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 void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
455 struct mem_cgroup_tree_per_node *mctz,
456 unsigned long new_usage_in_excess)
458 struct rb_node **p = &mctz->rb_root.rb_node;
459 struct rb_node *parent = NULL;
460 struct mem_cgroup_per_node *mz_node;
461 bool rightmost = true;
466 mz->usage_in_excess = new_usage_in_excess;
467 if (!mz->usage_in_excess)
471 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
473 if (mz->usage_in_excess < mz_node->usage_in_excess) {
482 mctz->rb_rightmost = &mz->tree_node;
484 rb_link_node(&mz->tree_node, parent, p);
485 rb_insert_color(&mz->tree_node, &mctz->rb_root);
489 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
490 struct mem_cgroup_tree_per_node *mctz)
495 if (&mz->tree_node == mctz->rb_rightmost)
496 mctz->rb_rightmost = rb_prev(&mz->tree_node);
498 rb_erase(&mz->tree_node, &mctz->rb_root);
502 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
503 struct mem_cgroup_tree_per_node *mctz)
507 spin_lock_irqsave(&mctz->lock, flags);
508 __mem_cgroup_remove_exceeded(mz, mctz);
509 spin_unlock_irqrestore(&mctz->lock, flags);
512 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
514 unsigned long nr_pages = page_counter_read(&memcg->memory);
515 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
516 unsigned long excess = 0;
518 if (nr_pages > soft_limit)
519 excess = nr_pages - soft_limit;
524 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, int nid)
526 unsigned long excess;
527 struct mem_cgroup_per_node *mz;
528 struct mem_cgroup_tree_per_node *mctz;
530 mctz = soft_limit_tree.rb_tree_per_node[nid];
534 * Necessary to update all ancestors when hierarchy is used.
535 * because their event counter is not touched.
537 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
538 mz = memcg->nodeinfo[nid];
539 excess = soft_limit_excess(memcg);
541 * We have to update the tree if mz is on RB-tree or
542 * mem is over its softlimit.
544 if (excess || mz->on_tree) {
547 spin_lock_irqsave(&mctz->lock, flags);
548 /* if on-tree, remove it */
550 __mem_cgroup_remove_exceeded(mz, mctz);
552 * Insert again. mz->usage_in_excess will be updated.
553 * If excess is 0, no tree ops.
555 __mem_cgroup_insert_exceeded(mz, mctz, excess);
556 spin_unlock_irqrestore(&mctz->lock, flags);
561 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
563 struct mem_cgroup_tree_per_node *mctz;
564 struct mem_cgroup_per_node *mz;
568 mz = memcg->nodeinfo[nid];
569 mctz = soft_limit_tree.rb_tree_per_node[nid];
571 mem_cgroup_remove_exceeded(mz, mctz);
575 static struct mem_cgroup_per_node *
576 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
578 struct mem_cgroup_per_node *mz;
582 if (!mctz->rb_rightmost)
583 goto done; /* Nothing to reclaim from */
585 mz = rb_entry(mctz->rb_rightmost,
586 struct mem_cgroup_per_node, tree_node);
588 * Remove the node now but someone else can add it back,
589 * we will to add it back at the end of reclaim to its correct
590 * position in the tree.
592 __mem_cgroup_remove_exceeded(mz, mctz);
593 if (!soft_limit_excess(mz->memcg) ||
594 !css_tryget(&mz->memcg->css))
600 static struct mem_cgroup_per_node *
601 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
603 struct mem_cgroup_per_node *mz;
605 spin_lock_irq(&mctz->lock);
606 mz = __mem_cgroup_largest_soft_limit_node(mctz);
607 spin_unlock_irq(&mctz->lock);
612 * memcg and lruvec stats flushing
614 * Many codepaths leading to stats update or read are performance sensitive and
615 * adding stats flushing in such codepaths is not desirable. So, to optimize the
616 * flushing the kernel does:
618 * 1) Periodically and asynchronously flush the stats every 2 seconds to not let
619 * rstat update tree grow unbounded.
621 * 2) Flush the stats synchronously on reader side only when there are more than
622 * (MEMCG_CHARGE_BATCH * nr_cpus) update events. Though this optimization
623 * will let stats be out of sync by atmost (MEMCG_CHARGE_BATCH * nr_cpus) but
624 * only for 2 seconds due to (1).
626 static void flush_memcg_stats_dwork(struct work_struct *w);
627 static DECLARE_DEFERRABLE_WORK(stats_flush_dwork, flush_memcg_stats_dwork);
628 static DEFINE_SPINLOCK(stats_flush_lock);
629 static DEFINE_PER_CPU(unsigned int, stats_updates);
630 static atomic_t stats_flush_threshold = ATOMIC_INIT(0);
632 static inline void memcg_rstat_updated(struct mem_cgroup *memcg)
634 cgroup_rstat_updated(memcg->css.cgroup, smp_processor_id());
635 if (!(__this_cpu_inc_return(stats_updates) % MEMCG_CHARGE_BATCH))
636 atomic_inc(&stats_flush_threshold);
639 static void __mem_cgroup_flush_stats(void)
643 if (!spin_trylock_irqsave(&stats_flush_lock, flag))
646 cgroup_rstat_flush_irqsafe(root_mem_cgroup->css.cgroup);
647 atomic_set(&stats_flush_threshold, 0);
648 spin_unlock_irqrestore(&stats_flush_lock, flag);
651 void mem_cgroup_flush_stats(void)
653 if (atomic_read(&stats_flush_threshold) > num_online_cpus())
654 __mem_cgroup_flush_stats();
657 static void flush_memcg_stats_dwork(struct work_struct *w)
659 mem_cgroup_flush_stats();
660 queue_delayed_work(system_unbound_wq, &stats_flush_dwork, 2UL*HZ);
664 * __mod_memcg_state - update cgroup memory statistics
665 * @memcg: the memory cgroup
666 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
667 * @val: delta to add to the counter, can be negative
669 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
671 if (mem_cgroup_disabled())
674 __this_cpu_add(memcg->vmstats_percpu->state[idx], val);
675 memcg_rstat_updated(memcg);
678 /* idx can be of type enum memcg_stat_item or node_stat_item. */
679 static unsigned long memcg_page_state_local(struct mem_cgroup *memcg, int idx)
684 for_each_possible_cpu(cpu)
685 x += per_cpu(memcg->vmstats_percpu->state[idx], cpu);
693 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
696 struct mem_cgroup_per_node *pn;
697 struct mem_cgroup *memcg;
699 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
703 __this_cpu_add(memcg->vmstats_percpu->state[idx], val);
706 __this_cpu_add(pn->lruvec_stats_percpu->state[idx], val);
708 memcg_rstat_updated(memcg);
712 * __mod_lruvec_state - update lruvec memory statistics
713 * @lruvec: the lruvec
714 * @idx: the stat item
715 * @val: delta to add to the counter, can be negative
717 * The lruvec is the intersection of the NUMA node and a cgroup. This
718 * function updates the all three counters that are affected by a
719 * change of state at this level: per-node, per-cgroup, per-lruvec.
721 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
725 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
727 /* Update memcg and lruvec */
728 if (!mem_cgroup_disabled())
729 __mod_memcg_lruvec_state(lruvec, idx, val);
732 void __mod_lruvec_page_state(struct page *page, enum node_stat_item idx,
735 struct page *head = compound_head(page); /* rmap on tail pages */
736 struct mem_cgroup *memcg;
737 pg_data_t *pgdat = page_pgdat(page);
738 struct lruvec *lruvec;
741 memcg = page_memcg(head);
742 /* Untracked pages have no memcg, no lruvec. Update only the node */
745 __mod_node_page_state(pgdat, idx, val);
749 lruvec = mem_cgroup_lruvec(memcg, pgdat);
750 __mod_lruvec_state(lruvec, idx, val);
753 EXPORT_SYMBOL(__mod_lruvec_page_state);
755 void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val)
757 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
758 struct mem_cgroup *memcg;
759 struct lruvec *lruvec;
762 memcg = mem_cgroup_from_obj(p);
765 * Untracked pages have no memcg, no lruvec. Update only the
766 * node. If we reparent the slab objects to the root memcg,
767 * when we free the slab object, we need to update the per-memcg
768 * vmstats to keep it correct for the root memcg.
771 __mod_node_page_state(pgdat, idx, val);
773 lruvec = mem_cgroup_lruvec(memcg, pgdat);
774 __mod_lruvec_state(lruvec, idx, val);
780 * __count_memcg_events - account VM events in a cgroup
781 * @memcg: the memory cgroup
782 * @idx: the event item
783 * @count: the number of events that occurred
785 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
788 if (mem_cgroup_disabled())
791 __this_cpu_add(memcg->vmstats_percpu->events[idx], count);
792 memcg_rstat_updated(memcg);
795 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
797 return READ_ONCE(memcg->vmstats.events[event]);
800 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
805 for_each_possible_cpu(cpu)
806 x += per_cpu(memcg->vmstats_percpu->events[event], cpu);
810 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
813 /* pagein of a big page is an event. So, ignore page size */
815 __count_memcg_events(memcg, PGPGIN, 1);
817 __count_memcg_events(memcg, PGPGOUT, 1);
818 nr_pages = -nr_pages; /* for event */
821 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
824 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
825 enum mem_cgroup_events_target target)
827 unsigned long val, next;
829 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
830 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
831 /* from time_after() in jiffies.h */
832 if ((long)(next - val) < 0) {
834 case MEM_CGROUP_TARGET_THRESH:
835 next = val + THRESHOLDS_EVENTS_TARGET;
837 case MEM_CGROUP_TARGET_SOFTLIMIT:
838 next = val + SOFTLIMIT_EVENTS_TARGET;
843 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
850 * Check events in order.
853 static void memcg_check_events(struct mem_cgroup *memcg, int nid)
855 /* threshold event is triggered in finer grain than soft limit */
856 if (unlikely(mem_cgroup_event_ratelimit(memcg,
857 MEM_CGROUP_TARGET_THRESH))) {
860 do_softlimit = mem_cgroup_event_ratelimit(memcg,
861 MEM_CGROUP_TARGET_SOFTLIMIT);
862 mem_cgroup_threshold(memcg);
863 if (unlikely(do_softlimit))
864 mem_cgroup_update_tree(memcg, nid);
868 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
871 * mm_update_next_owner() may clear mm->owner to NULL
872 * if it races with swapoff, page migration, etc.
873 * So this can be called with p == NULL.
878 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
880 EXPORT_SYMBOL(mem_cgroup_from_task);
882 static __always_inline struct mem_cgroup *active_memcg(void)
885 return this_cpu_read(int_active_memcg);
887 return current->active_memcg;
891 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
892 * @mm: mm from which memcg should be extracted. It can be NULL.
894 * Obtain a reference on mm->memcg and returns it if successful. If mm
895 * is NULL, then the memcg is chosen as follows:
896 * 1) The active memcg, if set.
897 * 2) current->mm->memcg, if available
899 * If mem_cgroup is disabled, NULL is returned.
901 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
903 struct mem_cgroup *memcg;
905 if (mem_cgroup_disabled())
909 * Page cache insertions can happen without an
910 * actual mm context, e.g. during disk probing
911 * on boot, loopback IO, acct() writes etc.
913 * No need to css_get on root memcg as the reference
914 * counting is disabled on the root level in the
915 * cgroup core. See CSS_NO_REF.
918 memcg = active_memcg();
919 if (unlikely(memcg)) {
920 /* remote memcg must hold a ref */
921 css_get(&memcg->css);
926 return root_mem_cgroup;
931 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
932 if (unlikely(!memcg))
933 memcg = root_mem_cgroup;
934 } while (!css_tryget(&memcg->css));
938 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
940 static __always_inline bool memcg_kmem_bypass(void)
942 /* Allow remote memcg charging from any context. */
943 if (unlikely(active_memcg()))
946 /* Memcg to charge can't be determined. */
947 if (!in_task() || !current->mm || (current->flags & PF_KTHREAD))
954 * mem_cgroup_iter - iterate over memory cgroup hierarchy
955 * @root: hierarchy root
956 * @prev: previously returned memcg, NULL on first invocation
957 * @reclaim: cookie for shared reclaim walks, NULL for full walks
959 * Returns references to children of the hierarchy below @root, or
960 * @root itself, or %NULL after a full round-trip.
962 * Caller must pass the return value in @prev on subsequent
963 * invocations for reference counting, or use mem_cgroup_iter_break()
964 * to cancel a hierarchy walk before the round-trip is complete.
966 * Reclaimers can specify a node in @reclaim to divide up the memcgs
967 * in the hierarchy among all concurrent reclaimers operating on the
970 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
971 struct mem_cgroup *prev,
972 struct mem_cgroup_reclaim_cookie *reclaim)
974 struct mem_cgroup_reclaim_iter *iter;
975 struct cgroup_subsys_state *css = NULL;
976 struct mem_cgroup *memcg = NULL;
977 struct mem_cgroup *pos = NULL;
979 if (mem_cgroup_disabled())
983 root = root_mem_cgroup;
985 if (prev && !reclaim)
991 struct mem_cgroup_per_node *mz;
993 mz = root->nodeinfo[reclaim->pgdat->node_id];
996 if (prev && reclaim->generation != iter->generation)
1000 pos = READ_ONCE(iter->position);
1001 if (!pos || css_tryget(&pos->css))
1004 * css reference reached zero, so iter->position will
1005 * be cleared by ->css_released. However, we should not
1006 * rely on this happening soon, because ->css_released
1007 * is called from a work queue, and by busy-waiting we
1008 * might block it. So we clear iter->position right
1011 (void)cmpxchg(&iter->position, pos, NULL);
1019 css = css_next_descendant_pre(css, &root->css);
1022 * Reclaimers share the hierarchy walk, and a
1023 * new one might jump in right at the end of
1024 * the hierarchy - make sure they see at least
1025 * one group and restart from the beginning.
1033 * Verify the css and acquire a reference. The root
1034 * is provided by the caller, so we know it's alive
1035 * and kicking, and don't take an extra reference.
1037 memcg = mem_cgroup_from_css(css);
1039 if (css == &root->css)
1042 if (css_tryget(css))
1050 * The position could have already been updated by a competing
1051 * thread, so check that the value hasn't changed since we read
1052 * it to avoid reclaiming from the same cgroup twice.
1054 (void)cmpxchg(&iter->position, pos, memcg);
1062 reclaim->generation = iter->generation;
1067 if (prev && prev != root)
1068 css_put(&prev->css);
1074 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1075 * @root: hierarchy root
1076 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1078 void mem_cgroup_iter_break(struct mem_cgroup *root,
1079 struct mem_cgroup *prev)
1082 root = root_mem_cgroup;
1083 if (prev && prev != root)
1084 css_put(&prev->css);
1087 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1088 struct mem_cgroup *dead_memcg)
1090 struct mem_cgroup_reclaim_iter *iter;
1091 struct mem_cgroup_per_node *mz;
1094 for_each_node(nid) {
1095 mz = from->nodeinfo[nid];
1097 cmpxchg(&iter->position, dead_memcg, NULL);
1101 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1103 struct mem_cgroup *memcg = dead_memcg;
1104 struct mem_cgroup *last;
1107 __invalidate_reclaim_iterators(memcg, dead_memcg);
1109 } while ((memcg = parent_mem_cgroup(memcg)));
1112 * When cgruop1 non-hierarchy mode is used,
1113 * parent_mem_cgroup() does not walk all the way up to the
1114 * cgroup root (root_mem_cgroup). So we have to handle
1115 * dead_memcg from cgroup root separately.
1117 if (last != root_mem_cgroup)
1118 __invalidate_reclaim_iterators(root_mem_cgroup,
1123 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1124 * @memcg: hierarchy root
1125 * @fn: function to call for each task
1126 * @arg: argument passed to @fn
1128 * This function iterates over tasks attached to @memcg or to any of its
1129 * descendants and calls @fn for each task. If @fn returns a non-zero
1130 * value, the function breaks the iteration loop and returns the value.
1131 * Otherwise, it will iterate over all tasks and return 0.
1133 * This function must not be called for the root memory cgroup.
1135 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1136 int (*fn)(struct task_struct *, void *), void *arg)
1138 struct mem_cgroup *iter;
1141 BUG_ON(memcg == root_mem_cgroup);
1143 for_each_mem_cgroup_tree(iter, memcg) {
1144 struct css_task_iter it;
1145 struct task_struct *task;
1147 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1148 while (!ret && (task = css_task_iter_next(&it)))
1149 ret = fn(task, arg);
1150 css_task_iter_end(&it);
1152 mem_cgroup_iter_break(memcg, iter);
1159 #ifdef CONFIG_DEBUG_VM
1160 void lruvec_memcg_debug(struct lruvec *lruvec, struct folio *folio)
1162 struct mem_cgroup *memcg;
1164 if (mem_cgroup_disabled())
1167 memcg = folio_memcg(folio);
1170 VM_BUG_ON_FOLIO(lruvec_memcg(lruvec) != root_mem_cgroup, folio);
1172 VM_BUG_ON_FOLIO(lruvec_memcg(lruvec) != memcg, folio);
1177 * folio_lruvec_lock - Lock the lruvec for a folio.
1178 * @folio: Pointer to the folio.
1180 * These functions are safe to use under any of the following conditions:
1182 * - folio_test_lru false
1183 * - folio_memcg_lock()
1184 * - folio frozen (refcount of 0)
1186 * Return: The lruvec this folio is on with its lock held.
1188 struct lruvec *folio_lruvec_lock(struct folio *folio)
1190 struct lruvec *lruvec = folio_lruvec(folio);
1192 spin_lock(&lruvec->lru_lock);
1193 lruvec_memcg_debug(lruvec, folio);
1199 * folio_lruvec_lock_irq - Lock the lruvec for a folio.
1200 * @folio: Pointer to the folio.
1202 * These functions are safe to use under any of the following conditions:
1204 * - folio_test_lru false
1205 * - folio_memcg_lock()
1206 * - folio frozen (refcount of 0)
1208 * Return: The lruvec this folio is on with its lock held and interrupts
1211 struct lruvec *folio_lruvec_lock_irq(struct folio *folio)
1213 struct lruvec *lruvec = folio_lruvec(folio);
1215 spin_lock_irq(&lruvec->lru_lock);
1216 lruvec_memcg_debug(lruvec, folio);
1222 * folio_lruvec_lock_irqsave - Lock the lruvec for a folio.
1223 * @folio: Pointer to the folio.
1224 * @flags: Pointer to irqsave flags.
1226 * These functions are safe to use under any of the following conditions:
1228 * - folio_test_lru false
1229 * - folio_memcg_lock()
1230 * - folio frozen (refcount of 0)
1232 * Return: The lruvec this folio is on with its lock held and interrupts
1235 struct lruvec *folio_lruvec_lock_irqsave(struct folio *folio,
1236 unsigned long *flags)
1238 struct lruvec *lruvec = folio_lruvec(folio);
1240 spin_lock_irqsave(&lruvec->lru_lock, *flags);
1241 lruvec_memcg_debug(lruvec, folio);
1247 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1248 * @lruvec: mem_cgroup per zone lru vector
1249 * @lru: index of lru list the page is sitting on
1250 * @zid: zone id of the accounted pages
1251 * @nr_pages: positive when adding or negative when removing
1253 * This function must be called under lru_lock, just before a page is added
1254 * to or just after a page is removed from an lru list (that ordering being
1255 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1257 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1258 int zid, int nr_pages)
1260 struct mem_cgroup_per_node *mz;
1261 unsigned long *lru_size;
1264 if (mem_cgroup_disabled())
1267 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1268 lru_size = &mz->lru_zone_size[zid][lru];
1271 *lru_size += nr_pages;
1274 if (WARN_ONCE(size < 0,
1275 "%s(%p, %d, %d): lru_size %ld\n",
1276 __func__, lruvec, lru, nr_pages, size)) {
1282 *lru_size += nr_pages;
1286 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1287 * @memcg: the memory cgroup
1289 * Returns the maximum amount of memory @mem can be charged with, in
1292 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1294 unsigned long margin = 0;
1295 unsigned long count;
1296 unsigned long limit;
1298 count = page_counter_read(&memcg->memory);
1299 limit = READ_ONCE(memcg->memory.max);
1301 margin = limit - count;
1303 if (do_memsw_account()) {
1304 count = page_counter_read(&memcg->memsw);
1305 limit = READ_ONCE(memcg->memsw.max);
1307 margin = min(margin, limit - count);
1316 * A routine for checking "mem" is under move_account() or not.
