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
24 * Per memcg lru locking
25 * Copyright (C) 2020 Alibaba, Inc, Alex Shi
28 #include <linux/page_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
31 #include <linux/pagewalk.h>
32 #include <linux/sched/mm.h>
33 #include <linux/shmem_fs.h>
34 #include <linux/hugetlb.h>
35 #include <linux/pagemap.h>
36 #include <linux/vm_event_item.h>
37 #include <linux/smp.h>
38 #include <linux/page-flags.h>
39 #include <linux/backing-dev.h>
40 #include <linux/bit_spinlock.h>
41 #include <linux/rcupdate.h>
42 #include <linux/limits.h>
43 #include <linux/export.h>
44 #include <linux/mutex.h>
45 #include <linux/rbtree.h>
46 #include <linux/slab.h>
47 #include <linux/swap.h>
48 #include <linux/swapops.h>
49 #include <linux/spinlock.h>
50 #include <linux/eventfd.h>
51 #include <linux/poll.h>
52 #include <linux/sort.h>
54 #include <linux/seq_file.h>
55 #include <linux/vmpressure.h>
56 #include <linux/mm_inline.h>
57 #include <linux/swap_cgroup.h>
58 #include <linux/cpu.h>
59 #include <linux/oom.h>
60 #include <linux/lockdep.h>
61 #include <linux/file.h>
62 #include <linux/tracehook.h>
63 #include <linux/psi.h>
64 #include <linux/seq_buf.h>
70 #include <linux/uaccess.h>
72 #include <trace/events/vmscan.h>
74 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
75 EXPORT_SYMBOL(memory_cgrp_subsys);
77 struct mem_cgroup *root_mem_cgroup __read_mostly;
79 /* Active memory cgroup to use from an interrupt context */
80 DEFINE_PER_CPU(struct mem_cgroup *, int_active_memcg);
81 EXPORT_PER_CPU_SYMBOL_GPL(int_active_memcg);
83 /* Socket memory accounting disabled? */
84 static bool cgroup_memory_nosocket __ro_after_init;
86 /* Kernel memory accounting disabled? */
87 static bool cgroup_memory_nokmem __ro_after_init;
89 /* Whether the swap controller is active */
90 #ifdef CONFIG_MEMCG_SWAP
91 bool cgroup_memory_noswap __ro_after_init;
93 #define cgroup_memory_noswap 1
96 #ifdef CONFIG_CGROUP_WRITEBACK
97 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
100 /* Whether legacy memory+swap accounting is active */
101 static bool do_memsw_account(void)
103 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_noswap;
106 #define THRESHOLDS_EVENTS_TARGET 128
107 #define SOFTLIMIT_EVENTS_TARGET 1024
110 * Cgroups above their limits are maintained in a RB-Tree, independent of
111 * their hierarchy representation
114 struct mem_cgroup_tree_per_node {
115 struct rb_root rb_root;
116 struct rb_node *rb_rightmost;
120 struct mem_cgroup_tree {
121 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
124 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
127 struct mem_cgroup_eventfd_list {
128 struct list_head list;
129 struct eventfd_ctx *eventfd;
133 * cgroup_event represents events which userspace want to receive.
135 struct mem_cgroup_event {
137 * memcg which the event belongs to.
139 struct mem_cgroup *memcg;
141 * eventfd to signal userspace about the event.
143 struct eventfd_ctx *eventfd;
145 * Each of these stored in a list by the cgroup.
147 struct list_head list;
149 * register_event() callback will be used to add new userspace
150 * waiter for changes related to this event. Use eventfd_signal()
151 * on eventfd to send notification to userspace.
153 int (*register_event)(struct mem_cgroup *memcg,
154 struct eventfd_ctx *eventfd, const char *args);
156 * unregister_event() callback will be called when userspace closes
157 * the eventfd or on cgroup removing. This callback must be set,
158 * if you want provide notification functionality.
160 void (*unregister_event)(struct mem_cgroup *memcg,
161 struct eventfd_ctx *eventfd);
163 * All fields below needed to unregister event when
164 * userspace closes eventfd.
167 wait_queue_head_t *wqh;
168 wait_queue_entry_t wait;
169 struct work_struct remove;
172 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
173 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
175 /* Stuffs for move charges at task migration. */
177 * Types of charges to be moved.
179 #define MOVE_ANON 0x1U
180 #define MOVE_FILE 0x2U
181 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
183 /* "mc" and its members are protected by cgroup_mutex */
184 static struct move_charge_struct {
185 spinlock_t lock; /* for from, to */
186 struct mm_struct *mm;
187 struct mem_cgroup *from;
188 struct mem_cgroup *to;
190 unsigned long precharge;
191 unsigned long moved_charge;
192 unsigned long moved_swap;
193 struct task_struct *moving_task; /* a task moving charges */
194 wait_queue_head_t waitq; /* a waitq for other context */
196 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
197 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
201 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
202 * limit reclaim to prevent infinite loops, if they ever occur.
204 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
205 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
207 /* for encoding cft->private value on file */
216 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
217 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
218 #define MEMFILE_ATTR(val) ((val) & 0xffff)
219 /* Used for OOM notifier */
220 #define OOM_CONTROL (0)
223 * Iteration constructs for visiting all cgroups (under a tree). If
224 * loops are exited prematurely (break), mem_cgroup_iter_break() must
225 * be used for reference counting.
227 #define for_each_mem_cgroup_tree(iter, root) \
228 for (iter = mem_cgroup_iter(root, NULL, NULL); \
230 iter = mem_cgroup_iter(root, iter, NULL))
232 #define for_each_mem_cgroup(iter) \
233 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
235 iter = mem_cgroup_iter(NULL, iter, NULL))
237 static inline bool task_is_dying(void)
239 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
240 (current->flags & PF_EXITING);
243 /* Some nice accessors for the vmpressure. */
244 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
247 memcg = root_mem_cgroup;
248 return &memcg->vmpressure;
251 struct mem_cgroup *vmpressure_to_memcg(struct vmpressure *vmpr)
253 return container_of(vmpr, struct mem_cgroup, vmpressure);
256 #ifdef CONFIG_MEMCG_KMEM
257 static DEFINE_SPINLOCK(objcg_lock);
259 bool mem_cgroup_kmem_disabled(void)
261 return cgroup_memory_nokmem;
264 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
265 unsigned int nr_pages);
267 static void obj_cgroup_release(struct percpu_ref *ref)
269 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
270 unsigned int nr_bytes;
271 unsigned int nr_pages;
275 * At this point all allocated objects are freed, and
276 * objcg->nr_charged_bytes can't have an arbitrary byte value.
277 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
279 * The following sequence can lead to it:
280 * 1) CPU0: objcg == stock->cached_objcg
281 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
282 * PAGE_SIZE bytes are charged
283 * 3) CPU1: a process from another memcg is allocating something,
284 * the stock if flushed,
285 * objcg->nr_charged_bytes = PAGE_SIZE - 92
286 * 5) CPU0: we do release this object,
287 * 92 bytes are added to stock->nr_bytes
288 * 6) CPU0: stock is flushed,
289 * 92 bytes are added to objcg->nr_charged_bytes
291 * In the result, nr_charged_bytes == PAGE_SIZE.
292 * This page will be uncharged in obj_cgroup_release().
294 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
295 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
296 nr_pages = nr_bytes >> PAGE_SHIFT;
299 obj_cgroup_uncharge_pages(objcg, nr_pages);
301 spin_lock_irqsave(&objcg_lock, flags);
302 list_del(&objcg->list);
303 spin_unlock_irqrestore(&objcg_lock, flags);
305 percpu_ref_exit(ref);
306 kfree_rcu(objcg, rcu);
309 static struct obj_cgroup *obj_cgroup_alloc(void)
311 struct obj_cgroup *objcg;
314 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
318 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
324 INIT_LIST_HEAD(&objcg->list);
328 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
329 struct mem_cgroup *parent)
331 struct obj_cgroup *objcg, *iter;
333 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
335 spin_lock_irq(&objcg_lock);
337 /* 1) Ready to reparent active objcg. */
338 list_add(&objcg->list, &memcg->objcg_list);
339 /* 2) Reparent active objcg and already reparented objcgs to parent. */
340 list_for_each_entry(iter, &memcg->objcg_list, list)
341 WRITE_ONCE(iter->memcg, parent);
342 /* 3) Move already reparented objcgs to the parent's list */
343 list_splice(&memcg->objcg_list, &parent->objcg_list);
345 spin_unlock_irq(&objcg_lock);
347 percpu_ref_kill(&objcg->refcnt);
351 * This will be used as a shrinker list's index.
352 * The main reason for not using cgroup id for this:
353 * this works better in sparse environments, where we have a lot of memcgs,
354 * but only a few kmem-limited. Or also, if we have, for instance, 200
355 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
356 * 200 entry array for that.
358 * The current size of the caches array is stored in memcg_nr_cache_ids. It
359 * will double each time we have to increase it.
361 static DEFINE_IDA(memcg_cache_ida);
362 int memcg_nr_cache_ids;
364 /* Protects memcg_nr_cache_ids */
365 static DECLARE_RWSEM(memcg_cache_ids_sem);
367 void memcg_get_cache_ids(void)
369 down_read(&memcg_cache_ids_sem);
372 void memcg_put_cache_ids(void)
374 up_read(&memcg_cache_ids_sem);
378 * MIN_SIZE is different than 1, because we would like to avoid going through
379 * the alloc/free process all the time. In a small machine, 4 kmem-limited
380 * cgroups is a reasonable guess. In the future, it could be a parameter or
381 * tunable, but that is strictly not necessary.
383 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
384 * this constant directly from cgroup, but it is understandable that this is
385 * better kept as an internal representation in cgroup.c. In any case, the
386 * cgrp_id space is not getting any smaller, and we don't have to necessarily
387 * increase ours as well if it increases.
389 #define MEMCG_CACHES_MIN_SIZE 4
390 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
393 * A lot of the calls to the cache allocation functions are expected to be
394 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
395 * conditional to this static branch, we'll have to allow modules that does
396 * kmem_cache_alloc and the such to see this symbol as well
398 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
399 EXPORT_SYMBOL(memcg_kmem_enabled_key);
403 * mem_cgroup_css_from_page - css of the memcg associated with a page
404 * @page: page of interest
406 * If memcg is bound to the default hierarchy, css of the memcg associated
407 * with @page is returned. The returned css remains associated with @page
408 * until it is released.
410 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
413 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
415 struct mem_cgroup *memcg;
417 memcg = page_memcg(page);
419 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
420 memcg = root_mem_cgroup;
426 * page_cgroup_ino - return inode number of the memcg a page is charged to
429 * Look up the closest online ancestor of the memory cgroup @page is charged to
430 * and return its inode number or 0 if @page is not charged to any cgroup. It
431 * is safe to call this function without holding a reference to @page.
433 * Note, this function is inherently racy, because there is nothing to prevent
434 * the cgroup inode from getting torn down and potentially reallocated a moment
435 * after page_cgroup_ino() returns, so it only should be used by callers that
436 * do not care (such as procfs interfaces).
438 ino_t page_cgroup_ino(struct page *page)
440 struct mem_cgroup *memcg;
441 unsigned long ino = 0;
444 memcg = page_memcg_check(page);
446 while (memcg && !(memcg->css.flags & CSS_ONLINE))
447 memcg = parent_mem_cgroup(memcg);
449 ino = cgroup_ino(memcg->css.cgroup);
454 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
455 struct mem_cgroup_tree_per_node *mctz,
456 unsigned long new_usage_in_excess)
458 struct rb_node **p = &mctz->rb_root.rb_node;
459 struct rb_node *parent = NULL;
460 struct mem_cgroup_per_node *mz_node;
461 bool rightmost = true;
466 mz->usage_in_excess = new_usage_in_excess;
467 if (!mz->usage_in_excess)
471 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
473 if (mz->usage_in_excess < mz_node->usage_in_excess) {
482 mctz->rb_rightmost = &mz->tree_node;
484 rb_link_node(&mz->tree_node, parent, p);
485 rb_insert_color(&mz->tree_node, &mctz->rb_root);
489 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
490 struct mem_cgroup_tree_per_node *mctz)
495 if (&mz->tree_node == mctz->rb_rightmost)
496 mctz->rb_rightmost = rb_prev(&mz->tree_node);
498 rb_erase(&mz->tree_node, &mctz->rb_root);
502 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
503 struct mem_cgroup_tree_per_node *mctz)
507 spin_lock_irqsave(&mctz->lock, flags);
508 __mem_cgroup_remove_exceeded(mz, mctz);
509 spin_unlock_irqrestore(&mctz->lock, flags);
512 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
514 unsigned long nr_pages = page_counter_read(&memcg->memory);
515 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
516 unsigned long excess = 0;
518 if (nr_pages > soft_limit)
519 excess = nr_pages - soft_limit;
524 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, int nid)
526 unsigned long excess;
527 struct mem_cgroup_per_node *mz;
528 struct mem_cgroup_tree_per_node *mctz;
530 mctz = soft_limit_tree.rb_tree_per_node[nid];
534 * Necessary to update all ancestors when hierarchy is used.
535 * because their event counter is not touched.
537 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
538 mz = memcg->nodeinfo[nid];
539 excess = soft_limit_excess(memcg);
541 * We have to update the tree if mz is on RB-tree or
542 * mem is over its softlimit.
544 if (excess || mz->on_tree) {
547 spin_lock_irqsave(&mctz->lock, flags);
548 /* if on-tree, remove it */
550 __mem_cgroup_remove_exceeded(mz, mctz);
552 * Insert again. mz->usage_in_excess will be updated.
553 * If excess is 0, no tree ops.
555 __mem_cgroup_insert_exceeded(mz, mctz, excess);
556 spin_unlock_irqrestore(&mctz->lock, flags);
561 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
563 struct mem_cgroup_tree_per_node *mctz;
564 struct mem_cgroup_per_node *mz;
568 mz = memcg->nodeinfo[nid];
569 mctz = soft_limit_tree.rb_tree_per_node[nid];
571 mem_cgroup_remove_exceeded(mz, mctz);
575 static struct mem_cgroup_per_node *
576 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
578 struct mem_cgroup_per_node *mz;
582 if (!mctz->rb_rightmost)
583 goto done; /* Nothing to reclaim from */
585 mz = rb_entry(mctz->rb_rightmost,
586 struct mem_cgroup_per_node, tree_node);
588 * Remove the node now but someone else can add it back,
589 * we will to add it back at the end of reclaim to its correct
590 * position in the tree.
592 __mem_cgroup_remove_exceeded(mz, mctz);
593 if (!soft_limit_excess(mz->memcg) ||
594 !css_tryget(&mz->memcg->css))
600 static struct mem_cgroup_per_node *
601 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
603 struct mem_cgroup_per_node *mz;
605 spin_lock_irq(&mctz->lock);
606 mz = __mem_cgroup_largest_soft_limit_node(mctz);
607 spin_unlock_irq(&mctz->lock);
612 * memcg and lruvec stats flushing
614 * Many codepaths leading to stats update or read are performance sensitive and
615 * adding stats flushing in such codepaths is not desirable. So, to optimize the
616 * flushing the kernel does:
618 * 1) Periodically and asynchronously flush the stats every 2 seconds to not let
619 * rstat update tree grow unbounded.
621 * 2) Flush the stats synchronously on reader side only when there are more than
622 * (MEMCG_CHARGE_BATCH * nr_cpus) update events. Though this optimization
623 * will let stats be out of sync by atmost (MEMCG_CHARGE_BATCH * nr_cpus) but
624 * only for 2 seconds due to (1).
626 static void flush_memcg_stats_dwork(struct work_struct *w);
627 static DECLARE_DEFERRABLE_WORK(stats_flush_dwork, flush_memcg_stats_dwork);
628 static DEFINE_SPINLOCK(stats_flush_lock);
629 static DEFINE_PER_CPU(unsigned int, stats_updates);
630 static atomic_t stats_flush_threshold = ATOMIC_INIT(0);
633 * Accessors to ensure that preemption is disabled on PREEMPT_RT because it can
634 * not rely on this as part of an acquired spinlock_t lock. These functions are
635 * never used in hardirq context on PREEMPT_RT and therefore disabling preemtion
638 static void memcg_stats_lock(void)
640 #ifdef CONFIG_PREEMPT_RT
643 VM_BUG_ON(!irqs_disabled());
647 static void __memcg_stats_lock(void)
649 #ifdef CONFIG_PREEMPT_RT
654 static void memcg_stats_unlock(void)
656 #ifdef CONFIG_PREEMPT_RT
661 static inline void memcg_rstat_updated(struct mem_cgroup *memcg, int val)
665 cgroup_rstat_updated(memcg->css.cgroup, smp_processor_id());
667 x = __this_cpu_add_return(stats_updates, abs(val));
668 if (x > MEMCG_CHARGE_BATCH) {
669 atomic_add(x / MEMCG_CHARGE_BATCH, &stats_flush_threshold);
670 __this_cpu_write(stats_updates, 0);
674 static void __mem_cgroup_flush_stats(void)
678 if (!spin_trylock_irqsave(&stats_flush_lock, flag))
681 cgroup_rstat_flush_irqsafe(root_mem_cgroup->css.cgroup);
682 atomic_set(&stats_flush_threshold, 0);
683 spin_unlock_irqrestore(&stats_flush_lock, flag);
686 void mem_cgroup_flush_stats(void)
688 if (atomic_read(&stats_flush_threshold) > num_online_cpus())
689 __mem_cgroup_flush_stats();
692 static void flush_memcg_stats_dwork(struct work_struct *w)
694 __mem_cgroup_flush_stats();
695 queue_delayed_work(system_unbound_wq, &stats_flush_dwork, 2UL*HZ);
699 * __mod_memcg_state - update cgroup memory statistics
700 * @memcg: the memory cgroup
701 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
702 * @val: delta to add to the counter, can be negative
704 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
706 if (mem_cgroup_disabled())
709 __this_cpu_add(memcg->vmstats_percpu->state[idx], val);
710 memcg_rstat_updated(memcg, val);
713 /* idx can be of type enum memcg_stat_item or node_stat_item. */
714 static unsigned long memcg_page_state_local(struct mem_cgroup *memcg, int idx)
719 for_each_possible_cpu(cpu)
720 x += per_cpu(memcg->vmstats_percpu->state[idx], cpu);
728 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
731 struct mem_cgroup_per_node *pn;
732 struct mem_cgroup *memcg;
734 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
738 * The caller from rmap relay on disabled preemption becase they never
739 * update their counter from in-interrupt context. For these two
740 * counters we check that the update is never performed from an
741 * interrupt context while other caller need to have disabled interrupt.
743 __memcg_stats_lock();
744 if (IS_ENABLED(CONFIG_DEBUG_VM) && !IS_ENABLED(CONFIG_PREEMPT_RT)) {
749 case NR_SHMEM_PMDMAPPED:
750 case NR_FILE_PMDMAPPED:
751 WARN_ON_ONCE(!in_task());
754 WARN_ON_ONCE(!irqs_disabled());
759 __this_cpu_add(memcg->vmstats_percpu->state[idx], val);
762 __this_cpu_add(pn->lruvec_stats_percpu->state[idx], val);
764 memcg_rstat_updated(memcg, val);
765 memcg_stats_unlock();
769 * __mod_lruvec_state - update lruvec memory statistics
770 * @lruvec: the lruvec
771 * @idx: the stat item
772 * @val: delta to add to the counter, can be negative
774 * The lruvec is the intersection of the NUMA node and a cgroup. This
775 * function updates the all three counters that are affected by a
776 * change of state at this level: per-node, per-cgroup, per-lruvec.
778 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
782 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
784 /* Update memcg and lruvec */
785 if (!mem_cgroup_disabled())
786 __mod_memcg_lruvec_state(lruvec, idx, val);
789 void __mod_lruvec_page_state(struct page *page, enum node_stat_item idx,
792 struct page *head = compound_head(page); /* rmap on tail pages */
793 struct mem_cgroup *memcg;
794 pg_data_t *pgdat = page_pgdat(page);
795 struct lruvec *lruvec;
798 memcg = page_memcg(head);
799 /* Untracked pages have no memcg, no lruvec. Update only the node */
802 __mod_node_page_state(pgdat, idx, val);
806 lruvec = mem_cgroup_lruvec(memcg, pgdat);
807 __mod_lruvec_state(lruvec, idx, val);
810 EXPORT_SYMBOL(__mod_lruvec_page_state);
812 void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val)
814 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
815 struct mem_cgroup *memcg;
816 struct lruvec *lruvec;
819 memcg = mem_cgroup_from_obj(p);
822 * Untracked pages have no memcg, no lruvec. Update only the
823 * node. If we reparent the slab objects to the root memcg,
824 * when we free the slab object, we need to update the per-memcg
825 * vmstats to keep it correct for the root memcg.
828 __mod_node_page_state(pgdat, idx, val);
830 lruvec = mem_cgroup_lruvec(memcg, pgdat);
831 __mod_lruvec_state(lruvec, idx, val);
837 * __count_memcg_events - account VM events in a cgroup
838 * @memcg: the memory cgroup
839 * @idx: the event item
840 * @count: the number of events that occurred
842 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
845 if (mem_cgroup_disabled())
849 __this_cpu_add(memcg->vmstats_percpu->events[idx], count);
850 memcg_rstat_updated(memcg, count);
851 memcg_stats_unlock();
854 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
856 return READ_ONCE(memcg->vmstats.events[event]);
859 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
864 for_each_possible_cpu(cpu)
865 x += per_cpu(memcg->vmstats_percpu->events[event], cpu);
869 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
872 /* pagein of a big page is an event. So, ignore page size */
874 __count_memcg_events(memcg, PGPGIN, 1);
876 __count_memcg_events(memcg, PGPGOUT, 1);
877 nr_pages = -nr_pages; /* for event */
880 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
883 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
884 enum mem_cgroup_events_target target)
886 unsigned long val, next;
888 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
889 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
890 /* from time_after() in jiffies.h */
891 if ((long)(next - val) < 0) {
893 case MEM_CGROUP_TARGET_THRESH:
894 next = val + THRESHOLDS_EVENTS_TARGET;
896 case MEM_CGROUP_TARGET_SOFTLIMIT:
897 next = val + SOFTLIMIT_EVENTS_TARGET;
902 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
909 * Check events in order.
