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 #define THRESHOLDS_EVENTS_TARGET 128
107 #define SOFTLIMIT_EVENTS_TARGET 1024
110 * Cgroups above their limits are maintained in a RB-Tree, independent of
111 * their hierarchy representation
114 struct mem_cgroup_tree_per_node {
115 struct rb_root rb_root;
116 struct rb_node *rb_rightmost;
120 struct mem_cgroup_tree {
121 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
124 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
127 struct mem_cgroup_eventfd_list {
128 struct list_head list;
129 struct eventfd_ctx *eventfd;
133 * cgroup_event represents events which userspace want to receive.
135 struct mem_cgroup_event {
137 * memcg which the event belongs to.
139 struct mem_cgroup *memcg;
141 * eventfd to signal userspace about the event.
143 struct eventfd_ctx *eventfd;
145 * Each of these stored in a list by the cgroup.
147 struct list_head list;
149 * register_event() callback will be used to add new userspace
150 * waiter for changes related to this event. Use eventfd_signal()
151 * on eventfd to send notification to userspace.
153 int (*register_event)(struct mem_cgroup *memcg,
154 struct eventfd_ctx *eventfd, const char *args);
156 * unregister_event() callback will be called when userspace closes
157 * the eventfd or on cgroup removing. This callback must be set,
158 * if you want provide notification functionality.
160 void (*unregister_event)(struct mem_cgroup *memcg,
161 struct eventfd_ctx *eventfd);
163 * All fields below needed to unregister event when
164 * userspace closes eventfd.
167 wait_queue_head_t *wqh;
168 wait_queue_entry_t wait;
169 struct work_struct remove;
172 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
173 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
175 /* Stuffs for move charges at task migration. */
177 * Types of charges to be moved.
179 #define MOVE_ANON 0x1U
180 #define MOVE_FILE 0x2U
181 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
183 /* "mc" and its members are protected by cgroup_mutex */
184 static struct move_charge_struct {
185 spinlock_t lock; /* for from, to */
186 struct mm_struct *mm;
187 struct mem_cgroup *from;
188 struct mem_cgroup *to;
190 unsigned long precharge;
191 unsigned long moved_charge;
192 unsigned long moved_swap;
193 struct task_struct *moving_task; /* a task moving charges */
194 wait_queue_head_t waitq; /* a waitq for other context */
196 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
197 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
201 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
202 * limit reclaim to prevent infinite loops, if they ever occur.
204 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
205 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
207 /* for encoding cft->private value on file */
216 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
217 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
218 #define MEMFILE_ATTR(val) ((val) & 0xffff)
219 /* Used for OOM notifier */
220 #define OOM_CONTROL (0)
223 * Iteration constructs for visiting all cgroups (under a tree). If
224 * loops are exited prematurely (break), mem_cgroup_iter_break() must
225 * be used for reference counting.
227 #define for_each_mem_cgroup_tree(iter, root) \
228 for (iter = mem_cgroup_iter(root, NULL, NULL); \
230 iter = mem_cgroup_iter(root, iter, NULL))
232 #define for_each_mem_cgroup(iter) \
233 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
235 iter = mem_cgroup_iter(NULL, iter, NULL))
237 static inline bool task_is_dying(void)
239 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
240 (current->flags & PF_EXITING);
243 /* Some nice accessors for the vmpressure. */
244 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
247 memcg = root_mem_cgroup;
248 return &memcg->vmpressure;
251 struct mem_cgroup *vmpressure_to_memcg(struct vmpressure *vmpr)
253 return container_of(vmpr, struct mem_cgroup, vmpressure);
256 #ifdef CONFIG_MEMCG_KMEM
257 extern spinlock_t css_set_lock;
259 bool mem_cgroup_kmem_disabled(void)
261 return cgroup_memory_nokmem;
264 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
265 unsigned int nr_pages);
267 static void obj_cgroup_release(struct percpu_ref *ref)
269 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
270 unsigned int nr_bytes;
271 unsigned int nr_pages;
275 * At this point all allocated objects are freed, and
276 * objcg->nr_charged_bytes can't have an arbitrary byte value.
277 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
279 * The following sequence can lead to it:
280 * 1) CPU0: objcg == stock->cached_objcg
281 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
282 * PAGE_SIZE bytes are charged
283 * 3) CPU1: a process from another memcg is allocating something,
284 * the stock if flushed,
285 * objcg->nr_charged_bytes = PAGE_SIZE - 92
286 * 5) CPU0: we do release this object,
287 * 92 bytes are added to stock->nr_bytes
288 * 6) CPU0: stock is flushed,
289 * 92 bytes are added to objcg->nr_charged_bytes
291 * In the result, nr_charged_bytes == PAGE_SIZE.
292 * This page will be uncharged in obj_cgroup_release().
294 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
295 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
296 nr_pages = nr_bytes >> PAGE_SHIFT;
299 obj_cgroup_uncharge_pages(objcg, nr_pages);
301 spin_lock_irqsave(&css_set_lock, flags);
302 list_del(&objcg->list);
303 spin_unlock_irqrestore(&css_set_lock, flags);
305 percpu_ref_exit(ref);
306 kfree_rcu(objcg, rcu);
309 static struct obj_cgroup *obj_cgroup_alloc(void)
311 struct obj_cgroup *objcg;
314 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
318 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
324 INIT_LIST_HEAD(&objcg->list);
328 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
329 struct mem_cgroup *parent)
331 struct obj_cgroup *objcg, *iter;
333 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
335 spin_lock_irq(&css_set_lock);
337 /* 1) Ready to reparent active objcg. */
338 list_add(&objcg->list, &memcg->objcg_list);
339 /* 2) Reparent active objcg and already reparented objcgs to parent. */
340 list_for_each_entry(iter, &memcg->objcg_list, list)
341 WRITE_ONCE(iter->memcg, parent);
342 /* 3) Move already reparented objcgs to the parent's list */
343 list_splice(&memcg->objcg_list, &parent->objcg_list);
345 spin_unlock_irq(&css_set_lock);
347 percpu_ref_kill(&objcg->refcnt);
351 * This will be used as a shrinker list's index.
352 * The main reason for not using cgroup id for this:
353 * this works better in sparse environments, where we have a lot of memcgs,
354 * but only a few kmem-limited. Or also, if we have, for instance, 200
355 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
356 * 200 entry array for that.
358 * The current size of the caches array is stored in memcg_nr_cache_ids. It
359 * will double each time we have to increase it.
361 static DEFINE_IDA(memcg_cache_ida);
362 int memcg_nr_cache_ids;
364 /* Protects memcg_nr_cache_ids */
365 static DECLARE_RWSEM(memcg_cache_ids_sem);
367 void memcg_get_cache_ids(void)
369 down_read(&memcg_cache_ids_sem);
372 void memcg_put_cache_ids(void)
374 up_read(&memcg_cache_ids_sem);
378 * MIN_SIZE is different than 1, because we would like to avoid going through
379 * the alloc/free process all the time. In a small machine, 4 kmem-limited
380 * cgroups is a reasonable guess. In the future, it could be a parameter or
381 * tunable, but that is strictly not necessary.
383 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
384 * this constant directly from cgroup, but it is understandable that this is
385 * better kept as an internal representation in cgroup.c. In any case, the
386 * cgrp_id space is not getting any smaller, and we don't have to necessarily
387 * increase ours as well if it increases.
389 #define MEMCG_CACHES_MIN_SIZE 4
390 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
393 * A lot of the calls to the cache allocation functions are expected to be
394 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
395 * conditional to this static branch, we'll have to allow modules that does
396 * kmem_cache_alloc and the such to see this symbol as well
398 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
399 EXPORT_SYMBOL(memcg_kmem_enabled_key);
403 * mem_cgroup_css_from_page - css of the memcg associated with a page
404 * @page: page of interest
406 * If memcg is bound to the default hierarchy, css of the memcg associated
407 * with @page is returned. The returned css remains associated with @page
408 * until it is released.
410 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
413 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
415 struct mem_cgroup *memcg;
417 memcg = page_memcg(page);
419 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
420 memcg = root_mem_cgroup;
426 * page_cgroup_ino - return inode number of the memcg a page is charged to
429 * Look up the closest online ancestor of the memory cgroup @page is charged to
430 * and return its inode number or 0 if @page is not charged to any cgroup. It
431 * is safe to call this function without holding a reference to @page.
433 * Note, this function is inherently racy, because there is nothing to prevent
434 * the cgroup inode from getting torn down and potentially reallocated a moment
435 * after page_cgroup_ino() returns, so it only should be used by callers that
436 * do not care (such as procfs interfaces).
438 ino_t page_cgroup_ino(struct page *page)
440 struct mem_cgroup *memcg;
441 unsigned long ino = 0;
444 memcg = page_memcg_check(page);
446 while (memcg && !(memcg->css.flags & CSS_ONLINE))
447 memcg = parent_mem_cgroup(memcg);
449 ino = cgroup_ino(memcg->css.cgroup);
454 static struct mem_cgroup_per_node *
455 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
457 int nid = page_to_nid(page);
459 return memcg->nodeinfo[nid];
462 static struct mem_cgroup_tree_per_node *
463 soft_limit_tree_node(int nid)
465 return soft_limit_tree.rb_tree_per_node[nid];
468 static struct mem_cgroup_tree_per_node *
469 soft_limit_tree_from_page(struct page *page)
471 int nid = page_to_nid(page);
473 return soft_limit_tree.rb_tree_per_node[nid];
476 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
477 struct mem_cgroup_tree_per_node *mctz,
478 unsigned long new_usage_in_excess)
480 struct rb_node **p = &mctz->rb_root.rb_node;
481 struct rb_node *parent = NULL;
482 struct mem_cgroup_per_node *mz_node;
483 bool rightmost = true;
488 mz->usage_in_excess = new_usage_in_excess;
489 if (!mz->usage_in_excess)
493 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
495 if (mz->usage_in_excess < mz_node->usage_in_excess) {
504 mctz->rb_rightmost = &mz->tree_node;
506 rb_link_node(&mz->tree_node, parent, p);
507 rb_insert_color(&mz->tree_node, &mctz->rb_root);
511 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
512 struct mem_cgroup_tree_per_node *mctz)
517 if (&mz->tree_node == mctz->rb_rightmost)
518 mctz->rb_rightmost = rb_prev(&mz->tree_node);
520 rb_erase(&mz->tree_node, &mctz->rb_root);
524 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
525 struct mem_cgroup_tree_per_node *mctz)
529 spin_lock_irqsave(&mctz->lock, flags);
530 __mem_cgroup_remove_exceeded(mz, mctz);
531 spin_unlock_irqrestore(&mctz->lock, flags);
534 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
536 unsigned long nr_pages = page_counter_read(&memcg->memory);
537 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
538 unsigned long excess = 0;
540 if (nr_pages > soft_limit)
541 excess = nr_pages - soft_limit;
546 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
548 unsigned long excess;
549 struct mem_cgroup_per_node *mz;
550 struct mem_cgroup_tree_per_node *mctz;
552 mctz = soft_limit_tree_from_page(page);
556 * Necessary to update all ancestors when hierarchy is used.
557 * because their event counter is not touched.
559 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
560 mz = mem_cgroup_page_nodeinfo(memcg, page);
561 excess = soft_limit_excess(memcg);
563 * We have to update the tree if mz is on RB-tree or
564 * mem is over its softlimit.
566 if (excess || mz->on_tree) {
569 spin_lock_irqsave(&mctz->lock, flags);
570 /* if on-tree, remove it */
572 __mem_cgroup_remove_exceeded(mz, mctz);
574 * Insert again. mz->usage_in_excess will be updated.
575 * If excess is 0, no tree ops.
577 __mem_cgroup_insert_exceeded(mz, mctz, excess);
578 spin_unlock_irqrestore(&mctz->lock, flags);
583 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
585 struct mem_cgroup_tree_per_node *mctz;
586 struct mem_cgroup_per_node *mz;
590 mz = memcg->nodeinfo[nid];
591 mctz = soft_limit_tree_node(nid);
593 mem_cgroup_remove_exceeded(mz, mctz);
597 static struct mem_cgroup_per_node *
598 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
600 struct mem_cgroup_per_node *mz;
604 if (!mctz->rb_rightmost)
605 goto done; /* Nothing to reclaim from */
607 mz = rb_entry(mctz->rb_rightmost,
608 struct mem_cgroup_per_node, tree_node);
610 * Remove the node now but someone else can add it back,
611 * we will to add it back at the end of reclaim to its correct
612 * position in the tree.
614 __mem_cgroup_remove_exceeded(mz, mctz);
615 if (!soft_limit_excess(mz->memcg) ||
616 !css_tryget(&mz->memcg->css))
622 static struct mem_cgroup_per_node *
623 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
625 struct mem_cgroup_per_node *mz;
627 spin_lock_irq(&mctz->lock);
628 mz = __mem_cgroup_largest_soft_limit_node(mctz);
629 spin_unlock_irq(&mctz->lock);
634 * memcg and lruvec stats flushing
636 * Many codepaths leading to stats update or read are performance sensitive and
637 * adding stats flushing in such codepaths is not desirable. So, to optimize the
638 * flushing the kernel does:
640 * 1) Periodically and asynchronously flush the stats every 2 seconds to not let
641 * rstat update tree grow unbounded.
643 * 2) Flush the stats synchronously on reader side only when there are more than
644 * (MEMCG_CHARGE_BATCH * nr_cpus) update events. Though this optimization
645 * will let stats be out of sync by atmost (MEMCG_CHARGE_BATCH * nr_cpus) but
646 * only for 2 seconds due to (1).
648 static void flush_memcg_stats_dwork(struct work_struct *w);
649 static DECLARE_DEFERRABLE_WORK(stats_flush_dwork, flush_memcg_stats_dwork);
650 static DEFINE_SPINLOCK(stats_flush_lock);
651 static DEFINE_PER_CPU(unsigned int, stats_updates);
652 static atomic_t stats_flush_threshold = ATOMIC_INIT(0);
654 static inline void memcg_rstat_updated(struct mem_cgroup *memcg)
656 cgroup_rstat_updated(memcg->css.cgroup, smp_processor_id());
657 if (!(__this_cpu_inc_return(stats_updates) % MEMCG_CHARGE_BATCH))
658 atomic_inc(&stats_flush_threshold);
661 static void __mem_cgroup_flush_stats(void)
665 if (!spin_trylock_irqsave(&stats_flush_lock, flag))
668 cgroup_rstat_flush_irqsafe(root_mem_cgroup->css.cgroup);
669 atomic_set(&stats_flush_threshold, 0);
670 spin_unlock_irqrestore(&stats_flush_lock, flag);
673 void mem_cgroup_flush_stats(void)
675 if (atomic_read(&stats_flush_threshold) > num_online_cpus())
676 __mem_cgroup_flush_stats();
679 static void flush_memcg_stats_dwork(struct work_struct *w)
681 mem_cgroup_flush_stats();
682 queue_delayed_work(system_unbound_wq, &stats_flush_dwork, 2UL*HZ);
686 * __mod_memcg_state - update cgroup memory statistics
687 * @memcg: the memory cgroup
688 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
689 * @val: delta to add to the counter, can be negative
691 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
693 if (mem_cgroup_disabled())
696 __this_cpu_add(memcg->vmstats_percpu->state[idx], val);
697 memcg_rstat_updated(memcg);
700 /* idx can be of type enum memcg_stat_item or node_stat_item. */
701 static unsigned long memcg_page_state_local(struct mem_cgroup *memcg, int idx)
706 for_each_possible_cpu(cpu)
707 x += per_cpu(memcg->vmstats_percpu->state[idx], cpu);
715 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
718 struct mem_cgroup_per_node *pn;
719 struct mem_cgroup *memcg;
721 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
725 __this_cpu_add(memcg->vmstats_percpu->state[idx], val);
728 __this_cpu_add(pn->lruvec_stats_percpu->state[idx], val);
730 memcg_rstat_updated(memcg);
734 * __mod_lruvec_state - update lruvec memory statistics
735 * @lruvec: the lruvec
736 * @idx: the stat item
737 * @val: delta to add to the counter, can be negative
739 * The lruvec is the intersection of the NUMA node and a cgroup. This
740 * function updates the all three counters that are affected by a
741 * change of state at this level: per-node, per-cgroup, per-lruvec.
743 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
747 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
749 /* Update memcg and lruvec */
750 if (!mem_cgroup_disabled())
751 __mod_memcg_lruvec_state(lruvec, idx, val);
754 void __mod_lruvec_page_state(struct page *page, enum node_stat_item idx,
757 struct page *head = compound_head(page); /* rmap on tail pages */
758 struct mem_cgroup *memcg;
759 pg_data_t *pgdat = page_pgdat(page);
760 struct lruvec *lruvec;
763 memcg = page_memcg(head);
764 /* Untracked pages have no memcg, no lruvec. Update only the node */
767 __mod_node_page_state(pgdat, idx, val);
771 lruvec = mem_cgroup_lruvec(memcg, pgdat);
772 __mod_lruvec_state(lruvec, idx, val);
775 EXPORT_SYMBOL(__mod_lruvec_page_state);
777 void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val)
779 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
780 struct mem_cgroup *memcg;
781 struct lruvec *lruvec;
784 memcg = mem_cgroup_from_obj(p);
787 * Untracked pages have no memcg, no lruvec. Update only the
788 * node. If we reparent the slab objects to the root memcg,
789 * when we free the slab object, we need to update the per-memcg
790 * vmstats to keep it correct for the root memcg.
793 __mod_node_page_state(pgdat, idx, val);
795 lruvec = mem_cgroup_lruvec(memcg, pgdat);
796 __mod_lruvec_state(lruvec, idx, val);
802 * mod_objcg_mlstate() may be called with irq enabled, so
803 * mod_memcg_lruvec_state() should be used.
805 static inline void mod_objcg_mlstate(struct obj_cgroup *objcg,
806 struct pglist_data *pgdat,
807 enum node_stat_item idx, int nr)
809 struct mem_cgroup *memcg;
810 struct lruvec *lruvec;
813 memcg = obj_cgroup_memcg(objcg);
814 lruvec = mem_cgroup_lruvec(memcg, pgdat);
815 mod_memcg_lruvec_state(lruvec, idx, nr);
820 * __count_memcg_events - account VM events in a cgroup
821 * @memcg: the memory cgroup
822 * @idx: the event item
823 * @count: the number of events that occurred
825 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
828 if (mem_cgroup_disabled())
831 __this_cpu_add(memcg->vmstats_percpu->events[idx], count);
832 memcg_rstat_updated(memcg);
835 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
837 return READ_ONCE(memcg->vmstats.events[event]);
840 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
845 for_each_possible_cpu(cpu)
846 x += per_cpu(memcg->vmstats_percpu->events[event], cpu);
850 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
854 /* pagein of a big page is an event. So, ignore page size */
856 __count_memcg_events(memcg, PGPGIN, 1);
858 __count_memcg_events(memcg, PGPGOUT, 1);
859 nr_pages = -nr_pages; /* for event */
862 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
865 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
866 enum mem_cgroup_events_target target)
868 unsigned long val, next;
870 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
871 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
872 /* from time_after() in jiffies.h */
873 if ((long)(next - val) < 0) {
875 case MEM_CGROUP_TARGET_THRESH:
876 next = val + THRESHOLDS_EVENTS_TARGET;
878 case MEM_CGROUP_TARGET_SOFTLIMIT:
879 next = val + SOFTLIMIT_EVENTS_TARGET;
884 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
891 * Check events in order.
894 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
896 /* threshold event is triggered in finer grain than soft limit */
897 if (unlikely(mem_cgroup_event_ratelimit(memcg,
898 MEM_CGROUP_TARGET_THRESH))) {
901 do_softlimit = mem_cgroup_event_ratelimit(memcg,
902 MEM_CGROUP_TARGET_SOFTLIMIT);
903 mem_cgroup_threshold(memcg);
904 if (unlikely(do_softlimit))
905 mem_cgroup_update_tree(memcg, page);
909 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
912 * mm_update_next_owner() may clear mm->owner to NULL
913 * if it races with swapoff, page migration, etc.
914 * So this can be called with p == NULL.
919 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
921 EXPORT_SYMBOL(mem_cgroup_from_task);
923 static __always_inline struct mem_cgroup *active_memcg(void)
926 return this_cpu_read(int_active_memcg);
928 return current->active_memcg;
932 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
933 * @mm: mm from which memcg should be extracted. It can be NULL.
935 * Obtain a reference on mm->memcg and returns it if successful. If mm
936 * is NULL, then the memcg is chosen as follows:
937 * 1) The active memcg, if set.
938 * 2) current->mm->memcg, if available
940 * If mem_cgroup is disabled, NULL is returned.
942 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
944 struct mem_cgroup *memcg;
946 if (mem_cgroup_disabled())
950 * Page cache insertions can happen without an
951 * actual mm context, e.g. during disk probing
952 * on boot, loopback IO, acct() writes etc.
954 * No need to css_get on root memcg as the reference
955 * counting is disabled on the root level in the
956 * cgroup core. See CSS_NO_REF.
