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 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 /* memcg and lruvec stats flushing */
107 static void flush_memcg_stats_dwork(struct work_struct *w);
108 static DECLARE_DEFERRABLE_WORK(stats_flush_dwork, flush_memcg_stats_dwork);
109 static void flush_memcg_stats_work(struct work_struct *w);
110 static DECLARE_WORK(stats_flush_work, flush_memcg_stats_work);
111 static DEFINE_PER_CPU(unsigned int, stats_flush_threshold);
112 static DEFINE_SPINLOCK(stats_flush_lock);
114 #define THRESHOLDS_EVENTS_TARGET 128
115 #define SOFTLIMIT_EVENTS_TARGET 1024
118 * Cgroups above their limits are maintained in a RB-Tree, independent of
119 * their hierarchy representation
122 struct mem_cgroup_tree_per_node {
123 struct rb_root rb_root;
124 struct rb_node *rb_rightmost;
128 struct mem_cgroup_tree {
129 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
132 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
135 struct mem_cgroup_eventfd_list {
136 struct list_head list;
137 struct eventfd_ctx *eventfd;
141 * cgroup_event represents events which userspace want to receive.
143 struct mem_cgroup_event {
145 * memcg which the event belongs to.
147 struct mem_cgroup *memcg;
149 * eventfd to signal userspace about the event.
151 struct eventfd_ctx *eventfd;
153 * Each of these stored in a list by the cgroup.
155 struct list_head list;
157 * register_event() callback will be used to add new userspace
158 * waiter for changes related to this event. Use eventfd_signal()
159 * on eventfd to send notification to userspace.
161 int (*register_event)(struct mem_cgroup *memcg,
162 struct eventfd_ctx *eventfd, const char *args);
164 * unregister_event() callback will be called when userspace closes
165 * the eventfd or on cgroup removing. This callback must be set,
166 * if you want provide notification functionality.
168 void (*unregister_event)(struct mem_cgroup *memcg,
169 struct eventfd_ctx *eventfd);
171 * All fields below needed to unregister event when
172 * userspace closes eventfd.
175 wait_queue_head_t *wqh;
176 wait_queue_entry_t wait;
177 struct work_struct remove;
180 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
181 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
183 /* Stuffs for move charges at task migration. */
185 * Types of charges to be moved.
187 #define MOVE_ANON 0x1U
188 #define MOVE_FILE 0x2U
189 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
191 /* "mc" and its members are protected by cgroup_mutex */
192 static struct move_charge_struct {
193 spinlock_t lock; /* for from, to */
194 struct mm_struct *mm;
195 struct mem_cgroup *from;
196 struct mem_cgroup *to;
198 unsigned long precharge;
199 unsigned long moved_charge;
200 unsigned long moved_swap;
201 struct task_struct *moving_task; /* a task moving charges */
202 wait_queue_head_t waitq; /* a waitq for other context */
204 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
205 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
209 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
210 * limit reclaim to prevent infinite loops, if they ever occur.
212 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
213 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
215 /* for encoding cft->private value on file */
224 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
225 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
226 #define MEMFILE_ATTR(val) ((val) & 0xffff)
227 /* Used for OOM notifier */
228 #define OOM_CONTROL (0)
231 * Iteration constructs for visiting all cgroups (under a tree). If
232 * loops are exited prematurely (break), mem_cgroup_iter_break() must
233 * be used for reference counting.
235 #define for_each_mem_cgroup_tree(iter, root) \
236 for (iter = mem_cgroup_iter(root, NULL, NULL); \
238 iter = mem_cgroup_iter(root, iter, NULL))
240 #define for_each_mem_cgroup(iter) \
241 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
243 iter = mem_cgroup_iter(NULL, iter, NULL))
245 static inline bool should_force_charge(void)
247 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
248 (current->flags & PF_EXITING);
251 /* Some nice accessors for the vmpressure. */
252 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
255 memcg = root_mem_cgroup;
256 return &memcg->vmpressure;
259 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
261 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
264 #ifdef CONFIG_MEMCG_KMEM
265 extern spinlock_t css_set_lock;
267 bool mem_cgroup_kmem_disabled(void)
269 return cgroup_memory_nokmem;
272 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
273 unsigned int nr_pages);
275 static void obj_cgroup_release(struct percpu_ref *ref)
277 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
278 unsigned int nr_bytes;
279 unsigned int nr_pages;
283 * At this point all allocated objects are freed, and
284 * objcg->nr_charged_bytes can't have an arbitrary byte value.
285 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
287 * The following sequence can lead to it:
288 * 1) CPU0: objcg == stock->cached_objcg
289 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
290 * PAGE_SIZE bytes are charged
291 * 3) CPU1: a process from another memcg is allocating something,
292 * the stock if flushed,
293 * objcg->nr_charged_bytes = PAGE_SIZE - 92
294 * 5) CPU0: we do release this object,
295 * 92 bytes are added to stock->nr_bytes
296 * 6) CPU0: stock is flushed,
297 * 92 bytes are added to objcg->nr_charged_bytes
299 * In the result, nr_charged_bytes == PAGE_SIZE.
300 * This page will be uncharged in obj_cgroup_release().
302 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
303 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
304 nr_pages = nr_bytes >> PAGE_SHIFT;
307 obj_cgroup_uncharge_pages(objcg, nr_pages);
309 spin_lock_irqsave(&css_set_lock, flags);
310 list_del(&objcg->list);
311 spin_unlock_irqrestore(&css_set_lock, flags);
313 percpu_ref_exit(ref);
314 kfree_rcu(objcg, rcu);
317 static struct obj_cgroup *obj_cgroup_alloc(void)
319 struct obj_cgroup *objcg;
322 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
326 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
332 INIT_LIST_HEAD(&objcg->list);
336 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
337 struct mem_cgroup *parent)
339 struct obj_cgroup *objcg, *iter;
341 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
343 spin_lock_irq(&css_set_lock);
345 /* 1) Ready to reparent active objcg. */
346 list_add(&objcg->list, &memcg->objcg_list);
347 /* 2) Reparent active objcg and already reparented objcgs to parent. */
348 list_for_each_entry(iter, &memcg->objcg_list, list)
349 WRITE_ONCE(iter->memcg, parent);
350 /* 3) Move already reparented objcgs to the parent's list */
351 list_splice(&memcg->objcg_list, &parent->objcg_list);
353 spin_unlock_irq(&css_set_lock);
355 percpu_ref_kill(&objcg->refcnt);
359 * This will be used as a shrinker list's index.
360 * The main reason for not using cgroup id for this:
361 * this works better in sparse environments, where we have a lot of memcgs,
362 * but only a few kmem-limited. Or also, if we have, for instance, 200
363 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
364 * 200 entry array for that.
366 * The current size of the caches array is stored in memcg_nr_cache_ids. It
367 * will double each time we have to increase it.
369 static DEFINE_IDA(memcg_cache_ida);
370 int memcg_nr_cache_ids;
372 /* Protects memcg_nr_cache_ids */
373 static DECLARE_RWSEM(memcg_cache_ids_sem);
375 void memcg_get_cache_ids(void)
377 down_read(&memcg_cache_ids_sem);
380 void memcg_put_cache_ids(void)
382 up_read(&memcg_cache_ids_sem);
386 * MIN_SIZE is different than 1, because we would like to avoid going through
387 * the alloc/free process all the time. In a small machine, 4 kmem-limited
388 * cgroups is a reasonable guess. In the future, it could be a parameter or
389 * tunable, but that is strictly not necessary.
391 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
392 * this constant directly from cgroup, but it is understandable that this is
393 * better kept as an internal representation in cgroup.c. In any case, the
394 * cgrp_id space is not getting any smaller, and we don't have to necessarily
395 * increase ours as well if it increases.
397 #define MEMCG_CACHES_MIN_SIZE 4
398 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
401 * A lot of the calls to the cache allocation functions are expected to be
402 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
403 * conditional to this static branch, we'll have to allow modules that does
404 * kmem_cache_alloc and the such to see this symbol as well
406 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
407 EXPORT_SYMBOL(memcg_kmem_enabled_key);
411 * mem_cgroup_css_from_page - css of the memcg associated with a page
412 * @page: page of interest
414 * If memcg is bound to the default hierarchy, css of the memcg associated
415 * with @page is returned. The returned css remains associated with @page
416 * until it is released.
418 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
421 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
423 struct mem_cgroup *memcg;
425 memcg = page_memcg(page);
427 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
428 memcg = root_mem_cgroup;
434 * page_cgroup_ino - return inode number of the memcg a page is charged to
437 * Look up the closest online ancestor of the memory cgroup @page is charged to
438 * and return its inode number or 0 if @page is not charged to any cgroup. It
439 * is safe to call this function without holding a reference to @page.
441 * Note, this function is inherently racy, because there is nothing to prevent
442 * the cgroup inode from getting torn down and potentially reallocated a moment
443 * after page_cgroup_ino() returns, so it only should be used by callers that
444 * do not care (such as procfs interfaces).
446 ino_t page_cgroup_ino(struct page *page)
448 struct mem_cgroup *memcg;
449 unsigned long ino = 0;
452 memcg = page_memcg_check(page);
454 while (memcg && !(memcg->css.flags & CSS_ONLINE))
455 memcg = parent_mem_cgroup(memcg);
457 ino = cgroup_ino(memcg->css.cgroup);
462 static struct mem_cgroup_per_node *
463 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
465 int nid = page_to_nid(page);
467 return memcg->nodeinfo[nid];
470 static struct mem_cgroup_tree_per_node *
471 soft_limit_tree_node(int nid)
473 return soft_limit_tree.rb_tree_per_node[nid];
476 static struct mem_cgroup_tree_per_node *
477 soft_limit_tree_from_page(struct page *page)
479 int nid = page_to_nid(page);
481 return soft_limit_tree.rb_tree_per_node[nid];
484 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
485 struct mem_cgroup_tree_per_node *mctz,
486 unsigned long new_usage_in_excess)
488 struct rb_node **p = &mctz->rb_root.rb_node;
489 struct rb_node *parent = NULL;
490 struct mem_cgroup_per_node *mz_node;
491 bool rightmost = true;
496 mz->usage_in_excess = new_usage_in_excess;
497 if (!mz->usage_in_excess)
501 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
503 if (mz->usage_in_excess < mz_node->usage_in_excess) {
512 mctz->rb_rightmost = &mz->tree_node;
514 rb_link_node(&mz->tree_node, parent, p);
515 rb_insert_color(&mz->tree_node, &mctz->rb_root);
519 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
520 struct mem_cgroup_tree_per_node *mctz)
525 if (&mz->tree_node == mctz->rb_rightmost)
526 mctz->rb_rightmost = rb_prev(&mz->tree_node);
528 rb_erase(&mz->tree_node, &mctz->rb_root);
532 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
533 struct mem_cgroup_tree_per_node *mctz)
537 spin_lock_irqsave(&mctz->lock, flags);
538 __mem_cgroup_remove_exceeded(mz, mctz);
539 spin_unlock_irqrestore(&mctz->lock, flags);
542 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
544 unsigned long nr_pages = page_counter_read(&memcg->memory);
545 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
546 unsigned long excess = 0;
548 if (nr_pages > soft_limit)
549 excess = nr_pages - soft_limit;
554 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
556 unsigned long excess;
557 struct mem_cgroup_per_node *mz;
558 struct mem_cgroup_tree_per_node *mctz;
560 mctz = soft_limit_tree_from_page(page);
564 * Necessary to update all ancestors when hierarchy is used.
565 * because their event counter is not touched.
567 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
568 mz = mem_cgroup_page_nodeinfo(memcg, page);
569 excess = soft_limit_excess(memcg);
571 * We have to update the tree if mz is on RB-tree or
572 * mem is over its softlimit.
574 if (excess || mz->on_tree) {
577 spin_lock_irqsave(&mctz->lock, flags);
578 /* if on-tree, remove it */
580 __mem_cgroup_remove_exceeded(mz, mctz);
582 * Insert again. mz->usage_in_excess will be updated.
583 * If excess is 0, no tree ops.
585 __mem_cgroup_insert_exceeded(mz, mctz, excess);
586 spin_unlock_irqrestore(&mctz->lock, flags);
591 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
593 struct mem_cgroup_tree_per_node *mctz;
594 struct mem_cgroup_per_node *mz;
598 mz = memcg->nodeinfo[nid];
599 mctz = soft_limit_tree_node(nid);
601 mem_cgroup_remove_exceeded(mz, mctz);
605 static struct mem_cgroup_per_node *
606 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
608 struct mem_cgroup_per_node *mz;
612 if (!mctz->rb_rightmost)
613 goto done; /* Nothing to reclaim from */
615 mz = rb_entry(mctz->rb_rightmost,
616 struct mem_cgroup_per_node, tree_node);
618 * Remove the node now but someone else can add it back,
619 * we will to add it back at the end of reclaim to its correct
620 * position in the tree.
622 __mem_cgroup_remove_exceeded(mz, mctz);
623 if (!soft_limit_excess(mz->memcg) ||
624 !css_tryget(&mz->memcg->css))
630 static struct mem_cgroup_per_node *
631 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
633 struct mem_cgroup_per_node *mz;
635 spin_lock_irq(&mctz->lock);
636 mz = __mem_cgroup_largest_soft_limit_node(mctz);
637 spin_unlock_irq(&mctz->lock);
642 * __mod_memcg_state - update cgroup memory statistics
643 * @memcg: the memory cgroup
644 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
645 * @val: delta to add to the counter, can be negative
647 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
649 if (mem_cgroup_disabled())
652 __this_cpu_add(memcg->vmstats_percpu->state[idx], val);
653 cgroup_rstat_updated(memcg->css.cgroup, smp_processor_id());
656 /* idx can be of type enum memcg_stat_item or node_stat_item. */
657 static unsigned long memcg_page_state_local(struct mem_cgroup *memcg, int idx)
662 for_each_possible_cpu(cpu)
663 x += per_cpu(memcg->vmstats_percpu->state[idx], cpu);
671 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
674 struct mem_cgroup_per_node *pn;
675 struct mem_cgroup *memcg;
677 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
681 __mod_memcg_state(memcg, idx, val);
684 __this_cpu_add(pn->lruvec_stats_percpu->state[idx], val);
685 if (!(__this_cpu_inc_return(stats_flush_threshold) % MEMCG_CHARGE_BATCH))
686 queue_work(system_unbound_wq, &stats_flush_work);
690 * __mod_lruvec_state - update lruvec memory statistics
691 * @lruvec: the lruvec
692 * @idx: the stat item
693 * @val: delta to add to the counter, can be negative
695 * The lruvec is the intersection of the NUMA node and a cgroup. This
696 * function updates the all three counters that are affected by a
697 * change of state at this level: per-node, per-cgroup, per-lruvec.
699 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
703 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
705 /* Update memcg and lruvec */
706 if (!mem_cgroup_disabled())
707 __mod_memcg_lruvec_state(lruvec, idx, val);
710 void __mod_lruvec_page_state(struct page *page, enum node_stat_item idx,
713 struct page *head = compound_head(page); /* rmap on tail pages */
714 struct mem_cgroup *memcg;
715 pg_data_t *pgdat = page_pgdat(page);
716 struct lruvec *lruvec;
719 memcg = page_memcg(head);
720 /* Untracked pages have no memcg, no lruvec. Update only the node */
723 __mod_node_page_state(pgdat, idx, val);
727 lruvec = mem_cgroup_lruvec(memcg, pgdat);
728 __mod_lruvec_state(lruvec, idx, val);
731 EXPORT_SYMBOL(__mod_lruvec_page_state);
733 void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val)
735 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
736 struct mem_cgroup *memcg;
737 struct lruvec *lruvec;
740 memcg = mem_cgroup_from_obj(p);
743 * Untracked pages have no memcg, no lruvec. Update only the
744 * node. If we reparent the slab objects to the root memcg,
745 * when we free the slab object, we need to update the per-memcg
746 * vmstats to keep it correct for the root memcg.
749 __mod_node_page_state(pgdat, idx, val);
751 lruvec = mem_cgroup_lruvec(memcg, pgdat);
752 __mod_lruvec_state(lruvec, idx, val);
758 * mod_objcg_mlstate() may be called with irq enabled, so
759 * mod_memcg_lruvec_state() should be used.
761 static inline void mod_objcg_mlstate(struct obj_cgroup *objcg,
762 struct pglist_data *pgdat,
763 enum node_stat_item idx, int nr)
765 struct mem_cgroup *memcg;
766 struct lruvec *lruvec;
769 memcg = obj_cgroup_memcg(objcg);
770 lruvec = mem_cgroup_lruvec(memcg, pgdat);
771 mod_memcg_lruvec_state(lruvec, idx, nr);
776 * __count_memcg_events - account VM events in a cgroup
777 * @memcg: the memory cgroup
778 * @idx: the event item
779 * @count: the number of events that occurred
781 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
784 if (mem_cgroup_disabled())
787 __this_cpu_add(memcg->vmstats_percpu->events[idx], count);
788 cgroup_rstat_updated(memcg->css.cgroup, smp_processor_id());
791 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
793 return READ_ONCE(memcg->vmstats.events[event]);
796 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
801 for_each_possible_cpu(cpu)
802 x += per_cpu(memcg->vmstats_percpu->events[event], cpu);
806 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
810 /* pagein of a big page is an event. So, ignore page size */
812 __count_memcg_events(memcg, PGPGIN, 1);
814 __count_memcg_events(memcg, PGPGOUT, 1);
815 nr_pages = -nr_pages; /* for event */
818 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
821 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
822 enum mem_cgroup_events_target target)
824 unsigned long val, next;
826 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
827 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
828 /* from time_after() in jiffies.h */
829 if ((long)(next - val) < 0) {
831 case MEM_CGROUP_TARGET_THRESH:
832 next = val + THRESHOLDS_EVENTS_TARGET;
834 case MEM_CGROUP_TARGET_SOFTLIMIT:
835 next = val + SOFTLIMIT_EVENTS_TARGET;
840 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
847 * Check events in order.
850 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
852 /* threshold event is triggered in finer grain than soft limit */
853 if (unlikely(mem_cgroup_event_ratelimit(memcg,
854 MEM_CGROUP_TARGET_THRESH))) {
857 do_softlimit = mem_cgroup_event_ratelimit(memcg,
858 MEM_CGROUP_TARGET_SOFTLIMIT);
859 mem_cgroup_threshold(memcg);
860 if (unlikely(do_softlimit))
861 mem_cgroup_update_tree(memcg, page);
865 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
868 * mm_update_next_owner() may clear mm->owner to NULL
869 * if it races with swapoff, page migration, etc.
870 * So this can be called with p == NULL.
875 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
877 EXPORT_SYMBOL(mem_cgroup_from_task);
879 static __always_inline struct mem_cgroup *active_memcg(void)
882 return this_cpu_read(int_active_memcg);
884 return current->active_memcg;
888 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
889 * @mm: mm from which memcg should be extracted. It can be NULL.
891 * Obtain a reference on mm->memcg and returns it if successful. If mm
892 * is NULL, then the memcg is chosen as follows:
893 * 1) The active memcg, if set.
894 * 2) current->mm->memcg, if available
896 * If mem_cgroup is disabled, NULL is returned.
898 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
900 struct mem_cgroup *memcg;
902 if (mem_cgroup_disabled())
906 * Page cache insertions can happen without an
907 * actual mm context, e.g. during disk probing
908 * on boot, loopback IO, acct() writes etc.
910 * No need to css_get on root memcg as the reference
911 * counting is disabled on the root level in the
912 * cgroup core. See CSS_NO_REF.
915 memcg = active_memcg();
916 if (unlikely(memcg)) {
917 /* remote memcg must hold a ref */
918 css_get(&memcg->css);
923 return root_mem_cgroup;
928 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
929 if (unlikely(!memcg))
930 memcg = root_mem_cgroup;
931 } while (!css_tryget(&memcg->css));
935 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
937 static __always_inline bool memcg_kmem_bypass(void)
939 /* Allow remote memcg charging from any context. */
940 if (unlikely(active_memcg()))
943 /* Memcg to charge can't be determined. */
944 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
951 * mem_cgroup_iter - iterate over memory cgroup hierarchy
952 * @root: hierarchy root
953 * @prev: previously returned memcg, NULL on first invocation
954 * @reclaim: cookie for shared reclaim walks, NULL for full walks
956 * Returns references to children of the hierarchy below @root, or
957 * @root itself, or %NULL after a full round-trip.
959 * Caller must pass the return value in @prev on subsequent
960 * invocations for reference counting, or use mem_cgroup_iter_break()
961 * to cancel a hierarchy walk before the round-trip is complete.