912 static void memcg_check_events(struct mem_cgroup *memcg, int nid)
914 if (IS_ENABLED(CONFIG_PREEMPT_RT))
917 /* threshold event is triggered in finer grain than soft limit */
918 if (unlikely(mem_cgroup_event_ratelimit(memcg,
919 MEM_CGROUP_TARGET_THRESH))) {
922 do_softlimit = mem_cgroup_event_ratelimit(memcg,
923 MEM_CGROUP_TARGET_SOFTLIMIT);
924 mem_cgroup_threshold(memcg);
925 if (unlikely(do_softlimit))
926 mem_cgroup_update_tree(memcg, nid);
930 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
933 * mm_update_next_owner() may clear mm->owner to NULL
934 * if it races with swapoff, page migration, etc.
935 * So this can be called with p == NULL.
940 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
942 EXPORT_SYMBOL(mem_cgroup_from_task);
944 static __always_inline struct mem_cgroup *active_memcg(void)
947 return this_cpu_read(int_active_memcg);
949 return current->active_memcg;
953 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
954 * @mm: mm from which memcg should be extracted. It can be NULL.
956 * Obtain a reference on mm->memcg and returns it if successful. If mm
957 * is NULL, then the memcg is chosen as follows:
958 * 1) The active memcg, if set.
959 * 2) current->mm->memcg, if available
961 * If mem_cgroup is disabled, NULL is returned.
963 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
965 struct mem_cgroup *memcg;
967 if (mem_cgroup_disabled())
971 * Page cache insertions can happen without an
972 * actual mm context, e.g. during disk probing
973 * on boot, loopback IO, acct() writes etc.
975 * No need to css_get on root memcg as the reference
976 * counting is disabled on the root level in the
977 * cgroup core. See CSS_NO_REF.
980 memcg = active_memcg();
981 if (unlikely(memcg)) {
982 /* remote memcg must hold a ref */
983 css_get(&memcg->css);
988 return root_mem_cgroup;
993 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
994 if (unlikely(!memcg))
995 memcg = root_mem_cgroup;
996 } while (!css_tryget(&memcg->css));
1000 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1002 static __always_inline bool memcg_kmem_bypass(void)
1004 /* Allow remote memcg charging from any context. */
1005 if (unlikely(active_memcg()))
1008 /* Memcg to charge can't be determined. */
1009 if (!in_task() || !current->mm || (current->flags & PF_KTHREAD))
1016 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1017 * @root: hierarchy root
1018 * @prev: previously returned memcg, NULL on first invocation
1019 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1021 * Returns references to children of the hierarchy below @root, or
1022 * @root itself, or %NULL after a full round-trip.
1024 * Caller must pass the return value in @prev on subsequent
1025 * invocations for reference counting, or use mem_cgroup_iter_break()
1026 * to cancel a hierarchy walk before the round-trip is complete.
1028 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1029 * in the hierarchy among all concurrent reclaimers operating on the
1032 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1033 struct mem_cgroup *prev,
1034 struct mem_cgroup_reclaim_cookie *reclaim)
1036 struct mem_cgroup_reclaim_iter *iter;
1037 struct cgroup_subsys_state *css = NULL;
1038 struct mem_cgroup *memcg = NULL;
1039 struct mem_cgroup *pos = NULL;
1041 if (mem_cgroup_disabled())
1045 root = root_mem_cgroup;
1047 if (prev && !reclaim)
1053 struct mem_cgroup_per_node *mz;
1055 mz = root->nodeinfo[reclaim->pgdat->node_id];
1058 if (prev && reclaim->generation != iter->generation)
1062 pos = READ_ONCE(iter->position);
1063 if (!pos || css_tryget(&pos->css))
1066 * css reference reached zero, so iter->position will
1067 * be cleared by ->css_released. However, we should not
1068 * rely on this happening soon, because ->css_released
1069 * is called from a work queue, and by busy-waiting we
1070 * might block it. So we clear iter->position right
1073 (void)cmpxchg(&iter->position, pos, NULL);
1081 css = css_next_descendant_pre(css, &root->css);
1084 * Reclaimers share the hierarchy walk, and a
1085 * new one might jump in right at the end of
1086 * the hierarchy - make sure they see at least
1087 * one group and restart from the beginning.
1095 * Verify the css and acquire a reference. The root
1096 * is provided by the caller, so we know it's alive
1097 * and kicking, and don't take an extra reference.
1099 memcg = mem_cgroup_from_css(css);
1101 if (css == &root->css)
1104 if (css_tryget(css))
1112 * The position could have already been updated by a competing
1113 * thread, so check that the value hasn't changed since we read
1114 * it to avoid reclaiming from the same cgroup twice.
1116 (void)cmpxchg(&iter->position, pos, memcg);
1124 reclaim->generation = iter->generation;
1129 if (prev && prev != root)
1130 css_put(&prev->css);
1136 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1137 * @root: hierarchy root
1138 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1140 void mem_cgroup_iter_break(struct mem_cgroup *root,
1141 struct mem_cgroup *prev)
1144 root = root_mem_cgroup;
1145 if (prev && prev != root)
1146 css_put(&prev->css);
1149 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1150 struct mem_cgroup *dead_memcg)
1152 struct mem_cgroup_reclaim_iter *iter;
1153 struct mem_cgroup_per_node *mz;
1156 for_each_node(nid) {
1157 mz = from->nodeinfo[nid];
1159 cmpxchg(&iter->position, dead_memcg, NULL);
1163 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1165 struct mem_cgroup *memcg = dead_memcg;
1166 struct mem_cgroup *last;
1169 __invalidate_reclaim_iterators(memcg, dead_memcg);
1171 } while ((memcg = parent_mem_cgroup(memcg)));
1174 * When cgruop1 non-hierarchy mode is used,
1175 * parent_mem_cgroup() does not walk all the way up to the
1176 * cgroup root (root_mem_cgroup). So we have to handle
1177 * dead_memcg from cgroup root separately.
1179 if (last != root_mem_cgroup)
1180 __invalidate_reclaim_iterators(root_mem_cgroup,
1185 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1186 * @memcg: hierarchy root
1187 * @fn: function to call for each task
1188 * @arg: argument passed to @fn
1190 * This function iterates over tasks attached to @memcg or to any of its
1191 * descendants and calls @fn for each task. If @fn returns a non-zero
1192 * value, the function breaks the iteration loop and returns the value.
1193 * Otherwise, it will iterate over all tasks and return 0.
1195 * This function must not be called for the root memory cgroup.
1197 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1198 int (*fn)(struct task_struct *, void *), void *arg)
1200 struct mem_cgroup *iter;
1203 BUG_ON(memcg == root_mem_cgroup);
1205 for_each_mem_cgroup_tree(iter, memcg) {
1206 struct css_task_iter it;
1207 struct task_struct *task;
1209 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1210 while (!ret && (task = css_task_iter_next(&it)))
1211 ret = fn(task, arg);
1212 css_task_iter_end(&it);
1214 mem_cgroup_iter_break(memcg, iter);
1221 #ifdef CONFIG_DEBUG_VM
1222 void lruvec_memcg_debug(struct lruvec *lruvec, struct folio *folio)
1224 struct mem_cgroup *memcg;
1226 if (mem_cgroup_disabled())
1229 memcg = folio_memcg(folio);
1232 VM_BUG_ON_FOLIO(lruvec_memcg(lruvec) != root_mem_cgroup, folio);
1234 VM_BUG_ON_FOLIO(lruvec_memcg(lruvec) != memcg, folio);
1239 * folio_lruvec_lock - Lock the lruvec for a folio.
1240 * @folio: Pointer to the folio.
1242 * These functions are safe to use under any of the following conditions:
1244 * - folio_test_lru false
1245 * - folio_memcg_lock()
1246 * - folio frozen (refcount of 0)
1248 * Return: The lruvec this folio is on with its lock held.
1250 struct lruvec *folio_lruvec_lock(struct folio *folio)
1252 struct lruvec *lruvec = folio_lruvec(folio);
1254 spin_lock(&lruvec->lru_lock);
1255 lruvec_memcg_debug(lruvec, folio);
1261 * folio_lruvec_lock_irq - Lock the lruvec for a folio.
1262 * @folio: Pointer to the folio.
1264 * These functions are safe to use under any of the following conditions:
1266 * - folio_test_lru false
1267 * - folio_memcg_lock()
1268 * - folio frozen (refcount of 0)
1270 * Return: The lruvec this folio is on with its lock held and interrupts
1273 struct lruvec *folio_lruvec_lock_irq(struct folio *folio)
1275 struct lruvec *lruvec = folio_lruvec(folio);
1277 spin_lock_irq(&lruvec->lru_lock);
1278 lruvec_memcg_debug(lruvec, folio);
1284 * folio_lruvec_lock_irqsave - Lock the lruvec for a folio.
1285 * @folio: Pointer to the folio.
1286 * @flags: Pointer to irqsave flags.
1288 * These functions are safe to use under any of the following conditions:
1290 * - folio_test_lru false
1291 * - folio_memcg_lock()
1292 * - folio frozen (refcount of 0)
1294 * Return: The lruvec this folio is on with its lock held and interrupts
1297 struct lruvec *folio_lruvec_lock_irqsave(struct folio *folio,
1298 unsigned long *flags)
1300 struct lruvec *lruvec = folio_lruvec(folio);
1302 spin_lock_irqsave(&lruvec->lru_lock, *flags);
1303 lruvec_memcg_debug(lruvec, folio);
1309 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1310 * @lruvec: mem_cgroup per zone lru vector
1311 * @lru: index of lru list the page is sitting on
1312 * @zid: zone id of the accounted pages
1313 * @nr_pages: positive when adding or negative when removing
1315 * This function must be called under lru_lock, just before a page is added
1316 * to or just after a page is removed from an lru list (that ordering being
1317 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1319 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1320 int zid, int nr_pages)
1322 struct mem_cgroup_per_node *mz;
1323 unsigned long *lru_size;
1326 if (mem_cgroup_disabled())
1329 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1330 lru_size = &mz->lru_zone_size[zid][lru];
1333 *lru_size += nr_pages;
1336 if (WARN_ONCE(size < 0,
1337 "%s(%p, %d, %d): lru_size %ld\n",
1338 __func__, lruvec, lru, nr_pages, size)) {
1344 *lru_size += nr_pages;
1348 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1349 * @memcg: the memory cgroup
1351 * Returns the maximum amount of memory @mem can be charged with, in
1354 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1356 unsigned long margin = 0;
1357 unsigned long count;
1358 unsigned long limit;
1360 count = page_counter_read(&memcg->memory);
1361 limit = READ_ONCE(memcg->memory.max);
1363 margin = limit - count;
1365 if (do_memsw_account()) {
1366 count = page_counter_read(&memcg->memsw);
1367 limit = READ_ONCE(memcg->memsw.max);
1369 margin = min(margin, limit - count);
1378 * A routine for checking "mem" is under move_account() or not.
1380 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1381 * moving cgroups. This is for waiting at high-memory pressure
1384 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1386 struct mem_cgroup *from;
1387 struct mem_cgroup *to;
1390 * Unlike task_move routines, we access mc.to, mc.from not under
1391 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1393 spin_lock(&mc.lock);
1399 ret = mem_cgroup_is_descendant(from, memcg) ||
1400 mem_cgroup_is_descendant(to, memcg);
1402 spin_unlock(&mc.lock);
1406 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1408 if (mc.moving_task && current != mc.moving_task) {
1409 if (mem_cgroup_under_move(memcg)) {
1411 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1412 /* moving charge context might have finished. */
1415 finish_wait(&mc.waitq, &wait);
1422 struct memory_stat {
1427 static const struct memory_stat memory_stats[] = {
1428 { "anon", NR_ANON_MAPPED },
1429 { "file", NR_FILE_PAGES },
1430 { "kernel", MEMCG_KMEM },
1431 { "kernel_stack", NR_KERNEL_STACK_KB },
1432 { "pagetables", NR_PAGETABLE },
1433 { "percpu", MEMCG_PERCPU_B },
1434 { "sock", MEMCG_SOCK },
1435 { "vmalloc", MEMCG_VMALLOC },
1436 { "shmem", NR_SHMEM },
1437 { "file_mapped", NR_FILE_MAPPED },
1438 { "file_dirty", NR_FILE_DIRTY },
1439 { "file_writeback", NR_WRITEBACK },
1441 { "swapcached", NR_SWAPCACHE },
1443 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1444 { "anon_thp", NR_ANON_THPS },
1445 { "file_thp", NR_FILE_THPS },
1446 { "shmem_thp", NR_SHMEM_THPS },
1448 { "inactive_anon", NR_INACTIVE_ANON },
1449 { "active_anon", NR_ACTIVE_ANON },
1450 { "inactive_file", NR_INACTIVE_FILE },
1451 { "active_file", NR_ACTIVE_FILE },
1452 { "unevictable", NR_UNEVICTABLE },
1453 { "slab_reclaimable", NR_SLAB_RECLAIMABLE_B },
1454 { "slab_unreclaimable", NR_SLAB_UNRECLAIMABLE_B },
1456 /* The memory events */
1457 { "workingset_refault_anon", WORKINGSET_REFAULT_ANON },
1458 { "workingset_refault_file", WORKINGSET_REFAULT_FILE },
1459 { "workingset_activate_anon", WORKINGSET_ACTIVATE_ANON },
1460 { "workingset_activate_file", WORKINGSET_ACTIVATE_FILE },
1461 { "workingset_restore_anon", WORKINGSET_RESTORE_ANON },
1462 { "workingset_restore_file", WORKINGSET_RESTORE_FILE },
1463 { "workingset_nodereclaim", WORKINGSET_NODERECLAIM },
1466 /* Translate stat items to the correct unit for memory.stat output */
1467 static int memcg_page_state_unit(int item)
1470 case MEMCG_PERCPU_B:
1471 case NR_SLAB_RECLAIMABLE_B:
1472 case NR_SLAB_UNRECLAIMABLE_B:
1473 case WORKINGSET_REFAULT_ANON:
1474 case WORKINGSET_REFAULT_FILE:
1475 case WORKINGSET_ACTIVATE_ANON:
1476 case WORKINGSET_ACTIVATE_FILE:
1477 case WORKINGSET_RESTORE_ANON:
1478 case WORKINGSET_RESTORE_FILE:
1479 case WORKINGSET_NODERECLAIM:
1481 case NR_KERNEL_STACK_KB:
1488 static inline unsigned long memcg_page_state_output(struct mem_cgroup *memcg,
1491 return memcg_page_state(memcg, item) * memcg_page_state_unit(item);
1494 static char *memory_stat_format(struct mem_cgroup *memcg)
1499 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1504 * Provide statistics on the state of the memory subsystem as
1505 * well as cumulative event counters that show past behavior.
1507 * This list is ordered following a combination of these gradients:
1508 * 1) generic big picture -> specifics and details
1509 * 2) reflecting userspace activity -> reflecting kernel heuristics
1511 * Current memory state:
1513 mem_cgroup_flush_stats();
1515 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1518 size = memcg_page_state_output(memcg, memory_stats[i].idx);
1519 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1521 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1522 size += memcg_page_state_output(memcg,
1523 NR_SLAB_RECLAIMABLE_B);
1524 seq_buf_printf(&s, "slab %llu\n", size);
1528 /* Accumulated memory events */
1530 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1531 memcg_events(memcg, PGFAULT));
1532 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1533 memcg_events(memcg, PGMAJFAULT));
1534 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1535 memcg_events(memcg, PGREFILL));
1536 seq_buf_printf(&s, "pgscan %lu\n",
1537 memcg_events(memcg, PGSCAN_KSWAPD) +
1538 memcg_events(memcg, PGSCAN_DIRECT));
1539 seq_buf_printf(&s, "pgsteal %lu\n",
1540 memcg_events(memcg, PGSTEAL_KSWAPD) +
1541 memcg_events(memcg, PGSTEAL_DIRECT));
1542 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1543 memcg_events(memcg, PGACTIVATE));
1544 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1545 memcg_events(memcg, PGDEACTIVATE));
1546 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1547 memcg_events(memcg, PGLAZYFREE));
1548 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1549 memcg_events(memcg, PGLAZYFREED));
1551 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1552 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1553 memcg_events(memcg, THP_FAULT_ALLOC));
1554 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1555 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1556 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1558 /* The above should easily fit into one page */
1559 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1564 #define K(x) ((x) << (PAGE_SHIFT-10))
1566 * mem_cgroup_print_oom_context: Print OOM information relevant to
1567 * memory controller.
1568 * @memcg: The memory cgroup that went over limit
1569 * @p: Task that is going to be killed
1571 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1574 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1579 pr_cont(",oom_memcg=");
1580 pr_cont_cgroup_path(memcg->css.cgroup);
1582 pr_cont(",global_oom");
1584 pr_cont(",task_memcg=");
1585 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1591 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1592 * memory controller.
1593 * @memcg: The memory cgroup that went over limit
1595 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1599 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1600 K((u64)page_counter_read(&memcg->memory)),
1601 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1602 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1603 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1604 K((u64)page_counter_read(&memcg->swap)),
1605 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1607 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1608 K((u64)page_counter_read(&memcg->memsw)),
1609 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1610 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1611 K((u64)page_counter_read(&memcg->kmem)),
1612 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1615 pr_info("Memory cgroup stats for ");
1616 pr_cont_cgroup_path(memcg->css.cgroup);
1618 buf = memory_stat_format(memcg);
1626 * Return the memory (and swap, if configured) limit for a memcg.
1628 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1630 unsigned long max = READ_ONCE(memcg->memory.max);
1632 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1633 if (mem_cgroup_swappiness(memcg))
1634 max += min(READ_ONCE(memcg->swap.max),
1635 (unsigned long)total_swap_pages);
1637 if (mem_cgroup_swappiness(memcg)) {
1638 /* Calculate swap excess capacity from memsw limit */
1639 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1641 max += min(swap, (unsigned long)total_swap_pages);
1647 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1649 return page_counter_read(&memcg->memory);
1652 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1655 struct oom_control oc = {
1659 .gfp_mask = gfp_mask,
1664 if (mutex_lock_killable(&oom_lock))
1667 if (mem_cgroup_margin(memcg) >= (1 << order))
1671 * A few threads which were not waiting at mutex_lock_killable() can
1672 * fail to bail out. Therefore, check again after holding oom_lock.
1674 ret = task_is_dying() || out_of_memory(&oc);
1677 mutex_unlock(&oom_lock);
1681 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1684 unsigned long *total_scanned)
1686 struct mem_cgroup *victim = NULL;
1689 unsigned long excess;
1690 unsigned long nr_scanned;
1691 struct mem_cgroup_reclaim_cookie reclaim = {
1695 excess = soft_limit_excess(root_memcg);
1698 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1703 * If we have not been able to reclaim
1704 * anything, it might because there are
1705 * no reclaimable pages under this hierarchy
1710 * We want to do more targeted reclaim.
1711 * excess >> 2 is not to excessive so as to
1712 * reclaim too much, nor too less that we keep
1713 * coming back to reclaim from this cgroup
1715 if (total >= (excess >> 2) ||
1716 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1721 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1722 pgdat, &nr_scanned);
1723 *total_scanned += nr_scanned;
1724 if (!soft_limit_excess(root_memcg))
1727 mem_cgroup_iter_break(root_memcg, victim);
1731 #ifdef CONFIG_LOCKDEP
1732 static struct lockdep_map memcg_oom_lock_dep_map = {
1733 .name = "memcg_oom_lock",
1737 static DEFINE_SPINLOCK(memcg_oom_lock);
1740 * Check OOM-Killer is already running under our hierarchy.
1741 * If someone is running, return false.
1743 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1745 struct mem_cgroup *iter, *failed = NULL;
1747 spin_lock(&memcg_oom_lock);
1749 for_each_mem_cgroup_tree(iter, memcg) {
1750 if (iter->oom_lock) {
1752 * this subtree of our hierarchy is already locked
1753 * so we cannot give a lock.
1756 mem_cgroup_iter_break(memcg, iter);
1759 iter->oom_lock = true;
1764 * OK, we failed to lock the whole subtree so we have
1765 * to clean up what we set up to the failing subtree
1767 for_each_mem_cgroup_tree(iter, memcg) {
1768 if (iter == failed) {
1769 mem_cgroup_iter_break(memcg, iter);
1772 iter->oom_lock = false;
1775 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1777 spin_unlock(&memcg_oom_lock);
1782 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1784 struct mem_cgroup *iter;
1786 spin_lock(&memcg_oom_lock);
1787 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1788 for_each_mem_cgroup_tree(iter, memcg)
1789 iter->oom_lock = false;
1790 spin_unlock(&memcg_oom_lock);
1793 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1795 struct mem_cgroup *iter;
1797 spin_lock(&memcg_oom_lock);
1798 for_each_mem_cgroup_tree(iter, memcg)
1800 spin_unlock(&memcg_oom_lock);
1803 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1805 struct mem_cgroup *iter;
1808 * Be careful about under_oom underflows because a child memcg
1809 * could have been added after mem_cgroup_mark_under_oom.