1318 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1319 * moving cgroups. This is for waiting at high-memory pressure
1322 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1324 struct mem_cgroup *from;
1325 struct mem_cgroup *to;
1328 * Unlike task_move routines, we access mc.to, mc.from not under
1329 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1331 spin_lock(&mc.lock);
1337 ret = mem_cgroup_is_descendant(from, memcg) ||
1338 mem_cgroup_is_descendant(to, memcg);
1340 spin_unlock(&mc.lock);
1344 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1346 if (mc.moving_task && current != mc.moving_task) {
1347 if (mem_cgroup_under_move(memcg)) {
1349 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1350 /* moving charge context might have finished. */
1353 finish_wait(&mc.waitq, &wait);
1360 struct memory_stat {
1365 static const struct memory_stat memory_stats[] = {
1366 { "anon", NR_ANON_MAPPED },
1367 { "file", NR_FILE_PAGES },
1368 { "kernel_stack", NR_KERNEL_STACK_KB },
1369 { "pagetables", NR_PAGETABLE },
1370 { "percpu", MEMCG_PERCPU_B },
1371 { "sock", MEMCG_SOCK },
1372 { "shmem", NR_SHMEM },
1373 { "file_mapped", NR_FILE_MAPPED },
1374 { "file_dirty", NR_FILE_DIRTY },
1375 { "file_writeback", NR_WRITEBACK },
1377 { "swapcached", NR_SWAPCACHE },
1379 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1380 { "anon_thp", NR_ANON_THPS },
1381 { "file_thp", NR_FILE_THPS },
1382 { "shmem_thp", NR_SHMEM_THPS },
1384 { "inactive_anon", NR_INACTIVE_ANON },
1385 { "active_anon", NR_ACTIVE_ANON },
1386 { "inactive_file", NR_INACTIVE_FILE },
1387 { "active_file", NR_ACTIVE_FILE },
1388 { "unevictable", NR_UNEVICTABLE },
1389 { "slab_reclaimable", NR_SLAB_RECLAIMABLE_B },
1390 { "slab_unreclaimable", NR_SLAB_UNRECLAIMABLE_B },
1392 /* The memory events */
1393 { "workingset_refault_anon", WORKINGSET_REFAULT_ANON },
1394 { "workingset_refault_file", WORKINGSET_REFAULT_FILE },
1395 { "workingset_activate_anon", WORKINGSET_ACTIVATE_ANON },
1396 { "workingset_activate_file", WORKINGSET_ACTIVATE_FILE },
1397 { "workingset_restore_anon", WORKINGSET_RESTORE_ANON },
1398 { "workingset_restore_file", WORKINGSET_RESTORE_FILE },
1399 { "workingset_nodereclaim", WORKINGSET_NODERECLAIM },
1402 /* Translate stat items to the correct unit for memory.stat output */
1403 static int memcg_page_state_unit(int item)
1406 case MEMCG_PERCPU_B:
1407 case NR_SLAB_RECLAIMABLE_B:
1408 case NR_SLAB_UNRECLAIMABLE_B:
1409 case WORKINGSET_REFAULT_ANON:
1410 case WORKINGSET_REFAULT_FILE:
1411 case WORKINGSET_ACTIVATE_ANON:
1412 case WORKINGSET_ACTIVATE_FILE:
1413 case WORKINGSET_RESTORE_ANON:
1414 case WORKINGSET_RESTORE_FILE:
1415 case WORKINGSET_NODERECLAIM:
1417 case NR_KERNEL_STACK_KB:
1424 static inline unsigned long memcg_page_state_output(struct mem_cgroup *memcg,
1427 return memcg_page_state(memcg, item) * memcg_page_state_unit(item);
1430 static char *memory_stat_format(struct mem_cgroup *memcg)
1435 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1440 * Provide statistics on the state of the memory subsystem as
1441 * well as cumulative event counters that show past behavior.
1443 * This list is ordered following a combination of these gradients:
1444 * 1) generic big picture -> specifics and details
1445 * 2) reflecting userspace activity -> reflecting kernel heuristics
1447 * Current memory state:
1449 mem_cgroup_flush_stats();
1451 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1454 size = memcg_page_state_output(memcg, memory_stats[i].idx);
1455 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1457 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1458 size += memcg_page_state_output(memcg,
1459 NR_SLAB_RECLAIMABLE_B);
1460 seq_buf_printf(&s, "slab %llu\n", size);
1464 /* Accumulated memory events */
1466 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1467 memcg_events(memcg, PGFAULT));
1468 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1469 memcg_events(memcg, PGMAJFAULT));
1470 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1471 memcg_events(memcg, PGREFILL));
1472 seq_buf_printf(&s, "pgscan %lu\n",
1473 memcg_events(memcg, PGSCAN_KSWAPD) +
1474 memcg_events(memcg, PGSCAN_DIRECT));
1475 seq_buf_printf(&s, "pgsteal %lu\n",
1476 memcg_events(memcg, PGSTEAL_KSWAPD) +
1477 memcg_events(memcg, PGSTEAL_DIRECT));
1478 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1479 memcg_events(memcg, PGACTIVATE));
1480 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1481 memcg_events(memcg, PGDEACTIVATE));
1482 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1483 memcg_events(memcg, PGLAZYFREE));
1484 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1485 memcg_events(memcg, PGLAZYFREED));
1487 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1488 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1489 memcg_events(memcg, THP_FAULT_ALLOC));
1490 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1491 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1492 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1494 /* The above should easily fit into one page */
1495 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1500 #define K(x) ((x) << (PAGE_SHIFT-10))
1502 * mem_cgroup_print_oom_context: Print OOM information relevant to
1503 * memory controller.
1504 * @memcg: The memory cgroup that went over limit
1505 * @p: Task that is going to be killed
1507 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1510 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1515 pr_cont(",oom_memcg=");
1516 pr_cont_cgroup_path(memcg->css.cgroup);
1518 pr_cont(",global_oom");
1520 pr_cont(",task_memcg=");
1521 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1527 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1528 * memory controller.
1529 * @memcg: The memory cgroup that went over limit
1531 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1535 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1536 K((u64)page_counter_read(&memcg->memory)),
1537 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1538 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1539 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1540 K((u64)page_counter_read(&memcg->swap)),
1541 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1543 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1544 K((u64)page_counter_read(&memcg->memsw)),
1545 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1546 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1547 K((u64)page_counter_read(&memcg->kmem)),
1548 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1551 pr_info("Memory cgroup stats for ");
1552 pr_cont_cgroup_path(memcg->css.cgroup);
1554 buf = memory_stat_format(memcg);
1562 * Return the memory (and swap, if configured) limit for a memcg.
1564 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1566 unsigned long max = READ_ONCE(memcg->memory.max);
1568 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1569 if (mem_cgroup_swappiness(memcg))
1570 max += min(READ_ONCE(memcg->swap.max),
1571 (unsigned long)total_swap_pages);
1573 if (mem_cgroup_swappiness(memcg)) {
1574 /* Calculate swap excess capacity from memsw limit */
1575 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1577 max += min(swap, (unsigned long)total_swap_pages);
1583 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1585 return page_counter_read(&memcg->memory);
1588 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1591 struct oom_control oc = {
1595 .gfp_mask = gfp_mask,
1600 if (mutex_lock_killable(&oom_lock))
1603 if (mem_cgroup_margin(memcg) >= (1 << order))
1607 * A few threads which were not waiting at mutex_lock_killable() can
1608 * fail to bail out. Therefore, check again after holding oom_lock.
1610 ret = task_is_dying() || out_of_memory(&oc);
1613 mutex_unlock(&oom_lock);
1617 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1620 unsigned long *total_scanned)
1622 struct mem_cgroup *victim = NULL;
1625 unsigned long excess;
1626 unsigned long nr_scanned;
1627 struct mem_cgroup_reclaim_cookie reclaim = {
1631 excess = soft_limit_excess(root_memcg);
1634 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1639 * If we have not been able to reclaim
1640 * anything, it might because there are
1641 * no reclaimable pages under this hierarchy
1646 * We want to do more targeted reclaim.
1647 * excess >> 2 is not to excessive so as to
1648 * reclaim too much, nor too less that we keep
1649 * coming back to reclaim from this cgroup
1651 if (total >= (excess >> 2) ||
1652 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1657 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1658 pgdat, &nr_scanned);
1659 *total_scanned += nr_scanned;
1660 if (!soft_limit_excess(root_memcg))
1663 mem_cgroup_iter_break(root_memcg, victim);
1667 #ifdef CONFIG_LOCKDEP
1668 static struct lockdep_map memcg_oom_lock_dep_map = {
1669 .name = "memcg_oom_lock",
1673 static DEFINE_SPINLOCK(memcg_oom_lock);
1676 * Check OOM-Killer is already running under our hierarchy.
1677 * If someone is running, return false.
1679 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1681 struct mem_cgroup *iter, *failed = NULL;
1683 spin_lock(&memcg_oom_lock);
1685 for_each_mem_cgroup_tree(iter, memcg) {
1686 if (iter->oom_lock) {
1688 * this subtree of our hierarchy is already locked
1689 * so we cannot give a lock.
1692 mem_cgroup_iter_break(memcg, iter);
1695 iter->oom_lock = true;
1700 * OK, we failed to lock the whole subtree so we have
1701 * to clean up what we set up to the failing subtree
1703 for_each_mem_cgroup_tree(iter, memcg) {
1704 if (iter == failed) {
1705 mem_cgroup_iter_break(memcg, iter);
1708 iter->oom_lock = false;
1711 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1713 spin_unlock(&memcg_oom_lock);
1718 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1720 struct mem_cgroup *iter;
1722 spin_lock(&memcg_oom_lock);
1723 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1724 for_each_mem_cgroup_tree(iter, memcg)
1725 iter->oom_lock = false;
1726 spin_unlock(&memcg_oom_lock);
1729 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1731 struct mem_cgroup *iter;
1733 spin_lock(&memcg_oom_lock);
1734 for_each_mem_cgroup_tree(iter, memcg)
1736 spin_unlock(&memcg_oom_lock);
1739 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1741 struct mem_cgroup *iter;
1744 * Be careful about under_oom underflows because a child memcg
1745 * could have been added after mem_cgroup_mark_under_oom.
1747 spin_lock(&memcg_oom_lock);
1748 for_each_mem_cgroup_tree(iter, memcg)
1749 if (iter->under_oom > 0)
1751 spin_unlock(&memcg_oom_lock);
1754 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1756 struct oom_wait_info {
1757 struct mem_cgroup *memcg;
1758 wait_queue_entry_t wait;
1761 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1762 unsigned mode, int sync, void *arg)
1764 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1765 struct mem_cgroup *oom_wait_memcg;
1766 struct oom_wait_info *oom_wait_info;
1768 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1769 oom_wait_memcg = oom_wait_info->memcg;
1771 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1772 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1774 return autoremove_wake_function(wait, mode, sync, arg);
1777 static void memcg_oom_recover(struct mem_cgroup *memcg)
1780 * For the following lockless ->under_oom test, the only required
1781 * guarantee is that it must see the state asserted by an OOM when
1782 * this function is called as a result of userland actions
1783 * triggered by the notification of the OOM. This is trivially
1784 * achieved by invoking mem_cgroup_mark_under_oom() before
1785 * triggering notification.
1787 if (memcg && memcg->under_oom)
1788 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1798 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1800 enum oom_status ret;
1803 if (order > PAGE_ALLOC_COSTLY_ORDER)
1806 memcg_memory_event(memcg, MEMCG_OOM);
1809 * We are in the middle of the charge context here, so we
1810 * don't want to block when potentially sitting on a callstack
1811 * that holds all kinds of filesystem and mm locks.
1813 * cgroup1 allows disabling the OOM killer and waiting for outside
1814 * handling until the charge can succeed; remember the context and put
1815 * the task to sleep at the end of the page fault when all locks are
1818 * On the other hand, in-kernel OOM killer allows for an async victim
1819 * memory reclaim (oom_reaper) and that means that we are not solely
1820 * relying on the oom victim to make a forward progress and we can
1821 * invoke the oom killer here.
1823 * Please note that mem_cgroup_out_of_memory might fail to find a
1824 * victim and then we have to bail out from the charge path.
1826 if (memcg->oom_kill_disable) {
1827 if (!current->in_user_fault)
1829 css_get(&memcg->css);
1830 current->memcg_in_oom = memcg;
1831 current->memcg_oom_gfp_mask = mask;
1832 current->memcg_oom_order = order;
1837 mem_cgroup_mark_under_oom(memcg);
1839 locked = mem_cgroup_oom_trylock(memcg);
1842 mem_cgroup_oom_notify(memcg);
1844 mem_cgroup_unmark_under_oom(memcg);
1845 if (mem_cgroup_out_of_memory(memcg, mask, order))
1851 mem_cgroup_oom_unlock(memcg);
1857 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1858 * @handle: actually kill/wait or just clean up the OOM state
1860 * This has to be called at the end of a page fault if the memcg OOM
1861 * handler was enabled.
1863 * Memcg supports userspace OOM handling where failed allocations must
1864 * sleep on a waitqueue until the userspace task resolves the
1865 * situation. Sleeping directly in the charge context with all kinds
1866 * of locks held is not a good idea, instead we remember an OOM state
1867 * in the task and mem_cgroup_oom_synchronize() has to be called at
1868 * the end of the page fault to complete the OOM handling.
1870 * Returns %true if an ongoing memcg OOM situation was detected and
1871 * completed, %false otherwise.
1873 bool mem_cgroup_oom_synchronize(bool handle)
1875 struct mem_cgroup *memcg = current->memcg_in_oom;
1876 struct oom_wait_info owait;
1879 /* OOM is global, do not handle */
1886 owait.memcg = memcg;
1887 owait.wait.flags = 0;
1888 owait.wait.func = memcg_oom_wake_function;
1889 owait.wait.private = current;
1890 INIT_LIST_HEAD(&owait.wait.entry);
1892 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1893 mem_cgroup_mark_under_oom(memcg);
1895 locked = mem_cgroup_oom_trylock(memcg);
1898 mem_cgroup_oom_notify(memcg);
1900 if (locked && !memcg->oom_kill_disable) {
1901 mem_cgroup_unmark_under_oom(memcg);
1902 finish_wait(&memcg_oom_waitq, &owait.wait);
1903 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1904 current->memcg_oom_order);
1907 mem_cgroup_unmark_under_oom(memcg);
1908 finish_wait(&memcg_oom_waitq, &owait.wait);
1912 mem_cgroup_oom_unlock(memcg);
1914 * There is no guarantee that an OOM-lock contender
1915 * sees the wakeups triggered by the OOM kill
1916 * uncharges. Wake any sleepers explicitly.
1918 memcg_oom_recover(memcg);
1921 current->memcg_in_oom = NULL;
1922 css_put(&memcg->css);
1927 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1928 * @victim: task to be killed by the OOM killer
1929 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1931 * Returns a pointer to a memory cgroup, which has to be cleaned up
1932 * by killing all belonging OOM-killable tasks.
1934 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1936 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1937 struct mem_cgroup *oom_domain)
1939 struct mem_cgroup *oom_group = NULL;
1940 struct mem_cgroup *memcg;
1942 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
1946 oom_domain = root_mem_cgroup;
1950 memcg = mem_cgroup_from_task(victim);
1951 if (memcg == root_mem_cgroup)
1955 * If the victim task has been asynchronously moved to a different
1956 * memory cgroup, we might end up killing tasks outside oom_domain.
1957 * In this case it's better to ignore memory.group.oom.
1959 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
1963 * Traverse the memory cgroup hierarchy from the victim task's
1964 * cgroup up to the OOMing cgroup (or root) to find the
1965 * highest-level memory cgroup with oom.group set.
1967 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1968 if (memcg->oom_group)
1971 if (memcg == oom_domain)
1976 css_get(&oom_group->css);
1983 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
1985 pr_info("Tasks in ");
1986 pr_cont_cgroup_path(memcg->css.cgroup);
1987 pr_cont(" are going to be killed due to memory.oom.group set\n");
1991 * folio_memcg_lock - Bind a folio to its memcg.
1992 * @folio: The folio.
1994 * This function prevents unlocked LRU folios from being moved to
1997 * It ensures lifetime of the bound memcg. The caller is responsible
1998 * for the lifetime of the folio.
2000 void folio_memcg_lock(struct folio *folio)
2002 struct mem_cgroup *memcg;
2003 unsigned long flags;
2006 * The RCU lock is held throughout the transaction. The fast
2007 * path can get away without acquiring the memcg->move_lock
2008 * because page moving starts with an RCU grace period.
2012 if (mem_cgroup_disabled())
2015 memcg = folio_memcg(folio);
2016 if (unlikely(!memcg))
2019 #ifdef CONFIG_PROVE_LOCKING
2020 local_irq_save(flags);
2021 might_lock(&memcg->move_lock);
2022 local_irq_restore(flags);
2025 if (atomic_read(&memcg->moving_account) <= 0)
2028 spin_lock_irqsave(&memcg->move_lock, flags);
2029 if (memcg != folio_memcg(folio)) {
2030 spin_unlock_irqrestore(&memcg->move_lock, flags);
2035 * When charge migration first begins, we can have multiple
2036 * critical sections holding the fast-path RCU lock and one
2037 * holding the slowpath move_lock. Track the task who has the
2038 * move_lock for unlock_page_memcg().
2040 memcg->move_lock_task = current;
2041 memcg->move_lock_flags = flags;
2044 void lock_page_memcg(struct page *page)
2046 folio_memcg_lock(page_folio(page));
2049 static void __folio_memcg_unlock(struct mem_cgroup *memcg)
2051 if (memcg && memcg->move_lock_task == current) {
2052 unsigned long flags = memcg->move_lock_flags;
2054 memcg->move_lock_task = NULL;
2055 memcg->move_lock_flags = 0;
2057 spin_unlock_irqrestore(&memcg->move_lock, flags);
2064 * folio_memcg_unlock - Release the binding between a folio and its memcg.
2065 * @folio: The folio.
2067 * This releases the binding created by folio_memcg_lock(). This does
2068 * not change the accounting of this folio to its memcg, but it does
2069 * permit others to change it.
2071 void folio_memcg_unlock(struct folio *folio)
2073 __folio_memcg_unlock(folio_memcg(folio));
2076 void unlock_page_memcg(struct page *page)
2078 folio_memcg_unlock(page_folio(page));
2082 #ifdef CONFIG_MEMCG_KMEM
2083 struct obj_cgroup *cached_objcg;
2084 struct pglist_data *cached_pgdat;
2085 unsigned int nr_bytes;
2086 int nr_slab_reclaimable_b;
2087 int nr_slab_unreclaimable_b;
2093 struct memcg_stock_pcp {
2094 struct mem_cgroup *cached; /* this never be root cgroup */
2095 unsigned int nr_pages;
2096 struct obj_stock task_obj;
2097 struct obj_stock irq_obj;
2099 struct work_struct work;
2100 unsigned long flags;
2101 #define FLUSHING_CACHED_CHARGE 0
2103 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2104 static DEFINE_MUTEX(percpu_charge_mutex);
2106 #ifdef CONFIG_MEMCG_KMEM
2107 static void drain_obj_stock(struct obj_stock *stock);
2108 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2109 struct mem_cgroup *root_memcg);
2112 static inline void drain_obj_stock(struct obj_stock *stock)
2115 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2116 struct mem_cgroup *root_memcg)
2123 * consume_stock: Try to consume stocked charge on this cpu.
2124 * @memcg: memcg to consume from.
2125 * @nr_pages: how many pages to charge.