963 * Reclaimers can specify a node in @reclaim to divide up the memcgs
964 * in the hierarchy among all concurrent reclaimers operating on the
967 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
968 struct mem_cgroup *prev,
969 struct mem_cgroup_reclaim_cookie *reclaim)
971 struct mem_cgroup_reclaim_iter *iter;
972 struct cgroup_subsys_state *css = NULL;
973 struct mem_cgroup *memcg = NULL;
974 struct mem_cgroup *pos = NULL;
976 if (mem_cgroup_disabled())
980 root = root_mem_cgroup;
982 if (prev && !reclaim)
988 struct mem_cgroup_per_node *mz;
990 mz = root->nodeinfo[reclaim->pgdat->node_id];
993 if (prev && reclaim->generation != iter->generation)
997 pos = READ_ONCE(iter->position);
998 if (!pos || css_tryget(&pos->css))
1001 * css reference reached zero, so iter->position will
1002 * be cleared by ->css_released. However, we should not
1003 * rely on this happening soon, because ->css_released
1004 * is called from a work queue, and by busy-waiting we
1005 * might block it. So we clear iter->position right
1008 (void)cmpxchg(&iter->position, pos, NULL);
1016 css = css_next_descendant_pre(css, &root->css);
1019 * Reclaimers share the hierarchy walk, and a
1020 * new one might jump in right at the end of
1021 * the hierarchy - make sure they see at least
1022 * one group and restart from the beginning.
1030 * Verify the css and acquire a reference. The root
1031 * is provided by the caller, so we know it's alive
1032 * and kicking, and don't take an extra reference.
1034 memcg = mem_cgroup_from_css(css);
1036 if (css == &root->css)
1039 if (css_tryget(css))
1047 * The position could have already been updated by a competing
1048 * thread, so check that the value hasn't changed since we read
1049 * it to avoid reclaiming from the same cgroup twice.
1051 (void)cmpxchg(&iter->position, pos, memcg);
1059 reclaim->generation = iter->generation;
1064 if (prev && prev != root)
1065 css_put(&prev->css);
1071 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1072 * @root: hierarchy root
1073 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1075 void mem_cgroup_iter_break(struct mem_cgroup *root,
1076 struct mem_cgroup *prev)
1079 root = root_mem_cgroup;
1080 if (prev && prev != root)
1081 css_put(&prev->css);
1084 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1085 struct mem_cgroup *dead_memcg)
1087 struct mem_cgroup_reclaim_iter *iter;
1088 struct mem_cgroup_per_node *mz;
1091 for_each_node(nid) {
1092 mz = from->nodeinfo[nid];
1094 cmpxchg(&iter->position, dead_memcg, NULL);
1098 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1100 struct mem_cgroup *memcg = dead_memcg;
1101 struct mem_cgroup *last;
1104 __invalidate_reclaim_iterators(memcg, dead_memcg);
1106 } while ((memcg = parent_mem_cgroup(memcg)));
1109 * When cgruop1 non-hierarchy mode is used,
1110 * parent_mem_cgroup() does not walk all the way up to the
1111 * cgroup root (root_mem_cgroup). So we have to handle
1112 * dead_memcg from cgroup root separately.
1114 if (last != root_mem_cgroup)
1115 __invalidate_reclaim_iterators(root_mem_cgroup,
1120 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1121 * @memcg: hierarchy root
1122 * @fn: function to call for each task
1123 * @arg: argument passed to @fn
1125 * This function iterates over tasks attached to @memcg or to any of its
1126 * descendants and calls @fn for each task. If @fn returns a non-zero
1127 * value, the function breaks the iteration loop and returns the value.
1128 * Otherwise, it will iterate over all tasks and return 0.
1130 * This function must not be called for the root memory cgroup.
1132 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1133 int (*fn)(struct task_struct *, void *), void *arg)
1135 struct mem_cgroup *iter;
1138 BUG_ON(memcg == root_mem_cgroup);
1140 for_each_mem_cgroup_tree(iter, memcg) {
1141 struct css_task_iter it;
1142 struct task_struct *task;
1144 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1145 while (!ret && (task = css_task_iter_next(&it)))
1146 ret = fn(task, arg);
1147 css_task_iter_end(&it);
1149 mem_cgroup_iter_break(memcg, iter);
1156 #ifdef CONFIG_DEBUG_VM
1157 void lruvec_memcg_debug(struct lruvec *lruvec, struct page *page)
1159 struct mem_cgroup *memcg;
1161 if (mem_cgroup_disabled())
1164 memcg = page_memcg(page);
1167 VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != root_mem_cgroup, page);
1169 VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != memcg, page);
1174 * lock_page_lruvec - lock and return lruvec for a given page.
1177 * These functions are safe to use under any of the following conditions:
1180 * - lock_page_memcg()
1181 * - page->_refcount is zero
1183 struct lruvec *lock_page_lruvec(struct page *page)
1185 struct lruvec *lruvec;
1187 lruvec = mem_cgroup_page_lruvec(page);
1188 spin_lock(&lruvec->lru_lock);
1190 lruvec_memcg_debug(lruvec, page);
1195 struct lruvec *lock_page_lruvec_irq(struct page *page)
1197 struct lruvec *lruvec;
1199 lruvec = mem_cgroup_page_lruvec(page);
1200 spin_lock_irq(&lruvec->lru_lock);
1202 lruvec_memcg_debug(lruvec, page);
1207 struct lruvec *lock_page_lruvec_irqsave(struct page *page, unsigned long *flags)
1209 struct lruvec *lruvec;
1211 lruvec = mem_cgroup_page_lruvec(page);
1212 spin_lock_irqsave(&lruvec->lru_lock, *flags);
1214 lruvec_memcg_debug(lruvec, page);
1220 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1221 * @lruvec: mem_cgroup per zone lru vector
1222 * @lru: index of lru list the page is sitting on
1223 * @zid: zone id of the accounted pages
1224 * @nr_pages: positive when adding or negative when removing
1226 * This function must be called under lru_lock, just before a page is added
1227 * to or just after a page is removed from an lru list (that ordering being
1228 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1230 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1231 int zid, int nr_pages)
1233 struct mem_cgroup_per_node *mz;
1234 unsigned long *lru_size;
1237 if (mem_cgroup_disabled())
1240 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1241 lru_size = &mz->lru_zone_size[zid][lru];
1244 *lru_size += nr_pages;
1247 if (WARN_ONCE(size < 0,
1248 "%s(%p, %d, %d): lru_size %ld\n",
1249 __func__, lruvec, lru, nr_pages, size)) {
1255 *lru_size += nr_pages;
1259 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1260 * @memcg: the memory cgroup
1262 * Returns the maximum amount of memory @mem can be charged with, in
1265 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1267 unsigned long margin = 0;
1268 unsigned long count;
1269 unsigned long limit;
1271 count = page_counter_read(&memcg->memory);
1272 limit = READ_ONCE(memcg->memory.max);
1274 margin = limit - count;
1276 if (do_memsw_account()) {
1277 count = page_counter_read(&memcg->memsw);
1278 limit = READ_ONCE(memcg->memsw.max);
1280 margin = min(margin, limit - count);
1289 * A routine for checking "mem" is under move_account() or not.
1291 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1292 * moving cgroups. This is for waiting at high-memory pressure
1295 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1297 struct mem_cgroup *from;
1298 struct mem_cgroup *to;
1301 * Unlike task_move routines, we access mc.to, mc.from not under
1302 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1304 spin_lock(&mc.lock);
1310 ret = mem_cgroup_is_descendant(from, memcg) ||
1311 mem_cgroup_is_descendant(to, memcg);
1313 spin_unlock(&mc.lock);
1317 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1319 if (mc.moving_task && current != mc.moving_task) {
1320 if (mem_cgroup_under_move(memcg)) {
1322 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1323 /* moving charge context might have finished. */
1326 finish_wait(&mc.waitq, &wait);
1333 struct memory_stat {
1338 static const struct memory_stat memory_stats[] = {
1339 { "anon", NR_ANON_MAPPED },
1340 { "file", NR_FILE_PAGES },
1341 { "kernel_stack", NR_KERNEL_STACK_KB },
1342 { "pagetables", NR_PAGETABLE },
1343 { "percpu", MEMCG_PERCPU_B },
1344 { "sock", MEMCG_SOCK },
1345 { "shmem", NR_SHMEM },
1346 { "file_mapped", NR_FILE_MAPPED },
1347 { "file_dirty", NR_FILE_DIRTY },
1348 { "file_writeback", NR_WRITEBACK },
1350 { "swapcached", NR_SWAPCACHE },
1352 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1353 { "anon_thp", NR_ANON_THPS },
1354 { "file_thp", NR_FILE_THPS },
1355 { "shmem_thp", NR_SHMEM_THPS },
1357 { "inactive_anon", NR_INACTIVE_ANON },
1358 { "active_anon", NR_ACTIVE_ANON },
1359 { "inactive_file", NR_INACTIVE_FILE },
1360 { "active_file", NR_ACTIVE_FILE },
1361 { "unevictable", NR_UNEVICTABLE },
1362 { "slab_reclaimable", NR_SLAB_RECLAIMABLE_B },
1363 { "slab_unreclaimable", NR_SLAB_UNRECLAIMABLE_B },
1365 /* The memory events */
1366 { "workingset_refault_anon", WORKINGSET_REFAULT_ANON },
1367 { "workingset_refault_file", WORKINGSET_REFAULT_FILE },
1368 { "workingset_activate_anon", WORKINGSET_ACTIVATE_ANON },
1369 { "workingset_activate_file", WORKINGSET_ACTIVATE_FILE },
1370 { "workingset_restore_anon", WORKINGSET_RESTORE_ANON },
1371 { "workingset_restore_file", WORKINGSET_RESTORE_FILE },
1372 { "workingset_nodereclaim", WORKINGSET_NODERECLAIM },
1375 /* Translate stat items to the correct unit for memory.stat output */
1376 static int memcg_page_state_unit(int item)
1379 case MEMCG_PERCPU_B:
1380 case NR_SLAB_RECLAIMABLE_B:
1381 case NR_SLAB_UNRECLAIMABLE_B:
1382 case WORKINGSET_REFAULT_ANON:
1383 case WORKINGSET_REFAULT_FILE:
1384 case WORKINGSET_ACTIVATE_ANON:
1385 case WORKINGSET_ACTIVATE_FILE:
1386 case WORKINGSET_RESTORE_ANON:
1387 case WORKINGSET_RESTORE_FILE:
1388 case WORKINGSET_NODERECLAIM:
1390 case NR_KERNEL_STACK_KB:
1397 static inline unsigned long memcg_page_state_output(struct mem_cgroup *memcg,
1400 return memcg_page_state(memcg, item) * memcg_page_state_unit(item);
1403 static char *memory_stat_format(struct mem_cgroup *memcg)
1408 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1413 * Provide statistics on the state of the memory subsystem as
1414 * well as cumulative event counters that show past behavior.
1416 * This list is ordered following a combination of these gradients:
1417 * 1) generic big picture -> specifics and details
1418 * 2) reflecting userspace activity -> reflecting kernel heuristics
1420 * Current memory state:
1422 cgroup_rstat_flush(memcg->css.cgroup);
1424 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1427 size = memcg_page_state_output(memcg, memory_stats[i].idx);
1428 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1430 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1431 size += memcg_page_state_output(memcg,
1432 NR_SLAB_RECLAIMABLE_B);
1433 seq_buf_printf(&s, "slab %llu\n", size);
1437 /* Accumulated memory events */
1439 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1440 memcg_events(memcg, PGFAULT));
1441 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1442 memcg_events(memcg, PGMAJFAULT));
1443 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1444 memcg_events(memcg, PGREFILL));
1445 seq_buf_printf(&s, "pgscan %lu\n",
1446 memcg_events(memcg, PGSCAN_KSWAPD) +
1447 memcg_events(memcg, PGSCAN_DIRECT));
1448 seq_buf_printf(&s, "pgsteal %lu\n",
1449 memcg_events(memcg, PGSTEAL_KSWAPD) +
1450 memcg_events(memcg, PGSTEAL_DIRECT));
1451 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1452 memcg_events(memcg, PGACTIVATE));
1453 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1454 memcg_events(memcg, PGDEACTIVATE));
1455 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1456 memcg_events(memcg, PGLAZYFREE));
1457 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1458 memcg_events(memcg, PGLAZYFREED));
1460 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1461 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1462 memcg_events(memcg, THP_FAULT_ALLOC));
1463 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1464 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1465 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1467 /* The above should easily fit into one page */
1468 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1473 #define K(x) ((x) << (PAGE_SHIFT-10))
1475 * mem_cgroup_print_oom_context: Print OOM information relevant to
1476 * memory controller.
1477 * @memcg: The memory cgroup that went over limit
1478 * @p: Task that is going to be killed
1480 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1483 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1488 pr_cont(",oom_memcg=");
1489 pr_cont_cgroup_path(memcg->css.cgroup);
1491 pr_cont(",global_oom");
1493 pr_cont(",task_memcg=");
1494 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1500 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1501 * memory controller.
1502 * @memcg: The memory cgroup that went over limit
1504 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1508 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1509 K((u64)page_counter_read(&memcg->memory)),
1510 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1511 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1512 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1513 K((u64)page_counter_read(&memcg->swap)),
1514 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1516 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1517 K((u64)page_counter_read(&memcg->memsw)),
1518 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1519 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1520 K((u64)page_counter_read(&memcg->kmem)),
1521 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1524 pr_info("Memory cgroup stats for ");
1525 pr_cont_cgroup_path(memcg->css.cgroup);
1527 buf = memory_stat_format(memcg);
1535 * Return the memory (and swap, if configured) limit for a memcg.
1537 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1539 unsigned long max = READ_ONCE(memcg->memory.max);
1541 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1542 if (mem_cgroup_swappiness(memcg))
1543 max += min(READ_ONCE(memcg->swap.max),
1544 (unsigned long)total_swap_pages);
1546 if (mem_cgroup_swappiness(memcg)) {
1547 /* Calculate swap excess capacity from memsw limit */
1548 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1550 max += min(swap, (unsigned long)total_swap_pages);
1556 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1558 return page_counter_read(&memcg->memory);
1561 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1564 struct oom_control oc = {
1568 .gfp_mask = gfp_mask,
1573 if (mutex_lock_killable(&oom_lock))
1576 if (mem_cgroup_margin(memcg) >= (1 << order))
1580 * A few threads which were not waiting at mutex_lock_killable() can
1581 * fail to bail out. Therefore, check again after holding oom_lock.
1583 ret = should_force_charge() || out_of_memory(&oc);
1586 mutex_unlock(&oom_lock);
1590 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1593 unsigned long *total_scanned)
1595 struct mem_cgroup *victim = NULL;
1598 unsigned long excess;
1599 unsigned long nr_scanned;
1600 struct mem_cgroup_reclaim_cookie reclaim = {
1604 excess = soft_limit_excess(root_memcg);
1607 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1612 * If we have not been able to reclaim
1613 * anything, it might because there are
1614 * no reclaimable pages under this hierarchy
1619 * We want to do more targeted reclaim.
1620 * excess >> 2 is not to excessive so as to
1621 * reclaim too much, nor too less that we keep
1622 * coming back to reclaim from this cgroup
1624 if (total >= (excess >> 2) ||
1625 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1630 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1631 pgdat, &nr_scanned);
1632 *total_scanned += nr_scanned;
1633 if (!soft_limit_excess(root_memcg))
1636 mem_cgroup_iter_break(root_memcg, victim);
1640 #ifdef CONFIG_LOCKDEP
1641 static struct lockdep_map memcg_oom_lock_dep_map = {
1642 .name = "memcg_oom_lock",
1646 static DEFINE_SPINLOCK(memcg_oom_lock);
1649 * Check OOM-Killer is already running under our hierarchy.
1650 * If someone is running, return false.
1652 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1654 struct mem_cgroup *iter, *failed = NULL;
1656 spin_lock(&memcg_oom_lock);
1658 for_each_mem_cgroup_tree(iter, memcg) {
1659 if (iter->oom_lock) {
1661 * this subtree of our hierarchy is already locked
1662 * so we cannot give a lock.
1665 mem_cgroup_iter_break(memcg, iter);
1668 iter->oom_lock = true;
1673 * OK, we failed to lock the whole subtree so we have
1674 * to clean up what we set up to the failing subtree
1676 for_each_mem_cgroup_tree(iter, memcg) {
1677 if (iter == failed) {
1678 mem_cgroup_iter_break(memcg, iter);
1681 iter->oom_lock = false;
1684 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1686 spin_unlock(&memcg_oom_lock);
1691 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1693 struct mem_cgroup *iter;
1695 spin_lock(&memcg_oom_lock);
1696 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1697 for_each_mem_cgroup_tree(iter, memcg)
1698 iter->oom_lock = false;
1699 spin_unlock(&memcg_oom_lock);
1702 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1704 struct mem_cgroup *iter;
1706 spin_lock(&memcg_oom_lock);
1707 for_each_mem_cgroup_tree(iter, memcg)
1709 spin_unlock(&memcg_oom_lock);
1712 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1714 struct mem_cgroup *iter;
1717 * Be careful about under_oom underflows because a child memcg
1718 * could have been added after mem_cgroup_mark_under_oom.
1720 spin_lock(&memcg_oom_lock);
1721 for_each_mem_cgroup_tree(iter, memcg)
1722 if (iter->under_oom > 0)
1724 spin_unlock(&memcg_oom_lock);
1727 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1729 struct oom_wait_info {
1730 struct mem_cgroup *memcg;
1731 wait_queue_entry_t wait;
1734 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1735 unsigned mode, int sync, void *arg)
1737 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1738 struct mem_cgroup *oom_wait_memcg;
1739 struct oom_wait_info *oom_wait_info;
1741 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1742 oom_wait_memcg = oom_wait_info->memcg;
1744 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1745 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1747 return autoremove_wake_function(wait, mode, sync, arg);
1750 static void memcg_oom_recover(struct mem_cgroup *memcg)
1753 * For the following lockless ->under_oom test, the only required
1754 * guarantee is that it must see the state asserted by an OOM when
1755 * this function is called as a result of userland actions
1756 * triggered by the notification of the OOM. This is trivially
1757 * achieved by invoking mem_cgroup_mark_under_oom() before
1758 * triggering notification.
1760 if (memcg && memcg->under_oom)
1761 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1771 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1773 enum oom_status ret;
1776 if (order > PAGE_ALLOC_COSTLY_ORDER)
1779 memcg_memory_event(memcg, MEMCG_OOM);
1782 * We are in the middle of the charge context here, so we
1783 * don't want to block when potentially sitting on a callstack
1784 * that holds all kinds of filesystem and mm locks.
1786 * cgroup1 allows disabling the OOM killer and waiting for outside
1787 * handling until the charge can succeed; remember the context and put
1788 * the task to sleep at the end of the page fault when all locks are
1791 * On the other hand, in-kernel OOM killer allows for an async victim
1792 * memory reclaim (oom_reaper) and that means that we are not solely
1793 * relying on the oom victim to make a forward progress and we can
1794 * invoke the oom killer here.
1796 * Please note that mem_cgroup_out_of_memory might fail to find a
1797 * victim and then we have to bail out from the charge path.
1799 if (memcg->oom_kill_disable) {
1800 if (!current->in_user_fault)
1802 css_get(&memcg->css);
1803 current->memcg_in_oom = memcg;
1804 current->memcg_oom_gfp_mask = mask;
1805 current->memcg_oom_order = order;
1810 mem_cgroup_mark_under_oom(memcg);
1812 locked = mem_cgroup_oom_trylock(memcg);
1815 mem_cgroup_oom_notify(memcg);
1817 mem_cgroup_unmark_under_oom(memcg);
1818 if (mem_cgroup_out_of_memory(memcg, mask, order))
1824 mem_cgroup_oom_unlock(memcg);
1830 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1831 * @handle: actually kill/wait or just clean up the OOM state
1833 * This has to be called at the end of a page fault if the memcg OOM
1834 * handler was enabled.
1836 * Memcg supports userspace OOM handling where failed allocations must
1837 * sleep on a waitqueue until the userspace task resolves the
1838 * situation. Sleeping directly in the charge context with all kinds
1839 * of locks held is not a good idea, instead we remember an OOM state
1840 * in the task and mem_cgroup_oom_synchronize() has to be called at
1841 * the end of the page fault to complete the OOM handling.
1843 * Returns %true if an ongoing memcg OOM situation was detected and
1844 * completed, %false otherwise.
1846 bool mem_cgroup_oom_synchronize(bool handle)
1848 struct mem_cgroup *memcg = current->memcg_in_oom;
1849 struct oom_wait_info owait;
1852 /* OOM is global, do not handle */
1859 owait.memcg = memcg;
1860 owait.wait.flags = 0;
1861 owait.wait.func = memcg_oom_wake_function;
1862 owait.wait.private = current;
1863 INIT_LIST_HEAD(&owait.wait.entry);
1865 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1866 mem_cgroup_mark_under_oom(memcg);
1868 locked = mem_cgroup_oom_trylock(memcg);
1871 mem_cgroup_oom_notify(memcg);
1873 if (locked && !memcg->oom_kill_disable) {
1874 mem_cgroup_unmark_under_oom(memcg);
1875 finish_wait(&memcg_oom_waitq, &owait.wait);
1876 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1877 current->memcg_oom_order);
1880 mem_cgroup_unmark_under_oom(memcg);
1881 finish_wait(&memcg_oom_waitq, &owait.wait);
1885 mem_cgroup_oom_unlock(memcg);
1887 * There is no guarantee that an OOM-lock contender
1888 * sees the wakeups triggered by the OOM kill
1889 * uncharges. Wake any sleepers explicitly.