1811 spin_lock(&memcg_oom_lock);
1812 for_each_mem_cgroup_tree(iter, memcg)
1813 if (iter->under_oom > 0)
1815 spin_unlock(&memcg_oom_lock);
1818 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1820 struct oom_wait_info {
1821 struct mem_cgroup *memcg;
1822 wait_queue_entry_t wait;
1825 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1826 unsigned mode, int sync, void *arg)
1828 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1829 struct mem_cgroup *oom_wait_memcg;
1830 struct oom_wait_info *oom_wait_info;
1832 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1833 oom_wait_memcg = oom_wait_info->memcg;
1835 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1836 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1838 return autoremove_wake_function(wait, mode, sync, arg);
1841 static void memcg_oom_recover(struct mem_cgroup *memcg)
1844 * For the following lockless ->under_oom test, the only required
1845 * guarantee is that it must see the state asserted by an OOM when
1846 * this function is called as a result of userland actions
1847 * triggered by the notification of the OOM. This is trivially
1848 * achieved by invoking mem_cgroup_mark_under_oom() before
1849 * triggering notification.
1851 if (memcg && memcg->under_oom)
1852 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1856 * Returns true if successfully killed one or more processes. Though in some
1857 * corner cases it can return true even without killing any process.
1859 static bool mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1863 if (order > PAGE_ALLOC_COSTLY_ORDER)
1866 memcg_memory_event(memcg, MEMCG_OOM);
1869 * We are in the middle of the charge context here, so we
1870 * don't want to block when potentially sitting on a callstack
1871 * that holds all kinds of filesystem and mm locks.
1873 * cgroup1 allows disabling the OOM killer and waiting for outside
1874 * handling until the charge can succeed; remember the context and put
1875 * the task to sleep at the end of the page fault when all locks are
1878 * On the other hand, in-kernel OOM killer allows for an async victim
1879 * memory reclaim (oom_reaper) and that means that we are not solely
1880 * relying on the oom victim to make a forward progress and we can
1881 * invoke the oom killer here.
1883 * Please note that mem_cgroup_out_of_memory might fail to find a
1884 * victim and then we have to bail out from the charge path.
1886 if (memcg->oom_kill_disable) {
1887 if (current->in_user_fault) {
1888 css_get(&memcg->css);
1889 current->memcg_in_oom = memcg;
1890 current->memcg_oom_gfp_mask = mask;
1891 current->memcg_oom_order = order;
1896 mem_cgroup_mark_under_oom(memcg);
1898 locked = mem_cgroup_oom_trylock(memcg);
1901 mem_cgroup_oom_notify(memcg);
1903 mem_cgroup_unmark_under_oom(memcg);
1904 ret = mem_cgroup_out_of_memory(memcg, mask, order);
1907 mem_cgroup_oom_unlock(memcg);
1913 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1914 * @handle: actually kill/wait or just clean up the OOM state
1916 * This has to be called at the end of a page fault if the memcg OOM
1917 * handler was enabled.
1919 * Memcg supports userspace OOM handling where failed allocations must
1920 * sleep on a waitqueue until the userspace task resolves the
1921 * situation. Sleeping directly in the charge context with all kinds
1922 * of locks held is not a good idea, instead we remember an OOM state
1923 * in the task and mem_cgroup_oom_synchronize() has to be called at
1924 * the end of the page fault to complete the OOM handling.
1926 * Returns %true if an ongoing memcg OOM situation was detected and
1927 * completed, %false otherwise.
1929 bool mem_cgroup_oom_synchronize(bool handle)
1931 struct mem_cgroup *memcg = current->memcg_in_oom;
1932 struct oom_wait_info owait;
1935 /* OOM is global, do not handle */
1942 owait.memcg = memcg;
1943 owait.wait.flags = 0;
1944 owait.wait.func = memcg_oom_wake_function;
1945 owait.wait.private = current;
1946 INIT_LIST_HEAD(&owait.wait.entry);
1948 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1949 mem_cgroup_mark_under_oom(memcg);
1951 locked = mem_cgroup_oom_trylock(memcg);
1954 mem_cgroup_oom_notify(memcg);
1956 if (locked && !memcg->oom_kill_disable) {
1957 mem_cgroup_unmark_under_oom(memcg);
1958 finish_wait(&memcg_oom_waitq, &owait.wait);
1959 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1960 current->memcg_oom_order);
1963 mem_cgroup_unmark_under_oom(memcg);
1964 finish_wait(&memcg_oom_waitq, &owait.wait);
1968 mem_cgroup_oom_unlock(memcg);
1970 * There is no guarantee that an OOM-lock contender
1971 * sees the wakeups triggered by the OOM kill
1972 * uncharges. Wake any sleepers explicitly.
1974 memcg_oom_recover(memcg);
1977 current->memcg_in_oom = NULL;
1978 css_put(&memcg->css);
1983 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1984 * @victim: task to be killed by the OOM killer
1985 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1987 * Returns a pointer to a memory cgroup, which has to be cleaned up
1988 * by killing all belonging OOM-killable tasks.
1990 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1992 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1993 struct mem_cgroup *oom_domain)
1995 struct mem_cgroup *oom_group = NULL;
1996 struct mem_cgroup *memcg;
1998 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2002 oom_domain = root_mem_cgroup;
2006 memcg = mem_cgroup_from_task(victim);
2007 if (memcg == root_mem_cgroup)
2011 * If the victim task has been asynchronously moved to a different
2012 * memory cgroup, we might end up killing tasks outside oom_domain.
2013 * In this case it's better to ignore memory.group.oom.
2015 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2019 * Traverse the memory cgroup hierarchy from the victim task's
2020 * cgroup up to the OOMing cgroup (or root) to find the
2021 * highest-level memory cgroup with oom.group set.
2023 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2024 if (memcg->oom_group)
2027 if (memcg == oom_domain)
2032 css_get(&oom_group->css);
2039 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2041 pr_info("Tasks in ");
2042 pr_cont_cgroup_path(memcg->css.cgroup);
2043 pr_cont(" are going to be killed due to memory.oom.group set\n");
2047 * folio_memcg_lock - Bind a folio to its memcg.
2048 * @folio: The folio.
2050 * This function prevents unlocked LRU folios from being moved to
2053 * It ensures lifetime of the bound memcg. The caller is responsible
2054 * for the lifetime of the folio.
2056 void folio_memcg_lock(struct folio *folio)
2058 struct mem_cgroup *memcg;
2059 unsigned long flags;
2062 * The RCU lock is held throughout the transaction. The fast
2063 * path can get away without acquiring the memcg->move_lock
2064 * because page moving starts with an RCU grace period.
2068 if (mem_cgroup_disabled())
2071 memcg = folio_memcg(folio);
2072 if (unlikely(!memcg))
2075 #ifdef CONFIG_PROVE_LOCKING
2076 local_irq_save(flags);
2077 might_lock(&memcg->move_lock);
2078 local_irq_restore(flags);
2081 if (atomic_read(&memcg->moving_account) <= 0)
2084 spin_lock_irqsave(&memcg->move_lock, flags);
2085 if (memcg != folio_memcg(folio)) {
2086 spin_unlock_irqrestore(&memcg->move_lock, flags);
2091 * When charge migration first begins, we can have multiple
2092 * critical sections holding the fast-path RCU lock and one
2093 * holding the slowpath move_lock. Track the task who has the
2094 * move_lock for unlock_page_memcg().
2096 memcg->move_lock_task = current;
2097 memcg->move_lock_flags = flags;
2100 void lock_page_memcg(struct page *page)
2102 folio_memcg_lock(page_folio(page));
2105 static void __folio_memcg_unlock(struct mem_cgroup *memcg)
2107 if (memcg && memcg->move_lock_task == current) {
2108 unsigned long flags = memcg->move_lock_flags;
2110 memcg->move_lock_task = NULL;
2111 memcg->move_lock_flags = 0;
2113 spin_unlock_irqrestore(&memcg->move_lock, flags);
2120 * folio_memcg_unlock - Release the binding between a folio and its memcg.
2121 * @folio: The folio.
2123 * This releases the binding created by folio_memcg_lock(). This does
2124 * not change the accounting of this folio to its memcg, but it does
2125 * permit others to change it.
2127 void folio_memcg_unlock(struct folio *folio)
2129 __folio_memcg_unlock(folio_memcg(folio));
2132 void unlock_page_memcg(struct page *page)
2134 folio_memcg_unlock(page_folio(page));
2137 struct memcg_stock_pcp {
2138 local_lock_t stock_lock;
2139 struct mem_cgroup *cached; /* this never be root cgroup */
2140 unsigned int nr_pages;
2142 #ifdef CONFIG_MEMCG_KMEM
2143 struct obj_cgroup *cached_objcg;
2144 struct pglist_data *cached_pgdat;
2145 unsigned int nr_bytes;
2146 int nr_slab_reclaimable_b;
2147 int nr_slab_unreclaimable_b;
2150 struct work_struct work;
2151 unsigned long flags;
2152 #define FLUSHING_CACHED_CHARGE 0
2154 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock) = {
2155 .stock_lock = INIT_LOCAL_LOCK(stock_lock),
2157 static DEFINE_MUTEX(percpu_charge_mutex);
2159 #ifdef CONFIG_MEMCG_KMEM
2160 static struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock);
2161 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2162 struct mem_cgroup *root_memcg);
2163 static void memcg_account_kmem(struct mem_cgroup *memcg, int nr_pages);
2166 static inline struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock)
2170 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2171 struct mem_cgroup *root_memcg)
2175 static void memcg_account_kmem(struct mem_cgroup *memcg, int nr_pages)
2181 * consume_stock: Try to consume stocked charge on this cpu.
2182 * @memcg: memcg to consume from.
2183 * @nr_pages: how many pages to charge.
2185 * The charges will only happen if @memcg matches the current cpu's memcg
2186 * stock, and at least @nr_pages are available in that stock. Failure to
2187 * service an allocation will refill the stock.
2189 * returns true if successful, false otherwise.
2191 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2193 struct memcg_stock_pcp *stock;
2194 unsigned long flags;
2197 if (nr_pages > MEMCG_CHARGE_BATCH)
2200 local_lock_irqsave(&memcg_stock.stock_lock, flags);
2202 stock = this_cpu_ptr(&memcg_stock);
2203 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2204 stock->nr_pages -= nr_pages;
2208 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2214 * Returns stocks cached in percpu and reset cached information.
2216 static void drain_stock(struct memcg_stock_pcp *stock)
2218 struct mem_cgroup *old = stock->cached;
2223 if (stock->nr_pages) {
2224 page_counter_uncharge(&old->memory, stock->nr_pages);
2225 if (do_memsw_account())
2226 page_counter_uncharge(&old->memsw, stock->nr_pages);
2227 stock->nr_pages = 0;
2231 stock->cached = NULL;
2234 static void drain_local_stock(struct work_struct *dummy)
2236 struct memcg_stock_pcp *stock;
2237 struct obj_cgroup *old = NULL;
2238 unsigned long flags;
2241 * The only protection from cpu hotplug (memcg_hotplug_cpu_dead) vs.
2242 * drain_stock races is that we always operate on local CPU stock
2243 * here with IRQ disabled
2245 local_lock_irqsave(&memcg_stock.stock_lock, flags);
2247 stock = this_cpu_ptr(&memcg_stock);
2248 old = drain_obj_stock(stock);
2250 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2252 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2254 obj_cgroup_put(old);
2258 * Cache charges(val) to local per_cpu area.
2259 * This will be consumed by consume_stock() function, later.
2261 static void __refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2263 struct memcg_stock_pcp *stock;
2265 stock = this_cpu_ptr(&memcg_stock);
2266 if (stock->cached != memcg) { /* reset if necessary */
2268 css_get(&memcg->css);
2269 stock->cached = memcg;
2271 stock->nr_pages += nr_pages;
2273 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2277 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2279 unsigned long flags;
2281 local_lock_irqsave(&memcg_stock.stock_lock, flags);
2282 __refill_stock(memcg, nr_pages);
2283 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2287 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2288 * of the hierarchy under it.
2290 static void drain_all_stock(struct mem_cgroup *root_memcg)
2294 /* If someone's already draining, avoid adding running more workers. */
2295 if (!mutex_trylock(&percpu_charge_mutex))
2298 * Notify other cpus that system-wide "drain" is running
2299 * We do not care about races with the cpu hotplug because cpu down
2300 * as well as workers from this path always operate on the local
2301 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2304 curcpu = smp_processor_id();
2305 for_each_online_cpu(cpu) {
2306 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2307 struct mem_cgroup *memcg;
2311 memcg = stock->cached;
2312 if (memcg && stock->nr_pages &&
2313 mem_cgroup_is_descendant(memcg, root_memcg))
2315 else if (obj_stock_flush_required(stock, root_memcg))
2320 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2322 drain_local_stock(&stock->work);
2324 schedule_work_on(cpu, &stock->work);
2328 mutex_unlock(&percpu_charge_mutex);
2331 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2333 struct memcg_stock_pcp *stock;
2335 stock = &per_cpu(memcg_stock, cpu);
2341 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2342 unsigned int nr_pages,
2345 unsigned long nr_reclaimed = 0;
2348 unsigned long pflags;
2350 if (page_counter_read(&memcg->memory) <=
2351 READ_ONCE(memcg->memory.high))
2354 memcg_memory_event(memcg, MEMCG_HIGH);
2356 psi_memstall_enter(&pflags);
2357 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2359 psi_memstall_leave(&pflags);
2360 } while ((memcg = parent_mem_cgroup(memcg)) &&
2361 !mem_cgroup_is_root(memcg));
2363 return nr_reclaimed;
2366 static void high_work_func(struct work_struct *work)
2368 struct mem_cgroup *memcg;
2370 memcg = container_of(work, struct mem_cgroup, high_work);
2371 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2375 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2376 * enough to still cause a significant slowdown in most cases, while still
2377 * allowing diagnostics and tracing to proceed without becoming stuck.
2379 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2382 * When calculating the delay, we use these either side of the exponentiation to
2383 * maintain precision and scale to a reasonable number of jiffies (see the table
2386 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2387 * overage ratio to a delay.
2388 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2389 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2390 * to produce a reasonable delay curve.
2392 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2393 * reasonable delay curve compared to precision-adjusted overage, not
2394 * penalising heavily at first, but still making sure that growth beyond the
2395 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2396 * example, with a high of 100 megabytes:
2398 * +-------+------------------------+
2399 * | usage | time to allocate in ms |
2400 * +-------+------------------------+
2422 * +-------+------------------------+
2424 #define MEMCG_DELAY_PRECISION_SHIFT 20
2425 #define MEMCG_DELAY_SCALING_SHIFT 14
2427 static u64 calculate_overage(unsigned long usage, unsigned long high)
2435 * Prevent division by 0 in overage calculation by acting as if
2436 * it was a threshold of 1 page
2438 high = max(high, 1UL);
2440 overage = usage - high;
2441 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2442 return div64_u64(overage, high);
2445 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2447 u64 overage, max_overage = 0;
2450 overage = calculate_overage(page_counter_read(&memcg->memory),
2451 READ_ONCE(memcg->memory.high));
2452 max_overage = max(overage, max_overage);
2453 } while ((memcg = parent_mem_cgroup(memcg)) &&
2454 !mem_cgroup_is_root(memcg));
2459 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2461 u64 overage, max_overage = 0;
2464 overage = calculate_overage(page_counter_read(&memcg->swap),
2465 READ_ONCE(memcg->swap.high));
2467 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2468 max_overage = max(overage, max_overage);
2469 } while ((memcg = parent_mem_cgroup(memcg)) &&
2470 !mem_cgroup_is_root(memcg));
2476 * Get the number of jiffies that we should penalise a mischievous cgroup which
2477 * is exceeding its memory.high by checking both it and its ancestors.
2479 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2480 unsigned int nr_pages,
2483 unsigned long penalty_jiffies;
2489 * We use overage compared to memory.high to calculate the number of
2490 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2491 * fairly lenient on small overages, and increasingly harsh when the
2492 * memcg in question makes it clear that it has no intention of stopping
2493 * its crazy behaviour, so we exponentially increase the delay based on
2496 penalty_jiffies = max_overage * max_overage * HZ;
2497 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2498 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2501 * Factor in the task's own contribution to the overage, such that four
2502 * N-sized allocations are throttled approximately the same as one
2503 * 4N-sized allocation.
2505 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2506 * larger the current charge patch is than that.
2508 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2512 * Scheduled by try_charge() to be executed from the userland return path
2513 * and reclaims memory over the high limit.
2515 void mem_cgroup_handle_over_high(void)
2517 unsigned long penalty_jiffies;
2518 unsigned long pflags;
2519 unsigned long nr_reclaimed;
2520 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2521 int nr_retries = MAX_RECLAIM_RETRIES;
2522 struct mem_cgroup *memcg;
2523 bool in_retry = false;
2525 if (likely(!nr_pages))
2528 memcg = get_mem_cgroup_from_mm(current->mm);
2529 current->memcg_nr_pages_over_high = 0;
2533 * The allocating task should reclaim at least the batch size, but for
2534 * subsequent retries we only want to do what's necessary to prevent oom
2535 * or breaching resource isolation.
2537 * This is distinct from memory.max or page allocator behaviour because
2538 * memory.high is currently batched, whereas memory.max and the page
2539 * allocator run every time an allocation is made.
2541 nr_reclaimed = reclaim_high(memcg,
2542 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2546 * memory.high is breached and reclaim is unable to keep up. Throttle
2547 * allocators proactively to slow down excessive growth.
2549 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2550 mem_find_max_overage(memcg));
2552 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2553 swap_find_max_overage(memcg));
2556 * Clamp the max delay per usermode return so as to still keep the
2557 * application moving forwards and also permit diagnostics, albeit
2560 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2563 * Don't sleep if the amount of jiffies this memcg owes us is so low
2564 * that it's not even worth doing, in an attempt to be nice to those who
2565 * go only a small amount over their memory.high value and maybe haven't
2566 * been aggressively reclaimed enough yet.
2568 if (penalty_jiffies <= HZ / 100)
2572 * If reclaim is making forward progress but we're still over
2573 * memory.high, we want to encourage that rather than doing allocator
2576 if (nr_reclaimed || nr_retries--) {
2582 * If we exit early, we're guaranteed to die (since
2583 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2584 * need to account for any ill-begotten jiffies to pay them off later.
2586 psi_memstall_enter(&pflags);
2587 schedule_timeout_killable(penalty_jiffies);
2588 psi_memstall_leave(&pflags);
2591 css_put(&memcg->css);
2594 static int try_charge_memcg(struct mem_cgroup *memcg, gfp_t gfp_mask,
2595 unsigned int nr_pages)
2597 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2598 int nr_retries = MAX_RECLAIM_RETRIES;
2599 struct mem_cgroup *mem_over_limit;
2600 struct page_counter *counter;
2601 unsigned long nr_reclaimed;
2602 bool passed_oom = false;
2603 bool may_swap = true;
2604 bool drained = false;
2605 unsigned long pflags;
2608 if (consume_stock(memcg, nr_pages))
2611 if (!do_memsw_account() ||
2612 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2613 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2615 if (do_memsw_account())
2616 page_counter_uncharge(&memcg->memsw, batch);
2617 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2619 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2623 if (batch > nr_pages) {
2629 * Prevent unbounded recursion when reclaim operations need to
2630 * allocate memory. This might exceed the limits temporarily,
2631 * but we prefer facilitating memory reclaim and getting back
2632 * under the limit over triggering OOM kills in these cases.
2634 if (unlikely(current->flags & PF_MEMALLOC))
2637 if (unlikely(task_in_memcg_oom(current)))
2640 if (!gfpflags_allow_blocking(gfp_mask))
2643 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2645 psi_memstall_enter(&pflags);
2646 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2647 gfp_mask, may_swap);
2648 psi_memstall_leave(&pflags);
2650 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2654 drain_all_stock(mem_over_limit);
2659 if (gfp_mask & __GFP_NORETRY)
2662 * Even though the limit is exceeded at this point, reclaim
2663 * may have been able to free some pages. Retry the charge
2664 * before killing the task.
2666 * Only for regular pages, though: huge pages are rather
2667 * unlikely to succeed so close to the limit, and we fall back
2668 * to regular pages anyway in case of failure.
2670 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2673 * At task move, charge accounts can be doubly counted. So, it's
2674 * better to wait until the end of task_move if something is going on.
2676 if (mem_cgroup_wait_acct_move(mem_over_limit))
2682 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2685 /* Avoid endless loop for tasks bypassed by the oom killer */
2686 if (passed_oom && task_is_dying())
2690 * keep retrying as long as the memcg oom killer is able to make
2691 * a forward progress or bypass the charge if the oom killer
2692 * couldn't make any progress.
2694 if (mem_cgroup_oom(mem_over_limit, gfp_mask,
2695 get_order(nr_pages * PAGE_SIZE))) {
2697 nr_retries = MAX_RECLAIM_RETRIES;
2702 * Memcg doesn't have a dedicated reserve for atomic
2703 * allocations. But like the global atomic pool, we need to
2704 * put the burden of reclaim on regular allocation requests
2705 * and let these go through as privileged allocations.
2707 if (!(gfp_mask & (__GFP_NOFAIL | __GFP_HIGH)))
2711 * The allocation either can't fail or will lead to more memory
2712 * being freed very soon. Allow memory usage go over the limit
2713 * temporarily by force charging it.
