642 * We can't avoid mem cgroups that are over their soft
643 * limit by the same amount
645 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
650 mctz->rb_rightmost = &mz->tree_node;
652 rb_link_node(&mz->tree_node, parent, p);
653 rb_insert_color(&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)
663 if (&mz->tree_node == mctz->rb_rightmost)
664 mctz->rb_rightmost = rb_prev(&mz->tree_node);
666 rb_erase(&mz->tree_node, &mctz->rb_root);
670 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
671 struct mem_cgroup_tree_per_node *mctz)
675 spin_lock_irqsave(&mctz->lock, flags);
676 __mem_cgroup_remove_exceeded(mz, mctz);
677 spin_unlock_irqrestore(&mctz->lock, flags);
680 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
682 unsigned long nr_pages = page_counter_read(&memcg->memory);
683 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
684 unsigned long excess = 0;
686 if (nr_pages > soft_limit)
687 excess = nr_pages - soft_limit;
692 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
694 unsigned long excess;
695 struct mem_cgroup_per_node *mz;
696 struct mem_cgroup_tree_per_node *mctz;
698 mctz = soft_limit_tree_from_page(page);
702 * Necessary to update all ancestors when hierarchy is used.
703 * because their event counter is not touched.
705 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
706 mz = mem_cgroup_page_nodeinfo(memcg, page);
707 excess = soft_limit_excess(memcg);
709 * We have to update the tree if mz is on RB-tree or
710 * mem is over its softlimit.
712 if (excess || mz->on_tree) {
715 spin_lock_irqsave(&mctz->lock, flags);
716 /* if on-tree, remove it */
718 __mem_cgroup_remove_exceeded(mz, mctz);
720 * Insert again. mz->usage_in_excess will be updated.
721 * If excess is 0, no tree ops.
723 __mem_cgroup_insert_exceeded(mz, mctz, excess);
724 spin_unlock_irqrestore(&mctz->lock, flags);
729 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
731 struct mem_cgroup_tree_per_node *mctz;
732 struct mem_cgroup_per_node *mz;
736 mz = mem_cgroup_nodeinfo(memcg, nid);
737 mctz = soft_limit_tree_node(nid);
739 mem_cgroup_remove_exceeded(mz, mctz);
743 static struct mem_cgroup_per_node *
744 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
746 struct mem_cgroup_per_node *mz;
750 if (!mctz->rb_rightmost)
751 goto done; /* Nothing to reclaim from */
753 mz = rb_entry(mctz->rb_rightmost,
754 struct mem_cgroup_per_node, tree_node);
756 * Remove the node now but someone else can add it back,
757 * we will to add it back at the end of reclaim to its correct
758 * position in the tree.
760 __mem_cgroup_remove_exceeded(mz, mctz);
761 if (!soft_limit_excess(mz->memcg) ||
762 !css_tryget(&mz->memcg->css))
768 static struct mem_cgroup_per_node *
769 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
771 struct mem_cgroup_per_node *mz;
773 spin_lock_irq(&mctz->lock);
774 mz = __mem_cgroup_largest_soft_limit_node(mctz);
775 spin_unlock_irq(&mctz->lock);
780 * __mod_memcg_state - update cgroup memory statistics
781 * @memcg: the memory cgroup
782 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
783 * @val: delta to add to the counter, can be negative
785 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
787 long x, threshold = MEMCG_CHARGE_BATCH;
789 if (mem_cgroup_disabled())
792 if (vmstat_item_in_bytes(idx))
793 threshold <<= PAGE_SHIFT;
795 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
796 if (unlikely(abs(x) > threshold)) {
797 struct mem_cgroup *mi;
800 * Batch local counters to keep them in sync with
801 * the hierarchical ones.
803 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
804 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
805 atomic_long_add(x, &mi->vmstats[idx]);
808 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
811 static struct mem_cgroup_per_node *
812 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
814 struct mem_cgroup *parent;
816 parent = parent_mem_cgroup(pn->memcg);
819 return mem_cgroup_nodeinfo(parent, nid);
822 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
825 struct mem_cgroup_per_node *pn;
826 struct mem_cgroup *memcg;
827 long x, threshold = MEMCG_CHARGE_BATCH;
829 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
833 __mod_memcg_state(memcg, idx, val);
836 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
838 if (vmstat_item_in_bytes(idx))
839 threshold <<= PAGE_SHIFT;
841 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
842 if (unlikely(abs(x) > threshold)) {
843 pg_data_t *pgdat = lruvec_pgdat(lruvec);
844 struct mem_cgroup_per_node *pi;
846 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
847 atomic_long_add(x, &pi->lruvec_stat[idx]);
850 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
854 * __mod_lruvec_state - update lruvec memory statistics
855 * @lruvec: the lruvec
856 * @idx: the stat item
857 * @val: delta to add to the counter, can be negative
859 * The lruvec is the intersection of the NUMA node and a cgroup. This
860 * function updates the all three counters that are affected by a
861 * change of state at this level: per-node, per-cgroup, per-lruvec.
863 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
867 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
869 /* Update memcg and lruvec */
870 if (!mem_cgroup_disabled())
871 __mod_memcg_lruvec_state(lruvec, idx, val);
874 void __mod_lruvec_slab_state(void *p, enum node_stat_item idx, int val)
876 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
877 struct mem_cgroup *memcg;
878 struct lruvec *lruvec;
881 memcg = mem_cgroup_from_obj(p);
883 /* Untracked pages have no memcg, no lruvec. Update only the node */
884 if (!memcg || memcg == root_mem_cgroup) {
885 __mod_node_page_state(pgdat, idx, val);
887 lruvec = mem_cgroup_lruvec(memcg, pgdat);
888 __mod_lruvec_state(lruvec, idx, val);
893 void mod_memcg_obj_state(void *p, int idx, int val)
895 struct mem_cgroup *memcg;
898 memcg = mem_cgroup_from_obj(p);
900 mod_memcg_state(memcg, idx, val);
905 * __count_memcg_events - account VM events in a cgroup
906 * @memcg: the memory cgroup
907 * @idx: the event item
908 * @count: the number of events that occured
910 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
915 if (mem_cgroup_disabled())
918 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
919 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
920 struct mem_cgroup *mi;
923 * Batch local counters to keep them in sync with
924 * the hierarchical ones.
926 __this_cpu_add(memcg->vmstats_local->events[idx], x);
927 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
928 atomic_long_add(x, &mi->vmevents[idx]);
931 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
934 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
936 return atomic_long_read(&memcg->vmevents[event]);
939 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
944 for_each_possible_cpu(cpu)
945 x += per_cpu(memcg->vmstats_local->events[event], cpu);
949 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
953 /* pagein of a big page is an event. So, ignore page size */
955 __count_memcg_events(memcg, PGPGIN, 1);
957 __count_memcg_events(memcg, PGPGOUT, 1);
958 nr_pages = -nr_pages; /* for event */
961 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
964 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
965 enum mem_cgroup_events_target target)
967 unsigned long val, next;
969 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
970 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
971 /* from time_after() in jiffies.h */
972 if ((long)(next - val) < 0) {
974 case MEM_CGROUP_TARGET_THRESH:
975 next = val + THRESHOLDS_EVENTS_TARGET;
977 case MEM_CGROUP_TARGET_SOFTLIMIT:
978 next = val + SOFTLIMIT_EVENTS_TARGET;
983 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
990 * Check events in order.
993 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
995 /* threshold event is triggered in finer grain than soft limit */
996 if (unlikely(mem_cgroup_event_ratelimit(memcg,
997 MEM_CGROUP_TARGET_THRESH))) {
1000 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1001 MEM_CGROUP_TARGET_SOFTLIMIT);
1002 mem_cgroup_threshold(memcg);
1003 if (unlikely(do_softlimit))
1004 mem_cgroup_update_tree(memcg, page);
1008 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1011 * mm_update_next_owner() may clear mm->owner to NULL
1012 * if it races with swapoff, page migration, etc.
1013 * So this can be called with p == NULL.
1018 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1020 EXPORT_SYMBOL(mem_cgroup_from_task);
1023 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1024 * @mm: mm from which memcg should be extracted. It can be NULL.
1026 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
1027 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
1030 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1032 struct mem_cgroup *memcg;
1034 if (mem_cgroup_disabled())
1040 * Page cache insertions can happen withou an
1041 * actual mm context, e.g. during disk probing
1042 * on boot, loopback IO, acct() writes etc.
1045 memcg = root_mem_cgroup;
1047 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1048 if (unlikely(!memcg))
1049 memcg = root_mem_cgroup;
1051 } while (!css_tryget(&memcg->css));
1055 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1058 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
1059 * @page: page from which memcg should be extracted.
1061 * Obtain a reference on page->memcg and returns it if successful. Otherwise
1062 * root_mem_cgroup is returned.
1064 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
1066 struct mem_cgroup *memcg = page->mem_cgroup;
1068 if (mem_cgroup_disabled())
1072 /* Page should not get uncharged and freed memcg under us. */
1073 if (!memcg || WARN_ON_ONCE(!css_tryget(&memcg->css)))
1074 memcg = root_mem_cgroup;
1078 EXPORT_SYMBOL(get_mem_cgroup_from_page);
1081 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
1083 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
1085 if (unlikely(current->active_memcg)) {
1086 struct mem_cgroup *memcg;
1089 /* current->active_memcg must hold a ref. */
1090 if (WARN_ON_ONCE(!css_tryget(¤t->active_memcg->css)))
1091 memcg = root_mem_cgroup;
1093 memcg = current->active_memcg;
1097 return get_mem_cgroup_from_mm(current->mm);
1101 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1102 * @root: hierarchy root
1103 * @prev: previously returned memcg, NULL on first invocation
1104 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1106 * Returns references to children of the hierarchy below @root, or
1107 * @root itself, or %NULL after a full round-trip.
1109 * Caller must pass the return value in @prev on subsequent
1110 * invocations for reference counting, or use mem_cgroup_iter_break()
1111 * to cancel a hierarchy walk before the round-trip is complete.
1113 * Reclaimers can specify a node and a priority level in @reclaim to
1114 * divide up the memcgs in the hierarchy among all concurrent
1115 * reclaimers operating on the same node and priority.
1117 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1118 struct mem_cgroup *prev,
1119 struct mem_cgroup_reclaim_cookie *reclaim)
1121 struct mem_cgroup_reclaim_iter *iter;
1122 struct cgroup_subsys_state *css = NULL;
1123 struct mem_cgroup *memcg = NULL;
1124 struct mem_cgroup *pos = NULL;
1126 if (mem_cgroup_disabled())
1130 root = root_mem_cgroup;
1132 if (prev && !reclaim)
1135 if (!root->use_hierarchy && root != root_mem_cgroup) {
1144 struct mem_cgroup_per_node *mz;
1146 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1149 if (prev && reclaim->generation != iter->generation)
1153 pos = READ_ONCE(iter->position);
1154 if (!pos || css_tryget(&pos->css))
1157 * css reference reached zero, so iter->position will
1158 * be cleared by ->css_released. However, we should not
1159 * rely on this happening soon, because ->css_released
1160 * is called from a work queue, and by busy-waiting we
1161 * might block it. So we clear iter->position right
1164 (void)cmpxchg(&iter->position, pos, NULL);
1172 css = css_next_descendant_pre(css, &root->css);
1175 * Reclaimers share the hierarchy walk, and a
1176 * new one might jump in right at the end of
1177 * the hierarchy - make sure they see at least
1178 * one group and restart from the beginning.
1186 * Verify the css and acquire a reference. The root
1187 * is provided by the caller, so we know it's alive
1188 * and kicking, and don't take an extra reference.
1190 memcg = mem_cgroup_from_css(css);
1192 if (css == &root->css)
1195 if (css_tryget(css))
1203 * The position could have already been updated by a competing
1204 * thread, so check that the value hasn't changed since we read
1205 * it to avoid reclaiming from the same cgroup twice.
1207 (void)cmpxchg(&iter->position, pos, memcg);
1215 reclaim->generation = iter->generation;
1221 if (prev && prev != root)
1222 css_put(&prev->css);
1228 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1229 * @root: hierarchy root
1230 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1232 void mem_cgroup_iter_break(struct mem_cgroup *root,
1233 struct mem_cgroup *prev)
1236 root = root_mem_cgroup;
1237 if (prev && prev != root)
1238 css_put(&prev->css);
1241 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1242 struct mem_cgroup *dead_memcg)
1244 struct mem_cgroup_reclaim_iter *iter;
1245 struct mem_cgroup_per_node *mz;
1248 for_each_node(nid) {
1249 mz = mem_cgroup_nodeinfo(from, nid);
1251 cmpxchg(&iter->position, dead_memcg, NULL);
1255 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1257 struct mem_cgroup *memcg = dead_memcg;
1258 struct mem_cgroup *last;
1261 __invalidate_reclaim_iterators(memcg, dead_memcg);
1263 } while ((memcg = parent_mem_cgroup(memcg)));
1266 * When cgruop1 non-hierarchy mode is used,
1267 * parent_mem_cgroup() does not walk all the way up to the
1268 * cgroup root (root_mem_cgroup). So we have to handle
1269 * dead_memcg from cgroup root separately.
1271 if (last != root_mem_cgroup)
1272 __invalidate_reclaim_iterators(root_mem_cgroup,
1277 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1278 * @memcg: hierarchy root
1279 * @fn: function to call for each task
1280 * @arg: argument passed to @fn
1282 * This function iterates over tasks attached to @memcg or to any of its
1283 * descendants and calls @fn for each task. If @fn returns a non-zero
1284 * value, the function breaks the iteration loop and returns the value.
1285 * Otherwise, it will iterate over all tasks and return 0.
1287 * This function must not be called for the root memory cgroup.
1289 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1290 int (*fn)(struct task_struct *, void *), void *arg)
1292 struct mem_cgroup *iter;
1295 BUG_ON(memcg == root_mem_cgroup);
1297 for_each_mem_cgroup_tree(iter, memcg) {
1298 struct css_task_iter it;
1299 struct task_struct *task;
1301 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1302 while (!ret && (task = css_task_iter_next(&it)))
1303 ret = fn(task, arg);
1304 css_task_iter_end(&it);
1306 mem_cgroup_iter_break(memcg, iter);
1314 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1316 * @pgdat: pgdat of the page
1318 * This function relies on page->mem_cgroup being stable - see the
1319 * access rules in commit_charge().
1321 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1323 struct mem_cgroup_per_node *mz;
1324 struct mem_cgroup *memcg;
1325 struct lruvec *lruvec;
1327 if (mem_cgroup_disabled()) {
1328 lruvec = &pgdat->__lruvec;
1332 memcg = page->mem_cgroup;
1334 * Swapcache readahead pages are added to the LRU - and
1335 * possibly migrated - before they are charged.
1338 memcg = root_mem_cgroup;
1340 mz = mem_cgroup_page_nodeinfo(memcg, page);
1341 lruvec = &mz->lruvec;
1344 * Since a node can be onlined after the mem_cgroup was created,
1345 * we have to be prepared to initialize lruvec->zone here;
1346 * and if offlined then reonlined, we need to reinitialize it.
1348 if (unlikely(lruvec->pgdat != pgdat))
1349 lruvec->pgdat = pgdat;
1354 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1355 * @lruvec: mem_cgroup per zone lru vector
1356 * @lru: index of lru list the page is sitting on
1357 * @zid: zone id of the accounted pages
1358 * @nr_pages: positive when adding or negative when removing
1360 * This function must be called under lru_lock, just before a page is added
1361 * to or just after a page is removed from an lru list (that ordering being
1362 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1364 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1365 int zid, int nr_pages)
1367 struct mem_cgroup_per_node *mz;
1368 unsigned long *lru_size;
1371 if (mem_cgroup_disabled())
1374 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1375 lru_size = &mz->lru_zone_size[zid][lru];
1378 *lru_size += nr_pages;
1381 if (WARN_ONCE(size < 0,
1382 "%s(%p, %d, %d): lru_size %ld\n",
1383 __func__, lruvec, lru, nr_pages, size)) {
1389 *lru_size += nr_pages;
1393 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1394 * @memcg: the memory cgroup
1396 * Returns the maximum amount of memory @mem can be charged with, in
1399 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1401 unsigned long margin = 0;
1402 unsigned long count;
1403 unsigned long limit;
1405 count = page_counter_read(&memcg->memory);
1406 limit = READ_ONCE(memcg->memory.max);
1408 margin = limit - count;
1410 if (do_memsw_account()) {
1411 count = page_counter_read(&memcg->memsw);
1412 limit = READ_ONCE(memcg->memsw.max);
1414 margin = min(margin, limit - count);
1423 * A routine for checking "mem" is under move_account() or not.
1425 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1426 * moving cgroups. This is for waiting at high-memory pressure
1429 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1431 struct mem_cgroup *from;
1432 struct mem_cgroup *to;
1435 * Unlike task_move routines, we access mc.to, mc.from not under
1436 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1438 spin_lock(&mc.lock);
1444 ret = mem_cgroup_is_descendant(from, memcg) ||
1445 mem_cgroup_is_descendant(to, memcg);
1447 spin_unlock(&mc.lock);
1451 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1453 if (mc.moving_task && current != mc.moving_task) {
1454 if (mem_cgroup_under_move(memcg)) {
1456 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1457 /* moving charge context might have finished. */
1460 finish_wait(&mc.waitq, &wait);
1467 static char *memory_stat_format(struct mem_cgroup *memcg)
1472 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1477 * Provide statistics on the state of the memory subsystem as
1478 * well as cumulative event counters that show past behavior.
