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)
663 if (!spin_trylock(&stats_flush_lock))
666 cgroup_rstat_flush_irqsafe(root_mem_cgroup->css.cgroup);
667 atomic_set(&stats_flush_threshold, 0);
668 spin_unlock(&stats_flush_lock);
671 void mem_cgroup_flush_stats(void)
673 if (atomic_read(&stats_flush_threshold) > num_online_cpus())
674 __mem_cgroup_flush_stats();
677 static void flush_memcg_stats_dwork(struct work_struct *w)
679 mem_cgroup_flush_stats();
680 queue_delayed_work(system_unbound_wq, &stats_flush_dwork, 2UL*HZ);
684 * __mod_memcg_state - update cgroup memory statistics
685 * @memcg: the memory cgroup
686 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
687 * @val: delta to add to the counter, can be negative
689 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
691 if (mem_cgroup_disabled())
694 __this_cpu_add(memcg->vmstats_percpu->state[idx], val);
695 memcg_rstat_updated(memcg);
698 /* idx can be of type enum memcg_stat_item or node_stat_item. */
699 static unsigned long memcg_page_state_local(struct mem_cgroup *memcg, int idx)
704 for_each_possible_cpu(cpu)
705 x += per_cpu(memcg->vmstats_percpu->state[idx], cpu);
713 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
716 struct mem_cgroup_per_node *pn;
717 struct mem_cgroup *memcg;
719 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
723 __this_cpu_add(memcg->vmstats_percpu->state[idx], val);
726 __this_cpu_add(pn->lruvec_stats_percpu->state[idx], val);
728 memcg_rstat_updated(memcg);
732 * __mod_lruvec_state - update lruvec memory statistics
733 * @lruvec: the lruvec
734 * @idx: the stat item
735 * @val: delta to add to the counter, can be negative
737 * The lruvec is the intersection of the NUMA node and a cgroup. This
738 * function updates the all three counters that are affected by a
739 * change of state at this level: per-node, per-cgroup, per-lruvec.
741 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
745 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
747 /* Update memcg and lruvec */
748 if (!mem_cgroup_disabled())
749 __mod_memcg_lruvec_state(lruvec, idx, val);
752 void __mod_lruvec_page_state(struct page *page, enum node_stat_item idx,
755 struct page *head = compound_head(page); /* rmap on tail pages */
756 struct mem_cgroup *memcg;
757 pg_data_t *pgdat = page_pgdat(page);
758 struct lruvec *lruvec;
761 memcg = page_memcg(head);
762 /* Untracked pages have no memcg, no lruvec. Update only the node */
765 __mod_node_page_state(pgdat, idx, val);
769 lruvec = mem_cgroup_lruvec(memcg, pgdat);
770 __mod_lruvec_state(lruvec, idx, val);
773 EXPORT_SYMBOL(__mod_lruvec_page_state);
775 void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val)
777 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
778 struct mem_cgroup *memcg;
779 struct lruvec *lruvec;
782 memcg = mem_cgroup_from_obj(p);
785 * Untracked pages have no memcg, no lruvec. Update only the
786 * node. If we reparent the slab objects to the root memcg,
787 * when we free the slab object, we need to update the per-memcg
788 * vmstats to keep it correct for the root memcg.
791 __mod_node_page_state(pgdat, idx, val);
793 lruvec = mem_cgroup_lruvec(memcg, pgdat);
794 __mod_lruvec_state(lruvec, idx, val);
800 * mod_objcg_mlstate() may be called with irq enabled, so
801 * mod_memcg_lruvec_state() should be used.
803 static inline void mod_objcg_mlstate(struct obj_cgroup *objcg,
804 struct pglist_data *pgdat,
805 enum node_stat_item idx, int nr)
807 struct mem_cgroup *memcg;
808 struct lruvec *lruvec;
811 memcg = obj_cgroup_memcg(objcg);
812 lruvec = mem_cgroup_lruvec(memcg, pgdat);
813 mod_memcg_lruvec_state(lruvec, idx, nr);
818 * __count_memcg_events - account VM events in a cgroup
819 * @memcg: the memory cgroup
820 * @idx: the event item
821 * @count: the number of events that occurred
823 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
826 if (mem_cgroup_disabled())
829 __this_cpu_add(memcg->vmstats_percpu->events[idx], count);
830 memcg_rstat_updated(memcg);
833 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
835 return READ_ONCE(memcg->vmstats.events[event]);
838 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
843 for_each_possible_cpu(cpu)
844 x += per_cpu(memcg->vmstats_percpu->events[event], cpu);
848 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
852 /* pagein of a big page is an event. So, ignore page size */
854 __count_memcg_events(memcg, PGPGIN, 1);
856 __count_memcg_events(memcg, PGPGOUT, 1);
857 nr_pages = -nr_pages; /* for event */
860 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
863 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
864 enum mem_cgroup_events_target target)
866 unsigned long val, next;
868 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
869 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
870 /* from time_after() in jiffies.h */
871 if ((long)(next - val) < 0) {
873 case MEM_CGROUP_TARGET_THRESH:
874 next = val + THRESHOLDS_EVENTS_TARGET;
876 case MEM_CGROUP_TARGET_SOFTLIMIT:
877 next = val + SOFTLIMIT_EVENTS_TARGET;
882 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
889 * Check events in order.
892 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
894 /* threshold event is triggered in finer grain than soft limit */
895 if (unlikely(mem_cgroup_event_ratelimit(memcg,
896 MEM_CGROUP_TARGET_THRESH))) {
899 do_softlimit = mem_cgroup_event_ratelimit(memcg,
900 MEM_CGROUP_TARGET_SOFTLIMIT);
901 mem_cgroup_threshold(memcg);
902 if (unlikely(do_softlimit))
903 mem_cgroup_update_tree(memcg, page);
907 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
910 * mm_update_next_owner() may clear mm->owner to NULL
911 * if it races with swapoff, page migration, etc.
912 * So this can be called with p == NULL.
917 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
919 EXPORT_SYMBOL(mem_cgroup_from_task);
921 static __always_inline struct mem_cgroup *active_memcg(void)
924 return this_cpu_read(int_active_memcg);
926 return current->active_memcg;
930 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
931 * @mm: mm from which memcg should be extracted. It can be NULL.
933 * Obtain a reference on mm->memcg and returns it if successful. If mm
934 * is NULL, then the memcg is chosen as follows:
935 * 1) The active memcg, if set.
936 * 2) current->mm->memcg, if available
938 * If mem_cgroup is disabled, NULL is returned.
940 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
942 struct mem_cgroup *memcg;
944 if (mem_cgroup_disabled())
948 * Page cache insertions can happen without an
949 * actual mm context, e.g. during disk probing
950 * on boot, loopback IO, acct() writes etc.
952 * No need to css_get on root memcg as the reference
953 * counting is disabled on the root level in the
954 * cgroup core. See CSS_NO_REF.
957 memcg = active_memcg();
958 if (unlikely(memcg)) {
959 /* remote memcg must hold a ref */
960 css_get(&memcg->css);
965 return root_mem_cgroup;
970 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
971 if (unlikely(!memcg))
972 memcg = root_mem_cgroup;
973 } while (!css_tryget(&memcg->css));
977 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
979 static __always_inline bool memcg_kmem_bypass(void)
981 /* Allow remote memcg charging from any context. */
982 if (unlikely(active_memcg()))
985 /* Memcg to charge can't be determined. */
986 if (!in_task() || !current->mm || (current->flags & PF_KTHREAD))
993 * mem_cgroup_iter - iterate over memory cgroup hierarchy
994 * @root: hierarchy root
995 * @prev: previously returned memcg, NULL on first invocation
996 * @reclaim: cookie for shared reclaim walks, NULL for full walks
998 * Returns references to children of the hierarchy below @root, or
999 * @root itself, or %NULL after a full round-trip.
1001 * Caller must pass the return value in @prev on subsequent
1002 * invocations for reference counting, or use mem_cgroup_iter_break()
1003 * to cancel a hierarchy walk before the round-trip is complete.
1005 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1006 * in the hierarchy among all concurrent reclaimers operating on the
1009 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1010 struct mem_cgroup *prev,
1011 struct mem_cgroup_reclaim_cookie *reclaim)
1013 struct mem_cgroup_reclaim_iter *iter;
1014 struct cgroup_subsys_state *css = NULL;
1015 struct mem_cgroup *memcg = NULL;
1016 struct mem_cgroup *pos = NULL;
1018 if (mem_cgroup_disabled())
1022 root = root_mem_cgroup;
1024 if (prev && !reclaim)
1030 struct mem_cgroup_per_node *mz;
1032 mz = root->nodeinfo[reclaim->pgdat->node_id];
1035 if (prev && reclaim->generation != iter->generation)
1039 pos = READ_ONCE(iter->position);
1040 if (!pos || css_tryget(&pos->css))
1043 * css reference reached zero, so iter->position will
1044 * be cleared by ->css_released. However, we should not
1045 * rely on this happening soon, because ->css_released
1046 * is called from a work queue, and by busy-waiting we
1047 * might block it. So we clear iter->position right
1050 (void)cmpxchg(&iter->position, pos, NULL);
1058 css = css_next_descendant_pre(css, &root->css);
1061 * Reclaimers share the hierarchy walk, and a
1062 * new one might jump in right at the end of
1063 * the hierarchy - make sure they see at least
1064 * one group and restart from the beginning.
1072 * Verify the css and acquire a reference. The root
1073 * is provided by the caller, so we know it's alive
1074 * and kicking, and don't take an extra reference.
1076 memcg = mem_cgroup_from_css(css);
1078 if (css == &root->css)
1081 if (css_tryget(css))
1089 * The position could have already been updated by a competing
1090 * thread, so check that the value hasn't changed since we read
1091 * it to avoid reclaiming from the same cgroup twice.
1093 (void)cmpxchg(&iter->position, pos, memcg);
1101 reclaim->generation = iter->generation;
1106 if (prev && prev != root)
1107 css_put(&prev->css);
1113 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1114 * @root: hierarchy root
1115 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1117 void mem_cgroup_iter_break(struct mem_cgroup *root,
1118 struct mem_cgroup *prev)
1121 root = root_mem_cgroup;
1122 if (prev && prev != root)
1123 css_put(&prev->css);
1126 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1127 struct mem_cgroup *dead_memcg)
1129 struct mem_cgroup_reclaim_iter *iter;
1130 struct mem_cgroup_per_node *mz;
1133 for_each_node(nid) {
1134 mz = from->nodeinfo[nid];
1136 cmpxchg(&iter->position, dead_memcg, NULL);
1140 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1142 struct mem_cgroup *memcg = dead_memcg;
1143 struct mem_cgroup *last;
1146 __invalidate_reclaim_iterators(memcg, dead_memcg);
1148 } while ((memcg = parent_mem_cgroup(memcg)));
1151 * When cgruop1 non-hierarchy mode is used,
1152 * parent_mem_cgroup() does not walk all the way up to the
1153 * cgroup root (root_mem_cgroup). So we have to handle
1154 * dead_memcg from cgroup root separately.
1156 if (last != root_mem_cgroup)
1157 __invalidate_reclaim_iterators(root_mem_cgroup,
1162 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1163 * @memcg: hierarchy root
1164 * @fn: function to call for each task
1165 * @arg: argument passed to @fn
1167 * This function iterates over tasks attached to @memcg or to any of its
1168 * descendants and calls @fn for each task. If @fn returns a non-zero
1169 * value, the function breaks the iteration loop and returns the value.
1170 * Otherwise, it will iterate over all tasks and return 0.
1172 * This function must not be called for the root memory cgroup.
1174 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1175 int (*fn)(struct task_struct *, void *), void *arg)
1177 struct mem_cgroup *iter;
1180 BUG_ON(memcg == root_mem_cgroup);
1182 for_each_mem_cgroup_tree(iter, memcg) {
1183 struct css_task_iter it;
1184 struct task_struct *task;
1186 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1187 while (!ret && (task = css_task_iter_next(&it)))
1188 ret = fn(task, arg);
1189 css_task_iter_end(&it);
1191 mem_cgroup_iter_break(memcg, iter);
1198 #ifdef CONFIG_DEBUG_VM
1199 void lruvec_memcg_debug(struct lruvec *lruvec, struct page *page)
1201 struct mem_cgroup *memcg;
1203 if (mem_cgroup_disabled())
1206 memcg = page_memcg(page);
1209 VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != root_mem_cgroup, page);
1211 VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != memcg, page);
1216 * lock_page_lruvec - lock and return lruvec for a given page.
1219 * These functions are safe to use under any of the following conditions:
1222 * - lock_page_memcg()
1223 * - page->_refcount is zero
1225 struct lruvec *lock_page_lruvec(struct page *page)
1227 struct lruvec *lruvec;
1229 lruvec = mem_cgroup_page_lruvec(page);
1230 spin_lock(&lruvec->lru_lock);
1232 lruvec_memcg_debug(lruvec, page);
1237 struct lruvec *lock_page_lruvec_irq(struct page *page)
1239 struct lruvec *lruvec;
1241 lruvec = mem_cgroup_page_lruvec(page);
1242 spin_lock_irq(&lruvec->lru_lock);
1244 lruvec_memcg_debug(lruvec, page);
1249 struct lruvec *lock_page_lruvec_irqsave(struct page *page, unsigned long *flags)
1251 struct lruvec *lruvec;
1253 lruvec = mem_cgroup_page_lruvec(page);
1254 spin_lock_irqsave(&lruvec->lru_lock, *flags);
1256 lruvec_memcg_debug(lruvec, page);
1262 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1263 * @lruvec: mem_cgroup per zone lru vector
1264 * @lru: index of lru list the page is sitting on
1265 * @zid: zone id of the accounted pages
1266 * @nr_pages: positive when adding or negative when removing
1268 * This function must be called under lru_lock, just before a page is added
1269 * to or just after a page is removed from an lru list (that ordering being
1270 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1272 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1273 int zid, int nr_pages)
1275 struct mem_cgroup_per_node *mz;
1276 unsigned long *lru_size;
1279 if (mem_cgroup_disabled())
1282 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1283 lru_size = &mz->lru_zone_size[zid][lru];
1286 *lru_size += nr_pages;
1289 if (WARN_ONCE(size < 0,
1290 "%s(%p, %d, %d): lru_size %ld\n",
1291 __func__, lruvec, lru, nr_pages, size)) {
1297 *lru_size += nr_pages;
1301 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1302 * @memcg: the memory cgroup
1304 * Returns the maximum amount of memory @mem can be charged with, in
1307 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1309 unsigned long margin = 0;
1310 unsigned long count;
1311 unsigned long limit;
1313 count = page_counter_read(&memcg->memory);
1314 limit = READ_ONCE(memcg->memory.max);
1316 margin = limit - count;
1318 if (do_memsw_account()) {
1319 count = page_counter_read(&memcg->memsw);
1320 limit = READ_ONCE(memcg->memsw.max);
1322 margin = min(margin, limit - count);
1331 * A routine for checking "mem" is under move_account() or not.
1333 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1334 * moving cgroups. This is for waiting at high-memory pressure
1337 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1339 struct mem_cgroup *from;
1340 struct mem_cgroup *to;
1343 * Unlike task_move routines, we access mc.to, mc.from not under
1344 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1346 spin_lock(&mc.lock);
1352 ret = mem_cgroup_is_descendant(from, memcg) ||
1353 mem_cgroup_is_descendant(to, memcg);
1355 spin_unlock(&mc.lock);
1359 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1361 if (mc.moving_task && current != mc.moving_task) {
1362 if (mem_cgroup_under_move(memcg)) {
1364 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1365 /* moving charge context might have finished. */
1368 finish_wait(&mc.waitq, &wait);
1375 struct memory_stat {
1380 static const struct memory_stat memory_stats[] = {
1381 { "anon", NR_ANON_MAPPED },
1382 { "file", NR_FILE_PAGES },
1383 { "kernel_stack", NR_KERNEL_STACK_KB },
1384 { "pagetables", NR_PAGETABLE },
1385 { "percpu", MEMCG_PERCPU_B },
1386 { "sock", MEMCG_SOCK },
1387 { "shmem", NR_SHMEM },
1388 { "file_mapped", NR_FILE_MAPPED },
1389 { "file_dirty", NR_FILE_DIRTY },
1390 { "file_writeback", NR_WRITEBACK },
1392 { "swapcached", NR_SWAPCACHE },
1394 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1395 { "anon_thp", NR_ANON_THPS },
1396 { "file_thp", NR_FILE_THPS },
1397 { "shmem_thp", NR_SHMEM_THPS },
1399 { "inactive_anon", NR_INACTIVE_ANON },
1400 { "active_anon", NR_ACTIVE_ANON },
1401 { "inactive_file", NR_INACTIVE_FILE },
1402 { "active_file", NR_ACTIVE_FILE },
1403 { "unevictable", NR_UNEVICTABLE },
1404 { "slab_reclaimable", NR_SLAB_RECLAIMABLE_B },
1405 { "slab_unreclaimable", NR_SLAB_UNRECLAIMABLE_B },
1407 /* The memory events */
1408 { "workingset_refault_anon", WORKINGSET_REFAULT_ANON },
1409 { "workingset_refault_file", WORKINGSET_REFAULT_FILE },
1410 { "workingset_activate_anon", WORKINGSET_ACTIVATE_ANON },
1411 { "workingset_activate_file", WORKINGSET_ACTIVATE_FILE },
1412 { "workingset_restore_anon", WORKINGSET_RESTORE_ANON },
1413 { "workingset_restore_file", WORKINGSET_RESTORE_FILE },
1414 { "workingset_nodereclaim", WORKINGSET_NODERECLAIM },
1417 /* Translate stat items to the correct unit for memory.stat output */
1418 static int memcg_page_state_unit(int item)
1421 case MEMCG_PERCPU_B:
1422 case NR_SLAB_RECLAIMABLE_B:
1423 case NR_SLAB_UNRECLAIMABLE_B:
1424 case WORKINGSET_REFAULT_ANON:
1425 case WORKINGSET_REFAULT_FILE:
1426 case WORKINGSET_ACTIVATE_ANON:
1427 case WORKINGSET_ACTIVATE_FILE:
1428 case WORKINGSET_RESTORE_ANON:
1429 case WORKINGSET_RESTORE_FILE:
1430 case WORKINGSET_NODERECLAIM:
1432 case NR_KERNEL_STACK_KB:
1439 static inline unsigned long memcg_page_state_output(struct mem_cgroup *memcg,
1442 return memcg_page_state(memcg, item) * memcg_page_state_unit(item);
1445 static char *memory_stat_format(struct mem_cgroup *memcg)
1450 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1455 * Provide statistics on the state of the memory subsystem as
1456 * well as cumulative event counters that show past behavior.