1891 memcg_oom_recover(memcg);
1894 current->memcg_in_oom = NULL;
1895 css_put(&memcg->css);
1900 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1901 * @victim: task to be killed by the OOM killer
1902 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1904 * Returns a pointer to a memory cgroup, which has to be cleaned up
1905 * by killing all belonging OOM-killable tasks.
1907 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1909 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1910 struct mem_cgroup *oom_domain)
1912 struct mem_cgroup *oom_group = NULL;
1913 struct mem_cgroup *memcg;
1915 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
1919 oom_domain = root_mem_cgroup;
1923 memcg = mem_cgroup_from_task(victim);
1924 if (memcg == root_mem_cgroup)
1928 * If the victim task has been asynchronously moved to a different
1929 * memory cgroup, we might end up killing tasks outside oom_domain.
1930 * In this case it's better to ignore memory.group.oom.
1932 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
1936 * Traverse the memory cgroup hierarchy from the victim task's
1937 * cgroup up to the OOMing cgroup (or root) to find the
1938 * highest-level memory cgroup with oom.group set.
1940 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1941 if (memcg->oom_group)
1944 if (memcg == oom_domain)
1949 css_get(&oom_group->css);
1956 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
1958 pr_info("Tasks in ");
1959 pr_cont_cgroup_path(memcg->css.cgroup);
1960 pr_cont(" are going to be killed due to memory.oom.group set\n");
1964 * lock_page_memcg - lock a page and memcg binding
1967 * This function protects unlocked LRU pages from being moved to
1970 * It ensures lifetime of the locked memcg. Caller is responsible
1971 * for the lifetime of the page.
1973 void lock_page_memcg(struct page *page)
1975 struct page *head = compound_head(page); /* rmap on tail pages */
1976 struct mem_cgroup *memcg;
1977 unsigned long flags;
1980 * The RCU lock is held throughout the transaction. The fast
1981 * path can get away without acquiring the memcg->move_lock
1982 * because page moving starts with an RCU grace period.
1986 if (mem_cgroup_disabled())
1989 memcg = page_memcg(head);
1990 if (unlikely(!memcg))
1993 #ifdef CONFIG_PROVE_LOCKING
1994 local_irq_save(flags);
1995 might_lock(&memcg->move_lock);
1996 local_irq_restore(flags);
1999 if (atomic_read(&memcg->moving_account) <= 0)
2002 spin_lock_irqsave(&memcg->move_lock, flags);
2003 if (memcg != page_memcg(head)) {
2004 spin_unlock_irqrestore(&memcg->move_lock, flags);
2009 * When charge migration first begins, we can have multiple
2010 * critical sections holding the fast-path RCU lock and one
2011 * holding the slowpath move_lock. Track the task who has the
2012 * move_lock for unlock_page_memcg().
2014 memcg->move_lock_task = current;
2015 memcg->move_lock_flags = flags;
2017 EXPORT_SYMBOL(lock_page_memcg);
2019 static void __unlock_page_memcg(struct mem_cgroup *memcg)
2021 if (memcg && memcg->move_lock_task == current) {
2022 unsigned long flags = memcg->move_lock_flags;
2024 memcg->move_lock_task = NULL;
2025 memcg->move_lock_flags = 0;
2027 spin_unlock_irqrestore(&memcg->move_lock, flags);
2034 * unlock_page_memcg - unlock a page and memcg binding
2037 void unlock_page_memcg(struct page *page)
2039 struct page *head = compound_head(page);
2041 __unlock_page_memcg(page_memcg(head));
2043 EXPORT_SYMBOL(unlock_page_memcg);
2046 #ifdef CONFIG_MEMCG_KMEM
2047 struct obj_cgroup *cached_objcg;
2048 struct pglist_data *cached_pgdat;
2049 unsigned int nr_bytes;
2050 int nr_slab_reclaimable_b;
2051 int nr_slab_unreclaimable_b;
2057 struct memcg_stock_pcp {
2058 struct mem_cgroup *cached; /* this never be root cgroup */
2059 unsigned int nr_pages;
2060 struct obj_stock task_obj;
2061 struct obj_stock irq_obj;
2063 struct work_struct work;
2064 unsigned long flags;
2065 #define FLUSHING_CACHED_CHARGE 0
2067 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2068 static DEFINE_MUTEX(percpu_charge_mutex);
2070 #ifdef CONFIG_MEMCG_KMEM
2071 static void drain_obj_stock(struct obj_stock *stock);
2072 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2073 struct mem_cgroup *root_memcg);
2076 static inline void drain_obj_stock(struct obj_stock *stock)
2079 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2080 struct mem_cgroup *root_memcg)
2087 * Most kmem_cache_alloc() calls are from user context. The irq disable/enable
2088 * sequence used in this case to access content from object stock is slow.
2089 * To optimize for user context access, there are now two object stocks for
2090 * task context and interrupt context access respectively.
2092 * The task context object stock can be accessed by disabling preemption only
2093 * which is cheap in non-preempt kernel. The interrupt context object stock
2094 * can only be accessed after disabling interrupt. User context code can
2095 * access interrupt object stock, but not vice versa.
2097 static inline struct obj_stock *get_obj_stock(unsigned long *pflags)
2099 struct memcg_stock_pcp *stock;
2101 if (likely(in_task())) {
2104 stock = this_cpu_ptr(&memcg_stock);
2105 return &stock->task_obj;
2108 local_irq_save(*pflags);
2109 stock = this_cpu_ptr(&memcg_stock);
2110 return &stock->irq_obj;
2113 static inline void put_obj_stock(unsigned long flags)
2115 if (likely(in_task()))
2118 local_irq_restore(flags);
2122 * consume_stock: Try to consume stocked charge on this cpu.
2123 * @memcg: memcg to consume from.
2124 * @nr_pages: how many pages to charge.
2126 * The charges will only happen if @memcg matches the current cpu's memcg
2127 * stock, and at least @nr_pages are available in that stock. Failure to
2128 * service an allocation will refill the stock.
2130 * returns true if successful, false otherwise.
2132 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2134 struct memcg_stock_pcp *stock;
2135 unsigned long flags;
2138 if (nr_pages > MEMCG_CHARGE_BATCH)
2141 local_irq_save(flags);
2143 stock = this_cpu_ptr(&memcg_stock);
2144 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2145 stock->nr_pages -= nr_pages;
2149 local_irq_restore(flags);
2155 * Returns stocks cached in percpu and reset cached information.
2157 static void drain_stock(struct memcg_stock_pcp *stock)
2159 struct mem_cgroup *old = stock->cached;
2164 if (stock->nr_pages) {
2165 page_counter_uncharge(&old->memory, stock->nr_pages);
2166 if (do_memsw_account())
2167 page_counter_uncharge(&old->memsw, stock->nr_pages);
2168 stock->nr_pages = 0;
2172 stock->cached = NULL;
2175 static void drain_local_stock(struct work_struct *dummy)
2177 struct memcg_stock_pcp *stock;
2178 unsigned long flags;
2181 * The only protection from memory hotplug vs. drain_stock races is
2182 * that we always operate on local CPU stock here with IRQ disabled
2184 local_irq_save(flags);
2186 stock = this_cpu_ptr(&memcg_stock);
2187 drain_obj_stock(&stock->irq_obj);
2189 drain_obj_stock(&stock->task_obj);
2191 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2193 local_irq_restore(flags);
2197 * Cache charges(val) to local per_cpu area.
2198 * This will be consumed by consume_stock() function, later.
2200 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2202 struct memcg_stock_pcp *stock;
2203 unsigned long flags;
2205 local_irq_save(flags);
2207 stock = this_cpu_ptr(&memcg_stock);
2208 if (stock->cached != memcg) { /* reset if necessary */
2210 css_get(&memcg->css);
2211 stock->cached = memcg;
2213 stock->nr_pages += nr_pages;
2215 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2218 local_irq_restore(flags);
2222 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2223 * of the hierarchy under it.
2225 static void drain_all_stock(struct mem_cgroup *root_memcg)
2229 /* If someone's already draining, avoid adding running more workers. */
2230 if (!mutex_trylock(&percpu_charge_mutex))
2233 * Notify other cpus that system-wide "drain" is running
2234 * We do not care about races with the cpu hotplug because cpu down
2235 * as well as workers from this path always operate on the local
2236 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2239 for_each_online_cpu(cpu) {
2240 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2241 struct mem_cgroup *memcg;
2245 memcg = stock->cached;
2246 if (memcg && stock->nr_pages &&
2247 mem_cgroup_is_descendant(memcg, root_memcg))
2249 if (obj_stock_flush_required(stock, root_memcg))
2254 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2256 drain_local_stock(&stock->work);
2258 schedule_work_on(cpu, &stock->work);
2262 mutex_unlock(&percpu_charge_mutex);
2265 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2267 struct memcg_stock_pcp *stock;
2269 stock = &per_cpu(memcg_stock, cpu);
2275 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2276 unsigned int nr_pages,
2279 unsigned long nr_reclaimed = 0;
2282 unsigned long pflags;
2284 if (page_counter_read(&memcg->memory) <=
2285 READ_ONCE(memcg->memory.high))
2288 memcg_memory_event(memcg, MEMCG_HIGH);
2290 psi_memstall_enter(&pflags);
2291 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2293 psi_memstall_leave(&pflags);
2294 } while ((memcg = parent_mem_cgroup(memcg)) &&
2295 !mem_cgroup_is_root(memcg));
2297 return nr_reclaimed;
2300 static void high_work_func(struct work_struct *work)
2302 struct mem_cgroup *memcg;
2304 memcg = container_of(work, struct mem_cgroup, high_work);
2305 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2309 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2310 * enough to still cause a significant slowdown in most cases, while still
2311 * allowing diagnostics and tracing to proceed without becoming stuck.
2313 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2316 * When calculating the delay, we use these either side of the exponentiation to
2317 * maintain precision and scale to a reasonable number of jiffies (see the table
2320 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2321 * overage ratio to a delay.
2322 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2323 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2324 * to produce a reasonable delay curve.
2326 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2327 * reasonable delay curve compared to precision-adjusted overage, not
2328 * penalising heavily at first, but still making sure that growth beyond the
2329 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2330 * example, with a high of 100 megabytes:
2332 * +-------+------------------------+
2333 * | usage | time to allocate in ms |
2334 * +-------+------------------------+
2356 * +-------+------------------------+
2358 #define MEMCG_DELAY_PRECISION_SHIFT 20
2359 #define MEMCG_DELAY_SCALING_SHIFT 14
2361 static u64 calculate_overage(unsigned long usage, unsigned long high)
2369 * Prevent division by 0 in overage calculation by acting as if
2370 * it was a threshold of 1 page
2372 high = max(high, 1UL);
2374 overage = usage - high;
2375 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2376 return div64_u64(overage, high);
2379 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2381 u64 overage, max_overage = 0;
2384 overage = calculate_overage(page_counter_read(&memcg->memory),
2385 READ_ONCE(memcg->memory.high));
2386 max_overage = max(overage, max_overage);
2387 } while ((memcg = parent_mem_cgroup(memcg)) &&
2388 !mem_cgroup_is_root(memcg));
2393 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2395 u64 overage, max_overage = 0;
2398 overage = calculate_overage(page_counter_read(&memcg->swap),
2399 READ_ONCE(memcg->swap.high));
2401 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2402 max_overage = max(overage, max_overage);
2403 } while ((memcg = parent_mem_cgroup(memcg)) &&
2404 !mem_cgroup_is_root(memcg));
2410 * Get the number of jiffies that we should penalise a mischievous cgroup which
2411 * is exceeding its memory.high by checking both it and its ancestors.
2413 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2414 unsigned int nr_pages,
2417 unsigned long penalty_jiffies;
2423 * We use overage compared to memory.high to calculate the number of
2424 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2425 * fairly lenient on small overages, and increasingly harsh when the
2426 * memcg in question makes it clear that it has no intention of stopping
2427 * its crazy behaviour, so we exponentially increase the delay based on
2430 penalty_jiffies = max_overage * max_overage * HZ;
2431 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2432 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2435 * Factor in the task's own contribution to the overage, such that four
2436 * N-sized allocations are throttled approximately the same as one
2437 * 4N-sized allocation.
2439 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2440 * larger the current charge patch is than that.
2442 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2446 * Scheduled by try_charge() to be executed from the userland return path
2447 * and reclaims memory over the high limit.
2449 void mem_cgroup_handle_over_high(void)
2451 unsigned long penalty_jiffies;
2452 unsigned long pflags;
2453 unsigned long nr_reclaimed;
2454 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2455 int nr_retries = MAX_RECLAIM_RETRIES;
2456 struct mem_cgroup *memcg;
2457 bool in_retry = false;
2459 if (likely(!nr_pages))
2462 memcg = get_mem_cgroup_from_mm(current->mm);
2463 current->memcg_nr_pages_over_high = 0;
2467 * The allocating task should reclaim at least the batch size, but for
2468 * subsequent retries we only want to do what's necessary to prevent oom
2469 * or breaching resource isolation.
2471 * This is distinct from memory.max or page allocator behaviour because
2472 * memory.high is currently batched, whereas memory.max and the page
2473 * allocator run every time an allocation is made.
2475 nr_reclaimed = reclaim_high(memcg,
2476 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2480 * memory.high is breached and reclaim is unable to keep up. Throttle
2481 * allocators proactively to slow down excessive growth.
2483 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2484 mem_find_max_overage(memcg));
2486 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2487 swap_find_max_overage(memcg));
2490 * Clamp the max delay per usermode return so as to still keep the
2491 * application moving forwards and also permit diagnostics, albeit
2494 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2497 * Don't sleep if the amount of jiffies this memcg owes us is so low
2498 * that it's not even worth doing, in an attempt to be nice to those who
2499 * go only a small amount over their memory.high value and maybe haven't
2500 * been aggressively reclaimed enough yet.
2502 if (penalty_jiffies <= HZ / 100)
2506 * If reclaim is making forward progress but we're still over
2507 * memory.high, we want to encourage that rather than doing allocator
2510 if (nr_reclaimed || nr_retries--) {
2516 * If we exit early, we're guaranteed to die (since
2517 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2518 * need to account for any ill-begotten jiffies to pay them off later.
2520 psi_memstall_enter(&pflags);
2521 schedule_timeout_killable(penalty_jiffies);
2522 psi_memstall_leave(&pflags);
2525 css_put(&memcg->css);
2528 static int try_charge_memcg(struct mem_cgroup *memcg, gfp_t gfp_mask,
2529 unsigned int nr_pages)
2531 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2532 int nr_retries = MAX_RECLAIM_RETRIES;
2533 struct mem_cgroup *mem_over_limit;
2534 struct page_counter *counter;
2535 enum oom_status oom_status;
2536 unsigned long nr_reclaimed;
2537 bool may_swap = true;
2538 bool drained = false;
2539 unsigned long pflags;
2542 if (consume_stock(memcg, nr_pages))
2545 if (!do_memsw_account() ||
2546 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2547 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2549 if (do_memsw_account())
2550 page_counter_uncharge(&memcg->memsw, batch);
2551 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2553 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2557 if (batch > nr_pages) {
2563 * Memcg doesn't have a dedicated reserve for atomic
2564 * allocations. But like the global atomic pool, we need to
2565 * put the burden of reclaim on regular allocation requests
2566 * and let these go through as privileged allocations.
2568 if (gfp_mask & __GFP_ATOMIC)
2572 * Unlike in global OOM situations, memcg is not in a physical
2573 * memory shortage. Allow dying and OOM-killed tasks to
2574 * bypass the last charges so that they can exit quickly and
2575 * free their memory.
2577 if (unlikely(should_force_charge()))
2581 * Prevent unbounded recursion when reclaim operations need to
2582 * allocate memory. This might exceed the limits temporarily,
2583 * but we prefer facilitating memory reclaim and getting back
2584 * under the limit over triggering OOM kills in these cases.
2586 if (unlikely(current->flags & PF_MEMALLOC))
2589 if (unlikely(task_in_memcg_oom(current)))
2592 if (!gfpflags_allow_blocking(gfp_mask))
2595 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2597 psi_memstall_enter(&pflags);
2598 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2599 gfp_mask, may_swap);
2600 psi_memstall_leave(&pflags);
2602 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2606 drain_all_stock(mem_over_limit);
2611 if (gfp_mask & __GFP_NORETRY)
2614 * Even though the limit is exceeded at this point, reclaim
2615 * may have been able to free some pages. Retry the charge
2616 * before killing the task.
2618 * Only for regular pages, though: huge pages are rather
2619 * unlikely to succeed so close to the limit, and we fall back
2620 * to regular pages anyway in case of failure.
2622 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2625 * At task move, charge accounts can be doubly counted. So, it's
2626 * better to wait until the end of task_move if something is going on.
2628 if (mem_cgroup_wait_acct_move(mem_over_limit))
2634 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2637 if (fatal_signal_pending(current))
2641 * keep retrying as long as the memcg oom killer is able to make
2642 * a forward progress or bypass the charge if the oom killer
2643 * couldn't make any progress.
2645 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2646 get_order(nr_pages * PAGE_SIZE));
2647 switch (oom_status) {
2649 nr_retries = MAX_RECLAIM_RETRIES;
2657 if (!(gfp_mask & __GFP_NOFAIL))
2661 * The allocation either can't fail or will lead to more memory
2662 * being freed very soon. Allow memory usage go over the limit
2663 * temporarily by force charging it.
2665 page_counter_charge(&memcg->memory, nr_pages);
2666 if (do_memsw_account())
2667 page_counter_charge(&memcg->memsw, nr_pages);
2672 if (batch > nr_pages)
2673 refill_stock(memcg, batch - nr_pages);
2676 * If the hierarchy is above the normal consumption range, schedule
2677 * reclaim on returning to userland. We can perform reclaim here
2678 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2679 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2680 * not recorded as it most likely matches current's and won't
2681 * change in the meantime. As high limit is checked again before
2682 * reclaim, the cost of mismatch is negligible.
2685 bool mem_high, swap_high;
2687 mem_high = page_counter_read(&memcg->memory) >
2688 READ_ONCE(memcg->memory.high);
2689 swap_high = page_counter_read(&memcg->swap) >
2690 READ_ONCE(memcg->swap.high);
2692 /* Don't bother a random interrupted task */
2693 if (in_interrupt()) {
2695 schedule_work(&memcg->high_work);
2701 if (mem_high || swap_high) {
2703 * The allocating tasks in this cgroup will need to do
2704 * reclaim or be throttled to prevent further growth
2705 * of the memory or swap footprints.
2707 * Target some best-effort fairness between the tasks,
2708 * and distribute reclaim work and delay penalties
2709 * based on how much each task is actually allocating.
2711 current->memcg_nr_pages_over_high += batch;
2712 set_notify_resume(current);
2715 } while ((memcg = parent_mem_cgroup(memcg)));
2720 static inline int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2721 unsigned int nr_pages)
2723 if (mem_cgroup_is_root(memcg))
2726 return try_charge_memcg(memcg, gfp_mask, nr_pages);
2729 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2730 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2732 if (mem_cgroup_is_root(memcg))
2735 page_counter_uncharge(&memcg->memory, nr_pages);
2736 if (do_memsw_account())
2737 page_counter_uncharge(&memcg->memsw, nr_pages);
2741 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2743 VM_BUG_ON_PAGE(page_memcg(page), page);
2745 * Any of the following ensures page's memcg stability:
2749 * - lock_page_memcg()
2750 * - exclusive reference
2752 page->memcg_data = (unsigned long)memcg;
2755 static struct mem_cgroup *get_mem_cgroup_from_objcg(struct obj_cgroup *objcg)
2757 struct mem_cgroup *memcg;
2761 memcg = obj_cgroup_memcg(objcg);
2762 if (unlikely(!css_tryget(&memcg->css)))
2769 #ifdef CONFIG_MEMCG_KMEM
2771 * The allocated objcg pointers array is not accounted directly.
2772 * Moreover, it should not come from DMA buffer and is not readily
2773 * reclaimable. So those GFP bits should be masked off.
2775 #define OBJCGS_CLEAR_MASK (__GFP_DMA | __GFP_RECLAIMABLE | __GFP_ACCOUNT)
2777 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2778 gfp_t gfp, bool new_page)
2780 unsigned int objects = objs_per_slab_page(s, page);
2781 unsigned long memcg_data;
2784 gfp &= ~OBJCGS_CLEAR_MASK;
2785 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2790 memcg_data = (unsigned long) vec | MEMCG_DATA_OBJCGS;
2793 * If the slab page is brand new and nobody can yet access
2794 * it's memcg_data, no synchronization is required and
2795 * memcg_data can be simply assigned.
