1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* memcontrol.c - Memory Controller
4 * Copyright IBM Corporation, 2007
5 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
7 * Copyright 2007 OpenVZ SWsoft Inc
8 * Author: Pavel Emelianov <xemul@openvz.org>
11 * Copyright (C) 2009 Nokia Corporation
12 * Author: Kirill A. Shutemov
14 * Kernel Memory Controller
15 * Copyright (C) 2012 Parallels Inc. and Google Inc.
16 * Authors: Glauber Costa and Suleiman Souhlal
19 * Charge lifetime sanitation
20 * Lockless page tracking & accounting
21 * Unified hierarchy configuration model
22 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
24 * Per memcg lru locking
25 * Copyright (C) 2020 Alibaba, Inc, Alex Shi
28 #include <linux/page_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
31 #include <linux/pagewalk.h>
32 #include <linux/sched/mm.h>
33 #include <linux/shmem_fs.h>
34 #include <linux/hugetlb.h>
35 #include <linux/pagemap.h>
36 #include <linux/vm_event_item.h>
37 #include <linux/smp.h>
38 #include <linux/page-flags.h>
39 #include <linux/backing-dev.h>
40 #include <linux/bit_spinlock.h>
41 #include <linux/rcupdate.h>
42 #include <linux/limits.h>
43 #include <linux/export.h>
44 #include <linux/mutex.h>
45 #include <linux/rbtree.h>
46 #include <linux/slab.h>
47 #include <linux/swap.h>
48 #include <linux/swapops.h>
49 #include <linux/spinlock.h>
50 #include <linux/eventfd.h>
51 #include <linux/poll.h>
52 #include <linux/sort.h>
54 #include <linux/seq_file.h>
55 #include <linux/vmpressure.h>
56 #include <linux/mm_inline.h>
57 #include <linux/swap_cgroup.h>
58 #include <linux/cpu.h>
59 #include <linux/oom.h>
60 #include <linux/lockdep.h>
61 #include <linux/file.h>
62 #include <linux/tracehook.h>
63 #include <linux/psi.h>
64 #include <linux/seq_buf.h>
70 #include <linux/uaccess.h>
72 #include <trace/events/vmscan.h>
74 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
75 EXPORT_SYMBOL(memory_cgrp_subsys);
77 struct mem_cgroup *root_mem_cgroup __read_mostly;
79 /* Active memory cgroup to use from an interrupt context */
80 DEFINE_PER_CPU(struct mem_cgroup *, int_active_memcg);
81 EXPORT_PER_CPU_SYMBOL_GPL(int_active_memcg);
83 /* Socket memory accounting disabled? */
84 static bool cgroup_memory_nosocket __ro_after_init;
86 /* Kernel memory accounting disabled? */
87 static bool cgroup_memory_nokmem __ro_after_init;
89 /* Whether the swap controller is active */
90 #ifdef CONFIG_MEMCG_SWAP
91 bool cgroup_memory_noswap __ro_after_init;
93 #define cgroup_memory_noswap 1
96 #ifdef CONFIG_CGROUP_WRITEBACK
97 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
100 /* Whether legacy memory+swap accounting is active */
101 static bool do_memsw_account(void)
103 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_noswap;
106 #define THRESHOLDS_EVENTS_TARGET 128
107 #define SOFTLIMIT_EVENTS_TARGET 1024
110 * Cgroups above their limits are maintained in a RB-Tree, independent of
111 * their hierarchy representation
114 struct mem_cgroup_tree_per_node {
115 struct rb_root rb_root;
116 struct rb_node *rb_rightmost;
120 struct mem_cgroup_tree {
121 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
124 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
127 struct mem_cgroup_eventfd_list {
128 struct list_head list;
129 struct eventfd_ctx *eventfd;
133 * cgroup_event represents events which userspace want to receive.
135 struct mem_cgroup_event {
137 * memcg which the event belongs to.
139 struct mem_cgroup *memcg;
141 * eventfd to signal userspace about the event.
143 struct eventfd_ctx *eventfd;
145 * Each of these stored in a list by the cgroup.
147 struct list_head list;
149 * register_event() callback will be used to add new userspace
150 * waiter for changes related to this event. Use eventfd_signal()
151 * on eventfd to send notification to userspace.
153 int (*register_event)(struct mem_cgroup *memcg,
154 struct eventfd_ctx *eventfd, const char *args);
156 * unregister_event() callback will be called when userspace closes
157 * the eventfd or on cgroup removing. This callback must be set,
158 * if you want provide notification functionality.
160 void (*unregister_event)(struct mem_cgroup *memcg,
161 struct eventfd_ctx *eventfd);
163 * All fields below needed to unregister event when
164 * userspace closes eventfd.
167 wait_queue_head_t *wqh;
168 wait_queue_entry_t wait;
169 struct work_struct remove;
172 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
173 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
175 /* Stuffs for move charges at task migration. */
177 * Types of charges to be moved.
179 #define MOVE_ANON 0x1U
180 #define MOVE_FILE 0x2U
181 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
183 /* "mc" and its members are protected by cgroup_mutex */
184 static struct move_charge_struct {
185 spinlock_t lock; /* for from, to */
186 struct mm_struct *mm;
187 struct mem_cgroup *from;
188 struct mem_cgroup *to;
190 unsigned long precharge;
191 unsigned long moved_charge;
192 unsigned long moved_swap;
193 struct task_struct *moving_task; /* a task moving charges */
194 wait_queue_head_t waitq; /* a waitq for other context */
196 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
197 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
201 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
202 * limit reclaim to prevent infinite loops, if they ever occur.
204 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
205 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
207 /* for encoding cft->private value on file */
216 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
217 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
218 #define MEMFILE_ATTR(val) ((val) & 0xffff)
219 /* Used for OOM notifier */
220 #define OOM_CONTROL (0)
223 * Iteration constructs for visiting all cgroups (under a tree). If
224 * loops are exited prematurely (break), mem_cgroup_iter_break() must
225 * be used for reference counting.
227 #define for_each_mem_cgroup_tree(iter, root) \
228 for (iter = mem_cgroup_iter(root, NULL, NULL); \
230 iter = mem_cgroup_iter(root, iter, NULL))
232 #define for_each_mem_cgroup(iter) \
233 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
235 iter = mem_cgroup_iter(NULL, iter, NULL))
237 static inline bool task_is_dying(void)
239 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
240 (current->flags & PF_EXITING);
243 /* Some nice accessors for the vmpressure. */
244 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
247 memcg = root_mem_cgroup;
248 return &memcg->vmpressure;
251 struct mem_cgroup *vmpressure_to_memcg(struct vmpressure *vmpr)
253 return container_of(vmpr, struct mem_cgroup, vmpressure);
256 #ifdef CONFIG_MEMCG_KMEM
257 static DEFINE_SPINLOCK(objcg_lock);
259 bool mem_cgroup_kmem_disabled(void)
261 return cgroup_memory_nokmem;
264 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
265 unsigned int nr_pages);
267 static void obj_cgroup_release(struct percpu_ref *ref)
269 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
270 unsigned int nr_bytes;
271 unsigned int nr_pages;
275 * At this point all allocated objects are freed, and
276 * objcg->nr_charged_bytes can't have an arbitrary byte value.
277 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
279 * The following sequence can lead to it:
280 * 1) CPU0: objcg == stock->cached_objcg
281 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
282 * PAGE_SIZE bytes are charged
283 * 3) CPU1: a process from another memcg is allocating something,
284 * the stock if flushed,
285 * objcg->nr_charged_bytes = PAGE_SIZE - 92
286 * 5) CPU0: we do release this object,
287 * 92 bytes are added to stock->nr_bytes
288 * 6) CPU0: stock is flushed,
289 * 92 bytes are added to objcg->nr_charged_bytes
291 * In the result, nr_charged_bytes == PAGE_SIZE.
292 * This page will be uncharged in obj_cgroup_release().
294 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
295 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
296 nr_pages = nr_bytes >> PAGE_SHIFT;
299 obj_cgroup_uncharge_pages(objcg, nr_pages);
301 spin_lock_irqsave(&objcg_lock, flags);
302 list_del(&objcg->list);
303 spin_unlock_irqrestore(&objcg_lock, flags);
305 percpu_ref_exit(ref);
306 kfree_rcu(objcg, rcu);
309 static struct obj_cgroup *obj_cgroup_alloc(void)
311 struct obj_cgroup *objcg;
314 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
318 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
324 INIT_LIST_HEAD(&objcg->list);
328 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
329 struct mem_cgroup *parent)
331 struct obj_cgroup *objcg, *iter;
333 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
335 spin_lock_irq(&objcg_lock);
337 /* 1) Ready to reparent active objcg. */
338 list_add(&objcg->list, &memcg->objcg_list);
339 /* 2) Reparent active objcg and already reparented objcgs to parent. */
340 list_for_each_entry(iter, &memcg->objcg_list, list)
341 WRITE_ONCE(iter->memcg, parent);
342 /* 3) Move already reparented objcgs to the parent's list */
343 list_splice(&memcg->objcg_list, &parent->objcg_list);
345 spin_unlock_irq(&objcg_lock);
347 percpu_ref_kill(&objcg->refcnt);
351 * This will be used as a shrinker list's index.
352 * The main reason for not using cgroup id for this:
353 * this works better in sparse environments, where we have a lot of memcgs,
354 * but only a few kmem-limited.
356 static DEFINE_IDA(memcg_cache_ida);
359 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
360 * this constant directly from cgroup, but it is understandable that this is
361 * better kept as an internal representation in cgroup.c. In any case, the
362 * cgrp_id space is not getting any smaller, and we don't have to necessarily
363 * increase ours as well if it increases.
365 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
368 * A lot of the calls to the cache allocation functions are expected to be
369 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
370 * conditional to this static branch, we'll have to allow modules that does
371 * kmem_cache_alloc and the such to see this symbol as well
373 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
374 EXPORT_SYMBOL(memcg_kmem_enabled_key);
378 * mem_cgroup_css_from_page - css of the memcg associated with a page
379 * @page: page of interest
381 * If memcg is bound to the default hierarchy, css of the memcg associated
382 * with @page is returned. The returned css remains associated with @page
383 * until it is released.
385 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
388 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
390 struct mem_cgroup *memcg;
392 memcg = page_memcg(page);
394 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
395 memcg = root_mem_cgroup;
401 * page_cgroup_ino - return inode number of the memcg a page is charged to
404 * Look up the closest online ancestor of the memory cgroup @page is charged to
405 * and return its inode number or 0 if @page is not charged to any cgroup. It
406 * is safe to call this function without holding a reference to @page.
408 * Note, this function is inherently racy, because there is nothing to prevent
409 * the cgroup inode from getting torn down and potentially reallocated a moment
410 * after page_cgroup_ino() returns, so it only should be used by callers that
411 * do not care (such as procfs interfaces).
413 ino_t page_cgroup_ino(struct page *page)
415 struct mem_cgroup *memcg;
416 unsigned long ino = 0;
419 memcg = page_memcg_check(page);
421 while (memcg && !(memcg->css.flags & CSS_ONLINE))
422 memcg = parent_mem_cgroup(memcg);
424 ino = cgroup_ino(memcg->css.cgroup);
429 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
430 struct mem_cgroup_tree_per_node *mctz,
431 unsigned long new_usage_in_excess)
433 struct rb_node **p = &mctz->rb_root.rb_node;
434 struct rb_node *parent = NULL;
435 struct mem_cgroup_per_node *mz_node;
436 bool rightmost = true;
441 mz->usage_in_excess = new_usage_in_excess;
442 if (!mz->usage_in_excess)
446 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
448 if (mz->usage_in_excess < mz_node->usage_in_excess) {
457 mctz->rb_rightmost = &mz->tree_node;
459 rb_link_node(&mz->tree_node, parent, p);
460 rb_insert_color(&mz->tree_node, &mctz->rb_root);
464 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
465 struct mem_cgroup_tree_per_node *mctz)
470 if (&mz->tree_node == mctz->rb_rightmost)
471 mctz->rb_rightmost = rb_prev(&mz->tree_node);
473 rb_erase(&mz->tree_node, &mctz->rb_root);
477 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
478 struct mem_cgroup_tree_per_node *mctz)
482 spin_lock_irqsave(&mctz->lock, flags);
483 __mem_cgroup_remove_exceeded(mz, mctz);
484 spin_unlock_irqrestore(&mctz->lock, flags);
487 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
489 unsigned long nr_pages = page_counter_read(&memcg->memory);
490 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
491 unsigned long excess = 0;
493 if (nr_pages > soft_limit)
494 excess = nr_pages - soft_limit;
499 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, int nid)
501 unsigned long excess;
502 struct mem_cgroup_per_node *mz;
503 struct mem_cgroup_tree_per_node *mctz;
505 mctz = soft_limit_tree.rb_tree_per_node[nid];
509 * Necessary to update all ancestors when hierarchy is used.
510 * because their event counter is not touched.
512 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
513 mz = memcg->nodeinfo[nid];
514 excess = soft_limit_excess(memcg);
516 * We have to update the tree if mz is on RB-tree or
517 * mem is over its softlimit.
519 if (excess || mz->on_tree) {
522 spin_lock_irqsave(&mctz->lock, flags);
523 /* if on-tree, remove it */
525 __mem_cgroup_remove_exceeded(mz, mctz);
527 * Insert again. mz->usage_in_excess will be updated.
528 * If excess is 0, no tree ops.
530 __mem_cgroup_insert_exceeded(mz, mctz, excess);
531 spin_unlock_irqrestore(&mctz->lock, flags);
536 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
538 struct mem_cgroup_tree_per_node *mctz;
539 struct mem_cgroup_per_node *mz;
543 mz = memcg->nodeinfo[nid];
544 mctz = soft_limit_tree.rb_tree_per_node[nid];
546 mem_cgroup_remove_exceeded(mz, mctz);
550 static struct mem_cgroup_per_node *
551 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
553 struct mem_cgroup_per_node *mz;
557 if (!mctz->rb_rightmost)
558 goto done; /* Nothing to reclaim from */
560 mz = rb_entry(mctz->rb_rightmost,
561 struct mem_cgroup_per_node, tree_node);
563 * Remove the node now but someone else can add it back,
564 * we will to add it back at the end of reclaim to its correct
565 * position in the tree.
567 __mem_cgroup_remove_exceeded(mz, mctz);
568 if (!soft_limit_excess(mz->memcg) ||
569 !css_tryget(&mz->memcg->css))
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;
580 spin_lock_irq(&mctz->lock);
581 mz = __mem_cgroup_largest_soft_limit_node(mctz);
582 spin_unlock_irq(&mctz->lock);
587 * memcg and lruvec stats flushing
589 * Many codepaths leading to stats update or read are performance sensitive and
590 * adding stats flushing in such codepaths is not desirable. So, to optimize the
591 * flushing the kernel does:
593 * 1) Periodically and asynchronously flush the stats every 2 seconds to not let
594 * rstat update tree grow unbounded.
596 * 2) Flush the stats synchronously on reader side only when there are more than
597 * (MEMCG_CHARGE_BATCH * nr_cpus) update events. Though this optimization
598 * will let stats be out of sync by atmost (MEMCG_CHARGE_BATCH * nr_cpus) but
599 * only for 2 seconds due to (1).