2127 * The charges will only happen if @memcg matches the current cpu's memcg
2128 * stock, and at least @nr_pages are available in that stock. Failure to
2129 * service an allocation will refill the stock.
2131 * returns true if successful, false otherwise.
2133 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2135 struct memcg_stock_pcp *stock;
2136 unsigned long flags;
2139 if (nr_pages > MEMCG_CHARGE_BATCH)
2142 local_irq_save(flags);
2144 stock = this_cpu_ptr(&memcg_stock);
2145 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2146 stock->nr_pages -= nr_pages;
2150 local_irq_restore(flags);
2156 * Returns stocks cached in percpu and reset cached information.
2158 static void drain_stock(struct memcg_stock_pcp *stock)
2160 struct mem_cgroup *old = stock->cached;
2165 if (stock->nr_pages) {
2166 page_counter_uncharge(&old->memory, stock->nr_pages);
2167 if (do_memsw_account())
2168 page_counter_uncharge(&old->memsw, stock->nr_pages);
2169 stock->nr_pages = 0;
2173 stock->cached = NULL;
2176 static void drain_local_stock(struct work_struct *dummy)
2178 struct memcg_stock_pcp *stock;
2179 unsigned long flags;
2182 * The only protection from cpu hotplug (memcg_hotplug_cpu_dead) vs.
2183 * drain_stock races is that we always operate on local CPU stock
2184 * here with IRQ disabled
2186 local_irq_save(flags);
2188 stock = this_cpu_ptr(&memcg_stock);
2189 drain_obj_stock(&stock->irq_obj);
2191 drain_obj_stock(&stock->task_obj);
2193 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2195 local_irq_restore(flags);
2199 * Cache charges(val) to local per_cpu area.
2200 * This will be consumed by consume_stock() function, later.
2202 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2204 struct memcg_stock_pcp *stock;
2205 unsigned long flags;
2207 local_irq_save(flags);
2209 stock = this_cpu_ptr(&memcg_stock);
2210 if (stock->cached != memcg) { /* reset if necessary */
2212 css_get(&memcg->css);
2213 stock->cached = memcg;
2215 stock->nr_pages += nr_pages;
2217 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2220 local_irq_restore(flags);
2224 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2225 * of the hierarchy under it.
2227 static void drain_all_stock(struct mem_cgroup *root_memcg)
2231 /* If someone's already draining, avoid adding running more workers. */
2232 if (!mutex_trylock(&percpu_charge_mutex))
2235 * Notify other cpus that system-wide "drain" is running
2236 * We do not care about races with the cpu hotplug because cpu down
2237 * as well as workers from this path always operate on the local
2238 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2241 for_each_online_cpu(cpu) {
2242 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2243 struct mem_cgroup *memcg;
2247 memcg = stock->cached;
2248 if (memcg && stock->nr_pages &&
2249 mem_cgroup_is_descendant(memcg, root_memcg))
2251 else if (obj_stock_flush_required(stock, root_memcg))
2256 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2258 drain_local_stock(&stock->work);
2260 schedule_work_on(cpu, &stock->work);
2264 mutex_unlock(&percpu_charge_mutex);
2267 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2269 struct memcg_stock_pcp *stock;
2271 stock = &per_cpu(memcg_stock, cpu);
2277 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2278 unsigned int nr_pages,
2281 unsigned long nr_reclaimed = 0;
2284 unsigned long pflags;
2286 if (page_counter_read(&memcg->memory) <=
2287 READ_ONCE(memcg->memory.high))
2290 memcg_memory_event(memcg, MEMCG_HIGH);
2292 psi_memstall_enter(&pflags);
2293 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2295 psi_memstall_leave(&pflags);
2296 } while ((memcg = parent_mem_cgroup(memcg)) &&
2297 !mem_cgroup_is_root(memcg));
2299 return nr_reclaimed;
2302 static void high_work_func(struct work_struct *work)
2304 struct mem_cgroup *memcg;
2306 memcg = container_of(work, struct mem_cgroup, high_work);
2307 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2311 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2312 * enough to still cause a significant slowdown in most cases, while still
2313 * allowing diagnostics and tracing to proceed without becoming stuck.
2315 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2318 * When calculating the delay, we use these either side of the exponentiation to
2319 * maintain precision and scale to a reasonable number of jiffies (see the table
2322 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2323 * overage ratio to a delay.
2324 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2325 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2326 * to produce a reasonable delay curve.
2328 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2329 * reasonable delay curve compared to precision-adjusted overage, not
2330 * penalising heavily at first, but still making sure that growth beyond the
2331 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2332 * example, with a high of 100 megabytes:
2334 * +-------+------------------------+
2335 * | usage | time to allocate in ms |
2336 * +-------+------------------------+
2358 * +-------+------------------------+
2360 #define MEMCG_DELAY_PRECISION_SHIFT 20
2361 #define MEMCG_DELAY_SCALING_SHIFT 14
2363 static u64 calculate_overage(unsigned long usage, unsigned long high)
2371 * Prevent division by 0 in overage calculation by acting as if
2372 * it was a threshold of 1 page
2374 high = max(high, 1UL);
2376 overage = usage - high;
2377 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2378 return div64_u64(overage, high);
2381 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2383 u64 overage, max_overage = 0;
2386 overage = calculate_overage(page_counter_read(&memcg->memory),
2387 READ_ONCE(memcg->memory.high));
2388 max_overage = max(overage, max_overage);
2389 } while ((memcg = parent_mem_cgroup(memcg)) &&
2390 !mem_cgroup_is_root(memcg));
2395 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2397 u64 overage, max_overage = 0;
2400 overage = calculate_overage(page_counter_read(&memcg->swap),
2401 READ_ONCE(memcg->swap.high));
2403 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2404 max_overage = max(overage, max_overage);
2405 } while ((memcg = parent_mem_cgroup(memcg)) &&
2406 !mem_cgroup_is_root(memcg));
2412 * Get the number of jiffies that we should penalise a mischievous cgroup which
2413 * is exceeding its memory.high by checking both it and its ancestors.
2415 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2416 unsigned int nr_pages,
2419 unsigned long penalty_jiffies;
2425 * We use overage compared to memory.high to calculate the number of
2426 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2427 * fairly lenient on small overages, and increasingly harsh when the
2428 * memcg in question makes it clear that it has no intention of stopping
2429 * its crazy behaviour, so we exponentially increase the delay based on
2432 penalty_jiffies = max_overage * max_overage * HZ;
2433 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2434 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2437 * Factor in the task's own contribution to the overage, such that four
2438 * N-sized allocations are throttled approximately the same as one
2439 * 4N-sized allocation.
2441 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2442 * larger the current charge patch is than that.
2444 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2448 * Scheduled by try_charge() to be executed from the userland return path
2449 * and reclaims memory over the high limit.
2451 void mem_cgroup_handle_over_high(void)
2453 unsigned long penalty_jiffies;
2454 unsigned long pflags;
2455 unsigned long nr_reclaimed;
2456 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2457 int nr_retries = MAX_RECLAIM_RETRIES;
2458 struct mem_cgroup *memcg;
2459 bool in_retry = false;
2461 if (likely(!nr_pages))
2464 memcg = get_mem_cgroup_from_mm(current->mm);
2465 current->memcg_nr_pages_over_high = 0;
2469 * The allocating task should reclaim at least the batch size, but for
2470 * subsequent retries we only want to do what's necessary to prevent oom
2471 * or breaching resource isolation.
2473 * This is distinct from memory.max or page allocator behaviour because
2474 * memory.high is currently batched, whereas memory.max and the page
2475 * allocator run every time an allocation is made.
2477 nr_reclaimed = reclaim_high(memcg,
2478 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2482 * memory.high is breached and reclaim is unable to keep up. Throttle
2483 * allocators proactively to slow down excessive growth.
2485 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2486 mem_find_max_overage(memcg));
2488 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2489 swap_find_max_overage(memcg));
2492 * Clamp the max delay per usermode return so as to still keep the
2493 * application moving forwards and also permit diagnostics, albeit
2496 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2499 * Don't sleep if the amount of jiffies this memcg owes us is so low
2500 * that it's not even worth doing, in an attempt to be nice to those who
2501 * go only a small amount over their memory.high value and maybe haven't
2502 * been aggressively reclaimed enough yet.
2504 if (penalty_jiffies <= HZ / 100)
2508 * If reclaim is making forward progress but we're still over
2509 * memory.high, we want to encourage that rather than doing allocator
2512 if (nr_reclaimed || nr_retries--) {
2518 * If we exit early, we're guaranteed to die (since
2519 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2520 * need to account for any ill-begotten jiffies to pay them off later.
2522 psi_memstall_enter(&pflags);
2523 schedule_timeout_killable(penalty_jiffies);
2524 psi_memstall_leave(&pflags);
2527 css_put(&memcg->css);
2530 static int try_charge_memcg(struct mem_cgroup *memcg, gfp_t gfp_mask,
2531 unsigned int nr_pages)
2533 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2534 int nr_retries = MAX_RECLAIM_RETRIES;
2535 struct mem_cgroup *mem_over_limit;
2536 struct page_counter *counter;
2537 enum oom_status oom_status;
2538 unsigned long nr_reclaimed;
2539 bool passed_oom = false;
2540 bool may_swap = true;
2541 bool drained = false;
2542 unsigned long pflags;
2545 if (consume_stock(memcg, nr_pages))
2548 if (!do_memsw_account() ||
2549 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2550 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2552 if (do_memsw_account())
2553 page_counter_uncharge(&memcg->memsw, batch);
2554 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2556 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2560 if (batch > nr_pages) {
2566 * Memcg doesn't have a dedicated reserve for atomic
2567 * allocations. But like the global atomic pool, we need to
2568 * put the burden of reclaim on regular allocation requests
2569 * and let these go through as privileged allocations.
2571 if (gfp_mask & __GFP_ATOMIC)
2575 * Prevent unbounded recursion when reclaim operations need to
2576 * allocate memory. This might exceed the limits temporarily,
2577 * but we prefer facilitating memory reclaim and getting back
2578 * under the limit over triggering OOM kills in these cases.
2580 if (unlikely(current->flags & PF_MEMALLOC))
2583 if (unlikely(task_in_memcg_oom(current)))
2586 if (!gfpflags_allow_blocking(gfp_mask))
2589 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2591 psi_memstall_enter(&pflags);
2592 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2593 gfp_mask, may_swap);
2594 psi_memstall_leave(&pflags);
2596 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2600 drain_all_stock(mem_over_limit);
2605 if (gfp_mask & __GFP_NORETRY)
2608 * Even though the limit is exceeded at this point, reclaim
2609 * may have been able to free some pages. Retry the charge
2610 * before killing the task.
2612 * Only for regular pages, though: huge pages are rather
2613 * unlikely to succeed so close to the limit, and we fall back
2614 * to regular pages anyway in case of failure.
2616 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2619 * At task move, charge accounts can be doubly counted. So, it's
2620 * better to wait until the end of task_move if something is going on.
2622 if (mem_cgroup_wait_acct_move(mem_over_limit))
2628 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2631 /* Avoid endless loop for tasks bypassed by the oom killer */
2632 if (passed_oom && task_is_dying())
2636 * keep retrying as long as the memcg oom killer is able to make
2637 * a forward progress or bypass the charge if the oom killer
2638 * couldn't make any progress.
2640 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2641 get_order(nr_pages * PAGE_SIZE));
2642 if (oom_status == OOM_SUCCESS) {
2644 nr_retries = MAX_RECLAIM_RETRIES;
2648 if (!(gfp_mask & __GFP_NOFAIL))
2652 * The allocation either can't fail or will lead to more memory
2653 * being freed very soon. Allow memory usage go over the limit
2654 * temporarily by force charging it.
2656 page_counter_charge(&memcg->memory, nr_pages);
2657 if (do_memsw_account())
2658 page_counter_charge(&memcg->memsw, nr_pages);
2663 if (batch > nr_pages)
2664 refill_stock(memcg, batch - nr_pages);
2667 * If the hierarchy is above the normal consumption range, schedule
2668 * reclaim on returning to userland. We can perform reclaim here
2669 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2670 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2671 * not recorded as it most likely matches current's and won't
2672 * change in the meantime. As high limit is checked again before
2673 * reclaim, the cost of mismatch is negligible.
2676 bool mem_high, swap_high;
2678 mem_high = page_counter_read(&memcg->memory) >
2679 READ_ONCE(memcg->memory.high);
2680 swap_high = page_counter_read(&memcg->swap) >
2681 READ_ONCE(memcg->swap.high);
2683 /* Don't bother a random interrupted task */
2684 if (in_interrupt()) {
2686 schedule_work(&memcg->high_work);
2692 if (mem_high || swap_high) {
2694 * The allocating tasks in this cgroup will need to do
2695 * reclaim or be throttled to prevent further growth
2696 * of the memory or swap footprints.
2698 * Target some best-effort fairness between the tasks,
2699 * and distribute reclaim work and delay penalties
2700 * based on how much each task is actually allocating.
2702 current->memcg_nr_pages_over_high += batch;
2703 set_notify_resume(current);
2706 } while ((memcg = parent_mem_cgroup(memcg)));
2711 static inline int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2712 unsigned int nr_pages)
2714 if (mem_cgroup_is_root(memcg))
2717 return try_charge_memcg(memcg, gfp_mask, nr_pages);
2720 static inline void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2722 if (mem_cgroup_is_root(memcg))
2725 page_counter_uncharge(&memcg->memory, nr_pages);
2726 if (do_memsw_account())
2727 page_counter_uncharge(&memcg->memsw, nr_pages);
2730 static void commit_charge(struct folio *folio, struct mem_cgroup *memcg)
2732 VM_BUG_ON_FOLIO(folio_memcg(folio), folio);
2734 * Any of the following ensures page's memcg stability:
2738 * - lock_page_memcg()
2739 * - exclusive reference
2741 folio->memcg_data = (unsigned long)memcg;
2744 static struct mem_cgroup *get_mem_cgroup_from_objcg(struct obj_cgroup *objcg)
2746 struct mem_cgroup *memcg;
2750 memcg = obj_cgroup_memcg(objcg);
2751 if (unlikely(!css_tryget(&memcg->css)))
2758 #ifdef CONFIG_MEMCG_KMEM
2760 * The allocated objcg pointers array is not accounted directly.
2761 * Moreover, it should not come from DMA buffer and is not readily
2762 * reclaimable. So those GFP bits should be masked off.
2764 #define OBJCGS_CLEAR_MASK (__GFP_DMA | __GFP_RECLAIMABLE | __GFP_ACCOUNT)
2767 * Most kmem_cache_alloc() calls are from user context. The irq disable/enable
2768 * sequence used in this case to access content from object stock is slow.
2769 * To optimize for user context access, there are now two object stocks for
2770 * task context and interrupt context access respectively.
2772 * The task context object stock can be accessed by disabling preemption only
2773 * which is cheap in non-preempt kernel. The interrupt context object stock
2774 * can only be accessed after disabling interrupt. User context code can
2775 * access interrupt object stock, but not vice versa.
2777 static inline struct obj_stock *get_obj_stock(unsigned long *pflags)
2779 struct memcg_stock_pcp *stock;
2781 if (likely(in_task())) {
2784 stock = this_cpu_ptr(&memcg_stock);
2785 return &stock->task_obj;
2788 local_irq_save(*pflags);
2789 stock = this_cpu_ptr(&memcg_stock);
2790 return &stock->irq_obj;
2793 static inline void put_obj_stock(unsigned long flags)
2795 if (likely(in_task()))
2798 local_irq_restore(flags);
2802 * mod_objcg_mlstate() may be called with irq enabled, so
2803 * mod_memcg_lruvec_state() should be used.
2805 static inline void mod_objcg_mlstate(struct obj_cgroup *objcg,
2806 struct pglist_data *pgdat,
2807 enum node_stat_item idx, int nr)
2809 struct mem_cgroup *memcg;
2810 struct lruvec *lruvec;
2813 memcg = obj_cgroup_memcg(objcg);
2814 lruvec = mem_cgroup_lruvec(memcg, pgdat);
2815 mod_memcg_lruvec_state(lruvec, idx, nr);
2819 int memcg_alloc_slab_cgroups(struct slab *slab, struct kmem_cache *s,
2820 gfp_t gfp, bool new_slab)
2822 unsigned int objects = objs_per_slab(s, slab);
2823 unsigned long memcg_data;
2826 gfp &= ~OBJCGS_CLEAR_MASK;
2827 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2832 memcg_data = (unsigned long) vec | MEMCG_DATA_OBJCGS;
2835 * If the slab is brand new and nobody can yet access its
2836 * memcg_data, no synchronization is required and memcg_data can
2837 * be simply assigned.
2839 slab->memcg_data = memcg_data;
2840 } else if (cmpxchg(&slab->memcg_data, 0, memcg_data)) {
2842 * If the slab is already in use, somebody can allocate and
2843 * assign obj_cgroups in parallel. In this case the existing
2844 * objcg vector should be reused.
2850 kmemleak_not_leak(vec);
2855 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2857 * A passed kernel object can be a slab object or a generic kernel page, so
2858 * different mechanisms for getting the memory cgroup pointer should be used.
2859 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2860 * can not know for sure how the kernel object is implemented.
2861 * mem_cgroup_from_obj() can be safely used in such cases.
2863 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2864 * cgroup_mutex, etc.
2866 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2868 struct folio *folio;
2870 if (mem_cgroup_disabled())
2873 folio = virt_to_folio(p);
2876 * Slab objects are accounted individually, not per-page.
2877 * Memcg membership data for each individual object is saved in
2880 if (folio_test_slab(folio)) {
2881 struct obj_cgroup **objcgs;
2885 slab = folio_slab(folio);
2886 objcgs = slab_objcgs(slab);
2890 off = obj_to_index(slab->slab_cache, slab, p);
2892 return obj_cgroup_memcg(objcgs[off]);
2898 * page_memcg_check() is used here, because in theory we can encounter
2899 * a folio where the slab flag has been cleared already, but
2900 * slab->memcg_data has not been freed yet
2901 * page_memcg_check(page) will guarantee that a proper memory
2902 * cgroup pointer or NULL will be returned.
2904 return page_memcg_check(folio_page(folio, 0));
2907 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2909 struct obj_cgroup *objcg = NULL;
2910 struct mem_cgroup *memcg;
2912 if (memcg_kmem_bypass())
2916 if (unlikely(active_memcg()))
2917 memcg = active_memcg();
2919 memcg = mem_cgroup_from_task(current);
2921 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2922 objcg = rcu_dereference(memcg->objcg);
2923 if (objcg && obj_cgroup_tryget(objcg))
2932 static int memcg_alloc_cache_id(void)
2937 id = ida_simple_get(&memcg_cache_ida,
2938 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2942 if (id < memcg_nr_cache_ids)
2946 * There's no space for the new id in memcg_caches arrays,
2947 * so we have to grow them.
2949 down_write(&memcg_cache_ids_sem);
2951 size = 2 * (id + 1);
2952 if (size < MEMCG_CACHES_MIN_SIZE)
2953 size = MEMCG_CACHES_MIN_SIZE;
2954 else if (size > MEMCG_CACHES_MAX_SIZE)
2955 size = MEMCG_CACHES_MAX_SIZE;
2957 err = memcg_update_all_list_lrus(size);
2959 memcg_nr_cache_ids = size;
2961 up_write(&memcg_cache_ids_sem);
2964 ida_simple_remove(&memcg_cache_ida, id);
2970 static void memcg_free_cache_id(int id)
2972 ida_simple_remove(&memcg_cache_ida, id);
2976 * obj_cgroup_uncharge_pages: uncharge a number of kernel pages from a objcg
2977 * @objcg: object cgroup to uncharge
2978 * @nr_pages: number of pages to uncharge
2980 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
2981 unsigned int nr_pages)
2983 struct mem_cgroup *memcg;
2985 memcg = get_mem_cgroup_from_objcg(objcg);
2987 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2988 page_counter_uncharge(&memcg->kmem, nr_pages);
2989 refill_stock(memcg, nr_pages);
2991 css_put(&memcg->css);
2995 * obj_cgroup_charge_pages: charge a number of kernel pages to a objcg
2996 * @objcg: object cgroup to charge
2997 * @gfp: reclaim mode
2998 * @nr_pages: number of pages to charge
3000 * Returns 0 on success, an error code on failure.