2797 page->memcg_data = memcg_data;
2798 } else if (cmpxchg(&page->memcg_data, 0, memcg_data)) {
2800 * If the slab page is already in use, somebody can allocate
2801 * and assign obj_cgroups in parallel. In this case the existing
2802 * objcg vector should be reused.
2808 kmemleak_not_leak(vec);
2813 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2815 * A passed kernel object can be a slab object or a generic kernel page, so
2816 * different mechanisms for getting the memory cgroup pointer should be used.
2817 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2818 * can not know for sure how the kernel object is implemented.
2819 * mem_cgroup_from_obj() can be safely used in such cases.
2821 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2822 * cgroup_mutex, etc.
2824 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2828 if (mem_cgroup_disabled())
2831 page = virt_to_head_page(p);
2834 * Slab objects are accounted individually, not per-page.
2835 * Memcg membership data for each individual object is saved in
2836 * the page->obj_cgroups.
2838 if (page_objcgs_check(page)) {
2839 struct obj_cgroup *objcg;
2842 off = obj_to_index(page->slab_cache, page, p);
2843 objcg = page_objcgs(page)[off];
2845 return obj_cgroup_memcg(objcg);
2851 * page_memcg_check() is used here, because page_has_obj_cgroups()
2852 * check above could fail because the object cgroups vector wasn't set
2853 * at that moment, but it can be set concurrently.
2854 * page_memcg_check(page) will guarantee that a proper memory
2855 * cgroup pointer or NULL will be returned.
2857 return page_memcg_check(page);
2860 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2862 struct obj_cgroup *objcg = NULL;
2863 struct mem_cgroup *memcg;
2865 if (memcg_kmem_bypass())
2869 if (unlikely(active_memcg()))
2870 memcg = active_memcg();
2872 memcg = mem_cgroup_from_task(current);
2874 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2875 objcg = rcu_dereference(memcg->objcg);
2876 if (objcg && obj_cgroup_tryget(objcg))
2885 static int memcg_alloc_cache_id(void)
2890 id = ida_simple_get(&memcg_cache_ida,
2891 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2895 if (id < memcg_nr_cache_ids)
2899 * There's no space for the new id in memcg_caches arrays,
2900 * so we have to grow them.
2902 down_write(&memcg_cache_ids_sem);
2904 size = 2 * (id + 1);
2905 if (size < MEMCG_CACHES_MIN_SIZE)
2906 size = MEMCG_CACHES_MIN_SIZE;
2907 else if (size > MEMCG_CACHES_MAX_SIZE)
2908 size = MEMCG_CACHES_MAX_SIZE;
2910 err = memcg_update_all_list_lrus(size);
2912 memcg_nr_cache_ids = size;
2914 up_write(&memcg_cache_ids_sem);
2917 ida_simple_remove(&memcg_cache_ida, id);
2923 static void memcg_free_cache_id(int id)
2925 ida_simple_remove(&memcg_cache_ida, id);
2929 * obj_cgroup_uncharge_pages: uncharge a number of kernel pages from a objcg
2930 * @objcg: object cgroup to uncharge
2931 * @nr_pages: number of pages to uncharge
2933 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
2934 unsigned int nr_pages)
2936 struct mem_cgroup *memcg;
2938 memcg = get_mem_cgroup_from_objcg(objcg);
2940 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2941 page_counter_uncharge(&memcg->kmem, nr_pages);
2942 refill_stock(memcg, nr_pages);
2944 css_put(&memcg->css);
2948 * obj_cgroup_charge_pages: charge a number of kernel pages to a objcg
2949 * @objcg: object cgroup to charge
2950 * @gfp: reclaim mode
2951 * @nr_pages: number of pages to charge
2953 * Returns 0 on success, an error code on failure.
2955 static int obj_cgroup_charge_pages(struct obj_cgroup *objcg, gfp_t gfp,
2956 unsigned int nr_pages)
2958 struct page_counter *counter;
2959 struct mem_cgroup *memcg;
2962 memcg = get_mem_cgroup_from_objcg(objcg);
2964 ret = try_charge_memcg(memcg, gfp, nr_pages);
2968 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2969 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2972 * Enforce __GFP_NOFAIL allocation because callers are not
2973 * prepared to see failures and likely do not have any failure
2976 if (gfp & __GFP_NOFAIL) {
2977 page_counter_charge(&memcg->kmem, nr_pages);
2980 cancel_charge(memcg, nr_pages);
2984 css_put(&memcg->css);
2990 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
2991 * @page: page to charge
2992 * @gfp: reclaim mode
2993 * @order: allocation order
2995 * Returns 0 on success, an error code on failure.
2997 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
2999 struct obj_cgroup *objcg;
3002 objcg = get_obj_cgroup_from_current();
3004 ret = obj_cgroup_charge_pages(objcg, gfp, 1 << order);
3006 page->memcg_data = (unsigned long)objcg |
3010 obj_cgroup_put(objcg);
3016 * __memcg_kmem_uncharge_page: uncharge a kmem page
3017 * @page: page to uncharge
3018 * @order: allocation order
3020 void __memcg_kmem_uncharge_page(struct page *page, int order)
3022 struct obj_cgroup *objcg;
3023 unsigned int nr_pages = 1 << order;
3025 if (!PageMemcgKmem(page))
3028 objcg = __page_objcg(page);
3029 obj_cgroup_uncharge_pages(objcg, nr_pages);
3030 page->memcg_data = 0;
3031 obj_cgroup_put(objcg);
3034 void mod_objcg_state(struct obj_cgroup *objcg, struct pglist_data *pgdat,
3035 enum node_stat_item idx, int nr)
3037 unsigned long flags;
3038 struct obj_stock *stock = get_obj_stock(&flags);
3042 * Save vmstat data in stock and skip vmstat array update unless
3043 * accumulating over a page of vmstat data or when pgdat or idx
3046 if (stock->cached_objcg != objcg) {
3047 drain_obj_stock(stock);
3048 obj_cgroup_get(objcg);
3049 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3050 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3051 stock->cached_objcg = objcg;
3052 stock->cached_pgdat = pgdat;
3053 } else if (stock->cached_pgdat != pgdat) {
3054 /* Flush the existing cached vmstat data */
3055 struct pglist_data *oldpg = stock->cached_pgdat;
3057 if (stock->nr_slab_reclaimable_b) {
3058 mod_objcg_mlstate(objcg, oldpg, NR_SLAB_RECLAIMABLE_B,
3059 stock->nr_slab_reclaimable_b);
3060 stock->nr_slab_reclaimable_b = 0;
3062 if (stock->nr_slab_unreclaimable_b) {
3063 mod_objcg_mlstate(objcg, oldpg, NR_SLAB_UNRECLAIMABLE_B,
3064 stock->nr_slab_unreclaimable_b);
3065 stock->nr_slab_unreclaimable_b = 0;
3067 stock->cached_pgdat = pgdat;
3070 bytes = (idx == NR_SLAB_RECLAIMABLE_B) ? &stock->nr_slab_reclaimable_b
3071 : &stock->nr_slab_unreclaimable_b;
3073 * Even for large object >= PAGE_SIZE, the vmstat data will still be
3074 * cached locally at least once before pushing it out.
3081 if (abs(*bytes) > PAGE_SIZE) {
3089 mod_objcg_mlstate(objcg, pgdat, idx, nr);
3091 put_obj_stock(flags);
3094 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3096 unsigned long flags;
3097 struct obj_stock *stock = get_obj_stock(&flags);
3100 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3101 stock->nr_bytes -= nr_bytes;
3105 put_obj_stock(flags);
3110 static void drain_obj_stock(struct obj_stock *stock)
3112 struct obj_cgroup *old = stock->cached_objcg;
3117 if (stock->nr_bytes) {
3118 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3119 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3122 obj_cgroup_uncharge_pages(old, nr_pages);
3125 * The leftover is flushed to the centralized per-memcg value.
3126 * On the next attempt to refill obj stock it will be moved
3127 * to a per-cpu stock (probably, on an other CPU), see
3128 * refill_obj_stock().
3130 * How often it's flushed is a trade-off between the memory
3131 * limit enforcement accuracy and potential CPU contention,
3132 * so it might be changed in the future.
3134 atomic_add(nr_bytes, &old->nr_charged_bytes);
3135 stock->nr_bytes = 0;
3139 * Flush the vmstat data in current stock
3141 if (stock->nr_slab_reclaimable_b || stock->nr_slab_unreclaimable_b) {
3142 if (stock->nr_slab_reclaimable_b) {
3143 mod_objcg_mlstate(old, stock->cached_pgdat,
3144 NR_SLAB_RECLAIMABLE_B,
3145 stock->nr_slab_reclaimable_b);
3146 stock->nr_slab_reclaimable_b = 0;
3148 if (stock->nr_slab_unreclaimable_b) {
3149 mod_objcg_mlstate(old, stock->cached_pgdat,
3150 NR_SLAB_UNRECLAIMABLE_B,
3151 stock->nr_slab_unreclaimable_b);
3152 stock->nr_slab_unreclaimable_b = 0;
3154 stock->cached_pgdat = NULL;
3157 obj_cgroup_put(old);
3158 stock->cached_objcg = NULL;
3161 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3162 struct mem_cgroup *root_memcg)
3164 struct mem_cgroup *memcg;
3166 if (in_task() && stock->task_obj.cached_objcg) {
3167 memcg = obj_cgroup_memcg(stock->task_obj.cached_objcg);
3168 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3171 if (stock->irq_obj.cached_objcg) {
3172 memcg = obj_cgroup_memcg(stock->irq_obj.cached_objcg);
3173 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3180 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes,
3181 bool allow_uncharge)
3183 unsigned long flags;
3184 struct obj_stock *stock = get_obj_stock(&flags);
3185 unsigned int nr_pages = 0;
3187 if (stock->cached_objcg != objcg) { /* reset if necessary */
3188 drain_obj_stock(stock);
3189 obj_cgroup_get(objcg);
3190 stock->cached_objcg = objcg;
3191 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3192 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3193 allow_uncharge = true; /* Allow uncharge when objcg changes */
3195 stock->nr_bytes += nr_bytes;
3197 if (allow_uncharge && (stock->nr_bytes > PAGE_SIZE)) {
3198 nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3199 stock->nr_bytes &= (PAGE_SIZE - 1);
3202 put_obj_stock(flags);
3205 obj_cgroup_uncharge_pages(objcg, nr_pages);
3208 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3210 unsigned int nr_pages, nr_bytes;
3213 if (consume_obj_stock(objcg, size))
3217 * In theory, objcg->nr_charged_bytes can have enough
3218 * pre-charged bytes to satisfy the allocation. However,
3219 * flushing objcg->nr_charged_bytes requires two atomic
3220 * operations, and objcg->nr_charged_bytes can't be big.
3221 * The shared objcg->nr_charged_bytes can also become a
3222 * performance bottleneck if all tasks of the same memcg are
3223 * trying to update it. So it's better to ignore it and try
3224 * grab some new pages. The stock's nr_bytes will be flushed to
3225 * objcg->nr_charged_bytes later on when objcg changes.
3227 * The stock's nr_bytes may contain enough pre-charged bytes
3228 * to allow one less page from being charged, but we can't rely
3229 * on the pre-charged bytes not being changed outside of
3230 * consume_obj_stock() or refill_obj_stock(). So ignore those
3231 * pre-charged bytes as well when charging pages. To avoid a
3232 * page uncharge right after a page charge, we set the
3233 * allow_uncharge flag to false when calling refill_obj_stock()
3234 * to temporarily allow the pre-charged bytes to exceed the page
3235 * size limit. The maximum reachable value of the pre-charged
3236 * bytes is (sizeof(object) + PAGE_SIZE - 2) if there is no data
3239 nr_pages = size >> PAGE_SHIFT;
3240 nr_bytes = size & (PAGE_SIZE - 1);
3245 ret = obj_cgroup_charge_pages(objcg, gfp, nr_pages);
3246 if (!ret && nr_bytes)
3247 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes, false);
3252 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3254 refill_obj_stock(objcg, size, true);
3257 #endif /* CONFIG_MEMCG_KMEM */
3260 * Because page_memcg(head) is not set on tails, set it now.
3262 void split_page_memcg(struct page *head, unsigned int nr)
3264 struct mem_cgroup *memcg = page_memcg(head);
3267 if (mem_cgroup_disabled() || !memcg)
3270 for (i = 1; i < nr; i++)
3271 head[i].memcg_data = head->memcg_data;
3273 if (PageMemcgKmem(head))
3274 obj_cgroup_get_many(__page_objcg(head), nr - 1);
3276 css_get_many(&memcg->css, nr - 1);
3279 #ifdef CONFIG_MEMCG_SWAP
3281 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3282 * @entry: swap entry to be moved
3283 * @from: mem_cgroup which the entry is moved from
3284 * @to: mem_cgroup which the entry is moved to
3286 * It succeeds only when the swap_cgroup's record for this entry is the same
3287 * as the mem_cgroup's id of @from.
3289 * Returns 0 on success, -EINVAL on failure.
3291 * The caller must have charged to @to, IOW, called page_counter_charge() about
3292 * both res and memsw, and called css_get().
3294 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3295 struct mem_cgroup *from, struct mem_cgroup *to)
3297 unsigned short old_id, new_id;
3299 old_id = mem_cgroup_id(from);
3300 new_id = mem_cgroup_id(to);
3302 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3303 mod_memcg_state(from, MEMCG_SWAP, -1);
3304 mod_memcg_state(to, MEMCG_SWAP, 1);
3310 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3311 struct mem_cgroup *from, struct mem_cgroup *to)
3317 static DEFINE_MUTEX(memcg_max_mutex);
3319 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3320 unsigned long max, bool memsw)
3322 bool enlarge = false;
3323 bool drained = false;
3325 bool limits_invariant;
3326 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3329 if (signal_pending(current)) {
3334 mutex_lock(&memcg_max_mutex);
3336 * Make sure that the new limit (memsw or memory limit) doesn't
3337 * break our basic invariant rule memory.max <= memsw.max.
3339 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3340 max <= memcg->memsw.max;
3341 if (!limits_invariant) {
3342 mutex_unlock(&memcg_max_mutex);
3346 if (max > counter->max)
3348 ret = page_counter_set_max(counter, max);
3349 mutex_unlock(&memcg_max_mutex);
3355 drain_all_stock(memcg);
3360 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3361 GFP_KERNEL, !memsw)) {
3367 if (!ret && enlarge)
3368 memcg_oom_recover(memcg);
3373 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3375 unsigned long *total_scanned)
3377 unsigned long nr_reclaimed = 0;
3378 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3379 unsigned long reclaimed;
3381 struct mem_cgroup_tree_per_node *mctz;
3382 unsigned long excess;
3383 unsigned long nr_scanned;
3388 mctz = soft_limit_tree_node(pgdat->node_id);
3391 * Do not even bother to check the largest node if the root
3392 * is empty. Do it lockless to prevent lock bouncing. Races
3393 * are acceptable as soft limit is best effort anyway.
3395 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3399 * This loop can run a while, specially if mem_cgroup's continuously
3400 * keep exceeding their soft limit and putting the system under
3407 mz = mem_cgroup_largest_soft_limit_node(mctz);
3412 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3413 gfp_mask, &nr_scanned);
3414 nr_reclaimed += reclaimed;
3415 *total_scanned += nr_scanned;
3416 spin_lock_irq(&mctz->lock);
3417 __mem_cgroup_remove_exceeded(mz, mctz);
3420 * If we failed to reclaim anything from this memory cgroup
3421 * it is time to move on to the next cgroup
3425 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3427 excess = soft_limit_excess(mz->memcg);
3429 * One school of thought says that we should not add
3430 * back the node to the tree if reclaim returns 0.
3431 * But our reclaim could return 0, simply because due
3432 * to priority we are exposing a smaller subset of
3433 * memory to reclaim from. Consider this as a longer
3436 /* If excess == 0, no tree ops */
3437 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3438 spin_unlock_irq(&mctz->lock);
3439 css_put(&mz->memcg->css);
3442 * Could not reclaim anything and there are no more
3443 * mem cgroups to try or we seem to be looping without
3444 * reclaiming anything.
3446 if (!nr_reclaimed &&
3448 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3450 } while (!nr_reclaimed);
3452 css_put(&next_mz->memcg->css);
3453 return nr_reclaimed;
3457 * Reclaims as many pages from the given memcg as possible.
3459 * Caller is responsible for holding css reference for memcg.
3461 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3463 int nr_retries = MAX_RECLAIM_RETRIES;
3465 /* we call try-to-free pages for make this cgroup empty */
3466 lru_add_drain_all();
3468 drain_all_stock(memcg);
3470 /* try to free all pages in this cgroup */
3471 while (nr_retries && page_counter_read(&memcg->memory)) {
3474 if (signal_pending(current))
3477 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3481 /* maybe some writeback is necessary */
3482 congestion_wait(BLK_RW_ASYNC, HZ/10);
3490 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3491 char *buf, size_t nbytes,
3494 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3496 if (mem_cgroup_is_root(memcg))
3498 return mem_cgroup_force_empty(memcg) ?: nbytes;
3501 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3507 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3508 struct cftype *cft, u64 val)
3513 pr_warn_once("Non-hierarchical mode is deprecated. "
3514 "Please report your usecase to linux-mm@kvack.org if you "
3515 "depend on this functionality.\n");
3520 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3524 if (mem_cgroup_is_root(memcg)) {
3525 /* mem_cgroup_threshold() calls here from irqsafe context */
3526 cgroup_rstat_flush_irqsafe(memcg->css.cgroup);
3527 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3528 memcg_page_state(memcg, NR_ANON_MAPPED);
3530 val += memcg_page_state(memcg, MEMCG_SWAP);
3533 val = page_counter_read(&memcg->memory);
3535 val = page_counter_read(&memcg->memsw);
3548 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3551 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3552 struct page_counter *counter;
3554 switch (MEMFILE_TYPE(cft->private)) {
3556 counter = &memcg->memory;
3559 counter = &memcg->memsw;
3562 counter = &memcg->kmem;
3565 counter = &memcg->tcpmem;
3571 switch (MEMFILE_ATTR(cft->private)) {
3573 if (counter == &memcg->memory)
3574 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3575 if (counter == &memcg->memsw)
3576 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3577 return (u64)page_counter_read(counter) * PAGE_SIZE;
3579 return (u64)counter->max * PAGE_SIZE;
3581 return (u64)counter->watermark * PAGE_SIZE;
3583 return counter->failcnt;
3584 case RES_SOFT_LIMIT:
3585 return (u64)memcg->soft_limit * PAGE_SIZE;
3591 #ifdef CONFIG_MEMCG_KMEM
3592 static int memcg_online_kmem(struct mem_cgroup *memcg)
3594 struct obj_cgroup *objcg;
3597 if (cgroup_memory_nokmem)
3600 BUG_ON(memcg->kmemcg_id >= 0);
3601 BUG_ON(memcg->kmem_state);
3603 memcg_id = memcg_alloc_cache_id();
3607 objcg = obj_cgroup_alloc();
3609 memcg_free_cache_id(memcg_id);
3612 objcg->memcg = memcg;
3613 rcu_assign_pointer(memcg->objcg, objcg);
3615 static_branch_enable(&memcg_kmem_enabled_key);
3617 memcg->kmemcg_id = memcg_id;
3618 memcg->kmem_state = KMEM_ONLINE;
3623 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3625 struct cgroup_subsys_state *css;
3626 struct mem_cgroup *parent, *child;
3629 if (memcg->kmem_state != KMEM_ONLINE)
3632 memcg->kmem_state = KMEM_ALLOCATED;
3634 parent = parent_mem_cgroup(memcg);
3636 parent = root_mem_cgroup;
3638 memcg_reparent_objcgs(memcg, parent);
3640 kmemcg_id = memcg->kmemcg_id;
3641 BUG_ON(kmemcg_id < 0);
3644 * Change kmemcg_id of this cgroup and all its descendants to the
3645 * parent's id, and then move all entries from this cgroup's list_lrus
3646 * to ones of the parent. After we have finished, all list_lrus
3647 * corresponding to this cgroup are guaranteed to remain empty. The
3648 * ordering is imposed by list_lru_node->lock taken by
3649 * memcg_drain_all_list_lrus().
3651 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3652 css_for_each_descendant_pre(css, &memcg->css) {
3653 child = mem_cgroup_from_css(css);
3654 BUG_ON(child->kmemcg_id != kmemcg_id);
3655 child->kmemcg_id = parent->kmemcg_id;
3659 memcg_drain_all_list_lrus(kmemcg_id, parent);
3661 memcg_free_cache_id(kmemcg_id);
3664 static void memcg_free_kmem(struct mem_cgroup *memcg)
3666 /* css_alloc() failed, offlining didn't happen */
3667 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3668 memcg_offline_kmem(memcg);
3671 static int memcg_online_kmem(struct mem_cgroup *memcg)
3675 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3678 static void memcg_free_kmem(struct mem_cgroup *memcg)
3681 #endif /* CONFIG_MEMCG_KMEM */
3683 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3688 mutex_lock(&memcg_max_mutex);
3689 ret = page_counter_set_max(&memcg->kmem, max);
3690 mutex_unlock(&memcg_max_mutex);
3694 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3698 mutex_lock(&memcg_max_mutex);
3700 ret = page_counter_set_max(&memcg->tcpmem, max);
3704 if (!memcg->tcpmem_active) {
3706 * The active flag needs to be written after the static_key
3707 * update. This is what guarantees that the socket activation
3708 * function is the last one to run. See mem_cgroup_sk_alloc()
3709 * for details, and note that we don't mark any socket as
3710 * belonging to this memcg until that flag is up.