959 memcg = active_memcg();
960 if (unlikely(memcg)) {
961 /* remote memcg must hold a ref */
962 css_get(&memcg->css);
967 return root_mem_cgroup;
972 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
973 if (unlikely(!memcg))
974 memcg = root_mem_cgroup;
975 } while (!css_tryget(&memcg->css));
979 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
981 static __always_inline bool memcg_kmem_bypass(void)
983 /* Allow remote memcg charging from any context. */
984 if (unlikely(active_memcg()))
987 /* Memcg to charge can't be determined. */
988 if (!in_task() || !current->mm || (current->flags & PF_KTHREAD))
995 * mem_cgroup_iter - iterate over memory cgroup hierarchy
996 * @root: hierarchy root
997 * @prev: previously returned memcg, NULL on first invocation
998 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1000 * Returns references to children of the hierarchy below @root, or
1001 * @root itself, or %NULL after a full round-trip.
1003 * Caller must pass the return value in @prev on subsequent
1004 * invocations for reference counting, or use mem_cgroup_iter_break()
1005 * to cancel a hierarchy walk before the round-trip is complete.
1007 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1008 * in the hierarchy among all concurrent reclaimers operating on the
1011 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1012 struct mem_cgroup *prev,
1013 struct mem_cgroup_reclaim_cookie *reclaim)
1015 struct mem_cgroup_reclaim_iter *iter;
1016 struct cgroup_subsys_state *css = NULL;
1017 struct mem_cgroup *memcg = NULL;
1018 struct mem_cgroup *pos = NULL;
1020 if (mem_cgroup_disabled())
1024 root = root_mem_cgroup;
1026 if (prev && !reclaim)
1032 struct mem_cgroup_per_node *mz;
1034 mz = root->nodeinfo[reclaim->pgdat->node_id];
1037 if (prev && reclaim->generation != iter->generation)
1041 pos = READ_ONCE(iter->position);
1042 if (!pos || css_tryget(&pos->css))
1045 * css reference reached zero, so iter->position will
1046 * be cleared by ->css_released. However, we should not
1047 * rely on this happening soon, because ->css_released
1048 * is called from a work queue, and by busy-waiting we
1049 * might block it. So we clear iter->position right
1052 (void)cmpxchg(&iter->position, pos, NULL);
1060 css = css_next_descendant_pre(css, &root->css);
1063 * Reclaimers share the hierarchy walk, and a
1064 * new one might jump in right at the end of
1065 * the hierarchy - make sure they see at least
1066 * one group and restart from the beginning.
1074 * Verify the css and acquire a reference. The root
1075 * is provided by the caller, so we know it's alive
1076 * and kicking, and don't take an extra reference.
1078 memcg = mem_cgroup_from_css(css);
1080 if (css == &root->css)
1083 if (css_tryget(css))
1091 * The position could have already been updated by a competing
1092 * thread, so check that the value hasn't changed since we read
1093 * it to avoid reclaiming from the same cgroup twice.
1095 (void)cmpxchg(&iter->position, pos, memcg);
1103 reclaim->generation = iter->generation;
1108 if (prev && prev != root)
1109 css_put(&prev->css);
1115 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1116 * @root: hierarchy root
1117 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1119 void mem_cgroup_iter_break(struct mem_cgroup *root,
1120 struct mem_cgroup *prev)
1123 root = root_mem_cgroup;
1124 if (prev && prev != root)
1125 css_put(&prev->css);
1128 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1129 struct mem_cgroup *dead_memcg)
1131 struct mem_cgroup_reclaim_iter *iter;
1132 struct mem_cgroup_per_node *mz;
1135 for_each_node(nid) {
1136 mz = from->nodeinfo[nid];
1138 cmpxchg(&iter->position, dead_memcg, NULL);
1142 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1144 struct mem_cgroup *memcg = dead_memcg;
1145 struct mem_cgroup *last;
1148 __invalidate_reclaim_iterators(memcg, dead_memcg);
1150 } while ((memcg = parent_mem_cgroup(memcg)));
1153 * When cgruop1 non-hierarchy mode is used,
1154 * parent_mem_cgroup() does not walk all the way up to the
1155 * cgroup root (root_mem_cgroup). So we have to handle
1156 * dead_memcg from cgroup root separately.
1158 if (last != root_mem_cgroup)
1159 __invalidate_reclaim_iterators(root_mem_cgroup,
1164 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1165 * @memcg: hierarchy root
1166 * @fn: function to call for each task
1167 * @arg: argument passed to @fn
1169 * This function iterates over tasks attached to @memcg or to any of its
1170 * descendants and calls @fn for each task. If @fn returns a non-zero
1171 * value, the function breaks the iteration loop and returns the value.
1172 * Otherwise, it will iterate over all tasks and return 0.
1174 * This function must not be called for the root memory cgroup.
1176 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1177 int (*fn)(struct task_struct *, void *), void *arg)
1179 struct mem_cgroup *iter;
1182 BUG_ON(memcg == root_mem_cgroup);
1184 for_each_mem_cgroup_tree(iter, memcg) {
1185 struct css_task_iter it;
1186 struct task_struct *task;
1188 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1189 while (!ret && (task = css_task_iter_next(&it)))
1190 ret = fn(task, arg);
1191 css_task_iter_end(&it);
1193 mem_cgroup_iter_break(memcg, iter);
1200 #ifdef CONFIG_DEBUG_VM
1201 void lruvec_memcg_debug(struct lruvec *lruvec, struct page *page)
1203 struct mem_cgroup *memcg;
1205 if (mem_cgroup_disabled())
1208 memcg = page_memcg(page);
1211 VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != root_mem_cgroup, page);
1213 VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != memcg, page);
1218 * lock_page_lruvec - lock and return lruvec for a given page.
1221 * These functions are safe to use under any of the following conditions:
1224 * - lock_page_memcg()
1225 * - page->_refcount is zero
1227 struct lruvec *lock_page_lruvec(struct page *page)
1229 struct lruvec *lruvec;
1231 lruvec = mem_cgroup_page_lruvec(page);
1232 spin_lock(&lruvec->lru_lock);
1234 lruvec_memcg_debug(lruvec, page);
1239 struct lruvec *lock_page_lruvec_irq(struct page *page)
1241 struct lruvec *lruvec;
1243 lruvec = mem_cgroup_page_lruvec(page);
1244 spin_lock_irq(&lruvec->lru_lock);
1246 lruvec_memcg_debug(lruvec, page);
1251 struct lruvec *lock_page_lruvec_irqsave(struct page *page, unsigned long *flags)
1253 struct lruvec *lruvec;
1255 lruvec = mem_cgroup_page_lruvec(page);
1256 spin_lock_irqsave(&lruvec->lru_lock, *flags);
1258 lruvec_memcg_debug(lruvec, page);
1264 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1265 * @lruvec: mem_cgroup per zone lru vector
1266 * @lru: index of lru list the page is sitting on
1267 * @zid: zone id of the accounted pages
1268 * @nr_pages: positive when adding or negative when removing
1270 * This function must be called under lru_lock, just before a page is added
1271 * to or just after a page is removed from an lru list (that ordering being
1272 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1274 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1275 int zid, int nr_pages)
1277 struct mem_cgroup_per_node *mz;
1278 unsigned long *lru_size;
1281 if (mem_cgroup_disabled())
1284 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1285 lru_size = &mz->lru_zone_size[zid][lru];
1288 *lru_size += nr_pages;
1291 if (WARN_ONCE(size < 0,
1292 "%s(%p, %d, %d): lru_size %ld\n",
1293 __func__, lruvec, lru, nr_pages, size)) {
1299 *lru_size += nr_pages;
1303 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1304 * @memcg: the memory cgroup
1306 * Returns the maximum amount of memory @mem can be charged with, in
1309 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1311 unsigned long margin = 0;
1312 unsigned long count;
1313 unsigned long limit;
1315 count = page_counter_read(&memcg->memory);
1316 limit = READ_ONCE(memcg->memory.max);
1318 margin = limit - count;
1320 if (do_memsw_account()) {
1321 count = page_counter_read(&memcg->memsw);
1322 limit = READ_ONCE(memcg->memsw.max);
1324 margin = min(margin, limit - count);
1333 * A routine for checking "mem" is under move_account() or not.
1335 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1336 * moving cgroups. This is for waiting at high-memory pressure
1339 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1341 struct mem_cgroup *from;
1342 struct mem_cgroup *to;
1345 * Unlike task_move routines, we access mc.to, mc.from not under
1346 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1348 spin_lock(&mc.lock);
1354 ret = mem_cgroup_is_descendant(from, memcg) ||
1355 mem_cgroup_is_descendant(to, memcg);
1357 spin_unlock(&mc.lock);
1361 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1363 if (mc.moving_task && current != mc.moving_task) {
1364 if (mem_cgroup_under_move(memcg)) {
1366 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1367 /* moving charge context might have finished. */
1370 finish_wait(&mc.waitq, &wait);
1377 struct memory_stat {
1382 static const struct memory_stat memory_stats[] = {
1383 { "anon", NR_ANON_MAPPED },
1384 { "file", NR_FILE_PAGES },
1385 { "kernel_stack", NR_KERNEL_STACK_KB },
1386 { "pagetables", NR_PAGETABLE },
1387 { "percpu", MEMCG_PERCPU_B },
1388 { "sock", MEMCG_SOCK },
1389 { "shmem", NR_SHMEM },
1390 { "file_mapped", NR_FILE_MAPPED },
1391 { "file_dirty", NR_FILE_DIRTY },
1392 { "file_writeback", NR_WRITEBACK },
1394 { "swapcached", NR_SWAPCACHE },
1396 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1397 { "anon_thp", NR_ANON_THPS },
1398 { "file_thp", NR_FILE_THPS },
1399 { "shmem_thp", NR_SHMEM_THPS },
1401 { "inactive_anon", NR_INACTIVE_ANON },
1402 { "active_anon", NR_ACTIVE_ANON },
1403 { "inactive_file", NR_INACTIVE_FILE },
1404 { "active_file", NR_ACTIVE_FILE },
1405 { "unevictable", NR_UNEVICTABLE },
1406 { "slab_reclaimable", NR_SLAB_RECLAIMABLE_B },
1407 { "slab_unreclaimable", NR_SLAB_UNRECLAIMABLE_B },
1409 /* The memory events */
1410 { "workingset_refault_anon", WORKINGSET_REFAULT_ANON },
1411 { "workingset_refault_file", WORKINGSET_REFAULT_FILE },
1412 { "workingset_activate_anon", WORKINGSET_ACTIVATE_ANON },
1413 { "workingset_activate_file", WORKINGSET_ACTIVATE_FILE },
1414 { "workingset_restore_anon", WORKINGSET_RESTORE_ANON },
1415 { "workingset_restore_file", WORKINGSET_RESTORE_FILE },
1416 { "workingset_nodereclaim", WORKINGSET_NODERECLAIM },
1419 /* Translate stat items to the correct unit for memory.stat output */
1420 static int memcg_page_state_unit(int item)
1423 case MEMCG_PERCPU_B:
1424 case NR_SLAB_RECLAIMABLE_B:
1425 case NR_SLAB_UNRECLAIMABLE_B:
1426 case WORKINGSET_REFAULT_ANON:
1427 case WORKINGSET_REFAULT_FILE:
1428 case WORKINGSET_ACTIVATE_ANON:
1429 case WORKINGSET_ACTIVATE_FILE:
1430 case WORKINGSET_RESTORE_ANON:
1431 case WORKINGSET_RESTORE_FILE:
1432 case WORKINGSET_NODERECLAIM:
1434 case NR_KERNEL_STACK_KB:
1441 static inline unsigned long memcg_page_state_output(struct mem_cgroup *memcg,
1444 return memcg_page_state(memcg, item) * memcg_page_state_unit(item);
1447 static char *memory_stat_format(struct mem_cgroup *memcg)
1452 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1457 * Provide statistics on the state of the memory subsystem as
1458 * well as cumulative event counters that show past behavior.
1460 * This list is ordered following a combination of these gradients:
1461 * 1) generic big picture -> specifics and details
1462 * 2) reflecting userspace activity -> reflecting kernel heuristics
1464 * Current memory state:
1466 mem_cgroup_flush_stats();
1468 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1471 size = memcg_page_state_output(memcg, memory_stats[i].idx);
1472 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1474 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1475 size += memcg_page_state_output(memcg,
1476 NR_SLAB_RECLAIMABLE_B);
1477 seq_buf_printf(&s, "slab %llu\n", size);
1481 /* Accumulated memory events */
1483 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1484 memcg_events(memcg, PGFAULT));
1485 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1486 memcg_events(memcg, PGMAJFAULT));
1487 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1488 memcg_events(memcg, PGREFILL));
1489 seq_buf_printf(&s, "pgscan %lu\n",
1490 memcg_events(memcg, PGSCAN_KSWAPD) +
1491 memcg_events(memcg, PGSCAN_DIRECT));
1492 seq_buf_printf(&s, "pgsteal %lu\n",
1493 memcg_events(memcg, PGSTEAL_KSWAPD) +
1494 memcg_events(memcg, PGSTEAL_DIRECT));
1495 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1496 memcg_events(memcg, PGACTIVATE));
1497 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1498 memcg_events(memcg, PGDEACTIVATE));
1499 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1500 memcg_events(memcg, PGLAZYFREE));
1501 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1502 memcg_events(memcg, PGLAZYFREED));
1504 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1505 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1506 memcg_events(memcg, THP_FAULT_ALLOC));
1507 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1508 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1509 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1511 /* The above should easily fit into one page */
1512 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1517 #define K(x) ((x) << (PAGE_SHIFT-10))
1519 * mem_cgroup_print_oom_context: Print OOM information relevant to
1520 * memory controller.
1521 * @memcg: The memory cgroup that went over limit
1522 * @p: Task that is going to be killed
1524 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1527 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1532 pr_cont(",oom_memcg=");
1533 pr_cont_cgroup_path(memcg->css.cgroup);
1535 pr_cont(",global_oom");
1537 pr_cont(",task_memcg=");
1538 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1544 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1545 * memory controller.
1546 * @memcg: The memory cgroup that went over limit
1548 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1552 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1553 K((u64)page_counter_read(&memcg->memory)),
1554 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1555 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1556 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1557 K((u64)page_counter_read(&memcg->swap)),
1558 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1560 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1561 K((u64)page_counter_read(&memcg->memsw)),
1562 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1563 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1564 K((u64)page_counter_read(&memcg->kmem)),
1565 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1568 pr_info("Memory cgroup stats for ");
1569 pr_cont_cgroup_path(memcg->css.cgroup);
1571 buf = memory_stat_format(memcg);
1579 * Return the memory (and swap, if configured) limit for a memcg.
1581 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1583 unsigned long max = READ_ONCE(memcg->memory.max);
1585 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1586 if (mem_cgroup_swappiness(memcg))
1587 max += min(READ_ONCE(memcg->swap.max),
1588 (unsigned long)total_swap_pages);
1590 if (mem_cgroup_swappiness(memcg)) {
1591 /* Calculate swap excess capacity from memsw limit */
1592 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1594 max += min(swap, (unsigned long)total_swap_pages);
1600 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1602 return page_counter_read(&memcg->memory);
1605 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1608 struct oom_control oc = {
1612 .gfp_mask = gfp_mask,
1617 if (mutex_lock_killable(&oom_lock))
1620 if (mem_cgroup_margin(memcg) >= (1 << order))
1624 * A few threads which were not waiting at mutex_lock_killable() can
1625 * fail to bail out. Therefore, check again after holding oom_lock.
1627 ret = task_is_dying() || out_of_memory(&oc);
1630 mutex_unlock(&oom_lock);
1634 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1637 unsigned long *total_scanned)
1639 struct mem_cgroup *victim = NULL;
1642 unsigned long excess;
1643 unsigned long nr_scanned;
1644 struct mem_cgroup_reclaim_cookie reclaim = {
1648 excess = soft_limit_excess(root_memcg);
1651 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1656 * If we have not been able to reclaim
1657 * anything, it might because there are
1658 * no reclaimable pages under this hierarchy
1663 * We want to do more targeted reclaim.
1664 * excess >> 2 is not to excessive so as to
1665 * reclaim too much, nor too less that we keep
1666 * coming back to reclaim from this cgroup
1668 if (total >= (excess >> 2) ||
1669 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1674 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1675 pgdat, &nr_scanned);
1676 *total_scanned += nr_scanned;
1677 if (!soft_limit_excess(root_memcg))
1680 mem_cgroup_iter_break(root_memcg, victim);
1684 #ifdef CONFIG_LOCKDEP
1685 static struct lockdep_map memcg_oom_lock_dep_map = {
1686 .name = "memcg_oom_lock",
1690 static DEFINE_SPINLOCK(memcg_oom_lock);
1693 * Check OOM-Killer is already running under our hierarchy.
1694 * If someone is running, return false.
1696 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1698 struct mem_cgroup *iter, *failed = NULL;
1700 spin_lock(&memcg_oom_lock);
1702 for_each_mem_cgroup_tree(iter, memcg) {
1703 if (iter->oom_lock) {
1705 * this subtree of our hierarchy is already locked
1706 * so we cannot give a lock.
1709 mem_cgroup_iter_break(memcg, iter);
1712 iter->oom_lock = true;
1717 * OK, we failed to lock the whole subtree so we have
1718 * to clean up what we set up to the failing subtree
1720 for_each_mem_cgroup_tree(iter, memcg) {
1721 if (iter == failed) {
1722 mem_cgroup_iter_break(memcg, iter);
1725 iter->oom_lock = false;
1728 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1730 spin_unlock(&memcg_oom_lock);
1735 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1737 struct mem_cgroup *iter;
1739 spin_lock(&memcg_oom_lock);
1740 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1741 for_each_mem_cgroup_tree(iter, memcg)
1742 iter->oom_lock = false;
1743 spin_unlock(&memcg_oom_lock);
1746 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1748 struct mem_cgroup *iter;
1750 spin_lock(&memcg_oom_lock);
1751 for_each_mem_cgroup_tree(iter, memcg)
1753 spin_unlock(&memcg_oom_lock);
1756 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1758 struct mem_cgroup *iter;
1761 * Be careful about under_oom underflows because a child memcg
1762 * could have been added after mem_cgroup_mark_under_oom.
1764 spin_lock(&memcg_oom_lock);
1765 for_each_mem_cgroup_tree(iter, memcg)
1766 if (iter->under_oom > 0)
1768 spin_unlock(&memcg_oom_lock);
1771 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1773 struct oom_wait_info {
1774 struct mem_cgroup *memcg;
1775 wait_queue_entry_t wait;
1778 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1779 unsigned mode, int sync, void *arg)
1781 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1782 struct mem_cgroup *oom_wait_memcg;
1783 struct oom_wait_info *oom_wait_info;
1785 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1786 oom_wait_memcg = oom_wait_info->memcg;
1788 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1789 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1791 return autoremove_wake_function(wait, mode, sync, arg);
1794 static void memcg_oom_recover(struct mem_cgroup *memcg)
1797 * For the following lockless ->under_oom test, the only required
1798 * guarantee is that it must see the state asserted by an OOM when
1799 * this function is called as a result of userland actions
1800 * triggered by the notification of the OOM. This is trivially
1801 * achieved by invoking mem_cgroup_mark_under_oom() before
1802 * triggering notification.
1804 if (memcg && memcg->under_oom)
1805 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1815 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1817 enum oom_status ret;
1820 if (order > PAGE_ALLOC_COSTLY_ORDER)
1823 memcg_memory_event(memcg, MEMCG_OOM);
1826 * We are in the middle of the charge context here, so we
1827 * don't want to block when potentially sitting on a callstack
1828 * that holds all kinds of filesystem and mm locks.
1830 * cgroup1 allows disabling the OOM killer and waiting for outside
1831 * handling until the charge can succeed; remember the context and put
1832 * the task to sleep at the end of the page fault when all locks are
1835 * On the other hand, in-kernel OOM killer allows for an async victim
1836 * memory reclaim (oom_reaper) and that means that we are not solely
1837 * relying on the oom victim to make a forward progress and we can
1838 * invoke the oom killer here.
1840 * Please note that mem_cgroup_out_of_memory might fail to find a
1841 * victim and then we have to bail out from the charge path.
1843 if (memcg->oom_kill_disable) {
1844 if (!current->in_user_fault)
1846 css_get(&memcg->css);
1847 current->memcg_in_oom = memcg;
1848 current->memcg_oom_gfp_mask = mask;
1849 current->memcg_oom_order = order;
1854 mem_cgroup_mark_under_oom(memcg);
1856 locked = mem_cgroup_oom_trylock(memcg);
1859 mem_cgroup_oom_notify(memcg);
1861 mem_cgroup_unmark_under_oom(memcg);
1862 if (mem_cgroup_out_of_memory(memcg, mask, order))
1868 mem_cgroup_oom_unlock(memcg);
1874 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1875 * @handle: actually kill/wait or just clean up the OOM state
1877 * This has to be called at the end of a page fault if the memcg OOM
1878 * handler was enabled.