1480 * This list is ordered following a combination of these gradients:
1481 * 1) generic big picture -> specifics and details
1482 * 2) reflecting userspace activity -> reflecting kernel heuristics
1484 * Current memory state:
1487 seq_buf_printf(&s, "anon %llu\n",
1488 (u64)memcg_page_state(memcg, NR_ANON_MAPPED) *
1490 seq_buf_printf(&s, "file %llu\n",
1491 (u64)memcg_page_state(memcg, NR_FILE_PAGES) *
1493 seq_buf_printf(&s, "kernel_stack %llu\n",
1494 (u64)memcg_page_state(memcg, MEMCG_KERNEL_STACK_KB) *
1496 seq_buf_printf(&s, "slab %llu\n",
1497 (u64)(memcg_page_state(memcg, NR_SLAB_RECLAIMABLE_B) +
1498 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE_B)));
1499 seq_buf_printf(&s, "sock %llu\n",
1500 (u64)memcg_page_state(memcg, MEMCG_SOCK) *
1503 seq_buf_printf(&s, "shmem %llu\n",
1504 (u64)memcg_page_state(memcg, NR_SHMEM) *
1506 seq_buf_printf(&s, "file_mapped %llu\n",
1507 (u64)memcg_page_state(memcg, NR_FILE_MAPPED) *
1509 seq_buf_printf(&s, "file_dirty %llu\n",
1510 (u64)memcg_page_state(memcg, NR_FILE_DIRTY) *
1512 seq_buf_printf(&s, "file_writeback %llu\n",
1513 (u64)memcg_page_state(memcg, NR_WRITEBACK) *
1516 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1517 seq_buf_printf(&s, "anon_thp %llu\n",
1518 (u64)memcg_page_state(memcg, NR_ANON_THPS) *
1522 for (i = 0; i < NR_LRU_LISTS; i++)
1523 seq_buf_printf(&s, "%s %llu\n", lru_list_name(i),
1524 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
1527 seq_buf_printf(&s, "slab_reclaimable %llu\n",
1528 (u64)memcg_page_state(memcg, NR_SLAB_RECLAIMABLE_B));
1529 seq_buf_printf(&s, "slab_unreclaimable %llu\n",
1530 (u64)memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE_B));
1532 /* Accumulated memory events */
1534 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1535 memcg_events(memcg, PGFAULT));
1536 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1537 memcg_events(memcg, PGMAJFAULT));
1539 seq_buf_printf(&s, "workingset_refault %lu\n",
1540 memcg_page_state(memcg, WORKINGSET_REFAULT));
1541 seq_buf_printf(&s, "workingset_activate %lu\n",
1542 memcg_page_state(memcg, WORKINGSET_ACTIVATE));
1543 seq_buf_printf(&s, "workingset_restore %lu\n",
1544 memcg_page_state(memcg, WORKINGSET_RESTORE));
1545 seq_buf_printf(&s, "workingset_nodereclaim %lu\n",
1546 memcg_page_state(memcg, WORKINGSET_NODERECLAIM));
1548 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1549 memcg_events(memcg, PGREFILL));
1550 seq_buf_printf(&s, "pgscan %lu\n",
1551 memcg_events(memcg, PGSCAN_KSWAPD) +
1552 memcg_events(memcg, PGSCAN_DIRECT));
1553 seq_buf_printf(&s, "pgsteal %lu\n",
1554 memcg_events(memcg, PGSTEAL_KSWAPD) +
1555 memcg_events(memcg, PGSTEAL_DIRECT));
1556 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1557 memcg_events(memcg, PGACTIVATE));
1558 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1559 memcg_events(memcg, PGDEACTIVATE));
1560 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1561 memcg_events(memcg, PGLAZYFREE));
1562 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1563 memcg_events(memcg, PGLAZYFREED));
1565 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1566 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1567 memcg_events(memcg, THP_FAULT_ALLOC));
1568 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1569 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1570 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1572 /* The above should easily fit into one page */
1573 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1578 #define K(x) ((x) << (PAGE_SHIFT-10))
1580 * mem_cgroup_print_oom_context: Print OOM information relevant to
1581 * memory controller.
1582 * @memcg: The memory cgroup that went over limit
1583 * @p: Task that is going to be killed
1585 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1588 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1593 pr_cont(",oom_memcg=");
1594 pr_cont_cgroup_path(memcg->css.cgroup);
1596 pr_cont(",global_oom");
1598 pr_cont(",task_memcg=");
1599 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1605 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1606 * memory controller.
1607 * @memcg: The memory cgroup that went over limit
1609 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1613 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1614 K((u64)page_counter_read(&memcg->memory)),
1615 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1616 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1617 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1618 K((u64)page_counter_read(&memcg->swap)),
1619 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1621 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1622 K((u64)page_counter_read(&memcg->memsw)),
1623 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1624 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1625 K((u64)page_counter_read(&memcg->kmem)),
1626 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1629 pr_info("Memory cgroup stats for ");
1630 pr_cont_cgroup_path(memcg->css.cgroup);
1632 buf = memory_stat_format(memcg);
1640 * Return the memory (and swap, if configured) limit for a memcg.
1642 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1646 max = READ_ONCE(memcg->memory.max);
1647 if (mem_cgroup_swappiness(memcg)) {
1648 unsigned long memsw_max;
1649 unsigned long swap_max;
1651 memsw_max = memcg->memsw.max;
1652 swap_max = READ_ONCE(memcg->swap.max);
1653 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1654 max = min(max + swap_max, memsw_max);
1659 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1661 return page_counter_read(&memcg->memory);
1664 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1667 struct oom_control oc = {
1671 .gfp_mask = gfp_mask,
1676 if (mutex_lock_killable(&oom_lock))
1679 * A few threads which were not waiting at mutex_lock_killable() can
1680 * fail to bail out. Therefore, check again after holding oom_lock.
1682 ret = should_force_charge() || out_of_memory(&oc);
1683 mutex_unlock(&oom_lock);
1687 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1690 unsigned long *total_scanned)
1692 struct mem_cgroup *victim = NULL;
1695 unsigned long excess;
1696 unsigned long nr_scanned;
1697 struct mem_cgroup_reclaim_cookie reclaim = {
1701 excess = soft_limit_excess(root_memcg);
1704 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1709 * If we have not been able to reclaim
1710 * anything, it might because there are
1711 * no reclaimable pages under this hierarchy
1716 * We want to do more targeted reclaim.
1717 * excess >> 2 is not to excessive so as to
1718 * reclaim too much, nor too less that we keep
1719 * coming back to reclaim from this cgroup
1721 if (total >= (excess >> 2) ||
1722 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1727 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1728 pgdat, &nr_scanned);
1729 *total_scanned += nr_scanned;
1730 if (!soft_limit_excess(root_memcg))
1733 mem_cgroup_iter_break(root_memcg, victim);
1737 #ifdef CONFIG_LOCKDEP
1738 static struct lockdep_map memcg_oom_lock_dep_map = {
1739 .name = "memcg_oom_lock",
1743 static DEFINE_SPINLOCK(memcg_oom_lock);
1746 * Check OOM-Killer is already running under our hierarchy.
1747 * If someone is running, return false.
1749 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1751 struct mem_cgroup *iter, *failed = NULL;
1753 spin_lock(&memcg_oom_lock);
1755 for_each_mem_cgroup_tree(iter, memcg) {
1756 if (iter->oom_lock) {
1758 * this subtree of our hierarchy is already locked
1759 * so we cannot give a lock.
1762 mem_cgroup_iter_break(memcg, iter);
1765 iter->oom_lock = true;
1770 * OK, we failed to lock the whole subtree so we have
1771 * to clean up what we set up to the failing subtree
1773 for_each_mem_cgroup_tree(iter, memcg) {
1774 if (iter == failed) {
1775 mem_cgroup_iter_break(memcg, iter);
1778 iter->oom_lock = false;
1781 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1783 spin_unlock(&memcg_oom_lock);
1788 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1790 struct mem_cgroup *iter;
1792 spin_lock(&memcg_oom_lock);
1793 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1794 for_each_mem_cgroup_tree(iter, memcg)
1795 iter->oom_lock = false;
1796 spin_unlock(&memcg_oom_lock);
1799 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1801 struct mem_cgroup *iter;
1803 spin_lock(&memcg_oom_lock);
1804 for_each_mem_cgroup_tree(iter, memcg)
1806 spin_unlock(&memcg_oom_lock);
1809 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1811 struct mem_cgroup *iter;
1814 * When a new child is created while the hierarchy is under oom,
1815 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1817 spin_lock(&memcg_oom_lock);
1818 for_each_mem_cgroup_tree(iter, memcg)
1819 if (iter->under_oom > 0)
1821 spin_unlock(&memcg_oom_lock);
1824 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1826 struct oom_wait_info {
1827 struct mem_cgroup *memcg;
1828 wait_queue_entry_t wait;
1831 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1832 unsigned mode, int sync, void *arg)
1834 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1835 struct mem_cgroup *oom_wait_memcg;
1836 struct oom_wait_info *oom_wait_info;
1838 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1839 oom_wait_memcg = oom_wait_info->memcg;
1841 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1842 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1844 return autoremove_wake_function(wait, mode, sync, arg);
1847 static void memcg_oom_recover(struct mem_cgroup *memcg)
1850 * For the following lockless ->under_oom test, the only required
1851 * guarantee is that it must see the state asserted by an OOM when
1852 * this function is called as a result of userland actions
1853 * triggered by the notification of the OOM. This is trivially
1854 * achieved by invoking mem_cgroup_mark_under_oom() before
1855 * triggering notification.
1857 if (memcg && memcg->under_oom)
1858 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1868 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1870 enum oom_status ret;
1873 if (order > PAGE_ALLOC_COSTLY_ORDER)
1876 memcg_memory_event(memcg, MEMCG_OOM);
1879 * We are in the middle of the charge context here, so we
1880 * don't want to block when potentially sitting on a callstack
1881 * that holds all kinds of filesystem and mm locks.
1883 * cgroup1 allows disabling the OOM killer and waiting for outside
1884 * handling until the charge can succeed; remember the context and put
1885 * the task to sleep at the end of the page fault when all locks are
1888 * On the other hand, in-kernel OOM killer allows for an async victim
1889 * memory reclaim (oom_reaper) and that means that we are not solely
1890 * relying on the oom victim to make a forward progress and we can
1891 * invoke the oom killer here.
1893 * Please note that mem_cgroup_out_of_memory might fail to find a
1894 * victim and then we have to bail out from the charge path.
1896 if (memcg->oom_kill_disable) {
1897 if (!current->in_user_fault)
1899 css_get(&memcg->css);
1900 current->memcg_in_oom = memcg;
1901 current->memcg_oom_gfp_mask = mask;
1902 current->memcg_oom_order = order;
1907 mem_cgroup_mark_under_oom(memcg);
1909 locked = mem_cgroup_oom_trylock(memcg);
1912 mem_cgroup_oom_notify(memcg);
1914 mem_cgroup_unmark_under_oom(memcg);
1915 if (mem_cgroup_out_of_memory(memcg, mask, order))
1921 mem_cgroup_oom_unlock(memcg);
1927 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1928 * @handle: actually kill/wait or just clean up the OOM state
1930 * This has to be called at the end of a page fault if the memcg OOM
1931 * handler was enabled.
1933 * Memcg supports userspace OOM handling where failed allocations must
1934 * sleep on a waitqueue until the userspace task resolves the
1935 * situation. Sleeping directly in the charge context with all kinds
1936 * of locks held is not a good idea, instead we remember an OOM state
1937 * in the task and mem_cgroup_oom_synchronize() has to be called at
1938 * the end of the page fault to complete the OOM handling.
1940 * Returns %true if an ongoing memcg OOM situation was detected and
1941 * completed, %false otherwise.
1943 bool mem_cgroup_oom_synchronize(bool handle)
1945 struct mem_cgroup *memcg = current->memcg_in_oom;
1946 struct oom_wait_info owait;
1949 /* OOM is global, do not handle */
1956 owait.memcg = memcg;
1957 owait.wait.flags = 0;
1958 owait.wait.func = memcg_oom_wake_function;
1959 owait.wait.private = current;
1960 INIT_LIST_HEAD(&owait.wait.entry);
1962 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1963 mem_cgroup_mark_under_oom(memcg);
1965 locked = mem_cgroup_oom_trylock(memcg);
1968 mem_cgroup_oom_notify(memcg);
1970 if (locked && !memcg->oom_kill_disable) {
1971 mem_cgroup_unmark_under_oom(memcg);
1972 finish_wait(&memcg_oom_waitq, &owait.wait);
1973 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1974 current->memcg_oom_order);
1977 mem_cgroup_unmark_under_oom(memcg);
1978 finish_wait(&memcg_oom_waitq, &owait.wait);
1982 mem_cgroup_oom_unlock(memcg);
1984 * There is no guarantee that an OOM-lock contender
1985 * sees the wakeups triggered by the OOM kill
1986 * uncharges. Wake any sleepers explicitely.
1988 memcg_oom_recover(memcg);
1991 current->memcg_in_oom = NULL;
1992 css_put(&memcg->css);
1997 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1998 * @victim: task to be killed by the OOM killer
1999 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2001 * Returns a pointer to a memory cgroup, which has to be cleaned up
2002 * by killing all belonging OOM-killable tasks.
2004 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2006 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2007 struct mem_cgroup *oom_domain)
2009 struct mem_cgroup *oom_group = NULL;
2010 struct mem_cgroup *memcg;
2012 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2016 oom_domain = root_mem_cgroup;
2020 memcg = mem_cgroup_from_task(victim);
2021 if (memcg == root_mem_cgroup)
2025 * If the victim task has been asynchronously moved to a different
2026 * memory cgroup, we might end up killing tasks outside oom_domain.
2027 * In this case it's better to ignore memory.group.oom.
2029 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2033 * Traverse the memory cgroup hierarchy from the victim task's
2034 * cgroup up to the OOMing cgroup (or root) to find the
2035 * highest-level memory cgroup with oom.group set.
2037 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2038 if (memcg->oom_group)
2041 if (memcg == oom_domain)
2046 css_get(&oom_group->css);
2053 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2055 pr_info("Tasks in ");
2056 pr_cont_cgroup_path(memcg->css.cgroup);
2057 pr_cont(" are going to be killed due to memory.oom.group set\n");
2061 * lock_page_memcg - lock a page->mem_cgroup binding
2064 * This function protects unlocked LRU pages from being moved to
2067 * It ensures lifetime of the returned memcg. Caller is responsible
2068 * for the lifetime of the page; __unlock_page_memcg() is available
2069 * when @page might get freed inside the locked section.
2071 struct mem_cgroup *lock_page_memcg(struct page *page)
2073 struct page *head = compound_head(page); /* rmap on tail pages */
2074 struct mem_cgroup *memcg;
2075 unsigned long flags;
2078 * The RCU lock is held throughout the transaction. The fast
2079 * path can get away without acquiring the memcg->move_lock
2080 * because page moving starts with an RCU grace period.
2082 * The RCU lock also protects the memcg from being freed when
2083 * the page state that is going to change is the only thing
2084 * preventing the page itself from being freed. E.g. writeback
2085 * doesn't hold a page reference and relies on PG_writeback to
2086 * keep off truncation, migration and so forth.
2090 if (mem_cgroup_disabled())
2093 memcg = head->mem_cgroup;
2094 if (unlikely(!memcg))
2097 if (atomic_read(&memcg->moving_account) <= 0)
2100 spin_lock_irqsave(&memcg->move_lock, flags);
2101 if (memcg != head->mem_cgroup) {
2102 spin_unlock_irqrestore(&memcg->move_lock, flags);
2107 * When charge migration first begins, we can have locked and
2108 * unlocked page stat updates happening concurrently. Track
2109 * the task who has the lock for unlock_page_memcg().
2111 memcg->move_lock_task = current;
2112 memcg->move_lock_flags = flags;
2116 EXPORT_SYMBOL(lock_page_memcg);
2119 * __unlock_page_memcg - unlock and unpin a memcg
2122 * Unlock and unpin a memcg returned by lock_page_memcg().
2124 void __unlock_page_memcg(struct mem_cgroup *memcg)
2126 if (memcg && memcg->move_lock_task == current) {
2127 unsigned long flags = memcg->move_lock_flags;
2129 memcg->move_lock_task = NULL;
2130 memcg->move_lock_flags = 0;
2132 spin_unlock_irqrestore(&memcg->move_lock, flags);
2139 * unlock_page_memcg - unlock a page->mem_cgroup binding
2142 void unlock_page_memcg(struct page *page)
2144 struct page *head = compound_head(page);
2146 __unlock_page_memcg(head->mem_cgroup);
2148 EXPORT_SYMBOL(unlock_page_memcg);
2150 struct memcg_stock_pcp {
2151 struct mem_cgroup *cached; /* this never be root cgroup */
2152 unsigned int nr_pages;
2154 #ifdef CONFIG_MEMCG_KMEM
2155 struct obj_cgroup *cached_objcg;
2156 unsigned int nr_bytes;
2159 struct work_struct work;
2160 unsigned long flags;
2161 #define FLUSHING_CACHED_CHARGE 0
2163 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2164 static DEFINE_MUTEX(percpu_charge_mutex);
2166 #ifdef CONFIG_MEMCG_KMEM
2167 static void drain_obj_stock(struct memcg_stock_pcp *stock);
2168 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2169 struct mem_cgroup *root_memcg);
2172 static inline void drain_obj_stock(struct memcg_stock_pcp *stock)
2175 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2176 struct mem_cgroup *root_memcg)
2183 * consume_stock: Try to consume stocked charge on this cpu.
2184 * @memcg: memcg to consume from.
2185 * @nr_pages: how many pages to charge.
2187 * The charges will only happen if @memcg matches the current cpu's memcg
2188 * stock, and at least @nr_pages are available in that stock. Failure to
2189 * service an allocation will refill the stock.
2191 * returns true if successful, false otherwise.
2193 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2195 struct memcg_stock_pcp *stock;
2196 unsigned long flags;
2199 if (nr_pages > MEMCG_CHARGE_BATCH)
2202 local_irq_save(flags);
2204 stock = this_cpu_ptr(&memcg_stock);
2205 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2206 stock->nr_pages -= nr_pages;
2210 local_irq_restore(flags);
2216 * Returns stocks cached in percpu and reset cached information.
2218 static void drain_stock(struct memcg_stock_pcp *stock)
2220 struct mem_cgroup *old = stock->cached;
2225 if (stock->nr_pages) {
2226 page_counter_uncharge(&old->memory, stock->nr_pages);
2227 if (do_memsw_account())
2228 page_counter_uncharge(&old->memsw, stock->nr_pages);
2229 stock->nr_pages = 0;
2233 stock->cached = NULL;
2236 static void drain_local_stock(struct work_struct *dummy)
2238 struct memcg_stock_pcp *stock;
2239 unsigned long flags;
2242 * The only protection from memory hotplug vs. drain_stock races is
2243 * that we always operate on local CPU stock here with IRQ disabled
2245 local_irq_save(flags);
2247 stock = this_cpu_ptr(&memcg_stock);
2248 drain_obj_stock(stock);
2250 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2252 local_irq_restore(flags);
2256 * Cache charges(val) to local per_cpu area.
2257 * This will be consumed by consume_stock() function, later.
2259 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2261 struct memcg_stock_pcp *stock;
2262 unsigned long flags;
2264 local_irq_save(flags);
2266 stock = this_cpu_ptr(&memcg_stock);
2267 if (stock->cached != memcg) { /* reset if necessary */
2269 css_get(&memcg->css);
2270 stock->cached = memcg;
2272 stock->nr_pages += nr_pages;
2274 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2277 local_irq_restore(flags);
2281 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2282 * of the hierarchy under it.