3712 * We need to do this, because static_keys will span multiple
3713 * sites, but we can't control their order. If we mark a socket
3714 * as accounted, but the accounting functions are not patched in
3715 * yet, we'll lose accounting.
3717 * We never race with the readers in mem_cgroup_sk_alloc(),
3718 * because when this value change, the code to process it is not
3721 static_branch_inc(&memcg_sockets_enabled_key);
3722 memcg->tcpmem_active = true;
3725 mutex_unlock(&memcg_max_mutex);
3730 * The user of this function is...
3733 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3734 char *buf, size_t nbytes, loff_t off)
3736 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3737 unsigned long nr_pages;
3740 buf = strstrip(buf);
3741 ret = page_counter_memparse(buf, "-1", &nr_pages);
3745 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3747 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3751 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3753 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3756 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3759 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3760 "Please report your usecase to linux-mm@kvack.org if you "
3761 "depend on this functionality.\n");
3762 ret = memcg_update_kmem_max(memcg, nr_pages);
3765 ret = memcg_update_tcp_max(memcg, nr_pages);
3769 case RES_SOFT_LIMIT:
3770 memcg->soft_limit = nr_pages;
3774 return ret ?: nbytes;
3777 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3778 size_t nbytes, loff_t off)
3780 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3781 struct page_counter *counter;
3783 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3785 counter = &memcg->memory;
3788 counter = &memcg->memsw;
3791 counter = &memcg->kmem;
3794 counter = &memcg->tcpmem;
3800 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3802 page_counter_reset_watermark(counter);
3805 counter->failcnt = 0;
3814 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3817 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3821 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3822 struct cftype *cft, u64 val)
3824 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3826 if (val & ~MOVE_MASK)
3830 * No kind of locking is needed in here, because ->can_attach() will
3831 * check this value once in the beginning of the process, and then carry
3832 * on with stale data. This means that changes to this value will only
3833 * affect task migrations starting after the change.
3835 memcg->move_charge_at_immigrate = val;
3839 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3840 struct cftype *cft, u64 val)
3848 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3849 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3850 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3852 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3853 int nid, unsigned int lru_mask, bool tree)
3855 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3856 unsigned long nr = 0;
3859 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3862 if (!(BIT(lru) & lru_mask))
3865 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3867 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3872 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3873 unsigned int lru_mask,
3876 unsigned long nr = 0;
3880 if (!(BIT(lru) & lru_mask))
3883 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3885 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3890 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3894 unsigned int lru_mask;
3897 static const struct numa_stat stats[] = {
3898 { "total", LRU_ALL },
3899 { "file", LRU_ALL_FILE },
3900 { "anon", LRU_ALL_ANON },
3901 { "unevictable", BIT(LRU_UNEVICTABLE) },
3903 const struct numa_stat *stat;
3905 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3907 cgroup_rstat_flush(memcg->css.cgroup);
3909 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3910 seq_printf(m, "%s=%lu", stat->name,
3911 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3913 for_each_node_state(nid, N_MEMORY)
3914 seq_printf(m, " N%d=%lu", nid,
3915 mem_cgroup_node_nr_lru_pages(memcg, nid,
3916 stat->lru_mask, false));
3920 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3922 seq_printf(m, "hierarchical_%s=%lu", stat->name,
3923 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3925 for_each_node_state(nid, N_MEMORY)
3926 seq_printf(m, " N%d=%lu", nid,
3927 mem_cgroup_node_nr_lru_pages(memcg, nid,
3928 stat->lru_mask, true));
3934 #endif /* CONFIG_NUMA */
3936 static const unsigned int memcg1_stats[] = {
3939 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3949 static const char *const memcg1_stat_names[] = {
3952 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3962 /* Universal VM events cgroup1 shows, original sort order */
3963 static const unsigned int memcg1_events[] = {
3970 static int memcg_stat_show(struct seq_file *m, void *v)
3972 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3973 unsigned long memory, memsw;
3974 struct mem_cgroup *mi;
3977 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3979 cgroup_rstat_flush(memcg->css.cgroup);
3981 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3984 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3986 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
3987 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
3990 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3991 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
3992 memcg_events_local(memcg, memcg1_events[i]));
3994 for (i = 0; i < NR_LRU_LISTS; i++)
3995 seq_printf(m, "%s %lu\n", lru_list_name(i),
3996 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3999 /* Hierarchical information */
4000 memory = memsw = PAGE_COUNTER_MAX;
4001 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4002 memory = min(memory, READ_ONCE(mi->memory.max));
4003 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4005 seq_printf(m, "hierarchical_memory_limit %llu\n",
4006 (u64)memory * PAGE_SIZE);
4007 if (do_memsw_account())
4008 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4009 (u64)memsw * PAGE_SIZE);
4011 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4014 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4016 nr = memcg_page_state(memcg, memcg1_stats[i]);
4017 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4018 (u64)nr * PAGE_SIZE);
4021 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4022 seq_printf(m, "total_%s %llu\n",
4023 vm_event_name(memcg1_events[i]),
4024 (u64)memcg_events(memcg, memcg1_events[i]));
4026 for (i = 0; i < NR_LRU_LISTS; i++)
4027 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4028 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4031 #ifdef CONFIG_DEBUG_VM
4034 struct mem_cgroup_per_node *mz;
4035 unsigned long anon_cost = 0;
4036 unsigned long file_cost = 0;
4038 for_each_online_pgdat(pgdat) {
4039 mz = memcg->nodeinfo[pgdat->node_id];
4041 anon_cost += mz->lruvec.anon_cost;
4042 file_cost += mz->lruvec.file_cost;
4044 seq_printf(m, "anon_cost %lu\n", anon_cost);
4045 seq_printf(m, "file_cost %lu\n", file_cost);
4052 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4055 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4057 return mem_cgroup_swappiness(memcg);
4060 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4061 struct cftype *cft, u64 val)
4063 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4068 if (!mem_cgroup_is_root(memcg))
4069 memcg->swappiness = val;
4071 vm_swappiness = val;
4076 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4078 struct mem_cgroup_threshold_ary *t;
4079 unsigned long usage;
4084 t = rcu_dereference(memcg->thresholds.primary);
4086 t = rcu_dereference(memcg->memsw_thresholds.primary);
4091 usage = mem_cgroup_usage(memcg, swap);
4094 * current_threshold points to threshold just below or equal to usage.
4095 * If it's not true, a threshold was crossed after last
4096 * call of __mem_cgroup_threshold().
4098 i = t->current_threshold;
4101 * Iterate backward over array of thresholds starting from
4102 * current_threshold and check if a threshold is crossed.
4103 * If none of thresholds below usage is crossed, we read
4104 * only one element of the array here.
4106 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4107 eventfd_signal(t->entries[i].eventfd, 1);
4109 /* i = current_threshold + 1 */
4113 * Iterate forward over array of thresholds starting from
4114 * current_threshold+1 and check if a threshold is crossed.
4115 * If none of thresholds above usage is crossed, we read
4116 * only one element of the array here.
4118 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4119 eventfd_signal(t->entries[i].eventfd, 1);
4121 /* Update current_threshold */
4122 t->current_threshold = i - 1;
4127 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4130 __mem_cgroup_threshold(memcg, false);
4131 if (do_memsw_account())
4132 __mem_cgroup_threshold(memcg, true);
4134 memcg = parent_mem_cgroup(memcg);
4138 static int compare_thresholds(const void *a, const void *b)
4140 const struct mem_cgroup_threshold *_a = a;
4141 const struct mem_cgroup_threshold *_b = b;
4143 if (_a->threshold > _b->threshold)
4146 if (_a->threshold < _b->threshold)
4152 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4154 struct mem_cgroup_eventfd_list *ev;
4156 spin_lock(&memcg_oom_lock);
4158 list_for_each_entry(ev, &memcg->oom_notify, list)
4159 eventfd_signal(ev->eventfd, 1);
4161 spin_unlock(&memcg_oom_lock);
4165 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4167 struct mem_cgroup *iter;
4169 for_each_mem_cgroup_tree(iter, memcg)
4170 mem_cgroup_oom_notify_cb(iter);
4173 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4174 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4176 struct mem_cgroup_thresholds *thresholds;
4177 struct mem_cgroup_threshold_ary *new;
4178 unsigned long threshold;
4179 unsigned long usage;
4182 ret = page_counter_memparse(args, "-1", &threshold);
4186 mutex_lock(&memcg->thresholds_lock);
4189 thresholds = &memcg->thresholds;
4190 usage = mem_cgroup_usage(memcg, false);
4191 } else if (type == _MEMSWAP) {
4192 thresholds = &memcg->memsw_thresholds;
4193 usage = mem_cgroup_usage(memcg, true);
4197 /* Check if a threshold crossed before adding a new one */
4198 if (thresholds->primary)
4199 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4201 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4203 /* Allocate memory for new array of thresholds */
4204 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4211 /* Copy thresholds (if any) to new array */
4212 if (thresholds->primary)
4213 memcpy(new->entries, thresholds->primary->entries,
4214 flex_array_size(new, entries, size - 1));
4216 /* Add new threshold */
4217 new->entries[size - 1].eventfd = eventfd;
4218 new->entries[size - 1].threshold = threshold;
4220 /* Sort thresholds. Registering of new threshold isn't time-critical */
4221 sort(new->entries, size, sizeof(*new->entries),
4222 compare_thresholds, NULL);
4224 /* Find current threshold */
4225 new->current_threshold = -1;
4226 for (i = 0; i < size; i++) {
4227 if (new->entries[i].threshold <= usage) {
4229 * new->current_threshold will not be used until
4230 * rcu_assign_pointer(), so it's safe to increment
4233 ++new->current_threshold;
4238 /* Free old spare buffer and save old primary buffer as spare */
4239 kfree(thresholds->spare);
4240 thresholds->spare = thresholds->primary;
4242 rcu_assign_pointer(thresholds->primary, new);
4244 /* To be sure that nobody uses thresholds */
4248 mutex_unlock(&memcg->thresholds_lock);
4253 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4254 struct eventfd_ctx *eventfd, const char *args)
4256 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4259 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4260 struct eventfd_ctx *eventfd, const char *args)
4262 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4265 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4266 struct eventfd_ctx *eventfd, enum res_type type)
4268 struct mem_cgroup_thresholds *thresholds;
4269 struct mem_cgroup_threshold_ary *new;
4270 unsigned long usage;
4271 int i, j, size, entries;
4273 mutex_lock(&memcg->thresholds_lock);
4276 thresholds = &memcg->thresholds;
4277 usage = mem_cgroup_usage(memcg, false);
4278 } else if (type == _MEMSWAP) {
4279 thresholds = &memcg->memsw_thresholds;
4280 usage = mem_cgroup_usage(memcg, true);
4284 if (!thresholds->primary)
4287 /* Check if a threshold crossed before removing */
4288 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4290 /* Calculate new number of threshold */
4292 for (i = 0; i < thresholds->primary->size; i++) {
4293 if (thresholds->primary->entries[i].eventfd != eventfd)
4299 new = thresholds->spare;
4301 /* If no items related to eventfd have been cleared, nothing to do */
4305 /* Set thresholds array to NULL if we don't have thresholds */
4314 /* Copy thresholds and find current threshold */
4315 new->current_threshold = -1;
4316 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4317 if (thresholds->primary->entries[i].eventfd == eventfd)
4320 new->entries[j] = thresholds->primary->entries[i];
4321 if (new->entries[j].threshold <= usage) {
4323 * new->current_threshold will not be used
4324 * until rcu_assign_pointer(), so it's safe to increment
4327 ++new->current_threshold;
4333 /* Swap primary and spare array */
4334 thresholds->spare = thresholds->primary;
4336 rcu_assign_pointer(thresholds->primary, new);
4338 /* To be sure that nobody uses thresholds */
4341 /* If all events are unregistered, free the spare array */
4343 kfree(thresholds->spare);
4344 thresholds->spare = NULL;
4347 mutex_unlock(&memcg->thresholds_lock);
4350 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4351 struct eventfd_ctx *eventfd)
4353 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4356 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4357 struct eventfd_ctx *eventfd)
4359 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4362 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4363 struct eventfd_ctx *eventfd, const char *args)
4365 struct mem_cgroup_eventfd_list *event;
4367 event = kmalloc(sizeof(*event), GFP_KERNEL);
4371 spin_lock(&memcg_oom_lock);
4373 event->eventfd = eventfd;
4374 list_add(&event->list, &memcg->oom_notify);
4376 /* already in OOM ? */
4377 if (memcg->under_oom)
4378 eventfd_signal(eventfd, 1);
4379 spin_unlock(&memcg_oom_lock);
4384 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4385 struct eventfd_ctx *eventfd)
4387 struct mem_cgroup_eventfd_list *ev, *tmp;
4389 spin_lock(&memcg_oom_lock);
4391 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4392 if (ev->eventfd == eventfd) {
4393 list_del(&ev->list);
4398 spin_unlock(&memcg_oom_lock);
4401 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4403 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4405 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4406 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4407 seq_printf(sf, "oom_kill %lu\n",
4408 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4412 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4413 struct cftype *cft, u64 val)
4415 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4417 /* cannot set to root cgroup and only 0 and 1 are allowed */
4418 if (mem_cgroup_is_root(memcg) || !((val == 0) || (val == 1)))
4421 memcg->oom_kill_disable = val;
4423 memcg_oom_recover(memcg);
4428 #ifdef CONFIG_CGROUP_WRITEBACK
4430 #include <trace/events/writeback.h>
4432 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4434 return wb_domain_init(&memcg->cgwb_domain, gfp);
4437 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4439 wb_domain_exit(&memcg->cgwb_domain);
4442 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4444 wb_domain_size_changed(&memcg->cgwb_domain);
4447 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4449 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4451 if (!memcg->css.parent)
4454 return &memcg->cgwb_domain;
4458 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4459 * @wb: bdi_writeback in question
4460 * @pfilepages: out parameter for number of file pages
4461 * @pheadroom: out parameter for number of allocatable pages according to memcg
4462 * @pdirty: out parameter for number of dirty pages
4463 * @pwriteback: out parameter for number of pages under writeback
4465 * Determine the numbers of file, headroom, dirty, and writeback pages in
4466 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4467 * is a bit more involved.
4469 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4470 * headroom is calculated as the lowest headroom of itself and the
4471 * ancestors. Note that this doesn't consider the actual amount of
4472 * available memory in the system. The caller should further cap
4473 * *@pheadroom accordingly.
4475 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4476 unsigned long *pheadroom, unsigned long *pdirty,
4477 unsigned long *pwriteback)
4479 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4480 struct mem_cgroup *parent;
4482 cgroup_rstat_flush_irqsafe(memcg->css.cgroup);
4484 *pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
4485 *pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
4486 *pfilepages = memcg_page_state(memcg, NR_INACTIVE_FILE) +
4487 memcg_page_state(memcg, NR_ACTIVE_FILE);
4489 *pheadroom = PAGE_COUNTER_MAX;
4490 while ((parent = parent_mem_cgroup(memcg))) {
4491 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4492 READ_ONCE(memcg->memory.high));
4493 unsigned long used = page_counter_read(&memcg->memory);
4495 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4501 * Foreign dirty flushing
4503 * There's an inherent mismatch between memcg and writeback. The former
4504 * tracks ownership per-page while the latter per-inode. This was a
4505 * deliberate design decision because honoring per-page ownership in the
4506 * writeback path is complicated, may lead to higher CPU and IO overheads
4507 * and deemed unnecessary given that write-sharing an inode across
4508 * different cgroups isn't a common use-case.
4510 * Combined with inode majority-writer ownership switching, this works well
4511 * enough in most cases but there are some pathological cases. For
4512 * example, let's say there are two cgroups A and B which keep writing to
4513 * different but confined parts of the same inode. B owns the inode and
4514 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4515 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4516 * triggering background writeback. A will be slowed down without a way to
4517 * make writeback of the dirty pages happen.
4519 * Conditions like the above can lead to a cgroup getting repeatedly and
4520 * severely throttled after making some progress after each
4521 * dirty_expire_interval while the underlying IO device is almost
4524 * Solving this problem completely requires matching the ownership tracking
4525 * granularities between memcg and writeback in either direction. However,
4526 * the more egregious behaviors can be avoided by simply remembering the
4527 * most recent foreign dirtying events and initiating remote flushes on
4528 * them when local writeback isn't enough to keep the memory clean enough.
4530 * The following two functions implement such mechanism. When a foreign
4531 * page - a page whose memcg and writeback ownerships don't match - is
4532 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4533 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4534 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4535 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4536 * foreign bdi_writebacks which haven't expired. Both the numbers of
4537 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4538 * limited to MEMCG_CGWB_FRN_CNT.
4540 * The mechanism only remembers IDs and doesn't hold any object references.
4541 * As being wrong occasionally doesn't matter, updates and accesses to the
4542 * records are lockless and racy.
4544 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4545 struct bdi_writeback *wb)
4547 struct mem_cgroup *memcg = page_memcg(page);
4548 struct memcg_cgwb_frn *frn;
4549 u64 now = get_jiffies_64();
4550 u64 oldest_at = now;
4554 trace_track_foreign_dirty(page, wb);
4557 * Pick the slot to use. If there is already a slot for @wb, keep
4558 * using it. If not replace the oldest one which isn't being
4561 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4562 frn = &memcg->cgwb_frn[i];
4563 if (frn->bdi_id == wb->bdi->id &&
4564 frn->memcg_id == wb->memcg_css->id)
4566 if (time_before64(frn->at, oldest_at) &&
4567 atomic_read(&frn->done.cnt) == 1) {
4569 oldest_at = frn->at;
4573 if (i < MEMCG_CGWB_FRN_CNT) {
4575 * Re-using an existing one. Update timestamp lazily to
4576 * avoid making the cacheline hot. We want them to be
4577 * reasonably up-to-date and significantly shorter than
4578 * dirty_expire_interval as that's what expires the record.
4579 * Use the shorter of 1s and dirty_expire_interval / 8.
4581 unsigned long update_intv =
4582 min_t(unsigned long, HZ,
4583 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4585 if (time_before64(frn->at, now - update_intv))
4587 } else if (oldest >= 0) {
4588 /* replace the oldest free one */
4589 frn = &memcg->cgwb_frn[oldest];
4590 frn->bdi_id = wb->bdi->id;
4591 frn->memcg_id = wb->memcg_css->id;
4596 /* issue foreign writeback flushes for recorded foreign dirtying events */
4597 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4599 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4600 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4601 u64 now = jiffies_64;
4604 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4605 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4608 * If the record is older than dirty_expire_interval,
4609 * writeback on it has already started. No need to kick it
4610 * off again. Also, don't start a new one if there's
4611 * already one in flight.
4613 if (time_after64(frn->at, now - intv) &&
4614 atomic_read(&frn->done.cnt) == 1) {
4616 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4617 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id,
4618 WB_REASON_FOREIGN_FLUSH,
4624 #else /* CONFIG_CGROUP_WRITEBACK */
4626 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4631 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4635 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4639 #endif /* CONFIG_CGROUP_WRITEBACK */
4642 * DO NOT USE IN NEW FILES.
4644 * "cgroup.event_control" implementation.
4646 * This is way over-engineered. It tries to support fully configurable
4647 * events for each user. Such level of flexibility is completely
4648 * unnecessary especially in the light of the planned unified hierarchy.
4650 * Please deprecate this and replace with something simpler if at all
4655 * Unregister event and free resources.
4657 * Gets called from workqueue.
4659 static void memcg_event_remove(struct work_struct *work)
4661 struct mem_cgroup_event *event =
4662 container_of(work, struct mem_cgroup_event, remove);
4663 struct mem_cgroup *memcg = event->memcg;
4665 remove_wait_queue(event->wqh, &event->wait);
4667 event->unregister_event(memcg, event->eventfd);
4669 /* Notify userspace the event is going away. */
4670 eventfd_signal(event->eventfd, 1);
4672 eventfd_ctx_put(event->eventfd);
4674 css_put(&memcg->css);
4678 * Gets called on EPOLLHUP on eventfd when user closes it.
4680 * Called with wqh->lock held and interrupts disabled.