1880 * Memcg supports userspace OOM handling where failed allocations must
1881 * sleep on a waitqueue until the userspace task resolves the
1882 * situation. Sleeping directly in the charge context with all kinds
1883 * of locks held is not a good idea, instead we remember an OOM state
1884 * in the task and mem_cgroup_oom_synchronize() has to be called at
1885 * the end of the page fault to complete the OOM handling.
1887 * Returns %true if an ongoing memcg OOM situation was detected and
1888 * completed, %false otherwise.
1890 bool mem_cgroup_oom_synchronize(bool handle)
1892 struct mem_cgroup *memcg = current->memcg_in_oom;
1893 struct oom_wait_info owait;
1896 /* OOM is global, do not handle */
1903 owait.memcg = memcg;
1904 owait.wait.flags = 0;
1905 owait.wait.func = memcg_oom_wake_function;
1906 owait.wait.private = current;
1907 INIT_LIST_HEAD(&owait.wait.entry);
1909 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1910 mem_cgroup_mark_under_oom(memcg);
1912 locked = mem_cgroup_oom_trylock(memcg);
1915 mem_cgroup_oom_notify(memcg);
1917 if (locked && !memcg->oom_kill_disable) {
1918 mem_cgroup_unmark_under_oom(memcg);
1919 finish_wait(&memcg_oom_waitq, &owait.wait);
1920 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1921 current->memcg_oom_order);
1924 mem_cgroup_unmark_under_oom(memcg);
1925 finish_wait(&memcg_oom_waitq, &owait.wait);
1929 mem_cgroup_oom_unlock(memcg);
1931 * There is no guarantee that an OOM-lock contender
1932 * sees the wakeups triggered by the OOM kill
1933 * uncharges. Wake any sleepers explicitly.
1935 memcg_oom_recover(memcg);
1938 current->memcg_in_oom = NULL;
1939 css_put(&memcg->css);
1944 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1945 * @victim: task to be killed by the OOM killer
1946 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1948 * Returns a pointer to a memory cgroup, which has to be cleaned up
1949 * by killing all belonging OOM-killable tasks.
1951 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1953 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1954 struct mem_cgroup *oom_domain)
1956 struct mem_cgroup *oom_group = NULL;
1957 struct mem_cgroup *memcg;
1959 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
1963 oom_domain = root_mem_cgroup;
1967 memcg = mem_cgroup_from_task(victim);
1968 if (memcg == root_mem_cgroup)
1972 * If the victim task has been asynchronously moved to a different
1973 * memory cgroup, we might end up killing tasks outside oom_domain.
1974 * In this case it's better to ignore memory.group.oom.
1976 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
1980 * Traverse the memory cgroup hierarchy from the victim task's
1981 * cgroup up to the OOMing cgroup (or root) to find the
1982 * highest-level memory cgroup with oom.group set.
1984 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1985 if (memcg->oom_group)
1988 if (memcg == oom_domain)
1993 css_get(&oom_group->css);
2000 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2002 pr_info("Tasks in ");
2003 pr_cont_cgroup_path(memcg->css.cgroup);
2004 pr_cont(" are going to be killed due to memory.oom.group set\n");
2008 * lock_page_memcg - lock a page and memcg binding
2011 * This function protects unlocked LRU pages from being moved to
2014 * It ensures lifetime of the locked memcg. Caller is responsible
2015 * for the lifetime of the page.
2017 void lock_page_memcg(struct page *page)
2019 struct page *head = compound_head(page); /* rmap on tail pages */
2020 struct mem_cgroup *memcg;
2021 unsigned long flags;
2024 * The RCU lock is held throughout the transaction. The fast
2025 * path can get away without acquiring the memcg->move_lock
2026 * because page moving starts with an RCU grace period.
2030 if (mem_cgroup_disabled())
2033 memcg = page_memcg(head);
2034 if (unlikely(!memcg))
2037 #ifdef CONFIG_PROVE_LOCKING
2038 local_irq_save(flags);
2039 might_lock(&memcg->move_lock);
2040 local_irq_restore(flags);
2043 if (atomic_read(&memcg->moving_account) <= 0)
2046 spin_lock_irqsave(&memcg->move_lock, flags);
2047 if (memcg != page_memcg(head)) {
2048 spin_unlock_irqrestore(&memcg->move_lock, flags);
2053 * When charge migration first begins, we can have multiple
2054 * critical sections holding the fast-path RCU lock and one
2055 * holding the slowpath move_lock. Track the task who has the
2056 * move_lock for unlock_page_memcg().
2058 memcg->move_lock_task = current;
2059 memcg->move_lock_flags = flags;
2061 EXPORT_SYMBOL(lock_page_memcg);
2063 static void __unlock_page_memcg(struct mem_cgroup *memcg)
2065 if (memcg && memcg->move_lock_task == current) {
2066 unsigned long flags = memcg->move_lock_flags;
2068 memcg->move_lock_task = NULL;
2069 memcg->move_lock_flags = 0;
2071 spin_unlock_irqrestore(&memcg->move_lock, flags);
2078 * unlock_page_memcg - unlock a page and memcg binding
2081 void unlock_page_memcg(struct page *page)
2083 struct page *head = compound_head(page);
2085 __unlock_page_memcg(page_memcg(head));
2087 EXPORT_SYMBOL(unlock_page_memcg);
2090 #ifdef CONFIG_MEMCG_KMEM
2091 struct obj_cgroup *cached_objcg;
2092 struct pglist_data *cached_pgdat;
2093 unsigned int nr_bytes;
2094 int nr_slab_reclaimable_b;
2095 int nr_slab_unreclaimable_b;
2101 struct memcg_stock_pcp {
2102 struct mem_cgroup *cached; /* this never be root cgroup */
2103 unsigned int nr_pages;
2104 struct obj_stock task_obj;
2105 struct obj_stock irq_obj;
2107 struct work_struct work;
2108 unsigned long flags;
2109 #define FLUSHING_CACHED_CHARGE 0
2111 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2112 static DEFINE_MUTEX(percpu_charge_mutex);
2114 #ifdef CONFIG_MEMCG_KMEM
2115 static void drain_obj_stock(struct obj_stock *stock);
2116 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2117 struct mem_cgroup *root_memcg);
2120 static inline void drain_obj_stock(struct obj_stock *stock)
2123 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2124 struct mem_cgroup *root_memcg)
2131 * Most kmem_cache_alloc() calls are from user context. The irq disable/enable
2132 * sequence used in this case to access content from object stock is slow.
2133 * To optimize for user context access, there are now two object stocks for
2134 * task context and interrupt context access respectively.
2136 * The task context object stock can be accessed by disabling preemption only
2137 * which is cheap in non-preempt kernel. The interrupt context object stock
2138 * can only be accessed after disabling interrupt. User context code can
2139 * access interrupt object stock, but not vice versa.
2141 static inline struct obj_stock *get_obj_stock(unsigned long *pflags)
2143 struct memcg_stock_pcp *stock;
2145 if (likely(in_task())) {
2148 stock = this_cpu_ptr(&memcg_stock);
2149 return &stock->task_obj;
2152 local_irq_save(*pflags);
2153 stock = this_cpu_ptr(&memcg_stock);
2154 return &stock->irq_obj;
2157 static inline void put_obj_stock(unsigned long flags)
2159 if (likely(in_task()))
2162 local_irq_restore(flags);
2166 * consume_stock: Try to consume stocked charge on this cpu.
2167 * @memcg: memcg to consume from.
2168 * @nr_pages: how many pages to charge.
2170 * The charges will only happen if @memcg matches the current cpu's memcg
2171 * stock, and at least @nr_pages are available in that stock. Failure to
2172 * service an allocation will refill the stock.
2174 * returns true if successful, false otherwise.
2176 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2178 struct memcg_stock_pcp *stock;
2179 unsigned long flags;
2182 if (nr_pages > MEMCG_CHARGE_BATCH)
2185 local_irq_save(flags);
2187 stock = this_cpu_ptr(&memcg_stock);
2188 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2189 stock->nr_pages -= nr_pages;
2193 local_irq_restore(flags);
2199 * Returns stocks cached in percpu and reset cached information.
2201 static void drain_stock(struct memcg_stock_pcp *stock)
2203 struct mem_cgroup *old = stock->cached;
2208 if (stock->nr_pages) {
2209 page_counter_uncharge(&old->memory, stock->nr_pages);
2210 if (do_memsw_account())
2211 page_counter_uncharge(&old->memsw, stock->nr_pages);
2212 stock->nr_pages = 0;
2216 stock->cached = NULL;
2219 static void drain_local_stock(struct work_struct *dummy)
2221 struct memcg_stock_pcp *stock;
2222 unsigned long flags;
2225 * The only protection from cpu hotplug (memcg_hotplug_cpu_dead) vs.
2226 * drain_stock races is that we always operate on local CPU stock
2227 * here with IRQ disabled
2229 local_irq_save(flags);
2231 stock = this_cpu_ptr(&memcg_stock);
2232 drain_obj_stock(&stock->irq_obj);
2234 drain_obj_stock(&stock->task_obj);
2236 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2238 local_irq_restore(flags);
2242 * Cache charges(val) to local per_cpu area.
2243 * This will be consumed by consume_stock() function, later.
2245 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2247 struct memcg_stock_pcp *stock;
2248 unsigned long flags;
2250 local_irq_save(flags);
2252 stock = this_cpu_ptr(&memcg_stock);
2253 if (stock->cached != memcg) { /* reset if necessary */
2255 css_get(&memcg->css);
2256 stock->cached = memcg;
2258 stock->nr_pages += nr_pages;
2260 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2263 local_irq_restore(flags);
2267 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2268 * of the hierarchy under it.
2270 static void drain_all_stock(struct mem_cgroup *root_memcg)
2274 /* If someone's already draining, avoid adding running more workers. */
2275 if (!mutex_trylock(&percpu_charge_mutex))
2278 * Notify other cpus that system-wide "drain" is running
2279 * We do not care about races with the cpu hotplug because cpu down
2280 * as well as workers from this path always operate on the local
2281 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2284 for_each_online_cpu(cpu) {
2285 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2286 struct mem_cgroup *memcg;
2290 memcg = stock->cached;
2291 if (memcg && stock->nr_pages &&
2292 mem_cgroup_is_descendant(memcg, root_memcg))
2294 else if (obj_stock_flush_required(stock, root_memcg))
2299 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2301 drain_local_stock(&stock->work);
2303 schedule_work_on(cpu, &stock->work);
2307 mutex_unlock(&percpu_charge_mutex);
2310 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2312 struct memcg_stock_pcp *stock;
2314 stock = &per_cpu(memcg_stock, cpu);
2320 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2321 unsigned int nr_pages,
2324 unsigned long nr_reclaimed = 0;
2327 unsigned long pflags;
2329 if (page_counter_read(&memcg->memory) <=
2330 READ_ONCE(memcg->memory.high))
2333 memcg_memory_event(memcg, MEMCG_HIGH);
2335 psi_memstall_enter(&pflags);
2336 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2338 psi_memstall_leave(&pflags);
2339 } while ((memcg = parent_mem_cgroup(memcg)) &&
2340 !mem_cgroup_is_root(memcg));
2342 return nr_reclaimed;
2345 static void high_work_func(struct work_struct *work)
2347 struct mem_cgroup *memcg;
2349 memcg = container_of(work, struct mem_cgroup, high_work);
2350 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2354 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2355 * enough to still cause a significant slowdown in most cases, while still
2356 * allowing diagnostics and tracing to proceed without becoming stuck.
2358 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2361 * When calculating the delay, we use these either side of the exponentiation to
2362 * maintain precision and scale to a reasonable number of jiffies (see the table
2365 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2366 * overage ratio to a delay.
2367 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2368 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2369 * to produce a reasonable delay curve.
2371 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2372 * reasonable delay curve compared to precision-adjusted overage, not
2373 * penalising heavily at first, but still making sure that growth beyond the
2374 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2375 * example, with a high of 100 megabytes:
2377 * +-------+------------------------+
2378 * | usage | time to allocate in ms |
2379 * +-------+------------------------+
2401 * +-------+------------------------+
2403 #define MEMCG_DELAY_PRECISION_SHIFT 20
2404 #define MEMCG_DELAY_SCALING_SHIFT 14
2406 static u64 calculate_overage(unsigned long usage, unsigned long high)
2414 * Prevent division by 0 in overage calculation by acting as if
2415 * it was a threshold of 1 page
2417 high = max(high, 1UL);
2419 overage = usage - high;
2420 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2421 return div64_u64(overage, high);
2424 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2426 u64 overage, max_overage = 0;
2429 overage = calculate_overage(page_counter_read(&memcg->memory),
2430 READ_ONCE(memcg->memory.high));
2431 max_overage = max(overage, max_overage);
2432 } while ((memcg = parent_mem_cgroup(memcg)) &&
2433 !mem_cgroup_is_root(memcg));
2438 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2440 u64 overage, max_overage = 0;
2443 overage = calculate_overage(page_counter_read(&memcg->swap),
2444 READ_ONCE(memcg->swap.high));
2446 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2447 max_overage = max(overage, max_overage);
2448 } while ((memcg = parent_mem_cgroup(memcg)) &&
2449 !mem_cgroup_is_root(memcg));
2455 * Get the number of jiffies that we should penalise a mischievous cgroup which
2456 * is exceeding its memory.high by checking both it and its ancestors.
2458 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2459 unsigned int nr_pages,
2462 unsigned long penalty_jiffies;
2468 * We use overage compared to memory.high to calculate the number of
2469 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2470 * fairly lenient on small overages, and increasingly harsh when the
2471 * memcg in question makes it clear that it has no intention of stopping
2472 * its crazy behaviour, so we exponentially increase the delay based on
2475 penalty_jiffies = max_overage * max_overage * HZ;
2476 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2477 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2480 * Factor in the task's own contribution to the overage, such that four
2481 * N-sized allocations are throttled approximately the same as one
2482 * 4N-sized allocation.
2484 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2485 * larger the current charge patch is than that.
2487 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2491 * Scheduled by try_charge() to be executed from the userland return path
2492 * and reclaims memory over the high limit.
2494 void mem_cgroup_handle_over_high(void)
2496 unsigned long penalty_jiffies;
2497 unsigned long pflags;
2498 unsigned long nr_reclaimed;
2499 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2500 int nr_retries = MAX_RECLAIM_RETRIES;
2501 struct mem_cgroup *memcg;
2502 bool in_retry = false;
2504 if (likely(!nr_pages))
2507 memcg = get_mem_cgroup_from_mm(current->mm);
2508 current->memcg_nr_pages_over_high = 0;
2512 * The allocating task should reclaim at least the batch size, but for
2513 * subsequent retries we only want to do what's necessary to prevent oom
2514 * or breaching resource isolation.
2516 * This is distinct from memory.max or page allocator behaviour because
2517 * memory.high is currently batched, whereas memory.max and the page
2518 * allocator run every time an allocation is made.
2520 nr_reclaimed = reclaim_high(memcg,
2521 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2525 * memory.high is breached and reclaim is unable to keep up. Throttle
2526 * allocators proactively to slow down excessive growth.
2528 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2529 mem_find_max_overage(memcg));
2531 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2532 swap_find_max_overage(memcg));
2535 * Clamp the max delay per usermode return so as to still keep the
2536 * application moving forwards and also permit diagnostics, albeit
2539 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2542 * Don't sleep if the amount of jiffies this memcg owes us is so low
2543 * that it's not even worth doing, in an attempt to be nice to those who
2544 * go only a small amount over their memory.high value and maybe haven't
2545 * been aggressively reclaimed enough yet.
2547 if (penalty_jiffies <= HZ / 100)
2551 * If reclaim is making forward progress but we're still over
2552 * memory.high, we want to encourage that rather than doing allocator
2555 if (nr_reclaimed || nr_retries--) {
2561 * If we exit early, we're guaranteed to die (since
2562 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2563 * need to account for any ill-begotten jiffies to pay them off later.
2565 psi_memstall_enter(&pflags);
2566 schedule_timeout_killable(penalty_jiffies);
2567 psi_memstall_leave(&pflags);
2570 css_put(&memcg->css);
2573 static int try_charge_memcg(struct mem_cgroup *memcg, gfp_t gfp_mask,
2574 unsigned int nr_pages)
2576 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2577 int nr_retries = MAX_RECLAIM_RETRIES;
2578 struct mem_cgroup *mem_over_limit;
2579 struct page_counter *counter;
2580 enum oom_status oom_status;
2581 unsigned long nr_reclaimed;
2582 bool passed_oom = false;
2583 bool may_swap = true;
2584 bool drained = false;
2585 unsigned long pflags;
2588 if (consume_stock(memcg, nr_pages))
2591 if (!do_memsw_account() ||
2592 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2593 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2595 if (do_memsw_account())
2596 page_counter_uncharge(&memcg->memsw, batch);
2597 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2599 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2603 if (batch > nr_pages) {
2609 * Memcg doesn't have a dedicated reserve for atomic
2610 * allocations. But like the global atomic pool, we need to
2611 * put the burden of reclaim on regular allocation requests
2612 * and let these go through as privileged allocations.
2614 if (gfp_mask & __GFP_ATOMIC)
2618 * Prevent unbounded recursion when reclaim operations need to
2619 * allocate memory. This might exceed the limits temporarily,
2620 * but we prefer facilitating memory reclaim and getting back
2621 * under the limit over triggering OOM kills in these cases.
2623 if (unlikely(current->flags & PF_MEMALLOC))
2626 if (unlikely(task_in_memcg_oom(current)))
2629 if (!gfpflags_allow_blocking(gfp_mask))
2632 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2634 psi_memstall_enter(&pflags);
2635 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2636 gfp_mask, may_swap);
2637 psi_memstall_leave(&pflags);
2639 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2643 drain_all_stock(mem_over_limit);
2648 if (gfp_mask & __GFP_NORETRY)
2651 * Even though the limit is exceeded at this point, reclaim
2652 * may have been able to free some pages. Retry the charge
2653 * before killing the task.
2655 * Only for regular pages, though: huge pages are rather
2656 * unlikely to succeed so close to the limit, and we fall back
2657 * to regular pages anyway in case of failure.
2659 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2662 * At task move, charge accounts can be doubly counted. So, it's
2663 * better to wait until the end of task_move if something is going on.
2665 if (mem_cgroup_wait_acct_move(mem_over_limit))
2671 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2674 /* Avoid endless loop for tasks bypassed by the oom killer */
2675 if (passed_oom && task_is_dying())
2679 * keep retrying as long as the memcg oom killer is able to make
2680 * a forward progress or bypass the charge if the oom killer
2681 * couldn't make any progress.
2683 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2684 get_order(nr_pages * PAGE_SIZE));
2685 if (oom_status == OOM_SUCCESS) {
2687 nr_retries = MAX_RECLAIM_RETRIES;
2691 if (!(gfp_mask & __GFP_NOFAIL))
2695 * The allocation either can't fail or will lead to more memory
2696 * being freed very soon. Allow memory usage go over the limit
2697 * temporarily by force charging it.
2699 page_counter_charge(&memcg->memory, nr_pages);
2700 if (do_memsw_account())
2701 page_counter_charge(&memcg->memsw, nr_pages);
2706 if (batch > nr_pages)
2707 refill_stock(memcg, batch - nr_pages);
2710 * If the hierarchy is above the normal consumption range, schedule
2711 * reclaim on returning to userland. We can perform reclaim here
2712 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2713 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2714 * not recorded as it most likely matches current's and won't
2715 * change in the meantime. As high limit is checked again before
2716 * reclaim, the cost of mismatch is negligible.
2719 bool mem_high, swap_high;
2721 mem_high = page_counter_read(&memcg->memory) >
2722 READ_ONCE(memcg->memory.high);
2723 swap_high = page_counter_read(&memcg->swap) >
2724 READ_ONCE(memcg->swap.high);
2726 /* Don't bother a random interrupted task */
2727 if (in_interrupt()) {
2729 schedule_work(&memcg->high_work);
2735 if (mem_high || swap_high) {
2737 * The allocating tasks in this cgroup will need to do
2738 * reclaim or be throttled to prevent further growth
2739 * of the memory or swap footprints.
2741 * Target some best-effort fairness between the tasks,
2742 * and distribute reclaim work and delay penalties
2743 * based on how much each task is actually allocating.
2745 current->memcg_nr_pages_over_high += batch;
2746 set_notify_resume(current);
2749 } while ((memcg = parent_mem_cgroup(memcg)));
2754 static inline int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2755 unsigned int nr_pages)
2757 if (mem_cgroup_is_root(memcg))
2760 return try_charge_memcg(memcg, gfp_mask, nr_pages);
2763 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2764 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2766 if (mem_cgroup_is_root(memcg))
2769 page_counter_uncharge(&memcg->memory, nr_pages);
2770 if (do_memsw_account())
2771 page_counter_uncharge(&memcg->memsw, nr_pages);
2775 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2777 VM_BUG_ON_PAGE(page_memcg(page), page);
2779 * Any of the following ensures page's memcg stability:
2783 * - lock_page_memcg()
2784 * - exclusive reference
2786 page->memcg_data = (unsigned long)memcg;
2789 static struct mem_cgroup *get_mem_cgroup_from_objcg(struct obj_cgroup *objcg)
2791 struct mem_cgroup *memcg;
2795 memcg = obj_cgroup_memcg(objcg);
2796 if (unlikely(!css_tryget(&memcg->css)))
2803 #ifdef CONFIG_MEMCG_KMEM
2805 * The allocated objcg pointers array is not accounted directly.