2284 static void drain_all_stock(struct mem_cgroup *root_memcg)
2288 /* If someone's already draining, avoid adding running more workers. */
2289 if (!mutex_trylock(&percpu_charge_mutex))
2292 * Notify other cpus that system-wide "drain" is running
2293 * We do not care about races with the cpu hotplug because cpu down
2294 * as well as workers from this path always operate on the local
2295 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2298 for_each_online_cpu(cpu) {
2299 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2300 struct mem_cgroup *memcg;
2304 memcg = stock->cached;
2305 if (memcg && stock->nr_pages &&
2306 mem_cgroup_is_descendant(memcg, root_memcg))
2308 if (obj_stock_flush_required(stock, root_memcg))
2313 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2315 drain_local_stock(&stock->work);
2317 schedule_work_on(cpu, &stock->work);
2321 mutex_unlock(&percpu_charge_mutex);
2324 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2326 struct memcg_stock_pcp *stock;
2327 struct mem_cgroup *memcg, *mi;
2329 stock = &per_cpu(memcg_stock, cpu);
2332 for_each_mem_cgroup(memcg) {
2335 for (i = 0; i < MEMCG_NR_STAT; i++) {
2339 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2341 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2342 atomic_long_add(x, &memcg->vmstats[i]);
2344 if (i >= NR_VM_NODE_STAT_ITEMS)
2347 for_each_node(nid) {
2348 struct mem_cgroup_per_node *pn;
2350 pn = mem_cgroup_nodeinfo(memcg, nid);
2351 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2354 atomic_long_add(x, &pn->lruvec_stat[i]);
2355 } while ((pn = parent_nodeinfo(pn, nid)));
2359 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2362 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2364 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2365 atomic_long_add(x, &memcg->vmevents[i]);
2372 static void reclaim_high(struct mem_cgroup *memcg,
2373 unsigned int nr_pages,
2377 if (page_counter_read(&memcg->memory) <=
2378 READ_ONCE(memcg->memory.high))
2380 memcg_memory_event(memcg, MEMCG_HIGH);
2381 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2382 } while ((memcg = parent_mem_cgroup(memcg)) &&
2383 !mem_cgroup_is_root(memcg));
2386 static void high_work_func(struct work_struct *work)
2388 struct mem_cgroup *memcg;
2390 memcg = container_of(work, struct mem_cgroup, high_work);
2391 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2395 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2396 * enough to still cause a significant slowdown in most cases, while still
2397 * allowing diagnostics and tracing to proceed without becoming stuck.
2399 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2402 * When calculating the delay, we use these either side of the exponentiation to
2403 * maintain precision and scale to a reasonable number of jiffies (see the table
2406 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2407 * overage ratio to a delay.
2408 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down down the
2409 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2410 * to produce a reasonable delay curve.
2412 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2413 * reasonable delay curve compared to precision-adjusted overage, not
2414 * penalising heavily at first, but still making sure that growth beyond the
2415 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2416 * example, with a high of 100 megabytes:
2418 * +-------+------------------------+
2419 * | usage | time to allocate in ms |
2420 * +-------+------------------------+
2442 * +-------+------------------------+
2444 #define MEMCG_DELAY_PRECISION_SHIFT 20
2445 #define MEMCG_DELAY_SCALING_SHIFT 14
2447 static u64 calculate_overage(unsigned long usage, unsigned long high)
2455 * Prevent division by 0 in overage calculation by acting as if
2456 * it was a threshold of 1 page
2458 high = max(high, 1UL);
2460 overage = usage - high;
2461 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2462 return div64_u64(overage, high);
2465 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2467 u64 overage, max_overage = 0;
2470 overage = calculate_overage(page_counter_read(&memcg->memory),
2471 READ_ONCE(memcg->memory.high));
2472 max_overage = max(overage, max_overage);
2473 } while ((memcg = parent_mem_cgroup(memcg)) &&
2474 !mem_cgroup_is_root(memcg));
2479 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2481 u64 overage, max_overage = 0;
2484 overage = calculate_overage(page_counter_read(&memcg->swap),
2485 READ_ONCE(memcg->swap.high));
2487 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2488 max_overage = max(overage, max_overage);
2489 } while ((memcg = parent_mem_cgroup(memcg)) &&
2490 !mem_cgroup_is_root(memcg));
2496 * Get the number of jiffies that we should penalise a mischievous cgroup which
2497 * is exceeding its memory.high by checking both it and its ancestors.
2499 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2500 unsigned int nr_pages,
2503 unsigned long penalty_jiffies;
2509 * We use overage compared to memory.high to calculate the number of
2510 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2511 * fairly lenient on small overages, and increasingly harsh when the
2512 * memcg in question makes it clear that it has no intention of stopping
2513 * its crazy behaviour, so we exponentially increase the delay based on
2516 penalty_jiffies = max_overage * max_overage * HZ;
2517 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2518 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2521 * Factor in the task's own contribution to the overage, such that four
2522 * N-sized allocations are throttled approximately the same as one
2523 * 4N-sized allocation.
2525 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2526 * larger the current charge patch is than that.
2528 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2532 * Scheduled by try_charge() to be executed from the userland return path
2533 * and reclaims memory over the high limit.
2535 void mem_cgroup_handle_over_high(void)
2537 unsigned long penalty_jiffies;
2538 unsigned long pflags;
2539 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2540 struct mem_cgroup *memcg;
2542 if (likely(!nr_pages))
2545 memcg = get_mem_cgroup_from_mm(current->mm);
2546 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2547 current->memcg_nr_pages_over_high = 0;
2550 * memory.high is breached and reclaim is unable to keep up. Throttle
2551 * allocators proactively to slow down excessive growth.
2553 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2554 mem_find_max_overage(memcg));
2556 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2557 swap_find_max_overage(memcg));
2560 * Clamp the max delay per usermode return so as to still keep the
2561 * application moving forwards and also permit diagnostics, albeit
2564 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2567 * Don't sleep if the amount of jiffies this memcg owes us is so low
2568 * that it's not even worth doing, in an attempt to be nice to those who
2569 * go only a small amount over their memory.high value and maybe haven't
2570 * been aggressively reclaimed enough yet.
2572 if (penalty_jiffies <= HZ / 100)
2576 * If we exit early, we're guaranteed to die (since
2577 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2578 * need to account for any ill-begotten jiffies to pay them off later.
2580 psi_memstall_enter(&pflags);
2581 schedule_timeout_killable(penalty_jiffies);
2582 psi_memstall_leave(&pflags);
2585 css_put(&memcg->css);
2588 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2589 unsigned int nr_pages)
2591 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2592 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2593 struct mem_cgroup *mem_over_limit;
2594 struct page_counter *counter;
2595 unsigned long nr_reclaimed;
2596 bool may_swap = true;
2597 bool drained = false;
2598 enum oom_status oom_status;
2600 if (mem_cgroup_is_root(memcg))
2603 if (consume_stock(memcg, nr_pages))
2606 if (!do_memsw_account() ||
2607 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2608 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2610 if (do_memsw_account())
2611 page_counter_uncharge(&memcg->memsw, batch);
2612 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2614 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2618 if (batch > nr_pages) {
2624 * Memcg doesn't have a dedicated reserve for atomic
2625 * allocations. But like the global atomic pool, we need to
2626 * put the burden of reclaim on regular allocation requests
2627 * and let these go through as privileged allocations.
2629 if (gfp_mask & __GFP_ATOMIC)
2633 * Unlike in global OOM situations, memcg is not in a physical
2634 * memory shortage. Allow dying and OOM-killed tasks to
2635 * bypass the last charges so that they can exit quickly and
2636 * free their memory.
2638 if (unlikely(should_force_charge()))
2642 * Prevent unbounded recursion when reclaim operations need to
2643 * allocate memory. This might exceed the limits temporarily,
2644 * but we prefer facilitating memory reclaim and getting back
2645 * under the limit over triggering OOM kills in these cases.
2647 if (unlikely(current->flags & PF_MEMALLOC))
2650 if (unlikely(task_in_memcg_oom(current)))
2653 if (!gfpflags_allow_blocking(gfp_mask))
2656 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2658 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2659 gfp_mask, may_swap);
2661 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2665 drain_all_stock(mem_over_limit);
2670 if (gfp_mask & __GFP_NORETRY)
2673 * Even though the limit is exceeded at this point, reclaim
2674 * may have been able to free some pages. Retry the charge
2675 * before killing the task.
2677 * Only for regular pages, though: huge pages are rather
2678 * unlikely to succeed so close to the limit, and we fall back
2679 * to regular pages anyway in case of failure.
2681 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2684 * At task move, charge accounts can be doubly counted. So, it's
2685 * better to wait until the end of task_move if something is going on.
2687 if (mem_cgroup_wait_acct_move(mem_over_limit))
2693 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2696 if (gfp_mask & __GFP_NOFAIL)
2699 if (fatal_signal_pending(current))
2703 * keep retrying as long as the memcg oom killer is able to make
2704 * a forward progress or bypass the charge if the oom killer
2705 * couldn't make any progress.
2707 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2708 get_order(nr_pages * PAGE_SIZE));
2709 switch (oom_status) {
2711 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2719 if (!(gfp_mask & __GFP_NOFAIL))
2723 * The allocation either can't fail or will lead to more memory
2724 * being freed very soon. Allow memory usage go over the limit
2725 * temporarily by force charging it.
2727 page_counter_charge(&memcg->memory, nr_pages);
2728 if (do_memsw_account())
2729 page_counter_charge(&memcg->memsw, nr_pages);
2734 if (batch > nr_pages)
2735 refill_stock(memcg, batch - nr_pages);
2738 * If the hierarchy is above the normal consumption range, schedule
2739 * reclaim on returning to userland. We can perform reclaim here
2740 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2741 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2742 * not recorded as it most likely matches current's and won't
2743 * change in the meantime. As high limit is checked again before
2744 * reclaim, the cost of mismatch is negligible.
2747 bool mem_high, swap_high;
2749 mem_high = page_counter_read(&memcg->memory) >
2750 READ_ONCE(memcg->memory.high);
2751 swap_high = page_counter_read(&memcg->swap) >
2752 READ_ONCE(memcg->swap.high);
2754 /* Don't bother a random interrupted task */
2755 if (in_interrupt()) {
2757 schedule_work(&memcg->high_work);
2763 if (mem_high || swap_high) {
2765 * The allocating tasks in this cgroup will need to do
2766 * reclaim or be throttled to prevent further growth
2767 * of the memory or swap footprints.
2769 * Target some best-effort fairness between the tasks,
2770 * and distribute reclaim work and delay penalties
2771 * based on how much each task is actually allocating.
2773 current->memcg_nr_pages_over_high += batch;
2774 set_notify_resume(current);
2777 } while ((memcg = parent_mem_cgroup(memcg)));
2782 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2783 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2785 if (mem_cgroup_is_root(memcg))
2788 page_counter_uncharge(&memcg->memory, nr_pages);
2789 if (do_memsw_account())
2790 page_counter_uncharge(&memcg->memsw, nr_pages);
2794 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2796 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2798 * Any of the following ensures page->mem_cgroup stability:
2802 * - lock_page_memcg()
2803 * - exclusive reference
2805 page->mem_cgroup = memcg;
2808 #ifdef CONFIG_MEMCG_KMEM
2810 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2812 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2813 * cgroup_mutex, etc.
2815 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2819 if (mem_cgroup_disabled())
2822 page = virt_to_head_page(p);
2825 * Slab pages don't have page->mem_cgroup set because corresponding
2826 * kmem caches can be reparented during the lifetime. That's why
2827 * memcg_from_slab_page() should be used instead.
2830 return memcg_from_slab_page(page);
2832 /* All other pages use page->mem_cgroup */
2833 return page->mem_cgroup;
2836 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2838 struct obj_cgroup *objcg = NULL;
2839 struct mem_cgroup *memcg;
2841 if (unlikely(!current->mm && !current->active_memcg))
2845 if (unlikely(current->active_memcg))
2846 memcg = rcu_dereference(current->active_memcg);
2848 memcg = mem_cgroup_from_task(current);
2850 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2851 objcg = rcu_dereference(memcg->objcg);
2852 if (objcg && obj_cgroup_tryget(objcg))
2860 static int memcg_alloc_cache_id(void)
2865 id = ida_simple_get(&memcg_cache_ida,
2866 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2870 if (id < memcg_nr_cache_ids)
2874 * There's no space for the new id in memcg_caches arrays,
2875 * so we have to grow them.
2877 down_write(&memcg_cache_ids_sem);
2879 size = 2 * (id + 1);
2880 if (size < MEMCG_CACHES_MIN_SIZE)
2881 size = MEMCG_CACHES_MIN_SIZE;
2882 else if (size > MEMCG_CACHES_MAX_SIZE)
2883 size = MEMCG_CACHES_MAX_SIZE;
2885 err = memcg_update_all_caches(size);
2887 err = memcg_update_all_list_lrus(size);
2889 memcg_nr_cache_ids = size;
2891 up_write(&memcg_cache_ids_sem);
2894 ida_simple_remove(&memcg_cache_ida, id);
2900 static void memcg_free_cache_id(int id)
2902 ida_simple_remove(&memcg_cache_ida, id);
2905 struct memcg_kmem_cache_create_work {
2906 struct mem_cgroup *memcg;
2907 struct kmem_cache *cachep;
2908 struct work_struct work;
2911 static void memcg_kmem_cache_create_func(struct work_struct *w)
2913 struct memcg_kmem_cache_create_work *cw =
2914 container_of(w, struct memcg_kmem_cache_create_work, work);
2915 struct mem_cgroup *memcg = cw->memcg;
2916 struct kmem_cache *cachep = cw->cachep;
2918 memcg_create_kmem_cache(memcg, cachep);
2920 css_put(&memcg->css);
2925 * Enqueue the creation of a per-memcg kmem_cache.
2927 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2928 struct kmem_cache *cachep)
2930 struct memcg_kmem_cache_create_work *cw;
2932 if (!css_tryget_online(&memcg->css))
2935 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2937 css_put(&memcg->css);
2942 cw->cachep = cachep;
2943 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2945 queue_work(memcg_kmem_cache_wq, &cw->work);
2948 static inline bool memcg_kmem_bypass(void)
2953 /* Allow remote memcg charging in kthread contexts. */
2954 if ((!current->mm || (current->flags & PF_KTHREAD)) &&
2955 !current->active_memcg)
2961 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2962 * @cachep: the original global kmem cache
2964 * Return the kmem_cache we're supposed to use for a slab allocation.
2965 * We try to use the current memcg's version of the cache.
2967 * If the cache does not exist yet, if we are the first user of it, we
2968 * create it asynchronously in a workqueue and let the current allocation
2969 * go through with the original cache.
2971 * This function takes a reference to the cache it returns to assure it
2972 * won't get destroyed while we are working with it. Once the caller is
2973 * done with it, memcg_kmem_put_cache() must be called to release the
2976 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2978 struct mem_cgroup *memcg;
2979 struct kmem_cache *memcg_cachep;
2980 struct memcg_cache_array *arr;
2983 VM_BUG_ON(!is_root_cache(cachep));
2985 if (memcg_kmem_bypass())
2990 if (unlikely(current->active_memcg))
2991 memcg = current->active_memcg;
2993 memcg = mem_cgroup_from_task(current);
2995 if (!memcg || memcg == root_mem_cgroup)
2998 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
3002 arr = rcu_dereference(cachep->memcg_params.memcg_caches);
3005 * Make sure we will access the up-to-date value. The code updating
3006 * memcg_caches issues a write barrier to match the data dependency
3007 * barrier inside READ_ONCE() (see memcg_create_kmem_cache()).
3009 memcg_cachep = READ_ONCE(arr->entries[kmemcg_id]);
3012 * If we are in a safe context (can wait, and not in interrupt
3013 * context), we could be be predictable and return right away.
3014 * This would guarantee that the allocation being performed
3015 * already belongs in the new cache.
3017 * However, there are some clashes that can arrive from locking.
3018 * For instance, because we acquire the slab_mutex while doing
3019 * memcg_create_kmem_cache, this means no further allocation
3020 * could happen with the slab_mutex held. So it's better to
3023 * If the memcg is dying or memcg_cache is about to be released,
3024 * don't bother creating new kmem_caches. Because memcg_cachep
3025 * is ZEROed as the fist step of kmem offlining, we don't need
3026 * percpu_ref_tryget_live() here. css_tryget_online() check in
3027 * memcg_schedule_kmem_cache_create() will prevent us from
3028 * creation of a new kmem_cache.
3030 if (unlikely(!memcg_cachep))
3031 memcg_schedule_kmem_cache_create(memcg, cachep);
3032 else if (percpu_ref_tryget(&memcg_cachep->memcg_params.refcnt))
3033 cachep = memcg_cachep;
3040 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
3041 * @cachep: the cache returned by memcg_kmem_get_cache
3043 void memcg_kmem_put_cache(struct kmem_cache *cachep)
3045 if (!is_root_cache(cachep))
3046 percpu_ref_put(&cachep->memcg_params.refcnt);
3050 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
3051 * @memcg: memory cgroup to charge
3052 * @gfp: reclaim mode
3053 * @nr_pages: number of pages to charge
3055 * Returns 0 on success, an error code on failure.
3057 int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
3058 unsigned int nr_pages)
3060 struct page_counter *counter;
3063 ret = try_charge(memcg, gfp, nr_pages);
3067 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3068 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3071 * Enforce __GFP_NOFAIL allocation because callers are not
3072 * prepared to see failures and likely do not have any failure
3075 if (gfp & __GFP_NOFAIL) {
3076 page_counter_charge(&memcg->kmem, nr_pages);
3079 cancel_charge(memcg, nr_pages);
3086 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
3087 * @memcg: memcg to uncharge
3088 * @nr_pages: number of pages to uncharge
3090 void __memcg_kmem_uncharge(struct mem_cgroup *memcg, unsigned int nr_pages)
3092 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3093 page_counter_uncharge(&memcg->kmem, nr_pages);
3095 page_counter_uncharge(&memcg->memory, nr_pages);
3096 if (do_memsw_account())
3097 page_counter_uncharge(&memcg->memsw, nr_pages);
3101 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3102 * @page: page to charge
3103 * @gfp: reclaim mode
3104 * @order: allocation order
3106 * Returns 0 on success, an error code on failure.