1458 * This list is ordered following a combination of these gradients:
1459 * 1) generic big picture -> specifics and details
1460 * 2) reflecting userspace activity -> reflecting kernel heuristics
1462 * Current memory state:
1464 cgroup_rstat_flush(memcg->css.cgroup);
1466 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1469 size = memcg_page_state_output(memcg, memory_stats[i].idx);
1470 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1472 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1473 size += memcg_page_state_output(memcg,
1474 NR_SLAB_RECLAIMABLE_B);
1475 seq_buf_printf(&s, "slab %llu\n", size);
1479 /* Accumulated memory events */
1481 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1482 memcg_events(memcg, PGFAULT));
1483 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1484 memcg_events(memcg, PGMAJFAULT));
1485 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1486 memcg_events(memcg, PGREFILL));
1487 seq_buf_printf(&s, "pgscan %lu\n",
1488 memcg_events(memcg, PGSCAN_KSWAPD) +
1489 memcg_events(memcg, PGSCAN_DIRECT));
1490 seq_buf_printf(&s, "pgsteal %lu\n",
1491 memcg_events(memcg, PGSTEAL_KSWAPD) +
1492 memcg_events(memcg, PGSTEAL_DIRECT));
1493 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1494 memcg_events(memcg, PGACTIVATE));
1495 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1496 memcg_events(memcg, PGDEACTIVATE));
1497 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1498 memcg_events(memcg, PGLAZYFREE));
1499 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1500 memcg_events(memcg, PGLAZYFREED));
1502 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1503 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1504 memcg_events(memcg, THP_FAULT_ALLOC));
1505 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1506 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1507 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1509 /* The above should easily fit into one page */
1510 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1515 #define K(x) ((x) << (PAGE_SHIFT-10))
1517 * mem_cgroup_print_oom_context: Print OOM information relevant to
1518 * memory controller.
1519 * @memcg: The memory cgroup that went over limit
1520 * @p: Task that is going to be killed
1522 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1525 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1530 pr_cont(",oom_memcg=");
1531 pr_cont_cgroup_path(memcg->css.cgroup);
1533 pr_cont(",global_oom");
1535 pr_cont(",task_memcg=");
1536 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1542 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1543 * memory controller.
1544 * @memcg: The memory cgroup that went over limit
1546 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1550 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1551 K((u64)page_counter_read(&memcg->memory)),
1552 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1553 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1554 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1555 K((u64)page_counter_read(&memcg->swap)),
1556 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1558 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1559 K((u64)page_counter_read(&memcg->memsw)),
1560 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1561 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1562 K((u64)page_counter_read(&memcg->kmem)),
1563 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1566 pr_info("Memory cgroup stats for ");
1567 pr_cont_cgroup_path(memcg->css.cgroup);
1569 buf = memory_stat_format(memcg);
1577 * Return the memory (and swap, if configured) limit for a memcg.
1579 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1581 unsigned long max = READ_ONCE(memcg->memory.max);
1583 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1584 if (mem_cgroup_swappiness(memcg))
1585 max += min(READ_ONCE(memcg->swap.max),
1586 (unsigned long)total_swap_pages);
1588 if (mem_cgroup_swappiness(memcg)) {
1589 /* Calculate swap excess capacity from memsw limit */
1590 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1592 max += min(swap, (unsigned long)total_swap_pages);
1598 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1600 return page_counter_read(&memcg->memory);
1603 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1606 struct oom_control oc = {
1610 .gfp_mask = gfp_mask,
1615 if (mutex_lock_killable(&oom_lock))
1618 if (mem_cgroup_margin(memcg) >= (1 << order))
1622 * A few threads which were not waiting at mutex_lock_killable() can
1623 * fail to bail out. Therefore, check again after holding oom_lock.
1625 ret = task_is_dying() || out_of_memory(&oc);
1628 mutex_unlock(&oom_lock);
1632 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1635 unsigned long *total_scanned)
1637 struct mem_cgroup *victim = NULL;
1640 unsigned long excess;
1641 unsigned long nr_scanned;
1642 struct mem_cgroup_reclaim_cookie reclaim = {
1646 excess = soft_limit_excess(root_memcg);
1649 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1654 * If we have not been able to reclaim
1655 * anything, it might because there are
1656 * no reclaimable pages under this hierarchy
1661 * We want to do more targeted reclaim.
1662 * excess >> 2 is not to excessive so as to
1663 * reclaim too much, nor too less that we keep
1664 * coming back to reclaim from this cgroup
1666 if (total >= (excess >> 2) ||
1667 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1672 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1673 pgdat, &nr_scanned);
1674 *total_scanned += nr_scanned;
1675 if (!soft_limit_excess(root_memcg))
1678 mem_cgroup_iter_break(root_memcg, victim);
1682 #ifdef CONFIG_LOCKDEP
1683 static struct lockdep_map memcg_oom_lock_dep_map = {
1684 .name = "memcg_oom_lock",
1688 static DEFINE_SPINLOCK(memcg_oom_lock);
1691 * Check OOM-Killer is already running under our hierarchy.
1692 * If someone is running, return false.
1694 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1696 struct mem_cgroup *iter, *failed = NULL;
1698 spin_lock(&memcg_oom_lock);
1700 for_each_mem_cgroup_tree(iter, memcg) {
1701 if (iter->oom_lock) {
1703 * this subtree of our hierarchy is already locked
1704 * so we cannot give a lock.
1707 mem_cgroup_iter_break(memcg, iter);
1710 iter->oom_lock = true;
1715 * OK, we failed to lock the whole subtree so we have
1716 * to clean up what we set up to the failing subtree
1718 for_each_mem_cgroup_tree(iter, memcg) {
1719 if (iter == failed) {
1720 mem_cgroup_iter_break(memcg, iter);
1723 iter->oom_lock = false;
1726 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1728 spin_unlock(&memcg_oom_lock);
1733 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1735 struct mem_cgroup *iter;
1737 spin_lock(&memcg_oom_lock);
1738 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1739 for_each_mem_cgroup_tree(iter, memcg)
1740 iter->oom_lock = false;
1741 spin_unlock(&memcg_oom_lock);
1744 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1746 struct mem_cgroup *iter;
1748 spin_lock(&memcg_oom_lock);
1749 for_each_mem_cgroup_tree(iter, memcg)
1751 spin_unlock(&memcg_oom_lock);
1754 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1756 struct mem_cgroup *iter;
1759 * Be careful about under_oom underflows because a child memcg
1760 * could have been added after mem_cgroup_mark_under_oom.
1762 spin_lock(&memcg_oom_lock);
1763 for_each_mem_cgroup_tree(iter, memcg)
1764 if (iter->under_oom > 0)
1766 spin_unlock(&memcg_oom_lock);
1769 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1771 struct oom_wait_info {
1772 struct mem_cgroup *memcg;
1773 wait_queue_entry_t wait;
1776 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1777 unsigned mode, int sync, void *arg)
1779 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1780 struct mem_cgroup *oom_wait_memcg;
1781 struct oom_wait_info *oom_wait_info;
1783 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1784 oom_wait_memcg = oom_wait_info->memcg;
1786 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1787 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1789 return autoremove_wake_function(wait, mode, sync, arg);
1792 static void memcg_oom_recover(struct mem_cgroup *memcg)
1795 * For the following lockless ->under_oom test, the only required
1796 * guarantee is that it must see the state asserted by an OOM when
1797 * this function is called as a result of userland actions
1798 * triggered by the notification of the OOM. This is trivially
1799 * achieved by invoking mem_cgroup_mark_under_oom() before
1800 * triggering notification.
1802 if (memcg && memcg->under_oom)
1803 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1813 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1815 enum oom_status ret;
1818 if (order > PAGE_ALLOC_COSTLY_ORDER)
1821 memcg_memory_event(memcg, MEMCG_OOM);
1824 * We are in the middle of the charge context here, so we
1825 * don't want to block when potentially sitting on a callstack
1826 * that holds all kinds of filesystem and mm locks.
1828 * cgroup1 allows disabling the OOM killer and waiting for outside
1829 * handling until the charge can succeed; remember the context and put
1830 * the task to sleep at the end of the page fault when all locks are
1833 * On the other hand, in-kernel OOM killer allows for an async victim
1834 * memory reclaim (oom_reaper) and that means that we are not solely
1835 * relying on the oom victim to make a forward progress and we can
1836 * invoke the oom killer here.
1838 * Please note that mem_cgroup_out_of_memory might fail to find a
1839 * victim and then we have to bail out from the charge path.
1841 if (memcg->oom_kill_disable) {
1842 if (!current->in_user_fault)
1844 css_get(&memcg->css);
1845 current->memcg_in_oom = memcg;
1846 current->memcg_oom_gfp_mask = mask;
1847 current->memcg_oom_order = order;
1852 mem_cgroup_mark_under_oom(memcg);
1854 locked = mem_cgroup_oom_trylock(memcg);
1857 mem_cgroup_oom_notify(memcg);
1859 mem_cgroup_unmark_under_oom(memcg);
1860 if (mem_cgroup_out_of_memory(memcg, mask, order))
1866 mem_cgroup_oom_unlock(memcg);
1872 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1873 * @handle: actually kill/wait or just clean up the OOM state
1875 * This has to be called at the end of a page fault if the memcg OOM
1876 * handler was enabled.
1878 * Memcg supports userspace OOM handling where failed allocations must
1879 * sleep on a waitqueue until the userspace task resolves the
1880 * situation. Sleeping directly in the charge context with all kinds
1881 * of locks held is not a good idea, instead we remember an OOM state
1882 * in the task and mem_cgroup_oom_synchronize() has to be called at
1883 * the end of the page fault to complete the OOM handling.
1885 * Returns %true if an ongoing memcg OOM situation was detected and
1886 * completed, %false otherwise.
1888 bool mem_cgroup_oom_synchronize(bool handle)
1890 struct mem_cgroup *memcg = current->memcg_in_oom;
1891 struct oom_wait_info owait;
1894 /* OOM is global, do not handle */
1901 owait.memcg = memcg;
1902 owait.wait.flags = 0;
1903 owait.wait.func = memcg_oom_wake_function;
1904 owait.wait.private = current;
1905 INIT_LIST_HEAD(&owait.wait.entry);
1907 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1908 mem_cgroup_mark_under_oom(memcg);
1910 locked = mem_cgroup_oom_trylock(memcg);
1913 mem_cgroup_oom_notify(memcg);
1915 if (locked && !memcg->oom_kill_disable) {
1916 mem_cgroup_unmark_under_oom(memcg);
1917 finish_wait(&memcg_oom_waitq, &owait.wait);
1918 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1919 current->memcg_oom_order);
1922 mem_cgroup_unmark_under_oom(memcg);
1923 finish_wait(&memcg_oom_waitq, &owait.wait);
1927 mem_cgroup_oom_unlock(memcg);
1929 * There is no guarantee that an OOM-lock contender
1930 * sees the wakeups triggered by the OOM kill
1931 * uncharges. Wake any sleepers explicitly.
1933 memcg_oom_recover(memcg);
1936 current->memcg_in_oom = NULL;
1937 css_put(&memcg->css);
1942 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1943 * @victim: task to be killed by the OOM killer
1944 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1946 * Returns a pointer to a memory cgroup, which has to be cleaned up
1947 * by killing all belonging OOM-killable tasks.
1949 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1951 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1952 struct mem_cgroup *oom_domain)
1954 struct mem_cgroup *oom_group = NULL;
1955 struct mem_cgroup *memcg;
1957 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
1961 oom_domain = root_mem_cgroup;
1965 memcg = mem_cgroup_from_task(victim);
1966 if (memcg == root_mem_cgroup)
1970 * If the victim task has been asynchronously moved to a different
1971 * memory cgroup, we might end up killing tasks outside oom_domain.
1972 * In this case it's better to ignore memory.group.oom.
1974 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
1978 * Traverse the memory cgroup hierarchy from the victim task's
1979 * cgroup up to the OOMing cgroup (or root) to find the
1980 * highest-level memory cgroup with oom.group set.
1982 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1983 if (memcg->oom_group)
1986 if (memcg == oom_domain)
1991 css_get(&oom_group->css);
1998 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2000 pr_info("Tasks in ");
2001 pr_cont_cgroup_path(memcg->css.cgroup);
2002 pr_cont(" are going to be killed due to memory.oom.group set\n");
2006 * lock_page_memcg - lock a page and memcg binding
2009 * This function protects unlocked LRU pages from being moved to
2012 * It ensures lifetime of the locked memcg. Caller is responsible
2013 * for the lifetime of the page.
2015 void lock_page_memcg(struct page *page)
2017 struct page *head = compound_head(page); /* rmap on tail pages */
2018 struct mem_cgroup *memcg;
2019 unsigned long flags;
2022 * The RCU lock is held throughout the transaction. The fast
2023 * path can get away without acquiring the memcg->move_lock
2024 * because page moving starts with an RCU grace period.
2028 if (mem_cgroup_disabled())
2031 memcg = page_memcg(head);
2032 if (unlikely(!memcg))
2035 #ifdef CONFIG_PROVE_LOCKING
2036 local_irq_save(flags);
2037 might_lock(&memcg->move_lock);
2038 local_irq_restore(flags);
2041 if (atomic_read(&memcg->moving_account) <= 0)
2044 spin_lock_irqsave(&memcg->move_lock, flags);
2045 if (memcg != page_memcg(head)) {
2046 spin_unlock_irqrestore(&memcg->move_lock, flags);
2051 * When charge migration first begins, we can have multiple
2052 * critical sections holding the fast-path RCU lock and one
2053 * holding the slowpath move_lock. Track the task who has the
2054 * move_lock for unlock_page_memcg().
2056 memcg->move_lock_task = current;
2057 memcg->move_lock_flags = flags;
2059 EXPORT_SYMBOL(lock_page_memcg);
2061 static void __unlock_page_memcg(struct mem_cgroup *memcg)
2063 if (memcg && memcg->move_lock_task == current) {
2064 unsigned long flags = memcg->move_lock_flags;
2066 memcg->move_lock_task = NULL;
2067 memcg->move_lock_flags = 0;
2069 spin_unlock_irqrestore(&memcg->move_lock, flags);
2076 * unlock_page_memcg - unlock a page and memcg binding
2079 void unlock_page_memcg(struct page *page)
2081 struct page *head = compound_head(page);
2083 __unlock_page_memcg(page_memcg(head));
2085 EXPORT_SYMBOL(unlock_page_memcg);
2088 #ifdef CONFIG_MEMCG_KMEM
2089 struct obj_cgroup *cached_objcg;
2090 struct pglist_data *cached_pgdat;
2091 unsigned int nr_bytes;
2092 int nr_slab_reclaimable_b;
2093 int nr_slab_unreclaimable_b;
2099 struct memcg_stock_pcp {
2100 struct mem_cgroup *cached; /* this never be root cgroup */
2101 unsigned int nr_pages;
2102 struct obj_stock task_obj;
2103 struct obj_stock irq_obj;
2105 struct work_struct work;
2106 unsigned long flags;
2107 #define FLUSHING_CACHED_CHARGE 0
2109 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2110 static DEFINE_MUTEX(percpu_charge_mutex);
2112 #ifdef CONFIG_MEMCG_KMEM
2113 static void drain_obj_stock(struct obj_stock *stock);
2114 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2115 struct mem_cgroup *root_memcg);
2118 static inline void drain_obj_stock(struct obj_stock *stock)
2121 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2122 struct mem_cgroup *root_memcg)
2129 * Most kmem_cache_alloc() calls are from user context. The irq disable/enable
2130 * sequence used in this case to access content from object stock is slow.
2131 * To optimize for user context access, there are now two object stocks for
2132 * task context and interrupt context access respectively.
2134 * The task context object stock can be accessed by disabling preemption only
2135 * which is cheap in non-preempt kernel. The interrupt context object stock
2136 * can only be accessed after disabling interrupt. User context code can
2137 * access interrupt object stock, but not vice versa.
2139 static inline struct obj_stock *get_obj_stock(unsigned long *pflags)
2141 struct memcg_stock_pcp *stock;
2143 if (likely(in_task())) {
2146 stock = this_cpu_ptr(&memcg_stock);
2147 return &stock->task_obj;
2150 local_irq_save(*pflags);
2151 stock = this_cpu_ptr(&memcg_stock);
2152 return &stock->irq_obj;
2155 static inline void put_obj_stock(unsigned long flags)
2157 if (likely(in_task()))
2160 local_irq_restore(flags);
2164 * consume_stock: Try to consume stocked charge on this cpu.
2165 * @memcg: memcg to consume from.
2166 * @nr_pages: how many pages to charge.
2168 * The charges will only happen if @memcg matches the current cpu's memcg
2169 * stock, and at least @nr_pages are available in that stock. Failure to
2170 * service an allocation will refill the stock.
2172 * returns true if successful, false otherwise.
2174 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2176 struct memcg_stock_pcp *stock;
2177 unsigned long flags;
2180 if (nr_pages > MEMCG_CHARGE_BATCH)
2183 local_irq_save(flags);
2185 stock = this_cpu_ptr(&memcg_stock);
2186 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2187 stock->nr_pages -= nr_pages;
2191 local_irq_restore(flags);
2197 * Returns stocks cached in percpu and reset cached information.
2199 static void drain_stock(struct memcg_stock_pcp *stock)
2201 struct mem_cgroup *old = stock->cached;
2206 if (stock->nr_pages) {
2207 page_counter_uncharge(&old->memory, stock->nr_pages);
2208 if (do_memsw_account())
2209 page_counter_uncharge(&old->memsw, stock->nr_pages);
2210 stock->nr_pages = 0;
2214 stock->cached = NULL;
2217 static void drain_local_stock(struct work_struct *dummy)
2219 struct memcg_stock_pcp *stock;
2220 unsigned long flags;
2223 * The only protection from cpu hotplug (memcg_hotplug_cpu_dead) vs.
2224 * drain_stock races is that we always operate on local CPU stock
2225 * here with IRQ disabled
2227 local_irq_save(flags);
2229 stock = this_cpu_ptr(&memcg_stock);
2230 drain_obj_stock(&stock->irq_obj);
2232 drain_obj_stock(&stock->task_obj);
2234 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2236 local_irq_restore(flags);
2240 * Cache charges(val) to local per_cpu area.
2241 * This will be consumed by consume_stock() function, later.
2243 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2245 struct memcg_stock_pcp *stock;
2246 unsigned long flags;
2248 local_irq_save(flags);
2250 stock = this_cpu_ptr(&memcg_stock);
2251 if (stock->cached != memcg) { /* reset if necessary */
2253 css_get(&memcg->css);
2254 stock->cached = memcg;
2256 stock->nr_pages += nr_pages;
2258 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2261 local_irq_restore(flags);
2265 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2266 * of the hierarchy under it.
2268 static void drain_all_stock(struct mem_cgroup *root_memcg)
2272 /* If someone's already draining, avoid adding running more workers. */
2273 if (!mutex_trylock(&percpu_charge_mutex))
2276 * Notify other cpus that system-wide "drain" is running
2277 * We do not care about races with the cpu hotplug because cpu down
2278 * as well as workers from this path always operate on the local
2279 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2282 for_each_online_cpu(cpu) {
2283 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2284 struct mem_cgroup *memcg;
2288 memcg = stock->cached;
2289 if (memcg && stock->nr_pages &&
2290 mem_cgroup_is_descendant(memcg, root_memcg))
2292 else if (obj_stock_flush_required(stock, root_memcg))
2297 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2299 drain_local_stock(&stock->work);
2301 schedule_work_on(cpu, &stock->work);
2305 mutex_unlock(&percpu_charge_mutex);
2308 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2310 struct memcg_stock_pcp *stock;
2312 stock = &per_cpu(memcg_stock, cpu);
2318 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2319 unsigned int nr_pages,
2322 unsigned long nr_reclaimed = 0;
2325 unsigned long pflags;
2327 if (page_counter_read(&memcg->memory) <=
2328 READ_ONCE(memcg->memory.high))
2331 memcg_memory_event(memcg, MEMCG_HIGH);
2333 psi_memstall_enter(&pflags);
2334 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2336 psi_memstall_leave(&pflags);
2337 } while ((memcg = parent_mem_cgroup(memcg)) &&
2338 !mem_cgroup_is_root(memcg));
2340 return nr_reclaimed;
2343 static void high_work_func(struct work_struct *work)
2345 struct mem_cgroup *memcg;
2347 memcg = container_of(work, struct mem_cgroup, high_work);
2348 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2352 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2353 * enough to still cause a significant slowdown in most cases, while still
2354 * allowing diagnostics and tracing to proceed without becoming stuck.