601 static void flush_memcg_stats_dwork(struct work_struct *w);
602 static DECLARE_DEFERRABLE_WORK(stats_flush_dwork, flush_memcg_stats_dwork);
603 static DEFINE_SPINLOCK(stats_flush_lock);
604 static DEFINE_PER_CPU(unsigned int, stats_updates);
605 static atomic_t stats_flush_threshold = ATOMIC_INIT(0);
608 * Accessors to ensure that preemption is disabled on PREEMPT_RT because it can
609 * not rely on this as part of an acquired spinlock_t lock. These functions are
610 * never used in hardirq context on PREEMPT_RT and therefore disabling preemtion
613 static void memcg_stats_lock(void)
615 #ifdef CONFIG_PREEMPT_RT
618 VM_BUG_ON(!irqs_disabled());
622 static void __memcg_stats_lock(void)
624 #ifdef CONFIG_PREEMPT_RT
629 static void memcg_stats_unlock(void)
631 #ifdef CONFIG_PREEMPT_RT
636 static inline void memcg_rstat_updated(struct mem_cgroup *memcg, int val)
640 cgroup_rstat_updated(memcg->css.cgroup, smp_processor_id());
642 x = __this_cpu_add_return(stats_updates, abs(val));
643 if (x > MEMCG_CHARGE_BATCH) {
644 atomic_add(x / MEMCG_CHARGE_BATCH, &stats_flush_threshold);
645 __this_cpu_write(stats_updates, 0);
649 static void __mem_cgroup_flush_stats(void)
653 if (!spin_trylock_irqsave(&stats_flush_lock, flag))
656 cgroup_rstat_flush_irqsafe(root_mem_cgroup->css.cgroup);
657 atomic_set(&stats_flush_threshold, 0);
658 spin_unlock_irqrestore(&stats_flush_lock, flag);
661 void mem_cgroup_flush_stats(void)
663 if (atomic_read(&stats_flush_threshold) > num_online_cpus())
664 __mem_cgroup_flush_stats();
667 static void flush_memcg_stats_dwork(struct work_struct *w)
669 __mem_cgroup_flush_stats();
670 queue_delayed_work(system_unbound_wq, &stats_flush_dwork, 2UL*HZ);
674 * __mod_memcg_state - update cgroup memory statistics
675 * @memcg: the memory cgroup
676 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
677 * @val: delta to add to the counter, can be negative
679 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
681 if (mem_cgroup_disabled())
684 __this_cpu_add(memcg->vmstats_percpu->state[idx], val);
685 memcg_rstat_updated(memcg, val);
688 /* idx can be of type enum memcg_stat_item or node_stat_item. */
689 static unsigned long memcg_page_state_local(struct mem_cgroup *memcg, int idx)
694 for_each_possible_cpu(cpu)
695 x += per_cpu(memcg->vmstats_percpu->state[idx], cpu);
703 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
706 struct mem_cgroup_per_node *pn;
707 struct mem_cgroup *memcg;
709 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
713 * The caller from rmap relay on disabled preemption becase they never
714 * update their counter from in-interrupt context. For these two
715 * counters we check that the update is never performed from an
716 * interrupt context while other caller need to have disabled interrupt.
718 __memcg_stats_lock();
719 if (IS_ENABLED(CONFIG_DEBUG_VM) && !IS_ENABLED(CONFIG_PREEMPT_RT)) {
724 case NR_SHMEM_PMDMAPPED:
725 case NR_FILE_PMDMAPPED:
726 WARN_ON_ONCE(!in_task());
729 WARN_ON_ONCE(!irqs_disabled());
734 __this_cpu_add(memcg->vmstats_percpu->state[idx], val);
737 __this_cpu_add(pn->lruvec_stats_percpu->state[idx], val);
739 memcg_rstat_updated(memcg, val);
740 memcg_stats_unlock();
744 * __mod_lruvec_state - update lruvec memory statistics
745 * @lruvec: the lruvec
746 * @idx: the stat item
747 * @val: delta to add to the counter, can be negative
749 * The lruvec is the intersection of the NUMA node and a cgroup. This
750 * function updates the all three counters that are affected by a
751 * change of state at this level: per-node, per-cgroup, per-lruvec.
753 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
757 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
759 /* Update memcg and lruvec */
760 if (!mem_cgroup_disabled())
761 __mod_memcg_lruvec_state(lruvec, idx, val);
764 void __mod_lruvec_page_state(struct page *page, enum node_stat_item idx,
767 struct page *head = compound_head(page); /* rmap on tail pages */
768 struct mem_cgroup *memcg;
769 pg_data_t *pgdat = page_pgdat(page);
770 struct lruvec *lruvec;
773 memcg = page_memcg(head);
774 /* Untracked pages have no memcg, no lruvec. Update only the node */
777 __mod_node_page_state(pgdat, idx, val);
781 lruvec = mem_cgroup_lruvec(memcg, pgdat);
782 __mod_lruvec_state(lruvec, idx, val);
785 EXPORT_SYMBOL(__mod_lruvec_page_state);
787 void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val)
789 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
790 struct mem_cgroup *memcg;
791 struct lruvec *lruvec;
794 memcg = mem_cgroup_from_obj(p);
797 * Untracked pages have no memcg, no lruvec. Update only the
798 * node. If we reparent the slab objects to the root memcg,
799 * when we free the slab object, we need to update the per-memcg
800 * vmstats to keep it correct for the root memcg.
803 __mod_node_page_state(pgdat, idx, val);
805 lruvec = mem_cgroup_lruvec(memcg, pgdat);
806 __mod_lruvec_state(lruvec, idx, val);
812 * __count_memcg_events - account VM events in a cgroup
813 * @memcg: the memory cgroup
814 * @idx: the event item
815 * @count: the number of events that occurred
817 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
820 if (mem_cgroup_disabled())
824 __this_cpu_add(memcg->vmstats_percpu->events[idx], count);
825 memcg_rstat_updated(memcg, count);
826 memcg_stats_unlock();
829 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
831 return READ_ONCE(memcg->vmstats.events[event]);
834 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
839 for_each_possible_cpu(cpu)
840 x += per_cpu(memcg->vmstats_percpu->events[event], cpu);
844 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
847 /* pagein of a big page is an event. So, ignore page size */
849 __count_memcg_events(memcg, PGPGIN, 1);
851 __count_memcg_events(memcg, PGPGOUT, 1);
852 nr_pages = -nr_pages; /* for event */
855 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
858 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
859 enum mem_cgroup_events_target target)
861 unsigned long val, next;
863 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
864 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
865 /* from time_after() in jiffies.h */
866 if ((long)(next - val) < 0) {
868 case MEM_CGROUP_TARGET_THRESH:
869 next = val + THRESHOLDS_EVENTS_TARGET;
871 case MEM_CGROUP_TARGET_SOFTLIMIT:
872 next = val + SOFTLIMIT_EVENTS_TARGET;
877 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
884 * Check events in order.
887 static void memcg_check_events(struct mem_cgroup *memcg, int nid)
889 if (IS_ENABLED(CONFIG_PREEMPT_RT))
892 /* threshold event is triggered in finer grain than soft limit */
893 if (unlikely(mem_cgroup_event_ratelimit(memcg,
894 MEM_CGROUP_TARGET_THRESH))) {
897 do_softlimit = mem_cgroup_event_ratelimit(memcg,
898 MEM_CGROUP_TARGET_SOFTLIMIT);
899 mem_cgroup_threshold(memcg);
900 if (unlikely(do_softlimit))
901 mem_cgroup_update_tree(memcg, nid);
905 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
908 * mm_update_next_owner() may clear mm->owner to NULL
909 * if it races with swapoff, page migration, etc.
910 * So this can be called with p == NULL.
915 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
917 EXPORT_SYMBOL(mem_cgroup_from_task);
919 static __always_inline struct mem_cgroup *active_memcg(void)
922 return this_cpu_read(int_active_memcg);
924 return current->active_memcg;
928 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
929 * @mm: mm from which memcg should be extracted. It can be NULL.
931 * Obtain a reference on mm->memcg and returns it if successful. If mm
932 * is NULL, then the memcg is chosen as follows:
933 * 1) The active memcg, if set.
934 * 2) current->mm->memcg, if available
936 * If mem_cgroup is disabled, NULL is returned.
938 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
940 struct mem_cgroup *memcg;
942 if (mem_cgroup_disabled())
946 * Page cache insertions can happen without an
947 * actual mm context, e.g. during disk probing
948 * on boot, loopback IO, acct() writes etc.
950 * No need to css_get on root memcg as the reference
951 * counting is disabled on the root level in the
952 * cgroup core. See CSS_NO_REF.
955 memcg = active_memcg();
956 if (unlikely(memcg)) {
957 /* remote memcg must hold a ref */
958 css_get(&memcg->css);
963 return root_mem_cgroup;
968 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
969 if (unlikely(!memcg))
970 memcg = root_mem_cgroup;
971 } while (!css_tryget(&memcg->css));
975 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
977 static __always_inline bool memcg_kmem_bypass(void)
979 /* Allow remote memcg charging from any context. */
980 if (unlikely(active_memcg()))
983 /* Memcg to charge can't be determined. */
984 if (!in_task() || !current->mm || (current->flags & PF_KTHREAD))
991 * mem_cgroup_iter - iterate over memory cgroup hierarchy
992 * @root: hierarchy root
993 * @prev: previously returned memcg, NULL on first invocation
994 * @reclaim: cookie for shared reclaim walks, NULL for full walks
996 * Returns references to children of the hierarchy below @root, or
997 * @root itself, or %NULL after a full round-trip.
999 * Caller must pass the return value in @prev on subsequent
1000 * invocations for reference counting, or use mem_cgroup_iter_break()
1001 * to cancel a hierarchy walk before the round-trip is complete.
1003 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1004 * in the hierarchy among all concurrent reclaimers operating on the
1007 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1008 struct mem_cgroup *prev,
1009 struct mem_cgroup_reclaim_cookie *reclaim)
1011 struct mem_cgroup_reclaim_iter *iter;
1012 struct cgroup_subsys_state *css = NULL;
1013 struct mem_cgroup *memcg = NULL;
1014 struct mem_cgroup *pos = NULL;
1016 if (mem_cgroup_disabled())
1020 root = root_mem_cgroup;
1022 if (prev && !reclaim)
1028 struct mem_cgroup_per_node *mz;
1030 mz = root->nodeinfo[reclaim->pgdat->node_id];
1033 if (prev && reclaim->generation != iter->generation)
1037 pos = READ_ONCE(iter->position);
1038 if (!pos || css_tryget(&pos->css))
1041 * css reference reached zero, so iter->position will
1042 * be cleared by ->css_released. However, we should not
1043 * rely on this happening soon, because ->css_released
1044 * is called from a work queue, and by busy-waiting we
1045 * might block it. So we clear iter->position right
1048 (void)cmpxchg(&iter->position, pos, NULL);
1056 css = css_next_descendant_pre(css, &root->css);
1059 * Reclaimers share the hierarchy walk, and a
1060 * new one might jump in right at the end of
1061 * the hierarchy - make sure they see at least
1062 * one group and restart from the beginning.
1070 * Verify the css and acquire a reference. The root
1071 * is provided by the caller, so we know it's alive
1072 * and kicking, and don't take an extra reference.
1074 memcg = mem_cgroup_from_css(css);
1076 if (css == &root->css)
1079 if (css_tryget(css))
1087 * The position could have already been updated by a competing
1088 * thread, so check that the value hasn't changed since we read
1089 * it to avoid reclaiming from the same cgroup twice.
1091 (void)cmpxchg(&iter->position, pos, memcg);
1099 reclaim->generation = iter->generation;
1104 if (prev && prev != root)
1105 css_put(&prev->css);
1111 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1112 * @root: hierarchy root
1113 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1115 void mem_cgroup_iter_break(struct mem_cgroup *root,
1116 struct mem_cgroup *prev)
1119 root = root_mem_cgroup;
1120 if (prev && prev != root)
1121 css_put(&prev->css);
1124 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1125 struct mem_cgroup *dead_memcg)
1127 struct mem_cgroup_reclaim_iter *iter;
1128 struct mem_cgroup_per_node *mz;
1131 for_each_node(nid) {
1132 mz = from->nodeinfo[nid];
1134 cmpxchg(&iter->position, dead_memcg, NULL);
1138 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1140 struct mem_cgroup *memcg = dead_memcg;
1141 struct mem_cgroup *last;
1144 __invalidate_reclaim_iterators(memcg, dead_memcg);
1146 } while ((memcg = parent_mem_cgroup(memcg)));
1149 * When cgruop1 non-hierarchy mode is used,
1150 * parent_mem_cgroup() does not walk all the way up to the
1151 * cgroup root (root_mem_cgroup). So we have to handle
1152 * dead_memcg from cgroup root separately.
1154 if (last != root_mem_cgroup)
1155 __invalidate_reclaim_iterators(root_mem_cgroup,
1160 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1161 * @memcg: hierarchy root
1162 * @fn: function to call for each task
1163 * @arg: argument passed to @fn
1165 * This function iterates over tasks attached to @memcg or to any of its
1166 * descendants and calls @fn for each task. If @fn returns a non-zero
1167 * value, the function breaks the iteration loop and returns the value.
1168 * Otherwise, it will iterate over all tasks and return 0.
1170 * This function must not be called for the root memory cgroup.
1172 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1173 int (*fn)(struct task_struct *, void *), void *arg)
1175 struct mem_cgroup *iter;
1178 BUG_ON(memcg == root_mem_cgroup);
1180 for_each_mem_cgroup_tree(iter, memcg) {
1181 struct css_task_iter it;
1182 struct task_struct *task;
1184 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1185 while (!ret && (task = css_task_iter_next(&it)))
1186 ret = fn(task, arg);
1187 css_task_iter_end(&it);
1189 mem_cgroup_iter_break(memcg, iter);
1196 #ifdef CONFIG_DEBUG_VM
1197 void lruvec_memcg_debug(struct lruvec *lruvec, struct folio *folio)
1199 struct mem_cgroup *memcg;
1201 if (mem_cgroup_disabled())
1204 memcg = folio_memcg(folio);
1207 VM_BUG_ON_FOLIO(lruvec_memcg(lruvec) != root_mem_cgroup, folio);
1209 VM_BUG_ON_FOLIO(lruvec_memcg(lruvec) != memcg, folio);
1214 * folio_lruvec_lock - Lock the lruvec for a folio.
1215 * @folio: Pointer to the folio.
1217 * These functions are safe to use under any of the following conditions:
1219 * - folio_test_lru false
1220 * - folio_memcg_lock()
1221 * - folio frozen (refcount of 0)
1223 * Return: The lruvec this folio is on with its lock held.
1225 struct lruvec *folio_lruvec_lock(struct folio *folio)
1227 struct lruvec *lruvec = folio_lruvec(folio);
1229 spin_lock(&lruvec->lru_lock);
1230 lruvec_memcg_debug(lruvec, folio);
1236 * folio_lruvec_lock_irq - Lock the lruvec for a folio.
1237 * @folio: Pointer to the folio.
1239 * These functions are safe to use under any of the following conditions:
1241 * - folio_test_lru false
1242 * - folio_memcg_lock()
1243 * - folio frozen (refcount of 0)
1245 * Return: The lruvec this folio is on with its lock held and interrupts
1248 struct lruvec *folio_lruvec_lock_irq(struct folio *folio)
1250 struct lruvec *lruvec = folio_lruvec(folio);
1252 spin_lock_irq(&lruvec->lru_lock);
1253 lruvec_memcg_debug(lruvec, folio);
1259 * folio_lruvec_lock_irqsave - Lock the lruvec for a folio.
1260 * @folio: Pointer to the folio.
1261 * @flags: Pointer to irqsave flags.
1263 * These functions are safe to use under any of the following conditions:
1265 * - folio_test_lru false
1266 * - folio_memcg_lock()
1267 * - folio frozen (refcount of 0)
1269 * Return: The lruvec this folio is on with its lock held and interrupts
1272 struct lruvec *folio_lruvec_lock_irqsave(struct folio *folio,
1273 unsigned long *flags)
1275 struct lruvec *lruvec = folio_lruvec(folio);
1277 spin_lock_irqsave(&lruvec->lru_lock, *flags);
1278 lruvec_memcg_debug(lruvec, folio);
1284 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1285 * @lruvec: mem_cgroup per zone lru vector
1286 * @lru: index of lru list the page is sitting on
1287 * @zid: zone id of the accounted pages
1288 * @nr_pages: positive when adding or negative when removing
1290 * This function must be called under lru_lock, just before a page is added
1291 * to or just after a page is removed from an lru list (that ordering being
1292 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1294 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1295 int zid, int nr_pages)
1297 struct mem_cgroup_per_node *mz;
1298 unsigned long *lru_size;
1301 if (mem_cgroup_disabled())
1304 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1305 lru_size = &mz->lru_zone_size[zid][lru];
1308 *lru_size += nr_pages;
1311 if (WARN_ONCE(size < 0,
1312 "%s(%p, %d, %d): lru_size %ld\n",
1313 __func__, lruvec, lru, nr_pages, size)) {
1319 *lru_size += nr_pages;
1323 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1324 * @memcg: the memory cgroup
1326 * Returns the maximum amount of memory @mem can be charged with, in
1329 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1331 unsigned long margin = 0;
1332 unsigned long count;
1333 unsigned long limit;
1335 count = page_counter_read(&memcg->memory);
1336 limit = READ_ONCE(memcg->memory.max);
1338 margin = limit - count;
1340 if (do_memsw_account()) {
1341 count = page_counter_read(&memcg->memsw);
1342 limit = READ_ONCE(memcg->memsw.max);
1344 margin = min(margin, limit - count);
1353 * A routine for checking "mem" is under move_account() or not.