3002 static int obj_cgroup_charge_pages(struct obj_cgroup *objcg, gfp_t gfp,
3003 unsigned int nr_pages)
3005 struct mem_cgroup *memcg;
3008 memcg = get_mem_cgroup_from_objcg(objcg);
3010 ret = try_charge_memcg(memcg, gfp, nr_pages);
3014 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3015 page_counter_charge(&memcg->kmem, nr_pages);
3017 css_put(&memcg->css);
3023 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3024 * @page: page to charge
3025 * @gfp: reclaim mode
3026 * @order: allocation order
3028 * Returns 0 on success, an error code on failure.
3030 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3032 struct obj_cgroup *objcg;
3035 objcg = get_obj_cgroup_from_current();
3037 ret = obj_cgroup_charge_pages(objcg, gfp, 1 << order);
3039 page->memcg_data = (unsigned long)objcg |
3043 obj_cgroup_put(objcg);
3049 * __memcg_kmem_uncharge_page: uncharge a kmem page
3050 * @page: page to uncharge
3051 * @order: allocation order
3053 void __memcg_kmem_uncharge_page(struct page *page, int order)
3055 struct folio *folio = page_folio(page);
3056 struct obj_cgroup *objcg;
3057 unsigned int nr_pages = 1 << order;
3059 if (!folio_memcg_kmem(folio))
3062 objcg = __folio_objcg(folio);
3063 obj_cgroup_uncharge_pages(objcg, nr_pages);
3064 folio->memcg_data = 0;
3065 obj_cgroup_put(objcg);
3068 void mod_objcg_state(struct obj_cgroup *objcg, struct pglist_data *pgdat,
3069 enum node_stat_item idx, int nr)
3071 unsigned long flags;
3072 struct obj_stock *stock = get_obj_stock(&flags);
3076 * Save vmstat data in stock and skip vmstat array update unless
3077 * accumulating over a page of vmstat data or when pgdat or idx
3080 if (stock->cached_objcg != objcg) {
3081 drain_obj_stock(stock);
3082 obj_cgroup_get(objcg);
3083 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3084 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3085 stock->cached_objcg = objcg;
3086 stock->cached_pgdat = pgdat;
3087 } else if (stock->cached_pgdat != pgdat) {
3088 /* Flush the existing cached vmstat data */
3089 struct pglist_data *oldpg = stock->cached_pgdat;
3091 if (stock->nr_slab_reclaimable_b) {
3092 mod_objcg_mlstate(objcg, oldpg, NR_SLAB_RECLAIMABLE_B,
3093 stock->nr_slab_reclaimable_b);
3094 stock->nr_slab_reclaimable_b = 0;
3096 if (stock->nr_slab_unreclaimable_b) {
3097 mod_objcg_mlstate(objcg, oldpg, NR_SLAB_UNRECLAIMABLE_B,
3098 stock->nr_slab_unreclaimable_b);
3099 stock->nr_slab_unreclaimable_b = 0;
3101 stock->cached_pgdat = pgdat;
3104 bytes = (idx == NR_SLAB_RECLAIMABLE_B) ? &stock->nr_slab_reclaimable_b
3105 : &stock->nr_slab_unreclaimable_b;
3107 * Even for large object >= PAGE_SIZE, the vmstat data will still be
3108 * cached locally at least once before pushing it out.
3115 if (abs(*bytes) > PAGE_SIZE) {
3123 mod_objcg_mlstate(objcg, pgdat, idx, nr);
3125 put_obj_stock(flags);
3128 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3130 unsigned long flags;
3131 struct obj_stock *stock = get_obj_stock(&flags);
3134 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3135 stock->nr_bytes -= nr_bytes;
3139 put_obj_stock(flags);
3144 static void drain_obj_stock(struct obj_stock *stock)
3146 struct obj_cgroup *old = stock->cached_objcg;
3151 if (stock->nr_bytes) {
3152 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3153 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3156 obj_cgroup_uncharge_pages(old, nr_pages);
3159 * The leftover is flushed to the centralized per-memcg value.
3160 * On the next attempt to refill obj stock it will be moved
3161 * to a per-cpu stock (probably, on an other CPU), see
3162 * refill_obj_stock().
3164 * How often it's flushed is a trade-off between the memory
3165 * limit enforcement accuracy and potential CPU contention,
3166 * so it might be changed in the future.
3168 atomic_add(nr_bytes, &old->nr_charged_bytes);
3169 stock->nr_bytes = 0;
3173 * Flush the vmstat data in current stock
3175 if (stock->nr_slab_reclaimable_b || stock->nr_slab_unreclaimable_b) {
3176 if (stock->nr_slab_reclaimable_b) {
3177 mod_objcg_mlstate(old, stock->cached_pgdat,
3178 NR_SLAB_RECLAIMABLE_B,
3179 stock->nr_slab_reclaimable_b);
3180 stock->nr_slab_reclaimable_b = 0;
3182 if (stock->nr_slab_unreclaimable_b) {
3183 mod_objcg_mlstate(old, stock->cached_pgdat,
3184 NR_SLAB_UNRECLAIMABLE_B,
3185 stock->nr_slab_unreclaimable_b);
3186 stock->nr_slab_unreclaimable_b = 0;
3188 stock->cached_pgdat = NULL;
3191 obj_cgroup_put(old);
3192 stock->cached_objcg = NULL;
3195 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3196 struct mem_cgroup *root_memcg)
3198 struct mem_cgroup *memcg;
3200 if (in_task() && stock->task_obj.cached_objcg) {
3201 memcg = obj_cgroup_memcg(stock->task_obj.cached_objcg);
3202 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3205 if (stock->irq_obj.cached_objcg) {
3206 memcg = obj_cgroup_memcg(stock->irq_obj.cached_objcg);
3207 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3214 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes,
3215 bool allow_uncharge)
3217 unsigned long flags;
3218 struct obj_stock *stock = get_obj_stock(&flags);
3219 unsigned int nr_pages = 0;
3221 if (stock->cached_objcg != objcg) { /* reset if necessary */
3222 drain_obj_stock(stock);
3223 obj_cgroup_get(objcg);
3224 stock->cached_objcg = objcg;
3225 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3226 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3227 allow_uncharge = true; /* Allow uncharge when objcg changes */
3229 stock->nr_bytes += nr_bytes;
3231 if (allow_uncharge && (stock->nr_bytes > PAGE_SIZE)) {
3232 nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3233 stock->nr_bytes &= (PAGE_SIZE - 1);
3236 put_obj_stock(flags);
3239 obj_cgroup_uncharge_pages(objcg, nr_pages);
3242 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3244 unsigned int nr_pages, nr_bytes;
3247 if (consume_obj_stock(objcg, size))
3251 * In theory, objcg->nr_charged_bytes can have enough
3252 * pre-charged bytes to satisfy the allocation. However,
3253 * flushing objcg->nr_charged_bytes requires two atomic
3254 * operations, and objcg->nr_charged_bytes can't be big.
3255 * The shared objcg->nr_charged_bytes can also become a
3256 * performance bottleneck if all tasks of the same memcg are
3257 * trying to update it. So it's better to ignore it and try
3258 * grab some new pages. The stock's nr_bytes will be flushed to
3259 * objcg->nr_charged_bytes later on when objcg changes.
3261 * The stock's nr_bytes may contain enough pre-charged bytes
3262 * to allow one less page from being charged, but we can't rely
3263 * on the pre-charged bytes not being changed outside of
3264 * consume_obj_stock() or refill_obj_stock(). So ignore those
3265 * pre-charged bytes as well when charging pages. To avoid a
3266 * page uncharge right after a page charge, we set the
3267 * allow_uncharge flag to false when calling refill_obj_stock()
3268 * to temporarily allow the pre-charged bytes to exceed the page
3269 * size limit. The maximum reachable value of the pre-charged
3270 * bytes is (sizeof(object) + PAGE_SIZE - 2) if there is no data
3273 nr_pages = size >> PAGE_SHIFT;
3274 nr_bytes = size & (PAGE_SIZE - 1);
3279 ret = obj_cgroup_charge_pages(objcg, gfp, nr_pages);
3280 if (!ret && nr_bytes)
3281 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes, false);
3286 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3288 refill_obj_stock(objcg, size, true);
3291 #endif /* CONFIG_MEMCG_KMEM */
3294 * Because page_memcg(head) is not set on tails, set it now.
3296 void split_page_memcg(struct page *head, unsigned int nr)
3298 struct folio *folio = page_folio(head);
3299 struct mem_cgroup *memcg = folio_memcg(folio);
3302 if (mem_cgroup_disabled() || !memcg)
3305 for (i = 1; i < nr; i++)
3306 folio_page(folio, i)->memcg_data = folio->memcg_data;
3308 if (folio_memcg_kmem(folio))
3309 obj_cgroup_get_many(__folio_objcg(folio), nr - 1);
3311 css_get_many(&memcg->css, nr - 1);
3314 #ifdef CONFIG_MEMCG_SWAP
3316 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3317 * @entry: swap entry to be moved
3318 * @from: mem_cgroup which the entry is moved from
3319 * @to: mem_cgroup which the entry is moved to
3321 * It succeeds only when the swap_cgroup's record for this entry is the same
3322 * as the mem_cgroup's id of @from.
3324 * Returns 0 on success, -EINVAL on failure.
3326 * The caller must have charged to @to, IOW, called page_counter_charge() about
3327 * both res and memsw, and called css_get().
3329 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3330 struct mem_cgroup *from, struct mem_cgroup *to)
3332 unsigned short old_id, new_id;
3334 old_id = mem_cgroup_id(from);
3335 new_id = mem_cgroup_id(to);
3337 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3338 mod_memcg_state(from, MEMCG_SWAP, -1);
3339 mod_memcg_state(to, MEMCG_SWAP, 1);
3345 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3346 struct mem_cgroup *from, struct mem_cgroup *to)
3352 static DEFINE_MUTEX(memcg_max_mutex);
3354 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3355 unsigned long max, bool memsw)
3357 bool enlarge = false;
3358 bool drained = false;
3360 bool limits_invariant;
3361 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3364 if (signal_pending(current)) {
3369 mutex_lock(&memcg_max_mutex);
3371 * Make sure that the new limit (memsw or memory limit) doesn't
3372 * break our basic invariant rule memory.max <= memsw.max.
3374 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3375 max <= memcg->memsw.max;
3376 if (!limits_invariant) {
3377 mutex_unlock(&memcg_max_mutex);
3381 if (max > counter->max)
3383 ret = page_counter_set_max(counter, max);
3384 mutex_unlock(&memcg_max_mutex);
3390 drain_all_stock(memcg);
3395 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3396 GFP_KERNEL, !memsw)) {
3402 if (!ret && enlarge)
3403 memcg_oom_recover(memcg);
3408 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3410 unsigned long *total_scanned)
3412 unsigned long nr_reclaimed = 0;
3413 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3414 unsigned long reclaimed;
3416 struct mem_cgroup_tree_per_node *mctz;
3417 unsigned long excess;
3418 unsigned long nr_scanned;
3423 mctz = soft_limit_tree.rb_tree_per_node[pgdat->node_id];
3426 * Do not even bother to check the largest node if the root
3427 * is empty. Do it lockless to prevent lock bouncing. Races
3428 * are acceptable as soft limit is best effort anyway.
3430 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3434 * This loop can run a while, specially if mem_cgroup's continuously
3435 * keep exceeding their soft limit and putting the system under
3442 mz = mem_cgroup_largest_soft_limit_node(mctz);
3447 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3448 gfp_mask, &nr_scanned);
3449 nr_reclaimed += reclaimed;
3450 *total_scanned += nr_scanned;
3451 spin_lock_irq(&mctz->lock);
3452 __mem_cgroup_remove_exceeded(mz, mctz);
3455 * If we failed to reclaim anything from this memory cgroup
3456 * it is time to move on to the next cgroup
3460 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3462 excess = soft_limit_excess(mz->memcg);
3464 * One school of thought says that we should not add
3465 * back the node to the tree if reclaim returns 0.
3466 * But our reclaim could return 0, simply because due
3467 * to priority we are exposing a smaller subset of
3468 * memory to reclaim from. Consider this as a longer
3471 /* If excess == 0, no tree ops */
3472 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3473 spin_unlock_irq(&mctz->lock);
3474 css_put(&mz->memcg->css);
3477 * Could not reclaim anything and there are no more
3478 * mem cgroups to try or we seem to be looping without
3479 * reclaiming anything.
3481 if (!nr_reclaimed &&
3483 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3485 } while (!nr_reclaimed);
3487 css_put(&next_mz->memcg->css);
3488 return nr_reclaimed;
3492 * Reclaims as many pages from the given memcg as possible.
3494 * Caller is responsible for holding css reference for memcg.
3496 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3498 int nr_retries = MAX_RECLAIM_RETRIES;
3500 /* we call try-to-free pages for make this cgroup empty */
3501 lru_add_drain_all();
3503 drain_all_stock(memcg);
3505 /* try to free all pages in this cgroup */
3506 while (nr_retries && page_counter_read(&memcg->memory)) {
3507 if (signal_pending(current))
3510 if (!try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true))
3517 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3518 char *buf, size_t nbytes,
3521 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3523 if (mem_cgroup_is_root(memcg))
3525 return mem_cgroup_force_empty(memcg) ?: nbytes;
3528 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3534 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3535 struct cftype *cft, u64 val)
3540 pr_warn_once("Non-hierarchical mode is deprecated. "
3541 "Please report your usecase to linux-mm@kvack.org if you "
3542 "depend on this functionality.\n");
3547 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3551 if (mem_cgroup_is_root(memcg)) {
3552 mem_cgroup_flush_stats();
3553 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3554 memcg_page_state(memcg, NR_ANON_MAPPED);
3556 val += memcg_page_state(memcg, MEMCG_SWAP);
3559 val = page_counter_read(&memcg->memory);
3561 val = page_counter_read(&memcg->memsw);
3574 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3577 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3578 struct page_counter *counter;
3580 switch (MEMFILE_TYPE(cft->private)) {
3582 counter = &memcg->memory;
3585 counter = &memcg->memsw;
3588 counter = &memcg->kmem;
3591 counter = &memcg->tcpmem;
3597 switch (MEMFILE_ATTR(cft->private)) {
3599 if (counter == &memcg->memory)
3600 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3601 if (counter == &memcg->memsw)
3602 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3603 return (u64)page_counter_read(counter) * PAGE_SIZE;
3605 return (u64)counter->max * PAGE_SIZE;
3607 return (u64)counter->watermark * PAGE_SIZE;
3609 return counter->failcnt;
3610 case RES_SOFT_LIMIT:
3611 return (u64)memcg->soft_limit * PAGE_SIZE;
3617 #ifdef CONFIG_MEMCG_KMEM
3618 static int memcg_online_kmem(struct mem_cgroup *memcg)
3620 struct obj_cgroup *objcg;
3623 if (cgroup_memory_nokmem)
3626 BUG_ON(memcg->kmemcg_id >= 0);
3628 memcg_id = memcg_alloc_cache_id();
3632 objcg = obj_cgroup_alloc();
3634 memcg_free_cache_id(memcg_id);
3637 objcg->memcg = memcg;
3638 rcu_assign_pointer(memcg->objcg, objcg);
3640 static_branch_enable(&memcg_kmem_enabled_key);
3642 memcg->kmemcg_id = memcg_id;
3647 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3649 struct mem_cgroup *parent;
3652 if (memcg->kmemcg_id == -1)
3655 parent = parent_mem_cgroup(memcg);
3657 parent = root_mem_cgroup;
3659 memcg_reparent_objcgs(memcg, parent);
3661 kmemcg_id = memcg->kmemcg_id;
3662 BUG_ON(kmemcg_id < 0);
3665 * After we have finished memcg_reparent_objcgs(), all list_lrus
3666 * corresponding to this cgroup are guaranteed to remain empty.
3667 * The ordering is imposed by list_lru_node->lock taken by
3668 * memcg_drain_all_list_lrus().
3670 memcg_drain_all_list_lrus(kmemcg_id, parent);
3672 memcg_free_cache_id(kmemcg_id);
3673 memcg->kmemcg_id = -1;
3676 static int memcg_online_kmem(struct mem_cgroup *memcg)
3680 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3683 #endif /* CONFIG_MEMCG_KMEM */
3685 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3689 mutex_lock(&memcg_max_mutex);
3691 ret = page_counter_set_max(&memcg->tcpmem, max);
3695 if (!memcg->tcpmem_active) {
3697 * The active flag needs to be written after the static_key
3698 * update. This is what guarantees that the socket activation
3699 * function is the last one to run. See mem_cgroup_sk_alloc()
3700 * for details, and note that we don't mark any socket as
3701 * belonging to this memcg until that flag is up.
3703 * We need to do this, because static_keys will span multiple
3704 * sites, but we can't control their order. If we mark a socket
3705 * as accounted, but the accounting functions are not patched in
3706 * yet, we'll lose accounting.
3708 * We never race with the readers in mem_cgroup_sk_alloc(),
3709 * because when this value change, the code to process it is not
3712 static_branch_inc(&memcg_sockets_enabled_key);
3713 memcg->tcpmem_active = true;
3716 mutex_unlock(&memcg_max_mutex);
3721 * The user of this function is...
3724 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3725 char *buf, size_t nbytes, loff_t off)
3727 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3728 unsigned long nr_pages;
3731 buf = strstrip(buf);
3732 ret = page_counter_memparse(buf, "-1", &nr_pages);
3736 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3738 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3742 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3744 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3747 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3750 /* kmem.limit_in_bytes is deprecated. */
3754 ret = memcg_update_tcp_max(memcg, nr_pages);
3758 case RES_SOFT_LIMIT:
3759 memcg->soft_limit = nr_pages;
3763 return ret ?: nbytes;
3766 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3767 size_t nbytes, loff_t off)
3769 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3770 struct page_counter *counter;
3772 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3774 counter = &memcg->memory;
3777 counter = &memcg->memsw;
3780 counter = &memcg->kmem;
3783 counter = &memcg->tcpmem;
3789 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3791 page_counter_reset_watermark(counter);
3794 counter->failcnt = 0;
3803 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3806 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3810 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3811 struct cftype *cft, u64 val)
3813 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3815 if (val & ~MOVE_MASK)
3819 * No kind of locking is needed in here, because ->can_attach() will
3820 * check this value once in the beginning of the process, and then carry
3821 * on with stale data. This means that changes to this value will only
3822 * affect task migrations starting after the change.