3108 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3110 struct mem_cgroup *memcg;
3113 if (memcg_kmem_bypass())
3116 memcg = get_mem_cgroup_from_current();
3117 if (!mem_cgroup_is_root(memcg)) {
3118 ret = __memcg_kmem_charge(memcg, gfp, 1 << order);
3120 page->mem_cgroup = memcg;
3121 __SetPageKmemcg(page);
3125 css_put(&memcg->css);
3130 * __memcg_kmem_uncharge_page: uncharge a kmem page
3131 * @page: page to uncharge
3132 * @order: allocation order
3134 void __memcg_kmem_uncharge_page(struct page *page, int order)
3136 struct mem_cgroup *memcg = page->mem_cgroup;
3137 unsigned int nr_pages = 1 << order;
3142 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3143 __memcg_kmem_uncharge(memcg, nr_pages);
3144 page->mem_cgroup = NULL;
3145 css_put(&memcg->css);
3147 /* slab pages do not have PageKmemcg flag set */
3148 if (PageKmemcg(page))
3149 __ClearPageKmemcg(page);
3152 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3154 struct memcg_stock_pcp *stock;
3155 unsigned long flags;
3158 local_irq_save(flags);
3160 stock = this_cpu_ptr(&memcg_stock);
3161 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3162 stock->nr_bytes -= nr_bytes;
3166 local_irq_restore(flags);
3171 static void drain_obj_stock(struct memcg_stock_pcp *stock)
3173 struct obj_cgroup *old = stock->cached_objcg;
3178 if (stock->nr_bytes) {
3179 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3180 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3184 __memcg_kmem_uncharge(obj_cgroup_memcg(old), nr_pages);
3189 * The leftover is flushed to the centralized per-memcg value.
3190 * On the next attempt to refill obj stock it will be moved
3191 * to a per-cpu stock (probably, on an other CPU), see
3192 * refill_obj_stock().
3194 * How often it's flushed is a trade-off between the memory
3195 * limit enforcement accuracy and potential CPU contention,
3196 * so it might be changed in the future.
3198 atomic_add(nr_bytes, &old->nr_charged_bytes);
3199 stock->nr_bytes = 0;
3202 obj_cgroup_put(old);
3203 stock->cached_objcg = NULL;
3206 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3207 struct mem_cgroup *root_memcg)
3209 struct mem_cgroup *memcg;
3211 if (stock->cached_objcg) {
3212 memcg = obj_cgroup_memcg(stock->cached_objcg);
3213 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3220 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3222 struct memcg_stock_pcp *stock;
3223 unsigned long flags;
3225 local_irq_save(flags);
3227 stock = this_cpu_ptr(&memcg_stock);
3228 if (stock->cached_objcg != objcg) { /* reset if necessary */
3229 drain_obj_stock(stock);
3230 obj_cgroup_get(objcg);
3231 stock->cached_objcg = objcg;
3232 stock->nr_bytes = atomic_xchg(&objcg->nr_charged_bytes, 0);
3234 stock->nr_bytes += nr_bytes;
3236 if (stock->nr_bytes > PAGE_SIZE)
3237 drain_obj_stock(stock);
3239 local_irq_restore(flags);
3242 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3244 struct mem_cgroup *memcg;
3245 unsigned int nr_pages, nr_bytes;
3248 if (consume_obj_stock(objcg, size))
3252 * In theory, memcg->nr_charged_bytes can have enough
3253 * pre-charged bytes to satisfy the allocation. However,
3254 * flushing memcg->nr_charged_bytes requires two atomic
3255 * operations, and memcg->nr_charged_bytes can't be big,
3256 * so it's better to ignore it and try grab some new pages.
3257 * memcg->nr_charged_bytes will be flushed in
3258 * refill_obj_stock(), called from this function or
3259 * independently later.
3262 memcg = obj_cgroup_memcg(objcg);
3263 css_get(&memcg->css);
3266 nr_pages = size >> PAGE_SHIFT;
3267 nr_bytes = size & (PAGE_SIZE - 1);
3272 ret = __memcg_kmem_charge(memcg, gfp, nr_pages);
3273 if (!ret && nr_bytes)
3274 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes);
3276 css_put(&memcg->css);
3280 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3282 refill_obj_stock(objcg, size);
3285 #endif /* CONFIG_MEMCG_KMEM */
3287 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3290 * Because tail pages are not marked as "used", set it. We're under
3291 * pgdat->lru_lock and migration entries setup in all page mappings.
3293 void mem_cgroup_split_huge_fixup(struct page *head)
3295 struct mem_cgroup *memcg = head->mem_cgroup;
3298 if (mem_cgroup_disabled())
3301 for (i = 1; i < HPAGE_PMD_NR; i++) {
3302 css_get(&memcg->css);
3303 head[i].mem_cgroup = memcg;
3306 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3308 #ifdef CONFIG_MEMCG_SWAP
3310 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3311 * @entry: swap entry to be moved
3312 * @from: mem_cgroup which the entry is moved from
3313 * @to: mem_cgroup which the entry is moved to
3315 * It succeeds only when the swap_cgroup's record for this entry is the same
3316 * as the mem_cgroup's id of @from.
3318 * Returns 0 on success, -EINVAL on failure.
3320 * The caller must have charged to @to, IOW, called page_counter_charge() about
3321 * both res and memsw, and called css_get().
3323 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3324 struct mem_cgroup *from, struct mem_cgroup *to)
3326 unsigned short old_id, new_id;
3328 old_id = mem_cgroup_id(from);
3329 new_id = mem_cgroup_id(to);
3331 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3332 mod_memcg_state(from, MEMCG_SWAP, -1);
3333 mod_memcg_state(to, MEMCG_SWAP, 1);
3339 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3340 struct mem_cgroup *from, struct mem_cgroup *to)
3346 static DEFINE_MUTEX(memcg_max_mutex);
3348 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3349 unsigned long max, bool memsw)
3351 bool enlarge = false;
3352 bool drained = false;
3354 bool limits_invariant;
3355 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3358 if (signal_pending(current)) {
3363 mutex_lock(&memcg_max_mutex);
3365 * Make sure that the new limit (memsw or memory limit) doesn't
3366 * break our basic invariant rule memory.max <= memsw.max.
3368 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3369 max <= memcg->memsw.max;
3370 if (!limits_invariant) {
3371 mutex_unlock(&memcg_max_mutex);
3375 if (max > counter->max)
3377 ret = page_counter_set_max(counter, max);
3378 mutex_unlock(&memcg_max_mutex);
3384 drain_all_stock(memcg);
3389 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3390 GFP_KERNEL, !memsw)) {
3396 if (!ret && enlarge)
3397 memcg_oom_recover(memcg);
3402 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3404 unsigned long *total_scanned)
3406 unsigned long nr_reclaimed = 0;
3407 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3408 unsigned long reclaimed;
3410 struct mem_cgroup_tree_per_node *mctz;
3411 unsigned long excess;
3412 unsigned long nr_scanned;
3417 mctz = soft_limit_tree_node(pgdat->node_id);
3420 * Do not even bother to check the largest node if the root
3421 * is empty. Do it lockless to prevent lock bouncing. Races
3422 * are acceptable as soft limit is best effort anyway.
3424 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3428 * This loop can run a while, specially if mem_cgroup's continuously
3429 * keep exceeding their soft limit and putting the system under
3436 mz = mem_cgroup_largest_soft_limit_node(mctz);
3441 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3442 gfp_mask, &nr_scanned);
3443 nr_reclaimed += reclaimed;
3444 *total_scanned += nr_scanned;
3445 spin_lock_irq(&mctz->lock);
3446 __mem_cgroup_remove_exceeded(mz, mctz);
3449 * If we failed to reclaim anything from this memory cgroup
3450 * it is time to move on to the next cgroup
3454 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3456 excess = soft_limit_excess(mz->memcg);
3458 * One school of thought says that we should not add
3459 * back the node to the tree if reclaim returns 0.
3460 * But our reclaim could return 0, simply because due
3461 * to priority we are exposing a smaller subset of
3462 * memory to reclaim from. Consider this as a longer
3465 /* If excess == 0, no tree ops */
3466 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3467 spin_unlock_irq(&mctz->lock);
3468 css_put(&mz->memcg->css);
3471 * Could not reclaim anything and there are no more
3472 * mem cgroups to try or we seem to be looping without
3473 * reclaiming anything.
3475 if (!nr_reclaimed &&
3477 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3479 } while (!nr_reclaimed);
3481 css_put(&next_mz->memcg->css);
3482 return nr_reclaimed;
3486 * Test whether @memcg has children, dead or alive. Note that this
3487 * function doesn't care whether @memcg has use_hierarchy enabled and
3488 * returns %true if there are child csses according to the cgroup
3489 * hierarchy. Testing use_hierarchy is the caller's responsibility.
3491 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3496 ret = css_next_child(NULL, &memcg->css);
3502 * Reclaims as many pages from the given memcg as possible.
3504 * Caller is responsible for holding css reference for memcg.
3506 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3508 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3510 /* we call try-to-free pages for make this cgroup empty */
3511 lru_add_drain_all();
3513 drain_all_stock(memcg);
3515 /* try to free all pages in this cgroup */
3516 while (nr_retries && page_counter_read(&memcg->memory)) {
3519 if (signal_pending(current))
3522 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3526 /* maybe some writeback is necessary */
3527 congestion_wait(BLK_RW_ASYNC, HZ/10);
3535 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3536 char *buf, size_t nbytes,
3539 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3541 if (mem_cgroup_is_root(memcg))
3543 return mem_cgroup_force_empty(memcg) ?: nbytes;
3546 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3549 return mem_cgroup_from_css(css)->use_hierarchy;
3552 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3553 struct cftype *cft, u64 val)
3556 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3557 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3559 if (memcg->use_hierarchy == val)
3563 * If parent's use_hierarchy is set, we can't make any modifications
3564 * in the child subtrees. If it is unset, then the change can
3565 * occur, provided the current cgroup has no children.
3567 * For the root cgroup, parent_mem is NULL, we allow value to be
3568 * set if there are no children.
3570 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3571 (val == 1 || val == 0)) {
3572 if (!memcg_has_children(memcg))
3573 memcg->use_hierarchy = val;
3582 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3586 if (mem_cgroup_is_root(memcg)) {
3587 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3588 memcg_page_state(memcg, NR_ANON_MAPPED);
3590 val += memcg_page_state(memcg, MEMCG_SWAP);
3593 val = page_counter_read(&memcg->memory);
3595 val = page_counter_read(&memcg->memsw);
3608 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3611 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3612 struct page_counter *counter;
3614 switch (MEMFILE_TYPE(cft->private)) {
3616 counter = &memcg->memory;
3619 counter = &memcg->memsw;
3622 counter = &memcg->kmem;
3625 counter = &memcg->tcpmem;
3631 switch (MEMFILE_ATTR(cft->private)) {
3633 if (counter == &memcg->memory)
3634 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3635 if (counter == &memcg->memsw)
3636 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3637 return (u64)page_counter_read(counter) * PAGE_SIZE;
3639 return (u64)counter->max * PAGE_SIZE;
3641 return (u64)counter->watermark * PAGE_SIZE;
3643 return counter->failcnt;
3644 case RES_SOFT_LIMIT:
3645 return (u64)memcg->soft_limit * PAGE_SIZE;
3651 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3653 unsigned long stat[MEMCG_NR_STAT] = {0};
3654 struct mem_cgroup *mi;
3657 for_each_online_cpu(cpu)
3658 for (i = 0; i < MEMCG_NR_STAT; i++)
3659 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3661 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3662 for (i = 0; i < MEMCG_NR_STAT; i++)
3663 atomic_long_add(stat[i], &mi->vmstats[i]);
3665 for_each_node(node) {
3666 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3667 struct mem_cgroup_per_node *pi;
3669 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3672 for_each_online_cpu(cpu)
3673 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3675 pn->lruvec_stat_cpu->count[i], cpu);
3677 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3678 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3679 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3683 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3685 unsigned long events[NR_VM_EVENT_ITEMS];
3686 struct mem_cgroup *mi;
3689 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3692 for_each_online_cpu(cpu)
3693 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3694 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3697 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3698 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3699 atomic_long_add(events[i], &mi->vmevents[i]);
3702 #ifdef CONFIG_MEMCG_KMEM
3703 static int memcg_online_kmem(struct mem_cgroup *memcg)
3705 struct obj_cgroup *objcg;
3708 if (cgroup_memory_nokmem)
3711 BUG_ON(memcg->kmemcg_id >= 0);
3712 BUG_ON(memcg->kmem_state);
3714 memcg_id = memcg_alloc_cache_id();
3718 objcg = obj_cgroup_alloc();
3720 memcg_free_cache_id(memcg_id);
3723 objcg->memcg = memcg;
3724 rcu_assign_pointer(memcg->objcg, objcg);
3726 static_branch_enable(&memcg_kmem_enabled_key);
3729 * A memory cgroup is considered kmem-online as soon as it gets
3730 * kmemcg_id. Setting the id after enabling static branching will
3731 * guarantee no one starts accounting before all call sites are
3734 memcg->kmemcg_id = memcg_id;
3735 memcg->kmem_state = KMEM_ONLINE;
3736 INIT_LIST_HEAD(&memcg->kmem_caches);
3741 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3743 struct cgroup_subsys_state *css;
3744 struct mem_cgroup *parent, *child;
3747 if (memcg->kmem_state != KMEM_ONLINE)
3750 * Clear the online state before clearing memcg_caches array
3751 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3752 * guarantees that no cache will be created for this cgroup
3753 * after we are done (see memcg_create_kmem_cache()).
3755 memcg->kmem_state = KMEM_ALLOCATED;
3757 parent = parent_mem_cgroup(memcg);
3759 parent = root_mem_cgroup;
3762 * Deactivate and reparent kmem_caches and objcgs.
3764 memcg_deactivate_kmem_caches(memcg, parent);
3765 memcg_reparent_objcgs(memcg, parent);
3767 kmemcg_id = memcg->kmemcg_id;
3768 BUG_ON(kmemcg_id < 0);
3771 * Change kmemcg_id of this cgroup and all its descendants to the
3772 * parent's id, and then move all entries from this cgroup's list_lrus
3773 * to ones of the parent. After we have finished, all list_lrus
3774 * corresponding to this cgroup are guaranteed to remain empty. The
3775 * ordering is imposed by list_lru_node->lock taken by
3776 * memcg_drain_all_list_lrus().
3778 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3779 css_for_each_descendant_pre(css, &memcg->css) {
3780 child = mem_cgroup_from_css(css);
3781 BUG_ON(child->kmemcg_id != kmemcg_id);
3782 child->kmemcg_id = parent->kmemcg_id;
3783 if (!memcg->use_hierarchy)
3788 memcg_drain_all_list_lrus(kmemcg_id, parent);
3790 memcg_free_cache_id(kmemcg_id);
3793 static void memcg_free_kmem(struct mem_cgroup *memcg)
3795 /* css_alloc() failed, offlining didn't happen */
3796 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3797 memcg_offline_kmem(memcg);
3800 static int memcg_online_kmem(struct mem_cgroup *memcg)
3804 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3807 static void memcg_free_kmem(struct mem_cgroup *memcg)
3810 #endif /* CONFIG_MEMCG_KMEM */
3812 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3817 mutex_lock(&memcg_max_mutex);
3818 ret = page_counter_set_max(&memcg->kmem, max);
3819 mutex_unlock(&memcg_max_mutex);
3823 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3827 mutex_lock(&memcg_max_mutex);
3829 ret = page_counter_set_max(&memcg->tcpmem, max);
3833 if (!memcg->tcpmem_active) {
3835 * The active flag needs to be written after the static_key
3836 * update. This is what guarantees that the socket activation
3837 * function is the last one to run. See mem_cgroup_sk_alloc()
3838 * for details, and note that we don't mark any socket as
3839 * belonging to this memcg until that flag is up.
3841 * We need to do this, because static_keys will span multiple
3842 * sites, but we can't control their order. If we mark a socket
3843 * as accounted, but the accounting functions are not patched in
3844 * yet, we'll lose accounting.
3846 * We never race with the readers in mem_cgroup_sk_alloc(),
3847 * because when this value change, the code to process it is not
3850 static_branch_inc(&memcg_sockets_enabled_key);
3851 memcg->tcpmem_active = true;
3854 mutex_unlock(&memcg_max_mutex);
3859 * The user of this function is...
3862 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3863 char *buf, size_t nbytes, loff_t off)
3865 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3866 unsigned long nr_pages;
3869 buf = strstrip(buf);
3870 ret = page_counter_memparse(buf, "-1", &nr_pages);
3874 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3876 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3880 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3882 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3885 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3888 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3889 "Please report your usecase to linux-mm@kvack.org if you "
3890 "depend on this functionality.\n");
3891 ret = memcg_update_kmem_max(memcg, nr_pages);
3894 ret = memcg_update_tcp_max(memcg, nr_pages);
3898 case RES_SOFT_LIMIT:
3899 memcg->soft_limit = nr_pages;
3903 return ret ?: nbytes;
3906 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3907 size_t nbytes, loff_t off)
3909 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3910 struct page_counter *counter;
3912 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3914 counter = &memcg->memory;
3917 counter = &memcg->memsw;
3920 counter = &memcg->kmem;
3923 counter = &memcg->tcpmem;
3929 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3931 page_counter_reset_watermark(counter);
3934 counter->failcnt = 0;
3943 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3946 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3950 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3951 struct cftype *cft, u64 val)
3953 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3955 if (val & ~MOVE_MASK)
3959 * No kind of locking is needed in here, because ->can_attach() will
3960 * check this value once in the beginning of the process, and then carry
3961 * on with stale data. This means that changes to this value will only
3962 * affect task migrations starting after the change.