2806 * Moreover, it should not come from DMA buffer and is not readily
2807 * reclaimable. So those GFP bits should be masked off.
2809 #define OBJCGS_CLEAR_MASK (__GFP_DMA | __GFP_RECLAIMABLE | __GFP_ACCOUNT)
2811 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2812 gfp_t gfp, bool new_page)
2814 unsigned int objects = objs_per_slab_page(s, page);
2815 unsigned long memcg_data;
2818 gfp &= ~OBJCGS_CLEAR_MASK;
2819 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2824 memcg_data = (unsigned long) vec | MEMCG_DATA_OBJCGS;
2827 * If the slab page is brand new and nobody can yet access
2828 * it's memcg_data, no synchronization is required and
2829 * memcg_data can be simply assigned.
2831 page->memcg_data = memcg_data;
2832 } else if (cmpxchg(&page->memcg_data, 0, memcg_data)) {
2834 * If the slab page is already in use, somebody can allocate
2835 * and assign obj_cgroups in parallel. In this case the existing
2836 * objcg vector should be reused.
2842 kmemleak_not_leak(vec);
2847 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2849 * A passed kernel object can be a slab object or a generic kernel page, so
2850 * different mechanisms for getting the memory cgroup pointer should be used.
2851 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2852 * can not know for sure how the kernel object is implemented.
2853 * mem_cgroup_from_obj() can be safely used in such cases.
2855 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2856 * cgroup_mutex, etc.
2858 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2862 if (mem_cgroup_disabled())
2865 page = virt_to_head_page(p);
2868 * Slab objects are accounted individually, not per-page.
2869 * Memcg membership data for each individual object is saved in
2870 * the page->obj_cgroups.
2872 if (page_objcgs_check(page)) {
2873 struct obj_cgroup *objcg;
2876 off = obj_to_index(page->slab_cache, page, p);
2877 objcg = page_objcgs(page)[off];
2879 return obj_cgroup_memcg(objcg);
2885 * page_memcg_check() is used here, because page_has_obj_cgroups()
2886 * check above could fail because the object cgroups vector wasn't set
2887 * at that moment, but it can be set concurrently.
2888 * page_memcg_check(page) will guarantee that a proper memory
2889 * cgroup pointer or NULL will be returned.
2891 return page_memcg_check(page);
2894 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2896 struct obj_cgroup *objcg = NULL;
2897 struct mem_cgroup *memcg;
2899 if (memcg_kmem_bypass())
2903 if (unlikely(active_memcg()))
2904 memcg = active_memcg();
2906 memcg = mem_cgroup_from_task(current);
2908 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2909 objcg = rcu_dereference(memcg->objcg);
2910 if (objcg && obj_cgroup_tryget(objcg))
2919 static int memcg_alloc_cache_id(void)
2924 id = ida_simple_get(&memcg_cache_ida,
2925 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2929 if (id < memcg_nr_cache_ids)
2933 * There's no space for the new id in memcg_caches arrays,
2934 * so we have to grow them.
2936 down_write(&memcg_cache_ids_sem);
2938 size = 2 * (id + 1);
2939 if (size < MEMCG_CACHES_MIN_SIZE)
2940 size = MEMCG_CACHES_MIN_SIZE;
2941 else if (size > MEMCG_CACHES_MAX_SIZE)
2942 size = MEMCG_CACHES_MAX_SIZE;
2944 err = memcg_update_all_list_lrus(size);
2946 memcg_nr_cache_ids = size;
2948 up_write(&memcg_cache_ids_sem);
2951 ida_simple_remove(&memcg_cache_ida, id);
2957 static void memcg_free_cache_id(int id)
2959 ida_simple_remove(&memcg_cache_ida, id);
2963 * obj_cgroup_uncharge_pages: uncharge a number of kernel pages from a objcg
2964 * @objcg: object cgroup to uncharge
2965 * @nr_pages: number of pages to uncharge
2967 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
2968 unsigned int nr_pages)
2970 struct mem_cgroup *memcg;
2972 memcg = get_mem_cgroup_from_objcg(objcg);
2974 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2975 page_counter_uncharge(&memcg->kmem, nr_pages);
2976 refill_stock(memcg, nr_pages);
2978 css_put(&memcg->css);
2982 * obj_cgroup_charge_pages: charge a number of kernel pages to a objcg
2983 * @objcg: object cgroup to charge
2984 * @gfp: reclaim mode
2985 * @nr_pages: number of pages to charge
2987 * Returns 0 on success, an error code on failure.
2989 static int obj_cgroup_charge_pages(struct obj_cgroup *objcg, gfp_t gfp,
2990 unsigned int nr_pages)
2992 struct page_counter *counter;
2993 struct mem_cgroup *memcg;
2996 memcg = get_mem_cgroup_from_objcg(objcg);
2998 ret = try_charge_memcg(memcg, gfp, nr_pages);
3002 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3003 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3006 * Enforce __GFP_NOFAIL allocation because callers are not
3007 * prepared to see failures and likely do not have any failure
3010 if (gfp & __GFP_NOFAIL) {
3011 page_counter_charge(&memcg->kmem, nr_pages);
3014 cancel_charge(memcg, nr_pages);
3018 css_put(&memcg->css);
3024 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3025 * @page: page to charge
3026 * @gfp: reclaim mode
3027 * @order: allocation order
3029 * Returns 0 on success, an error code on failure.
3031 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3033 struct obj_cgroup *objcg;
3036 objcg = get_obj_cgroup_from_current();
3038 ret = obj_cgroup_charge_pages(objcg, gfp, 1 << order);
3040 page->memcg_data = (unsigned long)objcg |
3044 obj_cgroup_put(objcg);
3050 * __memcg_kmem_uncharge_page: uncharge a kmem page
3051 * @page: page to uncharge
3052 * @order: allocation order
3054 void __memcg_kmem_uncharge_page(struct page *page, int order)
3056 struct obj_cgroup *objcg;
3057 unsigned int nr_pages = 1 << order;
3059 if (!PageMemcgKmem(page))
3062 objcg = __page_objcg(page);
3063 obj_cgroup_uncharge_pages(objcg, nr_pages);
3064 page->memcg_data = 0;
3065 obj_cgroup_put(objcg);
3068 void mod_objcg_state(struct obj_cgroup *objcg, struct pglist_data *pgdat,
3069 enum node_stat_item idx, int nr)
3071 unsigned long flags;
3072 struct obj_stock *stock = get_obj_stock(&flags);
3076 * Save vmstat data in stock and skip vmstat array update unless
3077 * accumulating over a page of vmstat data or when pgdat or idx
3080 if (stock->cached_objcg != objcg) {
3081 drain_obj_stock(stock);
3082 obj_cgroup_get(objcg);
3083 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3084 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3085 stock->cached_objcg = objcg;
3086 stock->cached_pgdat = pgdat;
3087 } else if (stock->cached_pgdat != pgdat) {
3088 /* Flush the existing cached vmstat data */
3089 struct pglist_data *oldpg = stock->cached_pgdat;
3091 if (stock->nr_slab_reclaimable_b) {
3092 mod_objcg_mlstate(objcg, oldpg, NR_SLAB_RECLAIMABLE_B,
3093 stock->nr_slab_reclaimable_b);
3094 stock->nr_slab_reclaimable_b = 0;
3096 if (stock->nr_slab_unreclaimable_b) {
3097 mod_objcg_mlstate(objcg, oldpg, NR_SLAB_UNRECLAIMABLE_B,
3098 stock->nr_slab_unreclaimable_b);
3099 stock->nr_slab_unreclaimable_b = 0;
3101 stock->cached_pgdat = pgdat;
3104 bytes = (idx == NR_SLAB_RECLAIMABLE_B) ? &stock->nr_slab_reclaimable_b
3105 : &stock->nr_slab_unreclaimable_b;
3107 * Even for large object >= PAGE_SIZE, the vmstat data will still be
3108 * cached locally at least once before pushing it out.
3115 if (abs(*bytes) > PAGE_SIZE) {
3123 mod_objcg_mlstate(objcg, pgdat, idx, nr);
3125 put_obj_stock(flags);
3128 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3130 unsigned long flags;
3131 struct obj_stock *stock = get_obj_stock(&flags);
3134 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3135 stock->nr_bytes -= nr_bytes;
3139 put_obj_stock(flags);
3144 static void drain_obj_stock(struct obj_stock *stock)
3146 struct obj_cgroup *old = stock->cached_objcg;
3151 if (stock->nr_bytes) {
3152 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3153 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3156 obj_cgroup_uncharge_pages(old, nr_pages);
3159 * The leftover is flushed to the centralized per-memcg value.
3160 * On the next attempt to refill obj stock it will be moved
3161 * to a per-cpu stock (probably, on an other CPU), see
3162 * refill_obj_stock().
3164 * How often it's flushed is a trade-off between the memory
3165 * limit enforcement accuracy and potential CPU contention,
3166 * so it might be changed in the future.
3168 atomic_add(nr_bytes, &old->nr_charged_bytes);
3169 stock->nr_bytes = 0;
3173 * Flush the vmstat data in current stock
3175 if (stock->nr_slab_reclaimable_b || stock->nr_slab_unreclaimable_b) {
3176 if (stock->nr_slab_reclaimable_b) {
3177 mod_objcg_mlstate(old, stock->cached_pgdat,
3178 NR_SLAB_RECLAIMABLE_B,
3179 stock->nr_slab_reclaimable_b);
3180 stock->nr_slab_reclaimable_b = 0;
3182 if (stock->nr_slab_unreclaimable_b) {
3183 mod_objcg_mlstate(old, stock->cached_pgdat,
3184 NR_SLAB_UNRECLAIMABLE_B,
3185 stock->nr_slab_unreclaimable_b);
3186 stock->nr_slab_unreclaimable_b = 0;
3188 stock->cached_pgdat = NULL;
3191 obj_cgroup_put(old);
3192 stock->cached_objcg = NULL;
3195 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3196 struct mem_cgroup *root_memcg)
3198 struct mem_cgroup *memcg;
3200 if (in_task() && stock->task_obj.cached_objcg) {
3201 memcg = obj_cgroup_memcg(stock->task_obj.cached_objcg);
3202 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3205 if (stock->irq_obj.cached_objcg) {
3206 memcg = obj_cgroup_memcg(stock->irq_obj.cached_objcg);
3207 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3214 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes,
3215 bool allow_uncharge)
3217 unsigned long flags;
3218 struct obj_stock *stock = get_obj_stock(&flags);
3219 unsigned int nr_pages = 0;
3221 if (stock->cached_objcg != objcg) { /* reset if necessary */
3222 drain_obj_stock(stock);
3223 obj_cgroup_get(objcg);
3224 stock->cached_objcg = objcg;
3225 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3226 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3227 allow_uncharge = true; /* Allow uncharge when objcg changes */
3229 stock->nr_bytes += nr_bytes;
3231 if (allow_uncharge && (stock->nr_bytes > PAGE_SIZE)) {
3232 nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3233 stock->nr_bytes &= (PAGE_SIZE - 1);
3236 put_obj_stock(flags);
3239 obj_cgroup_uncharge_pages(objcg, nr_pages);
3242 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3244 unsigned int nr_pages, nr_bytes;
3247 if (consume_obj_stock(objcg, size))
3251 * In theory, objcg->nr_charged_bytes can have enough
3252 * pre-charged bytes to satisfy the allocation. However,
3253 * flushing objcg->nr_charged_bytes requires two atomic
3254 * operations, and objcg->nr_charged_bytes can't be big.
3255 * The shared objcg->nr_charged_bytes can also become a
3256 * performance bottleneck if all tasks of the same memcg are
3257 * trying to update it. So it's better to ignore it and try
3258 * grab some new pages. The stock's nr_bytes will be flushed to
3259 * objcg->nr_charged_bytes later on when objcg changes.
3261 * The stock's nr_bytes may contain enough pre-charged bytes
3262 * to allow one less page from being charged, but we can't rely
3263 * on the pre-charged bytes not being changed outside of
3264 * consume_obj_stock() or refill_obj_stock(). So ignore those
3265 * pre-charged bytes as well when charging pages. To avoid a
3266 * page uncharge right after a page charge, we set the
3267 * allow_uncharge flag to false when calling refill_obj_stock()
3268 * to temporarily allow the pre-charged bytes to exceed the page
3269 * size limit. The maximum reachable value of the pre-charged
3270 * bytes is (sizeof(object) + PAGE_SIZE - 2) if there is no data
3273 nr_pages = size >> PAGE_SHIFT;
3274 nr_bytes = size & (PAGE_SIZE - 1);
3279 ret = obj_cgroup_charge_pages(objcg, gfp, nr_pages);
3280 if (!ret && nr_bytes)
3281 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes, false);
3286 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3288 refill_obj_stock(objcg, size, true);
3291 #endif /* CONFIG_MEMCG_KMEM */
3294 * Because page_memcg(head) is not set on tails, set it now.
3296 void split_page_memcg(struct page *head, unsigned int nr)
3298 struct mem_cgroup *memcg = page_memcg(head);
3301 if (mem_cgroup_disabled() || !memcg)
3304 for (i = 1; i < nr; i++)
3305 head[i].memcg_data = head->memcg_data;
3307 if (PageMemcgKmem(head))
3308 obj_cgroup_get_many(__page_objcg(head), nr - 1);
3310 css_get_many(&memcg->css, nr - 1);
3313 #ifdef CONFIG_MEMCG_SWAP
3315 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3316 * @entry: swap entry to be moved
3317 * @from: mem_cgroup which the entry is moved from
3318 * @to: mem_cgroup which the entry is moved to
3320 * It succeeds only when the swap_cgroup's record for this entry is the same
3321 * as the mem_cgroup's id of @from.
3323 * Returns 0 on success, -EINVAL on failure.
3325 * The caller must have charged to @to, IOW, called page_counter_charge() about
3326 * both res and memsw, and called css_get().
3328 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3329 struct mem_cgroup *from, struct mem_cgroup *to)
3331 unsigned short old_id, new_id;
3333 old_id = mem_cgroup_id(from);
3334 new_id = mem_cgroup_id(to);
3336 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3337 mod_memcg_state(from, MEMCG_SWAP, -1);
3338 mod_memcg_state(to, MEMCG_SWAP, 1);
3344 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3345 struct mem_cgroup *from, struct mem_cgroup *to)
3351 static DEFINE_MUTEX(memcg_max_mutex);
3353 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3354 unsigned long max, bool memsw)
3356 bool enlarge = false;
3357 bool drained = false;
3359 bool limits_invariant;
3360 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3363 if (signal_pending(current)) {
3368 mutex_lock(&memcg_max_mutex);
3370 * Make sure that the new limit (memsw or memory limit) doesn't
3371 * break our basic invariant rule memory.max <= memsw.max.
3373 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3374 max <= memcg->memsw.max;
3375 if (!limits_invariant) {
3376 mutex_unlock(&memcg_max_mutex);
3380 if (max > counter->max)
3382 ret = page_counter_set_max(counter, max);
3383 mutex_unlock(&memcg_max_mutex);
3389 drain_all_stock(memcg);
3394 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3395 GFP_KERNEL, !memsw)) {
3401 if (!ret && enlarge)
3402 memcg_oom_recover(memcg);
3407 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3409 unsigned long *total_scanned)
3411 unsigned long nr_reclaimed = 0;
3412 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3413 unsigned long reclaimed;
3415 struct mem_cgroup_tree_per_node *mctz;
3416 unsigned long excess;
3417 unsigned long nr_scanned;
3422 mctz = soft_limit_tree_node(pgdat->node_id);
3425 * Do not even bother to check the largest node if the root
3426 * is empty. Do it lockless to prevent lock bouncing. Races
3427 * are acceptable as soft limit is best effort anyway.
3429 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3433 * This loop can run a while, specially if mem_cgroup's continuously
3434 * keep exceeding their soft limit and putting the system under
3441 mz = mem_cgroup_largest_soft_limit_node(mctz);
3446 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3447 gfp_mask, &nr_scanned);
3448 nr_reclaimed += reclaimed;
3449 *total_scanned += nr_scanned;
3450 spin_lock_irq(&mctz->lock);
3451 __mem_cgroup_remove_exceeded(mz, mctz);
3454 * If we failed to reclaim anything from this memory cgroup
3455 * it is time to move on to the next cgroup
3459 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3461 excess = soft_limit_excess(mz->memcg);
3463 * One school of thought says that we should not add
3464 * back the node to the tree if reclaim returns 0.
3465 * But our reclaim could return 0, simply because due
3466 * to priority we are exposing a smaller subset of
3467 * memory to reclaim from. Consider this as a longer
3470 /* If excess == 0, no tree ops */
3471 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3472 spin_unlock_irq(&mctz->lock);
3473 css_put(&mz->memcg->css);
3476 * Could not reclaim anything and there are no more
3477 * mem cgroups to try or we seem to be looping without
3478 * reclaiming anything.
3480 if (!nr_reclaimed &&
3482 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3484 } while (!nr_reclaimed);
3486 css_put(&next_mz->memcg->css);
3487 return nr_reclaimed;
3491 * Reclaims as many pages from the given memcg as possible.
3493 * Caller is responsible for holding css reference for memcg.
3495 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3497 int nr_retries = MAX_RECLAIM_RETRIES;
3499 /* we call try-to-free pages for make this cgroup empty */
3500 lru_add_drain_all();
3502 drain_all_stock(memcg);
3504 /* try to free all pages in this cgroup */
3505 while (nr_retries && page_counter_read(&memcg->memory)) {
3508 if (signal_pending(current))
3511 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3515 /* maybe some writeback is necessary */
3516 congestion_wait(BLK_RW_ASYNC, HZ/10);
3524 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3525 char *buf, size_t nbytes,
3528 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3530 if (mem_cgroup_is_root(memcg))
3532 return mem_cgroup_force_empty(memcg) ?: nbytes;
3535 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3541 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3542 struct cftype *cft, u64 val)
3547 pr_warn_once("Non-hierarchical mode is deprecated. "
3548 "Please report your usecase to linux-mm@kvack.org if you "
3549 "depend on this functionality.\n");
3554 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3558 if (mem_cgroup_is_root(memcg)) {
3559 mem_cgroup_flush_stats();
3560 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3561 memcg_page_state(memcg, NR_ANON_MAPPED);
3563 val += memcg_page_state(memcg, MEMCG_SWAP);
3566 val = page_counter_read(&memcg->memory);
3568 val = page_counter_read(&memcg->memsw);
3581 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3584 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3585 struct page_counter *counter;
3587 switch (MEMFILE_TYPE(cft->private)) {
3589 counter = &memcg->memory;
3592 counter = &memcg->memsw;
3595 counter = &memcg->kmem;
3598 counter = &memcg->tcpmem;
3604 switch (MEMFILE_ATTR(cft->private)) {
3606 if (counter == &memcg->memory)
3607 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3608 if (counter == &memcg->memsw)
3609 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3610 return (u64)page_counter_read(counter) * PAGE_SIZE;
3612 return (u64)counter->max * PAGE_SIZE;
3614 return (u64)counter->watermark * PAGE_SIZE;
3616 return counter->failcnt;
3617 case RES_SOFT_LIMIT:
3618 return (u64)memcg->soft_limit * PAGE_SIZE;
3624 #ifdef CONFIG_MEMCG_KMEM
3625 static int memcg_online_kmem(struct mem_cgroup *memcg)
3627 struct obj_cgroup *objcg;
3630 if (cgroup_memory_nokmem)
3633 BUG_ON(memcg->kmemcg_id >= 0);
3634 BUG_ON(memcg->kmem_state);
3636 memcg_id = memcg_alloc_cache_id();
3640 objcg = obj_cgroup_alloc();
3642 memcg_free_cache_id(memcg_id);
3645 objcg->memcg = memcg;
3646 rcu_assign_pointer(memcg->objcg, objcg);
3648 static_branch_enable(&memcg_kmem_enabled_key);
3650 memcg->kmemcg_id = memcg_id;
3651 memcg->kmem_state = KMEM_ONLINE;
3656 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3658 struct cgroup_subsys_state *css;
3659 struct mem_cgroup *parent, *child;
3662 if (memcg->kmem_state != KMEM_ONLINE)
3665 memcg->kmem_state = KMEM_ALLOCATED;
3667 parent = parent_mem_cgroup(memcg);
3669 parent = root_mem_cgroup;
3671 memcg_reparent_objcgs(memcg, parent);
3673 kmemcg_id = memcg->kmemcg_id;
3674 BUG_ON(kmemcg_id < 0);
3677 * Change kmemcg_id of this cgroup and all its descendants to the
3678 * parent's id, and then move all entries from this cgroup's list_lrus
3679 * to ones of the parent. After we have finished, all list_lrus
3680 * corresponding to this cgroup are guaranteed to remain empty. The
3681 * ordering is imposed by list_lru_node->lock taken by
3682 * memcg_drain_all_list_lrus().