3824 memcg->move_charge_at_immigrate = val;
3828 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3829 struct cftype *cft, u64 val)
3837 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3838 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3839 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3841 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3842 int nid, unsigned int lru_mask, bool tree)
3844 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3845 unsigned long nr = 0;
3848 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3851 if (!(BIT(lru) & lru_mask))
3854 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3856 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3861 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3862 unsigned int lru_mask,
3865 unsigned long nr = 0;
3869 if (!(BIT(lru) & lru_mask))
3872 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3874 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3879 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3883 unsigned int lru_mask;
3886 static const struct numa_stat stats[] = {
3887 { "total", LRU_ALL },
3888 { "file", LRU_ALL_FILE },
3889 { "anon", LRU_ALL_ANON },
3890 { "unevictable", BIT(LRU_UNEVICTABLE) },
3892 const struct numa_stat *stat;
3894 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3896 mem_cgroup_flush_stats();
3898 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3899 seq_printf(m, "%s=%lu", stat->name,
3900 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3902 for_each_node_state(nid, N_MEMORY)
3903 seq_printf(m, " N%d=%lu", nid,
3904 mem_cgroup_node_nr_lru_pages(memcg, nid,
3905 stat->lru_mask, false));
3909 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3911 seq_printf(m, "hierarchical_%s=%lu", stat->name,
3912 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3914 for_each_node_state(nid, N_MEMORY)
3915 seq_printf(m, " N%d=%lu", nid,
3916 mem_cgroup_node_nr_lru_pages(memcg, nid,
3917 stat->lru_mask, true));
3923 #endif /* CONFIG_NUMA */
3925 static const unsigned int memcg1_stats[] = {
3928 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3938 static const char *const memcg1_stat_names[] = {
3941 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3951 /* Universal VM events cgroup1 shows, original sort order */
3952 static const unsigned int memcg1_events[] = {
3959 static int memcg_stat_show(struct seq_file *m, void *v)
3961 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3962 unsigned long memory, memsw;
3963 struct mem_cgroup *mi;
3966 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3968 mem_cgroup_flush_stats();
3970 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3973 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3975 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
3976 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
3979 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3980 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
3981 memcg_events_local(memcg, memcg1_events[i]));
3983 for (i = 0; i < NR_LRU_LISTS; i++)
3984 seq_printf(m, "%s %lu\n", lru_list_name(i),
3985 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3988 /* Hierarchical information */
3989 memory = memsw = PAGE_COUNTER_MAX;
3990 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3991 memory = min(memory, READ_ONCE(mi->memory.max));
3992 memsw = min(memsw, READ_ONCE(mi->memsw.max));
3994 seq_printf(m, "hierarchical_memory_limit %llu\n",
3995 (u64)memory * PAGE_SIZE);
3996 if (do_memsw_account())
3997 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3998 (u64)memsw * PAGE_SIZE);
4000 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4003 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4005 nr = memcg_page_state(memcg, memcg1_stats[i]);
4006 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4007 (u64)nr * PAGE_SIZE);
4010 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4011 seq_printf(m, "total_%s %llu\n",
4012 vm_event_name(memcg1_events[i]),
4013 (u64)memcg_events(memcg, memcg1_events[i]));
4015 for (i = 0; i < NR_LRU_LISTS; i++)
4016 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4017 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4020 #ifdef CONFIG_DEBUG_VM
4023 struct mem_cgroup_per_node *mz;
4024 unsigned long anon_cost = 0;
4025 unsigned long file_cost = 0;
4027 for_each_online_pgdat(pgdat) {
4028 mz = memcg->nodeinfo[pgdat->node_id];
4030 anon_cost += mz->lruvec.anon_cost;
4031 file_cost += mz->lruvec.file_cost;
4033 seq_printf(m, "anon_cost %lu\n", anon_cost);
4034 seq_printf(m, "file_cost %lu\n", file_cost);
4041 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4044 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4046 return mem_cgroup_swappiness(memcg);
4049 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4050 struct cftype *cft, u64 val)
4052 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4057 if (!mem_cgroup_is_root(memcg))
4058 memcg->swappiness = val;
4060 vm_swappiness = val;
4065 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4067 struct mem_cgroup_threshold_ary *t;
4068 unsigned long usage;
4073 t = rcu_dereference(memcg->thresholds.primary);
4075 t = rcu_dereference(memcg->memsw_thresholds.primary);
4080 usage = mem_cgroup_usage(memcg, swap);
4083 * current_threshold points to threshold just below or equal to usage.
4084 * If it's not true, a threshold was crossed after last
4085 * call of __mem_cgroup_threshold().
4087 i = t->current_threshold;
4090 * Iterate backward over array of thresholds starting from
4091 * current_threshold and check if a threshold is crossed.
4092 * If none of thresholds below usage is crossed, we read
4093 * only one element of the array here.
4095 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4096 eventfd_signal(t->entries[i].eventfd, 1);
4098 /* i = current_threshold + 1 */
4102 * Iterate forward over array of thresholds starting from
4103 * current_threshold+1 and check if a threshold is crossed.
4104 * If none of thresholds above usage is crossed, we read
4105 * only one element of the array here.
4107 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4108 eventfd_signal(t->entries[i].eventfd, 1);
4110 /* Update current_threshold */
4111 t->current_threshold = i - 1;
4116 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4119 __mem_cgroup_threshold(memcg, false);
4120 if (do_memsw_account())
4121 __mem_cgroup_threshold(memcg, true);
4123 memcg = parent_mem_cgroup(memcg);
4127 static int compare_thresholds(const void *a, const void *b)
4129 const struct mem_cgroup_threshold *_a = a;
4130 const struct mem_cgroup_threshold *_b = b;
4132 if (_a->threshold > _b->threshold)
4135 if (_a->threshold < _b->threshold)
4141 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4143 struct mem_cgroup_eventfd_list *ev;
4145 spin_lock(&memcg_oom_lock);
4147 list_for_each_entry(ev, &memcg->oom_notify, list)
4148 eventfd_signal(ev->eventfd, 1);
4150 spin_unlock(&memcg_oom_lock);
4154 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4156 struct mem_cgroup *iter;
4158 for_each_mem_cgroup_tree(iter, memcg)
4159 mem_cgroup_oom_notify_cb(iter);
4162 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4163 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4165 struct mem_cgroup_thresholds *thresholds;
4166 struct mem_cgroup_threshold_ary *new;
4167 unsigned long threshold;
4168 unsigned long usage;
4171 ret = page_counter_memparse(args, "-1", &threshold);
4175 mutex_lock(&memcg->thresholds_lock);
4178 thresholds = &memcg->thresholds;
4179 usage = mem_cgroup_usage(memcg, false);
4180 } else if (type == _MEMSWAP) {
4181 thresholds = &memcg->memsw_thresholds;
4182 usage = mem_cgroup_usage(memcg, true);
4186 /* Check if a threshold crossed before adding a new one */
4187 if (thresholds->primary)
4188 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4190 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4192 /* Allocate memory for new array of thresholds */
4193 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4200 /* Copy thresholds (if any) to new array */
4201 if (thresholds->primary)
4202 memcpy(new->entries, thresholds->primary->entries,
4203 flex_array_size(new, entries, size - 1));
4205 /* Add new threshold */
4206 new->entries[size - 1].eventfd = eventfd;
4207 new->entries[size - 1].threshold = threshold;
4209 /* Sort thresholds. Registering of new threshold isn't time-critical */
4210 sort(new->entries, size, sizeof(*new->entries),
4211 compare_thresholds, NULL);
4213 /* Find current threshold */
4214 new->current_threshold = -1;
4215 for (i = 0; i < size; i++) {
4216 if (new->entries[i].threshold <= usage) {
4218 * new->current_threshold will not be used until
4219 * rcu_assign_pointer(), so it's safe to increment
4222 ++new->current_threshold;
4227 /* Free old spare buffer and save old primary buffer as spare */
4228 kfree(thresholds->spare);
4229 thresholds->spare = thresholds->primary;
4231 rcu_assign_pointer(thresholds->primary, new);
4233 /* To be sure that nobody uses thresholds */
4237 mutex_unlock(&memcg->thresholds_lock);
4242 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4243 struct eventfd_ctx *eventfd, const char *args)
4245 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4248 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4249 struct eventfd_ctx *eventfd, const char *args)
4251 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4254 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4255 struct eventfd_ctx *eventfd, enum res_type type)
4257 struct mem_cgroup_thresholds *thresholds;
4258 struct mem_cgroup_threshold_ary *new;
4259 unsigned long usage;
4260 int i, j, size, entries;
4262 mutex_lock(&memcg->thresholds_lock);
4265 thresholds = &memcg->thresholds;
4266 usage = mem_cgroup_usage(memcg, false);
4267 } else if (type == _MEMSWAP) {
4268 thresholds = &memcg->memsw_thresholds;
4269 usage = mem_cgroup_usage(memcg, true);
4273 if (!thresholds->primary)
4276 /* Check if a threshold crossed before removing */
4277 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4279 /* Calculate new number of threshold */
4281 for (i = 0; i < thresholds->primary->size; i++) {
4282 if (thresholds->primary->entries[i].eventfd != eventfd)
4288 new = thresholds->spare;
4290 /* If no items related to eventfd have been cleared, nothing to do */
4294 /* Set thresholds array to NULL if we don't have thresholds */
4303 /* Copy thresholds and find current threshold */
4304 new->current_threshold = -1;
4305 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4306 if (thresholds->primary->entries[i].eventfd == eventfd)
4309 new->entries[j] = thresholds->primary->entries[i];
4310 if (new->entries[j].threshold <= usage) {
4312 * new->current_threshold will not be used
4313 * until rcu_assign_pointer(), so it's safe to increment
4316 ++new->current_threshold;
4322 /* Swap primary and spare array */
4323 thresholds->spare = thresholds->primary;
4325 rcu_assign_pointer(thresholds->primary, new);
4327 /* To be sure that nobody uses thresholds */
4330 /* If all events are unregistered, free the spare array */
4332 kfree(thresholds->spare);
4333 thresholds->spare = NULL;
4336 mutex_unlock(&memcg->thresholds_lock);
4339 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4340 struct eventfd_ctx *eventfd)
4342 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4345 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4346 struct eventfd_ctx *eventfd)
4348 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4351 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4352 struct eventfd_ctx *eventfd, const char *args)
4354 struct mem_cgroup_eventfd_list *event;
4356 event = kmalloc(sizeof(*event), GFP_KERNEL);
4360 spin_lock(&memcg_oom_lock);
4362 event->eventfd = eventfd;
4363 list_add(&event->list, &memcg->oom_notify);
4365 /* already in OOM ? */
4366 if (memcg->under_oom)
4367 eventfd_signal(eventfd, 1);
4368 spin_unlock(&memcg_oom_lock);
4373 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4374 struct eventfd_ctx *eventfd)
4376 struct mem_cgroup_eventfd_list *ev, *tmp;
4378 spin_lock(&memcg_oom_lock);
4380 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4381 if (ev->eventfd == eventfd) {
4382 list_del(&ev->list);
4387 spin_unlock(&memcg_oom_lock);
4390 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4392 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4394 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4395 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4396 seq_printf(sf, "oom_kill %lu\n",
4397 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4401 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4402 struct cftype *cft, u64 val)
4404 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4406 /* cannot set to root cgroup and only 0 and 1 are allowed */
4407 if (mem_cgroup_is_root(memcg) || !((val == 0) || (val == 1)))
4410 memcg->oom_kill_disable = val;
4412 memcg_oom_recover(memcg);
4417 #ifdef CONFIG_CGROUP_WRITEBACK
4419 #include <trace/events/writeback.h>
4421 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4423 return wb_domain_init(&memcg->cgwb_domain, gfp);
4426 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4428 wb_domain_exit(&memcg->cgwb_domain);
4431 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4433 wb_domain_size_changed(&memcg->cgwb_domain);
4436 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4438 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4440 if (!memcg->css.parent)
4443 return &memcg->cgwb_domain;
4447 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4448 * @wb: bdi_writeback in question
4449 * @pfilepages: out parameter for number of file pages
4450 * @pheadroom: out parameter for number of allocatable pages according to memcg
4451 * @pdirty: out parameter for number of dirty pages
4452 * @pwriteback: out parameter for number of pages under writeback
4454 * Determine the numbers of file, headroom, dirty, and writeback pages in
4455 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4456 * is a bit more involved.
4458 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4459 * headroom is calculated as the lowest headroom of itself and the
4460 * ancestors. Note that this doesn't consider the actual amount of
4461 * available memory in the system. The caller should further cap
4462 * *@pheadroom accordingly.
4464 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4465 unsigned long *pheadroom, unsigned long *pdirty,
4466 unsigned long *pwriteback)
4468 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4469 struct mem_cgroup *parent;
4471 mem_cgroup_flush_stats();
4473 *pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
4474 *pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
4475 *pfilepages = memcg_page_state(memcg, NR_INACTIVE_FILE) +
4476 memcg_page_state(memcg, NR_ACTIVE_FILE);
4478 *pheadroom = PAGE_COUNTER_MAX;
4479 while ((parent = parent_mem_cgroup(memcg))) {
4480 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4481 READ_ONCE(memcg->memory.high));
4482 unsigned long used = page_counter_read(&memcg->memory);
4484 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4490 * Foreign dirty flushing
4492 * There's an inherent mismatch between memcg and writeback. The former
4493 * tracks ownership per-page while the latter per-inode. This was a
4494 * deliberate design decision because honoring per-page ownership in the
4495 * writeback path is complicated, may lead to higher CPU and IO overheads
4496 * and deemed unnecessary given that write-sharing an inode across
4497 * different cgroups isn't a common use-case.
4499 * Combined with inode majority-writer ownership switching, this works well
4500 * enough in most cases but there are some pathological cases. For
4501 * example, let's say there are two cgroups A and B which keep writing to
4502 * different but confined parts of the same inode. B owns the inode and
4503 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4504 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4505 * triggering background writeback. A will be slowed down without a way to
4506 * make writeback of the dirty pages happen.
4508 * Conditions like the above can lead to a cgroup getting repeatedly and
4509 * severely throttled after making some progress after each
4510 * dirty_expire_interval while the underlying IO device is almost
4513 * Solving this problem completely requires matching the ownership tracking
4514 * granularities between memcg and writeback in either direction. However,
4515 * the more egregious behaviors can be avoided by simply remembering the
4516 * most recent foreign dirtying events and initiating remote flushes on
4517 * them when local writeback isn't enough to keep the memory clean enough.
4519 * The following two functions implement such mechanism. When a foreign
4520 * page - a page whose memcg and writeback ownerships don't match - is
4521 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4522 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4523 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4524 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4525 * foreign bdi_writebacks which haven't expired. Both the numbers of
4526 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4527 * limited to MEMCG_CGWB_FRN_CNT.
4529 * The mechanism only remembers IDs and doesn't hold any object references.
4530 * As being wrong occasionally doesn't matter, updates and accesses to the
4531 * records are lockless and racy.
4533 void mem_cgroup_track_foreign_dirty_slowpath(struct folio *folio,
4534 struct bdi_writeback *wb)
4536 struct mem_cgroup *memcg = folio_memcg(folio);
4537 struct memcg_cgwb_frn *frn;
4538 u64 now = get_jiffies_64();
4539 u64 oldest_at = now;
4543 trace_track_foreign_dirty(folio, wb);
4546 * Pick the slot to use. If there is already a slot for @wb, keep
4547 * using it. If not replace the oldest one which isn't being
4550 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4551 frn = &memcg->cgwb_frn[i];
4552 if (frn->bdi_id == wb->bdi->id &&
4553 frn->memcg_id == wb->memcg_css->id)
4555 if (time_before64(frn->at, oldest_at) &&
4556 atomic_read(&frn->done.cnt) == 1) {
4558 oldest_at = frn->at;
4562 if (i < MEMCG_CGWB_FRN_CNT) {
4564 * Re-using an existing one. Update timestamp lazily to
4565 * avoid making the cacheline hot. We want them to be
4566 * reasonably up-to-date and significantly shorter than
4567 * dirty_expire_interval as that's what expires the record.
4568 * Use the shorter of 1s and dirty_expire_interval / 8.
4570 unsigned long update_intv =
4571 min_t(unsigned long, HZ,
4572 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4574 if (time_before64(frn->at, now - update_intv))
4576 } else if (oldest >= 0) {
4577 /* replace the oldest free one */
4578 frn = &memcg->cgwb_frn[oldest];
4579 frn->bdi_id = wb->bdi->id;
4580 frn->memcg_id = wb->memcg_css->id;
4585 /* issue foreign writeback flushes for recorded foreign dirtying events */
4586 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4588 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4589 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4590 u64 now = jiffies_64;
4593 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4594 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4597 * If the record is older than dirty_expire_interval,
4598 * writeback on it has already started. No need to kick it
4599 * off again. Also, don't start a new one if there's
4600 * already one in flight.
4602 if (time_after64(frn->at, now - intv) &&
4603 atomic_read(&frn->done.cnt) == 1) {
4605 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4606 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id,
4607 WB_REASON_FOREIGN_FLUSH,
4613 #else /* CONFIG_CGROUP_WRITEBACK */
4615 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4620 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4624 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4628 #endif /* CONFIG_CGROUP_WRITEBACK */
4631 * DO NOT USE IN NEW FILES.
4633 * "cgroup.event_control" implementation.
4635 * This is way over-engineered. It tries to support fully configurable
4636 * events for each user. Such level of flexibility is completely
4637 * unnecessary especially in the light of the planned unified hierarchy.
4639 * Please deprecate this and replace with something simpler if at all
4644 * Unregister event and free resources.
4646 * Gets called from workqueue.
4648 static void memcg_event_remove(struct work_struct *work)
4650 struct mem_cgroup_event *event =
4651 container_of(work, struct mem_cgroup_event, remove);
4652 struct mem_cgroup *memcg = event->memcg;
4654 remove_wait_queue(event->wqh, &event->wait);
4656 event->unregister_event(memcg, event->eventfd);
4658 /* Notify userspace the event is going away. */
4659 eventfd_signal(event->eventfd, 1);
4661 eventfd_ctx_put(event->eventfd);
4663 css_put(&memcg->css);
4667 * Gets called on EPOLLHUP on eventfd when user closes it.
4669 * Called with wqh->lock held and interrupts disabled.
4671 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4672 int sync, void *key)
4674 struct mem_cgroup_event *event =
4675 container_of(wait, struct mem_cgroup_event, wait);
4676 struct mem_cgroup *memcg = event->memcg;
4677 __poll_t flags = key_to_poll(key);
4679 if (flags & EPOLLHUP) {
4681 * If the event has been detached at cgroup removal, we
4682 * can simply return knowing the other side will cleanup
4685 * We can't race against event freeing since the other
4686 * side will require wqh->lock via remove_wait_queue(),
4689 spin_lock(&memcg->event_list_lock);
4690 if (!list_empty(&event->list)) {
4691 list_del_init(&event->list);
4693 * We are in atomic context, but cgroup_event_remove()
4694 * may sleep, so we have to call it in workqueue.
4696 schedule_work(&event->remove);
4698 spin_unlock(&memcg->event_list_lock);
4704 static void memcg_event_ptable_queue_proc(struct file *file,
4705 wait_queue_head_t *wqh, poll_table *pt)
4707 struct mem_cgroup_event *event =
4708 container_of(pt, struct mem_cgroup_event, pt);
4711 add_wait_queue(wqh, &event->wait);
4715 * DO NOT USE IN NEW FILES.
4717 * Parse input and register new cgroup event handler.
4719 * Input must be in format '<event_fd> <control_fd> <args>'.
4720 * Interpretation of args is defined by control file implementation.
4722 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4723 char *buf, size_t nbytes, loff_t off)
4725 struct cgroup_subsys_state *css = of_css(of);
4726 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4727 struct mem_cgroup_event *event;
4728 struct cgroup_subsys_state *cfile_css;
4729 unsigned int efd, cfd;
4736 buf = strstrip(buf);
4738 efd = simple_strtoul(buf, &endp, 10);
4743 cfd = simple_strtoul(buf, &endp, 10);
4744 if ((*endp != ' ') && (*endp != '\0'))
4748 event = kzalloc(sizeof(*event), GFP_KERNEL);
4752 event->memcg = memcg;
4753 INIT_LIST_HEAD(&event->list);
4754 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4755 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4756 INIT_WORK(&event->remove, memcg_event_remove);
4764 event->eventfd = eventfd_ctx_fileget(efile.file);
4765 if (IS_ERR(event->eventfd)) {
4766 ret = PTR_ERR(event->eventfd);
4773 goto out_put_eventfd;
4776 /* the process need read permission on control file */
4777 /* AV: shouldn't we check that it's been opened for read instead? */
4778 ret = file_permission(cfile.file, MAY_READ);
4783 * Determine the event callbacks and set them in @event. This used
4784 * to be done via struct cftype but cgroup core no longer knows
4785 * about these events. The following is crude but the whole thing
4786 * is for compatibility anyway.
4788 * DO NOT ADD NEW FILES.