3964 memcg->move_charge_at_immigrate = val;
3968 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3969 struct cftype *cft, u64 val)
3977 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3978 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3979 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3981 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3982 int nid, unsigned int lru_mask, bool tree)
3984 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3985 unsigned long nr = 0;
3988 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3991 if (!(BIT(lru) & lru_mask))
3994 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3996 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
4001 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
4002 unsigned int lru_mask,
4005 unsigned long nr = 0;
4009 if (!(BIT(lru) & lru_mask))
4012 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
4014 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
4019 static int memcg_numa_stat_show(struct seq_file *m, void *v)
4023 unsigned int lru_mask;
4026 static const struct numa_stat stats[] = {
4027 { "total", LRU_ALL },
4028 { "file", LRU_ALL_FILE },
4029 { "anon", LRU_ALL_ANON },
4030 { "unevictable", BIT(LRU_UNEVICTABLE) },
4032 const struct numa_stat *stat;
4034 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4036 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4037 seq_printf(m, "%s=%lu", stat->name,
4038 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4040 for_each_node_state(nid, N_MEMORY)
4041 seq_printf(m, " N%d=%lu", nid,
4042 mem_cgroup_node_nr_lru_pages(memcg, nid,
4043 stat->lru_mask, false));
4047 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4049 seq_printf(m, "hierarchical_%s=%lu", stat->name,
4050 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4052 for_each_node_state(nid, N_MEMORY)
4053 seq_printf(m, " N%d=%lu", nid,
4054 mem_cgroup_node_nr_lru_pages(memcg, nid,
4055 stat->lru_mask, true));
4061 #endif /* CONFIG_NUMA */
4063 static const unsigned int memcg1_stats[] = {
4066 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4076 static const char *const memcg1_stat_names[] = {
4079 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4089 /* Universal VM events cgroup1 shows, original sort order */
4090 static const unsigned int memcg1_events[] = {
4097 static int memcg_stat_show(struct seq_file *m, void *v)
4099 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4100 unsigned long memory, memsw;
4101 struct mem_cgroup *mi;
4104 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4106 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4109 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4111 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4112 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4113 if (memcg1_stats[i] == NR_ANON_THPS)
4116 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4119 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4120 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4121 memcg_events_local(memcg, memcg1_events[i]));
4123 for (i = 0; i < NR_LRU_LISTS; i++)
4124 seq_printf(m, "%s %lu\n", lru_list_name(i),
4125 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4128 /* Hierarchical information */
4129 memory = memsw = PAGE_COUNTER_MAX;
4130 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4131 memory = min(memory, READ_ONCE(mi->memory.max));
4132 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4134 seq_printf(m, "hierarchical_memory_limit %llu\n",
4135 (u64)memory * PAGE_SIZE);
4136 if (do_memsw_account())
4137 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4138 (u64)memsw * PAGE_SIZE);
4140 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4141 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4143 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4144 (u64)memcg_page_state(memcg, memcg1_stats[i]) *
4148 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4149 seq_printf(m, "total_%s %llu\n",
4150 vm_event_name(memcg1_events[i]),
4151 (u64)memcg_events(memcg, memcg1_events[i]));
4153 for (i = 0; i < NR_LRU_LISTS; i++)
4154 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4155 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4158 #ifdef CONFIG_DEBUG_VM
4161 struct mem_cgroup_per_node *mz;
4162 unsigned long anon_cost = 0;
4163 unsigned long file_cost = 0;
4165 for_each_online_pgdat(pgdat) {
4166 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
4168 anon_cost += mz->lruvec.anon_cost;
4169 file_cost += mz->lruvec.file_cost;
4171 seq_printf(m, "anon_cost %lu\n", anon_cost);
4172 seq_printf(m, "file_cost %lu\n", file_cost);
4179 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4182 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4184 return mem_cgroup_swappiness(memcg);
4187 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4188 struct cftype *cft, u64 val)
4190 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4196 memcg->swappiness = val;
4198 vm_swappiness = val;
4203 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4205 struct mem_cgroup_threshold_ary *t;
4206 unsigned long usage;
4211 t = rcu_dereference(memcg->thresholds.primary);
4213 t = rcu_dereference(memcg->memsw_thresholds.primary);
4218 usage = mem_cgroup_usage(memcg, swap);
4221 * current_threshold points to threshold just below or equal to usage.
4222 * If it's not true, a threshold was crossed after last
4223 * call of __mem_cgroup_threshold().
4225 i = t->current_threshold;
4228 * Iterate backward over array of thresholds starting from
4229 * current_threshold and check if a threshold is crossed.
4230 * If none of thresholds below usage is crossed, we read
4231 * only one element of the array here.
4233 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4234 eventfd_signal(t->entries[i].eventfd, 1);
4236 /* i = current_threshold + 1 */
4240 * Iterate forward over array of thresholds starting from
4241 * current_threshold+1 and check if a threshold is crossed.
4242 * If none of thresholds above usage is crossed, we read
4243 * only one element of the array here.
4245 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4246 eventfd_signal(t->entries[i].eventfd, 1);
4248 /* Update current_threshold */
4249 t->current_threshold = i - 1;
4254 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4257 __mem_cgroup_threshold(memcg, false);
4258 if (do_memsw_account())
4259 __mem_cgroup_threshold(memcg, true);
4261 memcg = parent_mem_cgroup(memcg);
4265 static int compare_thresholds(const void *a, const void *b)
4267 const struct mem_cgroup_threshold *_a = a;
4268 const struct mem_cgroup_threshold *_b = b;
4270 if (_a->threshold > _b->threshold)
4273 if (_a->threshold < _b->threshold)
4279 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4281 struct mem_cgroup_eventfd_list *ev;
4283 spin_lock(&memcg_oom_lock);
4285 list_for_each_entry(ev, &memcg->oom_notify, list)
4286 eventfd_signal(ev->eventfd, 1);
4288 spin_unlock(&memcg_oom_lock);
4292 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4294 struct mem_cgroup *iter;
4296 for_each_mem_cgroup_tree(iter, memcg)
4297 mem_cgroup_oom_notify_cb(iter);
4300 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4301 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4303 struct mem_cgroup_thresholds *thresholds;
4304 struct mem_cgroup_threshold_ary *new;
4305 unsigned long threshold;
4306 unsigned long usage;
4309 ret = page_counter_memparse(args, "-1", &threshold);
4313 mutex_lock(&memcg->thresholds_lock);
4316 thresholds = &memcg->thresholds;
4317 usage = mem_cgroup_usage(memcg, false);
4318 } else if (type == _MEMSWAP) {
4319 thresholds = &memcg->memsw_thresholds;
4320 usage = mem_cgroup_usage(memcg, true);
4324 /* Check if a threshold crossed before adding a new one */
4325 if (thresholds->primary)
4326 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4328 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4330 /* Allocate memory for new array of thresholds */
4331 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4338 /* Copy thresholds (if any) to new array */
4339 if (thresholds->primary) {
4340 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4341 sizeof(struct mem_cgroup_threshold));
4344 /* Add new threshold */
4345 new->entries[size - 1].eventfd = eventfd;
4346 new->entries[size - 1].threshold = threshold;
4348 /* Sort thresholds. Registering of new threshold isn't time-critical */
4349 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4350 compare_thresholds, NULL);
4352 /* Find current threshold */
4353 new->current_threshold = -1;
4354 for (i = 0; i < size; i++) {
4355 if (new->entries[i].threshold <= usage) {
4357 * new->current_threshold will not be used until
4358 * rcu_assign_pointer(), so it's safe to increment
4361 ++new->current_threshold;
4366 /* Free old spare buffer and save old primary buffer as spare */
4367 kfree(thresholds->spare);
4368 thresholds->spare = thresholds->primary;
4370 rcu_assign_pointer(thresholds->primary, new);
4372 /* To be sure that nobody uses thresholds */
4376 mutex_unlock(&memcg->thresholds_lock);
4381 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4382 struct eventfd_ctx *eventfd, const char *args)
4384 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4387 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4388 struct eventfd_ctx *eventfd, const char *args)
4390 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4393 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4394 struct eventfd_ctx *eventfd, enum res_type type)
4396 struct mem_cgroup_thresholds *thresholds;
4397 struct mem_cgroup_threshold_ary *new;
4398 unsigned long usage;
4399 int i, j, size, entries;
4401 mutex_lock(&memcg->thresholds_lock);
4404 thresholds = &memcg->thresholds;
4405 usage = mem_cgroup_usage(memcg, false);
4406 } else if (type == _MEMSWAP) {
4407 thresholds = &memcg->memsw_thresholds;
4408 usage = mem_cgroup_usage(memcg, true);
4412 if (!thresholds->primary)
4415 /* Check if a threshold crossed before removing */
4416 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4418 /* Calculate new number of threshold */
4420 for (i = 0; i < thresholds->primary->size; i++) {
4421 if (thresholds->primary->entries[i].eventfd != eventfd)
4427 new = thresholds->spare;
4429 /* If no items related to eventfd have been cleared, nothing to do */
4433 /* Set thresholds array to NULL if we don't have thresholds */
4442 /* Copy thresholds and find current threshold */
4443 new->current_threshold = -1;
4444 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4445 if (thresholds->primary->entries[i].eventfd == eventfd)
4448 new->entries[j] = thresholds->primary->entries[i];
4449 if (new->entries[j].threshold <= usage) {
4451 * new->current_threshold will not be used
4452 * until rcu_assign_pointer(), so it's safe to increment
4455 ++new->current_threshold;
4461 /* Swap primary and spare array */
4462 thresholds->spare = thresholds->primary;
4464 rcu_assign_pointer(thresholds->primary, new);
4466 /* To be sure that nobody uses thresholds */
4469 /* If all events are unregistered, free the spare array */
4471 kfree(thresholds->spare);
4472 thresholds->spare = NULL;
4475 mutex_unlock(&memcg->thresholds_lock);
4478 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4479 struct eventfd_ctx *eventfd)
4481 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4484 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4485 struct eventfd_ctx *eventfd)
4487 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4490 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4491 struct eventfd_ctx *eventfd, const char *args)
4493 struct mem_cgroup_eventfd_list *event;
4495 event = kmalloc(sizeof(*event), GFP_KERNEL);
4499 spin_lock(&memcg_oom_lock);
4501 event->eventfd = eventfd;
4502 list_add(&event->list, &memcg->oom_notify);
4504 /* already in OOM ? */
4505 if (memcg->under_oom)
4506 eventfd_signal(eventfd, 1);
4507 spin_unlock(&memcg_oom_lock);
4512 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4513 struct eventfd_ctx *eventfd)
4515 struct mem_cgroup_eventfd_list *ev, *tmp;
4517 spin_lock(&memcg_oom_lock);
4519 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4520 if (ev->eventfd == eventfd) {
4521 list_del(&ev->list);
4526 spin_unlock(&memcg_oom_lock);
4529 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4531 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4533 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4534 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4535 seq_printf(sf, "oom_kill %lu\n",
4536 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4540 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4541 struct cftype *cft, u64 val)
4543 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4545 /* cannot set to root cgroup and only 0 and 1 are allowed */
4546 if (!css->parent || !((val == 0) || (val == 1)))
4549 memcg->oom_kill_disable = val;
4551 memcg_oom_recover(memcg);
4556 #ifdef CONFIG_CGROUP_WRITEBACK
4558 #include <trace/events/writeback.h>
4560 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4562 return wb_domain_init(&memcg->cgwb_domain, gfp);
4565 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4567 wb_domain_exit(&memcg->cgwb_domain);
4570 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4572 wb_domain_size_changed(&memcg->cgwb_domain);
4575 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4577 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4579 if (!memcg->css.parent)
4582 return &memcg->cgwb_domain;
4586 * idx can be of type enum memcg_stat_item or node_stat_item.
4587 * Keep in sync with memcg_exact_page().
4589 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4591 long x = atomic_long_read(&memcg->vmstats[idx]);
4594 for_each_online_cpu(cpu)
4595 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4602 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4603 * @wb: bdi_writeback in question
4604 * @pfilepages: out parameter for number of file pages
4605 * @pheadroom: out parameter for number of allocatable pages according to memcg
4606 * @pdirty: out parameter for number of dirty pages
4607 * @pwriteback: out parameter for number of pages under writeback
4609 * Determine the numbers of file, headroom, dirty, and writeback pages in
4610 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4611 * is a bit more involved.
4613 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4614 * headroom is calculated as the lowest headroom of itself and the
4615 * ancestors. Note that this doesn't consider the actual amount of
4616 * available memory in the system. The caller should further cap
4617 * *@pheadroom accordingly.
4619 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4620 unsigned long *pheadroom, unsigned long *pdirty,
4621 unsigned long *pwriteback)
4623 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4624 struct mem_cgroup *parent;
4626 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4628 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4629 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4630 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4631 *pheadroom = PAGE_COUNTER_MAX;
4633 while ((parent = parent_mem_cgroup(memcg))) {
4634 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4635 READ_ONCE(memcg->memory.high));
4636 unsigned long used = page_counter_read(&memcg->memory);
4638 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4644 * Foreign dirty flushing
4646 * There's an inherent mismatch between memcg and writeback. The former
4647 * trackes ownership per-page while the latter per-inode. This was a
4648 * deliberate design decision because honoring per-page ownership in the
4649 * writeback path is complicated, may lead to higher CPU and IO overheads
4650 * and deemed unnecessary given that write-sharing an inode across
4651 * different cgroups isn't a common use-case.
4653 * Combined with inode majority-writer ownership switching, this works well
4654 * enough in most cases but there are some pathological cases. For
4655 * example, let's say there are two cgroups A and B which keep writing to
4656 * different but confined parts of the same inode. B owns the inode and
4657 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4658 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4659 * triggering background writeback. A will be slowed down without a way to
4660 * make writeback of the dirty pages happen.
4662 * Conditions like the above can lead to a cgroup getting repatedly and
4663 * severely throttled after making some progress after each
4664 * dirty_expire_interval while the underyling IO device is almost
4667 * Solving this problem completely requires matching the ownership tracking
4668 * granularities between memcg and writeback in either direction. However,
4669 * the more egregious behaviors can be avoided by simply remembering the
4670 * most recent foreign dirtying events and initiating remote flushes on
4671 * them when local writeback isn't enough to keep the memory clean enough.
4673 * The following two functions implement such mechanism. When a foreign
4674 * page - a page whose memcg and writeback ownerships don't match - is
4675 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4676 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4677 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4678 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4679 * foreign bdi_writebacks which haven't expired. Both the numbers of
4680 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4681 * limited to MEMCG_CGWB_FRN_CNT.
4683 * The mechanism only remembers IDs and doesn't hold any object references.
4684 * As being wrong occasionally doesn't matter, updates and accesses to the
4685 * records are lockless and racy.
4687 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4688 struct bdi_writeback *wb)
4690 struct mem_cgroup *memcg = page->mem_cgroup;
4691 struct memcg_cgwb_frn *frn;
4692 u64 now = get_jiffies_64();
4693 u64 oldest_at = now;
4697 trace_track_foreign_dirty(page, wb);
4700 * Pick the slot to use. If there is already a slot for @wb, keep
4701 * using it. If not replace the oldest one which isn't being
4704 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4705 frn = &memcg->cgwb_frn[i];
4706 if (frn->bdi_id == wb->bdi->id &&
4707 frn->memcg_id == wb->memcg_css->id)
4709 if (time_before64(frn->at, oldest_at) &&
4710 atomic_read(&frn->done.cnt) == 1) {
4712 oldest_at = frn->at;
4716 if (i < MEMCG_CGWB_FRN_CNT) {
4718 * Re-using an existing one. Update timestamp lazily to
4719 * avoid making the cacheline hot. We want them to be
4720 * reasonably up-to-date and significantly shorter than
4721 * dirty_expire_interval as that's what expires the record.
4722 * Use the shorter of 1s and dirty_expire_interval / 8.
4724 unsigned long update_intv =
4725 min_t(unsigned long, HZ,
4726 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4728 if (time_before64(frn->at, now - update_intv))
4730 } else if (oldest >= 0) {
4731 /* replace the oldest free one */
4732 frn = &memcg->cgwb_frn[oldest];
4733 frn->bdi_id = wb->bdi->id;
4734 frn->memcg_id = wb->memcg_css->id;
4739 /* issue foreign writeback flushes for recorded foreign dirtying events */
4740 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4742 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4743 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4744 u64 now = jiffies_64;
4747 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4748 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4751 * If the record is older than dirty_expire_interval,
4752 * writeback on it has already started. No need to kick it
4753 * off again. Also, don't start a new one if there's
4754 * already one in flight.
4756 if (time_after64(frn->at, now - intv) &&
4757 atomic_read(&frn->done.cnt) == 1) {
4759 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4760 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4761 WB_REASON_FOREIGN_FLUSH,
4767 #else /* CONFIG_CGROUP_WRITEBACK */
4769 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4774 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4778 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4782 #endif /* CONFIG_CGROUP_WRITEBACK */
4785 * DO NOT USE IN NEW FILES.
4787 * "cgroup.event_control" implementation.
4789 * This is way over-engineered. It tries to support fully configurable
4790 * events for each user. Such level of flexibility is completely
4791 * unnecessary especially in the light of the planned unified hierarchy.
4793 * Please deprecate this and replace with something simpler if at all
4798 * Unregister event and free resources.
4800 * Gets called from workqueue.
4802 static void memcg_event_remove(struct work_struct *work)
4804 struct mem_cgroup_event *event =
4805 container_of(work, struct mem_cgroup_event, remove);
4806 struct mem_cgroup *memcg = event->memcg;
4808 remove_wait_queue(event->wqh, &event->wait);
4810 event->unregister_event(memcg, event->eventfd);
4812 /* Notify userspace the event is going away. */
4813 eventfd_signal(event->eventfd, 1);
4815 eventfd_ctx_put(event->eventfd);
4817 css_put(&memcg->css);
4821 * Gets called on EPOLLHUP on eventfd when user closes it.
4823 * Called with wqh->lock held and interrupts disabled.
4825 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4826 int sync, void *key)
4828 struct mem_cgroup_event *event =
4829 container_of(wait, struct mem_cgroup_event, wait);
4830 struct mem_cgroup *memcg = event->memcg;
4831 __poll_t flags = key_to_poll(key);
4833 if (flags & EPOLLHUP) {
4835 * If the event has been detached at cgroup removal, we
4836 * can simply return knowing the other side will cleanup
4839 * We can't race against event freeing since the other
4840 * side will require wqh->lock via remove_wait_queue(),
4843 spin_lock(&memcg->event_list_lock);
4844 if (!list_empty(&event->list)) {
4845 list_del_init(&event->list);
4847 * We are in atomic context, but cgroup_event_remove()
4848 * may sleep, so we have to call it in workqueue.
4850 schedule_work(&event->remove);
4852 spin_unlock(&memcg->event_list_lock);
4858 static void memcg_event_ptable_queue_proc(struct file *file,
4859 wait_queue_head_t *wqh, poll_table *pt)
4861 struct mem_cgroup_event *event =
4862 container_of(pt, struct mem_cgroup_event, pt);
4865 add_wait_queue(wqh, &event->wait);
4869 * DO NOT USE IN NEW FILES.
4871 * Parse input and register new cgroup event handler.
4873 * Input must be in format '<event_fd> <control_fd> <args>'.
4874 * Interpretation of args is defined by control file implementation.
4876 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4877 char *buf, size_t nbytes, loff_t off)
4879 struct cgroup_subsys_state *css = of_css(of);
4880 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4881 struct mem_cgroup_event *event;
4882 struct cgroup_subsys_state *cfile_css;
4883 unsigned int efd, cfd;
4890 buf = strstrip(buf);
4892 efd = simple_strtoul(buf, &endp, 10);
4897 cfd = simple_strtoul(buf, &endp, 10);
4898 if ((*endp != ' ') && (*endp != '\0'))
4902 event = kzalloc(sizeof(*event), GFP_KERNEL);
4906 event->memcg = memcg;
4907 INIT_LIST_HEAD(&event->list);
4908 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4909 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4910 INIT_WORK(&event->remove, memcg_event_remove);
4918 event->eventfd = eventfd_ctx_fileget(efile.file);
4919 if (IS_ERR(event->eventfd)) {
4920 ret = PTR_ERR(event->eventfd);
4927 goto out_put_eventfd;
4930 /* the process need read permission on control file */
4931 /* AV: shouldn't we check that it's been opened for read instead? */
4932 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4937 * Determine the event callbacks and set them in @event. This used
4938 * to be done via struct cftype but cgroup core no longer knows
4939 * about these events. The following is crude but the whole thing
4940 * is for compatibility anyway.