2356 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2359 * When calculating the delay, we use these either side of the exponentiation to
2360 * maintain precision and scale to a reasonable number of jiffies (see the table
2363 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2364 * overage ratio to a delay.
2365 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2366 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2367 * to produce a reasonable delay curve.
2369 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2370 * reasonable delay curve compared to precision-adjusted overage, not
2371 * penalising heavily at first, but still making sure that growth beyond the
2372 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2373 * example, with a high of 100 megabytes:
2375 * +-------+------------------------+
2376 * | usage | time to allocate in ms |
2377 * +-------+------------------------+
2399 * +-------+------------------------+
2401 #define MEMCG_DELAY_PRECISION_SHIFT 20
2402 #define MEMCG_DELAY_SCALING_SHIFT 14
2404 static u64 calculate_overage(unsigned long usage, unsigned long high)
2412 * Prevent division by 0 in overage calculation by acting as if
2413 * it was a threshold of 1 page
2415 high = max(high, 1UL);
2417 overage = usage - high;
2418 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2419 return div64_u64(overage, high);
2422 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2424 u64 overage, max_overage = 0;
2427 overage = calculate_overage(page_counter_read(&memcg->memory),
2428 READ_ONCE(memcg->memory.high));
2429 max_overage = max(overage, max_overage);
2430 } while ((memcg = parent_mem_cgroup(memcg)) &&
2431 !mem_cgroup_is_root(memcg));
2436 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2438 u64 overage, max_overage = 0;
2441 overage = calculate_overage(page_counter_read(&memcg->swap),
2442 READ_ONCE(memcg->swap.high));
2444 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2445 max_overage = max(overage, max_overage);
2446 } while ((memcg = parent_mem_cgroup(memcg)) &&
2447 !mem_cgroup_is_root(memcg));
2453 * Get the number of jiffies that we should penalise a mischievous cgroup which
2454 * is exceeding its memory.high by checking both it and its ancestors.
2456 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2457 unsigned int nr_pages,
2460 unsigned long penalty_jiffies;
2466 * We use overage compared to memory.high to calculate the number of
2467 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2468 * fairly lenient on small overages, and increasingly harsh when the
2469 * memcg in question makes it clear that it has no intention of stopping
2470 * its crazy behaviour, so we exponentially increase the delay based on
2473 penalty_jiffies = max_overage * max_overage * HZ;
2474 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2475 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2478 * Factor in the task's own contribution to the overage, such that four
2479 * N-sized allocations are throttled approximately the same as one
2480 * 4N-sized allocation.
2482 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2483 * larger the current charge patch is than that.
2485 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2489 * Scheduled by try_charge() to be executed from the userland return path
2490 * and reclaims memory over the high limit.
2492 void mem_cgroup_handle_over_high(void)
2494 unsigned long penalty_jiffies;
2495 unsigned long pflags;
2496 unsigned long nr_reclaimed;
2497 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2498 int nr_retries = MAX_RECLAIM_RETRIES;
2499 struct mem_cgroup *memcg;
2500 bool in_retry = false;
2502 if (likely(!nr_pages))
2505 memcg = get_mem_cgroup_from_mm(current->mm);
2506 current->memcg_nr_pages_over_high = 0;
2510 * The allocating task should reclaim at least the batch size, but for
2511 * subsequent retries we only want to do what's necessary to prevent oom
2512 * or breaching resource isolation.
2514 * This is distinct from memory.max or page allocator behaviour because
2515 * memory.high is currently batched, whereas memory.max and the page
2516 * allocator run every time an allocation is made.
2518 nr_reclaimed = reclaim_high(memcg,
2519 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2523 * memory.high is breached and reclaim is unable to keep up. Throttle
2524 * allocators proactively to slow down excessive growth.
2526 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2527 mem_find_max_overage(memcg));
2529 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2530 swap_find_max_overage(memcg));
2533 * Clamp the max delay per usermode return so as to still keep the
2534 * application moving forwards and also permit diagnostics, albeit
2537 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2540 * Don't sleep if the amount of jiffies this memcg owes us is so low
2541 * that it's not even worth doing, in an attempt to be nice to those who
2542 * go only a small amount over their memory.high value and maybe haven't
2543 * been aggressively reclaimed enough yet.
2545 if (penalty_jiffies <= HZ / 100)
2549 * If reclaim is making forward progress but we're still over
2550 * memory.high, we want to encourage that rather than doing allocator
2553 if (nr_reclaimed || nr_retries--) {
2559 * If we exit early, we're guaranteed to die (since
2560 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2561 * need to account for any ill-begotten jiffies to pay them off later.
2563 psi_memstall_enter(&pflags);
2564 schedule_timeout_killable(penalty_jiffies);
2565 psi_memstall_leave(&pflags);
2568 css_put(&memcg->css);
2571 static int try_charge_memcg(struct mem_cgroup *memcg, gfp_t gfp_mask,
2572 unsigned int nr_pages)
2574 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2575 int nr_retries = MAX_RECLAIM_RETRIES;
2576 struct mem_cgroup *mem_over_limit;
2577 struct page_counter *counter;
2578 enum oom_status oom_status;
2579 unsigned long nr_reclaimed;
2580 bool passed_oom = false;
2581 bool may_swap = true;
2582 bool drained = false;
2583 unsigned long pflags;
2586 if (consume_stock(memcg, nr_pages))
2589 if (!do_memsw_account() ||
2590 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2591 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2593 if (do_memsw_account())
2594 page_counter_uncharge(&memcg->memsw, batch);
2595 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2597 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2601 if (batch > nr_pages) {
2607 * Memcg doesn't have a dedicated reserve for atomic
2608 * allocations. But like the global atomic pool, we need to
2609 * put the burden of reclaim on regular allocation requests
2610 * and let these go through as privileged allocations.
2612 if (gfp_mask & __GFP_ATOMIC)
2616 * Prevent unbounded recursion when reclaim operations need to
2617 * allocate memory. This might exceed the limits temporarily,
2618 * but we prefer facilitating memory reclaim and getting back
2619 * under the limit over triggering OOM kills in these cases.
2621 if (unlikely(current->flags & PF_MEMALLOC))
2624 if (unlikely(task_in_memcg_oom(current)))
2627 if (!gfpflags_allow_blocking(gfp_mask))
2630 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2632 psi_memstall_enter(&pflags);
2633 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2634 gfp_mask, may_swap);
2635 psi_memstall_leave(&pflags);
2637 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2641 drain_all_stock(mem_over_limit);
2646 if (gfp_mask & __GFP_NORETRY)
2649 * Even though the limit is exceeded at this point, reclaim
2650 * may have been able to free some pages. Retry the charge
2651 * before killing the task.
2653 * Only for regular pages, though: huge pages are rather
2654 * unlikely to succeed so close to the limit, and we fall back
2655 * to regular pages anyway in case of failure.
2657 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2660 * At task move, charge accounts can be doubly counted. So, it's
2661 * better to wait until the end of task_move if something is going on.
2663 if (mem_cgroup_wait_acct_move(mem_over_limit))
2669 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2672 /* Avoid endless loop for tasks bypassed by the oom killer */
2673 if (passed_oom && task_is_dying())
2677 * keep retrying as long as the memcg oom killer is able to make
2678 * a forward progress or bypass the charge if the oom killer
2679 * couldn't make any progress.
2681 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2682 get_order(nr_pages * PAGE_SIZE));
2683 if (oom_status == OOM_SUCCESS) {
2685 nr_retries = MAX_RECLAIM_RETRIES;
2689 if (!(gfp_mask & __GFP_NOFAIL))
2693 * The allocation either can't fail or will lead to more memory
2694 * being freed very soon. Allow memory usage go over the limit
2695 * temporarily by force charging it.
2697 page_counter_charge(&memcg->memory, nr_pages);
2698 if (do_memsw_account())
2699 page_counter_charge(&memcg->memsw, nr_pages);
2704 if (batch > nr_pages)
2705 refill_stock(memcg, batch - nr_pages);
2708 * If the hierarchy is above the normal consumption range, schedule
2709 * reclaim on returning to userland. We can perform reclaim here
2710 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2711 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2712 * not recorded as it most likely matches current's and won't
2713 * change in the meantime. As high limit is checked again before
2714 * reclaim, the cost of mismatch is negligible.
2717 bool mem_high, swap_high;
2719 mem_high = page_counter_read(&memcg->memory) >
2720 READ_ONCE(memcg->memory.high);
2721 swap_high = page_counter_read(&memcg->swap) >
2722 READ_ONCE(memcg->swap.high);
2724 /* Don't bother a random interrupted task */
2725 if (in_interrupt()) {
2727 schedule_work(&memcg->high_work);
2733 if (mem_high || swap_high) {
2735 * The allocating tasks in this cgroup will need to do
2736 * reclaim or be throttled to prevent further growth
2737 * of the memory or swap footprints.
2739 * Target some best-effort fairness between the tasks,
2740 * and distribute reclaim work and delay penalties
2741 * based on how much each task is actually allocating.
2743 current->memcg_nr_pages_over_high += batch;
2744 set_notify_resume(current);
2747 } while ((memcg = parent_mem_cgroup(memcg)));
2752 static inline int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2753 unsigned int nr_pages)
2755 if (mem_cgroup_is_root(memcg))
2758 return try_charge_memcg(memcg, gfp_mask, nr_pages);
2761 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2762 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2764 if (mem_cgroup_is_root(memcg))
2767 page_counter_uncharge(&memcg->memory, nr_pages);
2768 if (do_memsw_account())
2769 page_counter_uncharge(&memcg->memsw, nr_pages);
2773 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2775 VM_BUG_ON_PAGE(page_memcg(page), page);
2777 * Any of the following ensures page's memcg stability:
2781 * - lock_page_memcg()
2782 * - exclusive reference
2784 page->memcg_data = (unsigned long)memcg;
2787 static struct mem_cgroup *get_mem_cgroup_from_objcg(struct obj_cgroup *objcg)
2789 struct mem_cgroup *memcg;
2793 memcg = obj_cgroup_memcg(objcg);
2794 if (unlikely(!css_tryget(&memcg->css)))
2801 #ifdef CONFIG_MEMCG_KMEM
2803 * The allocated objcg pointers array is not accounted directly.
2804 * Moreover, it should not come from DMA buffer and is not readily
2805 * reclaimable. So those GFP bits should be masked off.
2807 #define OBJCGS_CLEAR_MASK (__GFP_DMA | __GFP_RECLAIMABLE | __GFP_ACCOUNT)
2809 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2810 gfp_t gfp, bool new_page)
2812 unsigned int objects = objs_per_slab_page(s, page);
2813 unsigned long memcg_data;
2816 gfp &= ~OBJCGS_CLEAR_MASK;
2817 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2822 memcg_data = (unsigned long) vec | MEMCG_DATA_OBJCGS;
2825 * If the slab page is brand new and nobody can yet access
2826 * it's memcg_data, no synchronization is required and
2827 * memcg_data can be simply assigned.
2829 page->memcg_data = memcg_data;
2830 } else if (cmpxchg(&page->memcg_data, 0, memcg_data)) {
2832 * If the slab page is already in use, somebody can allocate
2833 * and assign obj_cgroups in parallel. In this case the existing
2834 * objcg vector should be reused.
2840 kmemleak_not_leak(vec);
2845 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2847 * A passed kernel object can be a slab object or a generic kernel page, so
2848 * different mechanisms for getting the memory cgroup pointer should be used.
2849 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2850 * can not know for sure how the kernel object is implemented.
2851 * mem_cgroup_from_obj() can be safely used in such cases.
2853 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2854 * cgroup_mutex, etc.
2856 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2860 if (mem_cgroup_disabled())
2863 page = virt_to_head_page(p);
2866 * Slab objects are accounted individually, not per-page.
2867 * Memcg membership data for each individual object is saved in
2868 * the page->obj_cgroups.
2870 if (page_objcgs_check(page)) {
2871 struct obj_cgroup *objcg;
2874 off = obj_to_index(page->slab_cache, page, p);
2875 objcg = page_objcgs(page)[off];
2877 return obj_cgroup_memcg(objcg);
2883 * page_memcg_check() is used here, because page_has_obj_cgroups()
2884 * check above could fail because the object cgroups vector wasn't set
2885 * at that moment, but it can be set concurrently.
2886 * page_memcg_check(page) will guarantee that a proper memory
2887 * cgroup pointer or NULL will be returned.
2889 return page_memcg_check(page);
2892 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2894 struct obj_cgroup *objcg = NULL;
2895 struct mem_cgroup *memcg;
2897 if (memcg_kmem_bypass())
2901 if (unlikely(active_memcg()))
2902 memcg = active_memcg();
2904 memcg = mem_cgroup_from_task(current);
2906 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2907 objcg = rcu_dereference(memcg->objcg);
2908 if (objcg && obj_cgroup_tryget(objcg))
2917 static int memcg_alloc_cache_id(void)
2922 id = ida_simple_get(&memcg_cache_ida,
2923 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2927 if (id < memcg_nr_cache_ids)
2931 * There's no space for the new id in memcg_caches arrays,
2932 * so we have to grow them.
2934 down_write(&memcg_cache_ids_sem);
2936 size = 2 * (id + 1);
2937 if (size < MEMCG_CACHES_MIN_SIZE)
2938 size = MEMCG_CACHES_MIN_SIZE;
2939 else if (size > MEMCG_CACHES_MAX_SIZE)
2940 size = MEMCG_CACHES_MAX_SIZE;
2942 err = memcg_update_all_list_lrus(size);
2944 memcg_nr_cache_ids = size;
2946 up_write(&memcg_cache_ids_sem);
2949 ida_simple_remove(&memcg_cache_ida, id);
2955 static void memcg_free_cache_id(int id)
2957 ida_simple_remove(&memcg_cache_ida, id);
2961 * obj_cgroup_uncharge_pages: uncharge a number of kernel pages from a objcg
2962 * @objcg: object cgroup to uncharge
2963 * @nr_pages: number of pages to uncharge
2965 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
2966 unsigned int nr_pages)
2968 struct mem_cgroup *memcg;
2970 memcg = get_mem_cgroup_from_objcg(objcg);
2972 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2973 page_counter_uncharge(&memcg->kmem, nr_pages);
2974 refill_stock(memcg, nr_pages);
2976 css_put(&memcg->css);
2980 * obj_cgroup_charge_pages: charge a number of kernel pages to a objcg
2981 * @objcg: object cgroup to charge
2982 * @gfp: reclaim mode
2983 * @nr_pages: number of pages to charge
2985 * Returns 0 on success, an error code on failure.
2987 static int obj_cgroup_charge_pages(struct obj_cgroup *objcg, gfp_t gfp,
2988 unsigned int nr_pages)
2990 struct page_counter *counter;
2991 struct mem_cgroup *memcg;
2994 memcg = get_mem_cgroup_from_objcg(objcg);
2996 ret = try_charge_memcg(memcg, gfp, nr_pages);
3000 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3001 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3004 * Enforce __GFP_NOFAIL allocation because callers are not
3005 * prepared to see failures and likely do not have any failure
3008 if (gfp & __GFP_NOFAIL) {
3009 page_counter_charge(&memcg->kmem, nr_pages);
3012 cancel_charge(memcg, nr_pages);
3016 css_put(&memcg->css);
3022 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3023 * @page: page to charge
3024 * @gfp: reclaim mode
3025 * @order: allocation order
3027 * Returns 0 on success, an error code on failure.
3029 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3031 struct obj_cgroup *objcg;
3034 objcg = get_obj_cgroup_from_current();
3036 ret = obj_cgroup_charge_pages(objcg, gfp, 1 << order);
3038 page->memcg_data = (unsigned long)objcg |
3042 obj_cgroup_put(objcg);
3048 * __memcg_kmem_uncharge_page: uncharge a kmem page
3049 * @page: page to uncharge
3050 * @order: allocation order
3052 void __memcg_kmem_uncharge_page(struct page *page, int order)
3054 struct obj_cgroup *objcg;
3055 unsigned int nr_pages = 1 << order;
3057 if (!PageMemcgKmem(page))
3060 objcg = __page_objcg(page);
3061 obj_cgroup_uncharge_pages(objcg, nr_pages);
3062 page->memcg_data = 0;
3063 obj_cgroup_put(objcg);
3066 void mod_objcg_state(struct obj_cgroup *objcg, struct pglist_data *pgdat,
3067 enum node_stat_item idx, int nr)
3069 unsigned long flags;
3070 struct obj_stock *stock = get_obj_stock(&flags);
3074 * Save vmstat data in stock and skip vmstat array update unless
3075 * accumulating over a page of vmstat data or when pgdat or idx
3078 if (stock->cached_objcg != objcg) {
3079 drain_obj_stock(stock);
3080 obj_cgroup_get(objcg);
3081 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3082 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3083 stock->cached_objcg = objcg;
3084 stock->cached_pgdat = pgdat;
3085 } else if (stock->cached_pgdat != pgdat) {
3086 /* Flush the existing cached vmstat data */
3087 struct pglist_data *oldpg = stock->cached_pgdat;
3089 if (stock->nr_slab_reclaimable_b) {
3090 mod_objcg_mlstate(objcg, oldpg, NR_SLAB_RECLAIMABLE_B,
3091 stock->nr_slab_reclaimable_b);
3092 stock->nr_slab_reclaimable_b = 0;
3094 if (stock->nr_slab_unreclaimable_b) {
3095 mod_objcg_mlstate(objcg, oldpg, NR_SLAB_UNRECLAIMABLE_B,
3096 stock->nr_slab_unreclaimable_b);
3097 stock->nr_slab_unreclaimable_b = 0;
3099 stock->cached_pgdat = pgdat;
3102 bytes = (idx == NR_SLAB_RECLAIMABLE_B) ? &stock->nr_slab_reclaimable_b
3103 : &stock->nr_slab_unreclaimable_b;
3105 * Even for large object >= PAGE_SIZE, the vmstat data will still be
3106 * cached locally at least once before pushing it out.
3113 if (abs(*bytes) > PAGE_SIZE) {
3121 mod_objcg_mlstate(objcg, pgdat, idx, nr);
3123 put_obj_stock(flags);
3126 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3128 unsigned long flags;
3129 struct obj_stock *stock = get_obj_stock(&flags);
3132 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3133 stock->nr_bytes -= nr_bytes;
3137 put_obj_stock(flags);
3142 static void drain_obj_stock(struct obj_stock *stock)
3144 struct obj_cgroup *old = stock->cached_objcg;
3149 if (stock->nr_bytes) {
3150 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3151 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3154 obj_cgroup_uncharge_pages(old, nr_pages);
3157 * The leftover is flushed to the centralized per-memcg value.
3158 * On the next attempt to refill obj stock it will be moved
3159 * to a per-cpu stock (probably, on an other CPU), see
3160 * refill_obj_stock().