4790 name = cfile.file->f_path.dentry->d_name.name;
4792 if (!strcmp(name, "memory.usage_in_bytes")) {
4793 event->register_event = mem_cgroup_usage_register_event;
4794 event->unregister_event = mem_cgroup_usage_unregister_event;
4795 } else if (!strcmp(name, "memory.oom_control")) {
4796 event->register_event = mem_cgroup_oom_register_event;
4797 event->unregister_event = mem_cgroup_oom_unregister_event;
4798 } else if (!strcmp(name, "memory.pressure_level")) {
4799 event->register_event = vmpressure_register_event;
4800 event->unregister_event = vmpressure_unregister_event;
4801 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4802 event->register_event = memsw_cgroup_usage_register_event;
4803 event->unregister_event = memsw_cgroup_usage_unregister_event;
4810 * Verify @cfile should belong to @css. Also, remaining events are
4811 * automatically removed on cgroup destruction but the removal is
4812 * asynchronous, so take an extra ref on @css.
4814 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4815 &memory_cgrp_subsys);
4817 if (IS_ERR(cfile_css))
4819 if (cfile_css != css) {
4824 ret = event->register_event(memcg, event->eventfd, buf);
4828 vfs_poll(efile.file, &event->pt);
4830 spin_lock_irq(&memcg->event_list_lock);
4831 list_add(&event->list, &memcg->event_list);
4832 spin_unlock_irq(&memcg->event_list_lock);
4844 eventfd_ctx_put(event->eventfd);
4853 static struct cftype mem_cgroup_legacy_files[] = {
4855 .name = "usage_in_bytes",
4856 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4857 .read_u64 = mem_cgroup_read_u64,
4860 .name = "max_usage_in_bytes",
4861 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4862 .write = mem_cgroup_reset,
4863 .read_u64 = mem_cgroup_read_u64,
4866 .name = "limit_in_bytes",
4867 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4868 .write = mem_cgroup_write,
4869 .read_u64 = mem_cgroup_read_u64,
4872 .name = "soft_limit_in_bytes",
4873 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4874 .write = mem_cgroup_write,
4875 .read_u64 = mem_cgroup_read_u64,
4879 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4880 .write = mem_cgroup_reset,
4881 .read_u64 = mem_cgroup_read_u64,
4885 .seq_show = memcg_stat_show,
4888 .name = "force_empty",
4889 .write = mem_cgroup_force_empty_write,
4892 .name = "use_hierarchy",
4893 .write_u64 = mem_cgroup_hierarchy_write,
4894 .read_u64 = mem_cgroup_hierarchy_read,
4897 .name = "cgroup.event_control", /* XXX: for compat */
4898 .write = memcg_write_event_control,
4899 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4902 .name = "swappiness",
4903 .read_u64 = mem_cgroup_swappiness_read,
4904 .write_u64 = mem_cgroup_swappiness_write,
4907 .name = "move_charge_at_immigrate",
4908 .read_u64 = mem_cgroup_move_charge_read,
4909 .write_u64 = mem_cgroup_move_charge_write,
4912 .name = "oom_control",
4913 .seq_show = mem_cgroup_oom_control_read,
4914 .write_u64 = mem_cgroup_oom_control_write,
4915 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4918 .name = "pressure_level",
4922 .name = "numa_stat",
4923 .seq_show = memcg_numa_stat_show,
4927 .name = "kmem.limit_in_bytes",
4928 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4929 .write = mem_cgroup_write,
4930 .read_u64 = mem_cgroup_read_u64,
4933 .name = "kmem.usage_in_bytes",
4934 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4935 .read_u64 = mem_cgroup_read_u64,
4938 .name = "kmem.failcnt",
4939 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4940 .write = mem_cgroup_reset,
4941 .read_u64 = mem_cgroup_read_u64,
4944 .name = "kmem.max_usage_in_bytes",
4945 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4946 .write = mem_cgroup_reset,
4947 .read_u64 = mem_cgroup_read_u64,
4949 #if defined(CONFIG_MEMCG_KMEM) && \
4950 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
4952 .name = "kmem.slabinfo",
4953 .seq_show = memcg_slab_show,
4957 .name = "kmem.tcp.limit_in_bytes",
4958 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4959 .write = mem_cgroup_write,
4960 .read_u64 = mem_cgroup_read_u64,
4963 .name = "kmem.tcp.usage_in_bytes",
4964 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4965 .read_u64 = mem_cgroup_read_u64,
4968 .name = "kmem.tcp.failcnt",
4969 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4970 .write = mem_cgroup_reset,
4971 .read_u64 = mem_cgroup_read_u64,
4974 .name = "kmem.tcp.max_usage_in_bytes",
4975 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4976 .write = mem_cgroup_reset,
4977 .read_u64 = mem_cgroup_read_u64,
4979 { }, /* terminate */
4983 * Private memory cgroup IDR
4985 * Swap-out records and page cache shadow entries need to store memcg
4986 * references in constrained space, so we maintain an ID space that is
4987 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4988 * memory-controlled cgroups to 64k.
4990 * However, there usually are many references to the offline CSS after
4991 * the cgroup has been destroyed, such as page cache or reclaimable
4992 * slab objects, that don't need to hang on to the ID. We want to keep
4993 * those dead CSS from occupying IDs, or we might quickly exhaust the
4994 * relatively small ID space and prevent the creation of new cgroups
4995 * even when there are much fewer than 64k cgroups - possibly none.
4997 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4998 * be freed and recycled when it's no longer needed, which is usually
4999 * when the CSS is offlined.
5001 * The only exception to that are records of swapped out tmpfs/shmem
5002 * pages that need to be attributed to live ancestors on swapin. But
5003 * those references are manageable from userspace.
5006 static DEFINE_IDR(mem_cgroup_idr);
5008 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5010 if (memcg->id.id > 0) {
5011 idr_remove(&mem_cgroup_idr, memcg->id.id);
5016 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5019 refcount_add(n, &memcg->id.ref);
5022 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5024 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5025 mem_cgroup_id_remove(memcg);
5027 /* Memcg ID pins CSS */
5028 css_put(&memcg->css);
5032 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5034 mem_cgroup_id_put_many(memcg, 1);
5038 * mem_cgroup_from_id - look up a memcg from a memcg id
5039 * @id: the memcg id to look up
5041 * Caller must hold rcu_read_lock().
5043 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5045 WARN_ON_ONCE(!rcu_read_lock_held());
5046 return idr_find(&mem_cgroup_idr, id);
5049 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5051 struct mem_cgroup_per_node *pn;
5054 * This routine is called against possible nodes.
5055 * But it's BUG to call kmalloc() against offline node.
5057 * TODO: this routine can waste much memory for nodes which will
5058 * never be onlined. It's better to use memory hotplug callback
5061 if (!node_state(node, N_NORMAL_MEMORY))
5063 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5067 pn->lruvec_stats_percpu = alloc_percpu_gfp(struct lruvec_stats_percpu,
5068 GFP_KERNEL_ACCOUNT);
5069 if (!pn->lruvec_stats_percpu) {
5074 lruvec_init(&pn->lruvec);
5075 pn->usage_in_excess = 0;
5076 pn->on_tree = false;
5079 memcg->nodeinfo[node] = pn;
5083 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5085 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5090 free_percpu(pn->lruvec_stats_percpu);
5094 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5099 free_mem_cgroup_per_node_info(memcg, node);
5100 free_percpu(memcg->vmstats_percpu);
5104 static void mem_cgroup_free(struct mem_cgroup *memcg)
5106 memcg_wb_domain_exit(memcg);
5107 __mem_cgroup_free(memcg);
5110 static struct mem_cgroup *mem_cgroup_alloc(void)
5112 struct mem_cgroup *memcg;
5115 int __maybe_unused i;
5116 long error = -ENOMEM;
5118 size = sizeof(struct mem_cgroup);
5119 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5121 memcg = kzalloc(size, GFP_KERNEL);
5123 return ERR_PTR(error);
5125 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5126 1, MEM_CGROUP_ID_MAX,
5128 if (memcg->id.id < 0) {
5129 error = memcg->id.id;
5133 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5134 GFP_KERNEL_ACCOUNT);
5135 if (!memcg->vmstats_percpu)
5139 if (alloc_mem_cgroup_per_node_info(memcg, node))
5142 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5145 INIT_WORK(&memcg->high_work, high_work_func);
5146 INIT_LIST_HEAD(&memcg->oom_notify);
5147 mutex_init(&memcg->thresholds_lock);
5148 spin_lock_init(&memcg->move_lock);
5149 vmpressure_init(&memcg->vmpressure);
5150 INIT_LIST_HEAD(&memcg->event_list);
5151 spin_lock_init(&memcg->event_list_lock);
5152 memcg->socket_pressure = jiffies;
5153 #ifdef CONFIG_MEMCG_KMEM
5154 memcg->kmemcg_id = -1;
5155 INIT_LIST_HEAD(&memcg->objcg_list);
5157 #ifdef CONFIG_CGROUP_WRITEBACK
5158 INIT_LIST_HEAD(&memcg->cgwb_list);
5159 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5160 memcg->cgwb_frn[i].done =
5161 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5163 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5164 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5165 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5166 memcg->deferred_split_queue.split_queue_len = 0;
5168 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5171 mem_cgroup_id_remove(memcg);
5172 __mem_cgroup_free(memcg);
5173 return ERR_PTR(error);
5176 static struct cgroup_subsys_state * __ref
5177 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5179 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5180 struct mem_cgroup *memcg, *old_memcg;
5181 long error = -ENOMEM;
5183 old_memcg = set_active_memcg(parent);
5184 memcg = mem_cgroup_alloc();
5185 set_active_memcg(old_memcg);
5187 return ERR_CAST(memcg);
5189 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5190 memcg->soft_limit = PAGE_COUNTER_MAX;
5191 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5193 memcg->swappiness = mem_cgroup_swappiness(parent);
5194 memcg->oom_kill_disable = parent->oom_kill_disable;
5196 page_counter_init(&memcg->memory, &parent->memory);
5197 page_counter_init(&memcg->swap, &parent->swap);
5198 page_counter_init(&memcg->kmem, &parent->kmem);
5199 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5201 page_counter_init(&memcg->memory, NULL);
5202 page_counter_init(&memcg->swap, NULL);
5203 page_counter_init(&memcg->kmem, NULL);
5204 page_counter_init(&memcg->tcpmem, NULL);
5206 root_mem_cgroup = memcg;
5210 /* The following stuff does not apply to the root */
5211 error = memcg_online_kmem(memcg);
5215 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5216 static_branch_inc(&memcg_sockets_enabled_key);
5220 mem_cgroup_id_remove(memcg);
5221 mem_cgroup_free(memcg);
5222 return ERR_PTR(error);
5225 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5227 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5230 * A memcg must be visible for expand_shrinker_info()
5231 * by the time the maps are allocated. So, we allocate maps
5232 * here, when for_each_mem_cgroup() can't skip it.
5234 if (alloc_shrinker_info(memcg)) {
5235 mem_cgroup_id_remove(memcg);
5239 /* Online state pins memcg ID, memcg ID pins CSS */
5240 refcount_set(&memcg->id.ref, 1);
5243 if (unlikely(mem_cgroup_is_root(memcg)))
5244 queue_delayed_work(system_unbound_wq, &stats_flush_dwork,
5249 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5251 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5252 struct mem_cgroup_event *event, *tmp;
5255 * Unregister events and notify userspace.
5256 * Notify userspace about cgroup removing only after rmdir of cgroup
5257 * directory to avoid race between userspace and kernelspace.
5259 spin_lock_irq(&memcg->event_list_lock);
5260 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5261 list_del_init(&event->list);
5262 schedule_work(&event->remove);
5264 spin_unlock_irq(&memcg->event_list_lock);
5266 page_counter_set_min(&memcg->memory, 0);
5267 page_counter_set_low(&memcg->memory, 0);
5269 memcg_offline_kmem(memcg);
5270 reparent_shrinker_deferred(memcg);
5271 wb_memcg_offline(memcg);
5273 drain_all_stock(memcg);
5275 mem_cgroup_id_put(memcg);
5278 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5280 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5282 invalidate_reclaim_iterators(memcg);
5285 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5287 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5288 int __maybe_unused i;
5290 #ifdef CONFIG_CGROUP_WRITEBACK
5291 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5292 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5294 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5295 static_branch_dec(&memcg_sockets_enabled_key);
5297 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5298 static_branch_dec(&memcg_sockets_enabled_key);
5300 vmpressure_cleanup(&memcg->vmpressure);
5301 cancel_work_sync(&memcg->high_work);
5302 mem_cgroup_remove_from_trees(memcg);
5303 free_shrinker_info(memcg);
5305 /* Need to offline kmem if online_css() fails */
5306 memcg_offline_kmem(memcg);
5307 mem_cgroup_free(memcg);
5311 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5312 * @css: the target css
5314 * Reset the states of the mem_cgroup associated with @css. This is
5315 * invoked when the userland requests disabling on the default hierarchy
5316 * but the memcg is pinned through dependency. The memcg should stop
5317 * applying policies and should revert to the vanilla state as it may be
5318 * made visible again.
5320 * The current implementation only resets the essential configurations.
5321 * This needs to be expanded to cover all the visible parts.
5323 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5325 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5327 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5328 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5329 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5330 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5331 page_counter_set_min(&memcg->memory, 0);
5332 page_counter_set_low(&memcg->memory, 0);
5333 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5334 memcg->soft_limit = PAGE_COUNTER_MAX;
5335 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5336 memcg_wb_domain_size_changed(memcg);
5339 static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state *css, int cpu)
5341 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5342 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5343 struct memcg_vmstats_percpu *statc;
5347 statc = per_cpu_ptr(memcg->vmstats_percpu, cpu);
5349 for (i = 0; i < MEMCG_NR_STAT; i++) {
5351 * Collect the aggregated propagation counts of groups
5352 * below us. We're in a per-cpu loop here and this is
5353 * a global counter, so the first cycle will get them.
5355 delta = memcg->vmstats.state_pending[i];
5357 memcg->vmstats.state_pending[i] = 0;
5359 /* Add CPU changes on this level since the last flush */
5360 v = READ_ONCE(statc->state[i]);
5361 if (v != statc->state_prev[i]) {
5362 delta += v - statc->state_prev[i];
5363 statc->state_prev[i] = v;
5369 /* Aggregate counts on this level and propagate upwards */
5370 memcg->vmstats.state[i] += delta;
5372 parent->vmstats.state_pending[i] += delta;
5375 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
5376 delta = memcg->vmstats.events_pending[i];
5378 memcg->vmstats.events_pending[i] = 0;
5380 v = READ_ONCE(statc->events[i]);
5381 if (v != statc->events_prev[i]) {
5382 delta += v - statc->events_prev[i];
5383 statc->events_prev[i] = v;
5389 memcg->vmstats.events[i] += delta;
5391 parent->vmstats.events_pending[i] += delta;
5394 for_each_node_state(nid, N_MEMORY) {
5395 struct mem_cgroup_per_node *pn = memcg->nodeinfo[nid];
5396 struct mem_cgroup_per_node *ppn = NULL;
5397 struct lruvec_stats_percpu *lstatc;
5400 ppn = parent->nodeinfo[nid];
5402 lstatc = per_cpu_ptr(pn->lruvec_stats_percpu, cpu);
5404 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
5405 delta = pn->lruvec_stats.state_pending[i];
5407 pn->lruvec_stats.state_pending[i] = 0;
5409 v = READ_ONCE(lstatc->state[i]);
5410 if (v != lstatc->state_prev[i]) {
5411 delta += v - lstatc->state_prev[i];
5412 lstatc->state_prev[i] = v;
5418 pn->lruvec_stats.state[i] += delta;
5420 ppn->lruvec_stats.state_pending[i] += delta;
5426 /* Handlers for move charge at task migration. */
5427 static int mem_cgroup_do_precharge(unsigned long count)
5431 /* Try a single bulk charge without reclaim first, kswapd may wake */
5432 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5434 mc.precharge += count;
5438 /* Try charges one by one with reclaim, but do not retry */
5440 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5454 enum mc_target_type {
5461 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5462 unsigned long addr, pte_t ptent)
5464 struct page *page = vm_normal_page(vma, addr, ptent);
5466 if (!page || !page_mapped(page))
5468 if (PageAnon(page)) {
5469 if (!(mc.flags & MOVE_ANON))
5472 if (!(mc.flags & MOVE_FILE))
5475 if (!get_page_unless_zero(page))
5481 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5482 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5483 pte_t ptent, swp_entry_t *entry)
5485 struct page *page = NULL;
5486 swp_entry_t ent = pte_to_swp_entry(ptent);
5488 if (!(mc.flags & MOVE_ANON))
5492 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5493 * a device and because they are not accessible by CPU they are store
5494 * as special swap entry in the CPU page table.
5496 if (is_device_private_entry(ent)) {
5497 page = pfn_swap_entry_to_page(ent);
5499 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5500 * a refcount of 1 when free (unlike normal page)
5502 if (!page_ref_add_unless(page, 1, 1))
5507 if (non_swap_entry(ent))
5511 * Because lookup_swap_cache() updates some statistics counter,
5512 * we call find_get_page() with swapper_space directly.
5514 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5515 entry->val = ent.val;
5520 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5521 pte_t ptent, swp_entry_t *entry)
5527 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5528 unsigned long addr, pte_t ptent)
5530 if (!vma->vm_file) /* anonymous vma */
5532 if (!(mc.flags & MOVE_FILE))
5535 /* page is moved even if it's not RSS of this task(page-faulted). */
5536 /* shmem/tmpfs may report page out on swap: account for that too. */
5537 return find_get_incore_page(vma->vm_file->f_mapping,
5538 linear_page_index(vma, addr));
5542 * mem_cgroup_move_account - move account of the page
5544 * @compound: charge the page as compound or small page
5545 * @from: mem_cgroup which the page is moved from.
5546 * @to: mem_cgroup which the page is moved to. @from != @to.
5548 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5550 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5553 static int mem_cgroup_move_account(struct page *page,
5555 struct mem_cgroup *from,
5556 struct mem_cgroup *to)
5558 struct folio *folio = page_folio(page);
5559 struct lruvec *from_vec, *to_vec;
5560 struct pglist_data *pgdat;
5561 unsigned int nr_pages = compound ? folio_nr_pages(folio) : 1;
5564 VM_BUG_ON(from == to);
5565 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
5566 VM_BUG_ON(compound && !folio_test_large(folio));
5569 * Prevent mem_cgroup_migrate() from looking at
5570 * page's memory cgroup of its source page while we change it.
5573 if (!folio_trylock(folio))
5577 if (folio_memcg(folio) != from)
5580 pgdat = folio_pgdat(folio);
5581 from_vec = mem_cgroup_lruvec(from, pgdat);
5582 to_vec = mem_cgroup_lruvec(to, pgdat);
5584 folio_memcg_lock(folio);
5586 if (folio_test_anon(folio)) {
5587 if (folio_mapped(folio)) {
5588 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5589 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5590 if (folio_test_transhuge(folio)) {
5591 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5593 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5598 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5599 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5601 if (folio_test_swapbacked(folio)) {
5602 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5603 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5606 if (folio_mapped(folio)) {
5607 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5608 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5611 if (folio_test_dirty(folio)) {
5612 struct address_space *mapping = folio_mapping(folio);
5614 if (mapping_can_writeback(mapping)) {
5615 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5617 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5623 if (folio_test_writeback(folio)) {
5624 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5625 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5629 * All state has been migrated, let's switch to the new memcg.
5631 * It is safe to change page's memcg here because the page
5632 * is referenced, charged, isolated, and locked: we can't race
5633 * with (un)charging, migration, LRU putback, or anything else
5634 * that would rely on a stable page's memory cgroup.
5636 * Note that lock_page_memcg is a memcg lock, not a page lock,
5637 * to save space. As soon as we switch page's memory cgroup to a
5638 * new memcg that isn't locked, the above state can change
5639 * concurrently again. Make sure we're truly done with it.