4942 * DO NOT ADD NEW FILES.
4944 name = cfile.file->f_path.dentry->d_name.name;
4946 if (!strcmp(name, "memory.usage_in_bytes")) {
4947 event->register_event = mem_cgroup_usage_register_event;
4948 event->unregister_event = mem_cgroup_usage_unregister_event;
4949 } else if (!strcmp(name, "memory.oom_control")) {
4950 event->register_event = mem_cgroup_oom_register_event;
4951 event->unregister_event = mem_cgroup_oom_unregister_event;
4952 } else if (!strcmp(name, "memory.pressure_level")) {
4953 event->register_event = vmpressure_register_event;
4954 event->unregister_event = vmpressure_unregister_event;
4955 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4956 event->register_event = memsw_cgroup_usage_register_event;
4957 event->unregister_event = memsw_cgroup_usage_unregister_event;
4964 * Verify @cfile should belong to @css. Also, remaining events are
4965 * automatically removed on cgroup destruction but the removal is
4966 * asynchronous, so take an extra ref on @css.
4968 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4969 &memory_cgrp_subsys);
4971 if (IS_ERR(cfile_css))
4973 if (cfile_css != css) {
4978 ret = event->register_event(memcg, event->eventfd, buf);
4982 vfs_poll(efile.file, &event->pt);
4984 spin_lock(&memcg->event_list_lock);
4985 list_add(&event->list, &memcg->event_list);
4986 spin_unlock(&memcg->event_list_lock);
4998 eventfd_ctx_put(event->eventfd);
5007 static struct cftype mem_cgroup_legacy_files[] = {
5009 .name = "usage_in_bytes",
5010 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5011 .read_u64 = mem_cgroup_read_u64,
5014 .name = "max_usage_in_bytes",
5015 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5016 .write = mem_cgroup_reset,
5017 .read_u64 = mem_cgroup_read_u64,
5020 .name = "limit_in_bytes",
5021 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5022 .write = mem_cgroup_write,
5023 .read_u64 = mem_cgroup_read_u64,
5026 .name = "soft_limit_in_bytes",
5027 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5028 .write = mem_cgroup_write,
5029 .read_u64 = mem_cgroup_read_u64,
5033 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5034 .write = mem_cgroup_reset,
5035 .read_u64 = mem_cgroup_read_u64,
5039 .seq_show = memcg_stat_show,
5042 .name = "force_empty",
5043 .write = mem_cgroup_force_empty_write,
5046 .name = "use_hierarchy",
5047 .write_u64 = mem_cgroup_hierarchy_write,
5048 .read_u64 = mem_cgroup_hierarchy_read,
5051 .name = "cgroup.event_control", /* XXX: for compat */
5052 .write = memcg_write_event_control,
5053 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
5056 .name = "swappiness",
5057 .read_u64 = mem_cgroup_swappiness_read,
5058 .write_u64 = mem_cgroup_swappiness_write,
5061 .name = "move_charge_at_immigrate",
5062 .read_u64 = mem_cgroup_move_charge_read,
5063 .write_u64 = mem_cgroup_move_charge_write,
5066 .name = "oom_control",
5067 .seq_show = mem_cgroup_oom_control_read,
5068 .write_u64 = mem_cgroup_oom_control_write,
5069 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5072 .name = "pressure_level",
5076 .name = "numa_stat",
5077 .seq_show = memcg_numa_stat_show,
5081 .name = "kmem.limit_in_bytes",
5082 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5083 .write = mem_cgroup_write,
5084 .read_u64 = mem_cgroup_read_u64,
5087 .name = "kmem.usage_in_bytes",
5088 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5089 .read_u64 = mem_cgroup_read_u64,
5092 .name = "kmem.failcnt",
5093 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5094 .write = mem_cgroup_reset,
5095 .read_u64 = mem_cgroup_read_u64,
5098 .name = "kmem.max_usage_in_bytes",
5099 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5100 .write = mem_cgroup_reset,
5101 .read_u64 = mem_cgroup_read_u64,
5103 #if defined(CONFIG_MEMCG_KMEM) && \
5104 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5106 .name = "kmem.slabinfo",
5107 .seq_start = memcg_slab_start,
5108 .seq_next = memcg_slab_next,
5109 .seq_stop = memcg_slab_stop,
5110 .seq_show = memcg_slab_show,
5114 .name = "kmem.tcp.limit_in_bytes",
5115 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5116 .write = mem_cgroup_write,
5117 .read_u64 = mem_cgroup_read_u64,
5120 .name = "kmem.tcp.usage_in_bytes",
5121 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5122 .read_u64 = mem_cgroup_read_u64,
5125 .name = "kmem.tcp.failcnt",
5126 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5127 .write = mem_cgroup_reset,
5128 .read_u64 = mem_cgroup_read_u64,
5131 .name = "kmem.tcp.max_usage_in_bytes",
5132 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5133 .write = mem_cgroup_reset,
5134 .read_u64 = mem_cgroup_read_u64,
5136 { }, /* terminate */
5140 * Private memory cgroup IDR
5142 * Swap-out records and page cache shadow entries need to store memcg
5143 * references in constrained space, so we maintain an ID space that is
5144 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5145 * memory-controlled cgroups to 64k.
5147 * However, there usually are many references to the offline CSS after
5148 * the cgroup has been destroyed, such as page cache or reclaimable
5149 * slab objects, that don't need to hang on to the ID. We want to keep
5150 * those dead CSS from occupying IDs, or we might quickly exhaust the
5151 * relatively small ID space and prevent the creation of new cgroups
5152 * even when there are much fewer than 64k cgroups - possibly none.
5154 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5155 * be freed and recycled when it's no longer needed, which is usually
5156 * when the CSS is offlined.
5158 * The only exception to that are records of swapped out tmpfs/shmem
5159 * pages that need to be attributed to live ancestors on swapin. But
5160 * those references are manageable from userspace.
5163 static DEFINE_IDR(mem_cgroup_idr);
5165 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5167 if (memcg->id.id > 0) {
5168 idr_remove(&mem_cgroup_idr, memcg->id.id);
5173 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5176 refcount_add(n, &memcg->id.ref);
5179 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5181 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5182 mem_cgroup_id_remove(memcg);
5184 /* Memcg ID pins CSS */
5185 css_put(&memcg->css);
5189 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5191 mem_cgroup_id_put_many(memcg, 1);
5195 * mem_cgroup_from_id - look up a memcg from a memcg id
5196 * @id: the memcg id to look up
5198 * Caller must hold rcu_read_lock().
5200 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5202 WARN_ON_ONCE(!rcu_read_lock_held());
5203 return idr_find(&mem_cgroup_idr, id);
5206 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5208 struct mem_cgroup_per_node *pn;
5211 * This routine is called against possible nodes.
5212 * But it's BUG to call kmalloc() against offline node.
5214 * TODO: this routine can waste much memory for nodes which will
5215 * never be onlined. It's better to use memory hotplug callback
5218 if (!node_state(node, N_NORMAL_MEMORY))
5220 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5224 pn->lruvec_stat_local = alloc_percpu(struct lruvec_stat);
5225 if (!pn->lruvec_stat_local) {
5230 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
5231 if (!pn->lruvec_stat_cpu) {
5232 free_percpu(pn->lruvec_stat_local);
5237 lruvec_init(&pn->lruvec);
5238 pn->usage_in_excess = 0;
5239 pn->on_tree = false;
5242 memcg->nodeinfo[node] = pn;
5246 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5248 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5253 free_percpu(pn->lruvec_stat_cpu);
5254 free_percpu(pn->lruvec_stat_local);
5258 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5263 free_mem_cgroup_per_node_info(memcg, node);
5264 free_percpu(memcg->vmstats_percpu);
5265 free_percpu(memcg->vmstats_local);
5269 static void mem_cgroup_free(struct mem_cgroup *memcg)
5271 memcg_wb_domain_exit(memcg);
5273 * Flush percpu vmstats and vmevents to guarantee the value correctness
5274 * on parent's and all ancestor levels.
5276 memcg_flush_percpu_vmstats(memcg);
5277 memcg_flush_percpu_vmevents(memcg);
5278 __mem_cgroup_free(memcg);
5281 static struct mem_cgroup *mem_cgroup_alloc(void)
5283 struct mem_cgroup *memcg;
5286 int __maybe_unused i;
5287 long error = -ENOMEM;
5289 size = sizeof(struct mem_cgroup);
5290 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5292 memcg = kzalloc(size, GFP_KERNEL);
5294 return ERR_PTR(error);
5296 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5297 1, MEM_CGROUP_ID_MAX,
5299 if (memcg->id.id < 0) {
5300 error = memcg->id.id;
5304 memcg->vmstats_local = alloc_percpu(struct memcg_vmstats_percpu);
5305 if (!memcg->vmstats_local)
5308 memcg->vmstats_percpu = alloc_percpu(struct memcg_vmstats_percpu);
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;
5355 long error = -ENOMEM;
5357 memcg = mem_cgroup_alloc();
5359 return ERR_CAST(memcg);
5361 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5362 memcg->soft_limit = PAGE_COUNTER_MAX;
5363 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5365 memcg->swappiness = mem_cgroup_swappiness(parent);
5366 memcg->oom_kill_disable = parent->oom_kill_disable;
5368 if (parent && parent->use_hierarchy) {
5369 memcg->use_hierarchy = true;
5370 page_counter_init(&memcg->memory, &parent->memory);
5371 page_counter_init(&memcg->swap, &parent->swap);
5372 page_counter_init(&memcg->memsw, &parent->memsw);
5373 page_counter_init(&memcg->kmem, &parent->kmem);
5374 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5376 page_counter_init(&memcg->memory, NULL);
5377 page_counter_init(&memcg->swap, NULL);
5378 page_counter_init(&memcg->memsw, NULL);
5379 page_counter_init(&memcg->kmem, NULL);
5380 page_counter_init(&memcg->tcpmem, NULL);
5382 * Deeper hierachy with use_hierarchy == false doesn't make
5383 * much sense so let cgroup subsystem know about this
5384 * unfortunate state in our controller.
5386 if (parent != root_mem_cgroup)
5387 memory_cgrp_subsys.broken_hierarchy = true;
5390 /* The following stuff does not apply to the root */
5392 #ifdef CONFIG_MEMCG_KMEM
5393 INIT_LIST_HEAD(&memcg->kmem_caches);
5395 root_mem_cgroup = memcg;
5399 error = memcg_online_kmem(memcg);
5403 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5404 static_branch_inc(&memcg_sockets_enabled_key);
5408 mem_cgroup_id_remove(memcg);
5409 mem_cgroup_free(memcg);
5410 return ERR_PTR(error);
5413 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5415 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5418 * A memcg must be visible for memcg_expand_shrinker_maps()
5419 * by the time the maps are allocated. So, we allocate maps
5420 * here, when for_each_mem_cgroup() can't skip it.
5422 if (memcg_alloc_shrinker_maps(memcg)) {
5423 mem_cgroup_id_remove(memcg);
5427 /* Online state pins memcg ID, memcg ID pins CSS */
5428 refcount_set(&memcg->id.ref, 1);
5433 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5435 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5436 struct mem_cgroup_event *event, *tmp;
5439 * Unregister events and notify userspace.
5440 * Notify userspace about cgroup removing only after rmdir of cgroup
5441 * directory to avoid race between userspace and kernelspace.
5443 spin_lock(&memcg->event_list_lock);
5444 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5445 list_del_init(&event->list);
5446 schedule_work(&event->remove);
5448 spin_unlock(&memcg->event_list_lock);
5450 page_counter_set_min(&memcg->memory, 0);
5451 page_counter_set_low(&memcg->memory, 0);
5453 memcg_offline_kmem(memcg);
5454 wb_memcg_offline(memcg);
5456 drain_all_stock(memcg);
5458 mem_cgroup_id_put(memcg);
5461 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5463 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5465 invalidate_reclaim_iterators(memcg);
5468 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5470 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5471 int __maybe_unused i;
5473 #ifdef CONFIG_CGROUP_WRITEBACK
5474 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5475 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5477 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5478 static_branch_dec(&memcg_sockets_enabled_key);
5480 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5481 static_branch_dec(&memcg_sockets_enabled_key);
5483 vmpressure_cleanup(&memcg->vmpressure);
5484 cancel_work_sync(&memcg->high_work);
5485 mem_cgroup_remove_from_trees(memcg);
5486 memcg_free_shrinker_maps(memcg);
5487 memcg_free_kmem(memcg);
5488 mem_cgroup_free(memcg);
5492 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5493 * @css: the target css
5495 * Reset the states of the mem_cgroup associated with @css. This is
5496 * invoked when the userland requests disabling on the default hierarchy
5497 * but the memcg is pinned through dependency. The memcg should stop
5498 * applying policies and should revert to the vanilla state as it may be
5499 * made visible again.
5501 * The current implementation only resets the essential configurations.
5502 * This needs to be expanded to cover all the visible parts.
5504 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5506 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5508 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5509 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5510 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
5511 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5512 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5513 page_counter_set_min(&memcg->memory, 0);
5514 page_counter_set_low(&memcg->memory, 0);
5515 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5516 memcg->soft_limit = PAGE_COUNTER_MAX;
5517 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5518 memcg_wb_domain_size_changed(memcg);
5522 /* Handlers for move charge at task migration. */
5523 static int mem_cgroup_do_precharge(unsigned long count)
5527 /* Try a single bulk charge without reclaim first, kswapd may wake */
5528 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5530 mc.precharge += count;
5534 /* Try charges one by one with reclaim, but do not retry */
5536 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5550 enum mc_target_type {
5557 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5558 unsigned long addr, pte_t ptent)
5560 struct page *page = vm_normal_page(vma, addr, ptent);
5562 if (!page || !page_mapped(page))
5564 if (PageAnon(page)) {
5565 if (!(mc.flags & MOVE_ANON))
5568 if (!(mc.flags & MOVE_FILE))
5571 if (!get_page_unless_zero(page))
5577 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5578 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5579 pte_t ptent, swp_entry_t *entry)
5581 struct page *page = NULL;
5582 swp_entry_t ent = pte_to_swp_entry(ptent);
5584 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
5588 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5589 * a device and because they are not accessible by CPU they are store
5590 * as special swap entry in the CPU page table.
5592 if (is_device_private_entry(ent)) {
5593 page = device_private_entry_to_page(ent);
5595 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5596 * a refcount of 1 when free (unlike normal page)
5598 if (!page_ref_add_unless(page, 1, 1))
5604 * Because lookup_swap_cache() updates some statistics counter,
5605 * we call find_get_page() with swapper_space directly.
5607 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5608 entry->val = ent.val;
5613 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5614 pte_t ptent, swp_entry_t *entry)
5620 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5621 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5623 struct page *page = NULL;
5624 struct address_space *mapping;
5627 if (!vma->vm_file) /* anonymous vma */
5629 if (!(mc.flags & MOVE_FILE))
5632 mapping = vma->vm_file->f_mapping;
5633 pgoff = linear_page_index(vma, addr);
5635 /* page is moved even if it's not RSS of this task(page-faulted). */
5637 /* shmem/tmpfs may report page out on swap: account for that too. */
5638 if (shmem_mapping(mapping)) {
5639 page = find_get_entry(mapping, pgoff);
5640 if (xa_is_value(page)) {
5641 swp_entry_t swp = radix_to_swp_entry(page);
5643 page = find_get_page(swap_address_space(swp),
5647 page = find_get_page(mapping, pgoff);
5649 page = find_get_page(mapping, pgoff);
5655 * mem_cgroup_move_account - move account of the page
5657 * @compound: charge the page as compound or small page
5658 * @from: mem_cgroup which the page is moved from.
5659 * @to: mem_cgroup which the page is moved to. @from != @to.
5661 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5663 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5666 static int mem_cgroup_move_account(struct page *page,
5668 struct mem_cgroup *from,
5669 struct mem_cgroup *to)
5671 struct lruvec *from_vec, *to_vec;
5672 struct pglist_data *pgdat;
5673 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5676 VM_BUG_ON(from == to);
5677 VM_BUG_ON_PAGE(PageLRU(page), page);
5678 VM_BUG_ON(compound && !PageTransHuge(page));
5681 * Prevent mem_cgroup_migrate() from looking at
5682 * page->mem_cgroup of its source page while we change it.
5685 if (!trylock_page(page))
5689 if (page->mem_cgroup != from)
5692 pgdat = page_pgdat(page);
5693 from_vec = mem_cgroup_lruvec(from, pgdat);
5694 to_vec = mem_cgroup_lruvec(to, pgdat);
5696 lock_page_memcg(page);
5698 if (PageAnon(page)) {
5699 if (page_mapped(page)) {
5700 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5701 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5702 if (PageTransHuge(page)) {
5703 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5705 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5711 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5712 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5714 if (PageSwapBacked(page)) {
5715 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5716 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5719 if (page_mapped(page)) {
5720 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5721 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5724 if (PageDirty(page)) {
5725 struct address_space *mapping = page_mapping(page);
5727 if (mapping_cap_account_dirty(mapping)) {
5728 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5730 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5736 if (PageWriteback(page)) {
5737 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5738 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5742 * All state has been migrated, let's switch to the new memcg.
5744 * It is safe to change page->mem_cgroup here because the page
5745 * is referenced, charged, isolated, and locked: we can't race
5746 * with (un)charging, migration, LRU putback, or anything else
5747 * that would rely on a stable page->mem_cgroup.
5749 * Note that lock_page_memcg is a memcg lock, not a page lock,
5750 * to save space. As soon as we switch page->mem_cgroup to a
5751 * new memcg that isn't locked, the above state can change
5752 * concurrently again. Make sure we're truly done with it.
5757 css_put(&from->css);
5759 page->mem_cgroup = to;
5761 __unlock_page_memcg(from);
5765 local_irq_disable();
5766 mem_cgroup_charge_statistics(to, page, nr_pages);
5767 memcg_check_events(to, page);
5768 mem_cgroup_charge_statistics(from, page, -nr_pages);
5769 memcg_check_events(from, page);
5778 * get_mctgt_type - get target type of moving charge
5779 * @vma: the vma the pte to be checked belongs
5780 * @addr: the address corresponding to the pte to be checked
5781 * @ptent: the pte to be checked
5782 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5785 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5786 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5787 * move charge. if @target is not NULL, the page is stored in target->page
5788 * with extra refcnt got(Callers should handle it).
5789 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5790 * target for charge migration. if @target is not NULL, the entry is stored
5792 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5793 * (so ZONE_DEVICE page and thus not on the lru).
5794 * For now we such page is charge like a regular page would be as for all
5795 * intent and purposes it is just special memory taking the place of a
5798 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5800 * Called with pte lock held.