4682 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4683 int sync, void *key)
4685 struct mem_cgroup_event *event =
4686 container_of(wait, struct mem_cgroup_event, wait);
4687 struct mem_cgroup *memcg = event->memcg;
4688 __poll_t flags = key_to_poll(key);
4690 if (flags & EPOLLHUP) {
4692 * If the event has been detached at cgroup removal, we
4693 * can simply return knowing the other side will cleanup
4696 * We can't race against event freeing since the other
4697 * side will require wqh->lock via remove_wait_queue(),
4700 spin_lock(&memcg->event_list_lock);
4701 if (!list_empty(&event->list)) {
4702 list_del_init(&event->list);
4704 * We are in atomic context, but cgroup_event_remove()
4705 * may sleep, so we have to call it in workqueue.
4707 schedule_work(&event->remove);
4709 spin_unlock(&memcg->event_list_lock);
4715 static void memcg_event_ptable_queue_proc(struct file *file,
4716 wait_queue_head_t *wqh, poll_table *pt)
4718 struct mem_cgroup_event *event =
4719 container_of(pt, struct mem_cgroup_event, pt);
4722 add_wait_queue(wqh, &event->wait);
4726 * DO NOT USE IN NEW FILES.
4728 * Parse input and register new cgroup event handler.
4730 * Input must be in format '<event_fd> <control_fd> <args>'.
4731 * Interpretation of args is defined by control file implementation.
4733 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4734 char *buf, size_t nbytes, loff_t off)
4736 struct cgroup_subsys_state *css = of_css(of);
4737 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4738 struct mem_cgroup_event *event;
4739 struct cgroup_subsys_state *cfile_css;
4740 unsigned int efd, cfd;
4747 buf = strstrip(buf);
4749 efd = simple_strtoul(buf, &endp, 10);
4754 cfd = simple_strtoul(buf, &endp, 10);
4755 if ((*endp != ' ') && (*endp != '\0'))
4759 event = kzalloc(sizeof(*event), GFP_KERNEL);
4763 event->memcg = memcg;
4764 INIT_LIST_HEAD(&event->list);
4765 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4766 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4767 INIT_WORK(&event->remove, memcg_event_remove);
4775 event->eventfd = eventfd_ctx_fileget(efile.file);
4776 if (IS_ERR(event->eventfd)) {
4777 ret = PTR_ERR(event->eventfd);
4784 goto out_put_eventfd;
4787 /* the process need read permission on control file */
4788 /* AV: shouldn't we check that it's been opened for read instead? */
4789 ret = file_permission(cfile.file, MAY_READ);
4794 * Determine the event callbacks and set them in @event. This used
4795 * to be done via struct cftype but cgroup core no longer knows
4796 * about these events. The following is crude but the whole thing
4797 * is for compatibility anyway.
4799 * DO NOT ADD NEW FILES.
4801 name = cfile.file->f_path.dentry->d_name.name;
4803 if (!strcmp(name, "memory.usage_in_bytes")) {
4804 event->register_event = mem_cgroup_usage_register_event;
4805 event->unregister_event = mem_cgroup_usage_unregister_event;
4806 } else if (!strcmp(name, "memory.oom_control")) {
4807 event->register_event = mem_cgroup_oom_register_event;
4808 event->unregister_event = mem_cgroup_oom_unregister_event;
4809 } else if (!strcmp(name, "memory.pressure_level")) {
4810 event->register_event = vmpressure_register_event;
4811 event->unregister_event = vmpressure_unregister_event;
4812 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4813 event->register_event = memsw_cgroup_usage_register_event;
4814 event->unregister_event = memsw_cgroup_usage_unregister_event;
4821 * Verify @cfile should belong to @css. Also, remaining events are
4822 * automatically removed on cgroup destruction but the removal is
4823 * asynchronous, so take an extra ref on @css.
4825 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4826 &memory_cgrp_subsys);
4828 if (IS_ERR(cfile_css))
4830 if (cfile_css != css) {
4835 ret = event->register_event(memcg, event->eventfd, buf);
4839 vfs_poll(efile.file, &event->pt);
4841 spin_lock(&memcg->event_list_lock);
4842 list_add(&event->list, &memcg->event_list);
4843 spin_unlock(&memcg->event_list_lock);
4855 eventfd_ctx_put(event->eventfd);
4864 static struct cftype mem_cgroup_legacy_files[] = {
4866 .name = "usage_in_bytes",
4867 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4868 .read_u64 = mem_cgroup_read_u64,
4871 .name = "max_usage_in_bytes",
4872 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4873 .write = mem_cgroup_reset,
4874 .read_u64 = mem_cgroup_read_u64,
4877 .name = "limit_in_bytes",
4878 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4879 .write = mem_cgroup_write,
4880 .read_u64 = mem_cgroup_read_u64,
4883 .name = "soft_limit_in_bytes",
4884 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4885 .write = mem_cgroup_write,
4886 .read_u64 = mem_cgroup_read_u64,
4890 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4891 .write = mem_cgroup_reset,
4892 .read_u64 = mem_cgroup_read_u64,
4896 .seq_show = memcg_stat_show,
4899 .name = "force_empty",
4900 .write = mem_cgroup_force_empty_write,
4903 .name = "use_hierarchy",
4904 .write_u64 = mem_cgroup_hierarchy_write,
4905 .read_u64 = mem_cgroup_hierarchy_read,
4908 .name = "cgroup.event_control", /* XXX: for compat */
4909 .write = memcg_write_event_control,
4910 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4913 .name = "swappiness",
4914 .read_u64 = mem_cgroup_swappiness_read,
4915 .write_u64 = mem_cgroup_swappiness_write,
4918 .name = "move_charge_at_immigrate",
4919 .read_u64 = mem_cgroup_move_charge_read,
4920 .write_u64 = mem_cgroup_move_charge_write,
4923 .name = "oom_control",
4924 .seq_show = mem_cgroup_oom_control_read,
4925 .write_u64 = mem_cgroup_oom_control_write,
4926 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4929 .name = "pressure_level",
4933 .name = "numa_stat",
4934 .seq_show = memcg_numa_stat_show,
4938 .name = "kmem.limit_in_bytes",
4939 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4940 .write = mem_cgroup_write,
4941 .read_u64 = mem_cgroup_read_u64,
4944 .name = "kmem.usage_in_bytes",
4945 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4946 .read_u64 = mem_cgroup_read_u64,
4949 .name = "kmem.failcnt",
4950 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4951 .write = mem_cgroup_reset,
4952 .read_u64 = mem_cgroup_read_u64,
4955 .name = "kmem.max_usage_in_bytes",
4956 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4957 .write = mem_cgroup_reset,
4958 .read_u64 = mem_cgroup_read_u64,
4960 #if defined(CONFIG_MEMCG_KMEM) && \
4961 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
4963 .name = "kmem.slabinfo",
4964 .seq_show = memcg_slab_show,
4968 .name = "kmem.tcp.limit_in_bytes",
4969 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4970 .write = mem_cgroup_write,
4971 .read_u64 = mem_cgroup_read_u64,
4974 .name = "kmem.tcp.usage_in_bytes",
4975 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4976 .read_u64 = mem_cgroup_read_u64,
4979 .name = "kmem.tcp.failcnt",
4980 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4981 .write = mem_cgroup_reset,
4982 .read_u64 = mem_cgroup_read_u64,
4985 .name = "kmem.tcp.max_usage_in_bytes",
4986 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4987 .write = mem_cgroup_reset,
4988 .read_u64 = mem_cgroup_read_u64,
4990 { }, /* terminate */
4994 * Private memory cgroup IDR
4996 * Swap-out records and page cache shadow entries need to store memcg
4997 * references in constrained space, so we maintain an ID space that is
4998 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4999 * memory-controlled cgroups to 64k.
5001 * However, there usually are many references to the offline CSS after
5002 * the cgroup has been destroyed, such as page cache or reclaimable
5003 * slab objects, that don't need to hang on to the ID. We want to keep
5004 * those dead CSS from occupying IDs, or we might quickly exhaust the
5005 * relatively small ID space and prevent the creation of new cgroups
5006 * even when there are much fewer than 64k cgroups - possibly none.
5008 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5009 * be freed and recycled when it's no longer needed, which is usually
5010 * when the CSS is offlined.
5012 * The only exception to that are records of swapped out tmpfs/shmem
5013 * pages that need to be attributed to live ancestors on swapin. But
5014 * those references are manageable from userspace.
5017 static DEFINE_IDR(mem_cgroup_idr);
5019 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5021 if (memcg->id.id > 0) {
5022 idr_remove(&mem_cgroup_idr, memcg->id.id);
5027 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5030 refcount_add(n, &memcg->id.ref);
5033 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5035 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5036 mem_cgroup_id_remove(memcg);
5038 /* Memcg ID pins CSS */
5039 css_put(&memcg->css);
5043 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5045 mem_cgroup_id_put_many(memcg, 1);
5049 * mem_cgroup_from_id - look up a memcg from a memcg id
5050 * @id: the memcg id to look up
5052 * Caller must hold rcu_read_lock().
5054 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5056 WARN_ON_ONCE(!rcu_read_lock_held());
5057 return idr_find(&mem_cgroup_idr, id);
5060 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5062 struct mem_cgroup_per_node *pn;
5065 * This routine is called against possible nodes.
5066 * But it's BUG to call kmalloc() against offline node.
5068 * TODO: this routine can waste much memory for nodes which will
5069 * never be onlined. It's better to use memory hotplug callback
5072 if (!node_state(node, N_NORMAL_MEMORY))
5074 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5078 pn->lruvec_stats_percpu = alloc_percpu_gfp(struct lruvec_stats_percpu,
5079 GFP_KERNEL_ACCOUNT);
5080 if (!pn->lruvec_stats_percpu) {
5085 lruvec_init(&pn->lruvec);
5086 pn->usage_in_excess = 0;
5087 pn->on_tree = false;
5090 memcg->nodeinfo[node] = pn;
5094 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5096 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5101 free_percpu(pn->lruvec_stats_percpu);
5105 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5110 free_mem_cgroup_per_node_info(memcg, node);
5111 free_percpu(memcg->vmstats_percpu);
5115 static void mem_cgroup_free(struct mem_cgroup *memcg)
5117 memcg_wb_domain_exit(memcg);
5118 __mem_cgroup_free(memcg);
5121 static struct mem_cgroup *mem_cgroup_alloc(void)
5123 struct mem_cgroup *memcg;
5126 int __maybe_unused i;
5127 long error = -ENOMEM;
5129 size = sizeof(struct mem_cgroup);
5130 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5132 memcg = kzalloc(size, GFP_KERNEL);
5134 return ERR_PTR(error);
5136 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5137 1, MEM_CGROUP_ID_MAX,
5139 if (memcg->id.id < 0) {
5140 error = memcg->id.id;
5144 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5145 GFP_KERNEL_ACCOUNT);
5146 if (!memcg->vmstats_percpu)
5150 if (alloc_mem_cgroup_per_node_info(memcg, node))
5153 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5156 INIT_WORK(&memcg->high_work, high_work_func);
5157 INIT_LIST_HEAD(&memcg->oom_notify);
5158 mutex_init(&memcg->thresholds_lock);
5159 spin_lock_init(&memcg->move_lock);
5160 vmpressure_init(&memcg->vmpressure);
5161 INIT_LIST_HEAD(&memcg->event_list);
5162 spin_lock_init(&memcg->event_list_lock);
5163 memcg->socket_pressure = jiffies;
5164 #ifdef CONFIG_MEMCG_KMEM
5165 memcg->kmemcg_id = -1;
5166 INIT_LIST_HEAD(&memcg->objcg_list);
5168 #ifdef CONFIG_CGROUP_WRITEBACK
5169 INIT_LIST_HEAD(&memcg->cgwb_list);
5170 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5171 memcg->cgwb_frn[i].done =
5172 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5174 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5175 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5176 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5177 memcg->deferred_split_queue.split_queue_len = 0;
5179 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5182 mem_cgroup_id_remove(memcg);
5183 __mem_cgroup_free(memcg);
5184 return ERR_PTR(error);
5187 static struct cgroup_subsys_state * __ref
5188 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5190 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5191 struct mem_cgroup *memcg, *old_memcg;
5192 long error = -ENOMEM;
5194 old_memcg = set_active_memcg(parent);
5195 memcg = mem_cgroup_alloc();
5196 set_active_memcg(old_memcg);
5198 return ERR_CAST(memcg);
5200 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5201 memcg->soft_limit = PAGE_COUNTER_MAX;
5202 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5204 memcg->swappiness = mem_cgroup_swappiness(parent);
5205 memcg->oom_kill_disable = parent->oom_kill_disable;
5207 page_counter_init(&memcg->memory, &parent->memory);
5208 page_counter_init(&memcg->swap, &parent->swap);
5209 page_counter_init(&memcg->kmem, &parent->kmem);
5210 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5212 page_counter_init(&memcg->memory, NULL);
5213 page_counter_init(&memcg->swap, NULL);
5214 page_counter_init(&memcg->kmem, NULL);
5215 page_counter_init(&memcg->tcpmem, NULL);
5217 root_mem_cgroup = memcg;
5221 /* The following stuff does not apply to the root */
5222 error = memcg_online_kmem(memcg);
5226 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5227 static_branch_inc(&memcg_sockets_enabled_key);
5231 mem_cgroup_id_remove(memcg);
5232 mem_cgroup_free(memcg);
5233 return ERR_PTR(error);
5236 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5238 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5241 * A memcg must be visible for expand_shrinker_info()
5242 * by the time the maps are allocated. So, we allocate maps
5243 * here, when for_each_mem_cgroup() can't skip it.
5245 if (alloc_shrinker_info(memcg)) {
5246 mem_cgroup_id_remove(memcg);
5250 /* Online state pins memcg ID, memcg ID pins CSS */
5251 refcount_set(&memcg->id.ref, 1);
5254 if (unlikely(mem_cgroup_is_root(memcg)))
5255 queue_delayed_work(system_unbound_wq, &stats_flush_dwork,
5260 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5262 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5263 struct mem_cgroup_event *event, *tmp;
5266 * Unregister events and notify userspace.
5267 * Notify userspace about cgroup removing only after rmdir of cgroup
5268 * directory to avoid race between userspace and kernelspace.
5270 spin_lock(&memcg->event_list_lock);
5271 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5272 list_del_init(&event->list);
5273 schedule_work(&event->remove);
5275 spin_unlock(&memcg->event_list_lock);
5277 page_counter_set_min(&memcg->memory, 0);
5278 page_counter_set_low(&memcg->memory, 0);
5280 memcg_offline_kmem(memcg);
5281 reparent_shrinker_deferred(memcg);
5282 wb_memcg_offline(memcg);
5284 drain_all_stock(memcg);
5286 mem_cgroup_id_put(memcg);
5289 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5291 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5293 invalidate_reclaim_iterators(memcg);
5296 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5298 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5299 int __maybe_unused i;
5301 #ifdef CONFIG_CGROUP_WRITEBACK
5302 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5303 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5305 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5306 static_branch_dec(&memcg_sockets_enabled_key);
5308 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5309 static_branch_dec(&memcg_sockets_enabled_key);
5311 vmpressure_cleanup(&memcg->vmpressure);
5312 cancel_work_sync(&memcg->high_work);
5313 mem_cgroup_remove_from_trees(memcg);
5314 free_shrinker_info(memcg);
5315 memcg_free_kmem(memcg);
5316 mem_cgroup_free(memcg);
5320 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5321 * @css: the target css
5323 * Reset the states of the mem_cgroup associated with @css. This is
5324 * invoked when the userland requests disabling on the default hierarchy
5325 * but the memcg is pinned through dependency. The memcg should stop
5326 * applying policies and should revert to the vanilla state as it may be
5327 * made visible again.
5329 * The current implementation only resets the essential configurations.
5330 * This needs to be expanded to cover all the visible parts.
5332 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5334 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5336 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5337 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5338 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5339 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5340 page_counter_set_min(&memcg->memory, 0);
5341 page_counter_set_low(&memcg->memory, 0);
5342 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5343 memcg->soft_limit = PAGE_COUNTER_MAX;
5344 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5345 memcg_wb_domain_size_changed(memcg);
5348 void mem_cgroup_flush_stats(void)
5350 if (!spin_trylock(&stats_flush_lock))
5353 cgroup_rstat_flush_irqsafe(root_mem_cgroup->css.cgroup);
5354 spin_unlock(&stats_flush_lock);
5357 static void flush_memcg_stats_dwork(struct work_struct *w)
5359 mem_cgroup_flush_stats();
5360 queue_delayed_work(system_unbound_wq, &stats_flush_dwork, 2UL*HZ);
5363 static void flush_memcg_stats_work(struct work_struct *w)
5365 mem_cgroup_flush_stats();
5368 static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state *css, int cpu)
5370 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5371 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5372 struct memcg_vmstats_percpu *statc;
5376 statc = per_cpu_ptr(memcg->vmstats_percpu, cpu);
5378 for (i = 0; i < MEMCG_NR_STAT; i++) {
5380 * Collect the aggregated propagation counts of groups
5381 * below us. We're in a per-cpu loop here and this is
5382 * a global counter, so the first cycle will get them.
5384 delta = memcg->vmstats.state_pending[i];
5386 memcg->vmstats.state_pending[i] = 0;
5388 /* Add CPU changes on this level since the last flush */
5389 v = READ_ONCE(statc->state[i]);
5390 if (v != statc->state_prev[i]) {
5391 delta += v - statc->state_prev[i];
5392 statc->state_prev[i] = v;
5398 /* Aggregate counts on this level and propagate upwards */
5399 memcg->vmstats.state[i] += delta;
5401 parent->vmstats.state_pending[i] += delta;
5404 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
5405 delta = memcg->vmstats.events_pending[i];
5407 memcg->vmstats.events_pending[i] = 0;
5409 v = READ_ONCE(statc->events[i]);
5410 if (v != statc->events_prev[i]) {
5411 delta += v - statc->events_prev[i];
5412 statc->events_prev[i] = v;
5418 memcg->vmstats.events[i] += delta;
5420 parent->vmstats.events_pending[i] += delta;
5423 for_each_node_state(nid, N_MEMORY) {
5424 struct mem_cgroup_per_node *pn = memcg->nodeinfo[nid];
5425 struct mem_cgroup_per_node *ppn = NULL;
5426 struct lruvec_stats_percpu *lstatc;
5429 ppn = parent->nodeinfo[nid];
5431 lstatc = per_cpu_ptr(pn->lruvec_stats_percpu, cpu);
5433 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
5434 delta = pn->lruvec_stats.state_pending[i];
5436 pn->lruvec_stats.state_pending[i] = 0;
5438 v = READ_ONCE(lstatc->state[i]);
5439 if (v != lstatc->state_prev[i]) {
5440 delta += v - lstatc->state_prev[i];
5441 lstatc->state_prev[i] = v;
5447 pn->lruvec_stats.state[i] += delta;
5449 ppn->lruvec_stats.state_pending[i] += delta;
5455 /* Handlers for move charge at task migration. */
5456 static int mem_cgroup_do_precharge(unsigned long count)
5460 /* Try a single bulk charge without reclaim first, kswapd may wake */
5461 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5463 mc.precharge += count;
5467 /* Try charges one by one with reclaim, but do not retry */
5469 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5483 enum mc_target_type {
5490 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5491 unsigned long addr, pte_t ptent)
5493 struct page *page = vm_normal_page(vma, addr, ptent);
5495 if (!page || !page_mapped(page))
5497 if (PageAnon(page)) {
5498 if (!(mc.flags & MOVE_ANON))
5501 if (!(mc.flags & MOVE_FILE))
5504 if (!get_page_unless_zero(page))
5510 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5511 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5512 pte_t ptent, swp_entry_t *entry)
5514 struct page *page = NULL;
5515 swp_entry_t ent = pte_to_swp_entry(ptent);
5517 if (!(mc.flags & MOVE_ANON))
5521 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5522 * a device and because they are not accessible by CPU they are store
5523 * as special swap entry in the CPU page table.
5525 if (is_device_private_entry(ent)) {
5526 page = pfn_swap_entry_to_page(ent);
5528 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5529 * a refcount of 1 when free (unlike normal page)
5531 if (!page_ref_add_unless(page, 1, 1))
5536 if (non_swap_entry(ent))
5540 * Because lookup_swap_cache() updates some statistics counter,
5541 * we call find_get_page() with swapper_space directly.
5543 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5544 entry->val = ent.val;
5549 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5550 pte_t ptent, swp_entry_t *entry)
5556 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5557 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5559 if (!vma->vm_file) /* anonymous vma */
5561 if (!(mc.flags & MOVE_FILE))
5564 /* page is moved even if it's not RSS of this task(page-faulted). */
5565 /* shmem/tmpfs may report page out on swap: account for that too. */
5566 return find_get_incore_page(vma->vm_file->f_mapping,
5567 linear_page_index(vma, addr));
5571 * mem_cgroup_move_account - move account of the page
5573 * @compound: charge the page as compound or small page
5574 * @from: mem_cgroup which the page is moved from.
5575 * @to: mem_cgroup which the page is moved to. @from != @to.