1355 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1356 * moving cgroups. This is for waiting at high-memory pressure
1359 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1361 struct mem_cgroup *from;
1362 struct mem_cgroup *to;
1365 * Unlike task_move routines, we access mc.to, mc.from not under
1366 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1368 spin_lock(&mc.lock);
1374 ret = mem_cgroup_is_descendant(from, memcg) ||
1375 mem_cgroup_is_descendant(to, memcg);
1377 spin_unlock(&mc.lock);
1381 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1383 if (mc.moving_task && current != mc.moving_task) {
1384 if (mem_cgroup_under_move(memcg)) {
1386 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1387 /* moving charge context might have finished. */
1390 finish_wait(&mc.waitq, &wait);
1397 struct memory_stat {
1402 static const struct memory_stat memory_stats[] = {
1403 { "anon", NR_ANON_MAPPED },
1404 { "file", NR_FILE_PAGES },
1405 { "kernel", MEMCG_KMEM },
1406 { "kernel_stack", NR_KERNEL_STACK_KB },
1407 { "pagetables", NR_PAGETABLE },
1408 { "percpu", MEMCG_PERCPU_B },
1409 { "sock", MEMCG_SOCK },
1410 { "vmalloc", MEMCG_VMALLOC },
1411 { "shmem", NR_SHMEM },
1412 { "file_mapped", NR_FILE_MAPPED },
1413 { "file_dirty", NR_FILE_DIRTY },
1414 { "file_writeback", NR_WRITEBACK },
1416 { "swapcached", NR_SWAPCACHE },
1418 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1419 { "anon_thp", NR_ANON_THPS },
1420 { "file_thp", NR_FILE_THPS },
1421 { "shmem_thp", NR_SHMEM_THPS },
1423 { "inactive_anon", NR_INACTIVE_ANON },
1424 { "active_anon", NR_ACTIVE_ANON },
1425 { "inactive_file", NR_INACTIVE_FILE },
1426 { "active_file", NR_ACTIVE_FILE },
1427 { "unevictable", NR_UNEVICTABLE },
1428 { "slab_reclaimable", NR_SLAB_RECLAIMABLE_B },
1429 { "slab_unreclaimable", NR_SLAB_UNRECLAIMABLE_B },
1431 /* The memory events */
1432 { "workingset_refault_anon", WORKINGSET_REFAULT_ANON },
1433 { "workingset_refault_file", WORKINGSET_REFAULT_FILE },
1434 { "workingset_activate_anon", WORKINGSET_ACTIVATE_ANON },
1435 { "workingset_activate_file", WORKINGSET_ACTIVATE_FILE },
1436 { "workingset_restore_anon", WORKINGSET_RESTORE_ANON },
1437 { "workingset_restore_file", WORKINGSET_RESTORE_FILE },
1438 { "workingset_nodereclaim", WORKINGSET_NODERECLAIM },
1441 /* Translate stat items to the correct unit for memory.stat output */
1442 static int memcg_page_state_unit(int item)
1445 case MEMCG_PERCPU_B:
1446 case NR_SLAB_RECLAIMABLE_B:
1447 case NR_SLAB_UNRECLAIMABLE_B:
1448 case WORKINGSET_REFAULT_ANON:
1449 case WORKINGSET_REFAULT_FILE:
1450 case WORKINGSET_ACTIVATE_ANON:
1451 case WORKINGSET_ACTIVATE_FILE:
1452 case WORKINGSET_RESTORE_ANON:
1453 case WORKINGSET_RESTORE_FILE:
1454 case WORKINGSET_NODERECLAIM:
1456 case NR_KERNEL_STACK_KB:
1463 static inline unsigned long memcg_page_state_output(struct mem_cgroup *memcg,
1466 return memcg_page_state(memcg, item) * memcg_page_state_unit(item);
1469 static char *memory_stat_format(struct mem_cgroup *memcg)
1474 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1479 * Provide statistics on the state of the memory subsystem as
1480 * well as cumulative event counters that show past behavior.
1482 * This list is ordered following a combination of these gradients:
1483 * 1) generic big picture -> specifics and details
1484 * 2) reflecting userspace activity -> reflecting kernel heuristics
1486 * Current memory state:
1488 mem_cgroup_flush_stats();
1490 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1493 size = memcg_page_state_output(memcg, memory_stats[i].idx);
1494 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1496 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1497 size += memcg_page_state_output(memcg,
1498 NR_SLAB_RECLAIMABLE_B);
1499 seq_buf_printf(&s, "slab %llu\n", size);
1503 /* Accumulated memory events */
1505 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1506 memcg_events(memcg, PGFAULT));
1507 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1508 memcg_events(memcg, PGMAJFAULT));
1509 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1510 memcg_events(memcg, PGREFILL));
1511 seq_buf_printf(&s, "pgscan %lu\n",
1512 memcg_events(memcg, PGSCAN_KSWAPD) +
1513 memcg_events(memcg, PGSCAN_DIRECT));
1514 seq_buf_printf(&s, "pgsteal %lu\n",
1515 memcg_events(memcg, PGSTEAL_KSWAPD) +
1516 memcg_events(memcg, PGSTEAL_DIRECT));
1517 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1518 memcg_events(memcg, PGACTIVATE));
1519 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1520 memcg_events(memcg, PGDEACTIVATE));
1521 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1522 memcg_events(memcg, PGLAZYFREE));
1523 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1524 memcg_events(memcg, PGLAZYFREED));
1526 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1527 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1528 memcg_events(memcg, THP_FAULT_ALLOC));
1529 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1530 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1531 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1533 /* The above should easily fit into one page */
1534 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1539 #define K(x) ((x) << (PAGE_SHIFT-10))
1541 * mem_cgroup_print_oom_context: Print OOM information relevant to
1542 * memory controller.
1543 * @memcg: The memory cgroup that went over limit
1544 * @p: Task that is going to be killed
1546 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1549 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1554 pr_cont(",oom_memcg=");
1555 pr_cont_cgroup_path(memcg->css.cgroup);
1557 pr_cont(",global_oom");
1559 pr_cont(",task_memcg=");
1560 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1566 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1567 * memory controller.
1568 * @memcg: The memory cgroup that went over limit
1570 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1574 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1575 K((u64)page_counter_read(&memcg->memory)),
1576 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1577 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1578 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1579 K((u64)page_counter_read(&memcg->swap)),
1580 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1582 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1583 K((u64)page_counter_read(&memcg->memsw)),
1584 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1585 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1586 K((u64)page_counter_read(&memcg->kmem)),
1587 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1590 pr_info("Memory cgroup stats for ");
1591 pr_cont_cgroup_path(memcg->css.cgroup);
1593 buf = memory_stat_format(memcg);
1601 * Return the memory (and swap, if configured) limit for a memcg.
1603 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1605 unsigned long max = READ_ONCE(memcg->memory.max);
1607 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1608 if (mem_cgroup_swappiness(memcg))
1609 max += min(READ_ONCE(memcg->swap.max),
1610 (unsigned long)total_swap_pages);
1612 if (mem_cgroup_swappiness(memcg)) {
1613 /* Calculate swap excess capacity from memsw limit */
1614 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1616 max += min(swap, (unsigned long)total_swap_pages);
1622 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1624 return page_counter_read(&memcg->memory);
1627 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1630 struct oom_control oc = {
1634 .gfp_mask = gfp_mask,
1639 if (mutex_lock_killable(&oom_lock))
1642 if (mem_cgroup_margin(memcg) >= (1 << order))
1646 * A few threads which were not waiting at mutex_lock_killable() can
1647 * fail to bail out. Therefore, check again after holding oom_lock.
1649 ret = task_is_dying() || out_of_memory(&oc);
1652 mutex_unlock(&oom_lock);
1656 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1659 unsigned long *total_scanned)
1661 struct mem_cgroup *victim = NULL;
1664 unsigned long excess;
1665 unsigned long nr_scanned;
1666 struct mem_cgroup_reclaim_cookie reclaim = {
1670 excess = soft_limit_excess(root_memcg);
1673 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1678 * If we have not been able to reclaim
1679 * anything, it might because there are
1680 * no reclaimable pages under this hierarchy
1685 * We want to do more targeted reclaim.
1686 * excess >> 2 is not to excessive so as to
1687 * reclaim too much, nor too less that we keep
1688 * coming back to reclaim from this cgroup
1690 if (total >= (excess >> 2) ||
1691 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1696 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1697 pgdat, &nr_scanned);
1698 *total_scanned += nr_scanned;
1699 if (!soft_limit_excess(root_memcg))
1702 mem_cgroup_iter_break(root_memcg, victim);
1706 #ifdef CONFIG_LOCKDEP
1707 static struct lockdep_map memcg_oom_lock_dep_map = {
1708 .name = "memcg_oom_lock",
1712 static DEFINE_SPINLOCK(memcg_oom_lock);
1715 * Check OOM-Killer is already running under our hierarchy.
1716 * If someone is running, return false.
1718 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1720 struct mem_cgroup *iter, *failed = NULL;
1722 spin_lock(&memcg_oom_lock);
1724 for_each_mem_cgroup_tree(iter, memcg) {
1725 if (iter->oom_lock) {
1727 * this subtree of our hierarchy is already locked
1728 * so we cannot give a lock.
1731 mem_cgroup_iter_break(memcg, iter);
1734 iter->oom_lock = true;
1739 * OK, we failed to lock the whole subtree so we have
1740 * to clean up what we set up to the failing subtree
1742 for_each_mem_cgroup_tree(iter, memcg) {
1743 if (iter == failed) {
1744 mem_cgroup_iter_break(memcg, iter);
1747 iter->oom_lock = false;
1750 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1752 spin_unlock(&memcg_oom_lock);
1757 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1759 struct mem_cgroup *iter;
1761 spin_lock(&memcg_oom_lock);
1762 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1763 for_each_mem_cgroup_tree(iter, memcg)
1764 iter->oom_lock = false;
1765 spin_unlock(&memcg_oom_lock);
1768 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1770 struct mem_cgroup *iter;
1772 spin_lock(&memcg_oom_lock);
1773 for_each_mem_cgroup_tree(iter, memcg)
1775 spin_unlock(&memcg_oom_lock);
1778 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1780 struct mem_cgroup *iter;
1783 * Be careful about under_oom underflows because a child memcg
1784 * could have been added after mem_cgroup_mark_under_oom.
1786 spin_lock(&memcg_oom_lock);
1787 for_each_mem_cgroup_tree(iter, memcg)
1788 if (iter->under_oom > 0)
1790 spin_unlock(&memcg_oom_lock);
1793 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1795 struct oom_wait_info {
1796 struct mem_cgroup *memcg;
1797 wait_queue_entry_t wait;
1800 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1801 unsigned mode, int sync, void *arg)
1803 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1804 struct mem_cgroup *oom_wait_memcg;
1805 struct oom_wait_info *oom_wait_info;
1807 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1808 oom_wait_memcg = oom_wait_info->memcg;
1810 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1811 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1813 return autoremove_wake_function(wait, mode, sync, arg);
1816 static void memcg_oom_recover(struct mem_cgroup *memcg)
1819 * For the following lockless ->under_oom test, the only required
1820 * guarantee is that it must see the state asserted by an OOM when
1821 * this function is called as a result of userland actions
1822 * triggered by the notification of the OOM. This is trivially
1823 * achieved by invoking mem_cgroup_mark_under_oom() before
1824 * triggering notification.
1826 if (memcg && memcg->under_oom)
1827 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1831 * Returns true if successfully killed one or more processes. Though in some
1832 * corner cases it can return true even without killing any process.
1834 static bool mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1838 if (order > PAGE_ALLOC_COSTLY_ORDER)
1841 memcg_memory_event(memcg, MEMCG_OOM);
1844 * We are in the middle of the charge context here, so we
1845 * don't want to block when potentially sitting on a callstack
1846 * that holds all kinds of filesystem and mm locks.
1848 * cgroup1 allows disabling the OOM killer and waiting for outside
1849 * handling until the charge can succeed; remember the context and put
1850 * the task to sleep at the end of the page fault when all locks are
1853 * On the other hand, in-kernel OOM killer allows for an async victim
1854 * memory reclaim (oom_reaper) and that means that we are not solely
1855 * relying on the oom victim to make a forward progress and we can
1856 * invoke the oom killer here.
1858 * Please note that mem_cgroup_out_of_memory might fail to find a
1859 * victim and then we have to bail out from the charge path.
1861 if (memcg->oom_kill_disable) {
1862 if (current->in_user_fault) {
1863 css_get(&memcg->css);
1864 current->memcg_in_oom = memcg;
1865 current->memcg_oom_gfp_mask = mask;
1866 current->memcg_oom_order = order;
1871 mem_cgroup_mark_under_oom(memcg);
1873 locked = mem_cgroup_oom_trylock(memcg);
1876 mem_cgroup_oom_notify(memcg);
1878 mem_cgroup_unmark_under_oom(memcg);
1879 ret = mem_cgroup_out_of_memory(memcg, mask, order);
1882 mem_cgroup_oom_unlock(memcg);
1888 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1889 * @handle: actually kill/wait or just clean up the OOM state
1891 * This has to be called at the end of a page fault if the memcg OOM
1892 * handler was enabled.
1894 * Memcg supports userspace OOM handling where failed allocations must
1895 * sleep on a waitqueue until the userspace task resolves the
1896 * situation. Sleeping directly in the charge context with all kinds
1897 * of locks held is not a good idea, instead we remember an OOM state
1898 * in the task and mem_cgroup_oom_synchronize() has to be called at
1899 * the end of the page fault to complete the OOM handling.
1901 * Returns %true if an ongoing memcg OOM situation was detected and
1902 * completed, %false otherwise.
1904 bool mem_cgroup_oom_synchronize(bool handle)
1906 struct mem_cgroup *memcg = current->memcg_in_oom;
1907 struct oom_wait_info owait;
1910 /* OOM is global, do not handle */
1917 owait.memcg = memcg;
1918 owait.wait.flags = 0;
1919 owait.wait.func = memcg_oom_wake_function;
1920 owait.wait.private = current;
1921 INIT_LIST_HEAD(&owait.wait.entry);
1923 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1924 mem_cgroup_mark_under_oom(memcg);
1926 locked = mem_cgroup_oom_trylock(memcg);
1929 mem_cgroup_oom_notify(memcg);
1931 if (locked && !memcg->oom_kill_disable) {
1932 mem_cgroup_unmark_under_oom(memcg);
1933 finish_wait(&memcg_oom_waitq, &owait.wait);
1934 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1935 current->memcg_oom_order);
1938 mem_cgroup_unmark_under_oom(memcg);
1939 finish_wait(&memcg_oom_waitq, &owait.wait);
1943 mem_cgroup_oom_unlock(memcg);
1945 * There is no guarantee that an OOM-lock contender
1946 * sees the wakeups triggered by the OOM kill
1947 * uncharges. Wake any sleepers explicitly.
1949 memcg_oom_recover(memcg);
1952 current->memcg_in_oom = NULL;
1953 css_put(&memcg->css);
1958 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1959 * @victim: task to be killed by the OOM killer
1960 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1962 * Returns a pointer to a memory cgroup, which has to be cleaned up
1963 * by killing all belonging OOM-killable tasks.
1965 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1967 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1968 struct mem_cgroup *oom_domain)
1970 struct mem_cgroup *oom_group = NULL;
1971 struct mem_cgroup *memcg;
1973 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
1977 oom_domain = root_mem_cgroup;
1981 memcg = mem_cgroup_from_task(victim);
1982 if (memcg == root_mem_cgroup)
1986 * If the victim task has been asynchronously moved to a different
1987 * memory cgroup, we might end up killing tasks outside oom_domain.
1988 * In this case it's better to ignore memory.group.oom.