2715 page_counter_charge(&memcg->memory, nr_pages);
2716 if (do_memsw_account())
2717 page_counter_charge(&memcg->memsw, nr_pages);
2722 if (batch > nr_pages)
2723 refill_stock(memcg, batch - nr_pages);
2726 * If the hierarchy is above the normal consumption range, schedule
2727 * reclaim on returning to userland. We can perform reclaim here
2728 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2729 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2730 * not recorded as it most likely matches current's and won't
2731 * change in the meantime. As high limit is checked again before
2732 * reclaim, the cost of mismatch is negligible.
2735 bool mem_high, swap_high;
2737 mem_high = page_counter_read(&memcg->memory) >
2738 READ_ONCE(memcg->memory.high);
2739 swap_high = page_counter_read(&memcg->swap) >
2740 READ_ONCE(memcg->swap.high);
2742 /* Don't bother a random interrupted task */
2745 schedule_work(&memcg->high_work);
2751 if (mem_high || swap_high) {
2753 * The allocating tasks in this cgroup will need to do
2754 * reclaim or be throttled to prevent further growth
2755 * of the memory or swap footprints.
2757 * Target some best-effort fairness between the tasks,
2758 * and distribute reclaim work and delay penalties
2759 * based on how much each task is actually allocating.
2761 current->memcg_nr_pages_over_high += batch;
2762 set_notify_resume(current);
2765 } while ((memcg = parent_mem_cgroup(memcg)));
2767 if (current->memcg_nr_pages_over_high > MEMCG_CHARGE_BATCH &&
2768 !(current->flags & PF_MEMALLOC) &&
2769 gfpflags_allow_blocking(gfp_mask)) {
2770 mem_cgroup_handle_over_high();
2775 static inline int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2776 unsigned int nr_pages)
2778 if (mem_cgroup_is_root(memcg))
2781 return try_charge_memcg(memcg, gfp_mask, nr_pages);
2784 static inline void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2786 if (mem_cgroup_is_root(memcg))
2789 page_counter_uncharge(&memcg->memory, nr_pages);
2790 if (do_memsw_account())
2791 page_counter_uncharge(&memcg->memsw, nr_pages);
2794 static void commit_charge(struct folio *folio, struct mem_cgroup *memcg)
2796 VM_BUG_ON_FOLIO(folio_memcg(folio), folio);
2798 * Any of the following ensures page's memcg stability:
2802 * - lock_page_memcg()
2803 * - exclusive reference
2805 folio->memcg_data = (unsigned long)memcg;
2808 #ifdef CONFIG_MEMCG_KMEM
2810 * The allocated objcg pointers array is not accounted directly.
2811 * Moreover, it should not come from DMA buffer and is not readily
2812 * reclaimable. So those GFP bits should be masked off.
2814 #define OBJCGS_CLEAR_MASK (__GFP_DMA | __GFP_RECLAIMABLE | __GFP_ACCOUNT)
2817 * mod_objcg_mlstate() may be called with irq enabled, so
2818 * mod_memcg_lruvec_state() should be used.
2820 static inline void mod_objcg_mlstate(struct obj_cgroup *objcg,
2821 struct pglist_data *pgdat,
2822 enum node_stat_item idx, int nr)
2824 struct mem_cgroup *memcg;
2825 struct lruvec *lruvec;
2828 memcg = obj_cgroup_memcg(objcg);
2829 lruvec = mem_cgroup_lruvec(memcg, pgdat);
2830 mod_memcg_lruvec_state(lruvec, idx, nr);
2834 int memcg_alloc_slab_cgroups(struct slab *slab, struct kmem_cache *s,
2835 gfp_t gfp, bool new_slab)
2837 unsigned int objects = objs_per_slab(s, slab);
2838 unsigned long memcg_data;
2841 gfp &= ~OBJCGS_CLEAR_MASK;
2842 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2847 memcg_data = (unsigned long) vec | MEMCG_DATA_OBJCGS;
2850 * If the slab is brand new and nobody can yet access its
2851 * memcg_data, no synchronization is required and memcg_data can
2852 * be simply assigned.
2854 slab->memcg_data = memcg_data;
2855 } else if (cmpxchg(&slab->memcg_data, 0, memcg_data)) {
2857 * If the slab is already in use, somebody can allocate and
2858 * assign obj_cgroups in parallel. In this case the existing
2859 * objcg vector should be reused.
2865 kmemleak_not_leak(vec);
2870 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2872 * A passed kernel object can be a slab object or a generic kernel page, so
2873 * different mechanisms for getting the memory cgroup pointer should be used.
2874 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2875 * can not know for sure how the kernel object is implemented.
2876 * mem_cgroup_from_obj() can be safely used in such cases.
2878 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2879 * cgroup_mutex, etc.
2881 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2883 struct folio *folio;
2885 if (mem_cgroup_disabled())
2888 folio = virt_to_folio(p);
2891 * Slab objects are accounted individually, not per-page.
2892 * Memcg membership data for each individual object is saved in
2895 if (folio_test_slab(folio)) {
2896 struct obj_cgroup **objcgs;
2900 slab = folio_slab(folio);
2901 objcgs = slab_objcgs(slab);
2905 off = obj_to_index(slab->slab_cache, slab, p);
2907 return obj_cgroup_memcg(objcgs[off]);
2913 * page_memcg_check() is used here, because in theory we can encounter
2914 * a folio where the slab flag has been cleared already, but
2915 * slab->memcg_data has not been freed yet
2916 * page_memcg_check(page) will guarantee that a proper memory
2917 * cgroup pointer or NULL will be returned.
2919 return page_memcg_check(folio_page(folio, 0));
2922 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2924 struct obj_cgroup *objcg = NULL;
2925 struct mem_cgroup *memcg;
2927 if (memcg_kmem_bypass())
2931 if (unlikely(active_memcg()))
2932 memcg = active_memcg();
2934 memcg = mem_cgroup_from_task(current);
2936 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2937 objcg = rcu_dereference(memcg->objcg);
2938 if (objcg && obj_cgroup_tryget(objcg))
2947 static int memcg_alloc_cache_id(void)
2952 id = ida_simple_get(&memcg_cache_ida,
2953 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2957 if (id < memcg_nr_cache_ids)
2961 * There's no space for the new id in memcg_caches arrays,
2962 * so we have to grow them.
2964 down_write(&memcg_cache_ids_sem);
2966 size = 2 * (id + 1);
2967 if (size < MEMCG_CACHES_MIN_SIZE)
2968 size = MEMCG_CACHES_MIN_SIZE;
2969 else if (size > MEMCG_CACHES_MAX_SIZE)
2970 size = MEMCG_CACHES_MAX_SIZE;
2972 err = memcg_update_all_list_lrus(size);
2974 memcg_nr_cache_ids = size;
2976 up_write(&memcg_cache_ids_sem);
2979 ida_simple_remove(&memcg_cache_ida, id);
2985 static void memcg_free_cache_id(int id)
2987 ida_simple_remove(&memcg_cache_ida, id);
2990 static void memcg_account_kmem(struct mem_cgroup *memcg, int nr_pages)
2992 mod_memcg_state(memcg, MEMCG_KMEM, nr_pages);
2993 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
2995 page_counter_charge(&memcg->kmem, nr_pages);
2997 page_counter_uncharge(&memcg->kmem, -nr_pages);
3003 * obj_cgroup_uncharge_pages: uncharge a number of kernel pages from a objcg
3004 * @objcg: object cgroup to uncharge
3005 * @nr_pages: number of pages to uncharge
3007 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
3008 unsigned int nr_pages)
3010 struct mem_cgroup *memcg;
3012 memcg = get_mem_cgroup_from_objcg(objcg);
3014 memcg_account_kmem(memcg, -nr_pages);
3015 refill_stock(memcg, nr_pages);
3017 css_put(&memcg->css);
3021 * obj_cgroup_charge_pages: charge a number of kernel pages to a objcg
3022 * @objcg: object cgroup to charge
3023 * @gfp: reclaim mode
3024 * @nr_pages: number of pages to charge
3026 * Returns 0 on success, an error code on failure.
3028 static int obj_cgroup_charge_pages(struct obj_cgroup *objcg, gfp_t gfp,
3029 unsigned int nr_pages)
3031 struct mem_cgroup *memcg;
3034 memcg = get_mem_cgroup_from_objcg(objcg);
3036 ret = try_charge_memcg(memcg, gfp, nr_pages);
3040 memcg_account_kmem(memcg, nr_pages);
3042 css_put(&memcg->css);
3048 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3049 * @page: page to charge
3050 * @gfp: reclaim mode
3051 * @order: allocation order
3053 * Returns 0 on success, an error code on failure.
3055 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3057 struct obj_cgroup *objcg;
3060 objcg = get_obj_cgroup_from_current();
3062 ret = obj_cgroup_charge_pages(objcg, gfp, 1 << order);
3064 page->memcg_data = (unsigned long)objcg |
3068 obj_cgroup_put(objcg);
3074 * __memcg_kmem_uncharge_page: uncharge a kmem page
3075 * @page: page to uncharge
3076 * @order: allocation order
3078 void __memcg_kmem_uncharge_page(struct page *page, int order)
3080 struct folio *folio = page_folio(page);
3081 struct obj_cgroup *objcg;
3082 unsigned int nr_pages = 1 << order;
3084 if (!folio_memcg_kmem(folio))
3087 objcg = __folio_objcg(folio);
3088 obj_cgroup_uncharge_pages(objcg, nr_pages);
3089 folio->memcg_data = 0;
3090 obj_cgroup_put(objcg);
3093 void mod_objcg_state(struct obj_cgroup *objcg, struct pglist_data *pgdat,
3094 enum node_stat_item idx, int nr)
3096 struct memcg_stock_pcp *stock;
3097 struct obj_cgroup *old = NULL;
3098 unsigned long flags;
3101 local_lock_irqsave(&memcg_stock.stock_lock, flags);
3102 stock = this_cpu_ptr(&memcg_stock);
3105 * Save vmstat data in stock and skip vmstat array update unless
3106 * accumulating over a page of vmstat data or when pgdat or idx
3109 if (stock->cached_objcg != objcg) {
3110 old = drain_obj_stock(stock);
3111 obj_cgroup_get(objcg);
3112 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3113 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3114 stock->cached_objcg = objcg;
3115 stock->cached_pgdat = pgdat;
3116 } else if (stock->cached_pgdat != pgdat) {
3117 /* Flush the existing cached vmstat data */
3118 struct pglist_data *oldpg = stock->cached_pgdat;
3120 if (stock->nr_slab_reclaimable_b) {
3121 mod_objcg_mlstate(objcg, oldpg, NR_SLAB_RECLAIMABLE_B,
3122 stock->nr_slab_reclaimable_b);
3123 stock->nr_slab_reclaimable_b = 0;
3125 if (stock->nr_slab_unreclaimable_b) {
3126 mod_objcg_mlstate(objcg, oldpg, NR_SLAB_UNRECLAIMABLE_B,
3127 stock->nr_slab_unreclaimable_b);
3128 stock->nr_slab_unreclaimable_b = 0;
3130 stock->cached_pgdat = pgdat;
3133 bytes = (idx == NR_SLAB_RECLAIMABLE_B) ? &stock->nr_slab_reclaimable_b
3134 : &stock->nr_slab_unreclaimable_b;
3136 * Even for large object >= PAGE_SIZE, the vmstat data will still be
3137 * cached locally at least once before pushing it out.
3144 if (abs(*bytes) > PAGE_SIZE) {
3152 mod_objcg_mlstate(objcg, pgdat, idx, nr);
3154 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3156 obj_cgroup_put(old);
3159 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3161 struct memcg_stock_pcp *stock;
3162 unsigned long flags;
3165 local_lock_irqsave(&memcg_stock.stock_lock, flags);
3167 stock = this_cpu_ptr(&memcg_stock);
3168 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3169 stock->nr_bytes -= nr_bytes;
3173 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3178 static struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock)
3180 struct obj_cgroup *old = stock->cached_objcg;
3185 if (stock->nr_bytes) {
3186 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3187 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3190 struct mem_cgroup *memcg;
3192 memcg = get_mem_cgroup_from_objcg(old);
3194 memcg_account_kmem(memcg, -nr_pages);
3195 __refill_stock(memcg, nr_pages);
3197 css_put(&memcg->css);
3201 * The leftover is flushed to the centralized per-memcg value.
3202 * On the next attempt to refill obj stock it will be moved
3203 * to a per-cpu stock (probably, on an other CPU), see
3204 * refill_obj_stock().
3206 * How often it's flushed is a trade-off between the memory
3207 * limit enforcement accuracy and potential CPU contention,
3208 * so it might be changed in the future.
3210 atomic_add(nr_bytes, &old->nr_charged_bytes);
3211 stock->nr_bytes = 0;
3215 * Flush the vmstat data in current stock
3217 if (stock->nr_slab_reclaimable_b || stock->nr_slab_unreclaimable_b) {
3218 if (stock->nr_slab_reclaimable_b) {
3219 mod_objcg_mlstate(old, stock->cached_pgdat,
3220 NR_SLAB_RECLAIMABLE_B,
3221 stock->nr_slab_reclaimable_b);
3222 stock->nr_slab_reclaimable_b = 0;
3224 if (stock->nr_slab_unreclaimable_b) {
3225 mod_objcg_mlstate(old, stock->cached_pgdat,
3226 NR_SLAB_UNRECLAIMABLE_B,
3227 stock->nr_slab_unreclaimable_b);
3228 stock->nr_slab_unreclaimable_b = 0;
3230 stock->cached_pgdat = NULL;
3233 stock->cached_objcg = NULL;
3235 * The `old' objects needs to be released by the caller via
3236 * obj_cgroup_put() outside of memcg_stock_pcp::stock_lock.
3241 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3242 struct mem_cgroup *root_memcg)
3244 struct mem_cgroup *memcg;
3246 if (stock->cached_objcg) {
3247 memcg = obj_cgroup_memcg(stock->cached_objcg);
3248 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3255 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes,
3256 bool allow_uncharge)
3258 struct memcg_stock_pcp *stock;
3259 struct obj_cgroup *old = NULL;
3260 unsigned long flags;
3261 unsigned int nr_pages = 0;
3263 local_lock_irqsave(&memcg_stock.stock_lock, flags);
3265 stock = this_cpu_ptr(&memcg_stock);
3266 if (stock->cached_objcg != objcg) { /* reset if necessary */
3267 old = drain_obj_stock(stock);
3268 obj_cgroup_get(objcg);
3269 stock->cached_objcg = objcg;
3270 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3271 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3272 allow_uncharge = true; /* Allow uncharge when objcg changes */
3274 stock->nr_bytes += nr_bytes;
3276 if (allow_uncharge && (stock->nr_bytes > PAGE_SIZE)) {
3277 nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3278 stock->nr_bytes &= (PAGE_SIZE - 1);
3281 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3283 obj_cgroup_put(old);
3286 obj_cgroup_uncharge_pages(objcg, nr_pages);
3289 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3291 unsigned int nr_pages, nr_bytes;
3294 if (consume_obj_stock(objcg, size))
3298 * In theory, objcg->nr_charged_bytes can have enough
3299 * pre-charged bytes to satisfy the allocation. However,
3300 * flushing objcg->nr_charged_bytes requires two atomic
3301 * operations, and objcg->nr_charged_bytes can't be big.
3302 * The shared objcg->nr_charged_bytes can also become a
3303 * performance bottleneck if all tasks of the same memcg are
3304 * trying to update it. So it's better to ignore it and try
3305 * grab some new pages. The stock's nr_bytes will be flushed to
3306 * objcg->nr_charged_bytes later on when objcg changes.
3308 * The stock's nr_bytes may contain enough pre-charged bytes
3309 * to allow one less page from being charged, but we can't rely
3310 * on the pre-charged bytes not being changed outside of
3311 * consume_obj_stock() or refill_obj_stock(). So ignore those
3312 * pre-charged bytes as well when charging pages. To avoid a
3313 * page uncharge right after a page charge, we set the
3314 * allow_uncharge flag to false when calling refill_obj_stock()
3315 * to temporarily allow the pre-charged bytes to exceed the page
3316 * size limit. The maximum reachable value of the pre-charged
3317 * bytes is (sizeof(object) + PAGE_SIZE - 2) if there is no data
3320 nr_pages = size >> PAGE_SHIFT;
3321 nr_bytes = size & (PAGE_SIZE - 1);
3326 ret = obj_cgroup_charge_pages(objcg, gfp, nr_pages);
3327 if (!ret && nr_bytes)
3328 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes, false);
3333 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3335 refill_obj_stock(objcg, size, true);
3338 #endif /* CONFIG_MEMCG_KMEM */
3341 * Because page_memcg(head) is not set on tails, set it now.
3343 void split_page_memcg(struct page *head, unsigned int nr)
3345 struct folio *folio = page_folio(head);
3346 struct mem_cgroup *memcg = folio_memcg(folio);
3349 if (mem_cgroup_disabled() || !memcg)
3352 for (i = 1; i < nr; i++)
3353 folio_page(folio, i)->memcg_data = folio->memcg_data;
3355 if (folio_memcg_kmem(folio))
3356 obj_cgroup_get_many(__folio_objcg(folio), nr - 1);
3358 css_get_many(&memcg->css, nr - 1);
3361 #ifdef CONFIG_MEMCG_SWAP
3363 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3364 * @entry: swap entry to be moved
3365 * @from: mem_cgroup which the entry is moved from
3366 * @to: mem_cgroup which the entry is moved to
3368 * It succeeds only when the swap_cgroup's record for this entry is the same
3369 * as the mem_cgroup's id of @from.
3371 * Returns 0 on success, -EINVAL on failure.
3373 * The caller must have charged to @to, IOW, called page_counter_charge() about
3374 * both res and memsw, and called css_get().
3376 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3377 struct mem_cgroup *from, struct mem_cgroup *to)
3379 unsigned short old_id, new_id;
3381 old_id = mem_cgroup_id(from);
3382 new_id = mem_cgroup_id(to);
3384 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3385 mod_memcg_state(from, MEMCG_SWAP, -1);
3386 mod_memcg_state(to, MEMCG_SWAP, 1);
3392 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3393 struct mem_cgroup *from, struct mem_cgroup *to)
3399 static DEFINE_MUTEX(memcg_max_mutex);
3401 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3402 unsigned long max, bool memsw)
3404 bool enlarge = false;
3405 bool drained = false;
3407 bool limits_invariant;
3408 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3411 if (signal_pending(current)) {
3416 mutex_lock(&memcg_max_mutex);
3418 * Make sure that the new limit (memsw or memory limit) doesn't
3419 * break our basic invariant rule memory.max <= memsw.max.
3421 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3422 max <= memcg->memsw.max;
3423 if (!limits_invariant) {
3424 mutex_unlock(&memcg_max_mutex);
3428 if (max > counter->max)
3430 ret = page_counter_set_max(counter, max);
3431 mutex_unlock(&memcg_max_mutex);
3437 drain_all_stock(memcg);
3442 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3443 GFP_KERNEL, !memsw)) {
3449 if (!ret && enlarge)
3450 memcg_oom_recover(memcg);
3455 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3457 unsigned long *total_scanned)
3459 unsigned long nr_reclaimed = 0;
3460 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3461 unsigned long reclaimed;
3463 struct mem_cgroup_tree_per_node *mctz;
3464 unsigned long excess;
3465 unsigned long nr_scanned;
3470 mctz = soft_limit_tree.rb_tree_per_node[pgdat->node_id];
3473 * Do not even bother to check the largest node if the root
3474 * is empty. Do it lockless to prevent lock bouncing. Races
3475 * are acceptable as soft limit is best effort anyway.
3477 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3481 * This loop can run a while, specially if mem_cgroup's continuously
3482 * keep exceeding their soft limit and putting the system under
3489 mz = mem_cgroup_largest_soft_limit_node(mctz);
3494 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3495 gfp_mask, &nr_scanned);
3496 nr_reclaimed += reclaimed;
3497 *total_scanned += nr_scanned;
3498 spin_lock_irq(&mctz->lock);
3499 __mem_cgroup_remove_exceeded(mz, mctz);
3502 * If we failed to reclaim anything from this memory cgroup
3503 * it is time to move on to the next cgroup
3507 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3509 excess = soft_limit_excess(mz->memcg);
3511 * One school of thought says that we should not add
3512 * back the node to the tree if reclaim returns 0.
3513 * But our reclaim could return 0, simply because due
3514 * to priority we are exposing a smaller subset of
3515 * memory to reclaim from. Consider this as a longer
3518 /* If excess == 0, no tree ops */
3519 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3520 spin_unlock_irq(&mctz->lock);
3521 css_put(&mz->memcg->css);
3524 * Could not reclaim anything and there are no more
3525 * mem cgroups to try or we seem to be looping without
3526 * reclaiming anything.
3528 if (!nr_reclaimed &&
3530 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3532 } while (!nr_reclaimed);
3534 css_put(&next_mz->memcg->css);
3535 return nr_reclaimed;
3539 * Reclaims as many pages from the given memcg as possible.
3541 * Caller is responsible for holding css reference for memcg.