5577 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5579 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5582 static int mem_cgroup_move_account(struct page *page,
5584 struct mem_cgroup *from,
5585 struct mem_cgroup *to)
5587 struct lruvec *from_vec, *to_vec;
5588 struct pglist_data *pgdat;
5589 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5592 VM_BUG_ON(from == to);
5593 VM_BUG_ON_PAGE(PageLRU(page), page);
5594 VM_BUG_ON(compound && !PageTransHuge(page));
5597 * Prevent mem_cgroup_migrate() from looking at
5598 * page's memory cgroup of its source page while we change it.
5601 if (!trylock_page(page))
5605 if (page_memcg(page) != from)
5608 pgdat = page_pgdat(page);
5609 from_vec = mem_cgroup_lruvec(from, pgdat);
5610 to_vec = mem_cgroup_lruvec(to, pgdat);
5612 lock_page_memcg(page);
5614 if (PageAnon(page)) {
5615 if (page_mapped(page)) {
5616 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5617 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5618 if (PageTransHuge(page)) {
5619 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5621 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5626 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5627 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5629 if (PageSwapBacked(page)) {
5630 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5631 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5634 if (page_mapped(page)) {
5635 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5636 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5639 if (PageDirty(page)) {
5640 struct address_space *mapping = page_mapping(page);
5642 if (mapping_can_writeback(mapping)) {
5643 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5645 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5651 if (PageWriteback(page)) {
5652 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5653 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5657 * All state has been migrated, let's switch to the new memcg.
5659 * It is safe to change page's memcg here because the page
5660 * is referenced, charged, isolated, and locked: we can't race
5661 * with (un)charging, migration, LRU putback, or anything else
5662 * that would rely on a stable page's memory cgroup.
5664 * Note that lock_page_memcg is a memcg lock, not a page lock,
5665 * to save space. As soon as we switch page's memory cgroup to a
5666 * new memcg that isn't locked, the above state can change
5667 * concurrently again. Make sure we're truly done with it.
5672 css_put(&from->css);
5674 page->memcg_data = (unsigned long)to;
5676 __unlock_page_memcg(from);
5680 local_irq_disable();
5681 mem_cgroup_charge_statistics(to, page, nr_pages);
5682 memcg_check_events(to, page);
5683 mem_cgroup_charge_statistics(from, page, -nr_pages);
5684 memcg_check_events(from, page);
5693 * get_mctgt_type - get target type of moving charge
5694 * @vma: the vma the pte to be checked belongs
5695 * @addr: the address corresponding to the pte to be checked
5696 * @ptent: the pte to be checked
5697 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5700 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5701 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5702 * move charge. if @target is not NULL, the page is stored in target->page
5703 * with extra refcnt got(Callers should handle it).
5704 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5705 * target for charge migration. if @target is not NULL, the entry is stored
5707 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5708 * (so ZONE_DEVICE page and thus not on the lru).
5709 * For now we such page is charge like a regular page would be as for all
5710 * intent and purposes it is just special memory taking the place of a
5713 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5715 * Called with pte lock held.
5718 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5719 unsigned long addr, pte_t ptent, union mc_target *target)
5721 struct page *page = NULL;
5722 enum mc_target_type ret = MC_TARGET_NONE;
5723 swp_entry_t ent = { .val = 0 };
5725 if (pte_present(ptent))
5726 page = mc_handle_present_pte(vma, addr, ptent);
5727 else if (is_swap_pte(ptent))
5728 page = mc_handle_swap_pte(vma, ptent, &ent);
5729 else if (pte_none(ptent))
5730 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5732 if (!page && !ent.val)
5736 * Do only loose check w/o serialization.
5737 * mem_cgroup_move_account() checks the page is valid or
5738 * not under LRU exclusion.
5740 if (page_memcg(page) == mc.from) {
5741 ret = MC_TARGET_PAGE;
5742 if (is_device_private_page(page))
5743 ret = MC_TARGET_DEVICE;
5745 target->page = page;
5747 if (!ret || !target)
5751 * There is a swap entry and a page doesn't exist or isn't charged.
5752 * But we cannot move a tail-page in a THP.
5754 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5755 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5756 ret = MC_TARGET_SWAP;
5763 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5765 * We don't consider PMD mapped swapping or file mapped pages because THP does
5766 * not support them for now.
5767 * Caller should make sure that pmd_trans_huge(pmd) is true.
5769 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5770 unsigned long addr, pmd_t pmd, union mc_target *target)
5772 struct page *page = NULL;
5773 enum mc_target_type ret = MC_TARGET_NONE;
5775 if (unlikely(is_swap_pmd(pmd))) {
5776 VM_BUG_ON(thp_migration_supported() &&
5777 !is_pmd_migration_entry(pmd));
5780 page = pmd_page(pmd);
5781 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5782 if (!(mc.flags & MOVE_ANON))
5784 if (page_memcg(page) == mc.from) {
5785 ret = MC_TARGET_PAGE;
5788 target->page = page;
5794 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5795 unsigned long addr, pmd_t pmd, union mc_target *target)
5797 return MC_TARGET_NONE;
5801 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5802 unsigned long addr, unsigned long end,
5803 struct mm_walk *walk)
5805 struct vm_area_struct *vma = walk->vma;
5809 ptl = pmd_trans_huge_lock(pmd, vma);
5812 * Note their can not be MC_TARGET_DEVICE for now as we do not
5813 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5814 * this might change.
5816 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5817 mc.precharge += HPAGE_PMD_NR;
5822 if (pmd_trans_unstable(pmd))
5824 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5825 for (; addr != end; pte++, addr += PAGE_SIZE)
5826 if (get_mctgt_type(vma, addr, *pte, NULL))
5827 mc.precharge++; /* increment precharge temporarily */
5828 pte_unmap_unlock(pte - 1, ptl);
5834 static const struct mm_walk_ops precharge_walk_ops = {
5835 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5838 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5840 unsigned long precharge;
5843 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5844 mmap_read_unlock(mm);
5846 precharge = mc.precharge;
5852 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5854 unsigned long precharge = mem_cgroup_count_precharge(mm);
5856 VM_BUG_ON(mc.moving_task);
5857 mc.moving_task = current;
5858 return mem_cgroup_do_precharge(precharge);
5861 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5862 static void __mem_cgroup_clear_mc(void)
5864 struct mem_cgroup *from = mc.from;
5865 struct mem_cgroup *to = mc.to;
5867 /* we must uncharge all the leftover precharges from mc.to */
5869 cancel_charge(mc.to, mc.precharge);
5873 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5874 * we must uncharge here.
5876 if (mc.moved_charge) {
5877 cancel_charge(mc.from, mc.moved_charge);
5878 mc.moved_charge = 0;
5880 /* we must fixup refcnts and charges */
5881 if (mc.moved_swap) {
5882 /* uncharge swap account from the old cgroup */
5883 if (!mem_cgroup_is_root(mc.from))
5884 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5886 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5889 * we charged both to->memory and to->memsw, so we
5890 * should uncharge to->memory.
5892 if (!mem_cgroup_is_root(mc.to))
5893 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5897 memcg_oom_recover(from);
5898 memcg_oom_recover(to);
5899 wake_up_all(&mc.waitq);
5902 static void mem_cgroup_clear_mc(void)
5904 struct mm_struct *mm = mc.mm;
5907 * we must clear moving_task before waking up waiters at the end of
5910 mc.moving_task = NULL;
5911 __mem_cgroup_clear_mc();
5912 spin_lock(&mc.lock);
5916 spin_unlock(&mc.lock);
5921 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5923 struct cgroup_subsys_state *css;
5924 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5925 struct mem_cgroup *from;
5926 struct task_struct *leader, *p;
5927 struct mm_struct *mm;
5928 unsigned long move_flags;
5931 /* charge immigration isn't supported on the default hierarchy */
5932 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5936 * Multi-process migrations only happen on the default hierarchy
5937 * where charge immigration is not used. Perform charge
5938 * immigration if @tset contains a leader and whine if there are
5942 cgroup_taskset_for_each_leader(leader, css, tset) {
5945 memcg = mem_cgroup_from_css(css);
5951 * We are now committed to this value whatever it is. Changes in this
5952 * tunable will only affect upcoming migrations, not the current one.
5953 * So we need to save it, and keep it going.
5955 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5959 from = mem_cgroup_from_task(p);
5961 VM_BUG_ON(from == memcg);
5963 mm = get_task_mm(p);
5966 /* We move charges only when we move a owner of the mm */
5967 if (mm->owner == p) {
5970 VM_BUG_ON(mc.precharge);
5971 VM_BUG_ON(mc.moved_charge);
5972 VM_BUG_ON(mc.moved_swap);
5974 spin_lock(&mc.lock);
5978 mc.flags = move_flags;
5979 spin_unlock(&mc.lock);
5980 /* We set mc.moving_task later */
5982 ret = mem_cgroup_precharge_mc(mm);
5984 mem_cgroup_clear_mc();
5991 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5994 mem_cgroup_clear_mc();
5997 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5998 unsigned long addr, unsigned long end,
5999 struct mm_walk *walk)
6002 struct vm_area_struct *vma = walk->vma;
6005 enum mc_target_type target_type;
6006 union mc_target target;
6009 ptl = pmd_trans_huge_lock(pmd, vma);
6011 if (mc.precharge < HPAGE_PMD_NR) {
6015 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6016 if (target_type == MC_TARGET_PAGE) {
6018 if (!isolate_lru_page(page)) {
6019 if (!mem_cgroup_move_account(page, true,
6021 mc.precharge -= HPAGE_PMD_NR;
6022 mc.moved_charge += HPAGE_PMD_NR;
6024 putback_lru_page(page);
6027 } else if (target_type == MC_TARGET_DEVICE) {
6029 if (!mem_cgroup_move_account(page, true,
6031 mc.precharge -= HPAGE_PMD_NR;
6032 mc.moved_charge += HPAGE_PMD_NR;
6040 if (pmd_trans_unstable(pmd))
6043 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6044 for (; addr != end; addr += PAGE_SIZE) {
6045 pte_t ptent = *(pte++);
6046 bool device = false;
6052 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6053 case MC_TARGET_DEVICE:
6056 case MC_TARGET_PAGE:
6059 * We can have a part of the split pmd here. Moving it
6060 * can be done but it would be too convoluted so simply
6061 * ignore such a partial THP and keep it in original
6062 * memcg. There should be somebody mapping the head.
6064 if (PageTransCompound(page))
6066 if (!device && isolate_lru_page(page))
6068 if (!mem_cgroup_move_account(page, false,
6071 /* we uncharge from mc.from later. */
6075 putback_lru_page(page);
6076 put: /* get_mctgt_type() gets the page */
6079 case MC_TARGET_SWAP:
6081 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6083 mem_cgroup_id_get_many(mc.to, 1);
6084 /* we fixup other refcnts and charges later. */
6092 pte_unmap_unlock(pte - 1, ptl);
6097 * We have consumed all precharges we got in can_attach().
6098 * We try charge one by one, but don't do any additional
6099 * charges to mc.to if we have failed in charge once in attach()
6102 ret = mem_cgroup_do_precharge(1);
6110 static const struct mm_walk_ops charge_walk_ops = {
6111 .pmd_entry = mem_cgroup_move_charge_pte_range,
6114 static void mem_cgroup_move_charge(void)
6116 lru_add_drain_all();
6118 * Signal lock_page_memcg() to take the memcg's move_lock
6119 * while we're moving its pages to another memcg. Then wait
6120 * for already started RCU-only updates to finish.
6122 atomic_inc(&mc.from->moving_account);
6125 if (unlikely(!mmap_read_trylock(mc.mm))) {
6127 * Someone who are holding the mmap_lock might be waiting in
6128 * waitq. So we cancel all extra charges, wake up all waiters,
6129 * and retry. Because we cancel precharges, we might not be able
6130 * to move enough charges, but moving charge is a best-effort
6131 * feature anyway, so it wouldn't be a big problem.
6133 __mem_cgroup_clear_mc();
6138 * When we have consumed all precharges and failed in doing
6139 * additional charge, the page walk just aborts.
6141 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6144 mmap_read_unlock(mc.mm);
6145 atomic_dec(&mc.from->moving_account);
6148 static void mem_cgroup_move_task(void)
6151 mem_cgroup_move_charge();
6152 mem_cgroup_clear_mc();
6155 #else /* !CONFIG_MMU */
6156 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6160 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6163 static void mem_cgroup_move_task(void)
6168 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6170 if (value == PAGE_COUNTER_MAX)
6171 seq_puts(m, "max\n");
6173 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6178 static u64 memory_current_read(struct cgroup_subsys_state *css,
6181 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6183 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6186 static int memory_min_show(struct seq_file *m, void *v)
6188 return seq_puts_memcg_tunable(m,
6189 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6192 static ssize_t memory_min_write(struct kernfs_open_file *of,
6193 char *buf, size_t nbytes, loff_t off)
6195 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6199 buf = strstrip(buf);
6200 err = page_counter_memparse(buf, "max", &min);
6204 page_counter_set_min(&memcg->memory, min);
6209 static int memory_low_show(struct seq_file *m, void *v)
6211 return seq_puts_memcg_tunable(m,
6212 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6215 static ssize_t memory_low_write(struct kernfs_open_file *of,
6216 char *buf, size_t nbytes, loff_t off)
6218 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6222 buf = strstrip(buf);
6223 err = page_counter_memparse(buf, "max", &low);
6227 page_counter_set_low(&memcg->memory, low);
6232 static int memory_high_show(struct seq_file *m, void *v)
6234 return seq_puts_memcg_tunable(m,
6235 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6238 static ssize_t memory_high_write(struct kernfs_open_file *of,
6239 char *buf, size_t nbytes, loff_t off)
6241 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6242 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6243 bool drained = false;
6247 buf = strstrip(buf);
6248 err = page_counter_memparse(buf, "max", &high);
6252 page_counter_set_high(&memcg->memory, high);
6255 unsigned long nr_pages = page_counter_read(&memcg->memory);
6256 unsigned long reclaimed;
6258 if (nr_pages <= high)
6261 if (signal_pending(current))
6265 drain_all_stock(memcg);
6270 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6273 if (!reclaimed && !nr_retries--)
6277 memcg_wb_domain_size_changed(memcg);
6281 static int memory_max_show(struct seq_file *m, void *v)
6283 return seq_puts_memcg_tunable(m,
6284 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6287 static ssize_t memory_max_write(struct kernfs_open_file *of,
6288 char *buf, size_t nbytes, loff_t off)
6290 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6291 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6292 bool drained = false;
6296 buf = strstrip(buf);
6297 err = page_counter_memparse(buf, "max", &max);
6301 xchg(&memcg->memory.max, max);
6304 unsigned long nr_pages = page_counter_read(&memcg->memory);
6306 if (nr_pages <= max)
6309 if (signal_pending(current))
6313 drain_all_stock(memcg);
6319 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6325 memcg_memory_event(memcg, MEMCG_OOM);
6326 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6330 memcg_wb_domain_size_changed(memcg);
6334 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6336 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6337 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6338 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6339 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6340 seq_printf(m, "oom_kill %lu\n",
6341 atomic_long_read(&events[MEMCG_OOM_KILL]));
6344 static int memory_events_show(struct seq_file *m, void *v)
6346 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6348 __memory_events_show(m, memcg->memory_events);
6352 static int memory_events_local_show(struct seq_file *m, void *v)
6354 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6356 __memory_events_show(m, memcg->memory_events_local);
6360 static int memory_stat_show(struct seq_file *m, void *v)
6362 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6365 buf = memory_stat_format(memcg);
6374 static inline unsigned long lruvec_page_state_output(struct lruvec *lruvec,
6377 return lruvec_page_state(lruvec, item) * memcg_page_state_unit(item);
6380 static int memory_numa_stat_show(struct seq_file *m, void *v)
6383 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6385 cgroup_rstat_flush(memcg->css.cgroup);
6387 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6390 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6393 seq_printf(m, "%s", memory_stats[i].name);
6394 for_each_node_state(nid, N_MEMORY) {
6396 struct lruvec *lruvec;
6398 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6399 size = lruvec_page_state_output(lruvec,
6400 memory_stats[i].idx);
6401 seq_printf(m, " N%d=%llu", nid, size);
6410 static int memory_oom_group_show(struct seq_file *m, void *v)
6412 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6414 seq_printf(m, "%d\n", memcg->oom_group);
6419 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6420 char *buf, size_t nbytes, loff_t off)
6422 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6425 buf = strstrip(buf);
6429 ret = kstrtoint(buf, 0, &oom_group);
6433 if (oom_group != 0 && oom_group != 1)
6436 memcg->oom_group = oom_group;
6441 static struct cftype memory_files[] = {
6444 .flags = CFTYPE_NOT_ON_ROOT,
6445 .read_u64 = memory_current_read,
6449 .flags = CFTYPE_NOT_ON_ROOT,
6450 .seq_show = memory_min_show,
6451 .write = memory_min_write,
6455 .flags = CFTYPE_NOT_ON_ROOT,
6456 .seq_show = memory_low_show,
6457 .write = memory_low_write,
6461 .flags = CFTYPE_NOT_ON_ROOT,
6462 .seq_show = memory_high_show,
6463 .write = memory_high_write,
6467 .flags = CFTYPE_NOT_ON_ROOT,
6468 .seq_show = memory_max_show,
6469 .write = memory_max_write,
6473 .flags = CFTYPE_NOT_ON_ROOT,
6474 .file_offset = offsetof(struct mem_cgroup, events_file),
6475 .seq_show = memory_events_show,
6478 .name = "events.local",
6479 .flags = CFTYPE_NOT_ON_ROOT,
6480 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6481 .seq_show = memory_events_local_show,
6485 .seq_show = memory_stat_show,
6489 .name = "numa_stat",
6490 .seq_show = memory_numa_stat_show,
6494 .name = "oom.group",
6495 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6496 .seq_show = memory_oom_group_show,
6497 .write = memory_oom_group_write,
6502 struct cgroup_subsys memory_cgrp_subsys = {
6503 .css_alloc = mem_cgroup_css_alloc,
6504 .css_online = mem_cgroup_css_online,
6505 .css_offline = mem_cgroup_css_offline,
6506 .css_released = mem_cgroup_css_released,
6507 .css_free = mem_cgroup_css_free,
6508 .css_reset = mem_cgroup_css_reset,
6509 .css_rstat_flush = mem_cgroup_css_rstat_flush,
6510 .can_attach = mem_cgroup_can_attach,
6511 .cancel_attach = mem_cgroup_cancel_attach,
6512 .post_attach = mem_cgroup_move_task,
6513 .dfl_cftypes = memory_files,
6514 .legacy_cftypes = mem_cgroup_legacy_files,
6519 * This function calculates an individual cgroup's effective
6520 * protection which is derived from its own memory.min/low, its
6521 * parent's and siblings' settings, as well as the actual memory
6522 * distribution in the tree.
6524 * The following rules apply to the effective protection values:
6526 * 1. At the first level of reclaim, effective protection is equal to
6527 * the declared protection in memory.min and memory.low.
6529 * 2. To enable safe delegation of the protection configuration, at
6530 * subsequent levels the effective protection is capped to the
6531 * parent's effective protection.
6533 * 3. To make complex and dynamic subtrees easier to configure, the
6534 * user is allowed to overcommit the declared protection at a given
6535 * level. If that is the case, the parent's effective protection is
6536 * distributed to the children in proportion to how much protection
6537 * they have declared and how much of it they are utilizing.
6539 * This makes distribution proportional, but also work-conserving:
6540 * if one cgroup claims much more protection than it uses memory,
6541 * the unused remainder is available to its siblings.
6543 * 4. Conversely, when the declared protection is undercommitted at a
6544 * given level, the distribution of the larger parental protection
6545 * budget is NOT proportional. A cgroup's protection from a sibling
6546 * is capped to its own memory.min/low setting.
6548 * 5. However, to allow protecting recursive subtrees from each other
6549 * without having to declare each individual cgroup's fixed share
6550 * of the ancestor's claim to protection, any unutilized -
6551 * "floating" - protection from up the tree is distributed in
6552 * proportion to each cgroup's *usage*. This makes the protection
6553 * neutral wrt sibling cgroups and lets them compete freely over
6554 * the shared parental protection budget, but it protects the
6555 * subtree as a whole from neighboring subtrees.
6557 * Note that 4. and 5. are not in conflict: 4. is about protecting
6558 * against immediate siblings whereas 5. is about protecting against
6559 * neighboring subtrees.
6561 static unsigned long effective_protection(unsigned long usage,
6562 unsigned long parent_usage,
6563 unsigned long setting,
6564 unsigned long parent_effective,
6565 unsigned long siblings_protected)
6567 unsigned long protected;
6570 protected = min(usage, setting);
6572 * If all cgroups at this level combined claim and use more
6573 * protection then what the parent affords them, distribute
6574 * shares in proportion to utilization.
6576 * We are using actual utilization rather than the statically
6577 * claimed protection in order to be work-conserving: claimed
6578 * but unused protection is available to siblings that would
6579 * otherwise get a smaller chunk than what they claimed.