5803 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5804 unsigned long addr, pte_t ptent, union mc_target *target)
5806 struct page *page = NULL;
5807 enum mc_target_type ret = MC_TARGET_NONE;
5808 swp_entry_t ent = { .val = 0 };
5810 if (pte_present(ptent))
5811 page = mc_handle_present_pte(vma, addr, ptent);
5812 else if (is_swap_pte(ptent))
5813 page = mc_handle_swap_pte(vma, ptent, &ent);
5814 else if (pte_none(ptent))
5815 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5817 if (!page && !ent.val)
5821 * Do only loose check w/o serialization.
5822 * mem_cgroup_move_account() checks the page is valid or
5823 * not under LRU exclusion.
5825 if (page->mem_cgroup == mc.from) {
5826 ret = MC_TARGET_PAGE;
5827 if (is_device_private_page(page))
5828 ret = MC_TARGET_DEVICE;
5830 target->page = page;
5832 if (!ret || !target)
5836 * There is a swap entry and a page doesn't exist or isn't charged.
5837 * But we cannot move a tail-page in a THP.
5839 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5840 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5841 ret = MC_TARGET_SWAP;
5848 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5850 * We don't consider PMD mapped swapping or file mapped pages because THP does
5851 * not support them for now.
5852 * Caller should make sure that pmd_trans_huge(pmd) is true.
5854 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5855 unsigned long addr, pmd_t pmd, union mc_target *target)
5857 struct page *page = NULL;
5858 enum mc_target_type ret = MC_TARGET_NONE;
5860 if (unlikely(is_swap_pmd(pmd))) {
5861 VM_BUG_ON(thp_migration_supported() &&
5862 !is_pmd_migration_entry(pmd));
5865 page = pmd_page(pmd);
5866 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5867 if (!(mc.flags & MOVE_ANON))
5869 if (page->mem_cgroup == mc.from) {
5870 ret = MC_TARGET_PAGE;
5873 target->page = page;
5879 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5880 unsigned long addr, pmd_t pmd, union mc_target *target)
5882 return MC_TARGET_NONE;
5886 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5887 unsigned long addr, unsigned long end,
5888 struct mm_walk *walk)
5890 struct vm_area_struct *vma = walk->vma;
5894 ptl = pmd_trans_huge_lock(pmd, vma);
5897 * Note their can not be MC_TARGET_DEVICE for now as we do not
5898 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5899 * this might change.
5901 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5902 mc.precharge += HPAGE_PMD_NR;
5907 if (pmd_trans_unstable(pmd))
5909 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5910 for (; addr != end; pte++, addr += PAGE_SIZE)
5911 if (get_mctgt_type(vma, addr, *pte, NULL))
5912 mc.precharge++; /* increment precharge temporarily */
5913 pte_unmap_unlock(pte - 1, ptl);
5919 static const struct mm_walk_ops precharge_walk_ops = {
5920 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5923 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5925 unsigned long precharge;
5928 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5929 mmap_read_unlock(mm);
5931 precharge = mc.precharge;
5937 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5939 unsigned long precharge = mem_cgroup_count_precharge(mm);
5941 VM_BUG_ON(mc.moving_task);
5942 mc.moving_task = current;
5943 return mem_cgroup_do_precharge(precharge);
5946 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5947 static void __mem_cgroup_clear_mc(void)
5949 struct mem_cgroup *from = mc.from;
5950 struct mem_cgroup *to = mc.to;
5952 /* we must uncharge all the leftover precharges from mc.to */
5954 cancel_charge(mc.to, mc.precharge);
5958 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5959 * we must uncharge here.
5961 if (mc.moved_charge) {
5962 cancel_charge(mc.from, mc.moved_charge);
5963 mc.moved_charge = 0;
5965 /* we must fixup refcnts and charges */
5966 if (mc.moved_swap) {
5967 /* uncharge swap account from the old cgroup */
5968 if (!mem_cgroup_is_root(mc.from))
5969 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5971 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5974 * we charged both to->memory and to->memsw, so we
5975 * should uncharge to->memory.
5977 if (!mem_cgroup_is_root(mc.to))
5978 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5982 memcg_oom_recover(from);
5983 memcg_oom_recover(to);
5984 wake_up_all(&mc.waitq);
5987 static void mem_cgroup_clear_mc(void)
5989 struct mm_struct *mm = mc.mm;
5992 * we must clear moving_task before waking up waiters at the end of
5995 mc.moving_task = NULL;
5996 __mem_cgroup_clear_mc();
5997 spin_lock(&mc.lock);
6001 spin_unlock(&mc.lock);
6006 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6008 struct cgroup_subsys_state *css;
6009 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
6010 struct mem_cgroup *from;
6011 struct task_struct *leader, *p;
6012 struct mm_struct *mm;
6013 unsigned long move_flags;
6016 /* charge immigration isn't supported on the default hierarchy */
6017 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6021 * Multi-process migrations only happen on the default hierarchy
6022 * where charge immigration is not used. Perform charge
6023 * immigration if @tset contains a leader and whine if there are
6027 cgroup_taskset_for_each_leader(leader, css, tset) {
6030 memcg = mem_cgroup_from_css(css);
6036 * We are now commited to this value whatever it is. Changes in this
6037 * tunable will only affect upcoming migrations, not the current one.
6038 * So we need to save it, and keep it going.
6040 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
6044 from = mem_cgroup_from_task(p);
6046 VM_BUG_ON(from == memcg);
6048 mm = get_task_mm(p);
6051 /* We move charges only when we move a owner of the mm */
6052 if (mm->owner == p) {
6055 VM_BUG_ON(mc.precharge);
6056 VM_BUG_ON(mc.moved_charge);
6057 VM_BUG_ON(mc.moved_swap);
6059 spin_lock(&mc.lock);
6063 mc.flags = move_flags;
6064 spin_unlock(&mc.lock);
6065 /* We set mc.moving_task later */
6067 ret = mem_cgroup_precharge_mc(mm);
6069 mem_cgroup_clear_mc();
6076 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6079 mem_cgroup_clear_mc();
6082 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6083 unsigned long addr, unsigned long end,
6084 struct mm_walk *walk)
6087 struct vm_area_struct *vma = walk->vma;
6090 enum mc_target_type target_type;
6091 union mc_target target;
6094 ptl = pmd_trans_huge_lock(pmd, vma);
6096 if (mc.precharge < HPAGE_PMD_NR) {
6100 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6101 if (target_type == MC_TARGET_PAGE) {
6103 if (!isolate_lru_page(page)) {
6104 if (!mem_cgroup_move_account(page, true,
6106 mc.precharge -= HPAGE_PMD_NR;
6107 mc.moved_charge += HPAGE_PMD_NR;
6109 putback_lru_page(page);
6112 } else if (target_type == MC_TARGET_DEVICE) {
6114 if (!mem_cgroup_move_account(page, true,
6116 mc.precharge -= HPAGE_PMD_NR;
6117 mc.moved_charge += HPAGE_PMD_NR;
6125 if (pmd_trans_unstable(pmd))
6128 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6129 for (; addr != end; addr += PAGE_SIZE) {
6130 pte_t ptent = *(pte++);
6131 bool device = false;
6137 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6138 case MC_TARGET_DEVICE:
6141 case MC_TARGET_PAGE:
6144 * We can have a part of the split pmd here. Moving it
6145 * can be done but it would be too convoluted so simply
6146 * ignore such a partial THP and keep it in original
6147 * memcg. There should be somebody mapping the head.
6149 if (PageTransCompound(page))
6151 if (!device && isolate_lru_page(page))
6153 if (!mem_cgroup_move_account(page, false,
6156 /* we uncharge from mc.from later. */
6160 putback_lru_page(page);
6161 put: /* get_mctgt_type() gets the page */
6164 case MC_TARGET_SWAP:
6166 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6168 mem_cgroup_id_get_many(mc.to, 1);
6169 /* we fixup other refcnts and charges later. */
6177 pte_unmap_unlock(pte - 1, ptl);
6182 * We have consumed all precharges we got in can_attach().
6183 * We try charge one by one, but don't do any additional
6184 * charges to mc.to if we have failed in charge once in attach()
6187 ret = mem_cgroup_do_precharge(1);
6195 static const struct mm_walk_ops charge_walk_ops = {
6196 .pmd_entry = mem_cgroup_move_charge_pte_range,
6199 static void mem_cgroup_move_charge(void)
6201 lru_add_drain_all();
6203 * Signal lock_page_memcg() to take the memcg's move_lock
6204 * while we're moving its pages to another memcg. Then wait
6205 * for already started RCU-only updates to finish.
6207 atomic_inc(&mc.from->moving_account);
6210 if (unlikely(!mmap_read_trylock(mc.mm))) {
6212 * Someone who are holding the mmap_lock might be waiting in
6213 * waitq. So we cancel all extra charges, wake up all waiters,
6214 * and retry. Because we cancel precharges, we might not be able
6215 * to move enough charges, but moving charge is a best-effort
6216 * feature anyway, so it wouldn't be a big problem.
6218 __mem_cgroup_clear_mc();
6223 * When we have consumed all precharges and failed in doing
6224 * additional charge, the page walk just aborts.
6226 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6229 mmap_read_unlock(mc.mm);
6230 atomic_dec(&mc.from->moving_account);
6233 static void mem_cgroup_move_task(void)
6236 mem_cgroup_move_charge();
6237 mem_cgroup_clear_mc();
6240 #else /* !CONFIG_MMU */
6241 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6245 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6248 static void mem_cgroup_move_task(void)
6254 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6255 * to verify whether we're attached to the default hierarchy on each mount
6258 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6261 * use_hierarchy is forced on the default hierarchy. cgroup core
6262 * guarantees that @root doesn't have any children, so turning it
6263 * on for the root memcg is enough.
6265 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6266 root_mem_cgroup->use_hierarchy = true;
6268 root_mem_cgroup->use_hierarchy = false;
6271 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6273 if (value == PAGE_COUNTER_MAX)
6274 seq_puts(m, "max\n");
6276 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6281 static u64 memory_current_read(struct cgroup_subsys_state *css,
6284 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6286 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6289 static int memory_min_show(struct seq_file *m, void *v)
6291 return seq_puts_memcg_tunable(m,
6292 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6295 static ssize_t memory_min_write(struct kernfs_open_file *of,
6296 char *buf, size_t nbytes, loff_t off)
6298 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6302 buf = strstrip(buf);
6303 err = page_counter_memparse(buf, "max", &min);
6307 page_counter_set_min(&memcg->memory, min);
6312 static int memory_low_show(struct seq_file *m, void *v)
6314 return seq_puts_memcg_tunable(m,
6315 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6318 static ssize_t memory_low_write(struct kernfs_open_file *of,
6319 char *buf, size_t nbytes, loff_t off)
6321 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6325 buf = strstrip(buf);
6326 err = page_counter_memparse(buf, "max", &low);
6330 page_counter_set_low(&memcg->memory, low);
6335 static int memory_high_show(struct seq_file *m, void *v)
6337 return seq_puts_memcg_tunable(m,
6338 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6341 static ssize_t memory_high_write(struct kernfs_open_file *of,
6342 char *buf, size_t nbytes, loff_t off)
6344 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6345 unsigned int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
6346 bool drained = false;
6350 buf = strstrip(buf);
6351 err = page_counter_memparse(buf, "max", &high);
6355 page_counter_set_high(&memcg->memory, high);
6358 unsigned long nr_pages = page_counter_read(&memcg->memory);
6359 unsigned long reclaimed;
6361 if (nr_pages <= high)
6364 if (signal_pending(current))
6368 drain_all_stock(memcg);
6373 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6376 if (!reclaimed && !nr_retries--)
6383 static int memory_max_show(struct seq_file *m, void *v)
6385 return seq_puts_memcg_tunable(m,
6386 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6389 static ssize_t memory_max_write(struct kernfs_open_file *of,
6390 char *buf, size_t nbytes, loff_t off)
6392 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6393 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
6394 bool drained = false;
6398 buf = strstrip(buf);
6399 err = page_counter_memparse(buf, "max", &max);
6403 xchg(&memcg->memory.max, max);
6406 unsigned long nr_pages = page_counter_read(&memcg->memory);
6408 if (nr_pages <= max)
6411 if (signal_pending(current))
6415 drain_all_stock(memcg);
6421 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6427 memcg_memory_event(memcg, MEMCG_OOM);
6428 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6432 memcg_wb_domain_size_changed(memcg);
6436 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6438 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6439 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6440 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6441 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6442 seq_printf(m, "oom_kill %lu\n",
6443 atomic_long_read(&events[MEMCG_OOM_KILL]));
6446 static int memory_events_show(struct seq_file *m, void *v)
6448 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6450 __memory_events_show(m, memcg->memory_events);
6454 static int memory_events_local_show(struct seq_file *m, void *v)
6456 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6458 __memory_events_show(m, memcg->memory_events_local);
6462 static int memory_stat_show(struct seq_file *m, void *v)
6464 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6467 buf = memory_stat_format(memcg);
6475 static int memory_oom_group_show(struct seq_file *m, void *v)
6477 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6479 seq_printf(m, "%d\n", memcg->oom_group);
6484 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6485 char *buf, size_t nbytes, loff_t off)
6487 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6490 buf = strstrip(buf);
6494 ret = kstrtoint(buf, 0, &oom_group);
6498 if (oom_group != 0 && oom_group != 1)
6501 memcg->oom_group = oom_group;
6506 static struct cftype memory_files[] = {
6509 .flags = CFTYPE_NOT_ON_ROOT,
6510 .read_u64 = memory_current_read,
6514 .flags = CFTYPE_NOT_ON_ROOT,
6515 .seq_show = memory_min_show,
6516 .write = memory_min_write,
6520 .flags = CFTYPE_NOT_ON_ROOT,
6521 .seq_show = memory_low_show,
6522 .write = memory_low_write,
6526 .flags = CFTYPE_NOT_ON_ROOT,
6527 .seq_show = memory_high_show,
6528 .write = memory_high_write,
6532 .flags = CFTYPE_NOT_ON_ROOT,
6533 .seq_show = memory_max_show,
6534 .write = memory_max_write,
6538 .flags = CFTYPE_NOT_ON_ROOT,
6539 .file_offset = offsetof(struct mem_cgroup, events_file),
6540 .seq_show = memory_events_show,
6543 .name = "events.local",
6544 .flags = CFTYPE_NOT_ON_ROOT,
6545 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6546 .seq_show = memory_events_local_show,
6550 .seq_show = memory_stat_show,
6553 .name = "oom.group",
6554 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6555 .seq_show = memory_oom_group_show,
6556 .write = memory_oom_group_write,
6561 struct cgroup_subsys memory_cgrp_subsys = {
6562 .css_alloc = mem_cgroup_css_alloc,
6563 .css_online = mem_cgroup_css_online,
6564 .css_offline = mem_cgroup_css_offline,
6565 .css_released = mem_cgroup_css_released,
6566 .css_free = mem_cgroup_css_free,
6567 .css_reset = mem_cgroup_css_reset,
6568 .can_attach = mem_cgroup_can_attach,
6569 .cancel_attach = mem_cgroup_cancel_attach,
6570 .post_attach = mem_cgroup_move_task,
6571 .bind = mem_cgroup_bind,
6572 .dfl_cftypes = memory_files,
6573 .legacy_cftypes = mem_cgroup_legacy_files,
6578 * This function calculates an individual cgroup's effective
6579 * protection which is derived from its own memory.min/low, its
6580 * parent's and siblings' settings, as well as the actual memory
6581 * distribution in the tree.
6583 * The following rules apply to the effective protection values:
6585 * 1. At the first level of reclaim, effective protection is equal to
6586 * the declared protection in memory.min and memory.low.
6588 * 2. To enable safe delegation of the protection configuration, at
6589 * subsequent levels the effective protection is capped to the
6590 * parent's effective protection.
6592 * 3. To make complex and dynamic subtrees easier to configure, the
6593 * user is allowed to overcommit the declared protection at a given
6594 * level. If that is the case, the parent's effective protection is
6595 * distributed to the children in proportion to how much protection
6596 * they have declared and how much of it they are utilizing.
6598 * This makes distribution proportional, but also work-conserving:
6599 * if one cgroup claims much more protection than it uses memory,
6600 * the unused remainder is available to its siblings.
6602 * 4. Conversely, when the declared protection is undercommitted at a
6603 * given level, the distribution of the larger parental protection
6604 * budget is NOT proportional. A cgroup's protection from a sibling
6605 * is capped to its own memory.min/low setting.
6607 * 5. However, to allow protecting recursive subtrees from each other
6608 * without having to declare each individual cgroup's fixed share
6609 * of the ancestor's claim to protection, any unutilized -
6610 * "floating" - protection from up the tree is distributed in
6611 * proportion to each cgroup's *usage*. This makes the protection
6612 * neutral wrt sibling cgroups and lets them compete freely over
6613 * the shared parental protection budget, but it protects the
6614 * subtree as a whole from neighboring subtrees.
6616 * Note that 4. and 5. are not in conflict: 4. is about protecting
6617 * against immediate siblings whereas 5. is about protecting against
6618 * neighboring subtrees.
6620 static unsigned long effective_protection(unsigned long usage,
6621 unsigned long parent_usage,
6622 unsigned long setting,
6623 unsigned long parent_effective,
6624 unsigned long siblings_protected)
6626 unsigned long protected;
6629 protected = min(usage, setting);
6631 * If all cgroups at this level combined claim and use more
6632 * protection then what the parent affords them, distribute
6633 * shares in proportion to utilization.
6635 * We are using actual utilization rather than the statically
6636 * claimed protection in order to be work-conserving: claimed
6637 * but unused protection is available to siblings that would
6638 * otherwise get a smaller chunk than what they claimed.
6640 if (siblings_protected > parent_effective)
6641 return protected * parent_effective / siblings_protected;
6644 * Ok, utilized protection of all children is within what the
6645 * parent affords them, so we know whatever this child claims
6646 * and utilizes is effectively protected.
6648 * If there is unprotected usage beyond this value, reclaim
6649 * will apply pressure in proportion to that amount.
6651 * If there is unutilized protection, the cgroup will be fully
6652 * shielded from reclaim, but we do return a smaller value for
6653 * protection than what the group could enjoy in theory. This
6654 * is okay. With the overcommit distribution above, effective
6655 * protection is always dependent on how memory is actually
6656 * consumed among the siblings anyway.
6661 * If the children aren't claiming (all of) the protection
6662 * afforded to them by the parent, distribute the remainder in
6663 * proportion to the (unprotected) memory of each cgroup. That
6664 * way, cgroups that aren't explicitly prioritized wrt each
6665 * other compete freely over the allowance, but they are
6666 * collectively protected from neighboring trees.
6668 * We're using unprotected memory for the weight so that if
6669 * some cgroups DO claim explicit protection, we don't protect
6670 * the same bytes twice.