3543 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3545 int nr_retries = MAX_RECLAIM_RETRIES;
3547 /* we call try-to-free pages for make this cgroup empty */
3548 lru_add_drain_all();
3550 drain_all_stock(memcg);
3552 /* try to free all pages in this cgroup */
3553 while (nr_retries && page_counter_read(&memcg->memory)) {
3554 if (signal_pending(current))
3557 if (!try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true))
3564 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3565 char *buf, size_t nbytes,
3568 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3570 if (mem_cgroup_is_root(memcg))
3572 return mem_cgroup_force_empty(memcg) ?: nbytes;
3575 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3581 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3582 struct cftype *cft, u64 val)
3587 pr_warn_once("Non-hierarchical mode is deprecated. "
3588 "Please report your usecase to linux-mm@kvack.org if you "
3589 "depend on this functionality.\n");
3594 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3598 if (mem_cgroup_is_root(memcg)) {
3599 mem_cgroup_flush_stats();
3600 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3601 memcg_page_state(memcg, NR_ANON_MAPPED);
3603 val += memcg_page_state(memcg, MEMCG_SWAP);
3606 val = page_counter_read(&memcg->memory);
3608 val = page_counter_read(&memcg->memsw);
3621 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3624 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3625 struct page_counter *counter;
3627 switch (MEMFILE_TYPE(cft->private)) {
3629 counter = &memcg->memory;
3632 counter = &memcg->memsw;
3635 counter = &memcg->kmem;
3638 counter = &memcg->tcpmem;
3644 switch (MEMFILE_ATTR(cft->private)) {
3646 if (counter == &memcg->memory)
3647 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3648 if (counter == &memcg->memsw)
3649 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3650 return (u64)page_counter_read(counter) * PAGE_SIZE;
3652 return (u64)counter->max * PAGE_SIZE;
3654 return (u64)counter->watermark * PAGE_SIZE;
3656 return counter->failcnt;
3657 case RES_SOFT_LIMIT:
3658 return (u64)memcg->soft_limit * PAGE_SIZE;
3664 #ifdef CONFIG_MEMCG_KMEM
3665 static int memcg_online_kmem(struct mem_cgroup *memcg)
3667 struct obj_cgroup *objcg;
3670 if (cgroup_memory_nokmem)
3673 if (unlikely(mem_cgroup_is_root(memcg)))
3676 memcg_id = memcg_alloc_cache_id();
3680 objcg = obj_cgroup_alloc();
3682 memcg_free_cache_id(memcg_id);
3685 objcg->memcg = memcg;
3686 rcu_assign_pointer(memcg->objcg, objcg);
3688 static_branch_enable(&memcg_kmem_enabled_key);
3690 memcg->kmemcg_id = memcg_id;
3695 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3697 struct mem_cgroup *parent;
3700 if (cgroup_memory_nokmem)
3703 if (unlikely(mem_cgroup_is_root(memcg)))
3706 parent = parent_mem_cgroup(memcg);
3708 parent = root_mem_cgroup;
3710 memcg_reparent_objcgs(memcg, parent);
3712 kmemcg_id = memcg->kmemcg_id;
3715 * After we have finished memcg_reparent_objcgs(), all list_lrus
3716 * corresponding to this cgroup are guaranteed to remain empty.
3717 * The ordering is imposed by list_lru_node->lock taken by
3718 * memcg_drain_all_list_lrus().
3720 memcg_drain_all_list_lrus(kmemcg_id, parent);
3722 memcg_free_cache_id(kmemcg_id);
3725 static int memcg_online_kmem(struct mem_cgroup *memcg)
3729 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3732 #endif /* CONFIG_MEMCG_KMEM */
3734 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3738 mutex_lock(&memcg_max_mutex);
3740 ret = page_counter_set_max(&memcg->tcpmem, max);
3744 if (!memcg->tcpmem_active) {
3746 * The active flag needs to be written after the static_key
3747 * update. This is what guarantees that the socket activation
3748 * function is the last one to run. See mem_cgroup_sk_alloc()
3749 * for details, and note that we don't mark any socket as
3750 * belonging to this memcg until that flag is up.
3752 * We need to do this, because static_keys will span multiple
3753 * sites, but we can't control their order. If we mark a socket
3754 * as accounted, but the accounting functions are not patched in
3755 * yet, we'll lose accounting.
3757 * We never race with the readers in mem_cgroup_sk_alloc(),
3758 * because when this value change, the code to process it is not
3761 static_branch_inc(&memcg_sockets_enabled_key);
3762 memcg->tcpmem_active = true;
3765 mutex_unlock(&memcg_max_mutex);
3770 * The user of this function is...
3773 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3774 char *buf, size_t nbytes, loff_t off)
3776 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3777 unsigned long nr_pages;
3780 buf = strstrip(buf);
3781 ret = page_counter_memparse(buf, "-1", &nr_pages);
3785 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3787 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3791 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3793 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3796 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3799 /* kmem.limit_in_bytes is deprecated. */
3803 ret = memcg_update_tcp_max(memcg, nr_pages);
3807 case RES_SOFT_LIMIT:
3808 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
3811 memcg->soft_limit = nr_pages;
3816 return ret ?: nbytes;
3819 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3820 size_t nbytes, loff_t off)
3822 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3823 struct page_counter *counter;
3825 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3827 counter = &memcg->memory;
3830 counter = &memcg->memsw;
3833 counter = &memcg->kmem;
3836 counter = &memcg->tcpmem;
3842 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3844 page_counter_reset_watermark(counter);
3847 counter->failcnt = 0;
3856 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3859 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3863 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3864 struct cftype *cft, u64 val)
3866 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3868 if (val & ~MOVE_MASK)
3872 * No kind of locking is needed in here, because ->can_attach() will
3873 * check this value once in the beginning of the process, and then carry
3874 * on with stale data. This means that changes to this value will only
3875 * affect task migrations starting after the change.
3877 memcg->move_charge_at_immigrate = val;
3881 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3882 struct cftype *cft, u64 val)
3890 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3891 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3892 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3894 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3895 int nid, unsigned int lru_mask, bool tree)
3897 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3898 unsigned long nr = 0;
3901 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3904 if (!(BIT(lru) & lru_mask))
3907 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3909 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3914 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3915 unsigned int lru_mask,
3918 unsigned long nr = 0;
3922 if (!(BIT(lru) & lru_mask))
3925 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3927 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3932 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3936 unsigned int lru_mask;
3939 static const struct numa_stat stats[] = {
3940 { "total", LRU_ALL },
3941 { "file", LRU_ALL_FILE },
3942 { "anon", LRU_ALL_ANON },
3943 { "unevictable", BIT(LRU_UNEVICTABLE) },
3945 const struct numa_stat *stat;
3947 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3949 mem_cgroup_flush_stats();
3951 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3952 seq_printf(m, "%s=%lu", stat->name,
3953 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3955 for_each_node_state(nid, N_MEMORY)
3956 seq_printf(m, " N%d=%lu", nid,
3957 mem_cgroup_node_nr_lru_pages(memcg, nid,
3958 stat->lru_mask, false));
3962 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3964 seq_printf(m, "hierarchical_%s=%lu", stat->name,
3965 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3967 for_each_node_state(nid, N_MEMORY)
3968 seq_printf(m, " N%d=%lu", nid,
3969 mem_cgroup_node_nr_lru_pages(memcg, nid,
3970 stat->lru_mask, true));
3976 #endif /* CONFIG_NUMA */
3978 static const unsigned int memcg1_stats[] = {
3981 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3991 static const char *const memcg1_stat_names[] = {
3994 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4004 /* Universal VM events cgroup1 shows, original sort order */
4005 static const unsigned int memcg1_events[] = {
4012 static int memcg_stat_show(struct seq_file *m, void *v)
4014 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4015 unsigned long memory, memsw;
4016 struct mem_cgroup *mi;
4019 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4021 mem_cgroup_flush_stats();
4023 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4026 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4028 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4029 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4032 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4033 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4034 memcg_events_local(memcg, memcg1_events[i]));
4036 for (i = 0; i < NR_LRU_LISTS; i++)
4037 seq_printf(m, "%s %lu\n", lru_list_name(i),
4038 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4041 /* Hierarchical information */
4042 memory = memsw = PAGE_COUNTER_MAX;
4043 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4044 memory = min(memory, READ_ONCE(mi->memory.max));
4045 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4047 seq_printf(m, "hierarchical_memory_limit %llu\n",
4048 (u64)memory * PAGE_SIZE);
4049 if (do_memsw_account())
4050 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4051 (u64)memsw * PAGE_SIZE);
4053 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4056 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4058 nr = memcg_page_state(memcg, memcg1_stats[i]);
4059 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4060 (u64)nr * PAGE_SIZE);
4063 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4064 seq_printf(m, "total_%s %llu\n",
4065 vm_event_name(memcg1_events[i]),
4066 (u64)memcg_events(memcg, memcg1_events[i]));
4068 for (i = 0; i < NR_LRU_LISTS; i++)
4069 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4070 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4073 #ifdef CONFIG_DEBUG_VM
4076 struct mem_cgroup_per_node *mz;
4077 unsigned long anon_cost = 0;
4078 unsigned long file_cost = 0;
4080 for_each_online_pgdat(pgdat) {
4081 mz = memcg->nodeinfo[pgdat->node_id];
4083 anon_cost += mz->lruvec.anon_cost;
4084 file_cost += mz->lruvec.file_cost;
4086 seq_printf(m, "anon_cost %lu\n", anon_cost);
4087 seq_printf(m, "file_cost %lu\n", file_cost);
4094 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4097 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4099 return mem_cgroup_swappiness(memcg);
4102 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4103 struct cftype *cft, u64 val)
4105 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4110 if (!mem_cgroup_is_root(memcg))
4111 memcg->swappiness = val;
4113 vm_swappiness = val;
4118 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4120 struct mem_cgroup_threshold_ary *t;
4121 unsigned long usage;
4126 t = rcu_dereference(memcg->thresholds.primary);
4128 t = rcu_dereference(memcg->memsw_thresholds.primary);
4133 usage = mem_cgroup_usage(memcg, swap);
4136 * current_threshold points to threshold just below or equal to usage.
4137 * If it's not true, a threshold was crossed after last
4138 * call of __mem_cgroup_threshold().
4140 i = t->current_threshold;
4143 * Iterate backward over array of thresholds starting from
4144 * current_threshold and check if a threshold is crossed.
4145 * If none of thresholds below usage is crossed, we read
4146 * only one element of the array here.
4148 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4149 eventfd_signal(t->entries[i].eventfd, 1);
4151 /* i = current_threshold + 1 */
4155 * Iterate forward over array of thresholds starting from
4156 * current_threshold+1 and check if a threshold is crossed.
4157 * If none of thresholds above usage is crossed, we read
4158 * only one element of the array here.
4160 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4161 eventfd_signal(t->entries[i].eventfd, 1);
4163 /* Update current_threshold */
4164 t->current_threshold = i - 1;
4169 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4172 __mem_cgroup_threshold(memcg, false);
4173 if (do_memsw_account())
4174 __mem_cgroup_threshold(memcg, true);
4176 memcg = parent_mem_cgroup(memcg);
4180 static int compare_thresholds(const void *a, const void *b)
4182 const struct mem_cgroup_threshold *_a = a;
4183 const struct mem_cgroup_threshold *_b = b;
4185 if (_a->threshold > _b->threshold)
4188 if (_a->threshold < _b->threshold)
4194 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4196 struct mem_cgroup_eventfd_list *ev;
4198 spin_lock(&memcg_oom_lock);
4200 list_for_each_entry(ev, &memcg->oom_notify, list)
4201 eventfd_signal(ev->eventfd, 1);
4203 spin_unlock(&memcg_oom_lock);
4207 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4209 struct mem_cgroup *iter;
4211 for_each_mem_cgroup_tree(iter, memcg)
4212 mem_cgroup_oom_notify_cb(iter);
4215 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4216 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4218 struct mem_cgroup_thresholds *thresholds;
4219 struct mem_cgroup_threshold_ary *new;
4220 unsigned long threshold;
4221 unsigned long usage;
4224 ret = page_counter_memparse(args, "-1", &threshold);
4228 mutex_lock(&memcg->thresholds_lock);
4231 thresholds = &memcg->thresholds;
4232 usage = mem_cgroup_usage(memcg, false);
4233 } else if (type == _MEMSWAP) {
4234 thresholds = &memcg->memsw_thresholds;
4235 usage = mem_cgroup_usage(memcg, true);
4239 /* Check if a threshold crossed before adding a new one */
4240 if (thresholds->primary)
4241 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4243 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4245 /* Allocate memory for new array of thresholds */
4246 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4253 /* Copy thresholds (if any) to new array */
4254 if (thresholds->primary)
4255 memcpy(new->entries, thresholds->primary->entries,
4256 flex_array_size(new, entries, size - 1));
4258 /* Add new threshold */
4259 new->entries[size - 1].eventfd = eventfd;
4260 new->entries[size - 1].threshold = threshold;
4262 /* Sort thresholds. Registering of new threshold isn't time-critical */
4263 sort(new->entries, size, sizeof(*new->entries),
4264 compare_thresholds, NULL);
4266 /* Find current threshold */
4267 new->current_threshold = -1;
4268 for (i = 0; i < size; i++) {
4269 if (new->entries[i].threshold <= usage) {
4271 * new->current_threshold will not be used until
4272 * rcu_assign_pointer(), so it's safe to increment
4275 ++new->current_threshold;
4280 /* Free old spare buffer and save old primary buffer as spare */
4281 kfree(thresholds->spare);
4282 thresholds->spare = thresholds->primary;
4284 rcu_assign_pointer(thresholds->primary, new);
4286 /* To be sure that nobody uses thresholds */
4290 mutex_unlock(&memcg->thresholds_lock);
4295 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4296 struct eventfd_ctx *eventfd, const char *args)
4298 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4301 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4302 struct eventfd_ctx *eventfd, const char *args)
4304 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4307 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4308 struct eventfd_ctx *eventfd, enum res_type type)
4310 struct mem_cgroup_thresholds *thresholds;
4311 struct mem_cgroup_threshold_ary *new;
4312 unsigned long usage;
4313 int i, j, size, entries;
4315 mutex_lock(&memcg->thresholds_lock);
4318 thresholds = &memcg->thresholds;
4319 usage = mem_cgroup_usage(memcg, false);
4320 } else if (type == _MEMSWAP) {
4321 thresholds = &memcg->memsw_thresholds;
4322 usage = mem_cgroup_usage(memcg, true);
4326 if (!thresholds->primary)
4329 /* Check if a threshold crossed before removing */
4330 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4332 /* Calculate new number of threshold */
4334 for (i = 0; i < thresholds->primary->size; i++) {
4335 if (thresholds->primary->entries[i].eventfd != eventfd)
4341 new = thresholds->spare;
4343 /* If no items related to eventfd have been cleared, nothing to do */
4347 /* Set thresholds array to NULL if we don't have thresholds */
4356 /* Copy thresholds and find current threshold */
4357 new->current_threshold = -1;
4358 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4359 if (thresholds->primary->entries[i].eventfd == eventfd)
4362 new->entries[j] = thresholds->primary->entries[i];
4363 if (new->entries[j].threshold <= usage) {
4365 * new->current_threshold will not be used
4366 * until rcu_assign_pointer(), so it's safe to increment
4369 ++new->current_threshold;
4375 /* Swap primary and spare array */
4376 thresholds->spare = thresholds->primary;
4378 rcu_assign_pointer(thresholds->primary, new);
4380 /* To be sure that nobody uses thresholds */
4383 /* If all events are unregistered, free the spare array */
4385 kfree(thresholds->spare);
4386 thresholds->spare = NULL;
4389 mutex_unlock(&memcg->thresholds_lock);
4392 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4393 struct eventfd_ctx *eventfd)
4395 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4398 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4399 struct eventfd_ctx *eventfd)
4401 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4404 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4405 struct eventfd_ctx *eventfd, const char *args)
4407 struct mem_cgroup_eventfd_list *event;
4409 event = kmalloc(sizeof(*event), GFP_KERNEL);
4413 spin_lock(&memcg_oom_lock);
4415 event->eventfd = eventfd;
4416 list_add(&event->list, &memcg->oom_notify);
4418 /* already in OOM ? */
4419 if (memcg->under_oom)
4420 eventfd_signal(eventfd, 1);
4421 spin_unlock(&memcg_oom_lock);
4426 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4427 struct eventfd_ctx *eventfd)
4429 struct mem_cgroup_eventfd_list *ev, *tmp;
4431 spin_lock(&memcg_oom_lock);
4433 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4434 if (ev->eventfd == eventfd) {
4435 list_del(&ev->list);
4440 spin_unlock(&memcg_oom_lock);
4443 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4445 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4447 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4448 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4449 seq_printf(sf, "oom_kill %lu\n",
4450 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4454 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4455 struct cftype *cft, u64 val)
4457 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4459 /* cannot set to root cgroup and only 0 and 1 are allowed */
4460 if (mem_cgroup_is_root(memcg) || !((val == 0) || (val == 1)))
4463 memcg->oom_kill_disable = val;
4465 memcg_oom_recover(memcg);
4470 #ifdef CONFIG_CGROUP_WRITEBACK
4472 #include <trace/events/writeback.h>
4474 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4476 return wb_domain_init(&memcg->cgwb_domain, gfp);
4479 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4481 wb_domain_exit(&memcg->cgwb_domain);
4484 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4486 wb_domain_size_changed(&memcg->cgwb_domain);
4489 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4491 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4493 if (!memcg->css.parent)
4496 return &memcg->cgwb_domain;
4500 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4501 * @wb: bdi_writeback in question
4502 * @pfilepages: out parameter for number of file pages
4503 * @pheadroom: out parameter for number of allocatable pages according to memcg
4504 * @pdirty: out parameter for number of dirty pages
4505 * @pwriteback: out parameter for number of pages under writeback
4507 * Determine the numbers of file, headroom, dirty, and writeback pages in
4508 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4509 * is a bit more involved.
4511 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4512 * headroom is calculated as the lowest headroom of itself and the
4513 * ancestors. Note that this doesn't consider the actual amount of
4514 * available memory in the system. The caller should further cap
4515 * *@pheadroom accordingly.
4517 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4518 unsigned long *pheadroom, unsigned long *pdirty,
4519 unsigned long *pwriteback)
4521 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4522 struct mem_cgroup *parent;
4524 mem_cgroup_flush_stats();
4526 *pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
4527 *pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
4528 *pfilepages = memcg_page_state(memcg, NR_INACTIVE_FILE) +
4529 memcg_page_state(memcg, NR_ACTIVE_FILE);
4531 *pheadroom = PAGE_COUNTER_MAX;
4532 while ((parent = parent_mem_cgroup(memcg))) {
4533 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4534 READ_ONCE(memcg->memory.high));
4535 unsigned long used = page_counter_read(&memcg->memory);
4537 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4543 * Foreign dirty flushing
4545 * There's an inherent mismatch between memcg and writeback. The former
4546 * tracks ownership per-page while the latter per-inode. This was a
4547 * deliberate design decision because honoring per-page ownership in the
4548 * writeback path is complicated, may lead to higher CPU and IO overheads
4549 * and deemed unnecessary given that write-sharing an inode across
4550 * different cgroups isn't a common use-case.
4552 * Combined with inode majority-writer ownership switching, this works well
4553 * enough in most cases but there are some pathological cases. For
4554 * example, let's say there are two cgroups A and B which keep writing to
4555 * different but confined parts of the same inode. B owns the inode and
4556 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4557 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4558 * triggering background writeback. A will be slowed down without a way to
4559 * make writeback of the dirty pages happen.
4561 * Conditions like the above can lead to a cgroup getting repeatedly and
4562 * severely throttled after making some progress after each
4563 * dirty_expire_interval while the underlying IO device is almost
4566 * Solving this problem completely requires matching the ownership tracking
4567 * granularities between memcg and writeback in either direction. However,
4568 * the more egregious behaviors can be avoided by simply remembering the
4569 * most recent foreign dirtying events and initiating remote flushes on
4570 * them when local writeback isn't enough to keep the memory clean enough.
4572 * The following two functions implement such mechanism. When a foreign
4573 * page - a page whose memcg and writeback ownerships don't match - is
4574 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4575 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4576 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4577 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4578 * foreign bdi_writebacks which haven't expired. Both the numbers of
4579 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4580 * limited to MEMCG_CGWB_FRN_CNT.
4582 * The mechanism only remembers IDs and doesn't hold any object references.
4583 * As being wrong occasionally doesn't matter, updates and accesses to the
4584 * records are lockless and racy.
4586 void mem_cgroup_track_foreign_dirty_slowpath(struct folio *folio,
4587 struct bdi_writeback *wb)
4589 struct mem_cgroup *memcg = folio_memcg(folio);
4590 struct memcg_cgwb_frn *frn;
4591 u64 now = get_jiffies_64();
4592 u64 oldest_at = now;
4596 trace_track_foreign_dirty(folio, wb);
4599 * Pick the slot to use. If there is already a slot for @wb, keep
4600 * using it. If not replace the oldest one which isn't being
4603 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4604 frn = &memcg->cgwb_frn[i];
4605 if (frn->bdi_id == wb->bdi->id &&
4606 frn->memcg_id == wb->memcg_css->id)
4608 if (time_before64(frn->at, oldest_at) &&
4609 atomic_read(&frn->done.cnt) == 1) {
4611 oldest_at = frn->at;
4615 if (i < MEMCG_CGWB_FRN_CNT) {
4617 * Re-using an existing one. Update timestamp lazily to
4618 * avoid making the cacheline hot. We want them to be
4619 * reasonably up-to-date and significantly shorter than
4620 * dirty_expire_interval as that's what expires the record.
4621 * Use the shorter of 1s and dirty_expire_interval / 8.
4623 unsigned long update_intv =
4624 min_t(unsigned long, HZ,
4625 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4627 if (time_before64(frn->at, now - update_intv))
4629 } else if (oldest >= 0) {
4630 /* replace the oldest free one */
4631 frn = &memcg->cgwb_frn[oldest];
4632 frn->bdi_id = wb->bdi->id;
4633 frn->memcg_id = wb->memcg_css->id;
4638 /* issue foreign writeback flushes for recorded foreign dirtying events */
4639 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4641 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4642 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4643 u64 now = jiffies_64;
4646 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4647 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4650 * If the record is older than dirty_expire_interval,
4651 * writeback on it has already started. No need to kick it
4652 * off again. Also, don't start a new one if there's
4653 * already one in flight.