6581 if (siblings_protected > parent_effective)
6582 return protected * parent_effective / siblings_protected;
6585 * Ok, utilized protection of all children is within what the
6586 * parent affords them, so we know whatever this child claims
6587 * and utilizes is effectively protected.
6589 * If there is unprotected usage beyond this value, reclaim
6590 * will apply pressure in proportion to that amount.
6592 * If there is unutilized protection, the cgroup will be fully
6593 * shielded from reclaim, but we do return a smaller value for
6594 * protection than what the group could enjoy in theory. This
6595 * is okay. With the overcommit distribution above, effective
6596 * protection is always dependent on how memory is actually
6597 * consumed among the siblings anyway.
6602 * If the children aren't claiming (all of) the protection
6603 * afforded to them by the parent, distribute the remainder in
6604 * proportion to the (unprotected) memory of each cgroup. That
6605 * way, cgroups that aren't explicitly prioritized wrt each
6606 * other compete freely over the allowance, but they are
6607 * collectively protected from neighboring trees.
6609 * We're using unprotected memory for the weight so that if
6610 * some cgroups DO claim explicit protection, we don't protect
6611 * the same bytes twice.
6613 * Check both usage and parent_usage against the respective
6614 * protected values. One should imply the other, but they
6615 * aren't read atomically - make sure the division is sane.
6617 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6619 if (parent_effective > siblings_protected &&
6620 parent_usage > siblings_protected &&
6621 usage > protected) {
6622 unsigned long unclaimed;
6624 unclaimed = parent_effective - siblings_protected;
6625 unclaimed *= usage - protected;
6626 unclaimed /= parent_usage - siblings_protected;
6635 * mem_cgroup_calculate_protection - check if memory consumption is in the normal range
6636 * @root: the top ancestor of the sub-tree being checked
6637 * @memcg: the memory cgroup to check
6639 * WARNING: This function is not stateless! It can only be used as part
6640 * of a top-down tree iteration, not for isolated queries.
6642 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6643 struct mem_cgroup *memcg)
6645 unsigned long usage, parent_usage;
6646 struct mem_cgroup *parent;
6648 if (mem_cgroup_disabled())
6652 root = root_mem_cgroup;
6655 * Effective values of the reclaim targets are ignored so they
6656 * can be stale. Have a look at mem_cgroup_protection for more
6658 * TODO: calculation should be more robust so that we do not need
6659 * that special casing.
6664 usage = page_counter_read(&memcg->memory);
6668 parent = parent_mem_cgroup(memcg);
6669 /* No parent means a non-hierarchical mode on v1 memcg */
6673 if (parent == root) {
6674 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6675 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6679 parent_usage = page_counter_read(&parent->memory);
6681 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6682 READ_ONCE(memcg->memory.min),
6683 READ_ONCE(parent->memory.emin),
6684 atomic_long_read(&parent->memory.children_min_usage)));
6686 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6687 READ_ONCE(memcg->memory.low),
6688 READ_ONCE(parent->memory.elow),
6689 atomic_long_read(&parent->memory.children_low_usage)));
6692 static int charge_memcg(struct page *page, struct mem_cgroup *memcg, gfp_t gfp)
6694 unsigned int nr_pages = thp_nr_pages(page);
6697 ret = try_charge(memcg, gfp, nr_pages);
6701 css_get(&memcg->css);
6702 commit_charge(page, memcg);
6704 local_irq_disable();
6705 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6706 memcg_check_events(memcg, page);
6713 * __mem_cgroup_charge - charge a newly allocated page to a cgroup
6714 * @page: page to charge
6715 * @mm: mm context of the victim
6716 * @gfp_mask: reclaim mode
6718 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6719 * pages according to @gfp_mask if necessary. if @mm is NULL, try to
6720 * charge to the active memcg.
6722 * Do not use this for pages allocated for swapin.
6724 * Returns 0 on success. Otherwise, an error code is returned.
6726 int __mem_cgroup_charge(struct page *page, struct mm_struct *mm,
6729 struct mem_cgroup *memcg;
6732 memcg = get_mem_cgroup_from_mm(mm);
6733 ret = charge_memcg(page, memcg, gfp_mask);
6734 css_put(&memcg->css);
6740 * mem_cgroup_swapin_charge_page - charge a newly allocated page for swapin
6741 * @page: page to charge
6742 * @mm: mm context of the victim
6743 * @gfp: reclaim mode
6744 * @entry: swap entry for which the page is allocated
6746 * This function charges a page allocated for swapin. Please call this before
6747 * adding the page to the swapcache.
6749 * Returns 0 on success. Otherwise, an error code is returned.
6751 int mem_cgroup_swapin_charge_page(struct page *page, struct mm_struct *mm,
6752 gfp_t gfp, swp_entry_t entry)
6754 struct mem_cgroup *memcg;
6758 if (mem_cgroup_disabled())
6761 id = lookup_swap_cgroup_id(entry);
6763 memcg = mem_cgroup_from_id(id);
6764 if (!memcg || !css_tryget_online(&memcg->css))
6765 memcg = get_mem_cgroup_from_mm(mm);
6768 ret = charge_memcg(page, memcg, gfp);
6770 css_put(&memcg->css);
6775 * mem_cgroup_swapin_uncharge_swap - uncharge swap slot
6776 * @entry: swap entry for which the page is charged
6778 * Call this function after successfully adding the charged page to swapcache.
6780 * Note: This function assumes the page for which swap slot is being uncharged
6783 void mem_cgroup_swapin_uncharge_swap(swp_entry_t entry)
6786 * Cgroup1's unified memory+swap counter has been charged with the
6787 * new swapcache page, finish the transfer by uncharging the swap
6788 * slot. The swap slot would also get uncharged when it dies, but
6789 * it can stick around indefinitely and we'd count the page twice
6792 * Cgroup2 has separate resource counters for memory and swap,
6793 * so this is a non-issue here. Memory and swap charge lifetimes
6794 * correspond 1:1 to page and swap slot lifetimes: we charge the
6795 * page to memory here, and uncharge swap when the slot is freed.
6797 if (!mem_cgroup_disabled() && do_memsw_account()) {
6799 * The swap entry might not get freed for a long time,
6800 * let's not wait for it. The page already received a
6801 * memory+swap charge, drop the swap entry duplicate.
6803 mem_cgroup_uncharge_swap(entry, 1);
6807 struct uncharge_gather {
6808 struct mem_cgroup *memcg;
6809 unsigned long nr_memory;
6810 unsigned long pgpgout;
6811 unsigned long nr_kmem;
6812 struct page *dummy_page;
6815 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6817 memset(ug, 0, sizeof(*ug));
6820 static void uncharge_batch(const struct uncharge_gather *ug)
6822 unsigned long flags;
6824 if (ug->nr_memory) {
6825 page_counter_uncharge(&ug->memcg->memory, ug->nr_memory);
6826 if (do_memsw_account())
6827 page_counter_uncharge(&ug->memcg->memsw, ug->nr_memory);
6828 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6829 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6830 memcg_oom_recover(ug->memcg);
6833 local_irq_save(flags);
6834 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6835 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_memory);
6836 memcg_check_events(ug->memcg, ug->dummy_page);
6837 local_irq_restore(flags);
6839 /* drop reference from uncharge_page */
6840 css_put(&ug->memcg->css);
6843 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6845 unsigned long nr_pages;
6846 struct mem_cgroup *memcg;
6847 struct obj_cgroup *objcg;
6848 bool use_objcg = PageMemcgKmem(page);
6850 VM_BUG_ON_PAGE(PageLRU(page), page);
6853 * Nobody should be changing or seriously looking at
6854 * page memcg or objcg at this point, we have fully
6855 * exclusive access to the page.
6858 objcg = __page_objcg(page);
6860 * This get matches the put at the end of the function and
6861 * kmem pages do not hold memcg references anymore.
6863 memcg = get_mem_cgroup_from_objcg(objcg);
6865 memcg = __page_memcg(page);
6871 if (ug->memcg != memcg) {
6874 uncharge_gather_clear(ug);
6877 ug->dummy_page = page;
6879 /* pairs with css_put in uncharge_batch */
6880 css_get(&memcg->css);
6883 nr_pages = compound_nr(page);
6886 ug->nr_memory += nr_pages;
6887 ug->nr_kmem += nr_pages;
6889 page->memcg_data = 0;
6890 obj_cgroup_put(objcg);
6892 /* LRU pages aren't accounted at the root level */
6893 if (!mem_cgroup_is_root(memcg))
6894 ug->nr_memory += nr_pages;
6897 page->memcg_data = 0;
6900 css_put(&memcg->css);
6904 * __mem_cgroup_uncharge - uncharge a page
6905 * @page: page to uncharge
6907 * Uncharge a page previously charged with __mem_cgroup_charge().
6909 void __mem_cgroup_uncharge(struct page *page)
6911 struct uncharge_gather ug;
6913 /* Don't touch page->lru of any random page, pre-check: */
6914 if (!page_memcg(page))
6917 uncharge_gather_clear(&ug);
6918 uncharge_page(page, &ug);
6919 uncharge_batch(&ug);
6923 * __mem_cgroup_uncharge_list - uncharge a list of page
6924 * @page_list: list of pages to uncharge
6926 * Uncharge a list of pages previously charged with
6927 * __mem_cgroup_charge().
6929 void __mem_cgroup_uncharge_list(struct list_head *page_list)
6931 struct uncharge_gather ug;
6934 uncharge_gather_clear(&ug);
6935 list_for_each_entry(page, page_list, lru)
6936 uncharge_page(page, &ug);
6938 uncharge_batch(&ug);
6942 * mem_cgroup_migrate - charge a page's replacement
6943 * @oldpage: currently circulating page
6944 * @newpage: replacement page
6946 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6947 * be uncharged upon free.
6949 * Both pages must be locked, @newpage->mapping must be set up.
6951 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6953 struct mem_cgroup *memcg;
6954 unsigned int nr_pages;
6955 unsigned long flags;
6957 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6958 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6959 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6960 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6963 if (mem_cgroup_disabled())
6966 /* Page cache replacement: new page already charged? */
6967 if (page_memcg(newpage))
6970 memcg = page_memcg(oldpage);
6971 VM_WARN_ON_ONCE_PAGE(!memcg, oldpage);
6975 /* Force-charge the new page. The old one will be freed soon */
6976 nr_pages = thp_nr_pages(newpage);
6978 if (!mem_cgroup_is_root(memcg)) {
6979 page_counter_charge(&memcg->memory, nr_pages);
6980 if (do_memsw_account())
6981 page_counter_charge(&memcg->memsw, nr_pages);
6984 css_get(&memcg->css);
6985 commit_charge(newpage, memcg);
6987 local_irq_save(flags);
6988 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
6989 memcg_check_events(memcg, newpage);
6990 local_irq_restore(flags);
6993 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6994 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6996 void mem_cgroup_sk_alloc(struct sock *sk)
6998 struct mem_cgroup *memcg;
7000 if (!mem_cgroup_sockets_enabled)
7003 /* Do not associate the sock with unrelated interrupted task's memcg. */
7008 memcg = mem_cgroup_from_task(current);
7009 if (memcg == root_mem_cgroup)
7011 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
7013 if (css_tryget(&memcg->css))
7014 sk->sk_memcg = memcg;
7019 void mem_cgroup_sk_free(struct sock *sk)
7022 css_put(&sk->sk_memcg->css);
7026 * mem_cgroup_charge_skmem - charge socket memory
7027 * @memcg: memcg to charge
7028 * @nr_pages: number of pages to charge
7030 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7031 * @memcg's configured limit, %false if the charge had to be forced.
7033 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7035 gfp_t gfp_mask = GFP_KERNEL;
7037 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7038 struct page_counter *fail;
7040 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7041 memcg->tcpmem_pressure = 0;
7044 page_counter_charge(&memcg->tcpmem, nr_pages);
7045 memcg->tcpmem_pressure = 1;
7049 /* Don't block in the packet receive path */
7051 gfp_mask = GFP_NOWAIT;
7053 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7055 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
7058 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
7063 * mem_cgroup_uncharge_skmem - uncharge socket memory
7064 * @memcg: memcg to uncharge
7065 * @nr_pages: number of pages to uncharge
7067 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7069 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7070 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7074 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7076 refill_stock(memcg, nr_pages);
7079 static int __init cgroup_memory(char *s)
7083 while ((token = strsep(&s, ",")) != NULL) {
7086 if (!strcmp(token, "nosocket"))
7087 cgroup_memory_nosocket = true;
7088 if (!strcmp(token, "nokmem"))
7089 cgroup_memory_nokmem = true;
7093 __setup("cgroup.memory=", cgroup_memory);
7096 * subsys_initcall() for memory controller.
7098 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7099 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7100 * basically everything that doesn't depend on a specific mem_cgroup structure
7101 * should be initialized from here.
7103 static int __init mem_cgroup_init(void)
7108 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7109 * used for per-memcg-per-cpu caching of per-node statistics. In order
7110 * to work fine, we should make sure that the overfill threshold can't
7111 * exceed S32_MAX / PAGE_SIZE.
7113 BUILD_BUG_ON(MEMCG_CHARGE_BATCH > S32_MAX / PAGE_SIZE);
7115 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7116 memcg_hotplug_cpu_dead);
7118 for_each_possible_cpu(cpu)
7119 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7122 for_each_node(node) {
7123 struct mem_cgroup_tree_per_node *rtpn;
7125 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7126 node_online(node) ? node : NUMA_NO_NODE);
7128 rtpn->rb_root = RB_ROOT;
7129 rtpn->rb_rightmost = NULL;
7130 spin_lock_init(&rtpn->lock);
7131 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7136 subsys_initcall(mem_cgroup_init);
7138 #ifdef CONFIG_MEMCG_SWAP
7139 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7141 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7143 * The root cgroup cannot be destroyed, so it's refcount must
7146 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7150 memcg = parent_mem_cgroup(memcg);
7152 memcg = root_mem_cgroup;
7158 * mem_cgroup_swapout - transfer a memsw charge to swap
7159 * @page: page whose memsw charge to transfer
7160 * @entry: swap entry to move the charge to
7162 * Transfer the memsw charge of @page to @entry.
7164 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7166 struct mem_cgroup *memcg, *swap_memcg;
7167 unsigned int nr_entries;
7168 unsigned short oldid;
7170 VM_BUG_ON_PAGE(PageLRU(page), page);
7171 VM_BUG_ON_PAGE(page_count(page), page);
7173 if (mem_cgroup_disabled())
7176 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7179 memcg = page_memcg(page);
7181 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7186 * In case the memcg owning these pages has been offlined and doesn't
7187 * have an ID allocated to it anymore, charge the closest online
7188 * ancestor for the swap instead and transfer the memory+swap charge.
7190 swap_memcg = mem_cgroup_id_get_online(memcg);
7191 nr_entries = thp_nr_pages(page);
7192 /* Get references for the tail pages, too */
7194 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7195 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7197 VM_BUG_ON_PAGE(oldid, page);
7198 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7200 page->memcg_data = 0;
7202 if (!mem_cgroup_is_root(memcg))
7203 page_counter_uncharge(&memcg->memory, nr_entries);
7205 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7206 if (!mem_cgroup_is_root(swap_memcg))
7207 page_counter_charge(&swap_memcg->memsw, nr_entries);
7208 page_counter_uncharge(&memcg->memsw, nr_entries);
7212 * Interrupts should be disabled here because the caller holds the
7213 * i_pages lock which is taken with interrupts-off. It is
7214 * important here to have the interrupts disabled because it is the
7215 * only synchronisation we have for updating the per-CPU variables.
7217 VM_BUG_ON(!irqs_disabled());
7218 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7219 memcg_check_events(memcg, page);
7221 css_put(&memcg->css);
7225 * __mem_cgroup_try_charge_swap - try charging swap space for a page
7226 * @page: page being added to swap
7227 * @entry: swap entry to charge
7229 * Try to charge @page's memcg for the swap space at @entry.
7231 * Returns 0 on success, -ENOMEM on failure.
7233 int __mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7235 unsigned int nr_pages = thp_nr_pages(page);
7236 struct page_counter *counter;
7237 struct mem_cgroup *memcg;
7238 unsigned short oldid;
7240 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7243 memcg = page_memcg(page);
7245 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7250 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7254 memcg = mem_cgroup_id_get_online(memcg);
7256 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7257 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7258 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7259 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7260 mem_cgroup_id_put(memcg);
7264 /* Get references for the tail pages, too */
7266 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7267 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7268 VM_BUG_ON_PAGE(oldid, page);
7269 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7275 * __mem_cgroup_uncharge_swap - uncharge swap space
7276 * @entry: swap entry to uncharge
7277 * @nr_pages: the amount of swap space to uncharge
7279 void __mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7281 struct mem_cgroup *memcg;
7284 id = swap_cgroup_record(entry, 0, nr_pages);
7286 memcg = mem_cgroup_from_id(id);
7288 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7289 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7290 page_counter_uncharge(&memcg->swap, nr_pages);
7292 page_counter_uncharge(&memcg->memsw, nr_pages);
7294 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7295 mem_cgroup_id_put_many(memcg, nr_pages);
7300 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7302 long nr_swap_pages = get_nr_swap_pages();
7304 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7305 return nr_swap_pages;
7306 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7307 nr_swap_pages = min_t(long, nr_swap_pages,
7308 READ_ONCE(memcg->swap.max) -
7309 page_counter_read(&memcg->swap));
7310 return nr_swap_pages;
7313 bool mem_cgroup_swap_full(struct page *page)
7315 struct mem_cgroup *memcg;
7317 VM_BUG_ON_PAGE(!PageLocked(page), page);
7321 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7324 memcg = page_memcg(page);
7328 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7329 unsigned long usage = page_counter_read(&memcg->swap);
7331 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7332 usage * 2 >= READ_ONCE(memcg->swap.max))
7339 static int __init setup_swap_account(char *s)
7341 if (!strcmp(s, "1"))
7342 cgroup_memory_noswap = false;
7343 else if (!strcmp(s, "0"))
7344 cgroup_memory_noswap = true;
7347 __setup("swapaccount=", setup_swap_account);
7349 static u64 swap_current_read(struct cgroup_subsys_state *css,
7352 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7354 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7357 static int swap_high_show(struct seq_file *m, void *v)
7359 return seq_puts_memcg_tunable(m,
7360 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7363 static ssize_t swap_high_write(struct kernfs_open_file *of,
7364 char *buf, size_t nbytes, loff_t off)
7366 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7370 buf = strstrip(buf);
7371 err = page_counter_memparse(buf, "max", &high);
7375 page_counter_set_high(&memcg->swap, high);
7380 static int swap_max_show(struct seq_file *m, void *v)
7382 return seq_puts_memcg_tunable(m,
7383 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7386 static ssize_t swap_max_write(struct kernfs_open_file *of,
7387 char *buf, size_t nbytes, loff_t off)
7389 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7393 buf = strstrip(buf);
7394 err = page_counter_memparse(buf, "max", &max);
7398 xchg(&memcg->swap.max, max);
7403 static int swap_events_show(struct seq_file *m, void *v)
7405 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7407 seq_printf(m, "high %lu\n",
7408 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7409 seq_printf(m, "max %lu\n",
7410 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7411 seq_printf(m, "fail %lu\n",
7412 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7417 static struct cftype swap_files[] = {
7419 .name = "swap.current",
7420 .flags = CFTYPE_NOT_ON_ROOT,
7421 .read_u64 = swap_current_read,
7424 .name = "swap.high",
7425 .flags = CFTYPE_NOT_ON_ROOT,
7426 .seq_show = swap_high_show,
7427 .write = swap_high_write,
7431 .flags = CFTYPE_NOT_ON_ROOT,
7432 .seq_show = swap_max_show,
7433 .write = swap_max_write,
7436 .name = "swap.events",
7437 .flags = CFTYPE_NOT_ON_ROOT,
7438 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7439 .seq_show = swap_events_show,
7444 static struct cftype memsw_files[] = {
7446 .name = "memsw.usage_in_bytes",
7447 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7448 .read_u64 = mem_cgroup_read_u64,
7451 .name = "memsw.max_usage_in_bytes",
7452 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7453 .write = mem_cgroup_reset,
7454 .read_u64 = mem_cgroup_read_u64,
7457 .name = "memsw.limit_in_bytes",
7458 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7459 .write = mem_cgroup_write,
7460 .read_u64 = mem_cgroup_read_u64,
7463 .name = "memsw.failcnt",
7464 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7465 .write = mem_cgroup_reset,
7466 .read_u64 = mem_cgroup_read_u64,
7468 { }, /* terminate */
7472 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7473 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7474 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7475 * boot parameter. This may result in premature OOPS inside
7476 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7478 static int __init mem_cgroup_swap_init(void)
7480 /* No memory control -> no swap control */
7481 if (mem_cgroup_disabled())
7482 cgroup_memory_noswap = true;
7484 if (cgroup_memory_noswap)
7487 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7488 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7492 core_initcall(mem_cgroup_swap_init);
7494 #endif /* CONFIG_MEMCG_SWAP */