5644 css_put(&from->css);
5646 folio->memcg_data = (unsigned long)to;
5648 __folio_memcg_unlock(from);
5651 nid = folio_nid(folio);
5653 local_irq_disable();
5654 mem_cgroup_charge_statistics(to, nr_pages);
5655 memcg_check_events(to, nid);
5656 mem_cgroup_charge_statistics(from, -nr_pages);
5657 memcg_check_events(from, nid);
5660 folio_unlock(folio);
5666 * get_mctgt_type - get target type of moving charge
5667 * @vma: the vma the pte to be checked belongs
5668 * @addr: the address corresponding to the pte to be checked
5669 * @ptent: the pte to be checked
5670 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5673 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5674 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5675 * move charge. if @target is not NULL, the page is stored in target->page
5676 * with extra refcnt got(Callers should handle it).
5677 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5678 * target for charge migration. if @target is not NULL, the entry is stored
5680 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5681 * (so ZONE_DEVICE page and thus not on the lru).
5682 * For now we such page is charge like a regular page would be as for all
5683 * intent and purposes it is just special memory taking the place of a
5686 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5688 * Called with pte lock held.
5691 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5692 unsigned long addr, pte_t ptent, union mc_target *target)
5694 struct page *page = NULL;
5695 enum mc_target_type ret = MC_TARGET_NONE;
5696 swp_entry_t ent = { .val = 0 };
5698 if (pte_present(ptent))
5699 page = mc_handle_present_pte(vma, addr, ptent);
5700 else if (is_swap_pte(ptent))
5701 page = mc_handle_swap_pte(vma, ptent, &ent);
5702 else if (pte_none(ptent))
5703 page = mc_handle_file_pte(vma, addr, ptent);
5705 if (!page && !ent.val)
5709 * Do only loose check w/o serialization.
5710 * mem_cgroup_move_account() checks the page is valid or
5711 * not under LRU exclusion.
5713 if (page_memcg(page) == mc.from) {
5714 ret = MC_TARGET_PAGE;
5715 if (is_device_private_page(page))
5716 ret = MC_TARGET_DEVICE;
5718 target->page = page;
5720 if (!ret || !target)
5724 * There is a swap entry and a page doesn't exist or isn't charged.
5725 * But we cannot move a tail-page in a THP.
5727 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5728 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5729 ret = MC_TARGET_SWAP;
5736 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5738 * We don't consider PMD mapped swapping or file mapped pages because THP does
5739 * not support them for now.
5740 * Caller should make sure that pmd_trans_huge(pmd) is true.
5742 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5743 unsigned long addr, pmd_t pmd, union mc_target *target)
5745 struct page *page = NULL;
5746 enum mc_target_type ret = MC_TARGET_NONE;
5748 if (unlikely(is_swap_pmd(pmd))) {
5749 VM_BUG_ON(thp_migration_supported() &&
5750 !is_pmd_migration_entry(pmd));
5753 page = pmd_page(pmd);
5754 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5755 if (!(mc.flags & MOVE_ANON))
5757 if (page_memcg(page) == mc.from) {
5758 ret = MC_TARGET_PAGE;
5761 target->page = page;
5767 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5768 unsigned long addr, pmd_t pmd, union mc_target *target)
5770 return MC_TARGET_NONE;
5774 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5775 unsigned long addr, unsigned long end,
5776 struct mm_walk *walk)
5778 struct vm_area_struct *vma = walk->vma;
5782 ptl = pmd_trans_huge_lock(pmd, vma);
5785 * Note their can not be MC_TARGET_DEVICE for now as we do not
5786 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5787 * this might change.
5789 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5790 mc.precharge += HPAGE_PMD_NR;
5795 if (pmd_trans_unstable(pmd))
5797 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5798 for (; addr != end; pte++, addr += PAGE_SIZE)
5799 if (get_mctgt_type(vma, addr, *pte, NULL))
5800 mc.precharge++; /* increment precharge temporarily */
5801 pte_unmap_unlock(pte - 1, ptl);
5807 static const struct mm_walk_ops precharge_walk_ops = {
5808 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5811 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5813 unsigned long precharge;
5816 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5817 mmap_read_unlock(mm);
5819 precharge = mc.precharge;
5825 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5827 unsigned long precharge = mem_cgroup_count_precharge(mm);
5829 VM_BUG_ON(mc.moving_task);
5830 mc.moving_task = current;
5831 return mem_cgroup_do_precharge(precharge);
5834 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5835 static void __mem_cgroup_clear_mc(void)
5837 struct mem_cgroup *from = mc.from;
5838 struct mem_cgroup *to = mc.to;
5840 /* we must uncharge all the leftover precharges from mc.to */
5842 cancel_charge(mc.to, mc.precharge);
5846 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5847 * we must uncharge here.
5849 if (mc.moved_charge) {
5850 cancel_charge(mc.from, mc.moved_charge);
5851 mc.moved_charge = 0;
5853 /* we must fixup refcnts and charges */
5854 if (mc.moved_swap) {
5855 /* uncharge swap account from the old cgroup */
5856 if (!mem_cgroup_is_root(mc.from))
5857 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5859 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5862 * we charged both to->memory and to->memsw, so we
5863 * should uncharge to->memory.
5865 if (!mem_cgroup_is_root(mc.to))
5866 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5870 memcg_oom_recover(from);
5871 memcg_oom_recover(to);
5872 wake_up_all(&mc.waitq);
5875 static void mem_cgroup_clear_mc(void)
5877 struct mm_struct *mm = mc.mm;
5880 * we must clear moving_task before waking up waiters at the end of
5883 mc.moving_task = NULL;
5884 __mem_cgroup_clear_mc();
5885 spin_lock(&mc.lock);
5889 spin_unlock(&mc.lock);
5894 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5896 struct cgroup_subsys_state *css;
5897 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5898 struct mem_cgroup *from;
5899 struct task_struct *leader, *p;
5900 struct mm_struct *mm;
5901 unsigned long move_flags;
5904 /* charge immigration isn't supported on the default hierarchy */
5905 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5909 * Multi-process migrations only happen on the default hierarchy
5910 * where charge immigration is not used. Perform charge
5911 * immigration if @tset contains a leader and whine if there are
5915 cgroup_taskset_for_each_leader(leader, css, tset) {
5918 memcg = mem_cgroup_from_css(css);
5924 * We are now committed to this value whatever it is. Changes in this
5925 * tunable will only affect upcoming migrations, not the current one.
5926 * So we need to save it, and keep it going.
5928 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5932 from = mem_cgroup_from_task(p);
5934 VM_BUG_ON(from == memcg);
5936 mm = get_task_mm(p);
5939 /* We move charges only when we move a owner of the mm */
5940 if (mm->owner == p) {
5943 VM_BUG_ON(mc.precharge);
5944 VM_BUG_ON(mc.moved_charge);
5945 VM_BUG_ON(mc.moved_swap);
5947 spin_lock(&mc.lock);
5951 mc.flags = move_flags;
5952 spin_unlock(&mc.lock);
5953 /* We set mc.moving_task later */
5955 ret = mem_cgroup_precharge_mc(mm);
5957 mem_cgroup_clear_mc();
5964 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5967 mem_cgroup_clear_mc();
5970 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5971 unsigned long addr, unsigned long end,
5972 struct mm_walk *walk)
5975 struct vm_area_struct *vma = walk->vma;
5978 enum mc_target_type target_type;
5979 union mc_target target;
5982 ptl = pmd_trans_huge_lock(pmd, vma);
5984 if (mc.precharge < HPAGE_PMD_NR) {
5988 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5989 if (target_type == MC_TARGET_PAGE) {
5991 if (!isolate_lru_page(page)) {
5992 if (!mem_cgroup_move_account(page, true,
5994 mc.precharge -= HPAGE_PMD_NR;
5995 mc.moved_charge += HPAGE_PMD_NR;
5997 putback_lru_page(page);
6000 } else if (target_type == MC_TARGET_DEVICE) {
6002 if (!mem_cgroup_move_account(page, true,
6004 mc.precharge -= HPAGE_PMD_NR;
6005 mc.moved_charge += HPAGE_PMD_NR;
6013 if (pmd_trans_unstable(pmd))
6016 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6017 for (; addr != end; addr += PAGE_SIZE) {
6018 pte_t ptent = *(pte++);
6019 bool device = false;
6025 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6026 case MC_TARGET_DEVICE:
6029 case MC_TARGET_PAGE:
6032 * We can have a part of the split pmd here. Moving it
6033 * can be done but it would be too convoluted so simply
6034 * ignore such a partial THP and keep it in original
6035 * memcg. There should be somebody mapping the head.
6037 if (PageTransCompound(page))
6039 if (!device && isolate_lru_page(page))
6041 if (!mem_cgroup_move_account(page, false,
6044 /* we uncharge from mc.from later. */
6048 putback_lru_page(page);
6049 put: /* get_mctgt_type() gets the page */
6052 case MC_TARGET_SWAP:
6054 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6056 mem_cgroup_id_get_many(mc.to, 1);
6057 /* we fixup other refcnts and charges later. */
6065 pte_unmap_unlock(pte - 1, ptl);
6070 * We have consumed all precharges we got in can_attach().
6071 * We try charge one by one, but don't do any additional
6072 * charges to mc.to if we have failed in charge once in attach()
6075 ret = mem_cgroup_do_precharge(1);
6083 static const struct mm_walk_ops charge_walk_ops = {
6084 .pmd_entry = mem_cgroup_move_charge_pte_range,
6087 static void mem_cgroup_move_charge(void)
6089 lru_add_drain_all();
6091 * Signal lock_page_memcg() to take the memcg's move_lock
6092 * while we're moving its pages to another memcg. Then wait
6093 * for already started RCU-only updates to finish.
6095 atomic_inc(&mc.from->moving_account);
6098 if (unlikely(!mmap_read_trylock(mc.mm))) {
6100 * Someone who are holding the mmap_lock might be waiting in
6101 * waitq. So we cancel all extra charges, wake up all waiters,
6102 * and retry. Because we cancel precharges, we might not be able
6103 * to move enough charges, but moving charge is a best-effort
6104 * feature anyway, so it wouldn't be a big problem.
6106 __mem_cgroup_clear_mc();
6111 * When we have consumed all precharges and failed in doing
6112 * additional charge, the page walk just aborts.
6114 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6117 mmap_read_unlock(mc.mm);
6118 atomic_dec(&mc.from->moving_account);
6121 static void mem_cgroup_move_task(void)
6124 mem_cgroup_move_charge();
6125 mem_cgroup_clear_mc();
6128 #else /* !CONFIG_MMU */
6129 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6133 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6136 static void mem_cgroup_move_task(void)
6141 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6143 if (value == PAGE_COUNTER_MAX)
6144 seq_puts(m, "max\n");
6146 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6151 static u64 memory_current_read(struct cgroup_subsys_state *css,
6154 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6156 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6159 static int memory_min_show(struct seq_file *m, void *v)
6161 return seq_puts_memcg_tunable(m,
6162 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6165 static ssize_t memory_min_write(struct kernfs_open_file *of,
6166 char *buf, size_t nbytes, loff_t off)
6168 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6172 buf = strstrip(buf);
6173 err = page_counter_memparse(buf, "max", &min);
6177 page_counter_set_min(&memcg->memory, min);
6182 static int memory_low_show(struct seq_file *m, void *v)
6184 return seq_puts_memcg_tunable(m,
6185 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6188 static ssize_t memory_low_write(struct kernfs_open_file *of,
6189 char *buf, size_t nbytes, loff_t off)
6191 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6195 buf = strstrip(buf);
6196 err = page_counter_memparse(buf, "max", &low);
6200 page_counter_set_low(&memcg->memory, low);
6205 static int memory_high_show(struct seq_file *m, void *v)
6207 return seq_puts_memcg_tunable(m,
6208 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6211 static ssize_t memory_high_write(struct kernfs_open_file *of,
6212 char *buf, size_t nbytes, loff_t off)
6214 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6215 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6216 bool drained = false;
6220 buf = strstrip(buf);
6221 err = page_counter_memparse(buf, "max", &high);
6225 page_counter_set_high(&memcg->memory, high);
6228 unsigned long nr_pages = page_counter_read(&memcg->memory);
6229 unsigned long reclaimed;
6231 if (nr_pages <= high)
6234 if (signal_pending(current))
6238 drain_all_stock(memcg);
6243 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6246 if (!reclaimed && !nr_retries--)
6250 memcg_wb_domain_size_changed(memcg);
6254 static int memory_max_show(struct seq_file *m, void *v)
6256 return seq_puts_memcg_tunable(m,
6257 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6260 static ssize_t memory_max_write(struct kernfs_open_file *of,
6261 char *buf, size_t nbytes, loff_t off)
6263 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6264 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6265 bool drained = false;
6269 buf = strstrip(buf);
6270 err = page_counter_memparse(buf, "max", &max);
6274 xchg(&memcg->memory.max, max);
6277 unsigned long nr_pages = page_counter_read(&memcg->memory);
6279 if (nr_pages <= max)
6282 if (signal_pending(current))
6286 drain_all_stock(memcg);
6292 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6298 memcg_memory_event(memcg, MEMCG_OOM);
6299 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6303 memcg_wb_domain_size_changed(memcg);
6307 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6309 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6310 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6311 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6312 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6313 seq_printf(m, "oom_kill %lu\n",
6314 atomic_long_read(&events[MEMCG_OOM_KILL]));
6317 static int memory_events_show(struct seq_file *m, void *v)
6319 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6321 __memory_events_show(m, memcg->memory_events);
6325 static int memory_events_local_show(struct seq_file *m, void *v)
6327 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6329 __memory_events_show(m, memcg->memory_events_local);
6333 static int memory_stat_show(struct seq_file *m, void *v)
6335 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6338 buf = memory_stat_format(memcg);
6347 static inline unsigned long lruvec_page_state_output(struct lruvec *lruvec,
6350 return lruvec_page_state(lruvec, item) * memcg_page_state_unit(item);
6353 static int memory_numa_stat_show(struct seq_file *m, void *v)
6356 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6358 mem_cgroup_flush_stats();
6360 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6363 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6366 seq_printf(m, "%s", memory_stats[i].name);
6367 for_each_node_state(nid, N_MEMORY) {
6369 struct lruvec *lruvec;
6371 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6372 size = lruvec_page_state_output(lruvec,
6373 memory_stats[i].idx);
6374 seq_printf(m, " N%d=%llu", nid, size);
6383 static int memory_oom_group_show(struct seq_file *m, void *v)
6385 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6387 seq_printf(m, "%d\n", memcg->oom_group);
6392 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6393 char *buf, size_t nbytes, loff_t off)
6395 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6398 buf = strstrip(buf);
6402 ret = kstrtoint(buf, 0, &oom_group);
6406 if (oom_group != 0 && oom_group != 1)
6409 memcg->oom_group = oom_group;
6414 static struct cftype memory_files[] = {
6417 .flags = CFTYPE_NOT_ON_ROOT,
6418 .read_u64 = memory_current_read,
6422 .flags = CFTYPE_NOT_ON_ROOT,
6423 .seq_show = memory_min_show,
6424 .write = memory_min_write,
6428 .flags = CFTYPE_NOT_ON_ROOT,
6429 .seq_show = memory_low_show,
6430 .write = memory_low_write,
6434 .flags = CFTYPE_NOT_ON_ROOT,
6435 .seq_show = memory_high_show,
6436 .write = memory_high_write,
6440 .flags = CFTYPE_NOT_ON_ROOT,
6441 .seq_show = memory_max_show,
6442 .write = memory_max_write,
6446 .flags = CFTYPE_NOT_ON_ROOT,
6447 .file_offset = offsetof(struct mem_cgroup, events_file),
6448 .seq_show = memory_events_show,
6451 .name = "events.local",
6452 .flags = CFTYPE_NOT_ON_ROOT,
6453 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6454 .seq_show = memory_events_local_show,
6458 .seq_show = memory_stat_show,
6462 .name = "numa_stat",
6463 .seq_show = memory_numa_stat_show,
6467 .name = "oom.group",
6468 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6469 .seq_show = memory_oom_group_show,
6470 .write = memory_oom_group_write,
6475 struct cgroup_subsys memory_cgrp_subsys = {
6476 .css_alloc = mem_cgroup_css_alloc,
6477 .css_online = mem_cgroup_css_online,
6478 .css_offline = mem_cgroup_css_offline,
6479 .css_released = mem_cgroup_css_released,
6480 .css_free = mem_cgroup_css_free,
6481 .css_reset = mem_cgroup_css_reset,
6482 .css_rstat_flush = mem_cgroup_css_rstat_flush,
6483 .can_attach = mem_cgroup_can_attach,
6484 .cancel_attach = mem_cgroup_cancel_attach,
6485 .post_attach = mem_cgroup_move_task,
6486 .dfl_cftypes = memory_files,
6487 .legacy_cftypes = mem_cgroup_legacy_files,
6492 * This function calculates an individual cgroup's effective
6493 * protection which is derived from its own memory.min/low, its
6494 * parent's and siblings' settings, as well as the actual memory
6495 * distribution in the tree.
6497 * The following rules apply to the effective protection values:
6499 * 1. At the first level of reclaim, effective protection is equal to
6500 * the declared protection in memory.min and memory.low.
6502 * 2. To enable safe delegation of the protection configuration, at
6503 * subsequent levels the effective protection is capped to the
6504 * parent's effective protection.
6506 * 3. To make complex and dynamic subtrees easier to configure, the
6507 * user is allowed to overcommit the declared protection at a given
6508 * level. If that is the case, the parent's effective protection is
6509 * distributed to the children in proportion to how much protection
6510 * they have declared and how much of it they are utilizing.
6512 * This makes distribution proportional, but also work-conserving:
6513 * if one cgroup claims much more protection than it uses memory,
6514 * the unused remainder is available to its siblings.
6516 * 4. Conversely, when the declared protection is undercommitted at a
6517 * given level, the distribution of the larger parental protection
6518 * budget is NOT proportional. A cgroup's protection from a sibling
6519 * is capped to its own memory.min/low setting.
6521 * 5. However, to allow protecting recursive subtrees from each other
6522 * without having to declare each individual cgroup's fixed share
6523 * of the ancestor's claim to protection, any unutilized -
6524 * "floating" - protection from up the tree is distributed in
6525 * proportion to each cgroup's *usage*. This makes the protection
6526 * neutral wrt sibling cgroups and lets them compete freely over
6527 * the shared parental protection budget, but it protects the
6528 * subtree as a whole from neighboring subtrees.
6530 * Note that 4. and 5. are not in conflict: 4. is about protecting
6531 * against immediate siblings whereas 5. is about protecting against
6532 * neighboring subtrees.
6534 static unsigned long effective_protection(unsigned long usage,
6535 unsigned long parent_usage,
6536 unsigned long setting,
6537 unsigned long parent_effective,
6538 unsigned long siblings_protected)
6540 unsigned long protected;
6543 protected = min(usage, setting);
6545 * If all cgroups at this level combined claim and use more
6546 * protection then what the parent affords them, distribute
6547 * shares in proportion to utilization.
6549 * We are using actual utilization rather than the statically
6550 * claimed protection in order to be work-conserving: claimed
6551 * but unused protection is available to siblings that would
6552 * otherwise get a smaller chunk than what they claimed.
6554 if (siblings_protected > parent_effective)
6555 return protected * parent_effective / siblings_protected;
6558 * Ok, utilized protection of all children is within what the
6559 * parent affords them, so we know whatever this child claims
6560 * and utilizes is effectively protected.
6562 * If there is unprotected usage beyond this value, reclaim
6563 * will apply pressure in proportion to that amount.
6565 * If there is unutilized protection, the cgroup will be fully
6566 * shielded from reclaim, but we do return a smaller value for
6567 * protection than what the group could enjoy in theory. This
6568 * is okay. With the overcommit distribution above, effective
6569 * protection is always dependent on how memory is actually
6570 * consumed among the siblings anyway.
6575 * If the children aren't claiming (all of) the protection
6576 * afforded to them by the parent, distribute the remainder in
6577 * proportion to the (unprotected) memory of each cgroup. That
6578 * way, cgroups that aren't explicitly prioritized wrt each
6579 * other compete freely over the allowance, but they are
6580 * collectively protected from neighboring trees.
6582 * We're using unprotected memory for the weight so that if
6583 * some cgroups DO claim explicit protection, we don't protect
6584 * the same bytes twice.