3162 * How often it's flushed is a trade-off between the memory
3163 * limit enforcement accuracy and potential CPU contention,
3164 * so it might be changed in the future.
3166 atomic_add(nr_bytes, &old->nr_charged_bytes);
3167 stock->nr_bytes = 0;
3171 * Flush the vmstat data in current stock
3173 if (stock->nr_slab_reclaimable_b || stock->nr_slab_unreclaimable_b) {
3174 if (stock->nr_slab_reclaimable_b) {
3175 mod_objcg_mlstate(old, stock->cached_pgdat,
3176 NR_SLAB_RECLAIMABLE_B,
3177 stock->nr_slab_reclaimable_b);
3178 stock->nr_slab_reclaimable_b = 0;
3180 if (stock->nr_slab_unreclaimable_b) {
3181 mod_objcg_mlstate(old, stock->cached_pgdat,
3182 NR_SLAB_UNRECLAIMABLE_B,
3183 stock->nr_slab_unreclaimable_b);
3184 stock->nr_slab_unreclaimable_b = 0;
3186 stock->cached_pgdat = NULL;
3189 obj_cgroup_put(old);
3190 stock->cached_objcg = NULL;
3193 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3194 struct mem_cgroup *root_memcg)
3196 struct mem_cgroup *memcg;
3198 if (in_task() && stock->task_obj.cached_objcg) {
3199 memcg = obj_cgroup_memcg(stock->task_obj.cached_objcg);
3200 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3203 if (stock->irq_obj.cached_objcg) {
3204 memcg = obj_cgroup_memcg(stock->irq_obj.cached_objcg);
3205 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3212 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes,
3213 bool allow_uncharge)
3215 unsigned long flags;
3216 struct obj_stock *stock = get_obj_stock(&flags);
3217 unsigned int nr_pages = 0;
3219 if (stock->cached_objcg != objcg) { /* reset if necessary */
3220 drain_obj_stock(stock);
3221 obj_cgroup_get(objcg);
3222 stock->cached_objcg = objcg;
3223 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3224 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3225 allow_uncharge = true; /* Allow uncharge when objcg changes */
3227 stock->nr_bytes += nr_bytes;
3229 if (allow_uncharge && (stock->nr_bytes > PAGE_SIZE)) {
3230 nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3231 stock->nr_bytes &= (PAGE_SIZE - 1);
3234 put_obj_stock(flags);
3237 obj_cgroup_uncharge_pages(objcg, nr_pages);
3240 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3242 unsigned int nr_pages, nr_bytes;
3245 if (consume_obj_stock(objcg, size))
3249 * In theory, objcg->nr_charged_bytes can have enough
3250 * pre-charged bytes to satisfy the allocation. However,
3251 * flushing objcg->nr_charged_bytes requires two atomic
3252 * operations, and objcg->nr_charged_bytes can't be big.
3253 * The shared objcg->nr_charged_bytes can also become a
3254 * performance bottleneck if all tasks of the same memcg are
3255 * trying to update it. So it's better to ignore it and try
3256 * grab some new pages. The stock's nr_bytes will be flushed to
3257 * objcg->nr_charged_bytes later on when objcg changes.
3259 * The stock's nr_bytes may contain enough pre-charged bytes
3260 * to allow one less page from being charged, but we can't rely
3261 * on the pre-charged bytes not being changed outside of
3262 * consume_obj_stock() or refill_obj_stock(). So ignore those
3263 * pre-charged bytes as well when charging pages. To avoid a
3264 * page uncharge right after a page charge, we set the
3265 * allow_uncharge flag to false when calling refill_obj_stock()
3266 * to temporarily allow the pre-charged bytes to exceed the page
3267 * size limit. The maximum reachable value of the pre-charged
3268 * bytes is (sizeof(object) + PAGE_SIZE - 2) if there is no data
3271 nr_pages = size >> PAGE_SHIFT;
3272 nr_bytes = size & (PAGE_SIZE - 1);
3277 ret = obj_cgroup_charge_pages(objcg, gfp, nr_pages);
3278 if (!ret && nr_bytes)
3279 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes, false);
3284 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3286 refill_obj_stock(objcg, size, true);
3289 #endif /* CONFIG_MEMCG_KMEM */
3292 * Because page_memcg(head) is not set on tails, set it now.
3294 void split_page_memcg(struct page *head, unsigned int nr)
3296 struct mem_cgroup *memcg = page_memcg(head);
3299 if (mem_cgroup_disabled() || !memcg)
3302 for (i = 1; i < nr; i++)
3303 head[i].memcg_data = head->memcg_data;
3305 if (PageMemcgKmem(head))
3306 obj_cgroup_get_many(__page_objcg(head), nr - 1);
3308 css_get_many(&memcg->css, nr - 1);
3311 #ifdef CONFIG_MEMCG_SWAP
3313 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3314 * @entry: swap entry to be moved
3315 * @from: mem_cgroup which the entry is moved from
3316 * @to: mem_cgroup which the entry is moved to
3318 * It succeeds only when the swap_cgroup's record for this entry is the same
3319 * as the mem_cgroup's id of @from.
3321 * Returns 0 on success, -EINVAL on failure.
3323 * The caller must have charged to @to, IOW, called page_counter_charge() about
3324 * both res and memsw, and called css_get().
3326 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3327 struct mem_cgroup *from, struct mem_cgroup *to)
3329 unsigned short old_id, new_id;
3331 old_id = mem_cgroup_id(from);
3332 new_id = mem_cgroup_id(to);
3334 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3335 mod_memcg_state(from, MEMCG_SWAP, -1);
3336 mod_memcg_state(to, MEMCG_SWAP, 1);
3342 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3343 struct mem_cgroup *from, struct mem_cgroup *to)
3349 static DEFINE_MUTEX(memcg_max_mutex);
3351 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3352 unsigned long max, bool memsw)
3354 bool enlarge = false;
3355 bool drained = false;
3357 bool limits_invariant;
3358 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3361 if (signal_pending(current)) {
3366 mutex_lock(&memcg_max_mutex);
3368 * Make sure that the new limit (memsw or memory limit) doesn't
3369 * break our basic invariant rule memory.max <= memsw.max.
3371 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3372 max <= memcg->memsw.max;
3373 if (!limits_invariant) {
3374 mutex_unlock(&memcg_max_mutex);
3378 if (max > counter->max)
3380 ret = page_counter_set_max(counter, max);
3381 mutex_unlock(&memcg_max_mutex);
3387 drain_all_stock(memcg);
3392 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3393 GFP_KERNEL, !memsw)) {
3399 if (!ret && enlarge)
3400 memcg_oom_recover(memcg);
3405 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3407 unsigned long *total_scanned)
3409 unsigned long nr_reclaimed = 0;
3410 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3411 unsigned long reclaimed;
3413 struct mem_cgroup_tree_per_node *mctz;
3414 unsigned long excess;
3415 unsigned long nr_scanned;
3420 mctz = soft_limit_tree_node(pgdat->node_id);
3423 * Do not even bother to check the largest node if the root
3424 * is empty. Do it lockless to prevent lock bouncing. Races
3425 * are acceptable as soft limit is best effort anyway.
3427 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3431 * This loop can run a while, specially if mem_cgroup's continuously
3432 * keep exceeding their soft limit and putting the system under
3439 mz = mem_cgroup_largest_soft_limit_node(mctz);
3444 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3445 gfp_mask, &nr_scanned);
3446 nr_reclaimed += reclaimed;
3447 *total_scanned += nr_scanned;
3448 spin_lock_irq(&mctz->lock);
3449 __mem_cgroup_remove_exceeded(mz, mctz);
3452 * If we failed to reclaim anything from this memory cgroup
3453 * it is time to move on to the next cgroup
3457 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3459 excess = soft_limit_excess(mz->memcg);
3461 * One school of thought says that we should not add
3462 * back the node to the tree if reclaim returns 0.
3463 * But our reclaim could return 0, simply because due
3464 * to priority we are exposing a smaller subset of
3465 * memory to reclaim from. Consider this as a longer
3468 /* If excess == 0, no tree ops */
3469 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3470 spin_unlock_irq(&mctz->lock);
3471 css_put(&mz->memcg->css);
3474 * Could not reclaim anything and there are no more
3475 * mem cgroups to try or we seem to be looping without
3476 * reclaiming anything.
3478 if (!nr_reclaimed &&
3480 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3482 } while (!nr_reclaimed);
3484 css_put(&next_mz->memcg->css);
3485 return nr_reclaimed;
3489 * Reclaims as many pages from the given memcg as possible.
3491 * Caller is responsible for holding css reference for memcg.
3493 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3495 int nr_retries = MAX_RECLAIM_RETRIES;
3497 /* we call try-to-free pages for make this cgroup empty */
3498 lru_add_drain_all();
3500 drain_all_stock(memcg);
3502 /* try to free all pages in this cgroup */
3503 while (nr_retries && page_counter_read(&memcg->memory)) {
3506 if (signal_pending(current))
3509 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3513 /* maybe some writeback is necessary */
3514 congestion_wait(BLK_RW_ASYNC, HZ/10);
3522 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3523 char *buf, size_t nbytes,
3526 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3528 if (mem_cgroup_is_root(memcg))
3530 return mem_cgroup_force_empty(memcg) ?: nbytes;
3533 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3539 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3540 struct cftype *cft, u64 val)
3545 pr_warn_once("Non-hierarchical mode is deprecated. "
3546 "Please report your usecase to linux-mm@kvack.org if you "
3547 "depend on this functionality.\n");
3552 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3556 if (mem_cgroup_is_root(memcg)) {
3557 /* mem_cgroup_threshold() calls here from irqsafe context */
3558 cgroup_rstat_flush_irqsafe(memcg->css.cgroup);
3559 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3560 memcg_page_state(memcg, NR_ANON_MAPPED);
3562 val += memcg_page_state(memcg, MEMCG_SWAP);
3565 val = page_counter_read(&memcg->memory);
3567 val = page_counter_read(&memcg->memsw);
3580 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3583 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3584 struct page_counter *counter;
3586 switch (MEMFILE_TYPE(cft->private)) {
3588 counter = &memcg->memory;
3591 counter = &memcg->memsw;
3594 counter = &memcg->kmem;
3597 counter = &memcg->tcpmem;
3603 switch (MEMFILE_ATTR(cft->private)) {
3605 if (counter == &memcg->memory)
3606 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3607 if (counter == &memcg->memsw)
3608 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3609 return (u64)page_counter_read(counter) * PAGE_SIZE;
3611 return (u64)counter->max * PAGE_SIZE;
3613 return (u64)counter->watermark * PAGE_SIZE;
3615 return counter->failcnt;
3616 case RES_SOFT_LIMIT:
3617 return (u64)memcg->soft_limit * PAGE_SIZE;
3623 #ifdef CONFIG_MEMCG_KMEM
3624 static int memcg_online_kmem(struct mem_cgroup *memcg)
3626 struct obj_cgroup *objcg;
3629 if (cgroup_memory_nokmem)
3632 BUG_ON(memcg->kmemcg_id >= 0);
3633 BUG_ON(memcg->kmem_state);
3635 memcg_id = memcg_alloc_cache_id();
3639 objcg = obj_cgroup_alloc();
3641 memcg_free_cache_id(memcg_id);
3644 objcg->memcg = memcg;
3645 rcu_assign_pointer(memcg->objcg, objcg);
3647 static_branch_enable(&memcg_kmem_enabled_key);
3649 memcg->kmemcg_id = memcg_id;
3650 memcg->kmem_state = KMEM_ONLINE;
3655 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3657 struct cgroup_subsys_state *css;
3658 struct mem_cgroup *parent, *child;
3661 if (memcg->kmem_state != KMEM_ONLINE)
3664 memcg->kmem_state = KMEM_ALLOCATED;
3666 parent = parent_mem_cgroup(memcg);
3668 parent = root_mem_cgroup;
3670 memcg_reparent_objcgs(memcg, parent);
3672 kmemcg_id = memcg->kmemcg_id;
3673 BUG_ON(kmemcg_id < 0);
3676 * Change kmemcg_id of this cgroup and all its descendants to the
3677 * parent's id, and then move all entries from this cgroup's list_lrus
3678 * to ones of the parent. After we have finished, all list_lrus
3679 * corresponding to this cgroup are guaranteed to remain empty. The
3680 * ordering is imposed by list_lru_node->lock taken by
3681 * memcg_drain_all_list_lrus().
3683 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3684 css_for_each_descendant_pre(css, &memcg->css) {
3685 child = mem_cgroup_from_css(css);
3686 BUG_ON(child->kmemcg_id != kmemcg_id);
3687 child->kmemcg_id = parent->kmemcg_id;
3691 memcg_drain_all_list_lrus(kmemcg_id, parent);
3693 memcg_free_cache_id(kmemcg_id);
3696 static void memcg_free_kmem(struct mem_cgroup *memcg)
3698 /* css_alloc() failed, offlining didn't happen */
3699 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3700 memcg_offline_kmem(memcg);
3703 static int memcg_online_kmem(struct mem_cgroup *memcg)
3707 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3710 static void memcg_free_kmem(struct mem_cgroup *memcg)
3713 #endif /* CONFIG_MEMCG_KMEM */
3715 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3720 mutex_lock(&memcg_max_mutex);
3721 ret = page_counter_set_max(&memcg->kmem, max);
3722 mutex_unlock(&memcg_max_mutex);
3726 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3730 mutex_lock(&memcg_max_mutex);
3732 ret = page_counter_set_max(&memcg->tcpmem, max);
3736 if (!memcg->tcpmem_active) {
3738 * The active flag needs to be written after the static_key
3739 * update. This is what guarantees that the socket activation
3740 * function is the last one to run. See mem_cgroup_sk_alloc()
3741 * for details, and note that we don't mark any socket as
3742 * belonging to this memcg until that flag is up.
3744 * We need to do this, because static_keys will span multiple
3745 * sites, but we can't control their order. If we mark a socket
3746 * as accounted, but the accounting functions are not patched in
3747 * yet, we'll lose accounting.
3749 * We never race with the readers in mem_cgroup_sk_alloc(),
3750 * because when this value change, the code to process it is not
3753 static_branch_inc(&memcg_sockets_enabled_key);
3754 memcg->tcpmem_active = true;
3757 mutex_unlock(&memcg_max_mutex);
3762 * The user of this function is...
3765 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3766 char *buf, size_t nbytes, loff_t off)
3768 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3769 unsigned long nr_pages;
3772 buf = strstrip(buf);
3773 ret = page_counter_memparse(buf, "-1", &nr_pages);
3777 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3779 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3783 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3785 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3788 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3791 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3792 "Please report your usecase to linux-mm@kvack.org if you "
3793 "depend on this functionality.\n");
3794 ret = memcg_update_kmem_max(memcg, nr_pages);
3797 ret = memcg_update_tcp_max(memcg, nr_pages);
3801 case RES_SOFT_LIMIT:
3802 memcg->soft_limit = nr_pages;
3806 return ret ?: nbytes;
3809 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3810 size_t nbytes, loff_t off)
3812 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3813 struct page_counter *counter;
3815 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3817 counter = &memcg->memory;
3820 counter = &memcg->memsw;
3823 counter = &memcg->kmem;
3826 counter = &memcg->tcpmem;
3832 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3834 page_counter_reset_watermark(counter);
3837 counter->failcnt = 0;
3846 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3849 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3853 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3854 struct cftype *cft, u64 val)
3856 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3858 if (val & ~MOVE_MASK)
3862 * No kind of locking is needed in here, because ->can_attach() will
3863 * check this value once in the beginning of the process, and then carry
3864 * on with stale data. This means that changes to this value will only
3865 * affect task migrations starting after the change.
3867 memcg->move_charge_at_immigrate = val;
3871 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3872 struct cftype *cft, u64 val)
3880 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3881 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3882 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3884 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3885 int nid, unsigned int lru_mask, bool tree)
3887 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3888 unsigned long nr = 0;
3891 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3894 if (!(BIT(lru) & lru_mask))
3897 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3899 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3904 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3905 unsigned int lru_mask,
3908 unsigned long nr = 0;
3912 if (!(BIT(lru) & lru_mask))
3915 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3917 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3922 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3926 unsigned int lru_mask;
3929 static const struct numa_stat stats[] = {
3930 { "total", LRU_ALL },
3931 { "file", LRU_ALL_FILE },
3932 { "anon", LRU_ALL_ANON },
3933 { "unevictable", BIT(LRU_UNEVICTABLE) },
3935 const struct numa_stat *stat;
3937 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3939 cgroup_rstat_flush(memcg->css.cgroup);
3941 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3942 seq_printf(m, "%s=%lu", stat->name,
3943 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3945 for_each_node_state(nid, N_MEMORY)
3946 seq_printf(m, " N%d=%lu", nid,
3947 mem_cgroup_node_nr_lru_pages(memcg, nid,
3948 stat->lru_mask, false));
3952 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3954 seq_printf(m, "hierarchical_%s=%lu", stat->name,
3955 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3957 for_each_node_state(nid, N_MEMORY)
3958 seq_printf(m, " N%d=%lu", nid,
3959 mem_cgroup_node_nr_lru_pages(memcg, nid,
3960 stat->lru_mask, true));
3966 #endif /* CONFIG_NUMA */
3968 static const unsigned int memcg1_stats[] = {
3971 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3981 static const char *const memcg1_stat_names[] = {
3984 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3994 /* Universal VM events cgroup1 shows, original sort order */
3995 static const unsigned int memcg1_events[] = {
4002 static int memcg_stat_show(struct seq_file *m, void *v)
4004 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4005 unsigned long memory, memsw;
4006 struct mem_cgroup *mi;
4009 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4011 cgroup_rstat_flush(memcg->css.cgroup);
4013 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4016 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4018 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4019 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4022 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4023 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4024 memcg_events_local(memcg, memcg1_events[i]));
4026 for (i = 0; i < NR_LRU_LISTS; i++)
4027 seq_printf(m, "%s %lu\n", lru_list_name(i),
4028 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4031 /* Hierarchical information */
4032 memory = memsw = PAGE_COUNTER_MAX;
4033 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4034 memory = min(memory, READ_ONCE(mi->memory.max));
4035 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4037 seq_printf(m, "hierarchical_memory_limit %llu\n",
4038 (u64)memory * PAGE_SIZE);
4039 if (do_memsw_account())
4040 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4041 (u64)memsw * PAGE_SIZE);
4043 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4046 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4048 nr = memcg_page_state(memcg, memcg1_stats[i]);
4049 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4050 (u64)nr * PAGE_SIZE);
4053 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4054 seq_printf(m, "total_%s %llu\n",
4055 vm_event_name(memcg1_events[i]),
4056 (u64)memcg_events(memcg, memcg1_events[i]));
4058 for (i = 0; i < NR_LRU_LISTS; i++)
4059 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4060 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4063 #ifdef CONFIG_DEBUG_VM
4066 struct mem_cgroup_per_node *mz;
4067 unsigned long anon_cost = 0;
4068 unsigned long file_cost = 0;
4070 for_each_online_pgdat(pgdat) {
4071 mz = memcg->nodeinfo[pgdat->node_id];
4073 anon_cost += mz->lruvec.anon_cost;
4074 file_cost += mz->lruvec.file_cost;
4076 seq_printf(m, "anon_cost %lu\n", anon_cost);
4077 seq_printf(m, "file_cost %lu\n", file_cost);
4084 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4087 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4089 return mem_cgroup_swappiness(memcg);
4092 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4093 struct cftype *cft, u64 val)
4095 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4100 if (!mem_cgroup_is_root(memcg))
4101 memcg->swappiness = val;
4103 vm_swappiness = val;
4108 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4110 struct mem_cgroup_threshold_ary *t;
4111 unsigned long usage;
4116 t = rcu_dereference(memcg->thresholds.primary);
4118 t = rcu_dereference(memcg->memsw_thresholds.primary);
4123 usage = mem_cgroup_usage(memcg, swap);
4126 * current_threshold points to threshold just below or equal to usage.