3684 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3685 css_for_each_descendant_pre(css, &memcg->css) {
3686 child = mem_cgroup_from_css(css);
3687 BUG_ON(child->kmemcg_id != kmemcg_id);
3688 child->kmemcg_id = parent->kmemcg_id;
3692 memcg_drain_all_list_lrus(kmemcg_id, parent);
3694 memcg_free_cache_id(kmemcg_id);
3697 static void memcg_free_kmem(struct mem_cgroup *memcg)
3699 /* css_alloc() failed, offlining didn't happen */
3700 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3701 memcg_offline_kmem(memcg);
3704 static int memcg_online_kmem(struct mem_cgroup *memcg)
3708 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3711 static void memcg_free_kmem(struct mem_cgroup *memcg)
3714 #endif /* CONFIG_MEMCG_KMEM */
3716 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3721 mutex_lock(&memcg_max_mutex);
3722 ret = page_counter_set_max(&memcg->kmem, max);
3723 mutex_unlock(&memcg_max_mutex);
3727 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3731 mutex_lock(&memcg_max_mutex);
3733 ret = page_counter_set_max(&memcg->tcpmem, max);
3737 if (!memcg->tcpmem_active) {
3739 * The active flag needs to be written after the static_key
3740 * update. This is what guarantees that the socket activation
3741 * function is the last one to run. See mem_cgroup_sk_alloc()
3742 * for details, and note that we don't mark any socket as
3743 * belonging to this memcg until that flag is up.
3745 * We need to do this, because static_keys will span multiple
3746 * sites, but we can't control their order. If we mark a socket
3747 * as accounted, but the accounting functions are not patched in
3748 * yet, we'll lose accounting.
3750 * We never race with the readers in mem_cgroup_sk_alloc(),
3751 * because when this value change, the code to process it is not
3754 static_branch_inc(&memcg_sockets_enabled_key);
3755 memcg->tcpmem_active = true;
3758 mutex_unlock(&memcg_max_mutex);
3763 * The user of this function is...
3766 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3767 char *buf, size_t nbytes, loff_t off)
3769 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3770 unsigned long nr_pages;
3773 buf = strstrip(buf);
3774 ret = page_counter_memparse(buf, "-1", &nr_pages);
3778 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3780 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3784 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3786 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3789 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3792 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3793 "Please report your usecase to linux-mm@kvack.org if you "
3794 "depend on this functionality.\n");
3795 ret = memcg_update_kmem_max(memcg, nr_pages);
3798 ret = memcg_update_tcp_max(memcg, nr_pages);
3802 case RES_SOFT_LIMIT:
3803 memcg->soft_limit = nr_pages;
3807 return ret ?: nbytes;
3810 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3811 size_t nbytes, loff_t off)
3813 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3814 struct page_counter *counter;
3816 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3818 counter = &memcg->memory;
3821 counter = &memcg->memsw;
3824 counter = &memcg->kmem;
3827 counter = &memcg->tcpmem;
3833 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3835 page_counter_reset_watermark(counter);
3838 counter->failcnt = 0;
3847 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3850 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3854 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3855 struct cftype *cft, u64 val)
3857 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3859 if (val & ~MOVE_MASK)
3863 * No kind of locking is needed in here, because ->can_attach() will
3864 * check this value once in the beginning of the process, and then carry
3865 * on with stale data. This means that changes to this value will only
3866 * affect task migrations starting after the change.
3868 memcg->move_charge_at_immigrate = val;
3872 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3873 struct cftype *cft, u64 val)
3881 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3882 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3883 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3885 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3886 int nid, unsigned int lru_mask, bool tree)
3888 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3889 unsigned long nr = 0;
3892 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3895 if (!(BIT(lru) & lru_mask))
3898 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3900 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3905 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3906 unsigned int lru_mask,
3909 unsigned long nr = 0;
3913 if (!(BIT(lru) & lru_mask))
3916 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3918 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3923 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3927 unsigned int lru_mask;
3930 static const struct numa_stat stats[] = {
3931 { "total", LRU_ALL },
3932 { "file", LRU_ALL_FILE },
3933 { "anon", LRU_ALL_ANON },
3934 { "unevictable", BIT(LRU_UNEVICTABLE) },
3936 const struct numa_stat *stat;
3938 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3940 mem_cgroup_flush_stats();
3942 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3943 seq_printf(m, "%s=%lu", stat->name,
3944 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3946 for_each_node_state(nid, N_MEMORY)
3947 seq_printf(m, " N%d=%lu", nid,
3948 mem_cgroup_node_nr_lru_pages(memcg, nid,
3949 stat->lru_mask, false));
3953 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3955 seq_printf(m, "hierarchical_%s=%lu", stat->name,
3956 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3958 for_each_node_state(nid, N_MEMORY)
3959 seq_printf(m, " N%d=%lu", nid,
3960 mem_cgroup_node_nr_lru_pages(memcg, nid,
3961 stat->lru_mask, true));
3967 #endif /* CONFIG_NUMA */
3969 static const unsigned int memcg1_stats[] = {
3972 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3982 static const char *const memcg1_stat_names[] = {
3985 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3995 /* Universal VM events cgroup1 shows, original sort order */
3996 static const unsigned int memcg1_events[] = {
4003 static int memcg_stat_show(struct seq_file *m, void *v)
4005 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4006 unsigned long memory, memsw;
4007 struct mem_cgroup *mi;
4010 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4012 mem_cgroup_flush_stats();
4014 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4017 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4019 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4020 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4023 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4024 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4025 memcg_events_local(memcg, memcg1_events[i]));
4027 for (i = 0; i < NR_LRU_LISTS; i++)
4028 seq_printf(m, "%s %lu\n", lru_list_name(i),
4029 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4032 /* Hierarchical information */
4033 memory = memsw = PAGE_COUNTER_MAX;
4034 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4035 memory = min(memory, READ_ONCE(mi->memory.max));
4036 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4038 seq_printf(m, "hierarchical_memory_limit %llu\n",
4039 (u64)memory * PAGE_SIZE);
4040 if (do_memsw_account())
4041 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4042 (u64)memsw * PAGE_SIZE);
4044 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4047 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4049 nr = memcg_page_state(memcg, memcg1_stats[i]);
4050 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4051 (u64)nr * PAGE_SIZE);
4054 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4055 seq_printf(m, "total_%s %llu\n",
4056 vm_event_name(memcg1_events[i]),
4057 (u64)memcg_events(memcg, memcg1_events[i]));
4059 for (i = 0; i < NR_LRU_LISTS; i++)
4060 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4061 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4064 #ifdef CONFIG_DEBUG_VM
4067 struct mem_cgroup_per_node *mz;
4068 unsigned long anon_cost = 0;
4069 unsigned long file_cost = 0;
4071 for_each_online_pgdat(pgdat) {
4072 mz = memcg->nodeinfo[pgdat->node_id];
4074 anon_cost += mz->lruvec.anon_cost;
4075 file_cost += mz->lruvec.file_cost;
4077 seq_printf(m, "anon_cost %lu\n", anon_cost);
4078 seq_printf(m, "file_cost %lu\n", file_cost);
4085 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4088 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4090 return mem_cgroup_swappiness(memcg);
4093 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4094 struct cftype *cft, u64 val)
4096 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4101 if (!mem_cgroup_is_root(memcg))
4102 memcg->swappiness = val;
4104 vm_swappiness = val;
4109 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4111 struct mem_cgroup_threshold_ary *t;
4112 unsigned long usage;
4117 t = rcu_dereference(memcg->thresholds.primary);
4119 t = rcu_dereference(memcg->memsw_thresholds.primary);
4124 usage = mem_cgroup_usage(memcg, swap);
4127 * current_threshold points to threshold just below or equal to usage.
4128 * If it's not true, a threshold was crossed after last
4129 * call of __mem_cgroup_threshold().
4131 i = t->current_threshold;
4134 * Iterate backward over array of thresholds starting from
4135 * current_threshold and check if a threshold is crossed.
4136 * If none of thresholds below usage is crossed, we read
4137 * only one element of the array here.
4139 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4140 eventfd_signal(t->entries[i].eventfd, 1);
4142 /* i = current_threshold + 1 */
4146 * Iterate forward over array of thresholds starting from
4147 * current_threshold+1 and check if a threshold is crossed.
4148 * If none of thresholds above usage is crossed, we read
4149 * only one element of the array here.
4151 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4152 eventfd_signal(t->entries[i].eventfd, 1);
4154 /* Update current_threshold */
4155 t->current_threshold = i - 1;
4160 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4163 __mem_cgroup_threshold(memcg, false);
4164 if (do_memsw_account())
4165 __mem_cgroup_threshold(memcg, true);
4167 memcg = parent_mem_cgroup(memcg);
4171 static int compare_thresholds(const void *a, const void *b)
4173 const struct mem_cgroup_threshold *_a = a;
4174 const struct mem_cgroup_threshold *_b = b;
4176 if (_a->threshold > _b->threshold)
4179 if (_a->threshold < _b->threshold)
4185 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4187 struct mem_cgroup_eventfd_list *ev;
4189 spin_lock(&memcg_oom_lock);
4191 list_for_each_entry(ev, &memcg->oom_notify, list)
4192 eventfd_signal(ev->eventfd, 1);
4194 spin_unlock(&memcg_oom_lock);
4198 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4200 struct mem_cgroup *iter;
4202 for_each_mem_cgroup_tree(iter, memcg)
4203 mem_cgroup_oom_notify_cb(iter);
4206 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4207 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4209 struct mem_cgroup_thresholds *thresholds;
4210 struct mem_cgroup_threshold_ary *new;
4211 unsigned long threshold;
4212 unsigned long usage;
4215 ret = page_counter_memparse(args, "-1", &threshold);
4219 mutex_lock(&memcg->thresholds_lock);
4222 thresholds = &memcg->thresholds;
4223 usage = mem_cgroup_usage(memcg, false);
4224 } else if (type == _MEMSWAP) {
4225 thresholds = &memcg->memsw_thresholds;
4226 usage = mem_cgroup_usage(memcg, true);
4230 /* Check if a threshold crossed before adding a new one */
4231 if (thresholds->primary)
4232 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4234 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4236 /* Allocate memory for new array of thresholds */
4237 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4244 /* Copy thresholds (if any) to new array */
4245 if (thresholds->primary)
4246 memcpy(new->entries, thresholds->primary->entries,
4247 flex_array_size(new, entries, size - 1));
4249 /* Add new threshold */
4250 new->entries[size - 1].eventfd = eventfd;
4251 new->entries[size - 1].threshold = threshold;
4253 /* Sort thresholds. Registering of new threshold isn't time-critical */
4254 sort(new->entries, size, sizeof(*new->entries),
4255 compare_thresholds, NULL);
4257 /* Find current threshold */
4258 new->current_threshold = -1;
4259 for (i = 0; i < size; i++) {
4260 if (new->entries[i].threshold <= usage) {
4262 * new->current_threshold will not be used until
4263 * rcu_assign_pointer(), so it's safe to increment
4266 ++new->current_threshold;
4271 /* Free old spare buffer and save old primary buffer as spare */
4272 kfree(thresholds->spare);
4273 thresholds->spare = thresholds->primary;
4275 rcu_assign_pointer(thresholds->primary, new);
4277 /* To be sure that nobody uses thresholds */
4281 mutex_unlock(&memcg->thresholds_lock);
4286 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4287 struct eventfd_ctx *eventfd, const char *args)
4289 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4292 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4293 struct eventfd_ctx *eventfd, const char *args)
4295 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4298 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4299 struct eventfd_ctx *eventfd, enum res_type type)
4301 struct mem_cgroup_thresholds *thresholds;
4302 struct mem_cgroup_threshold_ary *new;
4303 unsigned long usage;
4304 int i, j, size, entries;
4306 mutex_lock(&memcg->thresholds_lock);
4309 thresholds = &memcg->thresholds;
4310 usage = mem_cgroup_usage(memcg, false);
4311 } else if (type == _MEMSWAP) {
4312 thresholds = &memcg->memsw_thresholds;
4313 usage = mem_cgroup_usage(memcg, true);
4317 if (!thresholds->primary)
4320 /* Check if a threshold crossed before removing */
4321 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4323 /* Calculate new number of threshold */
4325 for (i = 0; i < thresholds->primary->size; i++) {
4326 if (thresholds->primary->entries[i].eventfd != eventfd)
4332 new = thresholds->spare;
4334 /* If no items related to eventfd have been cleared, nothing to do */
4338 /* Set thresholds array to NULL if we don't have thresholds */
4347 /* Copy thresholds and find current threshold */
4348 new->current_threshold = -1;
4349 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4350 if (thresholds->primary->entries[i].eventfd == eventfd)
4353 new->entries[j] = thresholds->primary->entries[i];
4354 if (new->entries[j].threshold <= usage) {
4356 * new->current_threshold will not be used
4357 * until rcu_assign_pointer(), so it's safe to increment
4360 ++new->current_threshold;
4366 /* Swap primary and spare array */
4367 thresholds->spare = thresholds->primary;
4369 rcu_assign_pointer(thresholds->primary, new);
4371 /* To be sure that nobody uses thresholds */
4374 /* If all events are unregistered, free the spare array */
4376 kfree(thresholds->spare);
4377 thresholds->spare = NULL;
4380 mutex_unlock(&memcg->thresholds_lock);
4383 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4384 struct eventfd_ctx *eventfd)
4386 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4389 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4390 struct eventfd_ctx *eventfd)
4392 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4395 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4396 struct eventfd_ctx *eventfd, const char *args)
4398 struct mem_cgroup_eventfd_list *event;
4400 event = kmalloc(sizeof(*event), GFP_KERNEL);
4404 spin_lock(&memcg_oom_lock);
4406 event->eventfd = eventfd;
4407 list_add(&event->list, &memcg->oom_notify);
4409 /* already in OOM ? */
4410 if (memcg->under_oom)
4411 eventfd_signal(eventfd, 1);
4412 spin_unlock(&memcg_oom_lock);
4417 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4418 struct eventfd_ctx *eventfd)
4420 struct mem_cgroup_eventfd_list *ev, *tmp;
4422 spin_lock(&memcg_oom_lock);
4424 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4425 if (ev->eventfd == eventfd) {
4426 list_del(&ev->list);
4431 spin_unlock(&memcg_oom_lock);
4434 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4436 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4438 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4439 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4440 seq_printf(sf, "oom_kill %lu\n",
4441 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4445 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4446 struct cftype *cft, u64 val)
4448 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4450 /* cannot set to root cgroup and only 0 and 1 are allowed */
4451 if (mem_cgroup_is_root(memcg) || !((val == 0) || (val == 1)))
4454 memcg->oom_kill_disable = val;
4456 memcg_oom_recover(memcg);
4461 #ifdef CONFIG_CGROUP_WRITEBACK
4463 #include <trace/events/writeback.h>
4465 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4467 return wb_domain_init(&memcg->cgwb_domain, gfp);
4470 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4472 wb_domain_exit(&memcg->cgwb_domain);
4475 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4477 wb_domain_size_changed(&memcg->cgwb_domain);
4480 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4482 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4484 if (!memcg->css.parent)
4487 return &memcg->cgwb_domain;
4491 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4492 * @wb: bdi_writeback in question
4493 * @pfilepages: out parameter for number of file pages
4494 * @pheadroom: out parameter for number of allocatable pages according to memcg
4495 * @pdirty: out parameter for number of dirty pages
4496 * @pwriteback: out parameter for number of pages under writeback
4498 * Determine the numbers of file, headroom, dirty, and writeback pages in
4499 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4500 * is a bit more involved.
4502 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4503 * headroom is calculated as the lowest headroom of itself and the
4504 * ancestors. Note that this doesn't consider the actual amount of
4505 * available memory in the system. The caller should further cap
4506 * *@pheadroom accordingly.
4508 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4509 unsigned long *pheadroom, unsigned long *pdirty,
4510 unsigned long *pwriteback)
4512 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4513 struct mem_cgroup *parent;
4515 mem_cgroup_flush_stats();
4517 *pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
4518 *pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
4519 *pfilepages = memcg_page_state(memcg, NR_INACTIVE_FILE) +
4520 memcg_page_state(memcg, NR_ACTIVE_FILE);
4522 *pheadroom = PAGE_COUNTER_MAX;
4523 while ((parent = parent_mem_cgroup(memcg))) {
4524 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4525 READ_ONCE(memcg->memory.high));
4526 unsigned long used = page_counter_read(&memcg->memory);
4528 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4534 * Foreign dirty flushing
4536 * There's an inherent mismatch between memcg and writeback. The former
4537 * tracks ownership per-page while the latter per-inode. This was a
4538 * deliberate design decision because honoring per-page ownership in the
4539 * writeback path is complicated, may lead to higher CPU and IO overheads
4540 * and deemed unnecessary given that write-sharing an inode across
4541 * different cgroups isn't a common use-case.
4543 * Combined with inode majority-writer ownership switching, this works well
4544 * enough in most cases but there are some pathological cases. For
4545 * example, let's say there are two cgroups A and B which keep writing to
4546 * different but confined parts of the same inode. B owns the inode and
4547 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4548 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4549 * triggering background writeback. A will be slowed down without a way to
4550 * make writeback of the dirty pages happen.
4552 * Conditions like the above can lead to a cgroup getting repeatedly and
4553 * severely throttled after making some progress after each
4554 * dirty_expire_interval while the underlying IO device is almost
4557 * Solving this problem completely requires matching the ownership tracking
4558 * granularities between memcg and writeback in either direction. However,
4559 * the more egregious behaviors can be avoided by simply remembering the
4560 * most recent foreign dirtying events and initiating remote flushes on
4561 * them when local writeback isn't enough to keep the memory clean enough.
4563 * The following two functions implement such mechanism. When a foreign
4564 * page - a page whose memcg and writeback ownerships don't match - is
4565 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4566 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4567 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4568 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4569 * foreign bdi_writebacks which haven't expired. Both the numbers of
4570 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4571 * limited to MEMCG_CGWB_FRN_CNT.
4573 * The mechanism only remembers IDs and doesn't hold any object references.
4574 * As being wrong occasionally doesn't matter, updates and accesses to the
4575 * records are lockless and racy.
4577 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4578 struct bdi_writeback *wb)
4580 struct mem_cgroup *memcg = page_memcg(page);
4581 struct memcg_cgwb_frn *frn;
4582 u64 now = get_jiffies_64();
4583 u64 oldest_at = now;
4587 trace_track_foreign_dirty(page, wb);
4590 * Pick the slot to use. If there is already a slot for @wb, keep
4591 * using it. If not replace the oldest one which isn't being
4594 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4595 frn = &memcg->cgwb_frn[i];
4596 if (frn->bdi_id == wb->bdi->id &&
4597 frn->memcg_id == wb->memcg_css->id)
4599 if (time_before64(frn->at, oldest_at) &&
4600 atomic_read(&frn->done.cnt) == 1) {
4602 oldest_at = frn->at;
4606 if (i < MEMCG_CGWB_FRN_CNT) {
4608 * Re-using an existing one. Update timestamp lazily to
4609 * avoid making the cacheline hot. We want them to be
4610 * reasonably up-to-date and significantly shorter than
4611 * dirty_expire_interval as that's what expires the record.
4612 * Use the shorter of 1s and dirty_expire_interval / 8.
4614 unsigned long update_intv =
4615 min_t(unsigned long, HZ,
4616 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4618 if (time_before64(frn->at, now - update_intv))
4620 } else if (oldest >= 0) {
4621 /* replace the oldest free one */
4622 frn = &memcg->cgwb_frn[oldest];
4623 frn->bdi_id = wb->bdi->id;
4624 frn->memcg_id = wb->memcg_css->id;
4629 /* issue foreign writeback flushes for recorded foreign dirtying events */
4630 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4632 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4633 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4634 u64 now = jiffies_64;
4637 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4638 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4641 * If the record is older than dirty_expire_interval,
4642 * writeback on it has already started. No need to kick it
4643 * off again. Also, don't start a new one if there's
4644 * already one in flight.
4646 if (time_after64(frn->at, now - intv) &&
4647 atomic_read(&frn->done.cnt) == 1) {
4649 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4650 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id,
4651 WB_REASON_FOREIGN_FLUSH,
4657 #else /* CONFIG_CGROUP_WRITEBACK */
4659 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4664 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4668 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4672 #endif /* CONFIG_CGROUP_WRITEBACK */
4675 * DO NOT USE IN NEW FILES.
4677 * "cgroup.event_control" implementation.
4679 * This is way over-engineered. It tries to support fully configurable
4680 * events for each user. Such level of flexibility is completely
4681 * unnecessary especially in the light of the planned unified hierarchy.
4683 * Please deprecate this and replace with something simpler if at all
4688 * Unregister event and free resources.
4690 * Gets called from workqueue.