1990 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
1994 * Traverse the memory cgroup hierarchy from the victim task's
1995 * cgroup up to the OOMing cgroup (or root) to find the
1996 * highest-level memory cgroup with oom.group set.
1998 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1999 if (memcg->oom_group)
2002 if (memcg == oom_domain)
2007 css_get(&oom_group->css);
2014 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2016 pr_info("Tasks in ");
2017 pr_cont_cgroup_path(memcg->css.cgroup);
2018 pr_cont(" are going to be killed due to memory.oom.group set\n");
2022 * folio_memcg_lock - Bind a folio to its memcg.
2023 * @folio: The folio.
2025 * This function prevents unlocked LRU folios from being moved to
2028 * It ensures lifetime of the bound memcg. The caller is responsible
2029 * for the lifetime of the folio.
2031 void folio_memcg_lock(struct folio *folio)
2033 struct mem_cgroup *memcg;
2034 unsigned long flags;
2037 * The RCU lock is held throughout the transaction. The fast
2038 * path can get away without acquiring the memcg->move_lock
2039 * because page moving starts with an RCU grace period.
2043 if (mem_cgroup_disabled())
2046 memcg = folio_memcg(folio);
2047 if (unlikely(!memcg))
2050 #ifdef CONFIG_PROVE_LOCKING
2051 local_irq_save(flags);
2052 might_lock(&memcg->move_lock);
2053 local_irq_restore(flags);
2056 if (atomic_read(&memcg->moving_account) <= 0)
2059 spin_lock_irqsave(&memcg->move_lock, flags);
2060 if (memcg != folio_memcg(folio)) {
2061 spin_unlock_irqrestore(&memcg->move_lock, flags);
2066 * When charge migration first begins, we can have multiple
2067 * critical sections holding the fast-path RCU lock and one
2068 * holding the slowpath move_lock. Track the task who has the
2069 * move_lock for unlock_page_memcg().
2071 memcg->move_lock_task = current;
2072 memcg->move_lock_flags = flags;
2075 void lock_page_memcg(struct page *page)
2077 folio_memcg_lock(page_folio(page));
2080 static void __folio_memcg_unlock(struct mem_cgroup *memcg)
2082 if (memcg && memcg->move_lock_task == current) {
2083 unsigned long flags = memcg->move_lock_flags;
2085 memcg->move_lock_task = NULL;
2086 memcg->move_lock_flags = 0;
2088 spin_unlock_irqrestore(&memcg->move_lock, flags);
2095 * folio_memcg_unlock - Release the binding between a folio and its memcg.
2096 * @folio: The folio.
2098 * This releases the binding created by folio_memcg_lock(). This does
2099 * not change the accounting of this folio to its memcg, but it does
2100 * permit others to change it.
2102 void folio_memcg_unlock(struct folio *folio)
2104 __folio_memcg_unlock(folio_memcg(folio));
2107 void unlock_page_memcg(struct page *page)
2109 folio_memcg_unlock(page_folio(page));
2112 struct memcg_stock_pcp {
2113 local_lock_t stock_lock;
2114 struct mem_cgroup *cached; /* this never be root cgroup */
2115 unsigned int nr_pages;
2117 #ifdef CONFIG_MEMCG_KMEM
2118 struct obj_cgroup *cached_objcg;
2119 struct pglist_data *cached_pgdat;
2120 unsigned int nr_bytes;
2121 int nr_slab_reclaimable_b;
2122 int nr_slab_unreclaimable_b;
2125 struct work_struct work;
2126 unsigned long flags;
2127 #define FLUSHING_CACHED_CHARGE 0
2129 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock) = {
2130 .stock_lock = INIT_LOCAL_LOCK(stock_lock),
2132 static DEFINE_MUTEX(percpu_charge_mutex);
2134 #ifdef CONFIG_MEMCG_KMEM
2135 static struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock);
2136 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2137 struct mem_cgroup *root_memcg);
2138 static void memcg_account_kmem(struct mem_cgroup *memcg, int nr_pages);
2141 static inline struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock)
2145 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2146 struct mem_cgroup *root_memcg)
2150 static void memcg_account_kmem(struct mem_cgroup *memcg, int nr_pages)
2156 * consume_stock: Try to consume stocked charge on this cpu.
2157 * @memcg: memcg to consume from.
2158 * @nr_pages: how many pages to charge.
2160 * The charges will only happen if @memcg matches the current cpu's memcg
2161 * stock, and at least @nr_pages are available in that stock. Failure to
2162 * service an allocation will refill the stock.
2164 * returns true if successful, false otherwise.
2166 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2168 struct memcg_stock_pcp *stock;
2169 unsigned long flags;
2172 if (nr_pages > MEMCG_CHARGE_BATCH)
2175 local_lock_irqsave(&memcg_stock.stock_lock, flags);
2177 stock = this_cpu_ptr(&memcg_stock);
2178 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2179 stock->nr_pages -= nr_pages;
2183 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2189 * Returns stocks cached in percpu and reset cached information.
2191 static void drain_stock(struct memcg_stock_pcp *stock)
2193 struct mem_cgroup *old = stock->cached;
2198 if (stock->nr_pages) {
2199 page_counter_uncharge(&old->memory, stock->nr_pages);
2200 if (do_memsw_account())
2201 page_counter_uncharge(&old->memsw, stock->nr_pages);
2202 stock->nr_pages = 0;
2206 stock->cached = NULL;
2209 static void drain_local_stock(struct work_struct *dummy)
2211 struct memcg_stock_pcp *stock;
2212 struct obj_cgroup *old = NULL;
2213 unsigned long flags;
2216 * The only protection from cpu hotplug (memcg_hotplug_cpu_dead) vs.
2217 * drain_stock races is that we always operate on local CPU stock
2218 * here with IRQ disabled
2220 local_lock_irqsave(&memcg_stock.stock_lock, flags);
2222 stock = this_cpu_ptr(&memcg_stock);
2223 old = drain_obj_stock(stock);
2225 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2227 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2229 obj_cgroup_put(old);
2233 * Cache charges(val) to local per_cpu area.
2234 * This will be consumed by consume_stock() function, later.
2236 static void __refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2238 struct memcg_stock_pcp *stock;
2240 stock = this_cpu_ptr(&memcg_stock);
2241 if (stock->cached != memcg) { /* reset if necessary */
2243 css_get(&memcg->css);
2244 stock->cached = memcg;
2246 stock->nr_pages += nr_pages;
2248 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2252 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2254 unsigned long flags;
2256 local_lock_irqsave(&memcg_stock.stock_lock, flags);
2257 __refill_stock(memcg, nr_pages);
2258 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2262 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2263 * of the hierarchy under it.
2265 static void drain_all_stock(struct mem_cgroup *root_memcg)
2269 /* If someone's already draining, avoid adding running more workers. */
2270 if (!mutex_trylock(&percpu_charge_mutex))
2273 * Notify other cpus that system-wide "drain" is running
2274 * We do not care about races with the cpu hotplug because cpu down
2275 * as well as workers from this path always operate on the local
2276 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2279 curcpu = smp_processor_id();
2280 for_each_online_cpu(cpu) {
2281 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2282 struct mem_cgroup *memcg;
2286 memcg = stock->cached;
2287 if (memcg && stock->nr_pages &&
2288 mem_cgroup_is_descendant(memcg, root_memcg))
2290 else if (obj_stock_flush_required(stock, root_memcg))
2295 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2297 drain_local_stock(&stock->work);
2299 schedule_work_on(cpu, &stock->work);
2303 mutex_unlock(&percpu_charge_mutex);
2306 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2308 struct memcg_stock_pcp *stock;
2310 stock = &per_cpu(memcg_stock, cpu);
2316 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2317 unsigned int nr_pages,
2320 unsigned long nr_reclaimed = 0;
2323 unsigned long pflags;
2325 if (page_counter_read(&memcg->memory) <=
2326 READ_ONCE(memcg->memory.high))
2329 memcg_memory_event(memcg, MEMCG_HIGH);
2331 psi_memstall_enter(&pflags);
2332 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2334 psi_memstall_leave(&pflags);
2335 } while ((memcg = parent_mem_cgroup(memcg)) &&
2336 !mem_cgroup_is_root(memcg));
2338 return nr_reclaimed;
2341 static void high_work_func(struct work_struct *work)
2343 struct mem_cgroup *memcg;
2345 memcg = container_of(work, struct mem_cgroup, high_work);
2346 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2350 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2351 * enough to still cause a significant slowdown in most cases, while still
2352 * allowing diagnostics and tracing to proceed without becoming stuck.
2354 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2357 * When calculating the delay, we use these either side of the exponentiation to
2358 * maintain precision and scale to a reasonable number of jiffies (see the table
2361 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2362 * overage ratio to a delay.
2363 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2364 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2365 * to produce a reasonable delay curve.
2367 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2368 * reasonable delay curve compared to precision-adjusted overage, not
2369 * penalising heavily at first, but still making sure that growth beyond the
2370 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2371 * example, with a high of 100 megabytes:
2373 * +-------+------------------------+
2374 * | usage | time to allocate in ms |
2375 * +-------+------------------------+
2397 * +-------+------------------------+
2399 #define MEMCG_DELAY_PRECISION_SHIFT 20
2400 #define MEMCG_DELAY_SCALING_SHIFT 14
2402 static u64 calculate_overage(unsigned long usage, unsigned long high)
2410 * Prevent division by 0 in overage calculation by acting as if
2411 * it was a threshold of 1 page
2413 high = max(high, 1UL);
2415 overage = usage - high;
2416 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2417 return div64_u64(overage, high);
2420 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2422 u64 overage, max_overage = 0;
2425 overage = calculate_overage(page_counter_read(&memcg->memory),
2426 READ_ONCE(memcg->memory.high));
2427 max_overage = max(overage, max_overage);
2428 } while ((memcg = parent_mem_cgroup(memcg)) &&
2429 !mem_cgroup_is_root(memcg));
2434 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2436 u64 overage, max_overage = 0;
2439 overage = calculate_overage(page_counter_read(&memcg->swap),
2440 READ_ONCE(memcg->swap.high));
2442 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2443 max_overage = max(overage, max_overage);
2444 } while ((memcg = parent_mem_cgroup(memcg)) &&
2445 !mem_cgroup_is_root(memcg));
2451 * Get the number of jiffies that we should penalise a mischievous cgroup which
2452 * is exceeding its memory.high by checking both it and its ancestors.
2454 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2455 unsigned int nr_pages,
2458 unsigned long penalty_jiffies;
2464 * We use overage compared to memory.high to calculate the number of
2465 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2466 * fairly lenient on small overages, and increasingly harsh when the
2467 * memcg in question makes it clear that it has no intention of stopping
2468 * its crazy behaviour, so we exponentially increase the delay based on
2471 penalty_jiffies = max_overage * max_overage * HZ;
2472 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2473 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2476 * Factor in the task's own contribution to the overage, such that four
2477 * N-sized allocations are throttled approximately the same as one
2478 * 4N-sized allocation.
2480 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2481 * larger the current charge patch is than that.
2483 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2487 * Scheduled by try_charge() to be executed from the userland return path
2488 * and reclaims memory over the high limit.
2490 void mem_cgroup_handle_over_high(void)
2492 unsigned long penalty_jiffies;
2493 unsigned long pflags;
2494 unsigned long nr_reclaimed;
2495 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2496 int nr_retries = MAX_RECLAIM_RETRIES;
2497 struct mem_cgroup *memcg;
2498 bool in_retry = false;
2500 if (likely(!nr_pages))
2503 memcg = get_mem_cgroup_from_mm(current->mm);
2504 current->memcg_nr_pages_over_high = 0;
2508 * The allocating task should reclaim at least the batch size, but for
2509 * subsequent retries we only want to do what's necessary to prevent oom
2510 * or breaching resource isolation.
2512 * This is distinct from memory.max or page allocator behaviour because
2513 * memory.high is currently batched, whereas memory.max and the page
2514 * allocator run every time an allocation is made.
2516 nr_reclaimed = reclaim_high(memcg,
2517 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2521 * memory.high is breached and reclaim is unable to keep up. Throttle
2522 * allocators proactively to slow down excessive growth.
2524 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2525 mem_find_max_overage(memcg));
2527 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2528 swap_find_max_overage(memcg));
2531 * Clamp the max delay per usermode return so as to still keep the
2532 * application moving forwards and also permit diagnostics, albeit
2535 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2538 * Don't sleep if the amount of jiffies this memcg owes us is so low
2539 * that it's not even worth doing, in an attempt to be nice to those who
2540 * go only a small amount over their memory.high value and maybe haven't
2541 * been aggressively reclaimed enough yet.
2543 if (penalty_jiffies <= HZ / 100)
2547 * If reclaim is making forward progress but we're still over
2548 * memory.high, we want to encourage that rather than doing allocator
2551 if (nr_reclaimed || nr_retries--) {
2557 * If we exit early, we're guaranteed to die (since
2558 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2559 * need to account for any ill-begotten jiffies to pay them off later.
2561 psi_memstall_enter(&pflags);
2562 schedule_timeout_killable(penalty_jiffies);
2563 psi_memstall_leave(&pflags);
2566 css_put(&memcg->css);
2569 static int try_charge_memcg(struct mem_cgroup *memcg, gfp_t gfp_mask,
2570 unsigned int nr_pages)
2572 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2573 int nr_retries = MAX_RECLAIM_RETRIES;
2574 struct mem_cgroup *mem_over_limit;
2575 struct page_counter *counter;
2576 unsigned long nr_reclaimed;
2577 bool passed_oom = false;
2578 bool may_swap = true;
2579 bool drained = false;
2580 unsigned long pflags;
2583 if (consume_stock(memcg, nr_pages))
2586 if (!do_memsw_account() ||
2587 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2588 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2590 if (do_memsw_account())
2591 page_counter_uncharge(&memcg->memsw, batch);
2592 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2594 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2598 if (batch > nr_pages) {
2604 * Prevent unbounded recursion when reclaim operations need to
2605 * allocate memory. This might exceed the limits temporarily,
2606 * but we prefer facilitating memory reclaim and getting back
2607 * under the limit over triggering OOM kills in these cases.
2609 if (unlikely(current->flags & PF_MEMALLOC))
2612 if (unlikely(task_in_memcg_oom(current)))
2615 if (!gfpflags_allow_blocking(gfp_mask))
2618 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2620 psi_memstall_enter(&pflags);
2621 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2622 gfp_mask, may_swap);
2623 psi_memstall_leave(&pflags);
2625 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2629 drain_all_stock(mem_over_limit);
2634 if (gfp_mask & __GFP_NORETRY)
2637 * Even though the limit is exceeded at this point, reclaim
2638 * may have been able to free some pages. Retry the charge
2639 * before killing the task.
2641 * Only for regular pages, though: huge pages are rather
2642 * unlikely to succeed so close to the limit, and we fall back
2643 * to regular pages anyway in case of failure.
2645 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2648 * At task move, charge accounts can be doubly counted. So, it's
2649 * better to wait until the end of task_move if something is going on.
2651 if (mem_cgroup_wait_acct_move(mem_over_limit))
2657 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2660 /* Avoid endless loop for tasks bypassed by the oom killer */
2661 if (passed_oom && task_is_dying())
2665 * keep retrying as long as the memcg oom killer is able to make
2666 * a forward progress or bypass the charge if the oom killer
2667 * couldn't make any progress.
2669 if (mem_cgroup_oom(mem_over_limit, gfp_mask,
2670 get_order(nr_pages * PAGE_SIZE))) {
2672 nr_retries = MAX_RECLAIM_RETRIES;
2677 * Memcg doesn't have a dedicated reserve for atomic
2678 * allocations. But like the global atomic pool, we need to
2679 * put the burden of reclaim on regular allocation requests
2680 * and let these go through as privileged allocations.
2682 if (!(gfp_mask & (__GFP_NOFAIL | __GFP_HIGH)))
2686 * The allocation either can't fail or will lead to more memory
2687 * being freed very soon. Allow memory usage go over the limit
2688 * temporarily by force charging it.
2690 page_counter_charge(&memcg->memory, nr_pages);
2691 if (do_memsw_account())
2692 page_counter_charge(&memcg->memsw, nr_pages);
2697 if (batch > nr_pages)
2698 refill_stock(memcg, batch - nr_pages);
2701 * If the hierarchy is above the normal consumption range, schedule
2702 * reclaim on returning to userland. We can perform reclaim here
2703 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2704 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2705 * not recorded as it most likely matches current's and won't
2706 * change in the meantime. As high limit is checked again before
2707 * reclaim, the cost of mismatch is negligible.