4127 * If it's not true, a threshold was crossed after last
4128 * call of __mem_cgroup_threshold().
4130 i = t->current_threshold;
4133 * Iterate backward over array of thresholds starting from
4134 * current_threshold and check if a threshold is crossed.
4135 * If none of thresholds below usage is crossed, we read
4136 * only one element of the array here.
4138 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4139 eventfd_signal(t->entries[i].eventfd, 1);
4141 /* i = current_threshold + 1 */
4145 * Iterate forward over array of thresholds starting from
4146 * current_threshold+1 and check if a threshold is crossed.
4147 * If none of thresholds above usage is crossed, we read
4148 * only one element of the array here.
4150 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4151 eventfd_signal(t->entries[i].eventfd, 1);
4153 /* Update current_threshold */
4154 t->current_threshold = i - 1;
4159 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4162 __mem_cgroup_threshold(memcg, false);
4163 if (do_memsw_account())
4164 __mem_cgroup_threshold(memcg, true);
4166 memcg = parent_mem_cgroup(memcg);
4170 static int compare_thresholds(const void *a, const void *b)
4172 const struct mem_cgroup_threshold *_a = a;
4173 const struct mem_cgroup_threshold *_b = b;
4175 if (_a->threshold > _b->threshold)
4178 if (_a->threshold < _b->threshold)
4184 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4186 struct mem_cgroup_eventfd_list *ev;
4188 spin_lock(&memcg_oom_lock);
4190 list_for_each_entry(ev, &memcg->oom_notify, list)
4191 eventfd_signal(ev->eventfd, 1);
4193 spin_unlock(&memcg_oom_lock);
4197 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4199 struct mem_cgroup *iter;
4201 for_each_mem_cgroup_tree(iter, memcg)
4202 mem_cgroup_oom_notify_cb(iter);
4205 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4206 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4208 struct mem_cgroup_thresholds *thresholds;
4209 struct mem_cgroup_threshold_ary *new;
4210 unsigned long threshold;
4211 unsigned long usage;
4214 ret = page_counter_memparse(args, "-1", &threshold);
4218 mutex_lock(&memcg->thresholds_lock);
4221 thresholds = &memcg->thresholds;
4222 usage = mem_cgroup_usage(memcg, false);
4223 } else if (type == _MEMSWAP) {
4224 thresholds = &memcg->memsw_thresholds;
4225 usage = mem_cgroup_usage(memcg, true);
4229 /* Check if a threshold crossed before adding a new one */
4230 if (thresholds->primary)
4231 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4233 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4235 /* Allocate memory for new array of thresholds */
4236 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4243 /* Copy thresholds (if any) to new array */
4244 if (thresholds->primary)
4245 memcpy(new->entries, thresholds->primary->entries,
4246 flex_array_size(new, entries, size - 1));
4248 /* Add new threshold */
4249 new->entries[size - 1].eventfd = eventfd;
4250 new->entries[size - 1].threshold = threshold;
4252 /* Sort thresholds. Registering of new threshold isn't time-critical */
4253 sort(new->entries, size, sizeof(*new->entries),
4254 compare_thresholds, NULL);
4256 /* Find current threshold */
4257 new->current_threshold = -1;
4258 for (i = 0; i < size; i++) {
4259 if (new->entries[i].threshold <= usage) {
4261 * new->current_threshold will not be used until
4262 * rcu_assign_pointer(), so it's safe to increment
4265 ++new->current_threshold;
4270 /* Free old spare buffer and save old primary buffer as spare */
4271 kfree(thresholds->spare);
4272 thresholds->spare = thresholds->primary;
4274 rcu_assign_pointer(thresholds->primary, new);
4276 /* To be sure that nobody uses thresholds */
4280 mutex_unlock(&memcg->thresholds_lock);
4285 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4286 struct eventfd_ctx *eventfd, const char *args)
4288 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4291 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4292 struct eventfd_ctx *eventfd, const char *args)
4294 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4297 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4298 struct eventfd_ctx *eventfd, enum res_type type)
4300 struct mem_cgroup_thresholds *thresholds;
4301 struct mem_cgroup_threshold_ary *new;
4302 unsigned long usage;
4303 int i, j, size, entries;
4305 mutex_lock(&memcg->thresholds_lock);
4308 thresholds = &memcg->thresholds;
4309 usage = mem_cgroup_usage(memcg, false);
4310 } else if (type == _MEMSWAP) {
4311 thresholds = &memcg->memsw_thresholds;
4312 usage = mem_cgroup_usage(memcg, true);
4316 if (!thresholds->primary)
4319 /* Check if a threshold crossed before removing */
4320 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4322 /* Calculate new number of threshold */
4324 for (i = 0; i < thresholds->primary->size; i++) {
4325 if (thresholds->primary->entries[i].eventfd != eventfd)
4331 new = thresholds->spare;
4333 /* If no items related to eventfd have been cleared, nothing to do */
4337 /* Set thresholds array to NULL if we don't have thresholds */
4346 /* Copy thresholds and find current threshold */
4347 new->current_threshold = -1;
4348 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4349 if (thresholds->primary->entries[i].eventfd == eventfd)
4352 new->entries[j] = thresholds->primary->entries[i];
4353 if (new->entries[j].threshold <= usage) {
4355 * new->current_threshold will not be used
4356 * until rcu_assign_pointer(), so it's safe to increment
4359 ++new->current_threshold;
4365 /* Swap primary and spare array */
4366 thresholds->spare = thresholds->primary;
4368 rcu_assign_pointer(thresholds->primary, new);
4370 /* To be sure that nobody uses thresholds */
4373 /* If all events are unregistered, free the spare array */
4375 kfree(thresholds->spare);
4376 thresholds->spare = NULL;
4379 mutex_unlock(&memcg->thresholds_lock);
4382 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4383 struct eventfd_ctx *eventfd)
4385 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4388 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4389 struct eventfd_ctx *eventfd)
4391 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4394 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4395 struct eventfd_ctx *eventfd, const char *args)
4397 struct mem_cgroup_eventfd_list *event;
4399 event = kmalloc(sizeof(*event), GFP_KERNEL);
4403 spin_lock(&memcg_oom_lock);
4405 event->eventfd = eventfd;
4406 list_add(&event->list, &memcg->oom_notify);
4408 /* already in OOM ? */
4409 if (memcg->under_oom)
4410 eventfd_signal(eventfd, 1);
4411 spin_unlock(&memcg_oom_lock);
4416 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4417 struct eventfd_ctx *eventfd)
4419 struct mem_cgroup_eventfd_list *ev, *tmp;
4421 spin_lock(&memcg_oom_lock);
4423 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4424 if (ev->eventfd == eventfd) {
4425 list_del(&ev->list);
4430 spin_unlock(&memcg_oom_lock);
4433 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4435 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4437 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4438 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4439 seq_printf(sf, "oom_kill %lu\n",
4440 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4444 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4445 struct cftype *cft, u64 val)
4447 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4449 /* cannot set to root cgroup and only 0 and 1 are allowed */
4450 if (mem_cgroup_is_root(memcg) || !((val == 0) || (val == 1)))
4453 memcg->oom_kill_disable = val;
4455 memcg_oom_recover(memcg);
4460 #ifdef CONFIG_CGROUP_WRITEBACK
4462 #include <trace/events/writeback.h>
4464 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4466 return wb_domain_init(&memcg->cgwb_domain, gfp);
4469 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4471 wb_domain_exit(&memcg->cgwb_domain);
4474 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4476 wb_domain_size_changed(&memcg->cgwb_domain);
4479 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4481 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4483 if (!memcg->css.parent)
4486 return &memcg->cgwb_domain;
4490 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4491 * @wb: bdi_writeback in question
4492 * @pfilepages: out parameter for number of file pages
4493 * @pheadroom: out parameter for number of allocatable pages according to memcg
4494 * @pdirty: out parameter for number of dirty pages
4495 * @pwriteback: out parameter for number of pages under writeback
4497 * Determine the numbers of file, headroom, dirty, and writeback pages in
4498 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4499 * is a bit more involved.
4501 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4502 * headroom is calculated as the lowest headroom of itself and the
4503 * ancestors. Note that this doesn't consider the actual amount of
4504 * available memory in the system. The caller should further cap
4505 * *@pheadroom accordingly.
4507 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4508 unsigned long *pheadroom, unsigned long *pdirty,
4509 unsigned long *pwriteback)
4511 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4512 struct mem_cgroup *parent;
4514 cgroup_rstat_flush_irqsafe(memcg->css.cgroup);
4516 *pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
4517 *pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
4518 *pfilepages = memcg_page_state(memcg, NR_INACTIVE_FILE) +
4519 memcg_page_state(memcg, NR_ACTIVE_FILE);
4521 *pheadroom = PAGE_COUNTER_MAX;
4522 while ((parent = parent_mem_cgroup(memcg))) {
4523 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4524 READ_ONCE(memcg->memory.high));
4525 unsigned long used = page_counter_read(&memcg->memory);
4527 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4533 * Foreign dirty flushing
4535 * There's an inherent mismatch between memcg and writeback. The former
4536 * tracks ownership per-page while the latter per-inode. This was a
4537 * deliberate design decision because honoring per-page ownership in the
4538 * writeback path is complicated, may lead to higher CPU and IO overheads
4539 * and deemed unnecessary given that write-sharing an inode across
4540 * different cgroups isn't a common use-case.
4542 * Combined with inode majority-writer ownership switching, this works well
4543 * enough in most cases but there are some pathological cases. For
4544 * example, let's say there are two cgroups A and B which keep writing to
4545 * different but confined parts of the same inode. B owns the inode and
4546 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4547 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4548 * triggering background writeback. A will be slowed down without a way to
4549 * make writeback of the dirty pages happen.
4551 * Conditions like the above can lead to a cgroup getting repeatedly and
4552 * severely throttled after making some progress after each
4553 * dirty_expire_interval while the underlying IO device is almost
4556 * Solving this problem completely requires matching the ownership tracking
4557 * granularities between memcg and writeback in either direction. However,
4558 * the more egregious behaviors can be avoided by simply remembering the
4559 * most recent foreign dirtying events and initiating remote flushes on
4560 * them when local writeback isn't enough to keep the memory clean enough.
4562 * The following two functions implement such mechanism. When a foreign
4563 * page - a page whose memcg and writeback ownerships don't match - is
4564 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4565 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4566 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4567 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4568 * foreign bdi_writebacks which haven't expired. Both the numbers of
4569 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4570 * limited to MEMCG_CGWB_FRN_CNT.
4572 * The mechanism only remembers IDs and doesn't hold any object references.
4573 * As being wrong occasionally doesn't matter, updates and accesses to the
4574 * records are lockless and racy.
4576 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4577 struct bdi_writeback *wb)
4579 struct mem_cgroup *memcg = page_memcg(page);
4580 struct memcg_cgwb_frn *frn;
4581 u64 now = get_jiffies_64();
4582 u64 oldest_at = now;
4586 trace_track_foreign_dirty(page, wb);
4589 * Pick the slot to use. If there is already a slot for @wb, keep
4590 * using it. If not replace the oldest one which isn't being
4593 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4594 frn = &memcg->cgwb_frn[i];
4595 if (frn->bdi_id == wb->bdi->id &&
4596 frn->memcg_id == wb->memcg_css->id)
4598 if (time_before64(frn->at, oldest_at) &&
4599 atomic_read(&frn->done.cnt) == 1) {
4601 oldest_at = frn->at;
4605 if (i < MEMCG_CGWB_FRN_CNT) {
4607 * Re-using an existing one. Update timestamp lazily to
4608 * avoid making the cacheline hot. We want them to be
4609 * reasonably up-to-date and significantly shorter than
4610 * dirty_expire_interval as that's what expires the record.
4611 * Use the shorter of 1s and dirty_expire_interval / 8.
4613 unsigned long update_intv =
4614 min_t(unsigned long, HZ,
4615 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4617 if (time_before64(frn->at, now - update_intv))
4619 } else if (oldest >= 0) {
4620 /* replace the oldest free one */
4621 frn = &memcg->cgwb_frn[oldest];
4622 frn->bdi_id = wb->bdi->id;
4623 frn->memcg_id = wb->memcg_css->id;
4628 /* issue foreign writeback flushes for recorded foreign dirtying events */
4629 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4631 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4632 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4633 u64 now = jiffies_64;
4636 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4637 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4640 * If the record is older than dirty_expire_interval,
4641 * writeback on it has already started. No need to kick it
4642 * off again. Also, don't start a new one if there's
4643 * already one in flight.
4645 if (time_after64(frn->at, now - intv) &&
4646 atomic_read(&frn->done.cnt) == 1) {
4648 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4649 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id,
4650 WB_REASON_FOREIGN_FLUSH,
4656 #else /* CONFIG_CGROUP_WRITEBACK */
4658 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4663 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4667 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4671 #endif /* CONFIG_CGROUP_WRITEBACK */
4674 * DO NOT USE IN NEW FILES.
4676 * "cgroup.event_control" implementation.
4678 * This is way over-engineered. It tries to support fully configurable
4679 * events for each user. Such level of flexibility is completely
4680 * unnecessary especially in the light of the planned unified hierarchy.
4682 * Please deprecate this and replace with something simpler if at all
4687 * Unregister event and free resources.
4689 * Gets called from workqueue.
4691 static void memcg_event_remove(struct work_struct *work)
4693 struct mem_cgroup_event *event =
4694 container_of(work, struct mem_cgroup_event, remove);
4695 struct mem_cgroup *memcg = event->memcg;
4697 remove_wait_queue(event->wqh, &event->wait);
4699 event->unregister_event(memcg, event->eventfd);
4701 /* Notify userspace the event is going away. */
4702 eventfd_signal(event->eventfd, 1);
4704 eventfd_ctx_put(event->eventfd);
4706 css_put(&memcg->css);
4710 * Gets called on EPOLLHUP on eventfd when user closes it.
4712 * Called with wqh->lock held and interrupts disabled.
4714 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4715 int sync, void *key)
4717 struct mem_cgroup_event *event =
4718 container_of(wait, struct mem_cgroup_event, wait);
4719 struct mem_cgroup *memcg = event->memcg;
4720 __poll_t flags = key_to_poll(key);
4722 if (flags & EPOLLHUP) {
4724 * If the event has been detached at cgroup removal, we
4725 * can simply return knowing the other side will cleanup
4728 * We can't race against event freeing since the other
4729 * side will require wqh->lock via remove_wait_queue(),
4732 spin_lock(&memcg->event_list_lock);
4733 if (!list_empty(&event->list)) {
4734 list_del_init(&event->list);
4736 * We are in atomic context, but cgroup_event_remove()
4737 * may sleep, so we have to call it in workqueue.
4739 schedule_work(&event->remove);
4741 spin_unlock(&memcg->event_list_lock);
4747 static void memcg_event_ptable_queue_proc(struct file *file,
4748 wait_queue_head_t *wqh, poll_table *pt)
4750 struct mem_cgroup_event *event =
4751 container_of(pt, struct mem_cgroup_event, pt);
4754 add_wait_queue(wqh, &event->wait);
4758 * DO NOT USE IN NEW FILES.
4760 * Parse input and register new cgroup event handler.
4762 * Input must be in format '<event_fd> <control_fd> <args>'.
4763 * Interpretation of args is defined by control file implementation.
4765 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4766 char *buf, size_t nbytes, loff_t off)
4768 struct cgroup_subsys_state *css = of_css(of);
4769 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4770 struct mem_cgroup_event *event;
4771 struct cgroup_subsys_state *cfile_css;
4772 unsigned int efd, cfd;
4779 buf = strstrip(buf);
4781 efd = simple_strtoul(buf, &endp, 10);
4786 cfd = simple_strtoul(buf, &endp, 10);
4787 if ((*endp != ' ') && (*endp != '\0'))
4791 event = kzalloc(sizeof(*event), GFP_KERNEL);
4795 event->memcg = memcg;
4796 INIT_LIST_HEAD(&event->list);
4797 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4798 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4799 INIT_WORK(&event->remove, memcg_event_remove);
4807 event->eventfd = eventfd_ctx_fileget(efile.file);
4808 if (IS_ERR(event->eventfd)) {
4809 ret = PTR_ERR(event->eventfd);
4816 goto out_put_eventfd;
4819 /* the process need read permission on control file */
4820 /* AV: shouldn't we check that it's been opened for read instead? */
4821 ret = file_permission(cfile.file, MAY_READ);
4826 * Determine the event callbacks and set them in @event. This used
4827 * to be done via struct cftype but cgroup core no longer knows
4828 * about these events. The following is crude but the whole thing
4829 * is for compatibility anyway.
4831 * DO NOT ADD NEW FILES.
4833 name = cfile.file->f_path.dentry->d_name.name;
4835 if (!strcmp(name, "memory.usage_in_bytes")) {
4836 event->register_event = mem_cgroup_usage_register_event;
4837 event->unregister_event = mem_cgroup_usage_unregister_event;
4838 } else if (!strcmp(name, "memory.oom_control")) {
4839 event->register_event = mem_cgroup_oom_register_event;
4840 event->unregister_event = mem_cgroup_oom_unregister_event;
4841 } else if (!strcmp(name, "memory.pressure_level")) {
4842 event->register_event = vmpressure_register_event;
4843 event->unregister_event = vmpressure_unregister_event;
4844 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4845 event->register_event = memsw_cgroup_usage_register_event;
4846 event->unregister_event = memsw_cgroup_usage_unregister_event;
4853 * Verify @cfile should belong to @css. Also, remaining events are
4854 * automatically removed on cgroup destruction but the removal is
4855 * asynchronous, so take an extra ref on @css.