4692 static void memcg_event_remove(struct work_struct *work)
4694 struct mem_cgroup_event *event =
4695 container_of(work, struct mem_cgroup_event, remove);
4696 struct mem_cgroup *memcg = event->memcg;
4698 remove_wait_queue(event->wqh, &event->wait);
4700 event->unregister_event(memcg, event->eventfd);
4702 /* Notify userspace the event is going away. */
4703 eventfd_signal(event->eventfd, 1);
4705 eventfd_ctx_put(event->eventfd);
4707 css_put(&memcg->css);
4711 * Gets called on EPOLLHUP on eventfd when user closes it.
4713 * Called with wqh->lock held and interrupts disabled.
4715 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4716 int sync, void *key)
4718 struct mem_cgroup_event *event =
4719 container_of(wait, struct mem_cgroup_event, wait);
4720 struct mem_cgroup *memcg = event->memcg;
4721 __poll_t flags = key_to_poll(key);
4723 if (flags & EPOLLHUP) {
4725 * If the event has been detached at cgroup removal, we
4726 * can simply return knowing the other side will cleanup
4729 * We can't race against event freeing since the other
4730 * side will require wqh->lock via remove_wait_queue(),
4733 spin_lock(&memcg->event_list_lock);
4734 if (!list_empty(&event->list)) {
4735 list_del_init(&event->list);
4737 * We are in atomic context, but cgroup_event_remove()
4738 * may sleep, so we have to call it in workqueue.
4740 schedule_work(&event->remove);
4742 spin_unlock(&memcg->event_list_lock);
4748 static void memcg_event_ptable_queue_proc(struct file *file,
4749 wait_queue_head_t *wqh, poll_table *pt)
4751 struct mem_cgroup_event *event =
4752 container_of(pt, struct mem_cgroup_event, pt);
4755 add_wait_queue(wqh, &event->wait);
4759 * DO NOT USE IN NEW FILES.
4761 * Parse input and register new cgroup event handler.
4763 * Input must be in format '<event_fd> <control_fd> <args>'.
4764 * Interpretation of args is defined by control file implementation.
4766 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4767 char *buf, size_t nbytes, loff_t off)
4769 struct cgroup_subsys_state *css = of_css(of);
4770 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4771 struct mem_cgroup_event *event;
4772 struct cgroup_subsys_state *cfile_css;
4773 unsigned int efd, cfd;
4780 buf = strstrip(buf);
4782 efd = simple_strtoul(buf, &endp, 10);
4787 cfd = simple_strtoul(buf, &endp, 10);
4788 if ((*endp != ' ') && (*endp != '\0'))
4792 event = kzalloc(sizeof(*event), GFP_KERNEL);
4796 event->memcg = memcg;
4797 INIT_LIST_HEAD(&event->list);
4798 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4799 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4800 INIT_WORK(&event->remove, memcg_event_remove);
4808 event->eventfd = eventfd_ctx_fileget(efile.file);
4809 if (IS_ERR(event->eventfd)) {
4810 ret = PTR_ERR(event->eventfd);
4817 goto out_put_eventfd;
4820 /* the process need read permission on control file */
4821 /* AV: shouldn't we check that it's been opened for read instead? */
4822 ret = file_permission(cfile.file, MAY_READ);
4827 * Determine the event callbacks and set them in @event. This used
4828 * to be done via struct cftype but cgroup core no longer knows
4829 * about these events. The following is crude but the whole thing
4830 * is for compatibility anyway.
4832 * DO NOT ADD NEW FILES.
4834 name = cfile.file->f_path.dentry->d_name.name;
4836 if (!strcmp(name, "memory.usage_in_bytes")) {
4837 event->register_event = mem_cgroup_usage_register_event;
4838 event->unregister_event = mem_cgroup_usage_unregister_event;
4839 } else if (!strcmp(name, "memory.oom_control")) {
4840 event->register_event = mem_cgroup_oom_register_event;
4841 event->unregister_event = mem_cgroup_oom_unregister_event;
4842 } else if (!strcmp(name, "memory.pressure_level")) {
4843 event->register_event = vmpressure_register_event;
4844 event->unregister_event = vmpressure_unregister_event;
4845 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4846 event->register_event = memsw_cgroup_usage_register_event;
4847 event->unregister_event = memsw_cgroup_usage_unregister_event;
4854 * Verify @cfile should belong to @css. Also, remaining events are
4855 * automatically removed on cgroup destruction but the removal is
4856 * asynchronous, so take an extra ref on @css.
4858 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4859 &memory_cgrp_subsys);
4861 if (IS_ERR(cfile_css))
4863 if (cfile_css != css) {
4868 ret = event->register_event(memcg, event->eventfd, buf);
4872 vfs_poll(efile.file, &event->pt);
4874 spin_lock_irq(&memcg->event_list_lock);
4875 list_add(&event->list, &memcg->event_list);
4876 spin_unlock_irq(&memcg->event_list_lock);
4888 eventfd_ctx_put(event->eventfd);
4897 static struct cftype mem_cgroup_legacy_files[] = {
4899 .name = "usage_in_bytes",
4900 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4901 .read_u64 = mem_cgroup_read_u64,
4904 .name = "max_usage_in_bytes",
4905 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4906 .write = mem_cgroup_reset,
4907 .read_u64 = mem_cgroup_read_u64,
4910 .name = "limit_in_bytes",
4911 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4912 .write = mem_cgroup_write,
4913 .read_u64 = mem_cgroup_read_u64,
4916 .name = "soft_limit_in_bytes",
4917 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4918 .write = mem_cgroup_write,
4919 .read_u64 = mem_cgroup_read_u64,
4923 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4924 .write = mem_cgroup_reset,
4925 .read_u64 = mem_cgroup_read_u64,
4929 .seq_show = memcg_stat_show,
4932 .name = "force_empty",
4933 .write = mem_cgroup_force_empty_write,
4936 .name = "use_hierarchy",
4937 .write_u64 = mem_cgroup_hierarchy_write,
4938 .read_u64 = mem_cgroup_hierarchy_read,
4941 .name = "cgroup.event_control", /* XXX: for compat */
4942 .write = memcg_write_event_control,
4943 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4946 .name = "swappiness",
4947 .read_u64 = mem_cgroup_swappiness_read,
4948 .write_u64 = mem_cgroup_swappiness_write,
4951 .name = "move_charge_at_immigrate",
4952 .read_u64 = mem_cgroup_move_charge_read,
4953 .write_u64 = mem_cgroup_move_charge_write,
4956 .name = "oom_control",
4957 .seq_show = mem_cgroup_oom_control_read,
4958 .write_u64 = mem_cgroup_oom_control_write,
4959 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4962 .name = "pressure_level",
4966 .name = "numa_stat",
4967 .seq_show = memcg_numa_stat_show,
4971 .name = "kmem.limit_in_bytes",
4972 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4973 .write = mem_cgroup_write,
4974 .read_u64 = mem_cgroup_read_u64,
4977 .name = "kmem.usage_in_bytes",
4978 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4979 .read_u64 = mem_cgroup_read_u64,
4982 .name = "kmem.failcnt",
4983 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4984 .write = mem_cgroup_reset,
4985 .read_u64 = mem_cgroup_read_u64,
4988 .name = "kmem.max_usage_in_bytes",
4989 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4990 .write = mem_cgroup_reset,
4991 .read_u64 = mem_cgroup_read_u64,
4993 #if defined(CONFIG_MEMCG_KMEM) && \
4994 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
4996 .name = "kmem.slabinfo",
4997 .seq_show = memcg_slab_show,
5001 .name = "kmem.tcp.limit_in_bytes",
5002 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5003 .write = mem_cgroup_write,
5004 .read_u64 = mem_cgroup_read_u64,
5007 .name = "kmem.tcp.usage_in_bytes",
5008 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5009 .read_u64 = mem_cgroup_read_u64,
5012 .name = "kmem.tcp.failcnt",
5013 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5014 .write = mem_cgroup_reset,
5015 .read_u64 = mem_cgroup_read_u64,
5018 .name = "kmem.tcp.max_usage_in_bytes",
5019 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5020 .write = mem_cgroup_reset,
5021 .read_u64 = mem_cgroup_read_u64,
5023 { }, /* terminate */
5027 * Private memory cgroup IDR
5029 * Swap-out records and page cache shadow entries need to store memcg
5030 * references in constrained space, so we maintain an ID space that is
5031 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5032 * memory-controlled cgroups to 64k.
5034 * However, there usually are many references to the offline CSS after
5035 * the cgroup has been destroyed, such as page cache or reclaimable
5036 * slab objects, that don't need to hang on to the ID. We want to keep
5037 * those dead CSS from occupying IDs, or we might quickly exhaust the
5038 * relatively small ID space and prevent the creation of new cgroups
5039 * even when there are much fewer than 64k cgroups - possibly none.
5041 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5042 * be freed and recycled when it's no longer needed, which is usually
5043 * when the CSS is offlined.
5045 * The only exception to that are records of swapped out tmpfs/shmem
5046 * pages that need to be attributed to live ancestors on swapin. But
5047 * those references are manageable from userspace.
5050 static DEFINE_IDR(mem_cgroup_idr);
5052 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5054 if (memcg->id.id > 0) {
5055 idr_remove(&mem_cgroup_idr, memcg->id.id);
5060 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5063 refcount_add(n, &memcg->id.ref);
5066 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5068 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5069 mem_cgroup_id_remove(memcg);
5071 /* Memcg ID pins CSS */
5072 css_put(&memcg->css);
5076 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5078 mem_cgroup_id_put_many(memcg, 1);
5082 * mem_cgroup_from_id - look up a memcg from a memcg id
5083 * @id: the memcg id to look up
5085 * Caller must hold rcu_read_lock().
5087 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5089 WARN_ON_ONCE(!rcu_read_lock_held());
5090 return idr_find(&mem_cgroup_idr, id);
5093 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5095 struct mem_cgroup_per_node *pn;
5098 * This routine is called against possible nodes.
5099 * But it's BUG to call kmalloc() against offline node.
5101 * TODO: this routine can waste much memory for nodes which will
5102 * never be onlined. It's better to use memory hotplug callback
5105 if (!node_state(node, N_NORMAL_MEMORY))
5107 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5111 pn->lruvec_stats_percpu = alloc_percpu_gfp(struct lruvec_stats_percpu,
5112 GFP_KERNEL_ACCOUNT);
5113 if (!pn->lruvec_stats_percpu) {
5118 lruvec_init(&pn->lruvec);
5119 pn->usage_in_excess = 0;
5120 pn->on_tree = false;
5123 memcg->nodeinfo[node] = pn;
5127 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5129 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5134 free_percpu(pn->lruvec_stats_percpu);
5138 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5143 free_mem_cgroup_per_node_info(memcg, node);
5144 free_percpu(memcg->vmstats_percpu);
5148 static void mem_cgroup_free(struct mem_cgroup *memcg)
5150 memcg_wb_domain_exit(memcg);
5151 __mem_cgroup_free(memcg);
5154 static struct mem_cgroup *mem_cgroup_alloc(void)
5156 struct mem_cgroup *memcg;
5159 int __maybe_unused i;
5160 long error = -ENOMEM;
5162 size = sizeof(struct mem_cgroup);
5163 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5165 memcg = kzalloc(size, GFP_KERNEL);
5167 return ERR_PTR(error);
5169 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5170 1, MEM_CGROUP_ID_MAX,
5172 if (memcg->id.id < 0) {
5173 error = memcg->id.id;
5177 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5178 GFP_KERNEL_ACCOUNT);
5179 if (!memcg->vmstats_percpu)
5183 if (alloc_mem_cgroup_per_node_info(memcg, node))
5186 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5189 INIT_WORK(&memcg->high_work, high_work_func);
5190 INIT_LIST_HEAD(&memcg->oom_notify);
5191 mutex_init(&memcg->thresholds_lock);
5192 spin_lock_init(&memcg->move_lock);
5193 vmpressure_init(&memcg->vmpressure);
5194 INIT_LIST_HEAD(&memcg->event_list);
5195 spin_lock_init(&memcg->event_list_lock);
5196 memcg->socket_pressure = jiffies;
5197 #ifdef CONFIG_MEMCG_KMEM
5198 memcg->kmemcg_id = -1;
5199 INIT_LIST_HEAD(&memcg->objcg_list);
5201 #ifdef CONFIG_CGROUP_WRITEBACK
5202 INIT_LIST_HEAD(&memcg->cgwb_list);
5203 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5204 memcg->cgwb_frn[i].done =
5205 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5207 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5208 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5209 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5210 memcg->deferred_split_queue.split_queue_len = 0;
5212 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5215 mem_cgroup_id_remove(memcg);
5216 __mem_cgroup_free(memcg);
5217 return ERR_PTR(error);
5220 static struct cgroup_subsys_state * __ref
5221 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5223 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5224 struct mem_cgroup *memcg, *old_memcg;
5225 long error = -ENOMEM;
5227 old_memcg = set_active_memcg(parent);
5228 memcg = mem_cgroup_alloc();
5229 set_active_memcg(old_memcg);
5231 return ERR_CAST(memcg);
5233 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5234 memcg->soft_limit = PAGE_COUNTER_MAX;
5235 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5237 memcg->swappiness = mem_cgroup_swappiness(parent);
5238 memcg->oom_kill_disable = parent->oom_kill_disable;
5240 page_counter_init(&memcg->memory, &parent->memory);
5241 page_counter_init(&memcg->swap, &parent->swap);
5242 page_counter_init(&memcg->kmem, &parent->kmem);
5243 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5245 page_counter_init(&memcg->memory, NULL);
5246 page_counter_init(&memcg->swap, NULL);
5247 page_counter_init(&memcg->kmem, NULL);
5248 page_counter_init(&memcg->tcpmem, NULL);
5250 root_mem_cgroup = memcg;
5254 /* The following stuff does not apply to the root */
5255 error = memcg_online_kmem(memcg);
5259 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5260 static_branch_inc(&memcg_sockets_enabled_key);
5264 mem_cgroup_id_remove(memcg);
5265 mem_cgroup_free(memcg);
5266 return ERR_PTR(error);
5269 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5271 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5274 * A memcg must be visible for expand_shrinker_info()
5275 * by the time the maps are allocated. So, we allocate maps
5276 * here, when for_each_mem_cgroup() can't skip it.
5278 if (alloc_shrinker_info(memcg)) {
5279 mem_cgroup_id_remove(memcg);
5283 /* Online state pins memcg ID, memcg ID pins CSS */
5284 refcount_set(&memcg->id.ref, 1);
5287 if (unlikely(mem_cgroup_is_root(memcg)))
5288 queue_delayed_work(system_unbound_wq, &stats_flush_dwork,
5293 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5295 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5296 struct mem_cgroup_event *event, *tmp;
5299 * Unregister events and notify userspace.
5300 * Notify userspace about cgroup removing only after rmdir of cgroup
5301 * directory to avoid race between userspace and kernelspace.
5303 spin_lock_irq(&memcg->event_list_lock);
5304 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5305 list_del_init(&event->list);
5306 schedule_work(&event->remove);
5308 spin_unlock_irq(&memcg->event_list_lock);
5310 page_counter_set_min(&memcg->memory, 0);
5311 page_counter_set_low(&memcg->memory, 0);
5313 memcg_offline_kmem(memcg);
5314 reparent_shrinker_deferred(memcg);
5315 wb_memcg_offline(memcg);
5317 drain_all_stock(memcg);
5319 mem_cgroup_id_put(memcg);
5322 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5324 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5326 invalidate_reclaim_iterators(memcg);
5329 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5331 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5332 int __maybe_unused i;
5334 #ifdef CONFIG_CGROUP_WRITEBACK
5335 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5336 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5338 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5339 static_branch_dec(&memcg_sockets_enabled_key);
5341 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5342 static_branch_dec(&memcg_sockets_enabled_key);
5344 vmpressure_cleanup(&memcg->vmpressure);
5345 cancel_work_sync(&memcg->high_work);
5346 mem_cgroup_remove_from_trees(memcg);
5347 free_shrinker_info(memcg);
5348 memcg_free_kmem(memcg);
5349 mem_cgroup_free(memcg);
5353 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5354 * @css: the target css
5356 * Reset the states of the mem_cgroup associated with @css. This is
5357 * invoked when the userland requests disabling on the default hierarchy
5358 * but the memcg is pinned through dependency. The memcg should stop
5359 * applying policies and should revert to the vanilla state as it may be
5360 * made visible again.
5362 * The current implementation only resets the essential configurations.
5363 * This needs to be expanded to cover all the visible parts.
5365 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5367 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5369 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5370 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5371 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5372 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5373 page_counter_set_min(&memcg->memory, 0);
5374 page_counter_set_low(&memcg->memory, 0);
5375 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5376 memcg->soft_limit = PAGE_COUNTER_MAX;
5377 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5378 memcg_wb_domain_size_changed(memcg);
5381 static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state *css, int cpu)
5383 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5384 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5385 struct memcg_vmstats_percpu *statc;
5389 statc = per_cpu_ptr(memcg->vmstats_percpu, cpu);
5391 for (i = 0; i < MEMCG_NR_STAT; i++) {
5393 * Collect the aggregated propagation counts of groups
5394 * below us. We're in a per-cpu loop here and this is
5395 * a global counter, so the first cycle will get them.
5397 delta = memcg->vmstats.state_pending[i];
5399 memcg->vmstats.state_pending[i] = 0;
5401 /* Add CPU changes on this level since the last flush */
5402 v = READ_ONCE(statc->state[i]);
5403 if (v != statc->state_prev[i]) {
5404 delta += v - statc->state_prev[i];
5405 statc->state_prev[i] = v;
5411 /* Aggregate counts on this level and propagate upwards */
5412 memcg->vmstats.state[i] += delta;
5414 parent->vmstats.state_pending[i] += delta;
5417 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
5418 delta = memcg->vmstats.events_pending[i];
5420 memcg->vmstats.events_pending[i] = 0;
5422 v = READ_ONCE(statc->events[i]);
5423 if (v != statc->events_prev[i]) {
5424 delta += v - statc->events_prev[i];
5425 statc->events_prev[i] = v;
5431 memcg->vmstats.events[i] += delta;
5433 parent->vmstats.events_pending[i] += delta;
5436 for_each_node_state(nid, N_MEMORY) {
5437 struct mem_cgroup_per_node *pn = memcg->nodeinfo[nid];
5438 struct mem_cgroup_per_node *ppn = NULL;
5439 struct lruvec_stats_percpu *lstatc;
5442 ppn = parent->nodeinfo[nid];
5444 lstatc = per_cpu_ptr(pn->lruvec_stats_percpu, cpu);
5446 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
5447 delta = pn->lruvec_stats.state_pending[i];
5449 pn->lruvec_stats.state_pending[i] = 0;
5451 v = READ_ONCE(lstatc->state[i]);
5452 if (v != lstatc->state_prev[i]) {
5453 delta += v - lstatc->state_prev[i];
5454 lstatc->state_prev[i] = v;
5460 pn->lruvec_stats.state[i] += delta;
5462 ppn->lruvec_stats.state_pending[i] += delta;
5468 /* Handlers for move charge at task migration. */
5469 static int mem_cgroup_do_precharge(unsigned long count)
5473 /* Try a single bulk charge without reclaim first, kswapd may wake */
5474 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5476 mc.precharge += count;
5480 /* Try charges one by one with reclaim, but do not retry */
5482 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5496 enum mc_target_type {
5503 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5504 unsigned long addr, pte_t ptent)
5506 struct page *page = vm_normal_page(vma, addr, ptent);
5508 if (!page || !page_mapped(page))
5510 if (PageAnon(page)) {
5511 if (!(mc.flags & MOVE_ANON))
5514 if (!(mc.flags & MOVE_FILE))
5517 if (!get_page_unless_zero(page))
5523 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5524 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5525 pte_t ptent, swp_entry_t *entry)
5527 struct page *page = NULL;
5528 swp_entry_t ent = pte_to_swp_entry(ptent);
5530 if (!(mc.flags & MOVE_ANON))
5534 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5535 * a device and because they are not accessible by CPU they are store
5536 * as special swap entry in the CPU page table.
5538 if (is_device_private_entry(ent)) {
5539 page = pfn_swap_entry_to_page(ent);
5541 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5542 * a refcount of 1 when free (unlike normal page)
5544 if (!page_ref_add_unless(page, 1, 1))
5549 if (non_swap_entry(ent))
5553 * Because lookup_swap_cache() updates some statistics counter,
5554 * we call find_get_page() with swapper_space directly.
5556 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5557 entry->val = ent.val;
5562 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5563 pte_t ptent, swp_entry_t *entry)
5569 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5570 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5572 if (!vma->vm_file) /* anonymous vma */
5574 if (!(mc.flags & MOVE_FILE))
5577 /* page is moved even if it's not RSS of this task(page-faulted). */
5578 /* shmem/tmpfs may report page out on swap: account for that too. */
5579 return find_get_incore_page(vma->vm_file->f_mapping,
5580 linear_page_index(vma, addr));
5584 * mem_cgroup_move_account - move account of the page
5586 * @compound: charge the page as compound or small page
5587 * @from: mem_cgroup which the page is moved from.
5588 * @to: mem_cgroup which the page is moved to. @from != @to.