2710 bool mem_high, swap_high;
2712 mem_high = page_counter_read(&memcg->memory) >
2713 READ_ONCE(memcg->memory.high);
2714 swap_high = page_counter_read(&memcg->swap) >
2715 READ_ONCE(memcg->swap.high);
2717 /* Don't bother a random interrupted task */
2720 schedule_work(&memcg->high_work);
2726 if (mem_high || swap_high) {
2728 * The allocating tasks in this cgroup will need to do
2729 * reclaim or be throttled to prevent further growth
2730 * of the memory or swap footprints.
2732 * Target some best-effort fairness between the tasks,
2733 * and distribute reclaim work and delay penalties
2734 * based on how much each task is actually allocating.
2736 current->memcg_nr_pages_over_high += batch;
2737 set_notify_resume(current);
2740 } while ((memcg = parent_mem_cgroup(memcg)));
2742 if (current->memcg_nr_pages_over_high > MEMCG_CHARGE_BATCH &&
2743 !(current->flags & PF_MEMALLOC) &&
2744 gfpflags_allow_blocking(gfp_mask)) {
2745 mem_cgroup_handle_over_high();
2750 static inline int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2751 unsigned int nr_pages)
2753 if (mem_cgroup_is_root(memcg))
2756 return try_charge_memcg(memcg, gfp_mask, nr_pages);
2759 static inline void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2761 if (mem_cgroup_is_root(memcg))
2764 page_counter_uncharge(&memcg->memory, nr_pages);
2765 if (do_memsw_account())
2766 page_counter_uncharge(&memcg->memsw, nr_pages);
2769 static void commit_charge(struct folio *folio, struct mem_cgroup *memcg)
2771 VM_BUG_ON_FOLIO(folio_memcg(folio), folio);
2773 * Any of the following ensures page's memcg stability:
2777 * - lock_page_memcg()
2778 * - exclusive reference
2780 folio->memcg_data = (unsigned long)memcg;
2783 #ifdef CONFIG_MEMCG_KMEM
2785 * The allocated objcg pointers array is not accounted directly.
2786 * Moreover, it should not come from DMA buffer and is not readily
2787 * reclaimable. So those GFP bits should be masked off.
2789 #define OBJCGS_CLEAR_MASK (__GFP_DMA | __GFP_RECLAIMABLE | __GFP_ACCOUNT)
2792 * mod_objcg_mlstate() may be called with irq enabled, so
2793 * mod_memcg_lruvec_state() should be used.
2795 static inline void mod_objcg_mlstate(struct obj_cgroup *objcg,
2796 struct pglist_data *pgdat,
2797 enum node_stat_item idx, int nr)
2799 struct mem_cgroup *memcg;
2800 struct lruvec *lruvec;
2803 memcg = obj_cgroup_memcg(objcg);
2804 lruvec = mem_cgroup_lruvec(memcg, pgdat);
2805 mod_memcg_lruvec_state(lruvec, idx, nr);
2809 int memcg_alloc_slab_cgroups(struct slab *slab, struct kmem_cache *s,
2810 gfp_t gfp, bool new_slab)
2812 unsigned int objects = objs_per_slab(s, slab);
2813 unsigned long memcg_data;
2816 gfp &= ~OBJCGS_CLEAR_MASK;
2817 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2822 memcg_data = (unsigned long) vec | MEMCG_DATA_OBJCGS;
2825 * If the slab is brand new and nobody can yet access its
2826 * memcg_data, no synchronization is required and memcg_data can
2827 * be simply assigned.
2829 slab->memcg_data = memcg_data;
2830 } else if (cmpxchg(&slab->memcg_data, 0, memcg_data)) {
2832 * If the slab is already in use, somebody can allocate and
2833 * assign obj_cgroups in parallel. In this case the existing
2834 * objcg vector should be reused.
2840 kmemleak_not_leak(vec);
2845 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2847 * A passed kernel object can be a slab object or a generic kernel page, so
2848 * different mechanisms for getting the memory cgroup pointer should be used.
2849 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2850 * can not know for sure how the kernel object is implemented.
2851 * mem_cgroup_from_obj() can be safely used in such cases.
2853 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2854 * cgroup_mutex, etc.
2856 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2858 struct folio *folio;
2860 if (mem_cgroup_disabled())
2863 folio = virt_to_folio(p);
2866 * Slab objects are accounted individually, not per-page.
2867 * Memcg membership data for each individual object is saved in
2870 if (folio_test_slab(folio)) {
2871 struct obj_cgroup **objcgs;
2875 slab = folio_slab(folio);
2876 objcgs = slab_objcgs(slab);
2880 off = obj_to_index(slab->slab_cache, slab, p);
2882 return obj_cgroup_memcg(objcgs[off]);
2888 * page_memcg_check() is used here, because in theory we can encounter
2889 * a folio where the slab flag has been cleared already, but
2890 * slab->memcg_data has not been freed yet
2891 * page_memcg_check(page) will guarantee that a proper memory
2892 * cgroup pointer or NULL will be returned.
2894 return page_memcg_check(folio_page(folio, 0));
2897 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2899 struct obj_cgroup *objcg = NULL;
2900 struct mem_cgroup *memcg;
2902 if (memcg_kmem_bypass())
2906 if (unlikely(active_memcg()))
2907 memcg = active_memcg();
2909 memcg = mem_cgroup_from_task(current);
2911 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2912 objcg = rcu_dereference(memcg->objcg);
2913 if (objcg && obj_cgroup_tryget(objcg))
2922 static void memcg_account_kmem(struct mem_cgroup *memcg, int nr_pages)
2924 mod_memcg_state(memcg, MEMCG_KMEM, nr_pages);
2925 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
2927 page_counter_charge(&memcg->kmem, nr_pages);
2929 page_counter_uncharge(&memcg->kmem, -nr_pages);
2935 * obj_cgroup_uncharge_pages: uncharge a number of kernel pages from a objcg
2936 * @objcg: object cgroup to uncharge
2937 * @nr_pages: number of pages to uncharge
2939 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
2940 unsigned int nr_pages)
2942 struct mem_cgroup *memcg;
2944 memcg = get_mem_cgroup_from_objcg(objcg);
2946 memcg_account_kmem(memcg, -nr_pages);
2947 refill_stock(memcg, nr_pages);
2949 css_put(&memcg->css);
2953 * obj_cgroup_charge_pages: charge a number of kernel pages to a objcg
2954 * @objcg: object cgroup to charge
2955 * @gfp: reclaim mode
2956 * @nr_pages: number of pages to charge
2958 * Returns 0 on success, an error code on failure.
2960 static int obj_cgroup_charge_pages(struct obj_cgroup *objcg, gfp_t gfp,
2961 unsigned int nr_pages)
2963 struct mem_cgroup *memcg;
2966 memcg = get_mem_cgroup_from_objcg(objcg);
2968 ret = try_charge_memcg(memcg, gfp, nr_pages);
2972 memcg_account_kmem(memcg, nr_pages);
2974 css_put(&memcg->css);
2980 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
2981 * @page: page to charge
2982 * @gfp: reclaim mode
2983 * @order: allocation order
2985 * Returns 0 on success, an error code on failure.
2987 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
2989 struct obj_cgroup *objcg;
2992 objcg = get_obj_cgroup_from_current();
2994 ret = obj_cgroup_charge_pages(objcg, gfp, 1 << order);
2996 page->memcg_data = (unsigned long)objcg |
3000 obj_cgroup_put(objcg);
3006 * __memcg_kmem_uncharge_page: uncharge a kmem page
3007 * @page: page to uncharge
3008 * @order: allocation order
3010 void __memcg_kmem_uncharge_page(struct page *page, int order)
3012 struct folio *folio = page_folio(page);
3013 struct obj_cgroup *objcg;
3014 unsigned int nr_pages = 1 << order;
3016 if (!folio_memcg_kmem(folio))
3019 objcg = __folio_objcg(folio);
3020 obj_cgroup_uncharge_pages(objcg, nr_pages);
3021 folio->memcg_data = 0;
3022 obj_cgroup_put(objcg);
3025 void mod_objcg_state(struct obj_cgroup *objcg, struct pglist_data *pgdat,
3026 enum node_stat_item idx, int nr)
3028 struct memcg_stock_pcp *stock;
3029 struct obj_cgroup *old = NULL;
3030 unsigned long flags;
3033 local_lock_irqsave(&memcg_stock.stock_lock, flags);
3034 stock = this_cpu_ptr(&memcg_stock);
3037 * Save vmstat data in stock and skip vmstat array update unless
3038 * accumulating over a page of vmstat data or when pgdat or idx
3041 if (stock->cached_objcg != objcg) {
3042 old = drain_obj_stock(stock);
3043 obj_cgroup_get(objcg);
3044 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3045 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3046 stock->cached_objcg = objcg;
3047 stock->cached_pgdat = pgdat;
3048 } else if (stock->cached_pgdat != pgdat) {
3049 /* Flush the existing cached vmstat data */
3050 struct pglist_data *oldpg = stock->cached_pgdat;
3052 if (stock->nr_slab_reclaimable_b) {
3053 mod_objcg_mlstate(objcg, oldpg, NR_SLAB_RECLAIMABLE_B,
3054 stock->nr_slab_reclaimable_b);
3055 stock->nr_slab_reclaimable_b = 0;
3057 if (stock->nr_slab_unreclaimable_b) {
3058 mod_objcg_mlstate(objcg, oldpg, NR_SLAB_UNRECLAIMABLE_B,
3059 stock->nr_slab_unreclaimable_b);
3060 stock->nr_slab_unreclaimable_b = 0;
3062 stock->cached_pgdat = pgdat;
3065 bytes = (idx == NR_SLAB_RECLAIMABLE_B) ? &stock->nr_slab_reclaimable_b
3066 : &stock->nr_slab_unreclaimable_b;
3068 * Even for large object >= PAGE_SIZE, the vmstat data will still be
3069 * cached locally at least once before pushing it out.
3076 if (abs(*bytes) > PAGE_SIZE) {
3084 mod_objcg_mlstate(objcg, pgdat, idx, nr);
3086 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3088 obj_cgroup_put(old);
3091 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3093 struct memcg_stock_pcp *stock;
3094 unsigned long flags;
3097 local_lock_irqsave(&memcg_stock.stock_lock, flags);
3099 stock = this_cpu_ptr(&memcg_stock);
3100 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3101 stock->nr_bytes -= nr_bytes;
3105 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3110 static struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock)
3112 struct obj_cgroup *old = stock->cached_objcg;
3117 if (stock->nr_bytes) {
3118 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3119 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3122 struct mem_cgroup *memcg;
3124 memcg = get_mem_cgroup_from_objcg(old);
3126 memcg_account_kmem(memcg, -nr_pages);
3127 __refill_stock(memcg, nr_pages);
3129 css_put(&memcg->css);
3133 * The leftover is flushed to the centralized per-memcg value.
3134 * On the next attempt to refill obj stock it will be moved
3135 * to a per-cpu stock (probably, on an other CPU), see
3136 * refill_obj_stock().
3138 * How often it's flushed is a trade-off between the memory
3139 * limit enforcement accuracy and potential CPU contention,
3140 * so it might be changed in the future.
3142 atomic_add(nr_bytes, &old->nr_charged_bytes);
3143 stock->nr_bytes = 0;
3147 * Flush the vmstat data in current stock
3149 if (stock->nr_slab_reclaimable_b || stock->nr_slab_unreclaimable_b) {
3150 if (stock->nr_slab_reclaimable_b) {
3151 mod_objcg_mlstate(old, stock->cached_pgdat,
3152 NR_SLAB_RECLAIMABLE_B,
3153 stock->nr_slab_reclaimable_b);
3154 stock->nr_slab_reclaimable_b = 0;
3156 if (stock->nr_slab_unreclaimable_b) {
3157 mod_objcg_mlstate(old, stock->cached_pgdat,
3158 NR_SLAB_UNRECLAIMABLE_B,
3159 stock->nr_slab_unreclaimable_b);
3160 stock->nr_slab_unreclaimable_b = 0;
3162 stock->cached_pgdat = NULL;
3165 stock->cached_objcg = NULL;
3167 * The `old' objects needs to be released by the caller via
3168 * obj_cgroup_put() outside of memcg_stock_pcp::stock_lock.
3173 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3174 struct mem_cgroup *root_memcg)
3176 struct mem_cgroup *memcg;
3178 if (stock->cached_objcg) {
3179 memcg = obj_cgroup_memcg(stock->cached_objcg);
3180 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3187 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes,
3188 bool allow_uncharge)
3190 struct memcg_stock_pcp *stock;
3191 struct obj_cgroup *old = NULL;
3192 unsigned long flags;
3193 unsigned int nr_pages = 0;
3195 local_lock_irqsave(&memcg_stock.stock_lock, flags);
3197 stock = this_cpu_ptr(&memcg_stock);
3198 if (stock->cached_objcg != objcg) { /* reset if necessary */
3199 old = drain_obj_stock(stock);
3200 obj_cgroup_get(objcg);
3201 stock->cached_objcg = objcg;
3202 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3203 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3204 allow_uncharge = true; /* Allow uncharge when objcg changes */
3206 stock->nr_bytes += nr_bytes;
3208 if (allow_uncharge && (stock->nr_bytes > PAGE_SIZE)) {
3209 nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3210 stock->nr_bytes &= (PAGE_SIZE - 1);
3213 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3215 obj_cgroup_put(old);
3218 obj_cgroup_uncharge_pages(objcg, nr_pages);
3221 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3223 unsigned int nr_pages, nr_bytes;
3226 if (consume_obj_stock(objcg, size))
3230 * In theory, objcg->nr_charged_bytes can have enough
3231 * pre-charged bytes to satisfy the allocation. However,
3232 * flushing objcg->nr_charged_bytes requires two atomic
3233 * operations, and objcg->nr_charged_bytes can't be big.
3234 * The shared objcg->nr_charged_bytes can also become a
3235 * performance bottleneck if all tasks of the same memcg are
3236 * trying to update it. So it's better to ignore it and try
3237 * grab some new pages. The stock's nr_bytes will be flushed to
3238 * objcg->nr_charged_bytes later on when objcg changes.
3240 * The stock's nr_bytes may contain enough pre-charged bytes
3241 * to allow one less page from being charged, but we can't rely
3242 * on the pre-charged bytes not being changed outside of
3243 * consume_obj_stock() or refill_obj_stock(). So ignore those
3244 * pre-charged bytes as well when charging pages. To avoid a
3245 * page uncharge right after a page charge, we set the
3246 * allow_uncharge flag to false when calling refill_obj_stock()
3247 * to temporarily allow the pre-charged bytes to exceed the page
3248 * size limit. The maximum reachable value of the pre-charged
3249 * bytes is (sizeof(object) + PAGE_SIZE - 2) if there is no data
3252 nr_pages = size >> PAGE_SHIFT;
3253 nr_bytes = size & (PAGE_SIZE - 1);
3258 ret = obj_cgroup_charge_pages(objcg, gfp, nr_pages);
3259 if (!ret && nr_bytes)
3260 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes, false);
3265 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3267 refill_obj_stock(objcg, size, true);
3270 #endif /* CONFIG_MEMCG_KMEM */
3273 * Because page_memcg(head) is not set on tails, set it now.
3275 void split_page_memcg(struct page *head, unsigned int nr)
3277 struct folio *folio = page_folio(head);
3278 struct mem_cgroup *memcg = folio_memcg(folio);
3281 if (mem_cgroup_disabled() || !memcg)
3284 for (i = 1; i < nr; i++)
3285 folio_page(folio, i)->memcg_data = folio->memcg_data;
3287 if (folio_memcg_kmem(folio))
3288 obj_cgroup_get_many(__folio_objcg(folio), nr - 1);
3290 css_get_many(&memcg->css, nr - 1);
3293 #ifdef CONFIG_MEMCG_SWAP
3295 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3296 * @entry: swap entry to be moved
3297 * @from: mem_cgroup which the entry is moved from
3298 * @to: mem_cgroup which the entry is moved to
3300 * It succeeds only when the swap_cgroup's record for this entry is the same
3301 * as the mem_cgroup's id of @from.