6672 * Check both usage and parent_usage against the respective
6673 * protected values. One should imply the other, but they
6674 * aren't read atomically - make sure the division is sane.
6676 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6678 if (parent_effective > siblings_protected &&
6679 parent_usage > siblings_protected &&
6680 usage > protected) {
6681 unsigned long unclaimed;
6683 unclaimed = parent_effective - siblings_protected;
6684 unclaimed *= usage - protected;
6685 unclaimed /= parent_usage - siblings_protected;
6694 * mem_cgroup_protected - check if memory consumption is in the normal range
6695 * @root: the top ancestor of the sub-tree being checked
6696 * @memcg: the memory cgroup to check
6698 * WARNING: This function is not stateless! It can only be used as part
6699 * of a top-down tree iteration, not for isolated queries.
6701 * Returns one of the following:
6702 * MEMCG_PROT_NONE: cgroup memory is not protected
6703 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
6704 * an unprotected supply of reclaimable memory from other cgroups.
6705 * MEMCG_PROT_MIN: cgroup memory is protected
6707 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
6708 struct mem_cgroup *memcg)
6710 unsigned long usage, parent_usage;
6711 struct mem_cgroup *parent;
6713 if (mem_cgroup_disabled())
6714 return MEMCG_PROT_NONE;
6717 root = root_mem_cgroup;
6719 return MEMCG_PROT_NONE;
6721 usage = page_counter_read(&memcg->memory);
6723 return MEMCG_PROT_NONE;
6725 parent = parent_mem_cgroup(memcg);
6726 /* No parent means a non-hierarchical mode on v1 memcg */
6728 return MEMCG_PROT_NONE;
6730 if (parent == root) {
6731 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6732 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6736 parent_usage = page_counter_read(&parent->memory);
6738 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6739 READ_ONCE(memcg->memory.min),
6740 READ_ONCE(parent->memory.emin),
6741 atomic_long_read(&parent->memory.children_min_usage)));
6743 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6744 READ_ONCE(memcg->memory.low),
6745 READ_ONCE(parent->memory.elow),
6746 atomic_long_read(&parent->memory.children_low_usage)));
6749 if (usage <= memcg->memory.emin)
6750 return MEMCG_PROT_MIN;
6751 else if (usage <= memcg->memory.elow)
6752 return MEMCG_PROT_LOW;
6754 return MEMCG_PROT_NONE;
6758 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6759 * @page: page to charge
6760 * @mm: mm context of the victim
6761 * @gfp_mask: reclaim mode
6763 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6764 * pages according to @gfp_mask if necessary.
6766 * Returns 0 on success. Otherwise, an error code is returned.
6768 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6770 unsigned int nr_pages = hpage_nr_pages(page);
6771 struct mem_cgroup *memcg = NULL;
6774 if (mem_cgroup_disabled())
6777 if (PageSwapCache(page)) {
6778 swp_entry_t ent = { .val = page_private(page), };
6782 * Every swap fault against a single page tries to charge the
6783 * page, bail as early as possible. shmem_unuse() encounters
6784 * already charged pages, too. page->mem_cgroup is protected
6785 * by the page lock, which serializes swap cache removal, which
6786 * in turn serializes uncharging.
6788 VM_BUG_ON_PAGE(!PageLocked(page), page);
6789 if (compound_head(page)->mem_cgroup)
6792 id = lookup_swap_cgroup_id(ent);
6794 memcg = mem_cgroup_from_id(id);
6795 if (memcg && !css_tryget_online(&memcg->css))
6801 memcg = get_mem_cgroup_from_mm(mm);
6803 ret = try_charge(memcg, gfp_mask, nr_pages);
6807 css_get(&memcg->css);
6808 commit_charge(page, memcg);
6810 local_irq_disable();
6811 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6812 memcg_check_events(memcg, page);
6815 if (PageSwapCache(page)) {
6816 swp_entry_t entry = { .val = page_private(page) };
6818 * The swap entry might not get freed for a long time,
6819 * let's not wait for it. The page already received a
6820 * memory+swap charge, drop the swap entry duplicate.
6822 mem_cgroup_uncharge_swap(entry, nr_pages);
6826 css_put(&memcg->css);
6831 struct uncharge_gather {
6832 struct mem_cgroup *memcg;
6833 unsigned long nr_pages;
6834 unsigned long pgpgout;
6835 unsigned long nr_kmem;
6836 struct page *dummy_page;
6839 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6841 memset(ug, 0, sizeof(*ug));
6844 static void uncharge_batch(const struct uncharge_gather *ug)
6846 unsigned long flags;
6848 if (!mem_cgroup_is_root(ug->memcg)) {
6849 page_counter_uncharge(&ug->memcg->memory, ug->nr_pages);
6850 if (do_memsw_account())
6851 page_counter_uncharge(&ug->memcg->memsw, ug->nr_pages);
6852 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6853 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6854 memcg_oom_recover(ug->memcg);
6857 local_irq_save(flags);
6858 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6859 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_pages);
6860 memcg_check_events(ug->memcg, ug->dummy_page);
6861 local_irq_restore(flags);
6864 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6866 unsigned long nr_pages;
6868 VM_BUG_ON_PAGE(PageLRU(page), page);
6870 if (!page->mem_cgroup)
6874 * Nobody should be changing or seriously looking at
6875 * page->mem_cgroup at this point, we have fully
6876 * exclusive access to the page.
6879 if (ug->memcg != page->mem_cgroup) {
6882 uncharge_gather_clear(ug);
6884 ug->memcg = page->mem_cgroup;
6887 nr_pages = compound_nr(page);
6888 ug->nr_pages += nr_pages;
6890 if (!PageKmemcg(page)) {
6893 ug->nr_kmem += nr_pages;
6894 __ClearPageKmemcg(page);
6897 ug->dummy_page = page;
6898 page->mem_cgroup = NULL;
6899 css_put(&ug->memcg->css);
6902 static void uncharge_list(struct list_head *page_list)
6904 struct uncharge_gather ug;
6905 struct list_head *next;
6907 uncharge_gather_clear(&ug);
6910 * Note that the list can be a single page->lru; hence the
6911 * do-while loop instead of a simple list_for_each_entry().
6913 next = page_list->next;
6917 page = list_entry(next, struct page, lru);
6918 next = page->lru.next;
6920 uncharge_page(page, &ug);
6921 } while (next != page_list);
6924 uncharge_batch(&ug);
6928 * mem_cgroup_uncharge - uncharge a page
6929 * @page: page to uncharge
6931 * Uncharge a page previously charged with mem_cgroup_charge().
6933 void mem_cgroup_uncharge(struct page *page)
6935 struct uncharge_gather ug;
6937 if (mem_cgroup_disabled())
6940 /* Don't touch page->lru of any random page, pre-check: */
6941 if (!page->mem_cgroup)
6944 uncharge_gather_clear(&ug);
6945 uncharge_page(page, &ug);
6946 uncharge_batch(&ug);
6950 * mem_cgroup_uncharge_list - uncharge a list of page
6951 * @page_list: list of pages to uncharge
6953 * Uncharge a list of pages previously charged with
6954 * mem_cgroup_charge().
6956 void mem_cgroup_uncharge_list(struct list_head *page_list)
6958 if (mem_cgroup_disabled())
6961 if (!list_empty(page_list))
6962 uncharge_list(page_list);
6966 * mem_cgroup_migrate - charge a page's replacement
6967 * @oldpage: currently circulating page
6968 * @newpage: replacement page
6970 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6971 * be uncharged upon free.
6973 * Both pages must be locked, @newpage->mapping must be set up.
6975 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6977 struct mem_cgroup *memcg;
6978 unsigned int nr_pages;
6979 unsigned long flags;
6981 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6982 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6983 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6984 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6987 if (mem_cgroup_disabled())
6990 /* Page cache replacement: new page already charged? */
6991 if (newpage->mem_cgroup)
6994 /* Swapcache readahead pages can get replaced before being charged */
6995 memcg = oldpage->mem_cgroup;
6999 /* Force-charge the new page. The old one will be freed soon */
7000 nr_pages = hpage_nr_pages(newpage);
7002 page_counter_charge(&memcg->memory, nr_pages);
7003 if (do_memsw_account())
7004 page_counter_charge(&memcg->memsw, nr_pages);
7006 css_get(&memcg->css);
7007 commit_charge(newpage, memcg);
7009 local_irq_save(flags);
7010 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
7011 memcg_check_events(memcg, newpage);
7012 local_irq_restore(flags);
7015 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
7016 EXPORT_SYMBOL(memcg_sockets_enabled_key);
7018 void mem_cgroup_sk_alloc(struct sock *sk)
7020 struct mem_cgroup *memcg;
7022 if (!mem_cgroup_sockets_enabled)
7025 /* Do not associate the sock with unrelated interrupted task's memcg. */
7030 memcg = mem_cgroup_from_task(current);
7031 if (memcg == root_mem_cgroup)
7033 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
7035 if (css_tryget(&memcg->css))
7036 sk->sk_memcg = memcg;
7041 void mem_cgroup_sk_free(struct sock *sk)
7044 css_put(&sk->sk_memcg->css);
7048 * mem_cgroup_charge_skmem - charge socket memory
7049 * @memcg: memcg to charge
7050 * @nr_pages: number of pages to charge
7052 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7053 * @memcg's configured limit, %false if the charge had to be forced.
7055 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7057 gfp_t gfp_mask = GFP_KERNEL;
7059 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7060 struct page_counter *fail;
7062 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7063 memcg->tcpmem_pressure = 0;
7066 page_counter_charge(&memcg->tcpmem, nr_pages);
7067 memcg->tcpmem_pressure = 1;
7071 /* Don't block in the packet receive path */
7073 gfp_mask = GFP_NOWAIT;
7075 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7077 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
7080 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
7085 * mem_cgroup_uncharge_skmem - uncharge socket memory
7086 * @memcg: memcg to uncharge
7087 * @nr_pages: number of pages to uncharge
7089 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7091 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7092 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7096 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7098 refill_stock(memcg, nr_pages);
7101 static int __init cgroup_memory(char *s)
7105 while ((token = strsep(&s, ",")) != NULL) {
7108 if (!strcmp(token, "nosocket"))
7109 cgroup_memory_nosocket = true;
7110 if (!strcmp(token, "nokmem"))
7111 cgroup_memory_nokmem = true;
7115 __setup("cgroup.memory=", cgroup_memory);
7118 * subsys_initcall() for memory controller.
7120 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7121 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7122 * basically everything that doesn't depend on a specific mem_cgroup structure
7123 * should be initialized from here.
7125 static int __init mem_cgroup_init(void)
7129 #ifdef CONFIG_MEMCG_KMEM
7131 * Kmem cache creation is mostly done with the slab_mutex held,
7132 * so use a workqueue with limited concurrency to avoid stalling
7133 * all worker threads in case lots of cgroups are created and
7134 * destroyed simultaneously.
7136 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
7137 BUG_ON(!memcg_kmem_cache_wq);
7140 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7141 memcg_hotplug_cpu_dead);
7143 for_each_possible_cpu(cpu)
7144 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7147 for_each_node(node) {
7148 struct mem_cgroup_tree_per_node *rtpn;
7150 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7151 node_online(node) ? node : NUMA_NO_NODE);
7153 rtpn->rb_root = RB_ROOT;
7154 rtpn->rb_rightmost = NULL;
7155 spin_lock_init(&rtpn->lock);
7156 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7161 subsys_initcall(mem_cgroup_init);
7163 #ifdef CONFIG_MEMCG_SWAP
7164 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7166 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7168 * The root cgroup cannot be destroyed, so it's refcount must
7171 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7175 memcg = parent_mem_cgroup(memcg);
7177 memcg = root_mem_cgroup;
7183 * mem_cgroup_swapout - transfer a memsw charge to swap
7184 * @page: page whose memsw charge to transfer
7185 * @entry: swap entry to move the charge to
7187 * Transfer the memsw charge of @page to @entry.
7189 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7191 struct mem_cgroup *memcg, *swap_memcg;
7192 unsigned int nr_entries;
7193 unsigned short oldid;
7195 VM_BUG_ON_PAGE(PageLRU(page), page);
7196 VM_BUG_ON_PAGE(page_count(page), page);
7198 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7201 memcg = page->mem_cgroup;
7203 /* Readahead page, never charged */
7208 * In case the memcg owning these pages has been offlined and doesn't
7209 * have an ID allocated to it anymore, charge the closest online
7210 * ancestor for the swap instead and transfer the memory+swap charge.
7212 swap_memcg = mem_cgroup_id_get_online(memcg);
7213 nr_entries = hpage_nr_pages(page);
7214 /* Get references for the tail pages, too */
7216 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7217 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7219 VM_BUG_ON_PAGE(oldid, page);
7220 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7222 page->mem_cgroup = NULL;
7224 if (!mem_cgroup_is_root(memcg))
7225 page_counter_uncharge(&memcg->memory, nr_entries);
7227 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7228 if (!mem_cgroup_is_root(swap_memcg))
7229 page_counter_charge(&swap_memcg->memsw, nr_entries);
7230 page_counter_uncharge(&memcg->memsw, nr_entries);
7234 * Interrupts should be disabled here because the caller holds the
7235 * i_pages lock which is taken with interrupts-off. It is
7236 * important here to have the interrupts disabled because it is the
7237 * only synchronisation we have for updating the per-CPU variables.
7239 VM_BUG_ON(!irqs_disabled());
7240 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7241 memcg_check_events(memcg, page);
7243 css_put(&memcg->css);
7247 * mem_cgroup_try_charge_swap - try charging swap space for a page
7248 * @page: page being added to swap
7249 * @entry: swap entry to charge
7251 * Try to charge @page's memcg for the swap space at @entry.
7253 * Returns 0 on success, -ENOMEM on failure.
7255 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7257 unsigned int nr_pages = hpage_nr_pages(page);
7258 struct page_counter *counter;
7259 struct mem_cgroup *memcg;
7260 unsigned short oldid;
7262 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7265 memcg = page->mem_cgroup;
7267 /* Readahead page, never charged */
7272 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7276 memcg = mem_cgroup_id_get_online(memcg);
7278 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7279 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7280 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7281 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7282 mem_cgroup_id_put(memcg);
7286 /* Get references for the tail pages, too */
7288 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7289 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7290 VM_BUG_ON_PAGE(oldid, page);
7291 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7297 * mem_cgroup_uncharge_swap - uncharge swap space
7298 * @entry: swap entry to uncharge
7299 * @nr_pages: the amount of swap space to uncharge
7301 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7303 struct mem_cgroup *memcg;
7306 id = swap_cgroup_record(entry, 0, nr_pages);
7308 memcg = mem_cgroup_from_id(id);
7310 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7311 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7312 page_counter_uncharge(&memcg->swap, nr_pages);
7314 page_counter_uncharge(&memcg->memsw, nr_pages);
7316 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7317 mem_cgroup_id_put_many(memcg, nr_pages);
7322 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7324 long nr_swap_pages = get_nr_swap_pages();
7326 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7327 return nr_swap_pages;
7328 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7329 nr_swap_pages = min_t(long, nr_swap_pages,
7330 READ_ONCE(memcg->swap.max) -
7331 page_counter_read(&memcg->swap));
7332 return nr_swap_pages;
7335 bool mem_cgroup_swap_full(struct page *page)
7337 struct mem_cgroup *memcg;
7339 VM_BUG_ON_PAGE(!PageLocked(page), page);
7343 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7346 memcg = page->mem_cgroup;
7350 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7351 unsigned long usage = page_counter_read(&memcg->swap);
7353 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7354 usage * 2 >= READ_ONCE(memcg->swap.max))
7361 static int __init setup_swap_account(char *s)
7363 if (!strcmp(s, "1"))
7364 cgroup_memory_noswap = 0;
7365 else if (!strcmp(s, "0"))
7366 cgroup_memory_noswap = 1;
7369 __setup("swapaccount=", setup_swap_account);
7371 static u64 swap_current_read(struct cgroup_subsys_state *css,
7374 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7376 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7379 static int swap_high_show(struct seq_file *m, void *v)
7381 return seq_puts_memcg_tunable(m,
7382 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7385 static ssize_t swap_high_write(struct kernfs_open_file *of,
7386 char *buf, size_t nbytes, loff_t off)
7388 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7392 buf = strstrip(buf);
7393 err = page_counter_memparse(buf, "max", &high);
7397 page_counter_set_high(&memcg->swap, high);
7402 static int swap_max_show(struct seq_file *m, void *v)
7404 return seq_puts_memcg_tunable(m,
7405 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7408 static ssize_t swap_max_write(struct kernfs_open_file *of,
7409 char *buf, size_t nbytes, loff_t off)
7411 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7415 buf = strstrip(buf);
7416 err = page_counter_memparse(buf, "max", &max);
7420 xchg(&memcg->swap.max, max);
7425 static int swap_events_show(struct seq_file *m, void *v)
7427 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7429 seq_printf(m, "high %lu\n",
7430 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7431 seq_printf(m, "max %lu\n",
7432 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7433 seq_printf(m, "fail %lu\n",
7434 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7439 static struct cftype swap_files[] = {
7441 .name = "swap.current",
7442 .flags = CFTYPE_NOT_ON_ROOT,
7443 .read_u64 = swap_current_read,
7446 .name = "swap.high",
7447 .flags = CFTYPE_NOT_ON_ROOT,
7448 .seq_show = swap_high_show,
7449 .write = swap_high_write,
7453 .flags = CFTYPE_NOT_ON_ROOT,
7454 .seq_show = swap_max_show,
7455 .write = swap_max_write,
7458 .name = "swap.events",
7459 .flags = CFTYPE_NOT_ON_ROOT,
7460 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7461 .seq_show = swap_events_show,
7466 static struct cftype memsw_files[] = {
7468 .name = "memsw.usage_in_bytes",
7469 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7470 .read_u64 = mem_cgroup_read_u64,
7473 .name = "memsw.max_usage_in_bytes",
7474 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7475 .write = mem_cgroup_reset,
7476 .read_u64 = mem_cgroup_read_u64,
7479 .name = "memsw.limit_in_bytes",
7480 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7481 .write = mem_cgroup_write,
7482 .read_u64 = mem_cgroup_read_u64,
7485 .name = "memsw.failcnt",
7486 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7487 .write = mem_cgroup_reset,
7488 .read_u64 = mem_cgroup_read_u64,
7490 { }, /* terminate */
7494 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7495 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7496 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7497 * boot parameter. This may result in premature OOPS inside
7498 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7500 static int __init mem_cgroup_swap_init(void)
7502 /* No memory control -> no swap control */
7503 if (mem_cgroup_disabled())
7504 cgroup_memory_noswap = true;
7506 if (cgroup_memory_noswap)
7509 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7510 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7514 core_initcall(mem_cgroup_swap_init);
7516 #endif /* CONFIG_MEMCG_SWAP */