4655 if (time_after64(frn->at, now - intv) &&
4656 atomic_read(&frn->done.cnt) == 1) {
4658 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4659 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id,
4660 WB_REASON_FOREIGN_FLUSH,
4666 #else /* CONFIG_CGROUP_WRITEBACK */
4668 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4673 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4677 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4681 #endif /* CONFIG_CGROUP_WRITEBACK */
4684 * DO NOT USE IN NEW FILES.
4686 * "cgroup.event_control" implementation.
4688 * This is way over-engineered. It tries to support fully configurable
4689 * events for each user. Such level of flexibility is completely
4690 * unnecessary especially in the light of the planned unified hierarchy.
4692 * Please deprecate this and replace with something simpler if at all
4697 * Unregister event and free resources.
4699 * Gets called from workqueue.
4701 static void memcg_event_remove(struct work_struct *work)
4703 struct mem_cgroup_event *event =
4704 container_of(work, struct mem_cgroup_event, remove);
4705 struct mem_cgroup *memcg = event->memcg;
4707 remove_wait_queue(event->wqh, &event->wait);
4709 event->unregister_event(memcg, event->eventfd);
4711 /* Notify userspace the event is going away. */
4712 eventfd_signal(event->eventfd, 1);
4714 eventfd_ctx_put(event->eventfd);
4716 css_put(&memcg->css);
4720 * Gets called on EPOLLHUP on eventfd when user closes it.
4722 * Called with wqh->lock held and interrupts disabled.
4724 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4725 int sync, void *key)
4727 struct mem_cgroup_event *event =
4728 container_of(wait, struct mem_cgroup_event, wait);
4729 struct mem_cgroup *memcg = event->memcg;
4730 __poll_t flags = key_to_poll(key);
4732 if (flags & EPOLLHUP) {
4734 * If the event has been detached at cgroup removal, we
4735 * can simply return knowing the other side will cleanup
4738 * We can't race against event freeing since the other
4739 * side will require wqh->lock via remove_wait_queue(),
4742 spin_lock(&memcg->event_list_lock);
4743 if (!list_empty(&event->list)) {
4744 list_del_init(&event->list);
4746 * We are in atomic context, but cgroup_event_remove()
4747 * may sleep, so we have to call it in workqueue.
4749 schedule_work(&event->remove);
4751 spin_unlock(&memcg->event_list_lock);
4757 static void memcg_event_ptable_queue_proc(struct file *file,
4758 wait_queue_head_t *wqh, poll_table *pt)
4760 struct mem_cgroup_event *event =
4761 container_of(pt, struct mem_cgroup_event, pt);
4764 add_wait_queue(wqh, &event->wait);
4768 * DO NOT USE IN NEW FILES.
4770 * Parse input and register new cgroup event handler.
4772 * Input must be in format '<event_fd> <control_fd> <args>'.
4773 * Interpretation of args is defined by control file implementation.
4775 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4776 char *buf, size_t nbytes, loff_t off)
4778 struct cgroup_subsys_state *css = of_css(of);
4779 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4780 struct mem_cgroup_event *event;
4781 struct cgroup_subsys_state *cfile_css;
4782 unsigned int efd, cfd;
4789 if (IS_ENABLED(CONFIG_PREEMPT_RT))
4792 buf = strstrip(buf);
4794 efd = simple_strtoul(buf, &endp, 10);
4799 cfd = simple_strtoul(buf, &endp, 10);
4800 if ((*endp != ' ') && (*endp != '\0'))
4804 event = kzalloc(sizeof(*event), GFP_KERNEL);
4808 event->memcg = memcg;
4809 INIT_LIST_HEAD(&event->list);
4810 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4811 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4812 INIT_WORK(&event->remove, memcg_event_remove);
4820 event->eventfd = eventfd_ctx_fileget(efile.file);
4821 if (IS_ERR(event->eventfd)) {
4822 ret = PTR_ERR(event->eventfd);
4829 goto out_put_eventfd;
4832 /* the process need read permission on control file */
4833 /* AV: shouldn't we check that it's been opened for read instead? */
4834 ret = file_permission(cfile.file, MAY_READ);
4839 * Determine the event callbacks and set them in @event. This used
4840 * to be done via struct cftype but cgroup core no longer knows
4841 * about these events. The following is crude but the whole thing
4842 * is for compatibility anyway.
4844 * DO NOT ADD NEW FILES.
4846 name = cfile.file->f_path.dentry->d_name.name;
4848 if (!strcmp(name, "memory.usage_in_bytes")) {
4849 event->register_event = mem_cgroup_usage_register_event;
4850 event->unregister_event = mem_cgroup_usage_unregister_event;
4851 } else if (!strcmp(name, "memory.oom_control")) {
4852 event->register_event = mem_cgroup_oom_register_event;
4853 event->unregister_event = mem_cgroup_oom_unregister_event;
4854 } else if (!strcmp(name, "memory.pressure_level")) {
4855 event->register_event = vmpressure_register_event;
4856 event->unregister_event = vmpressure_unregister_event;
4857 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4858 event->register_event = memsw_cgroup_usage_register_event;
4859 event->unregister_event = memsw_cgroup_usage_unregister_event;
4866 * Verify @cfile should belong to @css. Also, remaining events are
4867 * automatically removed on cgroup destruction but the removal is
4868 * asynchronous, so take an extra ref on @css.
4870 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4871 &memory_cgrp_subsys);
4873 if (IS_ERR(cfile_css))
4875 if (cfile_css != css) {
4880 ret = event->register_event(memcg, event->eventfd, buf);
4884 vfs_poll(efile.file, &event->pt);
4886 spin_lock_irq(&memcg->event_list_lock);
4887 list_add(&event->list, &memcg->event_list);
4888 spin_unlock_irq(&memcg->event_list_lock);
4900 eventfd_ctx_put(event->eventfd);
4909 #if defined(CONFIG_MEMCG_KMEM) && (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
4910 static int mem_cgroup_slab_show(struct seq_file *m, void *p)
4914 * Please, take a look at tools/cgroup/slabinfo.py .
4920 static struct cftype mem_cgroup_legacy_files[] = {
4922 .name = "usage_in_bytes",
4923 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4924 .read_u64 = mem_cgroup_read_u64,
4927 .name = "max_usage_in_bytes",
4928 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4929 .write = mem_cgroup_reset,
4930 .read_u64 = mem_cgroup_read_u64,
4933 .name = "limit_in_bytes",
4934 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4935 .write = mem_cgroup_write,
4936 .read_u64 = mem_cgroup_read_u64,
4939 .name = "soft_limit_in_bytes",
4940 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4941 .write = mem_cgroup_write,
4942 .read_u64 = mem_cgroup_read_u64,
4946 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4947 .write = mem_cgroup_reset,
4948 .read_u64 = mem_cgroup_read_u64,
4952 .seq_show = memcg_stat_show,
4955 .name = "force_empty",
4956 .write = mem_cgroup_force_empty_write,
4959 .name = "use_hierarchy",
4960 .write_u64 = mem_cgroup_hierarchy_write,
4961 .read_u64 = mem_cgroup_hierarchy_read,
4964 .name = "cgroup.event_control", /* XXX: for compat */
4965 .write = memcg_write_event_control,
4966 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4969 .name = "swappiness",
4970 .read_u64 = mem_cgroup_swappiness_read,
4971 .write_u64 = mem_cgroup_swappiness_write,
4974 .name = "move_charge_at_immigrate",
4975 .read_u64 = mem_cgroup_move_charge_read,
4976 .write_u64 = mem_cgroup_move_charge_write,
4979 .name = "oom_control",
4980 .seq_show = mem_cgroup_oom_control_read,
4981 .write_u64 = mem_cgroup_oom_control_write,
4982 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4985 .name = "pressure_level",
4989 .name = "numa_stat",
4990 .seq_show = memcg_numa_stat_show,
4994 .name = "kmem.limit_in_bytes",
4995 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4996 .write = mem_cgroup_write,
4997 .read_u64 = mem_cgroup_read_u64,
5000 .name = "kmem.usage_in_bytes",
5001 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5002 .read_u64 = mem_cgroup_read_u64,
5005 .name = "kmem.failcnt",
5006 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5007 .write = mem_cgroup_reset,
5008 .read_u64 = mem_cgroup_read_u64,
5011 .name = "kmem.max_usage_in_bytes",
5012 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5013 .write = mem_cgroup_reset,
5014 .read_u64 = mem_cgroup_read_u64,
5016 #if defined(CONFIG_MEMCG_KMEM) && \
5017 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5019 .name = "kmem.slabinfo",
5020 .seq_show = mem_cgroup_slab_show,
5024 .name = "kmem.tcp.limit_in_bytes",
5025 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5026 .write = mem_cgroup_write,
5027 .read_u64 = mem_cgroup_read_u64,
5030 .name = "kmem.tcp.usage_in_bytes",
5031 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5032 .read_u64 = mem_cgroup_read_u64,
5035 .name = "kmem.tcp.failcnt",
5036 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5037 .write = mem_cgroup_reset,
5038 .read_u64 = mem_cgroup_read_u64,
5041 .name = "kmem.tcp.max_usage_in_bytes",
5042 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5043 .write = mem_cgroup_reset,
5044 .read_u64 = mem_cgroup_read_u64,
5046 { }, /* terminate */
5050 * Private memory cgroup IDR
5052 * Swap-out records and page cache shadow entries need to store memcg
5053 * references in constrained space, so we maintain an ID space that is
5054 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5055 * memory-controlled cgroups to 64k.
5057 * However, there usually are many references to the offline CSS after
5058 * the cgroup has been destroyed, such as page cache or reclaimable
5059 * slab objects, that don't need to hang on to the ID. We want to keep
5060 * those dead CSS from occupying IDs, or we might quickly exhaust the
5061 * relatively small ID space and prevent the creation of new cgroups
5062 * even when there are much fewer than 64k cgroups - possibly none.
5064 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5065 * be freed and recycled when it's no longer needed, which is usually
5066 * when the CSS is offlined.
5068 * The only exception to that are records of swapped out tmpfs/shmem
5069 * pages that need to be attributed to live ancestors on swapin. But
5070 * those references are manageable from userspace.
5073 static DEFINE_IDR(mem_cgroup_idr);
5075 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5077 if (memcg->id.id > 0) {
5078 idr_remove(&mem_cgroup_idr, memcg->id.id);
5083 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5086 refcount_add(n, &memcg->id.ref);
5089 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5091 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5092 mem_cgroup_id_remove(memcg);
5094 /* Memcg ID pins CSS */
5095 css_put(&memcg->css);
5099 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5101 mem_cgroup_id_put_many(memcg, 1);
5105 * mem_cgroup_from_id - look up a memcg from a memcg id
5106 * @id: the memcg id to look up
5108 * Caller must hold rcu_read_lock().
5110 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5112 WARN_ON_ONCE(!rcu_read_lock_held());
5113 return idr_find(&mem_cgroup_idr, id);
5116 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5118 struct mem_cgroup_per_node *pn;
5121 * This routine is called against possible nodes.
5122 * But it's BUG to call kmalloc() against offline node.
5124 * TODO: this routine can waste much memory for nodes which will
5125 * never be onlined. It's better to use memory hotplug callback
5128 if (!node_state(node, N_NORMAL_MEMORY))
5130 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5134 pn->lruvec_stats_percpu = alloc_percpu_gfp(struct lruvec_stats_percpu,
5135 GFP_KERNEL_ACCOUNT);
5136 if (!pn->lruvec_stats_percpu) {
5141 lruvec_init(&pn->lruvec);
5144 memcg->nodeinfo[node] = pn;
5148 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5150 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5155 free_percpu(pn->lruvec_stats_percpu);
5159 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5164 free_mem_cgroup_per_node_info(memcg, node);
5165 free_percpu(memcg->vmstats_percpu);
5169 static void mem_cgroup_free(struct mem_cgroup *memcg)
5171 memcg_wb_domain_exit(memcg);
5172 __mem_cgroup_free(memcg);
5175 static struct mem_cgroup *mem_cgroup_alloc(void)
5177 struct mem_cgroup *memcg;
5179 int __maybe_unused i;
5180 long error = -ENOMEM;
5182 memcg = kzalloc(struct_size(memcg, nodeinfo, nr_node_ids), GFP_KERNEL);
5184 return ERR_PTR(error);
5186 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5187 1, MEM_CGROUP_ID_MAX,
5189 if (memcg->id.id < 0) {
5190 error = memcg->id.id;
5194 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5195 GFP_KERNEL_ACCOUNT);
5196 if (!memcg->vmstats_percpu)
5200 if (alloc_mem_cgroup_per_node_info(memcg, node))
5203 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5206 INIT_WORK(&memcg->high_work, high_work_func);
5207 INIT_LIST_HEAD(&memcg->oom_notify);
5208 mutex_init(&memcg->thresholds_lock);
5209 spin_lock_init(&memcg->move_lock);
5210 vmpressure_init(&memcg->vmpressure);
5211 INIT_LIST_HEAD(&memcg->event_list);
5212 spin_lock_init(&memcg->event_list_lock);
5213 memcg->socket_pressure = jiffies;
5214 #ifdef CONFIG_MEMCG_KMEM
5215 memcg->kmemcg_id = -1;
5216 INIT_LIST_HEAD(&memcg->objcg_list);
5218 #ifdef CONFIG_CGROUP_WRITEBACK
5219 INIT_LIST_HEAD(&memcg->cgwb_list);
5220 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5221 memcg->cgwb_frn[i].done =
5222 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5224 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5225 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5226 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5227 memcg->deferred_split_queue.split_queue_len = 0;
5229 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5232 mem_cgroup_id_remove(memcg);
5233 __mem_cgroup_free(memcg);
5234 return ERR_PTR(error);
5237 static struct cgroup_subsys_state * __ref
5238 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5240 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5241 struct mem_cgroup *memcg, *old_memcg;
5243 old_memcg = set_active_memcg(parent);
5244 memcg = mem_cgroup_alloc();
5245 set_active_memcg(old_memcg);
5247 return ERR_CAST(memcg);
5249 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5250 memcg->soft_limit = PAGE_COUNTER_MAX;
5251 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5253 memcg->swappiness = mem_cgroup_swappiness(parent);
5254 memcg->oom_kill_disable = parent->oom_kill_disable;
5256 page_counter_init(&memcg->memory, &parent->memory);
5257 page_counter_init(&memcg->swap, &parent->swap);
5258 page_counter_init(&memcg->kmem, &parent->kmem);
5259 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5261 page_counter_init(&memcg->memory, NULL);
5262 page_counter_init(&memcg->swap, NULL);
5263 page_counter_init(&memcg->kmem, NULL);
5264 page_counter_init(&memcg->tcpmem, NULL);
5266 root_mem_cgroup = memcg;
5270 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5271 static_branch_inc(&memcg_sockets_enabled_key);
5276 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5278 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5280 if (memcg_online_kmem(memcg))
5284 * A memcg must be visible for expand_shrinker_info()
5285 * by the time the maps are allocated. So, we allocate maps
5286 * here, when for_each_mem_cgroup() can't skip it.
5288 if (alloc_shrinker_info(memcg))
5291 /* Online state pins memcg ID, memcg ID pins CSS */
5292 refcount_set(&memcg->id.ref, 1);
5295 if (unlikely(mem_cgroup_is_root(memcg)))
5296 queue_delayed_work(system_unbound_wq, &stats_flush_dwork,
5300 memcg_offline_kmem(memcg);
5302 mem_cgroup_id_remove(memcg);
5306 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5308 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5309 struct mem_cgroup_event *event, *tmp;
5312 * Unregister events and notify userspace.
5313 * Notify userspace about cgroup removing only after rmdir of cgroup
5314 * directory to avoid race between userspace and kernelspace.
5316 spin_lock_irq(&memcg->event_list_lock);
5317 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5318 list_del_init(&event->list);
5319 schedule_work(&event->remove);
5321 spin_unlock_irq(&memcg->event_list_lock);
5323 page_counter_set_min(&memcg->memory, 0);
5324 page_counter_set_low(&memcg->memory, 0);
5326 memcg_offline_kmem(memcg);
5327 reparent_shrinker_deferred(memcg);
5328 wb_memcg_offline(memcg);
5330 drain_all_stock(memcg);
5332 mem_cgroup_id_put(memcg);
5335 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5337 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5339 invalidate_reclaim_iterators(memcg);
5342 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5344 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5345 int __maybe_unused i;
5347 #ifdef CONFIG_CGROUP_WRITEBACK
5348 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5349 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5351 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5352 static_branch_dec(&memcg_sockets_enabled_key);
5354 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5355 static_branch_dec(&memcg_sockets_enabled_key);
5357 vmpressure_cleanup(&memcg->vmpressure);
5358 cancel_work_sync(&memcg->high_work);
5359 mem_cgroup_remove_from_trees(memcg);
5360 free_shrinker_info(memcg);
5361 mem_cgroup_free(memcg);
5365 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5366 * @css: the target css
5368 * Reset the states of the mem_cgroup associated with @css. This is
5369 * invoked when the userland requests disabling on the default hierarchy
5370 * but the memcg is pinned through dependency. The memcg should stop
5371 * applying policies and should revert to the vanilla state as it may be
5372 * made visible again.
5374 * The current implementation only resets the essential configurations.
5375 * This needs to be expanded to cover all the visible parts.
5377 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5379 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5381 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5382 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5383 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5384 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5385 page_counter_set_min(&memcg->memory, 0);
5386 page_counter_set_low(&memcg->memory, 0);
5387 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5388 memcg->soft_limit = PAGE_COUNTER_MAX;
5389 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5390 memcg_wb_domain_size_changed(memcg);
5393 static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state *css, int cpu)
5395 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5396 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5397 struct memcg_vmstats_percpu *statc;
5401 statc = per_cpu_ptr(memcg->vmstats_percpu, cpu);
5403 for (i = 0; i < MEMCG_NR_STAT; i++) {
5405 * Collect the aggregated propagation counts of groups
5406 * below us. We're in a per-cpu loop here and this is
5407 * a global counter, so the first cycle will get them.
5409 delta = memcg->vmstats.state_pending[i];
5411 memcg->vmstats.state_pending[i] = 0;
5413 /* Add CPU changes on this level since the last flush */
5414 v = READ_ONCE(statc->state[i]);
5415 if (v != statc->state_prev[i]) {
5416 delta += v - statc->state_prev[i];
5417 statc->state_prev[i] = v;
5423 /* Aggregate counts on this level and propagate upwards */
5424 memcg->vmstats.state[i] += delta;
5426 parent->vmstats.state_pending[i] += delta;
5429 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
5430 delta = memcg->vmstats.events_pending[i];
5432 memcg->vmstats.events_pending[i] = 0;
5434 v = READ_ONCE(statc->events[i]);
5435 if (v != statc->events_prev[i]) {
5436 delta += v - statc->events_prev[i];
5437 statc->events_prev[i] = v;
5443 memcg->vmstats.events[i] += delta;
5445 parent->vmstats.events_pending[i] += delta;
5448 for_each_node_state(nid, N_MEMORY) {
5449 struct mem_cgroup_per_node *pn = memcg->nodeinfo[nid];
5450 struct mem_cgroup_per_node *ppn = NULL;
5451 struct lruvec_stats_percpu *lstatc;
5454 ppn = parent->nodeinfo[nid];
5456 lstatc = per_cpu_ptr(pn->lruvec_stats_percpu, cpu);
5458 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
5459 delta = pn->lruvec_stats.state_pending[i];
5461 pn->lruvec_stats.state_pending[i] = 0;
5463 v = READ_ONCE(lstatc->state[i]);
5464 if (v != lstatc->state_prev[i]) {
5465 delta += v - lstatc->state_prev[i];
5466 lstatc->state_prev[i] = v;
5472 pn->lruvec_stats.state[i] += delta;
5474 ppn->lruvec_stats.state_pending[i] += delta;
5480 /* Handlers for move charge at task migration. */
5481 static int mem_cgroup_do_precharge(unsigned long count)
5485 /* Try a single bulk charge without reclaim first, kswapd may wake */
5486 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5488 mc.precharge += count;
5492 /* Try charges one by one with reclaim, but do not retry */
5494 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5508 enum mc_target_type {
5515 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5516 unsigned long addr, pte_t ptent)
5518 struct page *page = vm_normal_page(vma, addr, ptent);
5520 if (!page || !page_mapped(page))
5522 if (PageAnon(page)) {
5523 if (!(mc.flags & MOVE_ANON))
5526 if (!(mc.flags & MOVE_FILE))
5529 if (!get_page_unless_zero(page))
5535 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5536 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5537 pte_t ptent, swp_entry_t *entry)
5539 struct page *page = NULL;
5540 swp_entry_t ent = pte_to_swp_entry(ptent);
5542 if (!(mc.flags & MOVE_ANON))
5546 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5547 * a device and because they are not accessible by CPU they are store
5548 * as special swap entry in the CPU page table.
5550 if (is_device_private_entry(ent)) {
5551 page = pfn_swap_entry_to_page(ent);
5553 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5554 * a refcount of 1 when free (unlike normal page)
5556 if (!page_ref_add_unless(page, 1, 1))
5561 if (non_swap_entry(ent))
5565 * Because lookup_swap_cache() updates some statistics counter,
5566 * we call find_get_page() with swapper_space directly.