4857 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4858 &memory_cgrp_subsys);
4860 if (IS_ERR(cfile_css))
4862 if (cfile_css != css) {
4867 ret = event->register_event(memcg, event->eventfd, buf);
4871 vfs_poll(efile.file, &event->pt);
4873 spin_lock_irq(&memcg->event_list_lock);
4874 list_add(&event->list, &memcg->event_list);
4875 spin_unlock_irq(&memcg->event_list_lock);
4887 eventfd_ctx_put(event->eventfd);
4896 static struct cftype mem_cgroup_legacy_files[] = {
4898 .name = "usage_in_bytes",
4899 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4900 .read_u64 = mem_cgroup_read_u64,
4903 .name = "max_usage_in_bytes",
4904 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4905 .write = mem_cgroup_reset,
4906 .read_u64 = mem_cgroup_read_u64,
4909 .name = "limit_in_bytes",
4910 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4911 .write = mem_cgroup_write,
4912 .read_u64 = mem_cgroup_read_u64,
4915 .name = "soft_limit_in_bytes",
4916 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4917 .write = mem_cgroup_write,
4918 .read_u64 = mem_cgroup_read_u64,
4922 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4923 .write = mem_cgroup_reset,
4924 .read_u64 = mem_cgroup_read_u64,
4928 .seq_show = memcg_stat_show,
4931 .name = "force_empty",
4932 .write = mem_cgroup_force_empty_write,
4935 .name = "use_hierarchy",
4936 .write_u64 = mem_cgroup_hierarchy_write,
4937 .read_u64 = mem_cgroup_hierarchy_read,
4940 .name = "cgroup.event_control", /* XXX: for compat */
4941 .write = memcg_write_event_control,
4942 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4945 .name = "swappiness",
4946 .read_u64 = mem_cgroup_swappiness_read,
4947 .write_u64 = mem_cgroup_swappiness_write,
4950 .name = "move_charge_at_immigrate",
4951 .read_u64 = mem_cgroup_move_charge_read,
4952 .write_u64 = mem_cgroup_move_charge_write,
4955 .name = "oom_control",
4956 .seq_show = mem_cgroup_oom_control_read,
4957 .write_u64 = mem_cgroup_oom_control_write,
4958 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4961 .name = "pressure_level",
4965 .name = "numa_stat",
4966 .seq_show = memcg_numa_stat_show,
4970 .name = "kmem.limit_in_bytes",
4971 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4972 .write = mem_cgroup_write,
4973 .read_u64 = mem_cgroup_read_u64,
4976 .name = "kmem.usage_in_bytes",
4977 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4978 .read_u64 = mem_cgroup_read_u64,
4981 .name = "kmem.failcnt",
4982 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4983 .write = mem_cgroup_reset,
4984 .read_u64 = mem_cgroup_read_u64,
4987 .name = "kmem.max_usage_in_bytes",
4988 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4989 .write = mem_cgroup_reset,
4990 .read_u64 = mem_cgroup_read_u64,
4992 #if defined(CONFIG_MEMCG_KMEM) && \
4993 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
4995 .name = "kmem.slabinfo",
4996 .seq_show = memcg_slab_show,
5000 .name = "kmem.tcp.limit_in_bytes",
5001 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5002 .write = mem_cgroup_write,
5003 .read_u64 = mem_cgroup_read_u64,
5006 .name = "kmem.tcp.usage_in_bytes",
5007 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5008 .read_u64 = mem_cgroup_read_u64,
5011 .name = "kmem.tcp.failcnt",
5012 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5013 .write = mem_cgroup_reset,
5014 .read_u64 = mem_cgroup_read_u64,
5017 .name = "kmem.tcp.max_usage_in_bytes",
5018 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5019 .write = mem_cgroup_reset,
5020 .read_u64 = mem_cgroup_read_u64,
5022 { }, /* terminate */
5026 * Private memory cgroup IDR
5028 * Swap-out records and page cache shadow entries need to store memcg
5029 * references in constrained space, so we maintain an ID space that is
5030 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5031 * memory-controlled cgroups to 64k.
5033 * However, there usually are many references to the offline CSS after
5034 * the cgroup has been destroyed, such as page cache or reclaimable
5035 * slab objects, that don't need to hang on to the ID. We want to keep
5036 * those dead CSS from occupying IDs, or we might quickly exhaust the
5037 * relatively small ID space and prevent the creation of new cgroups
5038 * even when there are much fewer than 64k cgroups - possibly none.
5040 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5041 * be freed and recycled when it's no longer needed, which is usually
5042 * when the CSS is offlined.
5044 * The only exception to that are records of swapped out tmpfs/shmem
5045 * pages that need to be attributed to live ancestors on swapin. But
5046 * those references are manageable from userspace.
5049 static DEFINE_IDR(mem_cgroup_idr);
5051 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5053 if (memcg->id.id > 0) {
5054 idr_remove(&mem_cgroup_idr, memcg->id.id);
5059 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5062 refcount_add(n, &memcg->id.ref);
5065 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5067 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5068 mem_cgroup_id_remove(memcg);
5070 /* Memcg ID pins CSS */
5071 css_put(&memcg->css);
5075 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5077 mem_cgroup_id_put_many(memcg, 1);
5081 * mem_cgroup_from_id - look up a memcg from a memcg id
5082 * @id: the memcg id to look up
5084 * Caller must hold rcu_read_lock().
5086 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5088 WARN_ON_ONCE(!rcu_read_lock_held());
5089 return idr_find(&mem_cgroup_idr, id);
5092 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5094 struct mem_cgroup_per_node *pn;
5097 * This routine is called against possible nodes.
5098 * But it's BUG to call kmalloc() against offline node.
5100 * TODO: this routine can waste much memory for nodes which will
5101 * never be onlined. It's better to use memory hotplug callback
5104 if (!node_state(node, N_NORMAL_MEMORY))
5106 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5110 pn->lruvec_stats_percpu = alloc_percpu_gfp(struct lruvec_stats_percpu,
5111 GFP_KERNEL_ACCOUNT);
5112 if (!pn->lruvec_stats_percpu) {
5117 lruvec_init(&pn->lruvec);
5118 pn->usage_in_excess = 0;
5119 pn->on_tree = false;
5122 memcg->nodeinfo[node] = pn;
5126 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5128 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5133 free_percpu(pn->lruvec_stats_percpu);
5137 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5142 free_mem_cgroup_per_node_info(memcg, node);
5143 free_percpu(memcg->vmstats_percpu);
5147 static void mem_cgroup_free(struct mem_cgroup *memcg)
5149 memcg_wb_domain_exit(memcg);
5150 __mem_cgroup_free(memcg);
5153 static struct mem_cgroup *mem_cgroup_alloc(void)
5155 struct mem_cgroup *memcg;
5158 int __maybe_unused i;
5159 long error = -ENOMEM;
5161 size = sizeof(struct mem_cgroup);
5162 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5164 memcg = kzalloc(size, GFP_KERNEL);
5166 return ERR_PTR(error);
5168 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5169 1, MEM_CGROUP_ID_MAX,
5171 if (memcg->id.id < 0) {
5172 error = memcg->id.id;
5176 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5177 GFP_KERNEL_ACCOUNT);
5178 if (!memcg->vmstats_percpu)
5182 if (alloc_mem_cgroup_per_node_info(memcg, node))
5185 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5188 INIT_WORK(&memcg->high_work, high_work_func);
5189 INIT_LIST_HEAD(&memcg->oom_notify);
5190 mutex_init(&memcg->thresholds_lock);
5191 spin_lock_init(&memcg->move_lock);
5192 vmpressure_init(&memcg->vmpressure);
5193 INIT_LIST_HEAD(&memcg->event_list);
5194 spin_lock_init(&memcg->event_list_lock);
5195 memcg->socket_pressure = jiffies;
5196 #ifdef CONFIG_MEMCG_KMEM
5197 memcg->kmemcg_id = -1;
5198 INIT_LIST_HEAD(&memcg->objcg_list);
5200 #ifdef CONFIG_CGROUP_WRITEBACK
5201 INIT_LIST_HEAD(&memcg->cgwb_list);
5202 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5203 memcg->cgwb_frn[i].done =
5204 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5206 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5207 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5208 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5209 memcg->deferred_split_queue.split_queue_len = 0;
5211 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5214 mem_cgroup_id_remove(memcg);
5215 __mem_cgroup_free(memcg);
5216 return ERR_PTR(error);
5219 static struct cgroup_subsys_state * __ref
5220 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5222 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5223 struct mem_cgroup *memcg, *old_memcg;
5224 long error = -ENOMEM;
5226 old_memcg = set_active_memcg(parent);
5227 memcg = mem_cgroup_alloc();
5228 set_active_memcg(old_memcg);
5230 return ERR_CAST(memcg);
5232 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5233 memcg->soft_limit = PAGE_COUNTER_MAX;
5234 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5236 memcg->swappiness = mem_cgroup_swappiness(parent);
5237 memcg->oom_kill_disable = parent->oom_kill_disable;
5239 page_counter_init(&memcg->memory, &parent->memory);
5240 page_counter_init(&memcg->swap, &parent->swap);
5241 page_counter_init(&memcg->kmem, &parent->kmem);
5242 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5244 page_counter_init(&memcg->memory, NULL);
5245 page_counter_init(&memcg->swap, NULL);
5246 page_counter_init(&memcg->kmem, NULL);
5247 page_counter_init(&memcg->tcpmem, NULL);
5249 root_mem_cgroup = memcg;
5253 /* The following stuff does not apply to the root */
5254 error = memcg_online_kmem(memcg);
5258 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5259 static_branch_inc(&memcg_sockets_enabled_key);
5263 mem_cgroup_id_remove(memcg);
5264 mem_cgroup_free(memcg);
5265 return ERR_PTR(error);
5268 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5270 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5273 * A memcg must be visible for expand_shrinker_info()
5274 * by the time the maps are allocated. So, we allocate maps
5275 * here, when for_each_mem_cgroup() can't skip it.
5277 if (alloc_shrinker_info(memcg)) {
5278 mem_cgroup_id_remove(memcg);
5282 /* Online state pins memcg ID, memcg ID pins CSS */
5283 refcount_set(&memcg->id.ref, 1);
5286 if (unlikely(mem_cgroup_is_root(memcg)))
5287 queue_delayed_work(system_unbound_wq, &stats_flush_dwork,
5292 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5294 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5295 struct mem_cgroup_event *event, *tmp;
5298 * Unregister events and notify userspace.
5299 * Notify userspace about cgroup removing only after rmdir of cgroup
5300 * directory to avoid race between userspace and kernelspace.
5302 spin_lock_irq(&memcg->event_list_lock);
5303 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5304 list_del_init(&event->list);
5305 schedule_work(&event->remove);
5307 spin_unlock_irq(&memcg->event_list_lock);
5309 page_counter_set_min(&memcg->memory, 0);
5310 page_counter_set_low(&memcg->memory, 0);
5312 memcg_offline_kmem(memcg);
5313 reparent_shrinker_deferred(memcg);
5314 wb_memcg_offline(memcg);
5316 drain_all_stock(memcg);
5318 mem_cgroup_id_put(memcg);
5321 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5323 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5325 invalidate_reclaim_iterators(memcg);
5328 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5330 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5331 int __maybe_unused i;
5333 #ifdef CONFIG_CGROUP_WRITEBACK
5334 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5335 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5337 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5338 static_branch_dec(&memcg_sockets_enabled_key);
5340 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5341 static_branch_dec(&memcg_sockets_enabled_key);
5343 vmpressure_cleanup(&memcg->vmpressure);
5344 cancel_work_sync(&memcg->high_work);
5345 mem_cgroup_remove_from_trees(memcg);
5346 free_shrinker_info(memcg);
5347 memcg_free_kmem(memcg);
5348 mem_cgroup_free(memcg);
5352 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5353 * @css: the target css
5355 * Reset the states of the mem_cgroup associated with @css. This is
5356 * invoked when the userland requests disabling on the default hierarchy
5357 * but the memcg is pinned through dependency. The memcg should stop
5358 * applying policies and should revert to the vanilla state as it may be
5359 * made visible again.
5361 * The current implementation only resets the essential configurations.
5362 * This needs to be expanded to cover all the visible parts.
5364 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5366 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5368 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5369 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5370 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5371 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5372 page_counter_set_min(&memcg->memory, 0);
5373 page_counter_set_low(&memcg->memory, 0);
5374 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5375 memcg->soft_limit = PAGE_COUNTER_MAX;
5376 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5377 memcg_wb_domain_size_changed(memcg);
5380 static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state *css, int cpu)
5382 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5383 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5384 struct memcg_vmstats_percpu *statc;
5388 statc = per_cpu_ptr(memcg->vmstats_percpu, cpu);
5390 for (i = 0; i < MEMCG_NR_STAT; i++) {
5392 * Collect the aggregated propagation counts of groups
5393 * below us. We're in a per-cpu loop here and this is
5394 * a global counter, so the first cycle will get them.
5396 delta = memcg->vmstats.state_pending[i];
5398 memcg->vmstats.state_pending[i] = 0;
5400 /* Add CPU changes on this level since the last flush */
5401 v = READ_ONCE(statc->state[i]);
5402 if (v != statc->state_prev[i]) {
5403 delta += v - statc->state_prev[i];
5404 statc->state_prev[i] = v;
5410 /* Aggregate counts on this level and propagate upwards */
5411 memcg->vmstats.state[i] += delta;
5413 parent->vmstats.state_pending[i] += delta;
5416 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
5417 delta = memcg->vmstats.events_pending[i];
5419 memcg->vmstats.events_pending[i] = 0;
5421 v = READ_ONCE(statc->events[i]);
5422 if (v != statc->events_prev[i]) {
5423 delta += v - statc->events_prev[i];
5424 statc->events_prev[i] = v;
5430 memcg->vmstats.events[i] += delta;
5432 parent->vmstats.events_pending[i] += delta;
5435 for_each_node_state(nid, N_MEMORY) {
5436 struct mem_cgroup_per_node *pn = memcg->nodeinfo[nid];
5437 struct mem_cgroup_per_node *ppn = NULL;
5438 struct lruvec_stats_percpu *lstatc;
5441 ppn = parent->nodeinfo[nid];
5443 lstatc = per_cpu_ptr(pn->lruvec_stats_percpu, cpu);
5445 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
5446 delta = pn->lruvec_stats.state_pending[i];
5448 pn->lruvec_stats.state_pending[i] = 0;
5450 v = READ_ONCE(lstatc->state[i]);
5451 if (v != lstatc->state_prev[i]) {
5452 delta += v - lstatc->state_prev[i];
5453 lstatc->state_prev[i] = v;
5459 pn->lruvec_stats.state[i] += delta;
5461 ppn->lruvec_stats.state_pending[i] += delta;
5467 /* Handlers for move charge at task migration. */
5468 static int mem_cgroup_do_precharge(unsigned long count)
5472 /* Try a single bulk charge without reclaim first, kswapd may wake */
5473 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5475 mc.precharge += count;
5479 /* Try charges one by one with reclaim, but do not retry */
5481 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5495 enum mc_target_type {
5502 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5503 unsigned long addr, pte_t ptent)
5505 struct page *page = vm_normal_page(vma, addr, ptent);
5507 if (!page || !page_mapped(page))
5509 if (PageAnon(page)) {
5510 if (!(mc.flags & MOVE_ANON))
5513 if (!(mc.flags & MOVE_FILE))
5516 if (!get_page_unless_zero(page))
5522 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5523 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5524 pte_t ptent, swp_entry_t *entry)
5526 struct page *page = NULL;
5527 swp_entry_t ent = pte_to_swp_entry(ptent);
5529 if (!(mc.flags & MOVE_ANON))
5533 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5534 * a device and because they are not accessible by CPU they are store
5535 * as special swap entry in the CPU page table.
5537 if (is_device_private_entry(ent)) {
5538 page = pfn_swap_entry_to_page(ent);
5540 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5541 * a refcount of 1 when free (unlike normal page)
5543 if (!page_ref_add_unless(page, 1, 1))
5548 if (non_swap_entry(ent))
5552 * Because lookup_swap_cache() updates some statistics counter,
5553 * we call find_get_page() with swapper_space directly.
5555 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5556 entry->val = ent.val;
5561 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5562 pte_t ptent, swp_entry_t *entry)
5568 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5569 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5571 if (!vma->vm_file) /* anonymous vma */
5573 if (!(mc.flags & MOVE_FILE))
5576 /* page is moved even if it's not RSS of this task(page-faulted). */
5577 /* shmem/tmpfs may report page out on swap: account for that too. */
5578 return find_get_incore_page(vma->vm_file->f_mapping,
5579 linear_page_index(vma, addr));
5583 * mem_cgroup_move_account - move account of the page
5585 * @compound: charge the page as compound or small page
5586 * @from: mem_cgroup which the page is moved from.
5587 * @to: mem_cgroup which the page is moved to. @from != @to.
5589 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5591 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5594 static int mem_cgroup_move_account(struct page *page,
5596 struct mem_cgroup *from,
5597 struct mem_cgroup *to)
5599 struct lruvec *from_vec, *to_vec;
5600 struct pglist_data *pgdat;
5601 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5604 VM_BUG_ON(from == to);
5605 VM_BUG_ON_PAGE(PageLRU(page), page);
5606 VM_BUG_ON(compound && !PageTransHuge(page));
5609 * Prevent mem_cgroup_migrate() from looking at
5610 * page's memory cgroup of its source page while we change it.
5613 if (!trylock_page(page))
5617 if (page_memcg(page) != from)
5620 pgdat = page_pgdat(page);
5621 from_vec = mem_cgroup_lruvec(from, pgdat);
5622 to_vec = mem_cgroup_lruvec(to, pgdat);
5624 lock_page_memcg(page);
5626 if (PageAnon(page)) {
5627 if (page_mapped(page)) {
5628 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5629 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5630 if (PageTransHuge(page)) {
5631 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5633 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5638 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5639 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5641 if (PageSwapBacked(page)) {
5642 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5643 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5646 if (page_mapped(page)) {
5647 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5648 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5651 if (PageDirty(page)) {
5652 struct address_space *mapping = page_mapping(page);
5654 if (mapping_can_writeback(mapping)) {
5655 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5657 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5663 if (PageWriteback(page)) {
5664 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5665 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5669 * All state has been migrated, let's switch to the new memcg.
5671 * It is safe to change page's memcg here because the page
5672 * is referenced, charged, isolated, and locked: we can't race
5673 * with (un)charging, migration, LRU putback, or anything else
5674 * that would rely on a stable page's memory cgroup.
5676 * Note that lock_page_memcg is a memcg lock, not a page lock,
5677 * to save space. As soon as we switch page's memory cgroup to a
5678 * new memcg that isn't locked, the above state can change
5679 * concurrently again. Make sure we're truly done with it.
5684 css_put(&from->css);
5686 page->memcg_data = (unsigned long)to;
5688 __unlock_page_memcg(from);
5692 local_irq_disable();
5693 mem_cgroup_charge_statistics(to, page, nr_pages);
5694 memcg_check_events(to, page);
5695 mem_cgroup_charge_statistics(from, page, -nr_pages);
5696 memcg_check_events(from, page);
5705 * get_mctgt_type - get target type of moving charge
5706 * @vma: the vma the pte to be checked belongs
5707 * @addr: the address corresponding to the pte to be checked
5708 * @ptent: the pte to be checked
5709 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5712 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5713 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5714 * move charge. if @target is not NULL, the page is stored in target->page
5715 * with extra refcnt got(Callers should handle it).
5716 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5717 * target for charge migration. if @target is not NULL, the entry is stored
5719 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5720 * (so ZONE_DEVICE page and thus not on the lru).
5721 * For now we such page is charge like a regular page would be as for all
5722 * intent and purposes it is just special memory taking the place of a
5725 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5727 * Called with pte lock held.
5730 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5731 unsigned long addr, pte_t ptent, union mc_target *target)
5733 struct page *page = NULL;
5734 enum mc_target_type ret = MC_TARGET_NONE;
5735 swp_entry_t ent = { .val = 0 };
5737 if (pte_present(ptent))
5738 page = mc_handle_present_pte(vma, addr, ptent);
5739 else if (is_swap_pte(ptent))
5740 page = mc_handle_swap_pte(vma, ptent, &ent);
5741 else if (pte_none(ptent))
5742 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5744 if (!page && !ent.val)
5748 * Do only loose check w/o serialization.