3303 * Returns 0 on success, -EINVAL on failure.
3305 * The caller must have charged to @to, IOW, called page_counter_charge() about
3306 * both res and memsw, and called css_get().
3308 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3309 struct mem_cgroup *from, struct mem_cgroup *to)
3311 unsigned short old_id, new_id;
3313 old_id = mem_cgroup_id(from);
3314 new_id = mem_cgroup_id(to);
3316 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3317 mod_memcg_state(from, MEMCG_SWAP, -1);
3318 mod_memcg_state(to, MEMCG_SWAP, 1);
3324 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3325 struct mem_cgroup *from, struct mem_cgroup *to)
3331 static DEFINE_MUTEX(memcg_max_mutex);
3333 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3334 unsigned long max, bool memsw)
3336 bool enlarge = false;
3337 bool drained = false;
3339 bool limits_invariant;
3340 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3343 if (signal_pending(current)) {
3348 mutex_lock(&memcg_max_mutex);
3350 * Make sure that the new limit (memsw or memory limit) doesn't
3351 * break our basic invariant rule memory.max <= memsw.max.
3353 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3354 max <= memcg->memsw.max;
3355 if (!limits_invariant) {
3356 mutex_unlock(&memcg_max_mutex);
3360 if (max > counter->max)
3362 ret = page_counter_set_max(counter, max);
3363 mutex_unlock(&memcg_max_mutex);
3369 drain_all_stock(memcg);
3374 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3375 GFP_KERNEL, !memsw)) {
3381 if (!ret && enlarge)
3382 memcg_oom_recover(memcg);
3387 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3389 unsigned long *total_scanned)
3391 unsigned long nr_reclaimed = 0;
3392 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3393 unsigned long reclaimed;
3395 struct mem_cgroup_tree_per_node *mctz;
3396 unsigned long excess;
3397 unsigned long nr_scanned;
3402 mctz = soft_limit_tree.rb_tree_per_node[pgdat->node_id];
3405 * Do not even bother to check the largest node if the root
3406 * is empty. Do it lockless to prevent lock bouncing. Races
3407 * are acceptable as soft limit is best effort anyway.
3409 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3413 * This loop can run a while, specially if mem_cgroup's continuously
3414 * keep exceeding their soft limit and putting the system under
3421 mz = mem_cgroup_largest_soft_limit_node(mctz);
3426 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3427 gfp_mask, &nr_scanned);
3428 nr_reclaimed += reclaimed;
3429 *total_scanned += nr_scanned;
3430 spin_lock_irq(&mctz->lock);
3431 __mem_cgroup_remove_exceeded(mz, mctz);
3434 * If we failed to reclaim anything from this memory cgroup
3435 * it is time to move on to the next cgroup
3439 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3441 excess = soft_limit_excess(mz->memcg);
3443 * One school of thought says that we should not add
3444 * back the node to the tree if reclaim returns 0.
3445 * But our reclaim could return 0, simply because due
3446 * to priority we are exposing a smaller subset of
3447 * memory to reclaim from. Consider this as a longer
3450 /* If excess == 0, no tree ops */
3451 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3452 spin_unlock_irq(&mctz->lock);
3453 css_put(&mz->memcg->css);
3456 * Could not reclaim anything and there are no more
3457 * mem cgroups to try or we seem to be looping without
3458 * reclaiming anything.
3460 if (!nr_reclaimed &&
3462 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3464 } while (!nr_reclaimed);
3466 css_put(&next_mz->memcg->css);
3467 return nr_reclaimed;
3471 * Reclaims as many pages from the given memcg as possible.
3473 * Caller is responsible for holding css reference for memcg.
3475 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3477 int nr_retries = MAX_RECLAIM_RETRIES;
3479 /* we call try-to-free pages for make this cgroup empty */
3480 lru_add_drain_all();
3482 drain_all_stock(memcg);
3484 /* try to free all pages in this cgroup */
3485 while (nr_retries && page_counter_read(&memcg->memory)) {
3486 if (signal_pending(current))
3489 if (!try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true))
3496 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3497 char *buf, size_t nbytes,
3500 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3502 if (mem_cgroup_is_root(memcg))
3504 return mem_cgroup_force_empty(memcg) ?: nbytes;
3507 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3513 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3514 struct cftype *cft, u64 val)
3519 pr_warn_once("Non-hierarchical mode is deprecated. "
3520 "Please report your usecase to linux-mm@kvack.org if you "
3521 "depend on this functionality.\n");
3526 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3530 if (mem_cgroup_is_root(memcg)) {
3531 mem_cgroup_flush_stats();
3532 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3533 memcg_page_state(memcg, NR_ANON_MAPPED);
3535 val += memcg_page_state(memcg, MEMCG_SWAP);
3538 val = page_counter_read(&memcg->memory);
3540 val = page_counter_read(&memcg->memsw);
3553 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3556 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3557 struct page_counter *counter;
3559 switch (MEMFILE_TYPE(cft->private)) {
3561 counter = &memcg->memory;
3564 counter = &memcg->memsw;
3567 counter = &memcg->kmem;
3570 counter = &memcg->tcpmem;
3576 switch (MEMFILE_ATTR(cft->private)) {
3578 if (counter == &memcg->memory)
3579 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3580 if (counter == &memcg->memsw)
3581 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3582 return (u64)page_counter_read(counter) * PAGE_SIZE;
3584 return (u64)counter->max * PAGE_SIZE;
3586 return (u64)counter->watermark * PAGE_SIZE;
3588 return counter->failcnt;
3589 case RES_SOFT_LIMIT:
3590 return (u64)memcg->soft_limit * PAGE_SIZE;
3596 #ifdef CONFIG_MEMCG_KMEM
3597 static int memcg_online_kmem(struct mem_cgroup *memcg)
3599 struct obj_cgroup *objcg;
3602 if (cgroup_memory_nokmem)
3605 if (unlikely(mem_cgroup_is_root(memcg)))
3608 memcg_id = ida_alloc_max(&memcg_cache_ida, MEMCG_CACHES_MAX_SIZE - 1,
3613 objcg = obj_cgroup_alloc();
3615 ida_free(&memcg_cache_ida, memcg_id);
3618 objcg->memcg = memcg;
3619 rcu_assign_pointer(memcg->objcg, objcg);
3621 static_branch_enable(&memcg_kmem_enabled_key);
3623 memcg->kmemcg_id = memcg_id;
3628 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3630 struct mem_cgroup *parent;
3633 if (cgroup_memory_nokmem)
3636 if (unlikely(mem_cgroup_is_root(memcg)))
3639 parent = parent_mem_cgroup(memcg);
3641 parent = root_mem_cgroup;
3643 memcg_reparent_objcgs(memcg, parent);
3646 * memcg_reparent_list_lrus() can change memcg->kmemcg_id.
3647 * Cache it to local @kmemcg_id.
3649 kmemcg_id = memcg->kmemcg_id;
3652 * After we have finished memcg_reparent_objcgs(), all list_lrus
3653 * corresponding to this cgroup are guaranteed to remain empty.
3654 * The ordering is imposed by list_lru_node->lock taken by
3655 * memcg_reparent_list_lrus().
3657 memcg_reparent_list_lrus(memcg, parent);
3659 ida_free(&memcg_cache_ida, kmemcg_id);
3662 static int memcg_online_kmem(struct mem_cgroup *memcg)
3666 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3669 #endif /* CONFIG_MEMCG_KMEM */
3671 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3675 mutex_lock(&memcg_max_mutex);
3677 ret = page_counter_set_max(&memcg->tcpmem, max);
3681 if (!memcg->tcpmem_active) {
3683 * The active flag needs to be written after the static_key
3684 * update. This is what guarantees that the socket activation
3685 * function is the last one to run. See mem_cgroup_sk_alloc()
3686 * for details, and note that we don't mark any socket as
3687 * belonging to this memcg until that flag is up.
3689 * We need to do this, because static_keys will span multiple
3690 * sites, but we can't control their order. If we mark a socket
3691 * as accounted, but the accounting functions are not patched in
3692 * yet, we'll lose accounting.
3694 * We never race with the readers in mem_cgroup_sk_alloc(),
3695 * because when this value change, the code to process it is not
3698 static_branch_inc(&memcg_sockets_enabled_key);
3699 memcg->tcpmem_active = true;
3702 mutex_unlock(&memcg_max_mutex);
3707 * The user of this function is...
3710 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3711 char *buf, size_t nbytes, loff_t off)
3713 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3714 unsigned long nr_pages;
3717 buf = strstrip(buf);
3718 ret = page_counter_memparse(buf, "-1", &nr_pages);
3722 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3724 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3728 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3730 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3733 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3736 /* kmem.limit_in_bytes is deprecated. */
3740 ret = memcg_update_tcp_max(memcg, nr_pages);
3744 case RES_SOFT_LIMIT:
3745 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
3748 memcg->soft_limit = nr_pages;
3753 return ret ?: nbytes;
3756 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3757 size_t nbytes, loff_t off)
3759 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3760 struct page_counter *counter;
3762 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3764 counter = &memcg->memory;
3767 counter = &memcg->memsw;
3770 counter = &memcg->kmem;
3773 counter = &memcg->tcpmem;
3779 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3781 page_counter_reset_watermark(counter);
3784 counter->failcnt = 0;
3793 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3796 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3800 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3801 struct cftype *cft, u64 val)
3803 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3805 if (val & ~MOVE_MASK)
3809 * No kind of locking is needed in here, because ->can_attach() will
3810 * check this value once in the beginning of the process, and then carry
3811 * on with stale data. This means that changes to this value will only
3812 * affect task migrations starting after the change.
3814 memcg->move_charge_at_immigrate = val;
3818 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3819 struct cftype *cft, u64 val)
3827 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3828 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3829 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3831 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3832 int nid, unsigned int lru_mask, bool tree)
3834 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3835 unsigned long nr = 0;
3838 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3841 if (!(BIT(lru) & lru_mask))
3844 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3846 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3851 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3852 unsigned int lru_mask,
3855 unsigned long nr = 0;
3859 if (!(BIT(lru) & lru_mask))
3862 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3864 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3869 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3873 unsigned int lru_mask;
3876 static const struct numa_stat stats[] = {
3877 { "total", LRU_ALL },
3878 { "file", LRU_ALL_FILE },
3879 { "anon", LRU_ALL_ANON },
3880 { "unevictable", BIT(LRU_UNEVICTABLE) },
3882 const struct numa_stat *stat;
3884 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3886 mem_cgroup_flush_stats();
3888 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3889 seq_printf(m, "%s=%lu", stat->name,
3890 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3892 for_each_node_state(nid, N_MEMORY)
3893 seq_printf(m, " N%d=%lu", nid,
3894 mem_cgroup_node_nr_lru_pages(memcg, nid,
3895 stat->lru_mask, false));
3899 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3901 seq_printf(m, "hierarchical_%s=%lu", stat->name,
3902 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3904 for_each_node_state(nid, N_MEMORY)
3905 seq_printf(m, " N%d=%lu", nid,
3906 mem_cgroup_node_nr_lru_pages(memcg, nid,
3907 stat->lru_mask, true));
3913 #endif /* CONFIG_NUMA */
3915 static const unsigned int memcg1_stats[] = {
3918 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3928 static const char *const memcg1_stat_names[] = {
3931 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3941 /* Universal VM events cgroup1 shows, original sort order */
3942 static const unsigned int memcg1_events[] = {
3949 static int memcg_stat_show(struct seq_file *m, void *v)
3951 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3952 unsigned long memory, memsw;
3953 struct mem_cgroup *mi;
3956 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3958 mem_cgroup_flush_stats();
3960 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3963 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3965 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
3966 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
3969 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3970 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
3971 memcg_events_local(memcg, memcg1_events[i]));
3973 for (i = 0; i < NR_LRU_LISTS; i++)
3974 seq_printf(m, "%s %lu\n", lru_list_name(i),
3975 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3978 /* Hierarchical information */
3979 memory = memsw = PAGE_COUNTER_MAX;
3980 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3981 memory = min(memory, READ_ONCE(mi->memory.max));
3982 memsw = min(memsw, READ_ONCE(mi->memsw.max));
3984 seq_printf(m, "hierarchical_memory_limit %llu\n",
3985 (u64)memory * PAGE_SIZE);
3986 if (do_memsw_account())
3987 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3988 (u64)memsw * PAGE_SIZE);
3990 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3993 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3995 nr = memcg_page_state(memcg, memcg1_stats[i]);
3996 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3997 (u64)nr * PAGE_SIZE);
4000 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4001 seq_printf(m, "total_%s %llu\n",
4002 vm_event_name(memcg1_events[i]),
4003 (u64)memcg_events(memcg, memcg1_events[i]));
4005 for (i = 0; i < NR_LRU_LISTS; i++)
4006 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4007 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4010 #ifdef CONFIG_DEBUG_VM
4013 struct mem_cgroup_per_node *mz;
4014 unsigned long anon_cost = 0;
4015 unsigned long file_cost = 0;
4017 for_each_online_pgdat(pgdat) {
4018 mz = memcg->nodeinfo[pgdat->node_id];
4020 anon_cost += mz->lruvec.anon_cost;
4021 file_cost += mz->lruvec.file_cost;
4023 seq_printf(m, "anon_cost %lu\n", anon_cost);
4024 seq_printf(m, "file_cost %lu\n", file_cost);
4031 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4034 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4036 return mem_cgroup_swappiness(memcg);
4039 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4040 struct cftype *cft, u64 val)
4042 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4047 if (!mem_cgroup_is_root(memcg))
4048 memcg->swappiness = val;
4050 vm_swappiness = val;
4055 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4057 struct mem_cgroup_threshold_ary *t;
4058 unsigned long usage;
4063 t = rcu_dereference(memcg->thresholds.primary);
4065 t = rcu_dereference(memcg->memsw_thresholds.primary);
4070 usage = mem_cgroup_usage(memcg, swap);
4073 * current_threshold points to threshold just below or equal to usage.
4074 * If it's not true, a threshold was crossed after last
4075 * call of __mem_cgroup_threshold().
4077 i = t->current_threshold;
4080 * Iterate backward over array of thresholds starting from
4081 * current_threshold and check if a threshold is crossed.
4082 * If none of thresholds below usage is crossed, we read
4083 * only one element of the array here.
4085 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4086 eventfd_signal(t->entries[i].eventfd, 1);
4088 /* i = current_threshold + 1 */
4092 * Iterate forward over array of thresholds starting from
4093 * current_threshold+1 and check if a threshold is crossed.
4094 * If none of thresholds above usage is crossed, we read
4095 * only one element of the array here.