5568 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5569 entry->val = ent.val;
5574 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5575 pte_t ptent, swp_entry_t *entry)
5581 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5582 unsigned long addr, pte_t ptent)
5584 if (!vma->vm_file) /* anonymous vma */
5586 if (!(mc.flags & MOVE_FILE))
5589 /* page is moved even if it's not RSS of this task(page-faulted). */
5590 /* shmem/tmpfs may report page out on swap: account for that too. */
5591 return find_get_incore_page(vma->vm_file->f_mapping,
5592 linear_page_index(vma, addr));
5596 * mem_cgroup_move_account - move account of the page
5598 * @compound: charge the page as compound or small page
5599 * @from: mem_cgroup which the page is moved from.
5600 * @to: mem_cgroup which the page is moved to. @from != @to.
5602 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5604 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5607 static int mem_cgroup_move_account(struct page *page,
5609 struct mem_cgroup *from,
5610 struct mem_cgroup *to)
5612 struct folio *folio = page_folio(page);
5613 struct lruvec *from_vec, *to_vec;
5614 struct pglist_data *pgdat;
5615 unsigned int nr_pages = compound ? folio_nr_pages(folio) : 1;
5618 VM_BUG_ON(from == to);
5619 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
5620 VM_BUG_ON(compound && !folio_test_large(folio));
5623 * Prevent mem_cgroup_migrate() from looking at
5624 * page's memory cgroup of its source page while we change it.
5627 if (!folio_trylock(folio))
5631 if (folio_memcg(folio) != from)
5634 pgdat = folio_pgdat(folio);
5635 from_vec = mem_cgroup_lruvec(from, pgdat);
5636 to_vec = mem_cgroup_lruvec(to, pgdat);
5638 folio_memcg_lock(folio);
5640 if (folio_test_anon(folio)) {
5641 if (folio_mapped(folio)) {
5642 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5643 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5644 if (folio_test_transhuge(folio)) {
5645 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5647 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5652 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5653 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5655 if (folio_test_swapbacked(folio)) {
5656 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5657 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5660 if (folio_mapped(folio)) {
5661 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5662 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5665 if (folio_test_dirty(folio)) {
5666 struct address_space *mapping = folio_mapping(folio);
5668 if (mapping_can_writeback(mapping)) {
5669 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5671 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5677 if (folio_test_writeback(folio)) {
5678 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5679 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5683 * All state has been migrated, let's switch to the new memcg.
5685 * It is safe to change page's memcg here because the page
5686 * is referenced, charged, isolated, and locked: we can't race
5687 * with (un)charging, migration, LRU putback, or anything else
5688 * that would rely on a stable page's memory cgroup.
5690 * Note that lock_page_memcg is a memcg lock, not a page lock,
5691 * to save space. As soon as we switch page's memory cgroup to a
5692 * new memcg that isn't locked, the above state can change
5693 * concurrently again. Make sure we're truly done with it.
5698 css_put(&from->css);
5700 folio->memcg_data = (unsigned long)to;
5702 __folio_memcg_unlock(from);
5705 nid = folio_nid(folio);
5707 local_irq_disable();
5708 mem_cgroup_charge_statistics(to, nr_pages);
5709 memcg_check_events(to, nid);
5710 mem_cgroup_charge_statistics(from, -nr_pages);
5711 memcg_check_events(from, nid);
5714 folio_unlock(folio);
5720 * get_mctgt_type - get target type of moving charge
5721 * @vma: the vma the pte to be checked belongs
5722 * @addr: the address corresponding to the pte to be checked
5723 * @ptent: the pte to be checked
5724 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5727 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5728 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5729 * move charge. if @target is not NULL, the page is stored in target->page
5730 * with extra refcnt got(Callers should handle it).
5731 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5732 * target for charge migration. if @target is not NULL, the entry is stored
5734 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5735 * (so ZONE_DEVICE page and thus not on the lru).
5736 * For now we such page is charge like a regular page would be as for all
5737 * intent and purposes it is just special memory taking the place of a
5740 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5742 * Called with pte lock held.
5745 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5746 unsigned long addr, pte_t ptent, union mc_target *target)
5748 struct page *page = NULL;
5749 enum mc_target_type ret = MC_TARGET_NONE;
5750 swp_entry_t ent = { .val = 0 };
5752 if (pte_present(ptent))
5753 page = mc_handle_present_pte(vma, addr, ptent);
5754 else if (is_swap_pte(ptent))
5755 page = mc_handle_swap_pte(vma, ptent, &ent);
5756 else if (pte_none(ptent))
5757 page = mc_handle_file_pte(vma, addr, ptent);
5759 if (!page && !ent.val)
5763 * Do only loose check w/o serialization.
5764 * mem_cgroup_move_account() checks the page is valid or
5765 * not under LRU exclusion.
5767 if (page_memcg(page) == mc.from) {
5768 ret = MC_TARGET_PAGE;
5769 if (is_device_private_page(page))
5770 ret = MC_TARGET_DEVICE;
5772 target->page = page;
5774 if (!ret || !target)
5778 * There is a swap entry and a page doesn't exist or isn't charged.
5779 * But we cannot move a tail-page in a THP.
5781 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5782 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5783 ret = MC_TARGET_SWAP;
5790 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5792 * We don't consider PMD mapped swapping or file mapped pages because THP does
5793 * not support them for now.
5794 * Caller should make sure that pmd_trans_huge(pmd) is true.
5796 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5797 unsigned long addr, pmd_t pmd, union mc_target *target)
5799 struct page *page = NULL;
5800 enum mc_target_type ret = MC_TARGET_NONE;
5802 if (unlikely(is_swap_pmd(pmd))) {
5803 VM_BUG_ON(thp_migration_supported() &&
5804 !is_pmd_migration_entry(pmd));
5807 page = pmd_page(pmd);
5808 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5809 if (!(mc.flags & MOVE_ANON))
5811 if (page_memcg(page) == mc.from) {
5812 ret = MC_TARGET_PAGE;
5815 target->page = page;
5821 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5822 unsigned long addr, pmd_t pmd, union mc_target *target)
5824 return MC_TARGET_NONE;
5828 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5829 unsigned long addr, unsigned long end,
5830 struct mm_walk *walk)
5832 struct vm_area_struct *vma = walk->vma;
5836 ptl = pmd_trans_huge_lock(pmd, vma);
5839 * Note their can not be MC_TARGET_DEVICE for now as we do not
5840 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5841 * this might change.
5843 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5844 mc.precharge += HPAGE_PMD_NR;
5849 if (pmd_trans_unstable(pmd))
5851 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5852 for (; addr != end; pte++, addr += PAGE_SIZE)
5853 if (get_mctgt_type(vma, addr, *pte, NULL))
5854 mc.precharge++; /* increment precharge temporarily */
5855 pte_unmap_unlock(pte - 1, ptl);
5861 static const struct mm_walk_ops precharge_walk_ops = {
5862 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5865 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5867 unsigned long precharge;
5870 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5871 mmap_read_unlock(mm);
5873 precharge = mc.precharge;
5879 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5881 unsigned long precharge = mem_cgroup_count_precharge(mm);
5883 VM_BUG_ON(mc.moving_task);
5884 mc.moving_task = current;
5885 return mem_cgroup_do_precharge(precharge);
5888 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5889 static void __mem_cgroup_clear_mc(void)
5891 struct mem_cgroup *from = mc.from;
5892 struct mem_cgroup *to = mc.to;
5894 /* we must uncharge all the leftover precharges from mc.to */
5896 cancel_charge(mc.to, mc.precharge);
5900 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5901 * we must uncharge here.
5903 if (mc.moved_charge) {
5904 cancel_charge(mc.from, mc.moved_charge);
5905 mc.moved_charge = 0;
5907 /* we must fixup refcnts and charges */
5908 if (mc.moved_swap) {
5909 /* uncharge swap account from the old cgroup */
5910 if (!mem_cgroup_is_root(mc.from))
5911 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5913 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5916 * we charged both to->memory and to->memsw, so we
5917 * should uncharge to->memory.
5919 if (!mem_cgroup_is_root(mc.to))
5920 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5924 memcg_oom_recover(from);
5925 memcg_oom_recover(to);
5926 wake_up_all(&mc.waitq);
5929 static void mem_cgroup_clear_mc(void)
5931 struct mm_struct *mm = mc.mm;
5934 * we must clear moving_task before waking up waiters at the end of
5937 mc.moving_task = NULL;
5938 __mem_cgroup_clear_mc();
5939 spin_lock(&mc.lock);
5943 spin_unlock(&mc.lock);
5948 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5950 struct cgroup_subsys_state *css;
5951 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5952 struct mem_cgroup *from;
5953 struct task_struct *leader, *p;
5954 struct mm_struct *mm;
5955 unsigned long move_flags;
5958 /* charge immigration isn't supported on the default hierarchy */
5959 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5963 * Multi-process migrations only happen on the default hierarchy
5964 * where charge immigration is not used. Perform charge
5965 * immigration if @tset contains a leader and whine if there are
5969 cgroup_taskset_for_each_leader(leader, css, tset) {
5972 memcg = mem_cgroup_from_css(css);
5978 * We are now committed to this value whatever it is. Changes in this
5979 * tunable will only affect upcoming migrations, not the current one.
5980 * So we need to save it, and keep it going.
5982 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5986 from = mem_cgroup_from_task(p);
5988 VM_BUG_ON(from == memcg);
5990 mm = get_task_mm(p);
5993 /* We move charges only when we move a owner of the mm */
5994 if (mm->owner == p) {
5997 VM_BUG_ON(mc.precharge);
5998 VM_BUG_ON(mc.moved_charge);
5999 VM_BUG_ON(mc.moved_swap);
6001 spin_lock(&mc.lock);
6005 mc.flags = move_flags;
6006 spin_unlock(&mc.lock);
6007 /* We set mc.moving_task later */
6009 ret = mem_cgroup_precharge_mc(mm);
6011 mem_cgroup_clear_mc();
6018 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6021 mem_cgroup_clear_mc();
6024 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6025 unsigned long addr, unsigned long end,
6026 struct mm_walk *walk)
6029 struct vm_area_struct *vma = walk->vma;
6032 enum mc_target_type target_type;
6033 union mc_target target;
6036 ptl = pmd_trans_huge_lock(pmd, vma);
6038 if (mc.precharge < HPAGE_PMD_NR) {
6042 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6043 if (target_type == MC_TARGET_PAGE) {
6045 if (!isolate_lru_page(page)) {
6046 if (!mem_cgroup_move_account(page, true,
6048 mc.precharge -= HPAGE_PMD_NR;
6049 mc.moved_charge += HPAGE_PMD_NR;
6051 putback_lru_page(page);
6054 } else if (target_type == MC_TARGET_DEVICE) {
6056 if (!mem_cgroup_move_account(page, true,
6058 mc.precharge -= HPAGE_PMD_NR;
6059 mc.moved_charge += HPAGE_PMD_NR;
6067 if (pmd_trans_unstable(pmd))
6070 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6071 for (; addr != end; addr += PAGE_SIZE) {
6072 pte_t ptent = *(pte++);
6073 bool device = false;
6079 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6080 case MC_TARGET_DEVICE:
6083 case MC_TARGET_PAGE:
6086 * We can have a part of the split pmd here. Moving it
6087 * can be done but it would be too convoluted so simply
6088 * ignore such a partial THP and keep it in original
6089 * memcg. There should be somebody mapping the head.
6091 if (PageTransCompound(page))
6093 if (!device && isolate_lru_page(page))
6095 if (!mem_cgroup_move_account(page, false,
6098 /* we uncharge from mc.from later. */
6102 putback_lru_page(page);
6103 put: /* get_mctgt_type() gets the page */
6106 case MC_TARGET_SWAP:
6108 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6110 mem_cgroup_id_get_many(mc.to, 1);
6111 /* we fixup other refcnts and charges later. */
6119 pte_unmap_unlock(pte - 1, ptl);
6124 * We have consumed all precharges we got in can_attach().
6125 * We try charge one by one, but don't do any additional
6126 * charges to mc.to if we have failed in charge once in attach()
6129 ret = mem_cgroup_do_precharge(1);
6137 static const struct mm_walk_ops charge_walk_ops = {
6138 .pmd_entry = mem_cgroup_move_charge_pte_range,
6141 static void mem_cgroup_move_charge(void)
6143 lru_add_drain_all();
6145 * Signal lock_page_memcg() to take the memcg's move_lock
6146 * while we're moving its pages to another memcg. Then wait
6147 * for already started RCU-only updates to finish.
6149 atomic_inc(&mc.from->moving_account);
6152 if (unlikely(!mmap_read_trylock(mc.mm))) {
6154 * Someone who are holding the mmap_lock might be waiting in
6155 * waitq. So we cancel all extra charges, wake up all waiters,
6156 * and retry. Because we cancel precharges, we might not be able
6157 * to move enough charges, but moving charge is a best-effort
6158 * feature anyway, so it wouldn't be a big problem.
6160 __mem_cgroup_clear_mc();
6165 * When we have consumed all precharges and failed in doing
6166 * additional charge, the page walk just aborts.
6168 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6171 mmap_read_unlock(mc.mm);
6172 atomic_dec(&mc.from->moving_account);
6175 static void mem_cgroup_move_task(void)
6178 mem_cgroup_move_charge();
6179 mem_cgroup_clear_mc();
6182 #else /* !CONFIG_MMU */
6183 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6187 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6190 static void mem_cgroup_move_task(void)
6195 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6197 if (value == PAGE_COUNTER_MAX)
6198 seq_puts(m, "max\n");
6200 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6205 static u64 memory_current_read(struct cgroup_subsys_state *css,
6208 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6210 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6213 static int memory_min_show(struct seq_file *m, void *v)
6215 return seq_puts_memcg_tunable(m,
6216 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6219 static ssize_t memory_min_write(struct kernfs_open_file *of,
6220 char *buf, size_t nbytes, loff_t off)
6222 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6226 buf = strstrip(buf);
6227 err = page_counter_memparse(buf, "max", &min);
6231 page_counter_set_min(&memcg->memory, min);
6236 static int memory_low_show(struct seq_file *m, void *v)
6238 return seq_puts_memcg_tunable(m,
6239 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6242 static ssize_t memory_low_write(struct kernfs_open_file *of,
6243 char *buf, size_t nbytes, loff_t off)
6245 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6249 buf = strstrip(buf);
6250 err = page_counter_memparse(buf, "max", &low);
6254 page_counter_set_low(&memcg->memory, low);
6259 static int memory_high_show(struct seq_file *m, void *v)
6261 return seq_puts_memcg_tunable(m,
6262 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6265 static ssize_t memory_high_write(struct kernfs_open_file *of,
6266 char *buf, size_t nbytes, loff_t off)
6268 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6269 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6270 bool drained = false;
6274 buf = strstrip(buf);
6275 err = page_counter_memparse(buf, "max", &high);
6279 page_counter_set_high(&memcg->memory, high);
6282 unsigned long nr_pages = page_counter_read(&memcg->memory);
6283 unsigned long reclaimed;
6285 if (nr_pages <= high)
6288 if (signal_pending(current))
6292 drain_all_stock(memcg);
6297 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6300 if (!reclaimed && !nr_retries--)
6304 memcg_wb_domain_size_changed(memcg);
6308 static int memory_max_show(struct seq_file *m, void *v)
6310 return seq_puts_memcg_tunable(m,
6311 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6314 static ssize_t memory_max_write(struct kernfs_open_file *of,
6315 char *buf, size_t nbytes, loff_t off)
6317 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6318 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6319 bool drained = false;
6323 buf = strstrip(buf);
6324 err = page_counter_memparse(buf, "max", &max);
6328 xchg(&memcg->memory.max, max);
6331 unsigned long nr_pages = page_counter_read(&memcg->memory);
6333 if (nr_pages <= max)
6336 if (signal_pending(current))
6340 drain_all_stock(memcg);
6346 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6352 memcg_memory_event(memcg, MEMCG_OOM);
6353 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6357 memcg_wb_domain_size_changed(memcg);
6361 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6363 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6364 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6365 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6366 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6367 seq_printf(m, "oom_kill %lu\n",
6368 atomic_long_read(&events[MEMCG_OOM_KILL]));
6369 seq_printf(m, "oom_group_kill %lu\n",
6370 atomic_long_read(&events[MEMCG_OOM_GROUP_KILL]));
6373 static int memory_events_show(struct seq_file *m, void *v)
6375 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6377 __memory_events_show(m, memcg->memory_events);
6381 static int memory_events_local_show(struct seq_file *m, void *v)
6383 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6385 __memory_events_show(m, memcg->memory_events_local);
6389 static int memory_stat_show(struct seq_file *m, void *v)
6391 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6394 buf = memory_stat_format(memcg);
6403 static inline unsigned long lruvec_page_state_output(struct lruvec *lruvec,
6406 return lruvec_page_state(lruvec, item) * memcg_page_state_unit(item);
6409 static int memory_numa_stat_show(struct seq_file *m, void *v)
6412 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6414 mem_cgroup_flush_stats();
6416 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6419 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6422 seq_printf(m, "%s", memory_stats[i].name);
6423 for_each_node_state(nid, N_MEMORY) {
6425 struct lruvec *lruvec;
6427 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6428 size = lruvec_page_state_output(lruvec,
6429 memory_stats[i].idx);
6430 seq_printf(m, " N%d=%llu", nid, size);
6439 static int memory_oom_group_show(struct seq_file *m, void *v)
6441 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6443 seq_printf(m, "%d\n", memcg->oom_group);
6448 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6449 char *buf, size_t nbytes, loff_t off)
6451 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6454 buf = strstrip(buf);
6458 ret = kstrtoint(buf, 0, &oom_group);
6462 if (oom_group != 0 && oom_group != 1)
6465 memcg->oom_group = oom_group;
6470 static struct cftype memory_files[] = {
6473 .flags = CFTYPE_NOT_ON_ROOT,
6474 .read_u64 = memory_current_read,
6478 .flags = CFTYPE_NOT_ON_ROOT,
6479 .seq_show = memory_min_show,
6480 .write = memory_min_write,
6484 .flags = CFTYPE_NOT_ON_ROOT,
6485 .seq_show = memory_low_show,
6486 .write = memory_low_write,
6490 .flags = CFTYPE_NOT_ON_ROOT,
6491 .seq_show = memory_high_show,
6492 .write = memory_high_write,
6496 .flags = CFTYPE_NOT_ON_ROOT,
6497 .seq_show = memory_max_show,
6498 .write = memory_max_write,
6502 .flags = CFTYPE_NOT_ON_ROOT,
6503 .file_offset = offsetof(struct mem_cgroup, events_file),
6504 .seq_show = memory_events_show,
6507 .name = "events.local",
6508 .flags = CFTYPE_NOT_ON_ROOT,
6509 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6510 .seq_show = memory_events_local_show,
6514 .seq_show = memory_stat_show,
6518 .name = "numa_stat",
6519 .seq_show = memory_numa_stat_show,
6523 .name = "oom.group",
6524 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6525 .seq_show = memory_oom_group_show,
6526 .write = memory_oom_group_write,
6531 struct cgroup_subsys memory_cgrp_subsys = {
6532 .css_alloc = mem_cgroup_css_alloc,
6533 .css_online = mem_cgroup_css_online,
6534 .css_offline = mem_cgroup_css_offline,
6535 .css_released = mem_cgroup_css_released,
6536 .css_free = mem_cgroup_css_free,
6537 .css_reset = mem_cgroup_css_reset,
6538 .css_rstat_flush = mem_cgroup_css_rstat_flush,
6539 .can_attach = mem_cgroup_can_attach,
6540 .cancel_attach = mem_cgroup_cancel_attach,
6541 .post_attach = mem_cgroup_move_task,
6542 .dfl_cftypes = memory_files,
6543 .legacy_cftypes = mem_cgroup_legacy_files,
6548 * This function calculates an individual cgroup's effective
6549 * protection which is derived from its own memory.min/low, its
6550 * parent's and siblings' settings, as well as the actual memory
6551 * distribution in the tree.
6553 * The following rules apply to the effective protection values:
6555 * 1. At the first level of reclaim, effective protection is equal to
6556 * the declared protection in memory.min and memory.low.
6558 * 2. To enable safe delegation of the protection configuration, at
6559 * subsequent levels the effective protection is capped to the
6560 * parent's effective protection.
6562 * 3. To make complex and dynamic subtrees easier to configure, the
6563 * user is allowed to overcommit the declared protection at a given
6564 * level. If that is the case, the parent's effective protection is
6565 * distributed to the children in proportion to how much protection
6566 * they have declared and how much of it they are utilizing.
6568 * This makes distribution proportional, but also work-conserving:
6569 * if one cgroup claims much more protection than it uses memory,
6570 * the unused remainder is available to its siblings.
6572 * 4. Conversely, when the declared protection is undercommitted at a
6573 * given level, the distribution of the larger parental protection
6574 * budget is NOT proportional. A cgroup's protection from a sibling
6575 * is capped to its own memory.min/low setting.
6577 * 5. However, to allow protecting recursive subtrees from each other
6578 * without having to declare each individual cgroup's fixed share
6579 * of the ancestor's claim to protection, any unutilized -
6580 * "floating" - protection from up the tree is distributed in
6581 * proportion to each cgroup's *usage*. This makes the protection
6582 * neutral wrt sibling cgroups and lets them compete freely over
6583 * the shared parental protection budget, but it protects the
6584 * subtree as a whole from neighboring subtrees.
6586 * Note that 4. and 5. are not in conflict: 4. is about protecting
6587 * against immediate siblings whereas 5. is about protecting against
6588 * neighboring subtrees.
6590 static unsigned long effective_protection(unsigned long usage,
6591 unsigned long parent_usage,
6592 unsigned long setting,
6593 unsigned long parent_effective,
6594 unsigned long siblings_protected)
6596 unsigned long protected;
6599 protected = min(usage, setting);
6601 * If all cgroups at this level combined claim and use more
6602 * protection then what the parent affords them, distribute
6603 * shares in proportion to utilization.