6586 * Check both usage and parent_usage against the respective
6587 * protected values. One should imply the other, but they
6588 * aren't read atomically - make sure the division is sane.
6590 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6592 if (parent_effective > siblings_protected &&
6593 parent_usage > siblings_protected &&
6594 usage > protected) {
6595 unsigned long unclaimed;
6597 unclaimed = parent_effective - siblings_protected;
6598 unclaimed *= usage - protected;
6599 unclaimed /= parent_usage - siblings_protected;
6608 * mem_cgroup_calculate_protection - check if memory consumption is in the normal range
6609 * @root: the top ancestor of the sub-tree being checked
6610 * @memcg: the memory cgroup to check
6612 * WARNING: This function is not stateless! It can only be used as part
6613 * of a top-down tree iteration, not for isolated queries.
6615 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6616 struct mem_cgroup *memcg)
6618 unsigned long usage, parent_usage;
6619 struct mem_cgroup *parent;
6621 if (mem_cgroup_disabled())
6625 root = root_mem_cgroup;
6628 * Effective values of the reclaim targets are ignored so they
6629 * can be stale. Have a look at mem_cgroup_protection for more
6631 * TODO: calculation should be more robust so that we do not need
6632 * that special casing.
6637 usage = page_counter_read(&memcg->memory);
6641 parent = parent_mem_cgroup(memcg);
6642 /* No parent means a non-hierarchical mode on v1 memcg */
6646 if (parent == root) {
6647 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6648 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6652 parent_usage = page_counter_read(&parent->memory);
6654 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6655 READ_ONCE(memcg->memory.min),
6656 READ_ONCE(parent->memory.emin),
6657 atomic_long_read(&parent->memory.children_min_usage)));
6659 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6660 READ_ONCE(memcg->memory.low),
6661 READ_ONCE(parent->memory.elow),
6662 atomic_long_read(&parent->memory.children_low_usage)));
6665 static int charge_memcg(struct folio *folio, struct mem_cgroup *memcg,
6668 long nr_pages = folio_nr_pages(folio);
6671 ret = try_charge(memcg, gfp, nr_pages);
6675 css_get(&memcg->css);
6676 commit_charge(folio, memcg);
6678 local_irq_disable();
6679 mem_cgroup_charge_statistics(memcg, nr_pages);
6680 memcg_check_events(memcg, folio_nid(folio));
6686 int __mem_cgroup_charge(struct folio *folio, struct mm_struct *mm, gfp_t gfp)
6688 struct mem_cgroup *memcg;
6691 memcg = get_mem_cgroup_from_mm(mm);
6692 ret = charge_memcg(folio, memcg, gfp);
6693 css_put(&memcg->css);
6699 * mem_cgroup_swapin_charge_page - charge a newly allocated page for swapin
6700 * @page: page to charge
6701 * @mm: mm context of the victim
6702 * @gfp: reclaim mode
6703 * @entry: swap entry for which the page is allocated
6705 * This function charges a page allocated for swapin. Please call this before
6706 * adding the page to the swapcache.
6708 * Returns 0 on success. Otherwise, an error code is returned.
6710 int mem_cgroup_swapin_charge_page(struct page *page, struct mm_struct *mm,
6711 gfp_t gfp, swp_entry_t entry)
6713 struct folio *folio = page_folio(page);
6714 struct mem_cgroup *memcg;
6718 if (mem_cgroup_disabled())
6721 id = lookup_swap_cgroup_id(entry);
6723 memcg = mem_cgroup_from_id(id);
6724 if (!memcg || !css_tryget_online(&memcg->css))
6725 memcg = get_mem_cgroup_from_mm(mm);
6728 ret = charge_memcg(folio, memcg, gfp);
6730 css_put(&memcg->css);
6735 * mem_cgroup_swapin_uncharge_swap - uncharge swap slot
6736 * @entry: swap entry for which the page is charged
6738 * Call this function after successfully adding the charged page to swapcache.
6740 * Note: This function assumes the page for which swap slot is being uncharged
6743 void mem_cgroup_swapin_uncharge_swap(swp_entry_t entry)
6746 * Cgroup1's unified memory+swap counter has been charged with the
6747 * new swapcache page, finish the transfer by uncharging the swap
6748 * slot. The swap slot would also get uncharged when it dies, but
6749 * it can stick around indefinitely and we'd count the page twice
6752 * Cgroup2 has separate resource counters for memory and swap,
6753 * so this is a non-issue here. Memory and swap charge lifetimes
6754 * correspond 1:1 to page and swap slot lifetimes: we charge the
6755 * page to memory here, and uncharge swap when the slot is freed.
6757 if (!mem_cgroup_disabled() && do_memsw_account()) {
6759 * The swap entry might not get freed for a long time,
6760 * let's not wait for it. The page already received a
6761 * memory+swap charge, drop the swap entry duplicate.
6763 mem_cgroup_uncharge_swap(entry, 1);
6767 struct uncharge_gather {
6768 struct mem_cgroup *memcg;
6769 unsigned long nr_memory;
6770 unsigned long pgpgout;
6771 unsigned long nr_kmem;
6775 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6777 memset(ug, 0, sizeof(*ug));
6780 static void uncharge_batch(const struct uncharge_gather *ug)
6782 unsigned long flags;
6784 if (ug->nr_memory) {
6785 page_counter_uncharge(&ug->memcg->memory, ug->nr_memory);
6786 if (do_memsw_account())
6787 page_counter_uncharge(&ug->memcg->memsw, ug->nr_memory);
6788 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6789 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6790 memcg_oom_recover(ug->memcg);
6793 local_irq_save(flags);
6794 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6795 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_memory);
6796 memcg_check_events(ug->memcg, ug->nid);
6797 local_irq_restore(flags);
6799 /* drop reference from uncharge_folio */
6800 css_put(&ug->memcg->css);
6803 static void uncharge_folio(struct folio *folio, struct uncharge_gather *ug)
6806 struct mem_cgroup *memcg;
6807 struct obj_cgroup *objcg;
6808 bool use_objcg = folio_memcg_kmem(folio);
6810 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
6813 * Nobody should be changing or seriously looking at
6814 * folio memcg or objcg at this point, we have fully
6815 * exclusive access to the folio.
6818 objcg = __folio_objcg(folio);
6820 * This get matches the put at the end of the function and
6821 * kmem pages do not hold memcg references anymore.
6823 memcg = get_mem_cgroup_from_objcg(objcg);
6825 memcg = __folio_memcg(folio);
6831 if (ug->memcg != memcg) {
6834 uncharge_gather_clear(ug);
6837 ug->nid = folio_nid(folio);
6839 /* pairs with css_put in uncharge_batch */
6840 css_get(&memcg->css);
6843 nr_pages = folio_nr_pages(folio);
6846 ug->nr_memory += nr_pages;
6847 ug->nr_kmem += nr_pages;
6849 folio->memcg_data = 0;
6850 obj_cgroup_put(objcg);
6852 /* LRU pages aren't accounted at the root level */
6853 if (!mem_cgroup_is_root(memcg))
6854 ug->nr_memory += nr_pages;
6857 folio->memcg_data = 0;
6860 css_put(&memcg->css);
6863 void __mem_cgroup_uncharge(struct folio *folio)
6865 struct uncharge_gather ug;
6867 /* Don't touch folio->lru of any random page, pre-check: */
6868 if (!folio_memcg(folio))
6871 uncharge_gather_clear(&ug);
6872 uncharge_folio(folio, &ug);
6873 uncharge_batch(&ug);
6877 * __mem_cgroup_uncharge_list - uncharge a list of page
6878 * @page_list: list of pages to uncharge
6880 * Uncharge a list of pages previously charged with
6881 * __mem_cgroup_charge().
6883 void __mem_cgroup_uncharge_list(struct list_head *page_list)
6885 struct uncharge_gather ug;
6886 struct folio *folio;
6888 uncharge_gather_clear(&ug);
6889 list_for_each_entry(folio, page_list, lru)
6890 uncharge_folio(folio, &ug);
6892 uncharge_batch(&ug);
6896 * mem_cgroup_migrate - Charge a folio's replacement.
6897 * @old: Currently circulating folio.
6898 * @new: Replacement folio.
6900 * Charge @new as a replacement folio for @old. @old will
6901 * be uncharged upon free.
6903 * Both folios must be locked, @new->mapping must be set up.
6905 void mem_cgroup_migrate(struct folio *old, struct folio *new)
6907 struct mem_cgroup *memcg;
6908 long nr_pages = folio_nr_pages(new);
6909 unsigned long flags;
6911 VM_BUG_ON_FOLIO(!folio_test_locked(old), old);
6912 VM_BUG_ON_FOLIO(!folio_test_locked(new), new);
6913 VM_BUG_ON_FOLIO(folio_test_anon(old) != folio_test_anon(new), new);
6914 VM_BUG_ON_FOLIO(folio_nr_pages(old) != nr_pages, new);
6916 if (mem_cgroup_disabled())
6919 /* Page cache replacement: new folio already charged? */
6920 if (folio_memcg(new))
6923 memcg = folio_memcg(old);
6924 VM_WARN_ON_ONCE_FOLIO(!memcg, old);
6928 /* Force-charge the new page. The old one will be freed soon */
6929 if (!mem_cgroup_is_root(memcg)) {
6930 page_counter_charge(&memcg->memory, nr_pages);
6931 if (do_memsw_account())
6932 page_counter_charge(&memcg->memsw, nr_pages);
6935 css_get(&memcg->css);
6936 commit_charge(new, memcg);
6938 local_irq_save(flags);
6939 mem_cgroup_charge_statistics(memcg, nr_pages);
6940 memcg_check_events(memcg, folio_nid(new));
6941 local_irq_restore(flags);
6944 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6945 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6947 void mem_cgroup_sk_alloc(struct sock *sk)
6949 struct mem_cgroup *memcg;
6951 if (!mem_cgroup_sockets_enabled)
6954 /* Do not associate the sock with unrelated interrupted task's memcg. */
6959 memcg = mem_cgroup_from_task(current);
6960 if (memcg == root_mem_cgroup)
6962 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6964 if (css_tryget(&memcg->css))
6965 sk->sk_memcg = memcg;
6970 void mem_cgroup_sk_free(struct sock *sk)
6973 css_put(&sk->sk_memcg->css);
6977 * mem_cgroup_charge_skmem - charge socket memory
6978 * @memcg: memcg to charge
6979 * @nr_pages: number of pages to charge
6980 * @gfp_mask: reclaim mode
6982 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6983 * @memcg's configured limit, %false if it doesn't.
6985 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages,
6988 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6989 struct page_counter *fail;
6991 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6992 memcg->tcpmem_pressure = 0;
6995 memcg->tcpmem_pressure = 1;
6996 if (gfp_mask & __GFP_NOFAIL) {
6997 page_counter_charge(&memcg->tcpmem, nr_pages);
7003 if (try_charge(memcg, gfp_mask, nr_pages) == 0) {
7004 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7012 * mem_cgroup_uncharge_skmem - uncharge socket memory
7013 * @memcg: memcg to uncharge
7014 * @nr_pages: number of pages to uncharge
7016 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7018 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7019 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7023 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7025 refill_stock(memcg, nr_pages);
7028 static int __init cgroup_memory(char *s)
7032 while ((token = strsep(&s, ",")) != NULL) {
7035 if (!strcmp(token, "nosocket"))
7036 cgroup_memory_nosocket = true;
7037 if (!strcmp(token, "nokmem"))
7038 cgroup_memory_nokmem = true;
7042 __setup("cgroup.memory=", cgroup_memory);
7045 * subsys_initcall() for memory controller.
7047 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7048 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7049 * basically everything that doesn't depend on a specific mem_cgroup structure
7050 * should be initialized from here.
7052 static int __init mem_cgroup_init(void)
7057 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7058 * used for per-memcg-per-cpu caching of per-node statistics. In order
7059 * to work fine, we should make sure that the overfill threshold can't
7060 * exceed S32_MAX / PAGE_SIZE.
7062 BUILD_BUG_ON(MEMCG_CHARGE_BATCH > S32_MAX / PAGE_SIZE);
7064 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7065 memcg_hotplug_cpu_dead);
7067 for_each_possible_cpu(cpu)
7068 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7071 for_each_node(node) {
7072 struct mem_cgroup_tree_per_node *rtpn;
7074 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7075 node_online(node) ? node : NUMA_NO_NODE);
7077 rtpn->rb_root = RB_ROOT;
7078 rtpn->rb_rightmost = NULL;
7079 spin_lock_init(&rtpn->lock);
7080 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7085 subsys_initcall(mem_cgroup_init);
7087 #ifdef CONFIG_MEMCG_SWAP
7088 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7090 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7092 * The root cgroup cannot be destroyed, so it's refcount must
7095 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7099 memcg = parent_mem_cgroup(memcg);
7101 memcg = root_mem_cgroup;
7107 * mem_cgroup_swapout - transfer a memsw charge to swap
7108 * @page: page whose memsw charge to transfer
7109 * @entry: swap entry to move the charge to
7111 * Transfer the memsw charge of @page to @entry.
7113 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7115 struct mem_cgroup *memcg, *swap_memcg;
7116 unsigned int nr_entries;
7117 unsigned short oldid;
7119 VM_BUG_ON_PAGE(PageLRU(page), page);
7120 VM_BUG_ON_PAGE(page_count(page), page);
7122 if (mem_cgroup_disabled())
7125 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7128 memcg = page_memcg(page);
7130 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7135 * In case the memcg owning these pages has been offlined and doesn't
7136 * have an ID allocated to it anymore, charge the closest online
7137 * ancestor for the swap instead and transfer the memory+swap charge.
7139 swap_memcg = mem_cgroup_id_get_online(memcg);
7140 nr_entries = thp_nr_pages(page);
7141 /* Get references for the tail pages, too */
7143 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7144 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7146 VM_BUG_ON_PAGE(oldid, page);
7147 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7149 page->memcg_data = 0;
7151 if (!mem_cgroup_is_root(memcg))
7152 page_counter_uncharge(&memcg->memory, nr_entries);
7154 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7155 if (!mem_cgroup_is_root(swap_memcg))
7156 page_counter_charge(&swap_memcg->memsw, nr_entries);
7157 page_counter_uncharge(&memcg->memsw, nr_entries);
7161 * Interrupts should be disabled here because the caller holds the
7162 * i_pages lock which is taken with interrupts-off. It is
7163 * important here to have the interrupts disabled because it is the
7164 * only synchronisation we have for updating the per-CPU variables.
7166 VM_BUG_ON(!irqs_disabled());
7167 mem_cgroup_charge_statistics(memcg, -nr_entries);
7168 memcg_check_events(memcg, page_to_nid(page));
7170 css_put(&memcg->css);
7174 * __mem_cgroup_try_charge_swap - try charging swap space for a page
7175 * @page: page being added to swap
7176 * @entry: swap entry to charge
7178 * Try to charge @page's memcg for the swap space at @entry.
7180 * Returns 0 on success, -ENOMEM on failure.
7182 int __mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7184 unsigned int nr_pages = thp_nr_pages(page);
7185 struct page_counter *counter;
7186 struct mem_cgroup *memcg;
7187 unsigned short oldid;
7189 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7192 memcg = page_memcg(page);
7194 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7199 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7203 memcg = mem_cgroup_id_get_online(memcg);
7205 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7206 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7207 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7208 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7209 mem_cgroup_id_put(memcg);
7213 /* Get references for the tail pages, too */
7215 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7216 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7217 VM_BUG_ON_PAGE(oldid, page);
7218 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7224 * __mem_cgroup_uncharge_swap - uncharge swap space
7225 * @entry: swap entry to uncharge
7226 * @nr_pages: the amount of swap space to uncharge
7228 void __mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7230 struct mem_cgroup *memcg;
7233 id = swap_cgroup_record(entry, 0, nr_pages);
7235 memcg = mem_cgroup_from_id(id);
7237 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7238 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7239 page_counter_uncharge(&memcg->swap, nr_pages);
7241 page_counter_uncharge(&memcg->memsw, nr_pages);
7243 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7244 mem_cgroup_id_put_many(memcg, nr_pages);
7249 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7251 long nr_swap_pages = get_nr_swap_pages();
7253 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7254 return nr_swap_pages;
7255 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7256 nr_swap_pages = min_t(long, nr_swap_pages,
7257 READ_ONCE(memcg->swap.max) -
7258 page_counter_read(&memcg->swap));
7259 return nr_swap_pages;
7262 bool mem_cgroup_swap_full(struct page *page)
7264 struct mem_cgroup *memcg;
7266 VM_BUG_ON_PAGE(!PageLocked(page), page);
7270 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7273 memcg = page_memcg(page);
7277 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7278 unsigned long usage = page_counter_read(&memcg->swap);
7280 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7281 usage * 2 >= READ_ONCE(memcg->swap.max))
7288 static int __init setup_swap_account(char *s)
7290 if (!strcmp(s, "1"))
7291 cgroup_memory_noswap = false;
7292 else if (!strcmp(s, "0"))
7293 cgroup_memory_noswap = true;
7296 __setup("swapaccount=", setup_swap_account);
7298 static u64 swap_current_read(struct cgroup_subsys_state *css,
7301 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7303 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7306 static int swap_high_show(struct seq_file *m, void *v)
7308 return seq_puts_memcg_tunable(m,
7309 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7312 static ssize_t swap_high_write(struct kernfs_open_file *of,
7313 char *buf, size_t nbytes, loff_t off)
7315 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7319 buf = strstrip(buf);
7320 err = page_counter_memparse(buf, "max", &high);
7324 page_counter_set_high(&memcg->swap, high);
7329 static int swap_max_show(struct seq_file *m, void *v)
7331 return seq_puts_memcg_tunable(m,
7332 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7335 static ssize_t swap_max_write(struct kernfs_open_file *of,
7336 char *buf, size_t nbytes, loff_t off)
7338 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7342 buf = strstrip(buf);
7343 err = page_counter_memparse(buf, "max", &max);
7347 xchg(&memcg->swap.max, max);
7352 static int swap_events_show(struct seq_file *m, void *v)
7354 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7356 seq_printf(m, "high %lu\n",
7357 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7358 seq_printf(m, "max %lu\n",
7359 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7360 seq_printf(m, "fail %lu\n",
7361 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7366 static struct cftype swap_files[] = {
7368 .name = "swap.current",
7369 .flags = CFTYPE_NOT_ON_ROOT,
7370 .read_u64 = swap_current_read,
7373 .name = "swap.high",
7374 .flags = CFTYPE_NOT_ON_ROOT,
7375 .seq_show = swap_high_show,
7376 .write = swap_high_write,
7380 .flags = CFTYPE_NOT_ON_ROOT,
7381 .seq_show = swap_max_show,
7382 .write = swap_max_write,
7385 .name = "swap.events",
7386 .flags = CFTYPE_NOT_ON_ROOT,
7387 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7388 .seq_show = swap_events_show,
7393 static struct cftype memsw_files[] = {
7395 .name = "memsw.usage_in_bytes",
7396 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7397 .read_u64 = mem_cgroup_read_u64,
7400 .name = "memsw.max_usage_in_bytes",
7401 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7402 .write = mem_cgroup_reset,
7403 .read_u64 = mem_cgroup_read_u64,
7406 .name = "memsw.limit_in_bytes",
7407 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7408 .write = mem_cgroup_write,
7409 .read_u64 = mem_cgroup_read_u64,
7412 .name = "memsw.failcnt",
7413 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7414 .write = mem_cgroup_reset,
7415 .read_u64 = mem_cgroup_read_u64,
7417 { }, /* terminate */
7421 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7422 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7423 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7424 * boot parameter. This may result in premature OOPS inside
7425 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7427 static int __init mem_cgroup_swap_init(void)
7429 /* No memory control -> no swap control */
7430 if (mem_cgroup_disabled())
7431 cgroup_memory_noswap = true;
7433 if (cgroup_memory_noswap)
7436 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7437 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7441 core_initcall(mem_cgroup_swap_init);
7443 #endif /* CONFIG_MEMCG_SWAP */