5749 * mem_cgroup_move_account() checks the page is valid or
5750 * not under LRU exclusion.
5752 if (page_memcg(page) == mc.from) {
5753 ret = MC_TARGET_PAGE;
5754 if (is_device_private_page(page))
5755 ret = MC_TARGET_DEVICE;
5757 target->page = page;
5759 if (!ret || !target)
5763 * There is a swap entry and a page doesn't exist or isn't charged.
5764 * But we cannot move a tail-page in a THP.
5766 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5767 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5768 ret = MC_TARGET_SWAP;
5775 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5777 * We don't consider PMD mapped swapping or file mapped pages because THP does
5778 * not support them for now.
5779 * Caller should make sure that pmd_trans_huge(pmd) is true.
5781 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5782 unsigned long addr, pmd_t pmd, union mc_target *target)
5784 struct page *page = NULL;
5785 enum mc_target_type ret = MC_TARGET_NONE;
5787 if (unlikely(is_swap_pmd(pmd))) {
5788 VM_BUG_ON(thp_migration_supported() &&
5789 !is_pmd_migration_entry(pmd));
5792 page = pmd_page(pmd);
5793 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5794 if (!(mc.flags & MOVE_ANON))
5796 if (page_memcg(page) == mc.from) {
5797 ret = MC_TARGET_PAGE;
5800 target->page = page;
5806 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5807 unsigned long addr, pmd_t pmd, union mc_target *target)
5809 return MC_TARGET_NONE;
5813 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5814 unsigned long addr, unsigned long end,
5815 struct mm_walk *walk)
5817 struct vm_area_struct *vma = walk->vma;
5821 ptl = pmd_trans_huge_lock(pmd, vma);
5824 * Note their can not be MC_TARGET_DEVICE for now as we do not
5825 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5826 * this might change.
5828 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5829 mc.precharge += HPAGE_PMD_NR;
5834 if (pmd_trans_unstable(pmd))
5836 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5837 for (; addr != end; pte++, addr += PAGE_SIZE)
5838 if (get_mctgt_type(vma, addr, *pte, NULL))
5839 mc.precharge++; /* increment precharge temporarily */
5840 pte_unmap_unlock(pte - 1, ptl);
5846 static const struct mm_walk_ops precharge_walk_ops = {
5847 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5850 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5852 unsigned long precharge;
5855 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5856 mmap_read_unlock(mm);
5858 precharge = mc.precharge;
5864 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5866 unsigned long precharge = mem_cgroup_count_precharge(mm);
5868 VM_BUG_ON(mc.moving_task);
5869 mc.moving_task = current;
5870 return mem_cgroup_do_precharge(precharge);
5873 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5874 static void __mem_cgroup_clear_mc(void)
5876 struct mem_cgroup *from = mc.from;
5877 struct mem_cgroup *to = mc.to;
5879 /* we must uncharge all the leftover precharges from mc.to */
5881 cancel_charge(mc.to, mc.precharge);
5885 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5886 * we must uncharge here.
5888 if (mc.moved_charge) {
5889 cancel_charge(mc.from, mc.moved_charge);
5890 mc.moved_charge = 0;
5892 /* we must fixup refcnts and charges */
5893 if (mc.moved_swap) {
5894 /* uncharge swap account from the old cgroup */
5895 if (!mem_cgroup_is_root(mc.from))
5896 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5898 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5901 * we charged both to->memory and to->memsw, so we
5902 * should uncharge to->memory.
5904 if (!mem_cgroup_is_root(mc.to))
5905 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5909 memcg_oom_recover(from);
5910 memcg_oom_recover(to);
5911 wake_up_all(&mc.waitq);
5914 static void mem_cgroup_clear_mc(void)
5916 struct mm_struct *mm = mc.mm;
5919 * we must clear moving_task before waking up waiters at the end of
5922 mc.moving_task = NULL;
5923 __mem_cgroup_clear_mc();
5924 spin_lock(&mc.lock);
5928 spin_unlock(&mc.lock);
5933 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5935 struct cgroup_subsys_state *css;
5936 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5937 struct mem_cgroup *from;
5938 struct task_struct *leader, *p;
5939 struct mm_struct *mm;
5940 unsigned long move_flags;
5943 /* charge immigration isn't supported on the default hierarchy */
5944 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5948 * Multi-process migrations only happen on the default hierarchy
5949 * where charge immigration is not used. Perform charge
5950 * immigration if @tset contains a leader and whine if there are
5954 cgroup_taskset_for_each_leader(leader, css, tset) {
5957 memcg = mem_cgroup_from_css(css);
5963 * We are now committed to this value whatever it is. Changes in this
5964 * tunable will only affect upcoming migrations, not the current one.
5965 * So we need to save it, and keep it going.
5967 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5971 from = mem_cgroup_from_task(p);
5973 VM_BUG_ON(from == memcg);
5975 mm = get_task_mm(p);
5978 /* We move charges only when we move a owner of the mm */
5979 if (mm->owner == p) {
5982 VM_BUG_ON(mc.precharge);
5983 VM_BUG_ON(mc.moved_charge);
5984 VM_BUG_ON(mc.moved_swap);
5986 spin_lock(&mc.lock);
5990 mc.flags = move_flags;
5991 spin_unlock(&mc.lock);
5992 /* We set mc.moving_task later */
5994 ret = mem_cgroup_precharge_mc(mm);
5996 mem_cgroup_clear_mc();
6003 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6006 mem_cgroup_clear_mc();
6009 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6010 unsigned long addr, unsigned long end,
6011 struct mm_walk *walk)
6014 struct vm_area_struct *vma = walk->vma;
6017 enum mc_target_type target_type;
6018 union mc_target target;
6021 ptl = pmd_trans_huge_lock(pmd, vma);
6023 if (mc.precharge < HPAGE_PMD_NR) {
6027 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6028 if (target_type == MC_TARGET_PAGE) {
6030 if (!isolate_lru_page(page)) {
6031 if (!mem_cgroup_move_account(page, true,
6033 mc.precharge -= HPAGE_PMD_NR;
6034 mc.moved_charge += HPAGE_PMD_NR;
6036 putback_lru_page(page);
6039 } else if (target_type == MC_TARGET_DEVICE) {
6041 if (!mem_cgroup_move_account(page, true,
6043 mc.precharge -= HPAGE_PMD_NR;
6044 mc.moved_charge += HPAGE_PMD_NR;
6052 if (pmd_trans_unstable(pmd))
6055 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6056 for (; addr != end; addr += PAGE_SIZE) {
6057 pte_t ptent = *(pte++);
6058 bool device = false;
6064 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6065 case MC_TARGET_DEVICE:
6068 case MC_TARGET_PAGE:
6071 * We can have a part of the split pmd here. Moving it
6072 * can be done but it would be too convoluted so simply
6073 * ignore such a partial THP and keep it in original
6074 * memcg. There should be somebody mapping the head.
6076 if (PageTransCompound(page))
6078 if (!device && isolate_lru_page(page))
6080 if (!mem_cgroup_move_account(page, false,
6083 /* we uncharge from mc.from later. */
6087 putback_lru_page(page);
6088 put: /* get_mctgt_type() gets the page */
6091 case MC_TARGET_SWAP:
6093 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6095 mem_cgroup_id_get_many(mc.to, 1);
6096 /* we fixup other refcnts and charges later. */
6104 pte_unmap_unlock(pte - 1, ptl);
6109 * We have consumed all precharges we got in can_attach().
6110 * We try charge one by one, but don't do any additional
6111 * charges to mc.to if we have failed in charge once in attach()
6114 ret = mem_cgroup_do_precharge(1);
6122 static const struct mm_walk_ops charge_walk_ops = {
6123 .pmd_entry = mem_cgroup_move_charge_pte_range,
6126 static void mem_cgroup_move_charge(void)
6128 lru_add_drain_all();
6130 * Signal lock_page_memcg() to take the memcg's move_lock
6131 * while we're moving its pages to another memcg. Then wait
6132 * for already started RCU-only updates to finish.
6134 atomic_inc(&mc.from->moving_account);
6137 if (unlikely(!mmap_read_trylock(mc.mm))) {
6139 * Someone who are holding the mmap_lock might be waiting in
6140 * waitq. So we cancel all extra charges, wake up all waiters,
6141 * and retry. Because we cancel precharges, we might not be able
6142 * to move enough charges, but moving charge is a best-effort
6143 * feature anyway, so it wouldn't be a big problem.
6145 __mem_cgroup_clear_mc();
6150 * When we have consumed all precharges and failed in doing
6151 * additional charge, the page walk just aborts.
6153 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6156 mmap_read_unlock(mc.mm);
6157 atomic_dec(&mc.from->moving_account);
6160 static void mem_cgroup_move_task(void)
6163 mem_cgroup_move_charge();
6164 mem_cgroup_clear_mc();
6167 #else /* !CONFIG_MMU */
6168 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6172 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6175 static void mem_cgroup_move_task(void)
6180 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6182 if (value == PAGE_COUNTER_MAX)
6183 seq_puts(m, "max\n");
6185 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6190 static u64 memory_current_read(struct cgroup_subsys_state *css,
6193 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6195 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6198 static int memory_min_show(struct seq_file *m, void *v)
6200 return seq_puts_memcg_tunable(m,
6201 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6204 static ssize_t memory_min_write(struct kernfs_open_file *of,
6205 char *buf, size_t nbytes, loff_t off)
6207 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6211 buf = strstrip(buf);
6212 err = page_counter_memparse(buf, "max", &min);
6216 page_counter_set_min(&memcg->memory, min);
6221 static int memory_low_show(struct seq_file *m, void *v)
6223 return seq_puts_memcg_tunable(m,
6224 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6227 static ssize_t memory_low_write(struct kernfs_open_file *of,
6228 char *buf, size_t nbytes, loff_t off)
6230 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6234 buf = strstrip(buf);
6235 err = page_counter_memparse(buf, "max", &low);
6239 page_counter_set_low(&memcg->memory, low);
6244 static int memory_high_show(struct seq_file *m, void *v)
6246 return seq_puts_memcg_tunable(m,
6247 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6250 static ssize_t memory_high_write(struct kernfs_open_file *of,
6251 char *buf, size_t nbytes, loff_t off)
6253 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6254 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6255 bool drained = false;
6259 buf = strstrip(buf);
6260 err = page_counter_memparse(buf, "max", &high);
6264 page_counter_set_high(&memcg->memory, high);
6267 unsigned long nr_pages = page_counter_read(&memcg->memory);
6268 unsigned long reclaimed;
6270 if (nr_pages <= high)
6273 if (signal_pending(current))
6277 drain_all_stock(memcg);
6282 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6285 if (!reclaimed && !nr_retries--)
6289 memcg_wb_domain_size_changed(memcg);
6293 static int memory_max_show(struct seq_file *m, void *v)
6295 return seq_puts_memcg_tunable(m,
6296 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6299 static ssize_t memory_max_write(struct kernfs_open_file *of,
6300 char *buf, size_t nbytes, loff_t off)
6302 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6303 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6304 bool drained = false;
6308 buf = strstrip(buf);
6309 err = page_counter_memparse(buf, "max", &max);
6313 xchg(&memcg->memory.max, max);
6316 unsigned long nr_pages = page_counter_read(&memcg->memory);
6318 if (nr_pages <= max)
6321 if (signal_pending(current))
6325 drain_all_stock(memcg);
6331 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6337 memcg_memory_event(memcg, MEMCG_OOM);
6338 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6342 memcg_wb_domain_size_changed(memcg);
6346 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6348 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6349 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6350 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6351 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6352 seq_printf(m, "oom_kill %lu\n",
6353 atomic_long_read(&events[MEMCG_OOM_KILL]));
6356 static int memory_events_show(struct seq_file *m, void *v)
6358 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6360 __memory_events_show(m, memcg->memory_events);
6364 static int memory_events_local_show(struct seq_file *m, void *v)
6366 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6368 __memory_events_show(m, memcg->memory_events_local);
6372 static int memory_stat_show(struct seq_file *m, void *v)
6374 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6377 buf = memory_stat_format(memcg);
6386 static inline unsigned long lruvec_page_state_output(struct lruvec *lruvec,
6389 return lruvec_page_state(lruvec, item) * memcg_page_state_unit(item);
6392 static int memory_numa_stat_show(struct seq_file *m, void *v)
6395 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6397 cgroup_rstat_flush(memcg->css.cgroup);
6399 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6402 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6405 seq_printf(m, "%s", memory_stats[i].name);
6406 for_each_node_state(nid, N_MEMORY) {
6408 struct lruvec *lruvec;
6410 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6411 size = lruvec_page_state_output(lruvec,
6412 memory_stats[i].idx);
6413 seq_printf(m, " N%d=%llu", nid, size);
6422 static int memory_oom_group_show(struct seq_file *m, void *v)
6424 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6426 seq_printf(m, "%d\n", memcg->oom_group);
6431 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6432 char *buf, size_t nbytes, loff_t off)
6434 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6437 buf = strstrip(buf);
6441 ret = kstrtoint(buf, 0, &oom_group);
6445 if (oom_group != 0 && oom_group != 1)
6448 memcg->oom_group = oom_group;
6453 static struct cftype memory_files[] = {
6456 .flags = CFTYPE_NOT_ON_ROOT,
6457 .read_u64 = memory_current_read,
6461 .flags = CFTYPE_NOT_ON_ROOT,
6462 .seq_show = memory_min_show,
6463 .write = memory_min_write,
6467 .flags = CFTYPE_NOT_ON_ROOT,
6468 .seq_show = memory_low_show,
6469 .write = memory_low_write,
6473 .flags = CFTYPE_NOT_ON_ROOT,
6474 .seq_show = memory_high_show,
6475 .write = memory_high_write,
6479 .flags = CFTYPE_NOT_ON_ROOT,
6480 .seq_show = memory_max_show,
6481 .write = memory_max_write,
6485 .flags = CFTYPE_NOT_ON_ROOT,
6486 .file_offset = offsetof(struct mem_cgroup, events_file),
6487 .seq_show = memory_events_show,
6490 .name = "events.local",
6491 .flags = CFTYPE_NOT_ON_ROOT,
6492 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6493 .seq_show = memory_events_local_show,
6497 .seq_show = memory_stat_show,
6501 .name = "numa_stat",
6502 .seq_show = memory_numa_stat_show,
6506 .name = "oom.group",
6507 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6508 .seq_show = memory_oom_group_show,
6509 .write = memory_oom_group_write,
6514 struct cgroup_subsys memory_cgrp_subsys = {
6515 .css_alloc = mem_cgroup_css_alloc,
6516 .css_online = mem_cgroup_css_online,
6517 .css_offline = mem_cgroup_css_offline,
6518 .css_released = mem_cgroup_css_released,
6519 .css_free = mem_cgroup_css_free,
6520 .css_reset = mem_cgroup_css_reset,
6521 .css_rstat_flush = mem_cgroup_css_rstat_flush,
6522 .can_attach = mem_cgroup_can_attach,
6523 .cancel_attach = mem_cgroup_cancel_attach,
6524 .post_attach = mem_cgroup_move_task,
6525 .dfl_cftypes = memory_files,
6526 .legacy_cftypes = mem_cgroup_legacy_files,
6531 * This function calculates an individual cgroup's effective
6532 * protection which is derived from its own memory.min/low, its
6533 * parent's and siblings' settings, as well as the actual memory
6534 * distribution in the tree.
6536 * The following rules apply to the effective protection values:
6538 * 1. At the first level of reclaim, effective protection is equal to
6539 * the declared protection in memory.min and memory.low.
6541 * 2. To enable safe delegation of the protection configuration, at
6542 * subsequent levels the effective protection is capped to the
6543 * parent's effective protection.
6545 * 3. To make complex and dynamic subtrees easier to configure, the
6546 * user is allowed to overcommit the declared protection at a given
6547 * level. If that is the case, the parent's effective protection is
6548 * distributed to the children in proportion to how much protection
6549 * they have declared and how much of it they are utilizing.
6551 * This makes distribution proportional, but also work-conserving:
6552 * if one cgroup claims much more protection than it uses memory,
6553 * the unused remainder is available to its siblings.
6555 * 4. Conversely, when the declared protection is undercommitted at a
6556 * given level, the distribution of the larger parental protection
6557 * budget is NOT proportional. A cgroup's protection from a sibling
6558 * is capped to its own memory.min/low setting.
6560 * 5. However, to allow protecting recursive subtrees from each other
6561 * without having to declare each individual cgroup's fixed share
6562 * of the ancestor's claim to protection, any unutilized -
6563 * "floating" - protection from up the tree is distributed in
6564 * proportion to each cgroup's *usage*. This makes the protection
6565 * neutral wrt sibling cgroups and lets them compete freely over
6566 * the shared parental protection budget, but it protects the
6567 * subtree as a whole from neighboring subtrees.
6569 * Note that 4. and 5. are not in conflict: 4. is about protecting
6570 * against immediate siblings whereas 5. is about protecting against
6571 * neighboring subtrees.
6573 static unsigned long effective_protection(unsigned long usage,
6574 unsigned long parent_usage,
6575 unsigned long setting,
6576 unsigned long parent_effective,
6577 unsigned long siblings_protected)
6579 unsigned long protected;
6582 protected = min(usage, setting);
6584 * If all cgroups at this level combined claim and use more
6585 * protection then what the parent affords them, distribute
6586 * shares in proportion to utilization.
6588 * We are using actual utilization rather than the statically
6589 * claimed protection in order to be work-conserving: claimed
6590 * but unused protection is available to siblings that would
6591 * otherwise get a smaller chunk than what they claimed.
6593 if (siblings_protected > parent_effective)
6594 return protected * parent_effective / siblings_protected;
6597 * Ok, utilized protection of all children is within what the
6598 * parent affords them, so we know whatever this child claims
6599 * and utilizes is effectively protected.
6601 * If there is unprotected usage beyond this value, reclaim
6602 * will apply pressure in proportion to that amount.
6604 * If there is unutilized protection, the cgroup will be fully
6605 * shielded from reclaim, but we do return a smaller value for
6606 * protection than what the group could enjoy in theory. This
6607 * is okay. With the overcommit distribution above, effective
6608 * protection is always dependent on how memory is actually
6609 * consumed among the siblings anyway.
6614 * If the children aren't claiming (all of) the protection
6615 * afforded to them by the parent, distribute the remainder in
6616 * proportion to the (unprotected) memory of each cgroup. That
6617 * way, cgroups that aren't explicitly prioritized wrt each
6618 * other compete freely over the allowance, but they are
6619 * collectively protected from neighboring trees.
6621 * We're using unprotected memory for the weight so that if
6622 * some cgroups DO claim explicit protection, we don't protect
6623 * the same bytes twice.
6625 * Check both usage and parent_usage against the respective
6626 * protected values. One should imply the other, but they
6627 * aren't read atomically - make sure the division is sane.
6629 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6631 if (parent_effective > siblings_protected &&
6632 parent_usage > siblings_protected &&
6633 usage > protected) {
6634 unsigned long unclaimed;
6636 unclaimed = parent_effective - siblings_protected;
6637 unclaimed *= usage - protected;
6638 unclaimed /= parent_usage - siblings_protected;
6647 * mem_cgroup_calculate_protection - check if memory consumption is in the normal range
6648 * @root: the top ancestor of the sub-tree being checked
6649 * @memcg: the memory cgroup to check
6651 * WARNING: This function is not stateless! It can only be used as part
6652 * of a top-down tree iteration, not for isolated queries.