6605 * We are using actual utilization rather than the statically
6606 * claimed protection in order to be work-conserving: claimed
6607 * but unused protection is available to siblings that would
6608 * otherwise get a smaller chunk than what they claimed.
6610 if (siblings_protected > parent_effective)
6611 return protected * parent_effective / siblings_protected;
6614 * Ok, utilized protection of all children is within what the
6615 * parent affords them, so we know whatever this child claims
6616 * and utilizes is effectively protected.
6618 * If there is unprotected usage beyond this value, reclaim
6619 * will apply pressure in proportion to that amount.
6621 * If there is unutilized protection, the cgroup will be fully
6622 * shielded from reclaim, but we do return a smaller value for
6623 * protection than what the group could enjoy in theory. This
6624 * is okay. With the overcommit distribution above, effective
6625 * protection is always dependent on how memory is actually
6626 * consumed among the siblings anyway.
6631 * If the children aren't claiming (all of) the protection
6632 * afforded to them by the parent, distribute the remainder in
6633 * proportion to the (unprotected) memory of each cgroup. That
6634 * way, cgroups that aren't explicitly prioritized wrt each
6635 * other compete freely over the allowance, but they are
6636 * collectively protected from neighboring trees.
6638 * We're using unprotected memory for the weight so that if
6639 * some cgroups DO claim explicit protection, we don't protect
6640 * the same bytes twice.
6642 * Check both usage and parent_usage against the respective
6643 * protected values. One should imply the other, but they
6644 * aren't read atomically - make sure the division is sane.
6646 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6648 if (parent_effective > siblings_protected &&
6649 parent_usage > siblings_protected &&
6650 usage > protected) {
6651 unsigned long unclaimed;
6653 unclaimed = parent_effective - siblings_protected;
6654 unclaimed *= usage - protected;
6655 unclaimed /= parent_usage - siblings_protected;
6664 * mem_cgroup_calculate_protection - check if memory consumption is in the normal range
6665 * @root: the top ancestor of the sub-tree being checked
6666 * @memcg: the memory cgroup to check
6668 * WARNING: This function is not stateless! It can only be used as part
6669 * of a top-down tree iteration, not for isolated queries.
6671 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6672 struct mem_cgroup *memcg)
6674 unsigned long usage, parent_usage;
6675 struct mem_cgroup *parent;
6677 if (mem_cgroup_disabled())
6681 root = root_mem_cgroup;
6684 * Effective values of the reclaim targets are ignored so they
6685 * can be stale. Have a look at mem_cgroup_protection for more
6687 * TODO: calculation should be more robust so that we do not need
6688 * that special casing.
6693 usage = page_counter_read(&memcg->memory);
6697 parent = parent_mem_cgroup(memcg);
6698 /* No parent means a non-hierarchical mode on v1 memcg */
6702 if (parent == root) {
6703 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6704 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6708 parent_usage = page_counter_read(&parent->memory);
6710 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6711 READ_ONCE(memcg->memory.min),
6712 READ_ONCE(parent->memory.emin),
6713 atomic_long_read(&parent->memory.children_min_usage)));
6715 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6716 READ_ONCE(memcg->memory.low),
6717 READ_ONCE(parent->memory.elow),
6718 atomic_long_read(&parent->memory.children_low_usage)));
6721 static int charge_memcg(struct folio *folio, struct mem_cgroup *memcg,
6724 long nr_pages = folio_nr_pages(folio);
6727 ret = try_charge(memcg, gfp, nr_pages);
6731 css_get(&memcg->css);
6732 commit_charge(folio, memcg);
6734 local_irq_disable();
6735 mem_cgroup_charge_statistics(memcg, nr_pages);
6736 memcg_check_events(memcg, folio_nid(folio));
6742 int __mem_cgroup_charge(struct folio *folio, struct mm_struct *mm, gfp_t gfp)
6744 struct mem_cgroup *memcg;
6747 memcg = get_mem_cgroup_from_mm(mm);
6748 ret = charge_memcg(folio, memcg, gfp);
6749 css_put(&memcg->css);
6755 * mem_cgroup_swapin_charge_page - charge a newly allocated page for swapin
6756 * @page: page to charge
6757 * @mm: mm context of the victim
6758 * @gfp: reclaim mode
6759 * @entry: swap entry for which the page is allocated
6761 * This function charges a page allocated for swapin. Please call this before
6762 * adding the page to the swapcache.
6764 * Returns 0 on success. Otherwise, an error code is returned.
6766 int mem_cgroup_swapin_charge_page(struct page *page, struct mm_struct *mm,
6767 gfp_t gfp, swp_entry_t entry)
6769 struct folio *folio = page_folio(page);
6770 struct mem_cgroup *memcg;
6774 if (mem_cgroup_disabled())
6777 id = lookup_swap_cgroup_id(entry);
6779 memcg = mem_cgroup_from_id(id);
6780 if (!memcg || !css_tryget_online(&memcg->css))
6781 memcg = get_mem_cgroup_from_mm(mm);
6784 ret = charge_memcg(folio, memcg, gfp);
6786 css_put(&memcg->css);
6791 * mem_cgroup_swapin_uncharge_swap - uncharge swap slot
6792 * @entry: swap entry for which the page is charged
6794 * Call this function after successfully adding the charged page to swapcache.
6796 * Note: This function assumes the page for which swap slot is being uncharged
6799 void mem_cgroup_swapin_uncharge_swap(swp_entry_t entry)
6802 * Cgroup1's unified memory+swap counter has been charged with the
6803 * new swapcache page, finish the transfer by uncharging the swap
6804 * slot. The swap slot would also get uncharged when it dies, but
6805 * it can stick around indefinitely and we'd count the page twice
6808 * Cgroup2 has separate resource counters for memory and swap,
6809 * so this is a non-issue here. Memory and swap charge lifetimes
6810 * correspond 1:1 to page and swap slot lifetimes: we charge the
6811 * page to memory here, and uncharge swap when the slot is freed.
6813 if (!mem_cgroup_disabled() && do_memsw_account()) {
6815 * The swap entry might not get freed for a long time,
6816 * let's not wait for it. The page already received a
6817 * memory+swap charge, drop the swap entry duplicate.
6819 mem_cgroup_uncharge_swap(entry, 1);
6823 struct uncharge_gather {
6824 struct mem_cgroup *memcg;
6825 unsigned long nr_memory;
6826 unsigned long pgpgout;
6827 unsigned long nr_kmem;
6831 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6833 memset(ug, 0, sizeof(*ug));
6836 static void uncharge_batch(const struct uncharge_gather *ug)
6838 unsigned long flags;
6840 if (ug->nr_memory) {
6841 page_counter_uncharge(&ug->memcg->memory, ug->nr_memory);
6842 if (do_memsw_account())
6843 page_counter_uncharge(&ug->memcg->memsw, ug->nr_memory);
6845 memcg_account_kmem(ug->memcg, -ug->nr_kmem);
6846 memcg_oom_recover(ug->memcg);
6849 local_irq_save(flags);
6850 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6851 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_memory);
6852 memcg_check_events(ug->memcg, ug->nid);
6853 local_irq_restore(flags);
6855 /* drop reference from uncharge_folio */
6856 css_put(&ug->memcg->css);
6859 static void uncharge_folio(struct folio *folio, struct uncharge_gather *ug)
6862 struct mem_cgroup *memcg;
6863 struct obj_cgroup *objcg;
6865 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
6868 * Nobody should be changing or seriously looking at
6869 * folio memcg or objcg at this point, we have fully
6870 * exclusive access to the folio.
6872 if (folio_memcg_kmem(folio)) {
6873 objcg = __folio_objcg(folio);
6875 * This get matches the put at the end of the function and
6876 * kmem pages do not hold memcg references anymore.
6878 memcg = get_mem_cgroup_from_objcg(objcg);
6880 memcg = __folio_memcg(folio);
6886 if (ug->memcg != memcg) {
6889 uncharge_gather_clear(ug);
6892 ug->nid = folio_nid(folio);
6894 /* pairs with css_put in uncharge_batch */
6895 css_get(&memcg->css);
6898 nr_pages = folio_nr_pages(folio);
6900 if (folio_memcg_kmem(folio)) {
6901 ug->nr_memory += nr_pages;
6902 ug->nr_kmem += nr_pages;
6904 folio->memcg_data = 0;
6905 obj_cgroup_put(objcg);
6907 /* LRU pages aren't accounted at the root level */
6908 if (!mem_cgroup_is_root(memcg))
6909 ug->nr_memory += nr_pages;
6912 folio->memcg_data = 0;
6915 css_put(&memcg->css);
6918 void __mem_cgroup_uncharge(struct folio *folio)
6920 struct uncharge_gather ug;
6922 /* Don't touch folio->lru of any random page, pre-check: */
6923 if (!folio_memcg(folio))
6926 uncharge_gather_clear(&ug);
6927 uncharge_folio(folio, &ug);
6928 uncharge_batch(&ug);
6932 * __mem_cgroup_uncharge_list - uncharge a list of page
6933 * @page_list: list of pages to uncharge
6935 * Uncharge a list of pages previously charged with
6936 * __mem_cgroup_charge().
6938 void __mem_cgroup_uncharge_list(struct list_head *page_list)
6940 struct uncharge_gather ug;
6941 struct folio *folio;
6943 uncharge_gather_clear(&ug);
6944 list_for_each_entry(folio, page_list, lru)
6945 uncharge_folio(folio, &ug);
6947 uncharge_batch(&ug);
6951 * mem_cgroup_migrate - Charge a folio's replacement.
6952 * @old: Currently circulating folio.
6953 * @new: Replacement folio.
6955 * Charge @new as a replacement folio for @old. @old will
6956 * be uncharged upon free.
6958 * Both folios must be locked, @new->mapping must be set up.
6960 void mem_cgroup_migrate(struct folio *old, struct folio *new)
6962 struct mem_cgroup *memcg;
6963 long nr_pages = folio_nr_pages(new);
6964 unsigned long flags;
6966 VM_BUG_ON_FOLIO(!folio_test_locked(old), old);
6967 VM_BUG_ON_FOLIO(!folio_test_locked(new), new);
6968 VM_BUG_ON_FOLIO(folio_test_anon(old) != folio_test_anon(new), new);
6969 VM_BUG_ON_FOLIO(folio_nr_pages(old) != nr_pages, new);
6971 if (mem_cgroup_disabled())
6974 /* Page cache replacement: new folio already charged? */
6975 if (folio_memcg(new))
6978 memcg = folio_memcg(old);
6979 VM_WARN_ON_ONCE_FOLIO(!memcg, old);
6983 /* Force-charge the new page. The old one will be freed soon */
6984 if (!mem_cgroup_is_root(memcg)) {
6985 page_counter_charge(&memcg->memory, nr_pages);
6986 if (do_memsw_account())
6987 page_counter_charge(&memcg->memsw, nr_pages);
6990 css_get(&memcg->css);
6991 commit_charge(new, memcg);
6993 local_irq_save(flags);
6994 mem_cgroup_charge_statistics(memcg, nr_pages);
6995 memcg_check_events(memcg, folio_nid(new));
6996 local_irq_restore(flags);
6999 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
7000 EXPORT_SYMBOL(memcg_sockets_enabled_key);
7002 void mem_cgroup_sk_alloc(struct sock *sk)
7004 struct mem_cgroup *memcg;
7006 if (!mem_cgroup_sockets_enabled)
7009 /* Do not associate the sock with unrelated interrupted task's memcg. */
7014 memcg = mem_cgroup_from_task(current);
7015 if (memcg == root_mem_cgroup)
7017 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
7019 if (css_tryget(&memcg->css))
7020 sk->sk_memcg = memcg;
7025 void mem_cgroup_sk_free(struct sock *sk)
7028 css_put(&sk->sk_memcg->css);
7032 * mem_cgroup_charge_skmem - charge socket memory
7033 * @memcg: memcg to charge
7034 * @nr_pages: number of pages to charge
7035 * @gfp_mask: reclaim mode
7037 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7038 * @memcg's configured limit, %false if it doesn't.
7040 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages,
7043 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7044 struct page_counter *fail;
7046 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7047 memcg->tcpmem_pressure = 0;
7050 memcg->tcpmem_pressure = 1;
7051 if (gfp_mask & __GFP_NOFAIL) {
7052 page_counter_charge(&memcg->tcpmem, nr_pages);
7058 if (try_charge(memcg, gfp_mask, nr_pages) == 0) {
7059 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7067 * mem_cgroup_uncharge_skmem - uncharge socket memory
7068 * @memcg: memcg to uncharge
7069 * @nr_pages: number of pages to uncharge
7071 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7073 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7074 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7078 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7080 refill_stock(memcg, nr_pages);
7083 static int __init cgroup_memory(char *s)
7087 while ((token = strsep(&s, ",")) != NULL) {
7090 if (!strcmp(token, "nosocket"))
7091 cgroup_memory_nosocket = true;
7092 if (!strcmp(token, "nokmem"))
7093 cgroup_memory_nokmem = true;
7097 __setup("cgroup.memory=", cgroup_memory);
7100 * subsys_initcall() for memory controller.
7102 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7103 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7104 * basically everything that doesn't depend on a specific mem_cgroup structure
7105 * should be initialized from here.
7107 static int __init mem_cgroup_init(void)
7112 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7113 * used for per-memcg-per-cpu caching of per-node statistics. In order
7114 * to work fine, we should make sure that the overfill threshold can't
7115 * exceed S32_MAX / PAGE_SIZE.
7117 BUILD_BUG_ON(MEMCG_CHARGE_BATCH > S32_MAX / PAGE_SIZE);
7119 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7120 memcg_hotplug_cpu_dead);
7122 for_each_possible_cpu(cpu)
7123 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7126 for_each_node(node) {
7127 struct mem_cgroup_tree_per_node *rtpn;
7129 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7130 node_online(node) ? node : NUMA_NO_NODE);
7132 rtpn->rb_root = RB_ROOT;
7133 rtpn->rb_rightmost = NULL;
7134 spin_lock_init(&rtpn->lock);
7135 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7140 subsys_initcall(mem_cgroup_init);
7142 #ifdef CONFIG_MEMCG_SWAP
7143 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7145 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7147 * The root cgroup cannot be destroyed, so it's refcount must
7150 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7154 memcg = parent_mem_cgroup(memcg);
7156 memcg = root_mem_cgroup;
7162 * mem_cgroup_swapout - transfer a memsw charge to swap
7163 * @page: page whose memsw charge to transfer
7164 * @entry: swap entry to move the charge to
7166 * Transfer the memsw charge of @page to @entry.
7168 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7170 struct mem_cgroup *memcg, *swap_memcg;
7171 unsigned int nr_entries;
7172 unsigned short oldid;
7174 VM_BUG_ON_PAGE(PageLRU(page), page);
7175 VM_BUG_ON_PAGE(page_count(page), page);
7177 if (mem_cgroup_disabled())
7180 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7183 memcg = page_memcg(page);
7185 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7190 * In case the memcg owning these pages has been offlined and doesn't
7191 * have an ID allocated to it anymore, charge the closest online
7192 * ancestor for the swap instead and transfer the memory+swap charge.
7194 swap_memcg = mem_cgroup_id_get_online(memcg);
7195 nr_entries = thp_nr_pages(page);
7196 /* Get references for the tail pages, too */
7198 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7199 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7201 VM_BUG_ON_PAGE(oldid, page);
7202 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7204 page->memcg_data = 0;
7206 if (!mem_cgroup_is_root(memcg))
7207 page_counter_uncharge(&memcg->memory, nr_entries);
7209 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7210 if (!mem_cgroup_is_root(swap_memcg))
7211 page_counter_charge(&swap_memcg->memsw, nr_entries);
7212 page_counter_uncharge(&memcg->memsw, nr_entries);
7216 * Interrupts should be disabled here because the caller holds the
7217 * i_pages lock which is taken with interrupts-off. It is
7218 * important here to have the interrupts disabled because it is the
7219 * only synchronisation we have for updating the per-CPU variables.
7222 mem_cgroup_charge_statistics(memcg, -nr_entries);
7223 memcg_stats_unlock();
7224 memcg_check_events(memcg, page_to_nid(page));
7226 css_put(&memcg->css);
7230 * __mem_cgroup_try_charge_swap - try charging swap space for a page
7231 * @page: page being added to swap
7232 * @entry: swap entry to charge
7234 * Try to charge @page's memcg for the swap space at @entry.
7236 * Returns 0 on success, -ENOMEM on failure.
7238 int __mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7240 unsigned int nr_pages = thp_nr_pages(page);
7241 struct page_counter *counter;
7242 struct mem_cgroup *memcg;
7243 unsigned short oldid;
7245 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7248 memcg = page_memcg(page);
7250 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7255 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7259 memcg = mem_cgroup_id_get_online(memcg);
7261 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7262 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7263 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7264 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7265 mem_cgroup_id_put(memcg);
7269 /* Get references for the tail pages, too */
7271 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7272 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7273 VM_BUG_ON_PAGE(oldid, page);
7274 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7280 * __mem_cgroup_uncharge_swap - uncharge swap space
7281 * @entry: swap entry to uncharge
7282 * @nr_pages: the amount of swap space to uncharge
7284 void __mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7286 struct mem_cgroup *memcg;
7289 id = swap_cgroup_record(entry, 0, nr_pages);
7291 memcg = mem_cgroup_from_id(id);
7293 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7294 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7295 page_counter_uncharge(&memcg->swap, nr_pages);
7297 page_counter_uncharge(&memcg->memsw, nr_pages);
7299 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7300 mem_cgroup_id_put_many(memcg, nr_pages);
7305 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7307 long nr_swap_pages = get_nr_swap_pages();
7309 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7310 return nr_swap_pages;
7311 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7312 nr_swap_pages = min_t(long, nr_swap_pages,
7313 READ_ONCE(memcg->swap.max) -
7314 page_counter_read(&memcg->swap));
7315 return nr_swap_pages;
7318 bool mem_cgroup_swap_full(struct page *page)
7320 struct mem_cgroup *memcg;
7322 VM_BUG_ON_PAGE(!PageLocked(page), page);
7326 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7329 memcg = page_memcg(page);
7333 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7334 unsigned long usage = page_counter_read(&memcg->swap);
7336 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7337 usage * 2 >= READ_ONCE(memcg->swap.max))
7344 static int __init setup_swap_account(char *s)
7346 if (!strcmp(s, "1"))
7347 cgroup_memory_noswap = false;
7348 else if (!strcmp(s, "0"))
7349 cgroup_memory_noswap = true;
7352 __setup("swapaccount=", setup_swap_account);
7354 static u64 swap_current_read(struct cgroup_subsys_state *css,
7357 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7359 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7362 static int swap_high_show(struct seq_file *m, void *v)
7364 return seq_puts_memcg_tunable(m,
7365 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7368 static ssize_t swap_high_write(struct kernfs_open_file *of,
7369 char *buf, size_t nbytes, loff_t off)
7371 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7375 buf = strstrip(buf);
7376 err = page_counter_memparse(buf, "max", &high);
7380 page_counter_set_high(&memcg->swap, high);
7385 static int swap_max_show(struct seq_file *m, void *v)
7387 return seq_puts_memcg_tunable(m,
7388 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7391 static ssize_t swap_max_write(struct kernfs_open_file *of,
7392 char *buf, size_t nbytes, loff_t off)
7394 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7398 buf = strstrip(buf);
7399 err = page_counter_memparse(buf, "max", &max);
7403 xchg(&memcg->swap.max, max);
7408 static int swap_events_show(struct seq_file *m, void *v)
7410 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7412 seq_printf(m, "high %lu\n",
7413 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7414 seq_printf(m, "max %lu\n",
7415 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7416 seq_printf(m, "fail %lu\n",
7417 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7422 static struct cftype swap_files[] = {
7424 .name = "swap.current",
7425 .flags = CFTYPE_NOT_ON_ROOT,
7426 .read_u64 = swap_current_read,
7429 .name = "swap.high",
7430 .flags = CFTYPE_NOT_ON_ROOT,
7431 .seq_show = swap_high_show,
7432 .write = swap_high_write,
7436 .flags = CFTYPE_NOT_ON_ROOT,
7437 .seq_show = swap_max_show,
7438 .write = swap_max_write,
7441 .name = "swap.events",
7442 .flags = CFTYPE_NOT_ON_ROOT,
7443 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7444 .seq_show = swap_events_show,
7449 static struct cftype memsw_files[] = {
7451 .name = "memsw.usage_in_bytes",
7452 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7453 .read_u64 = mem_cgroup_read_u64,
7456 .name = "memsw.max_usage_in_bytes",
7457 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7458 .write = mem_cgroup_reset,
7459 .read_u64 = mem_cgroup_read_u64,
7462 .name = "memsw.limit_in_bytes",
7463 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7464 .write = mem_cgroup_write,
7465 .read_u64 = mem_cgroup_read_u64,
7468 .name = "memsw.failcnt",
7469 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7470 .write = mem_cgroup_reset,
7471 .read_u64 = mem_cgroup_read_u64,
7473 { }, /* terminate */
7477 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7478 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7479 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7480 * boot parameter. This may result in premature OOPS inside
7481 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7483 static int __init mem_cgroup_swap_init(void)
7485 /* No memory control -> no swap control */
7486 if (mem_cgroup_disabled())
7487 cgroup_memory_noswap = true;
7489 if (cgroup_memory_noswap)
7492 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7493 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7497 core_initcall(mem_cgroup_swap_init);
7499 #endif /* CONFIG_MEMCG_SWAP */