5590 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5592 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5595 static int mem_cgroup_move_account(struct page *page,
5597 struct mem_cgroup *from,
5598 struct mem_cgroup *to)
5600 struct lruvec *from_vec, *to_vec;
5601 struct pglist_data *pgdat;
5602 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5605 VM_BUG_ON(from == to);
5606 VM_BUG_ON_PAGE(PageLRU(page), page);
5607 VM_BUG_ON(compound && !PageTransHuge(page));
5610 * Prevent mem_cgroup_migrate() from looking at
5611 * page's memory cgroup of its source page while we change it.
5614 if (!trylock_page(page))
5618 if (page_memcg(page) != from)
5621 pgdat = page_pgdat(page);
5622 from_vec = mem_cgroup_lruvec(from, pgdat);
5623 to_vec = mem_cgroup_lruvec(to, pgdat);
5625 lock_page_memcg(page);
5627 if (PageAnon(page)) {
5628 if (page_mapped(page)) {
5629 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5630 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5631 if (PageTransHuge(page)) {
5632 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5634 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5639 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5640 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5642 if (PageSwapBacked(page)) {
5643 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5644 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5647 if (page_mapped(page)) {
5648 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5649 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5652 if (PageDirty(page)) {
5653 struct address_space *mapping = page_mapping(page);
5655 if (mapping_can_writeback(mapping)) {
5656 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5658 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5664 if (PageWriteback(page)) {
5665 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5666 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5670 * All state has been migrated, let's switch to the new memcg.
5672 * It is safe to change page's memcg here because the page
5673 * is referenced, charged, isolated, and locked: we can't race
5674 * with (un)charging, migration, LRU putback, or anything else
5675 * that would rely on a stable page's memory cgroup.
5677 * Note that lock_page_memcg is a memcg lock, not a page lock,
5678 * to save space. As soon as we switch page's memory cgroup to a
5679 * new memcg that isn't locked, the above state can change
5680 * concurrently again. Make sure we're truly done with it.
5685 css_put(&from->css);
5687 page->memcg_data = (unsigned long)to;
5689 __unlock_page_memcg(from);
5693 local_irq_disable();
5694 mem_cgroup_charge_statistics(to, page, nr_pages);
5695 memcg_check_events(to, page);
5696 mem_cgroup_charge_statistics(from, page, -nr_pages);
5697 memcg_check_events(from, page);
5706 * get_mctgt_type - get target type of moving charge
5707 * @vma: the vma the pte to be checked belongs
5708 * @addr: the address corresponding to the pte to be checked
5709 * @ptent: the pte to be checked
5710 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5713 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5714 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5715 * move charge. if @target is not NULL, the page is stored in target->page
5716 * with extra refcnt got(Callers should handle it).
5717 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5718 * target for charge migration. if @target is not NULL, the entry is stored
5720 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5721 * (so ZONE_DEVICE page and thus not on the lru).
5722 * For now we such page is charge like a regular page would be as for all
5723 * intent and purposes it is just special memory taking the place of a
5726 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5728 * Called with pte lock held.
5731 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5732 unsigned long addr, pte_t ptent, union mc_target *target)
5734 struct page *page = NULL;
5735 enum mc_target_type ret = MC_TARGET_NONE;
5736 swp_entry_t ent = { .val = 0 };
5738 if (pte_present(ptent))
5739 page = mc_handle_present_pte(vma, addr, ptent);
5740 else if (is_swap_pte(ptent))
5741 page = mc_handle_swap_pte(vma, ptent, &ent);
5742 else if (pte_none(ptent))
5743 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5745 if (!page && !ent.val)
5749 * Do only loose check w/o serialization.
5750 * mem_cgroup_move_account() checks the page is valid or
5751 * not under LRU exclusion.
5753 if (page_memcg(page) == mc.from) {
5754 ret = MC_TARGET_PAGE;
5755 if (is_device_private_page(page))
5756 ret = MC_TARGET_DEVICE;
5758 target->page = page;
5760 if (!ret || !target)
5764 * There is a swap entry and a page doesn't exist or isn't charged.
5765 * But we cannot move a tail-page in a THP.
5767 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5768 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5769 ret = MC_TARGET_SWAP;
5776 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5778 * We don't consider PMD mapped swapping or file mapped pages because THP does
5779 * not support them for now.
5780 * Caller should make sure that pmd_trans_huge(pmd) is true.
5782 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5783 unsigned long addr, pmd_t pmd, union mc_target *target)
5785 struct page *page = NULL;
5786 enum mc_target_type ret = MC_TARGET_NONE;
5788 if (unlikely(is_swap_pmd(pmd))) {
5789 VM_BUG_ON(thp_migration_supported() &&
5790 !is_pmd_migration_entry(pmd));
5793 page = pmd_page(pmd);
5794 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5795 if (!(mc.flags & MOVE_ANON))
5797 if (page_memcg(page) == mc.from) {
5798 ret = MC_TARGET_PAGE;
5801 target->page = page;
5807 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5808 unsigned long addr, pmd_t pmd, union mc_target *target)
5810 return MC_TARGET_NONE;
5814 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5815 unsigned long addr, unsigned long end,
5816 struct mm_walk *walk)
5818 struct vm_area_struct *vma = walk->vma;
5822 ptl = pmd_trans_huge_lock(pmd, vma);
5825 * Note their can not be MC_TARGET_DEVICE for now as we do not
5826 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5827 * this might change.
5829 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5830 mc.precharge += HPAGE_PMD_NR;
5835 if (pmd_trans_unstable(pmd))
5837 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5838 for (; addr != end; pte++, addr += PAGE_SIZE)
5839 if (get_mctgt_type(vma, addr, *pte, NULL))
5840 mc.precharge++; /* increment precharge temporarily */
5841 pte_unmap_unlock(pte - 1, ptl);
5847 static const struct mm_walk_ops precharge_walk_ops = {
5848 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5851 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5853 unsigned long precharge;
5856 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5857 mmap_read_unlock(mm);
5859 precharge = mc.precharge;
5865 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5867 unsigned long precharge = mem_cgroup_count_precharge(mm);
5869 VM_BUG_ON(mc.moving_task);
5870 mc.moving_task = current;
5871 return mem_cgroup_do_precharge(precharge);
5874 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5875 static void __mem_cgroup_clear_mc(void)
5877 struct mem_cgroup *from = mc.from;
5878 struct mem_cgroup *to = mc.to;
5880 /* we must uncharge all the leftover precharges from mc.to */
5882 cancel_charge(mc.to, mc.precharge);
5886 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5887 * we must uncharge here.
5889 if (mc.moved_charge) {
5890 cancel_charge(mc.from, mc.moved_charge);
5891 mc.moved_charge = 0;
5893 /* we must fixup refcnts and charges */
5894 if (mc.moved_swap) {
5895 /* uncharge swap account from the old cgroup */
5896 if (!mem_cgroup_is_root(mc.from))
5897 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5899 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5902 * we charged both to->memory and to->memsw, so we
5903 * should uncharge to->memory.
5905 if (!mem_cgroup_is_root(mc.to))
5906 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5910 memcg_oom_recover(from);
5911 memcg_oom_recover(to);
5912 wake_up_all(&mc.waitq);
5915 static void mem_cgroup_clear_mc(void)
5917 struct mm_struct *mm = mc.mm;
5920 * we must clear moving_task before waking up waiters at the end of
5923 mc.moving_task = NULL;
5924 __mem_cgroup_clear_mc();
5925 spin_lock(&mc.lock);
5929 spin_unlock(&mc.lock);
5934 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5936 struct cgroup_subsys_state *css;
5937 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5938 struct mem_cgroup *from;
5939 struct task_struct *leader, *p;
5940 struct mm_struct *mm;
5941 unsigned long move_flags;
5944 /* charge immigration isn't supported on the default hierarchy */
5945 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5949 * Multi-process migrations only happen on the default hierarchy
5950 * where charge immigration is not used. Perform charge
5951 * immigration if @tset contains a leader and whine if there are
5955 cgroup_taskset_for_each_leader(leader, css, tset) {
5958 memcg = mem_cgroup_from_css(css);
5964 * We are now committed to this value whatever it is. Changes in this
5965 * tunable will only affect upcoming migrations, not the current one.
5966 * So we need to save it, and keep it going.
5968 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5972 from = mem_cgroup_from_task(p);
5974 VM_BUG_ON(from == memcg);
5976 mm = get_task_mm(p);
5979 /* We move charges only when we move a owner of the mm */
5980 if (mm->owner == p) {
5983 VM_BUG_ON(mc.precharge);
5984 VM_BUG_ON(mc.moved_charge);
5985 VM_BUG_ON(mc.moved_swap);
5987 spin_lock(&mc.lock);
5991 mc.flags = move_flags;
5992 spin_unlock(&mc.lock);
5993 /* We set mc.moving_task later */
5995 ret = mem_cgroup_precharge_mc(mm);
5997 mem_cgroup_clear_mc();
6004 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6007 mem_cgroup_clear_mc();
6010 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6011 unsigned long addr, unsigned long end,
6012 struct mm_walk *walk)
6015 struct vm_area_struct *vma = walk->vma;
6018 enum mc_target_type target_type;
6019 union mc_target target;
6022 ptl = pmd_trans_huge_lock(pmd, vma);
6024 if (mc.precharge < HPAGE_PMD_NR) {
6028 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6029 if (target_type == MC_TARGET_PAGE) {
6031 if (!isolate_lru_page(page)) {
6032 if (!mem_cgroup_move_account(page, true,
6034 mc.precharge -= HPAGE_PMD_NR;
6035 mc.moved_charge += HPAGE_PMD_NR;
6037 putback_lru_page(page);
6040 } else if (target_type == MC_TARGET_DEVICE) {
6042 if (!mem_cgroup_move_account(page, true,
6044 mc.precharge -= HPAGE_PMD_NR;
6045 mc.moved_charge += HPAGE_PMD_NR;
6053 if (pmd_trans_unstable(pmd))
6056 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6057 for (; addr != end; addr += PAGE_SIZE) {
6058 pte_t ptent = *(pte++);
6059 bool device = false;
6065 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6066 case MC_TARGET_DEVICE:
6069 case MC_TARGET_PAGE:
6072 * We can have a part of the split pmd here. Moving it
6073 * can be done but it would be too convoluted so simply
6074 * ignore such a partial THP and keep it in original
6075 * memcg. There should be somebody mapping the head.
6077 if (PageTransCompound(page))
6079 if (!device && isolate_lru_page(page))
6081 if (!mem_cgroup_move_account(page, false,
6084 /* we uncharge from mc.from later. */
6088 putback_lru_page(page);
6089 put: /* get_mctgt_type() gets the page */
6092 case MC_TARGET_SWAP:
6094 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6096 mem_cgroup_id_get_many(mc.to, 1);
6097 /* we fixup other refcnts and charges later. */
6105 pte_unmap_unlock(pte - 1, ptl);
6110 * We have consumed all precharges we got in can_attach().
6111 * We try charge one by one, but don't do any additional
6112 * charges to mc.to if we have failed in charge once in attach()
6115 ret = mem_cgroup_do_precharge(1);
6123 static const struct mm_walk_ops charge_walk_ops = {
6124 .pmd_entry = mem_cgroup_move_charge_pte_range,
6127 static void mem_cgroup_move_charge(void)
6129 lru_add_drain_all();
6131 * Signal lock_page_memcg() to take the memcg's move_lock
6132 * while we're moving its pages to another memcg. Then wait
6133 * for already started RCU-only updates to finish.
6135 atomic_inc(&mc.from->moving_account);
6138 if (unlikely(!mmap_read_trylock(mc.mm))) {
6140 * Someone who are holding the mmap_lock might be waiting in
6141 * waitq. So we cancel all extra charges, wake up all waiters,
6142 * and retry. Because we cancel precharges, we might not be able
6143 * to move enough charges, but moving charge is a best-effort
6144 * feature anyway, so it wouldn't be a big problem.
6146 __mem_cgroup_clear_mc();
6151 * When we have consumed all precharges and failed in doing
6152 * additional charge, the page walk just aborts.
6154 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6157 mmap_read_unlock(mc.mm);
6158 atomic_dec(&mc.from->moving_account);
6161 static void mem_cgroup_move_task(void)
6164 mem_cgroup_move_charge();
6165 mem_cgroup_clear_mc();
6168 #else /* !CONFIG_MMU */
6169 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6173 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6176 static void mem_cgroup_move_task(void)
6181 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6183 if (value == PAGE_COUNTER_MAX)
6184 seq_puts(m, "max\n");
6186 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6191 static u64 memory_current_read(struct cgroup_subsys_state *css,
6194 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6196 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6199 static int memory_min_show(struct seq_file *m, void *v)
6201 return seq_puts_memcg_tunable(m,
6202 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6205 static ssize_t memory_min_write(struct kernfs_open_file *of,
6206 char *buf, size_t nbytes, loff_t off)
6208 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6212 buf = strstrip(buf);
6213 err = page_counter_memparse(buf, "max", &min);
6217 page_counter_set_min(&memcg->memory, min);
6222 static int memory_low_show(struct seq_file *m, void *v)
6224 return seq_puts_memcg_tunable(m,
6225 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6228 static ssize_t memory_low_write(struct kernfs_open_file *of,
6229 char *buf, size_t nbytes, loff_t off)
6231 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6235 buf = strstrip(buf);
6236 err = page_counter_memparse(buf, "max", &low);
6240 page_counter_set_low(&memcg->memory, low);
6245 static int memory_high_show(struct seq_file *m, void *v)
6247 return seq_puts_memcg_tunable(m,
6248 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6251 static ssize_t memory_high_write(struct kernfs_open_file *of,
6252 char *buf, size_t nbytes, loff_t off)
6254 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6255 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6256 bool drained = false;
6260 buf = strstrip(buf);
6261 err = page_counter_memparse(buf, "max", &high);
6265 page_counter_set_high(&memcg->memory, high);
6268 unsigned long nr_pages = page_counter_read(&memcg->memory);
6269 unsigned long reclaimed;
6271 if (nr_pages <= high)
6274 if (signal_pending(current))
6278 drain_all_stock(memcg);
6283 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6286 if (!reclaimed && !nr_retries--)
6290 memcg_wb_domain_size_changed(memcg);
6294 static int memory_max_show(struct seq_file *m, void *v)
6296 return seq_puts_memcg_tunable(m,
6297 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6300 static ssize_t memory_max_write(struct kernfs_open_file *of,
6301 char *buf, size_t nbytes, loff_t off)
6303 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6304 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6305 bool drained = false;
6309 buf = strstrip(buf);
6310 err = page_counter_memparse(buf, "max", &max);
6314 xchg(&memcg->memory.max, max);
6317 unsigned long nr_pages = page_counter_read(&memcg->memory);
6319 if (nr_pages <= max)
6322 if (signal_pending(current))
6326 drain_all_stock(memcg);
6332 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6338 memcg_memory_event(memcg, MEMCG_OOM);
6339 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6343 memcg_wb_domain_size_changed(memcg);
6347 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6349 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6350 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6351 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6352 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6353 seq_printf(m, "oom_kill %lu\n",
6354 atomic_long_read(&events[MEMCG_OOM_KILL]));
6357 static int memory_events_show(struct seq_file *m, void *v)
6359 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6361 __memory_events_show(m, memcg->memory_events);
6365 static int memory_events_local_show(struct seq_file *m, void *v)
6367 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6369 __memory_events_show(m, memcg->memory_events_local);
6373 static int memory_stat_show(struct seq_file *m, void *v)
6375 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6378 buf = memory_stat_format(memcg);
6387 static inline unsigned long lruvec_page_state_output(struct lruvec *lruvec,
6390 return lruvec_page_state(lruvec, item) * memcg_page_state_unit(item);
6393 static int memory_numa_stat_show(struct seq_file *m, void *v)
6396 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6398 mem_cgroup_flush_stats();
6400 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6403 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6406 seq_printf(m, "%s", memory_stats[i].name);
6407 for_each_node_state(nid, N_MEMORY) {
6409 struct lruvec *lruvec;
6411 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6412 size = lruvec_page_state_output(lruvec,
6413 memory_stats[i].idx);
6414 seq_printf(m, " N%d=%llu", nid, size);
6423 static int memory_oom_group_show(struct seq_file *m, void *v)
6425 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6427 seq_printf(m, "%d\n", memcg->oom_group);
6432 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6433 char *buf, size_t nbytes, loff_t off)
6435 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6438 buf = strstrip(buf);
6442 ret = kstrtoint(buf, 0, &oom_group);
6446 if (oom_group != 0 && oom_group != 1)
6449 memcg->oom_group = oom_group;
6454 static struct cftype memory_files[] = {
6457 .flags = CFTYPE_NOT_ON_ROOT,
6458 .read_u64 = memory_current_read,
6462 .flags = CFTYPE_NOT_ON_ROOT,
6463 .seq_show = memory_min_show,
6464 .write = memory_min_write,
6468 .flags = CFTYPE_NOT_ON_ROOT,
6469 .seq_show = memory_low_show,
6470 .write = memory_low_write,
6474 .flags = CFTYPE_NOT_ON_ROOT,
6475 .seq_show = memory_high_show,
6476 .write = memory_high_write,
6480 .flags = CFTYPE_NOT_ON_ROOT,
6481 .seq_show = memory_max_show,
6482 .write = memory_max_write,
6486 .flags = CFTYPE_NOT_ON_ROOT,
6487 .file_offset = offsetof(struct mem_cgroup, events_file),
6488 .seq_show = memory_events_show,
6491 .name = "events.local",
6492 .flags = CFTYPE_NOT_ON_ROOT,
6493 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6494 .seq_show = memory_events_local_show,
6498 .seq_show = memory_stat_show,
6502 .name = "numa_stat",
6503 .seq_show = memory_numa_stat_show,
6507 .name = "oom.group",
6508 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6509 .seq_show = memory_oom_group_show,
6510 .write = memory_oom_group_write,
6515 struct cgroup_subsys memory_cgrp_subsys = {
6516 .css_alloc = mem_cgroup_css_alloc,
6517 .css_online = mem_cgroup_css_online,
6518 .css_offline = mem_cgroup_css_offline,
6519 .css_released = mem_cgroup_css_released,
6520 .css_free = mem_cgroup_css_free,
6521 .css_reset = mem_cgroup_css_reset,
6522 .css_rstat_flush = mem_cgroup_css_rstat_flush,
6523 .can_attach = mem_cgroup_can_attach,
6524 .cancel_attach = mem_cgroup_cancel_attach,
6525 .post_attach = mem_cgroup_move_task,
6526 .dfl_cftypes = memory_files,
6527 .legacy_cftypes = mem_cgroup_legacy_files,
6532 * This function calculates an individual cgroup's effective
6533 * protection which is derived from its own memory.min/low, its
6534 * parent's and siblings' settings, as well as the actual memory
6535 * distribution in the tree.
6537 * The following rules apply to the effective protection values:
6539 * 1. At the first level of reclaim, effective protection is equal to
6540 * the declared protection in memory.min and memory.low.
6542 * 2. To enable safe delegation of the protection configuration, at
6543 * subsequent levels the effective protection is capped to the
6544 * parent's effective protection.
6546 * 3. To make complex and dynamic subtrees easier to configure, the
6547 * user is allowed to overcommit the declared protection at a given
6548 * level. If that is the case, the parent's effective protection is
6549 * distributed to the children in proportion to how much protection
6550 * they have declared and how much of it they are utilizing.
6552 * This makes distribution proportional, but also work-conserving:
6553 * if one cgroup claims much more protection than it uses memory,
6554 * the unused remainder is available to its siblings.
6556 * 4. Conversely, when the declared protection is undercommitted at a
6557 * given level, the distribution of the larger parental protection
6558 * budget is NOT proportional. A cgroup's protection from a sibling
6559 * is capped to its own memory.min/low setting.
6561 * 5. However, to allow protecting recursive subtrees from each other
6562 * without having to declare each individual cgroup's fixed share
6563 * of the ancestor's claim to protection, any unutilized -
6564 * "floating" - protection from up the tree is distributed in
6565 * proportion to each cgroup's *usage*. This makes the protection
6566 * neutral wrt sibling cgroups and lets them compete freely over
6567 * the shared parental protection budget, but it protects the
6568 * subtree as a whole from neighboring subtrees.
6570 * Note that 4. and 5. are not in conflict: 4. is about protecting
6571 * against immediate siblings whereas 5. is about protecting against
6572 * neighboring subtrees.
6574 static unsigned long effective_protection(unsigned long usage,
6575 unsigned long parent_usage,
6576 unsigned long setting,
6577 unsigned long parent_effective,
6578 unsigned long siblings_protected)
6580 unsigned long protected;
6583 protected = min(usage, setting);
6585 * If all cgroups at this level combined claim and use more
6586 * protection then what the parent affords them, distribute
6587 * shares in proportion to utilization.
6589 * We are using actual utilization rather than the statically
6590 * claimed protection in order to be work-conserving: claimed
6591 * but unused protection is available to siblings that would
6592 * otherwise get a smaller chunk than what they claimed.