6654 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6655 struct mem_cgroup *memcg)
6657 unsigned long usage, parent_usage;
6658 struct mem_cgroup *parent;
6660 if (mem_cgroup_disabled())
6664 root = root_mem_cgroup;
6667 * Effective values of the reclaim targets are ignored so they
6668 * can be stale. Have a look at mem_cgroup_protection for more
6670 * TODO: calculation should be more robust so that we do not need
6671 * that special casing.
6676 usage = page_counter_read(&memcg->memory);
6680 parent = parent_mem_cgroup(memcg);
6681 /* No parent means a non-hierarchical mode on v1 memcg */
6685 if (parent == root) {
6686 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6687 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6691 parent_usage = page_counter_read(&parent->memory);
6693 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6694 READ_ONCE(memcg->memory.min),
6695 READ_ONCE(parent->memory.emin),
6696 atomic_long_read(&parent->memory.children_min_usage)));
6698 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6699 READ_ONCE(memcg->memory.low),
6700 READ_ONCE(parent->memory.elow),
6701 atomic_long_read(&parent->memory.children_low_usage)));
6704 static int charge_memcg(struct page *page, struct mem_cgroup *memcg, gfp_t gfp)
6706 unsigned int nr_pages = thp_nr_pages(page);
6709 ret = try_charge(memcg, gfp, nr_pages);
6713 css_get(&memcg->css);
6714 commit_charge(page, memcg);
6716 local_irq_disable();
6717 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6718 memcg_check_events(memcg, page);
6725 * __mem_cgroup_charge - charge a newly allocated page to a cgroup
6726 * @page: page to charge
6727 * @mm: mm context of the victim
6728 * @gfp_mask: reclaim mode
6730 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6731 * pages according to @gfp_mask if necessary. if @mm is NULL, try to
6732 * charge to the active memcg.
6734 * Do not use this for pages allocated for swapin.
6736 * Returns 0 on success. Otherwise, an error code is returned.
6738 int __mem_cgroup_charge(struct page *page, struct mm_struct *mm,
6741 struct mem_cgroup *memcg;
6744 memcg = get_mem_cgroup_from_mm(mm);
6745 ret = charge_memcg(page, memcg, gfp_mask);
6746 css_put(&memcg->css);
6752 * mem_cgroup_swapin_charge_page - charge a newly allocated page for swapin
6753 * @page: page to charge
6754 * @mm: mm context of the victim
6755 * @gfp: reclaim mode
6756 * @entry: swap entry for which the page is allocated
6758 * This function charges a page allocated for swapin. Please call this before
6759 * adding the page to the swapcache.
6761 * Returns 0 on success. Otherwise, an error code is returned.
6763 int mem_cgroup_swapin_charge_page(struct page *page, struct mm_struct *mm,
6764 gfp_t gfp, swp_entry_t entry)
6766 struct mem_cgroup *memcg;
6770 if (mem_cgroup_disabled())
6773 id = lookup_swap_cgroup_id(entry);
6775 memcg = mem_cgroup_from_id(id);
6776 if (!memcg || !css_tryget_online(&memcg->css))
6777 memcg = get_mem_cgroup_from_mm(mm);
6780 ret = charge_memcg(page, memcg, gfp);
6782 css_put(&memcg->css);
6787 * mem_cgroup_swapin_uncharge_swap - uncharge swap slot
6788 * @entry: swap entry for which the page is charged
6790 * Call this function after successfully adding the charged page to swapcache.
6792 * Note: This function assumes the page for which swap slot is being uncharged
6795 void mem_cgroup_swapin_uncharge_swap(swp_entry_t entry)
6798 * Cgroup1's unified memory+swap counter has been charged with the
6799 * new swapcache page, finish the transfer by uncharging the swap
6800 * slot. The swap slot would also get uncharged when it dies, but
6801 * it can stick around indefinitely and we'd count the page twice
6804 * Cgroup2 has separate resource counters for memory and swap,
6805 * so this is a non-issue here. Memory and swap charge lifetimes
6806 * correspond 1:1 to page and swap slot lifetimes: we charge the
6807 * page to memory here, and uncharge swap when the slot is freed.
6809 if (!mem_cgroup_disabled() && do_memsw_account()) {
6811 * The swap entry might not get freed for a long time,
6812 * let's not wait for it. The page already received a
6813 * memory+swap charge, drop the swap entry duplicate.
6815 mem_cgroup_uncharge_swap(entry, 1);
6819 struct uncharge_gather {
6820 struct mem_cgroup *memcg;
6821 unsigned long nr_memory;
6822 unsigned long pgpgout;
6823 unsigned long nr_kmem;
6824 struct page *dummy_page;
6827 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6829 memset(ug, 0, sizeof(*ug));
6832 static void uncharge_batch(const struct uncharge_gather *ug)
6834 unsigned long flags;
6836 if (ug->nr_memory) {
6837 page_counter_uncharge(&ug->memcg->memory, ug->nr_memory);
6838 if (do_memsw_account())
6839 page_counter_uncharge(&ug->memcg->memsw, ug->nr_memory);
6840 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6841 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6842 memcg_oom_recover(ug->memcg);
6845 local_irq_save(flags);
6846 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6847 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_memory);
6848 memcg_check_events(ug->memcg, ug->dummy_page);
6849 local_irq_restore(flags);
6851 /* drop reference from uncharge_page */
6852 css_put(&ug->memcg->css);
6855 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6857 unsigned long nr_pages;
6858 struct mem_cgroup *memcg;
6859 struct obj_cgroup *objcg;
6860 bool use_objcg = PageMemcgKmem(page);
6862 VM_BUG_ON_PAGE(PageLRU(page), page);
6865 * Nobody should be changing or seriously looking at
6866 * page memcg or objcg at this point, we have fully
6867 * exclusive access to the page.
6870 objcg = __page_objcg(page);
6872 * This get matches the put at the end of the function and
6873 * kmem pages do not hold memcg references anymore.
6875 memcg = get_mem_cgroup_from_objcg(objcg);
6877 memcg = __page_memcg(page);
6883 if (ug->memcg != memcg) {
6886 uncharge_gather_clear(ug);
6889 ug->dummy_page = page;
6891 /* pairs with css_put in uncharge_batch */
6892 css_get(&memcg->css);
6895 nr_pages = compound_nr(page);
6898 ug->nr_memory += nr_pages;
6899 ug->nr_kmem += nr_pages;
6901 page->memcg_data = 0;
6902 obj_cgroup_put(objcg);
6904 /* LRU pages aren't accounted at the root level */
6905 if (!mem_cgroup_is_root(memcg))
6906 ug->nr_memory += nr_pages;
6909 page->memcg_data = 0;
6912 css_put(&memcg->css);
6916 * __mem_cgroup_uncharge - uncharge a page
6917 * @page: page to uncharge
6919 * Uncharge a page previously charged with __mem_cgroup_charge().
6921 void __mem_cgroup_uncharge(struct page *page)
6923 struct uncharge_gather ug;
6925 /* Don't touch page->lru of any random page, pre-check: */
6926 if (!page_memcg(page))
6929 uncharge_gather_clear(&ug);
6930 uncharge_page(page, &ug);
6931 uncharge_batch(&ug);
6935 * __mem_cgroup_uncharge_list - uncharge a list of page
6936 * @page_list: list of pages to uncharge
6938 * Uncharge a list of pages previously charged with
6939 * __mem_cgroup_charge().
6941 void __mem_cgroup_uncharge_list(struct list_head *page_list)
6943 struct uncharge_gather ug;
6946 uncharge_gather_clear(&ug);
6947 list_for_each_entry(page, page_list, lru)
6948 uncharge_page(page, &ug);
6950 uncharge_batch(&ug);
6954 * mem_cgroup_migrate - charge a page's replacement
6955 * @oldpage: currently circulating page
6956 * @newpage: replacement page
6958 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6959 * be uncharged upon free.
6961 * Both pages must be locked, @newpage->mapping must be set up.
6963 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6965 struct mem_cgroup *memcg;
6966 unsigned int nr_pages;
6967 unsigned long flags;
6969 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6970 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6971 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6972 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6975 if (mem_cgroup_disabled())
6978 /* Page cache replacement: new page already charged? */
6979 if (page_memcg(newpage))
6982 memcg = page_memcg(oldpage);
6983 VM_WARN_ON_ONCE_PAGE(!memcg, oldpage);
6987 /* Force-charge the new page. The old one will be freed soon */
6988 nr_pages = thp_nr_pages(newpage);
6990 if (!mem_cgroup_is_root(memcg)) {
6991 page_counter_charge(&memcg->memory, nr_pages);
6992 if (do_memsw_account())
6993 page_counter_charge(&memcg->memsw, nr_pages);
6996 css_get(&memcg->css);
6997 commit_charge(newpage, memcg);
6999 local_irq_save(flags);
7000 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
7001 memcg_check_events(memcg, newpage);
7002 local_irq_restore(flags);
7005 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
7006 EXPORT_SYMBOL(memcg_sockets_enabled_key);
7008 void mem_cgroup_sk_alloc(struct sock *sk)
7010 struct mem_cgroup *memcg;
7012 if (!mem_cgroup_sockets_enabled)
7015 /* Do not associate the sock with unrelated interrupted task's memcg. */
7020 memcg = mem_cgroup_from_task(current);
7021 if (memcg == root_mem_cgroup)
7023 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
7025 if (css_tryget(&memcg->css))
7026 sk->sk_memcg = memcg;
7031 void mem_cgroup_sk_free(struct sock *sk)
7034 css_put(&sk->sk_memcg->css);
7038 * mem_cgroup_charge_skmem - charge socket memory
7039 * @memcg: memcg to charge
7040 * @nr_pages: number of pages to charge
7041 * @gfp_mask: reclaim mode
7043 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7044 * @memcg's configured limit, %false if it doesn't.
7046 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages,
7049 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7050 struct page_counter *fail;
7052 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7053 memcg->tcpmem_pressure = 0;
7056 memcg->tcpmem_pressure = 1;
7057 if (gfp_mask & __GFP_NOFAIL) {
7058 page_counter_charge(&memcg->tcpmem, nr_pages);
7064 if (try_charge(memcg, gfp_mask, nr_pages) == 0) {
7065 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7073 * mem_cgroup_uncharge_skmem - uncharge socket memory
7074 * @memcg: memcg to uncharge
7075 * @nr_pages: number of pages to uncharge
7077 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7079 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7080 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7084 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7086 refill_stock(memcg, nr_pages);
7089 static int __init cgroup_memory(char *s)
7093 while ((token = strsep(&s, ",")) != NULL) {
7096 if (!strcmp(token, "nosocket"))
7097 cgroup_memory_nosocket = true;
7098 if (!strcmp(token, "nokmem"))
7099 cgroup_memory_nokmem = true;
7103 __setup("cgroup.memory=", cgroup_memory);
7106 * subsys_initcall() for memory controller.
7108 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7109 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7110 * basically everything that doesn't depend on a specific mem_cgroup structure
7111 * should be initialized from here.
7113 static int __init mem_cgroup_init(void)
7118 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7119 * used for per-memcg-per-cpu caching of per-node statistics. In order
7120 * to work fine, we should make sure that the overfill threshold can't
7121 * exceed S32_MAX / PAGE_SIZE.
7123 BUILD_BUG_ON(MEMCG_CHARGE_BATCH > S32_MAX / PAGE_SIZE);
7125 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7126 memcg_hotplug_cpu_dead);
7128 for_each_possible_cpu(cpu)
7129 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7132 for_each_node(node) {
7133 struct mem_cgroup_tree_per_node *rtpn;
7135 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7136 node_online(node) ? node : NUMA_NO_NODE);
7138 rtpn->rb_root = RB_ROOT;
7139 rtpn->rb_rightmost = NULL;
7140 spin_lock_init(&rtpn->lock);
7141 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7146 subsys_initcall(mem_cgroup_init);
7148 #ifdef CONFIG_MEMCG_SWAP
7149 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7151 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7153 * The root cgroup cannot be destroyed, so it's refcount must
7156 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7160 memcg = parent_mem_cgroup(memcg);
7162 memcg = root_mem_cgroup;
7168 * mem_cgroup_swapout - transfer a memsw charge to swap
7169 * @page: page whose memsw charge to transfer
7170 * @entry: swap entry to move the charge to
7172 * Transfer the memsw charge of @page to @entry.
7174 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7176 struct mem_cgroup *memcg, *swap_memcg;
7177 unsigned int nr_entries;
7178 unsigned short oldid;
7180 VM_BUG_ON_PAGE(PageLRU(page), page);
7181 VM_BUG_ON_PAGE(page_count(page), page);
7183 if (mem_cgroup_disabled())
7186 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7189 memcg = page_memcg(page);
7191 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7196 * In case the memcg owning these pages has been offlined and doesn't
7197 * have an ID allocated to it anymore, charge the closest online
7198 * ancestor for the swap instead and transfer the memory+swap charge.
7200 swap_memcg = mem_cgroup_id_get_online(memcg);
7201 nr_entries = thp_nr_pages(page);
7202 /* Get references for the tail pages, too */
7204 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7205 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7207 VM_BUG_ON_PAGE(oldid, page);
7208 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7210 page->memcg_data = 0;
7212 if (!mem_cgroup_is_root(memcg))
7213 page_counter_uncharge(&memcg->memory, nr_entries);
7215 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7216 if (!mem_cgroup_is_root(swap_memcg))
7217 page_counter_charge(&swap_memcg->memsw, nr_entries);
7218 page_counter_uncharge(&memcg->memsw, nr_entries);
7222 * Interrupts should be disabled here because the caller holds the
7223 * i_pages lock which is taken with interrupts-off. It is
7224 * important here to have the interrupts disabled because it is the
7225 * only synchronisation we have for updating the per-CPU variables.
7227 VM_BUG_ON(!irqs_disabled());
7228 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7229 memcg_check_events(memcg, page);
7231 css_put(&memcg->css);
7235 * __mem_cgroup_try_charge_swap - try charging swap space for a page
7236 * @page: page being added to swap
7237 * @entry: swap entry to charge
7239 * Try to charge @page's memcg for the swap space at @entry.
7241 * Returns 0 on success, -ENOMEM on failure.
7243 int __mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7245 unsigned int nr_pages = thp_nr_pages(page);
7246 struct page_counter *counter;
7247 struct mem_cgroup *memcg;
7248 unsigned short oldid;
7250 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7253 memcg = page_memcg(page);
7255 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7260 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7264 memcg = mem_cgroup_id_get_online(memcg);
7266 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7267 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7268 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7269 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7270 mem_cgroup_id_put(memcg);
7274 /* Get references for the tail pages, too */
7276 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7277 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7278 VM_BUG_ON_PAGE(oldid, page);
7279 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7285 * __mem_cgroup_uncharge_swap - uncharge swap space
7286 * @entry: swap entry to uncharge
7287 * @nr_pages: the amount of swap space to uncharge
7289 void __mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7291 struct mem_cgroup *memcg;
7294 id = swap_cgroup_record(entry, 0, nr_pages);
7296 memcg = mem_cgroup_from_id(id);
7298 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7299 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7300 page_counter_uncharge(&memcg->swap, nr_pages);
7302 page_counter_uncharge(&memcg->memsw, nr_pages);
7304 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7305 mem_cgroup_id_put_many(memcg, nr_pages);
7310 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7312 long nr_swap_pages = get_nr_swap_pages();
7314 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7315 return nr_swap_pages;
7316 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7317 nr_swap_pages = min_t(long, nr_swap_pages,
7318 READ_ONCE(memcg->swap.max) -
7319 page_counter_read(&memcg->swap));
7320 return nr_swap_pages;
7323 bool mem_cgroup_swap_full(struct page *page)
7325 struct mem_cgroup *memcg;
7327 VM_BUG_ON_PAGE(!PageLocked(page), page);
7331 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7334 memcg = page_memcg(page);
7338 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7339 unsigned long usage = page_counter_read(&memcg->swap);
7341 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7342 usage * 2 >= READ_ONCE(memcg->swap.max))
7349 static int __init setup_swap_account(char *s)
7351 if (!strcmp(s, "1"))
7352 cgroup_memory_noswap = false;
7353 else if (!strcmp(s, "0"))
7354 cgroup_memory_noswap = true;
7357 __setup("swapaccount=", setup_swap_account);
7359 static u64 swap_current_read(struct cgroup_subsys_state *css,
7362 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7364 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7367 static int swap_high_show(struct seq_file *m, void *v)
7369 return seq_puts_memcg_tunable(m,
7370 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7373 static ssize_t swap_high_write(struct kernfs_open_file *of,
7374 char *buf, size_t nbytes, loff_t off)
7376 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7380 buf = strstrip(buf);
7381 err = page_counter_memparse(buf, "max", &high);
7385 page_counter_set_high(&memcg->swap, high);
7390 static int swap_max_show(struct seq_file *m, void *v)
7392 return seq_puts_memcg_tunable(m,
7393 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7396 static ssize_t swap_max_write(struct kernfs_open_file *of,
7397 char *buf, size_t nbytes, loff_t off)
7399 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7403 buf = strstrip(buf);
7404 err = page_counter_memparse(buf, "max", &max);
7408 xchg(&memcg->swap.max, max);
7413 static int swap_events_show(struct seq_file *m, void *v)
7415 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7417 seq_printf(m, "high %lu\n",
7418 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7419 seq_printf(m, "max %lu\n",
7420 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7421 seq_printf(m, "fail %lu\n",
7422 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7427 static struct cftype swap_files[] = {
7429 .name = "swap.current",
7430 .flags = CFTYPE_NOT_ON_ROOT,
7431 .read_u64 = swap_current_read,
7434 .name = "swap.high",
7435 .flags = CFTYPE_NOT_ON_ROOT,
7436 .seq_show = swap_high_show,
7437 .write = swap_high_write,
7441 .flags = CFTYPE_NOT_ON_ROOT,
7442 .seq_show = swap_max_show,
7443 .write = swap_max_write,
7446 .name = "swap.events",
7447 .flags = CFTYPE_NOT_ON_ROOT,
7448 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7449 .seq_show = swap_events_show,
7454 static struct cftype memsw_files[] = {
7456 .name = "memsw.usage_in_bytes",
7457 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7458 .read_u64 = mem_cgroup_read_u64,
7461 .name = "memsw.max_usage_in_bytes",
7462 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7463 .write = mem_cgroup_reset,
7464 .read_u64 = mem_cgroup_read_u64,
7467 .name = "memsw.limit_in_bytes",
7468 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7469 .write = mem_cgroup_write,
7470 .read_u64 = mem_cgroup_read_u64,
7473 .name = "memsw.failcnt",
7474 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7475 .write = mem_cgroup_reset,
7476 .read_u64 = mem_cgroup_read_u64,
7478 { }, /* terminate */
7482 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7483 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7484 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7485 * boot parameter. This may result in premature OOPS inside
7486 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7488 static int __init mem_cgroup_swap_init(void)
7490 /* No memory control -> no swap control */
7491 if (mem_cgroup_disabled())
7492 cgroup_memory_noswap = true;
7494 if (cgroup_memory_noswap)
7497 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7498 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7502 core_initcall(mem_cgroup_swap_init);
7504 #endif /* CONFIG_MEMCG_SWAP */