4097 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4098 eventfd_signal(t->entries[i].eventfd, 1);
4100 /* Update current_threshold */
4101 t->current_threshold = i - 1;
4106 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4109 __mem_cgroup_threshold(memcg, false);
4110 if (do_memsw_account())
4111 __mem_cgroup_threshold(memcg, true);
4113 memcg = parent_mem_cgroup(memcg);
4117 static int compare_thresholds(const void *a, const void *b)
4119 const struct mem_cgroup_threshold *_a = a;
4120 const struct mem_cgroup_threshold *_b = b;
4122 if (_a->threshold > _b->threshold)
4125 if (_a->threshold < _b->threshold)
4131 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4133 struct mem_cgroup_eventfd_list *ev;
4135 spin_lock(&memcg_oom_lock);
4137 list_for_each_entry(ev, &memcg->oom_notify, list)
4138 eventfd_signal(ev->eventfd, 1);
4140 spin_unlock(&memcg_oom_lock);
4144 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4146 struct mem_cgroup *iter;
4148 for_each_mem_cgroup_tree(iter, memcg)
4149 mem_cgroup_oom_notify_cb(iter);
4152 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4153 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4155 struct mem_cgroup_thresholds *thresholds;
4156 struct mem_cgroup_threshold_ary *new;
4157 unsigned long threshold;
4158 unsigned long usage;
4161 ret = page_counter_memparse(args, "-1", &threshold);
4165 mutex_lock(&memcg->thresholds_lock);
4168 thresholds = &memcg->thresholds;
4169 usage = mem_cgroup_usage(memcg, false);
4170 } else if (type == _MEMSWAP) {
4171 thresholds = &memcg->memsw_thresholds;
4172 usage = mem_cgroup_usage(memcg, true);
4176 /* Check if a threshold crossed before adding a new one */
4177 if (thresholds->primary)
4178 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4180 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4182 /* Allocate memory for new array of thresholds */
4183 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4190 /* Copy thresholds (if any) to new array */
4191 if (thresholds->primary)
4192 memcpy(new->entries, thresholds->primary->entries,
4193 flex_array_size(new, entries, size - 1));
4195 /* Add new threshold */
4196 new->entries[size - 1].eventfd = eventfd;
4197 new->entries[size - 1].threshold = threshold;
4199 /* Sort thresholds. Registering of new threshold isn't time-critical */
4200 sort(new->entries, size, sizeof(*new->entries),
4201 compare_thresholds, NULL);
4203 /* Find current threshold */
4204 new->current_threshold = -1;
4205 for (i = 0; i < size; i++) {
4206 if (new->entries[i].threshold <= usage) {
4208 * new->current_threshold will not be used until
4209 * rcu_assign_pointer(), so it's safe to increment
4212 ++new->current_threshold;
4217 /* Free old spare buffer and save old primary buffer as spare */
4218 kfree(thresholds->spare);
4219 thresholds->spare = thresholds->primary;
4221 rcu_assign_pointer(thresholds->primary, new);
4223 /* To be sure that nobody uses thresholds */
4227 mutex_unlock(&memcg->thresholds_lock);
4232 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4233 struct eventfd_ctx *eventfd, const char *args)
4235 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4238 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4239 struct eventfd_ctx *eventfd, const char *args)
4241 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4244 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4245 struct eventfd_ctx *eventfd, enum res_type type)
4247 struct mem_cgroup_thresholds *thresholds;
4248 struct mem_cgroup_threshold_ary *new;
4249 unsigned long usage;
4250 int i, j, size, entries;
4252 mutex_lock(&memcg->thresholds_lock);
4255 thresholds = &memcg->thresholds;
4256 usage = mem_cgroup_usage(memcg, false);
4257 } else if (type == _MEMSWAP) {
4258 thresholds = &memcg->memsw_thresholds;
4259 usage = mem_cgroup_usage(memcg, true);
4263 if (!thresholds->primary)
4266 /* Check if a threshold crossed before removing */
4267 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4269 /* Calculate new number of threshold */
4271 for (i = 0; i < thresholds->primary->size; i++) {
4272 if (thresholds->primary->entries[i].eventfd != eventfd)
4278 new = thresholds->spare;
4280 /* If no items related to eventfd have been cleared, nothing to do */
4284 /* Set thresholds array to NULL if we don't have thresholds */
4293 /* Copy thresholds and find current threshold */
4294 new->current_threshold = -1;
4295 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4296 if (thresholds->primary->entries[i].eventfd == eventfd)
4299 new->entries[j] = thresholds->primary->entries[i];
4300 if (new->entries[j].threshold <= usage) {
4302 * new->current_threshold will not be used
4303 * until rcu_assign_pointer(), so it's safe to increment
4306 ++new->current_threshold;
4312 /* Swap primary and spare array */
4313 thresholds->spare = thresholds->primary;
4315 rcu_assign_pointer(thresholds->primary, new);
4317 /* To be sure that nobody uses thresholds */
4320 /* If all events are unregistered, free the spare array */
4322 kfree(thresholds->spare);
4323 thresholds->spare = NULL;
4326 mutex_unlock(&memcg->thresholds_lock);
4329 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4330 struct eventfd_ctx *eventfd)
4332 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4335 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4336 struct eventfd_ctx *eventfd)
4338 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4341 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4342 struct eventfd_ctx *eventfd, const char *args)
4344 struct mem_cgroup_eventfd_list *event;
4346 event = kmalloc(sizeof(*event), GFP_KERNEL);
4350 spin_lock(&memcg_oom_lock);
4352 event->eventfd = eventfd;
4353 list_add(&event->list, &memcg->oom_notify);
4355 /* already in OOM ? */
4356 if (memcg->under_oom)
4357 eventfd_signal(eventfd, 1);
4358 spin_unlock(&memcg_oom_lock);
4363 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4364 struct eventfd_ctx *eventfd)
4366 struct mem_cgroup_eventfd_list *ev, *tmp;
4368 spin_lock(&memcg_oom_lock);
4370 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4371 if (ev->eventfd == eventfd) {
4372 list_del(&ev->list);
4377 spin_unlock(&memcg_oom_lock);
4380 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4382 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4384 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4385 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4386 seq_printf(sf, "oom_kill %lu\n",
4387 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4391 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4392 struct cftype *cft, u64 val)
4394 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4396 /* cannot set to root cgroup and only 0 and 1 are allowed */
4397 if (mem_cgroup_is_root(memcg) || !((val == 0) || (val == 1)))
4400 memcg->oom_kill_disable = val;
4402 memcg_oom_recover(memcg);
4407 #ifdef CONFIG_CGROUP_WRITEBACK
4409 #include <trace/events/writeback.h>
4411 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4413 return wb_domain_init(&memcg->cgwb_domain, gfp);
4416 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4418 wb_domain_exit(&memcg->cgwb_domain);
4421 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4423 wb_domain_size_changed(&memcg->cgwb_domain);
4426 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4428 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4430 if (!memcg->css.parent)
4433 return &memcg->cgwb_domain;
4437 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4438 * @wb: bdi_writeback in question
4439 * @pfilepages: out parameter for number of file pages
4440 * @pheadroom: out parameter for number of allocatable pages according to memcg
4441 * @pdirty: out parameter for number of dirty pages
4442 * @pwriteback: out parameter for number of pages under writeback
4444 * Determine the numbers of file, headroom, dirty, and writeback pages in
4445 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4446 * is a bit more involved.
4448 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4449 * headroom is calculated as the lowest headroom of itself and the
4450 * ancestors. Note that this doesn't consider the actual amount of
4451 * available memory in the system. The caller should further cap
4452 * *@pheadroom accordingly.
4454 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4455 unsigned long *pheadroom, unsigned long *pdirty,
4456 unsigned long *pwriteback)
4458 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4459 struct mem_cgroup *parent;
4461 mem_cgroup_flush_stats();
4463 *pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
4464 *pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
4465 *pfilepages = memcg_page_state(memcg, NR_INACTIVE_FILE) +
4466 memcg_page_state(memcg, NR_ACTIVE_FILE);
4468 *pheadroom = PAGE_COUNTER_MAX;
4469 while ((parent = parent_mem_cgroup(memcg))) {
4470 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4471 READ_ONCE(memcg->memory.high));
4472 unsigned long used = page_counter_read(&memcg->memory);
4474 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4480 * Foreign dirty flushing
4482 * There's an inherent mismatch between memcg and writeback. The former
4483 * tracks ownership per-page while the latter per-inode. This was a
4484 * deliberate design decision because honoring per-page ownership in the
4485 * writeback path is complicated, may lead to higher CPU and IO overheads
4486 * and deemed unnecessary given that write-sharing an inode across
4487 * different cgroups isn't a common use-case.
4489 * Combined with inode majority-writer ownership switching, this works well
4490 * enough in most cases but there are some pathological cases. For
4491 * example, let's say there are two cgroups A and B which keep writing to
4492 * different but confined parts of the same inode. B owns the inode and
4493 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4494 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4495 * triggering background writeback. A will be slowed down without a way to
4496 * make writeback of the dirty pages happen.
4498 * Conditions like the above can lead to a cgroup getting repeatedly and
4499 * severely throttled after making some progress after each
4500 * dirty_expire_interval while the underlying IO device is almost
4503 * Solving this problem completely requires matching the ownership tracking
4504 * granularities between memcg and writeback in either direction. However,
4505 * the more egregious behaviors can be avoided by simply remembering the
4506 * most recent foreign dirtying events and initiating remote flushes on
4507 * them when local writeback isn't enough to keep the memory clean enough.
4509 * The following two functions implement such mechanism. When a foreign
4510 * page - a page whose memcg and writeback ownerships don't match - is
4511 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4512 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4513 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4514 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4515 * foreign bdi_writebacks which haven't expired. Both the numbers of
4516 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4517 * limited to MEMCG_CGWB_FRN_CNT.
4519 * The mechanism only remembers IDs and doesn't hold any object references.
4520 * As being wrong occasionally doesn't matter, updates and accesses to the
4521 * records are lockless and racy.
4523 void mem_cgroup_track_foreign_dirty_slowpath(struct folio *folio,
4524 struct bdi_writeback *wb)
4526 struct mem_cgroup *memcg = folio_memcg(folio);
4527 struct memcg_cgwb_frn *frn;
4528 u64 now = get_jiffies_64();
4529 u64 oldest_at = now;
4533 trace_track_foreign_dirty(folio, wb);
4536 * Pick the slot to use. If there is already a slot for @wb, keep
4537 * using it. If not replace the oldest one which isn't being
4540 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4541 frn = &memcg->cgwb_frn[i];
4542 if (frn->bdi_id == wb->bdi->id &&
4543 frn->memcg_id == wb->memcg_css->id)
4545 if (time_before64(frn->at, oldest_at) &&
4546 atomic_read(&frn->done.cnt) == 1) {
4548 oldest_at = frn->at;
4552 if (i < MEMCG_CGWB_FRN_CNT) {
4554 * Re-using an existing one. Update timestamp lazily to
4555 * avoid making the cacheline hot. We want them to be
4556 * reasonably up-to-date and significantly shorter than
4557 * dirty_expire_interval as that's what expires the record.
4558 * Use the shorter of 1s and dirty_expire_interval / 8.
4560 unsigned long update_intv =
4561 min_t(unsigned long, HZ,
4562 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4564 if (time_before64(frn->at, now - update_intv))
4566 } else if (oldest >= 0) {
4567 /* replace the oldest free one */
4568 frn = &memcg->cgwb_frn[oldest];
4569 frn->bdi_id = wb->bdi->id;
4570 frn->memcg_id = wb->memcg_css->id;
4575 /* issue foreign writeback flushes for recorded foreign dirtying events */
4576 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4578 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4579 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4580 u64 now = jiffies_64;
4583 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4584 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4587 * If the record is older than dirty_expire_interval,
4588 * writeback on it has already started. No need to kick it
4589 * off again. Also, don't start a new one if there's
4590 * already one in flight.
4592 if (time_after64(frn->at, now - intv) &&
4593 atomic_read(&frn->done.cnt) == 1) {
4595 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4596 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id,
4597 WB_REASON_FOREIGN_FLUSH,
4603 #else /* CONFIG_CGROUP_WRITEBACK */
4605 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4610 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4614 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4618 #endif /* CONFIG_CGROUP_WRITEBACK */
4621 * DO NOT USE IN NEW FILES.
4623 * "cgroup.event_control" implementation.
4625 * This is way over-engineered. It tries to support fully configurable
4626 * events for each user. Such level of flexibility is completely
4627 * unnecessary especially in the light of the planned unified hierarchy.
4629 * Please deprecate this and replace with something simpler if at all
4634 * Unregister event and free resources.
4636 * Gets called from workqueue.
4638 static void memcg_event_remove(struct work_struct *work)
4640 struct mem_cgroup_event *event =
4641 container_of(work, struct mem_cgroup_event, remove);
4642 struct mem_cgroup *memcg = event->memcg;
4644 remove_wait_queue(event->wqh, &event->wait);
4646 event->unregister_event(memcg, event->eventfd);
4648 /* Notify userspace the event is going away. */
4649 eventfd_signal(event->eventfd, 1);
4651 eventfd_ctx_put(event->eventfd);
4653 css_put(&memcg->css);
4657 * Gets called on EPOLLHUP on eventfd when user closes it.
4659 * Called with wqh->lock held and interrupts disabled.
4661 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4662 int sync, void *key)
4664 struct mem_cgroup_event *event =
4665 container_of(wait, struct mem_cgroup_event, wait);
4666 struct mem_cgroup *memcg = event->memcg;
4667 __poll_t flags = key_to_poll(key);
4669 if (flags & EPOLLHUP) {
4671 * If the event has been detached at cgroup removal, we
4672 * can simply return knowing the other side will cleanup
4675 * We can't race against event freeing since the other
4676 * side will require wqh->lock via remove_wait_queue(),
4679 spin_lock(&memcg->event_list_lock);
4680 if (!list_empty(&event->list)) {
4681 list_del_init(&event->list);
4683 * We are in atomic context, but cgroup_event_remove()
4684 * may sleep, so we have to call it in workqueue.
4686 schedule_work(&event->remove);
4688 spin_unlock(&memcg->event_list_lock);
4694 static void memcg_event_ptable_queue_proc(struct file *file,
4695 wait_queue_head_t *wqh, poll_table *pt)
4697 struct mem_cgroup_event *event =
4698 container_of(pt, struct mem_cgroup_event, pt);
4701 add_wait_queue(wqh, &event->wait);
4705 * DO NOT USE IN NEW FILES.
4707 * Parse input and register new cgroup event handler.
4709 * Input must be in format '<event_fd> <control_fd> <args>'.
4710 * Interpretation of args is defined by control file implementation.
4712 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4713 char *buf, size_t nbytes, loff_t off)
4715 struct cgroup_subsys_state *css = of_css(of);
4716 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4717 struct mem_cgroup_event *event;
4718 struct cgroup_subsys_state *cfile_css;
4719 unsigned int efd, cfd;
4726 if (IS_ENABLED(CONFIG_PREEMPT_RT))
4729 buf = strstrip(buf);
4731 efd = simple_strtoul(buf, &endp, 10);
4736 cfd = simple_strtoul(buf, &endp, 10);
4737 if ((*endp != ' ') && (*endp != '\0'))
4741 event = kzalloc(sizeof(*event), GFP_KERNEL);
4745 event->memcg = memcg;
4746 INIT_LIST_HEAD(&event->list);
4747 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4748 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4749 INIT_WORK(&event->remove, memcg_event_remove);
4757 event->eventfd = eventfd_ctx_fileget(efile.file);
4758 if (IS_ERR(event->eventfd)) {
4759 ret = PTR_ERR(event->eventfd);
4766 goto out_put_eventfd;
4769 /* the process need read permission on control file */
4770 /* AV: shouldn't we check that it's been opened for read instead? */
4771 ret = file_permission(cfile.file, MAY_READ);
4776 * Determine the event callbacks and set them in @event. This used
4777 * to be done via struct cftype but cgroup core no longer knows
4778 * about these events. The following is crude but the whole thing
4779 * is for compatibility anyway.
4781 * DO NOT ADD NEW FILES.
4783 name = cfile.file->f_path.dentry->d_name.name;
4785 if (!strcmp(name, "memory.usage_in_bytes")) {
4786 event->register_event = mem_cgroup_usage_register_event;
4787 event->unregister_event = mem_cgroup_usage_unregister_event;
4788 } else if (!strcmp(name, "memory.oom_control")) {
4789 event->register_event = mem_cgroup_oom_register_event;
4790 event->unregister_event = mem_cgroup_oom_unregister_event;
4791 } else if (!strcmp(name, "memory.pressure_level")) {
4792 event->register_event = vmpressure_register_event;
4793 event->unregister_event = vmpressure_unregister_event;
4794 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4795 event->register_event = memsw_cgroup_usage_register_event;
4796 event->unregister_event = memsw_cgroup_usage_unregister_event;
4803 * Verify @cfile should belong to @css. Also, remaining events are
4804 * automatically removed on cgroup destruction but the removal is
4805 * asynchronous, so take an extra ref on @css.
4807 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4808 &memory_cgrp_subsys);
4810 if (IS_ERR(cfile_css))
4812 if (cfile_css != css) {
4817 ret = event->register_event(memcg, event->eventfd, buf);
4821 vfs_poll(efile.file, &event->pt);
4823 spin_lock_irq(&memcg->event_list_lock);
4824 list_add(&event->list, &memcg->event_list);
4825 spin_unlock_irq(&memcg->event_list_lock);
4837 eventfd_ctx_put(event->eventfd);
4846 #if defined(CONFIG_MEMCG_KMEM) && (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
4847 static int mem_cgroup_slab_show(struct seq_file *m, void *p)
4851 * Please, take a look at tools/cgroup/slabinfo.py .