6594 if (siblings_protected > parent_effective)
6595 return protected * parent_effective / siblings_protected;
6598 * Ok, utilized protection of all children is within what the
6599 * parent affords them, so we know whatever this child claims
6600 * and utilizes is effectively protected.
6602 * If there is unprotected usage beyond this value, reclaim
6603 * will apply pressure in proportion to that amount.
6605 * If there is unutilized protection, the cgroup will be fully
6606 * shielded from reclaim, but we do return a smaller value for
6607 * protection than what the group could enjoy in theory. This
6608 * is okay. With the overcommit distribution above, effective
6609 * protection is always dependent on how memory is actually
6610 * consumed among the siblings anyway.
6615 * If the children aren't claiming (all of) the protection
6616 * afforded to them by the parent, distribute the remainder in
6617 * proportion to the (unprotected) memory of each cgroup. That
6618 * way, cgroups that aren't explicitly prioritized wrt each
6619 * other compete freely over the allowance, but they are
6620 * collectively protected from neighboring trees.
6622 * We're using unprotected memory for the weight so that if
6623 * some cgroups DO claim explicit protection, we don't protect
6624 * the same bytes twice.
6626 * Check both usage and parent_usage against the respective
6627 * protected values. One should imply the other, but they
6628 * aren't read atomically - make sure the division is sane.
6630 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6632 if (parent_effective > siblings_protected &&
6633 parent_usage > siblings_protected &&
6634 usage > protected) {
6635 unsigned long unclaimed;
6637 unclaimed = parent_effective - siblings_protected;
6638 unclaimed *= usage - protected;
6639 unclaimed /= parent_usage - siblings_protected;
6648 * mem_cgroup_calculate_protection - check if memory consumption is in the normal range
6649 * @root: the top ancestor of the sub-tree being checked
6650 * @memcg: the memory cgroup to check
6652 * WARNING: This function is not stateless! It can only be used as part
6653 * of a top-down tree iteration, not for isolated queries.
6655 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6656 struct mem_cgroup *memcg)
6658 unsigned long usage, parent_usage;
6659 struct mem_cgroup *parent;
6661 if (mem_cgroup_disabled())
6665 root = root_mem_cgroup;
6668 * Effective values of the reclaim targets are ignored so they
6669 * can be stale. Have a look at mem_cgroup_protection for more
6671 * TODO: calculation should be more robust so that we do not need
6672 * that special casing.
6677 usage = page_counter_read(&memcg->memory);
6681 parent = parent_mem_cgroup(memcg);
6682 /* No parent means a non-hierarchical mode on v1 memcg */
6686 if (parent == root) {
6687 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6688 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6692 parent_usage = page_counter_read(&parent->memory);
6694 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6695 READ_ONCE(memcg->memory.min),
6696 READ_ONCE(parent->memory.emin),
6697 atomic_long_read(&parent->memory.children_min_usage)));
6699 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6700 READ_ONCE(memcg->memory.low),
6701 READ_ONCE(parent->memory.elow),
6702 atomic_long_read(&parent->memory.children_low_usage)));
6705 static int charge_memcg(struct page *page, struct mem_cgroup *memcg, gfp_t gfp)
6707 unsigned int nr_pages = thp_nr_pages(page);
6710 ret = try_charge(memcg, gfp, nr_pages);
6714 css_get(&memcg->css);
6715 commit_charge(page, memcg);
6717 local_irq_disable();
6718 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6719 memcg_check_events(memcg, page);
6726 * __mem_cgroup_charge - charge a newly allocated page to a cgroup
6727 * @page: page to charge
6728 * @mm: mm context of the victim
6729 * @gfp_mask: reclaim mode
6731 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6732 * pages according to @gfp_mask if necessary. if @mm is NULL, try to
6733 * charge to the active memcg.
6735 * Do not use this for pages allocated for swapin.
6737 * Returns 0 on success. Otherwise, an error code is returned.
6739 int __mem_cgroup_charge(struct page *page, struct mm_struct *mm,
6742 struct mem_cgroup *memcg;
6745 memcg = get_mem_cgroup_from_mm(mm);
6746 ret = charge_memcg(page, memcg, gfp_mask);
6747 css_put(&memcg->css);
6753 * mem_cgroup_swapin_charge_page - charge a newly allocated page for swapin
6754 * @page: page to charge
6755 * @mm: mm context of the victim
6756 * @gfp: reclaim mode
6757 * @entry: swap entry for which the page is allocated
6759 * This function charges a page allocated for swapin. Please call this before
6760 * adding the page to the swapcache.
6762 * Returns 0 on success. Otherwise, an error code is returned.
6764 int mem_cgroup_swapin_charge_page(struct page *page, struct mm_struct *mm,
6765 gfp_t gfp, swp_entry_t entry)
6767 struct mem_cgroup *memcg;
6771 if (mem_cgroup_disabled())
6774 id = lookup_swap_cgroup_id(entry);
6776 memcg = mem_cgroup_from_id(id);
6777 if (!memcg || !css_tryget_online(&memcg->css))
6778 memcg = get_mem_cgroup_from_mm(mm);
6781 ret = charge_memcg(page, memcg, gfp);
6783 css_put(&memcg->css);
6788 * mem_cgroup_swapin_uncharge_swap - uncharge swap slot
6789 * @entry: swap entry for which the page is charged
6791 * Call this function after successfully adding the charged page to swapcache.
6793 * Note: This function assumes the page for which swap slot is being uncharged
6796 void mem_cgroup_swapin_uncharge_swap(swp_entry_t entry)
6799 * Cgroup1's unified memory+swap counter has been charged with the
6800 * new swapcache page, finish the transfer by uncharging the swap
6801 * slot. The swap slot would also get uncharged when it dies, but
6802 * it can stick around indefinitely and we'd count the page twice
6805 * Cgroup2 has separate resource counters for memory and swap,
6806 * so this is a non-issue here. Memory and swap charge lifetimes
6807 * correspond 1:1 to page and swap slot lifetimes: we charge the
6808 * page to memory here, and uncharge swap when the slot is freed.
6810 if (!mem_cgroup_disabled() && do_memsw_account()) {
6812 * The swap entry might not get freed for a long time,
6813 * let's not wait for it. The page already received a
6814 * memory+swap charge, drop the swap entry duplicate.
6816 mem_cgroup_uncharge_swap(entry, 1);
6820 struct uncharge_gather {
6821 struct mem_cgroup *memcg;
6822 unsigned long nr_memory;
6823 unsigned long pgpgout;
6824 unsigned long nr_kmem;
6825 struct page *dummy_page;
6828 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6830 memset(ug, 0, sizeof(*ug));
6833 static void uncharge_batch(const struct uncharge_gather *ug)
6835 unsigned long flags;
6837 if (ug->nr_memory) {
6838 page_counter_uncharge(&ug->memcg->memory, ug->nr_memory);
6839 if (do_memsw_account())
6840 page_counter_uncharge(&ug->memcg->memsw, ug->nr_memory);
6841 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6842 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6843 memcg_oom_recover(ug->memcg);
6846 local_irq_save(flags);
6847 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6848 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_memory);
6849 memcg_check_events(ug->memcg, ug->dummy_page);
6850 local_irq_restore(flags);
6852 /* drop reference from uncharge_page */
6853 css_put(&ug->memcg->css);
6856 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6858 unsigned long nr_pages;
6859 struct mem_cgroup *memcg;
6860 struct obj_cgroup *objcg;
6861 bool use_objcg = PageMemcgKmem(page);
6863 VM_BUG_ON_PAGE(PageLRU(page), page);
6866 * Nobody should be changing or seriously looking at
6867 * page memcg or objcg at this point, we have fully
6868 * exclusive access to the page.
6871 objcg = __page_objcg(page);
6873 * This get matches the put at the end of the function and
6874 * kmem pages do not hold memcg references anymore.
6876 memcg = get_mem_cgroup_from_objcg(objcg);
6878 memcg = __page_memcg(page);
6884 if (ug->memcg != memcg) {
6887 uncharge_gather_clear(ug);
6890 ug->dummy_page = page;
6892 /* pairs with css_put in uncharge_batch */
6893 css_get(&memcg->css);
6896 nr_pages = compound_nr(page);
6899 ug->nr_memory += nr_pages;
6900 ug->nr_kmem += nr_pages;
6902 page->memcg_data = 0;
6903 obj_cgroup_put(objcg);
6905 /* LRU pages aren't accounted at the root level */
6906 if (!mem_cgroup_is_root(memcg))
6907 ug->nr_memory += nr_pages;
6910 page->memcg_data = 0;
6913 css_put(&memcg->css);
6917 * __mem_cgroup_uncharge - uncharge a page
6918 * @page: page to uncharge
6920 * Uncharge a page previously charged with __mem_cgroup_charge().
6922 void __mem_cgroup_uncharge(struct page *page)
6924 struct uncharge_gather ug;
6926 /* Don't touch page->lru of any random page, pre-check: */
6927 if (!page_memcg(page))
6930 uncharge_gather_clear(&ug);
6931 uncharge_page(page, &ug);
6932 uncharge_batch(&ug);
6936 * __mem_cgroup_uncharge_list - uncharge a list of page
6937 * @page_list: list of pages to uncharge
6939 * Uncharge a list of pages previously charged with
6940 * __mem_cgroup_charge().
6942 void __mem_cgroup_uncharge_list(struct list_head *page_list)
6944 struct uncharge_gather ug;
6947 uncharge_gather_clear(&ug);
6948 list_for_each_entry(page, page_list, lru)
6949 uncharge_page(page, &ug);
6951 uncharge_batch(&ug);
6955 * mem_cgroup_migrate - charge a page's replacement
6956 * @oldpage: currently circulating page
6957 * @newpage: replacement page
6959 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6960 * be uncharged upon free.
6962 * Both pages must be locked, @newpage->mapping must be set up.
6964 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6966 struct mem_cgroup *memcg;
6967 unsigned int nr_pages;
6968 unsigned long flags;
6970 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6971 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6972 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6973 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6976 if (mem_cgroup_disabled())
6979 /* Page cache replacement: new page already charged? */
6980 if (page_memcg(newpage))
6983 memcg = page_memcg(oldpage);
6984 VM_WARN_ON_ONCE_PAGE(!memcg, oldpage);
6988 /* Force-charge the new page. The old one will be freed soon */
6989 nr_pages = thp_nr_pages(newpage);
6991 if (!mem_cgroup_is_root(memcg)) {
6992 page_counter_charge(&memcg->memory, nr_pages);
6993 if (do_memsw_account())
6994 page_counter_charge(&memcg->memsw, nr_pages);
6997 css_get(&memcg->css);
6998 commit_charge(newpage, memcg);
7000 local_irq_save(flags);
7001 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
7002 memcg_check_events(memcg, newpage);
7003 local_irq_restore(flags);
7006 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
7007 EXPORT_SYMBOL(memcg_sockets_enabled_key);
7009 void mem_cgroup_sk_alloc(struct sock *sk)
7011 struct mem_cgroup *memcg;
7013 if (!mem_cgroup_sockets_enabled)
7016 /* Do not associate the sock with unrelated interrupted task's memcg. */
7021 memcg = mem_cgroup_from_task(current);
7022 if (memcg == root_mem_cgroup)
7024 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
7026 if (css_tryget(&memcg->css))
7027 sk->sk_memcg = memcg;
7032 void mem_cgroup_sk_free(struct sock *sk)
7035 css_put(&sk->sk_memcg->css);
7039 * mem_cgroup_charge_skmem - charge socket memory
7040 * @memcg: memcg to charge
7041 * @nr_pages: number of pages to charge
7042 * @gfp_mask: reclaim mode
7044 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7045 * @memcg's configured limit, %false if it doesn't.
7047 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages,
7050 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7051 struct page_counter *fail;
7053 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7054 memcg->tcpmem_pressure = 0;
7057 memcg->tcpmem_pressure = 1;
7058 if (gfp_mask & __GFP_NOFAIL) {
7059 page_counter_charge(&memcg->tcpmem, nr_pages);
7065 if (try_charge(memcg, gfp_mask, nr_pages) == 0) {
7066 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7074 * mem_cgroup_uncharge_skmem - uncharge socket memory
7075 * @memcg: memcg to uncharge
7076 * @nr_pages: number of pages to uncharge
7078 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7080 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7081 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7085 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7087 refill_stock(memcg, nr_pages);
7090 static int __init cgroup_memory(char *s)
7094 while ((token = strsep(&s, ",")) != NULL) {
7097 if (!strcmp(token, "nosocket"))
7098 cgroup_memory_nosocket = true;
7099 if (!strcmp(token, "nokmem"))
7100 cgroup_memory_nokmem = true;
7104 __setup("cgroup.memory=", cgroup_memory);
7107 * subsys_initcall() for memory controller.
7109 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7110 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7111 * basically everything that doesn't depend on a specific mem_cgroup structure
7112 * should be initialized from here.
7114 static int __init mem_cgroup_init(void)
7119 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7120 * used for per-memcg-per-cpu caching of per-node statistics. In order
7121 * to work fine, we should make sure that the overfill threshold can't
7122 * exceed S32_MAX / PAGE_SIZE.
7124 BUILD_BUG_ON(MEMCG_CHARGE_BATCH > S32_MAX / PAGE_SIZE);
7126 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7127 memcg_hotplug_cpu_dead);
7129 for_each_possible_cpu(cpu)
7130 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7133 for_each_node(node) {
7134 struct mem_cgroup_tree_per_node *rtpn;
7136 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7137 node_online(node) ? node : NUMA_NO_NODE);
7139 rtpn->rb_root = RB_ROOT;
7140 rtpn->rb_rightmost = NULL;
7141 spin_lock_init(&rtpn->lock);
7142 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7147 subsys_initcall(mem_cgroup_init);
7149 #ifdef CONFIG_MEMCG_SWAP
7150 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7152 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7154 * The root cgroup cannot be destroyed, so it's refcount must
7157 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7161 memcg = parent_mem_cgroup(memcg);
7163 memcg = root_mem_cgroup;
7169 * mem_cgroup_swapout - transfer a memsw charge to swap
7170 * @page: page whose memsw charge to transfer
7171 * @entry: swap entry to move the charge to
7173 * Transfer the memsw charge of @page to @entry.
7175 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7177 struct mem_cgroup *memcg, *swap_memcg;
7178 unsigned int nr_entries;
7179 unsigned short oldid;
7181 VM_BUG_ON_PAGE(PageLRU(page), page);
7182 VM_BUG_ON_PAGE(page_count(page), page);
7184 if (mem_cgroup_disabled())
7187 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7190 memcg = page_memcg(page);
7192 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7197 * In case the memcg owning these pages has been offlined and doesn't
7198 * have an ID allocated to it anymore, charge the closest online
7199 * ancestor for the swap instead and transfer the memory+swap charge.
7201 swap_memcg = mem_cgroup_id_get_online(memcg);
7202 nr_entries = thp_nr_pages(page);
7203 /* Get references for the tail pages, too */
7205 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7206 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7208 VM_BUG_ON_PAGE(oldid, page);
7209 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7211 page->memcg_data = 0;
7213 if (!mem_cgroup_is_root(memcg))
7214 page_counter_uncharge(&memcg->memory, nr_entries);
7216 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7217 if (!mem_cgroup_is_root(swap_memcg))
7218 page_counter_charge(&swap_memcg->memsw, nr_entries);
7219 page_counter_uncharge(&memcg->memsw, nr_entries);
7223 * Interrupts should be disabled here because the caller holds the
7224 * i_pages lock which is taken with interrupts-off. It is
7225 * important here to have the interrupts disabled because it is the
7226 * only synchronisation we have for updating the per-CPU variables.
7228 VM_BUG_ON(!irqs_disabled());
7229 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7230 memcg_check_events(memcg, page);
7232 css_put(&memcg->css);
7236 * __mem_cgroup_try_charge_swap - try charging swap space for a page
7237 * @page: page being added to swap
7238 * @entry: swap entry to charge
7240 * Try to charge @page's memcg for the swap space at @entry.
7242 * Returns 0 on success, -ENOMEM on failure.
7244 int __mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7246 unsigned int nr_pages = thp_nr_pages(page);
7247 struct page_counter *counter;
7248 struct mem_cgroup *memcg;
7249 unsigned short oldid;
7251 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7254 memcg = page_memcg(page);
7256 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7261 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7265 memcg = mem_cgroup_id_get_online(memcg);
7267 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7268 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7269 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7270 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7271 mem_cgroup_id_put(memcg);
7275 /* Get references for the tail pages, too */
7277 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7278 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7279 VM_BUG_ON_PAGE(oldid, page);
7280 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7286 * __mem_cgroup_uncharge_swap - uncharge swap space
7287 * @entry: swap entry to uncharge
7288 * @nr_pages: the amount of swap space to uncharge
7290 void __mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7292 struct mem_cgroup *memcg;
7295 id = swap_cgroup_record(entry, 0, nr_pages);
7297 memcg = mem_cgroup_from_id(id);
7299 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7300 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7301 page_counter_uncharge(&memcg->swap, nr_pages);
7303 page_counter_uncharge(&memcg->memsw, nr_pages);
7305 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7306 mem_cgroup_id_put_many(memcg, nr_pages);
7311 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7313 long nr_swap_pages = get_nr_swap_pages();
7315 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7316 return nr_swap_pages;
7317 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7318 nr_swap_pages = min_t(long, nr_swap_pages,
7319 READ_ONCE(memcg->swap.max) -
7320 page_counter_read(&memcg->swap));
7321 return nr_swap_pages;
7324 bool mem_cgroup_swap_full(struct page *page)
7326 struct mem_cgroup *memcg;
7328 VM_BUG_ON_PAGE(!PageLocked(page), page);
7332 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7335 memcg = page_memcg(page);
7339 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7340 unsigned long usage = page_counter_read(&memcg->swap);
7342 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7343 usage * 2 >= READ_ONCE(memcg->swap.max))
7350 static int __init setup_swap_account(char *s)
7352 if (!strcmp(s, "1"))
7353 cgroup_memory_noswap = false;
7354 else if (!strcmp(s, "0"))
7355 cgroup_memory_noswap = true;
7358 __setup("swapaccount=", setup_swap_account);
7360 static u64 swap_current_read(struct cgroup_subsys_state *css,
7363 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7365 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7368 static int swap_high_show(struct seq_file *m, void *v)
7370 return seq_puts_memcg_tunable(m,
7371 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7374 static ssize_t swap_high_write(struct kernfs_open_file *of,
7375 char *buf, size_t nbytes, loff_t off)
7377 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7381 buf = strstrip(buf);
7382 err = page_counter_memparse(buf, "max", &high);
7386 page_counter_set_high(&memcg->swap, high);
7391 static int swap_max_show(struct seq_file *m, void *v)
7393 return seq_puts_memcg_tunable(m,
7394 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7397 static ssize_t swap_max_write(struct kernfs_open_file *of,
7398 char *buf, size_t nbytes, loff_t off)
7400 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7404 buf = strstrip(buf);
7405 err = page_counter_memparse(buf, "max", &max);
7409 xchg(&memcg->swap.max, max);
7414 static int swap_events_show(struct seq_file *m, void *v)
7416 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7418 seq_printf(m, "high %lu\n",
7419 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7420 seq_printf(m, "max %lu\n",
7421 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7422 seq_printf(m, "fail %lu\n",
7423 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7428 static struct cftype swap_files[] = {
7430 .name = "swap.current",
7431 .flags = CFTYPE_NOT_ON_ROOT,
7432 .read_u64 = swap_current_read,
7435 .name = "swap.high",
7436 .flags = CFTYPE_NOT_ON_ROOT,
7437 .seq_show = swap_high_show,
7438 .write = swap_high_write,
7442 .flags = CFTYPE_NOT_ON_ROOT,
7443 .seq_show = swap_max_show,
7444 .write = swap_max_write,
7447 .name = "swap.events",
7448 .flags = CFTYPE_NOT_ON_ROOT,
7449 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7450 .seq_show = swap_events_show,
7455 static struct cftype memsw_files[] = {
7457 .name = "memsw.usage_in_bytes",
7458 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7459 .read_u64 = mem_cgroup_read_u64,
7462 .name = "memsw.max_usage_in_bytes",
7463 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7464 .write = mem_cgroup_reset,
7465 .read_u64 = mem_cgroup_read_u64,
7468 .name = "memsw.limit_in_bytes",
7469 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7470 .write = mem_cgroup_write,
7471 .read_u64 = mem_cgroup_read_u64,
7474 .name = "memsw.failcnt",
7475 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7476 .write = mem_cgroup_reset,
7477 .read_u64 = mem_cgroup_read_u64,
7479 { }, /* terminate */
7483 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7484 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7485 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7486 * boot parameter. This may result in premature OOPS inside
7487 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7489 static int __init mem_cgroup_swap_init(void)
7491 /* No memory control -> no swap control */
7492 if (mem_cgroup_disabled())
7493 cgroup_memory_noswap = true;
7495 if (cgroup_memory_noswap)
7498 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7499 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7503 core_initcall(mem_cgroup_swap_init);
7505 #endif /* CONFIG_MEMCG_SWAP */