4857 static struct cftype mem_cgroup_legacy_files[] = {
4859 .name = "usage_in_bytes",
4860 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4861 .read_u64 = mem_cgroup_read_u64,
4864 .name = "max_usage_in_bytes",
4865 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4866 .write = mem_cgroup_reset,
4867 .read_u64 = mem_cgroup_read_u64,
4870 .name = "limit_in_bytes",
4871 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4872 .write = mem_cgroup_write,
4873 .read_u64 = mem_cgroup_read_u64,
4876 .name = "soft_limit_in_bytes",
4877 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4878 .write = mem_cgroup_write,
4879 .read_u64 = mem_cgroup_read_u64,
4883 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4884 .write = mem_cgroup_reset,
4885 .read_u64 = mem_cgroup_read_u64,
4889 .seq_show = memcg_stat_show,
4892 .name = "force_empty",
4893 .write = mem_cgroup_force_empty_write,
4896 .name = "use_hierarchy",
4897 .write_u64 = mem_cgroup_hierarchy_write,
4898 .read_u64 = mem_cgroup_hierarchy_read,
4901 .name = "cgroup.event_control", /* XXX: for compat */
4902 .write = memcg_write_event_control,
4903 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4906 .name = "swappiness",
4907 .read_u64 = mem_cgroup_swappiness_read,
4908 .write_u64 = mem_cgroup_swappiness_write,
4911 .name = "move_charge_at_immigrate",
4912 .read_u64 = mem_cgroup_move_charge_read,
4913 .write_u64 = mem_cgroup_move_charge_write,
4916 .name = "oom_control",
4917 .seq_show = mem_cgroup_oom_control_read,
4918 .write_u64 = mem_cgroup_oom_control_write,
4919 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4922 .name = "pressure_level",
4926 .name = "numa_stat",
4927 .seq_show = memcg_numa_stat_show,
4931 .name = "kmem.limit_in_bytes",
4932 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4933 .write = mem_cgroup_write,
4934 .read_u64 = mem_cgroup_read_u64,
4937 .name = "kmem.usage_in_bytes",
4938 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4939 .read_u64 = mem_cgroup_read_u64,
4942 .name = "kmem.failcnt",
4943 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4944 .write = mem_cgroup_reset,
4945 .read_u64 = mem_cgroup_read_u64,
4948 .name = "kmem.max_usage_in_bytes",
4949 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4950 .write = mem_cgroup_reset,
4951 .read_u64 = mem_cgroup_read_u64,
4953 #if defined(CONFIG_MEMCG_KMEM) && \
4954 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
4956 .name = "kmem.slabinfo",
4957 .seq_show = mem_cgroup_slab_show,
4961 .name = "kmem.tcp.limit_in_bytes",
4962 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4963 .write = mem_cgroup_write,
4964 .read_u64 = mem_cgroup_read_u64,
4967 .name = "kmem.tcp.usage_in_bytes",
4968 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4969 .read_u64 = mem_cgroup_read_u64,
4972 .name = "kmem.tcp.failcnt",
4973 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4974 .write = mem_cgroup_reset,
4975 .read_u64 = mem_cgroup_read_u64,
4978 .name = "kmem.tcp.max_usage_in_bytes",
4979 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4980 .write = mem_cgroup_reset,
4981 .read_u64 = mem_cgroup_read_u64,
4983 { }, /* terminate */
4987 * Private memory cgroup IDR
4989 * Swap-out records and page cache shadow entries need to store memcg
4990 * references in constrained space, so we maintain an ID space that is
4991 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4992 * memory-controlled cgroups to 64k.
4994 * However, there usually are many references to the offline CSS after
4995 * the cgroup has been destroyed, such as page cache or reclaimable
4996 * slab objects, that don't need to hang on to the ID. We want to keep
4997 * those dead CSS from occupying IDs, or we might quickly exhaust the
4998 * relatively small ID space and prevent the creation of new cgroups
4999 * even when there are much fewer than 64k cgroups - possibly none.
5001 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5002 * be freed and recycled when it's no longer needed, which is usually
5003 * when the CSS is offlined.
5005 * The only exception to that are records of swapped out tmpfs/shmem
5006 * pages that need to be attributed to live ancestors on swapin. But
5007 * those references are manageable from userspace.
5010 static DEFINE_IDR(mem_cgroup_idr);
5012 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5014 if (memcg->id.id > 0) {
5015 idr_remove(&mem_cgroup_idr, memcg->id.id);
5020 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5023 refcount_add(n, &memcg->id.ref);
5026 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5028 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5029 mem_cgroup_id_remove(memcg);
5031 /* Memcg ID pins CSS */
5032 css_put(&memcg->css);
5036 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5038 mem_cgroup_id_put_many(memcg, 1);
5042 * mem_cgroup_from_id - look up a memcg from a memcg id
5043 * @id: the memcg id to look up
5045 * Caller must hold rcu_read_lock().
5047 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5049 WARN_ON_ONCE(!rcu_read_lock_held());
5050 return idr_find(&mem_cgroup_idr, id);
5053 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5055 struct mem_cgroup_per_node *pn;
5058 * This routine is called against possible nodes.
5059 * But it's BUG to call kmalloc() against offline node.
5061 * TODO: this routine can waste much memory for nodes which will
5062 * never be onlined. It's better to use memory hotplug callback
5065 if (!node_state(node, N_NORMAL_MEMORY))
5067 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5071 pn->lruvec_stats_percpu = alloc_percpu_gfp(struct lruvec_stats_percpu,
5072 GFP_KERNEL_ACCOUNT);
5073 if (!pn->lruvec_stats_percpu) {
5078 lruvec_init(&pn->lruvec);
5081 memcg->nodeinfo[node] = pn;
5085 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5087 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5092 free_percpu(pn->lruvec_stats_percpu);
5096 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5101 free_mem_cgroup_per_node_info(memcg, node);
5102 free_percpu(memcg->vmstats_percpu);
5106 static void mem_cgroup_free(struct mem_cgroup *memcg)
5108 memcg_wb_domain_exit(memcg);
5109 __mem_cgroup_free(memcg);
5112 static struct mem_cgroup *mem_cgroup_alloc(void)
5114 struct mem_cgroup *memcg;
5116 int __maybe_unused i;
5117 long error = -ENOMEM;
5119 memcg = kzalloc(struct_size(memcg, nodeinfo, nr_node_ids), GFP_KERNEL);
5121 return ERR_PTR(error);
5123 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5124 1, MEM_CGROUP_ID_MAX,
5126 if (memcg->id.id < 0) {
5127 error = memcg->id.id;
5131 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5132 GFP_KERNEL_ACCOUNT);
5133 if (!memcg->vmstats_percpu)
5137 if (alloc_mem_cgroup_per_node_info(memcg, node))
5140 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5143 INIT_WORK(&memcg->high_work, high_work_func);
5144 INIT_LIST_HEAD(&memcg->oom_notify);
5145 mutex_init(&memcg->thresholds_lock);
5146 spin_lock_init(&memcg->move_lock);
5147 vmpressure_init(&memcg->vmpressure);
5148 INIT_LIST_HEAD(&memcg->event_list);
5149 spin_lock_init(&memcg->event_list_lock);
5150 memcg->socket_pressure = jiffies;
5151 #ifdef CONFIG_MEMCG_KMEM
5152 memcg->kmemcg_id = -1;
5153 INIT_LIST_HEAD(&memcg->objcg_list);
5155 #ifdef CONFIG_CGROUP_WRITEBACK
5156 INIT_LIST_HEAD(&memcg->cgwb_list);
5157 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5158 memcg->cgwb_frn[i].done =
5159 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5161 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5162 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5163 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5164 memcg->deferred_split_queue.split_queue_len = 0;
5166 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5169 mem_cgroup_id_remove(memcg);
5170 __mem_cgroup_free(memcg);
5171 return ERR_PTR(error);
5174 static struct cgroup_subsys_state * __ref
5175 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5177 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5178 struct mem_cgroup *memcg, *old_memcg;
5180 old_memcg = set_active_memcg(parent);
5181 memcg = mem_cgroup_alloc();
5182 set_active_memcg(old_memcg);
5184 return ERR_CAST(memcg);
5186 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5187 memcg->soft_limit = PAGE_COUNTER_MAX;
5188 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5190 memcg->swappiness = mem_cgroup_swappiness(parent);
5191 memcg->oom_kill_disable = parent->oom_kill_disable;
5193 page_counter_init(&memcg->memory, &parent->memory);
5194 page_counter_init(&memcg->swap, &parent->swap);
5195 page_counter_init(&memcg->kmem, &parent->kmem);
5196 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5198 page_counter_init(&memcg->memory, NULL);
5199 page_counter_init(&memcg->swap, NULL);
5200 page_counter_init(&memcg->kmem, NULL);
5201 page_counter_init(&memcg->tcpmem, NULL);
5203 root_mem_cgroup = memcg;
5207 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5208 static_branch_inc(&memcg_sockets_enabled_key);
5213 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5215 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5217 if (memcg_online_kmem(memcg))
5221 * A memcg must be visible for expand_shrinker_info()
5222 * by the time the maps are allocated. So, we allocate maps
5223 * here, when for_each_mem_cgroup() can't skip it.
5225 if (alloc_shrinker_info(memcg))
5228 /* Online state pins memcg ID, memcg ID pins CSS */
5229 refcount_set(&memcg->id.ref, 1);
5232 if (unlikely(mem_cgroup_is_root(memcg)))
5233 queue_delayed_work(system_unbound_wq, &stats_flush_dwork,
5237 memcg_offline_kmem(memcg);
5239 mem_cgroup_id_remove(memcg);
5243 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5245 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5246 struct mem_cgroup_event *event, *tmp;
5249 * Unregister events and notify userspace.
5250 * Notify userspace about cgroup removing only after rmdir of cgroup
5251 * directory to avoid race between userspace and kernelspace.
5253 spin_lock_irq(&memcg->event_list_lock);
5254 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5255 list_del_init(&event->list);
5256 schedule_work(&event->remove);
5258 spin_unlock_irq(&memcg->event_list_lock);
5260 page_counter_set_min(&memcg->memory, 0);
5261 page_counter_set_low(&memcg->memory, 0);
5263 memcg_offline_kmem(memcg);
5264 reparent_shrinker_deferred(memcg);
5265 wb_memcg_offline(memcg);
5267 drain_all_stock(memcg);
5269 mem_cgroup_id_put(memcg);
5272 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5274 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5276 invalidate_reclaim_iterators(memcg);
5279 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5281 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5282 int __maybe_unused i;
5284 #ifdef CONFIG_CGROUP_WRITEBACK
5285 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5286 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5288 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5289 static_branch_dec(&memcg_sockets_enabled_key);
5291 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5292 static_branch_dec(&memcg_sockets_enabled_key);
5294 vmpressure_cleanup(&memcg->vmpressure);
5295 cancel_work_sync(&memcg->high_work);
5296 mem_cgroup_remove_from_trees(memcg);
5297 free_shrinker_info(memcg);
5298 mem_cgroup_free(memcg);
5302 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5303 * @css: the target css
5305 * Reset the states of the mem_cgroup associated with @css. This is
5306 * invoked when the userland requests disabling on the default hierarchy
5307 * but the memcg is pinned through dependency. The memcg should stop
5308 * applying policies and should revert to the vanilla state as it may be
5309 * made visible again.
5311 * The current implementation only resets the essential configurations.
5312 * This needs to be expanded to cover all the visible parts.
5314 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5316 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5318 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5319 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5320 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5321 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5322 page_counter_set_min(&memcg->memory, 0);
5323 page_counter_set_low(&memcg->memory, 0);
5324 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5325 memcg->soft_limit = PAGE_COUNTER_MAX;
5326 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5327 memcg_wb_domain_size_changed(memcg);
5330 static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state *css, int cpu)
5332 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5333 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5334 struct memcg_vmstats_percpu *statc;
5338 statc = per_cpu_ptr(memcg->vmstats_percpu, cpu);
5340 for (i = 0; i < MEMCG_NR_STAT; i++) {
5342 * Collect the aggregated propagation counts of groups
5343 * below us. We're in a per-cpu loop here and this is
5344 * a global counter, so the first cycle will get them.
5346 delta = memcg->vmstats.state_pending[i];
5348 memcg->vmstats.state_pending[i] = 0;
5350 /* Add CPU changes on this level since the last flush */
5351 v = READ_ONCE(statc->state[i]);
5352 if (v != statc->state_prev[i]) {
5353 delta += v - statc->state_prev[i];
5354 statc->state_prev[i] = v;
5360 /* Aggregate counts on this level and propagate upwards */
5361 memcg->vmstats.state[i] += delta;
5363 parent->vmstats.state_pending[i] += delta;
5366 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
5367 delta = memcg->vmstats.events_pending[i];
5369 memcg->vmstats.events_pending[i] = 0;
5371 v = READ_ONCE(statc->events[i]);
5372 if (v != statc->events_prev[i]) {
5373 delta += v - statc->events_prev[i];
5374 statc->events_prev[i] = v;
5380 memcg->vmstats.events[i] += delta;
5382 parent->vmstats.events_pending[i] += delta;
5385 for_each_node_state(nid, N_MEMORY) {
5386 struct mem_cgroup_per_node *pn = memcg->nodeinfo[nid];
5387 struct mem_cgroup_per_node *ppn = NULL;
5388 struct lruvec_stats_percpu *lstatc;
5391 ppn = parent->nodeinfo[nid];
5393 lstatc = per_cpu_ptr(pn->lruvec_stats_percpu, cpu);
5395 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
5396 delta = pn->lruvec_stats.state_pending[i];
5398 pn->lruvec_stats.state_pending[i] = 0;
5400 v = READ_ONCE(lstatc->state[i]);
5401 if (v != lstatc->state_prev[i]) {
5402 delta += v - lstatc->state_prev[i];
5403 lstatc->state_prev[i] = v;
5409 pn->lruvec_stats.state[i] += delta;
5411 ppn->lruvec_stats.state_pending[i] += delta;
5417 /* Handlers for move charge at task migration. */
5418 static int mem_cgroup_do_precharge(unsigned long count)
5422 /* Try a single bulk charge without reclaim first, kswapd may wake */
5423 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5425 mc.precharge += count;
5429 /* Try charges one by one with reclaim, but do not retry */
5431 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5445 enum mc_target_type {
5452 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5453 unsigned long addr, pte_t ptent)
5455 struct page *page = vm_normal_page(vma, addr, ptent);
5457 if (!page || !page_mapped(page))
5459 if (PageAnon(page)) {
5460 if (!(mc.flags & MOVE_ANON))
5463 if (!(mc.flags & MOVE_FILE))
5466 if (!get_page_unless_zero(page))
5472 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5473 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5474 pte_t ptent, swp_entry_t *entry)
5476 struct page *page = NULL;
5477 swp_entry_t ent = pte_to_swp_entry(ptent);
5479 if (!(mc.flags & MOVE_ANON))
5483 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5484 * a device and because they are not accessible by CPU they are store
5485 * as special swap entry in the CPU page table.
5487 if (is_device_private_entry(ent)) {
5488 page = pfn_swap_entry_to_page(ent);
5490 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5491 * a refcount of 1 when free (unlike normal page)
5493 if (!page_ref_add_unless(page, 1, 1))
5498 if (non_swap_entry(ent))
5502 * Because lookup_swap_cache() updates some statistics counter,
5503 * we call find_get_page() with swapper_space directly.
5505 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5506 entry->val = ent.val;
5511 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5512 pte_t ptent, swp_entry_t *entry)
5518 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5519 unsigned long addr, pte_t ptent)
5521 if (!vma->vm_file) /* anonymous vma */
5523 if (!(mc.flags & MOVE_FILE))
5526 /* page is moved even if it's not RSS of this task(page-faulted). */
5527 /* shmem/tmpfs may report page out on swap: account for that too. */
5528 return find_get_incore_page(vma->vm_file->f_mapping,
5529 linear_page_index(vma, addr));
5533 * mem_cgroup_move_account - move account of the page
5535 * @compound: charge the page as compound or small page
5536 * @from: mem_cgroup which the page is moved from.
5537 * @to: mem_cgroup which the page is moved to. @from != @to.
5539 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5541 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5544 static int mem_cgroup_move_account(struct page *page,
5546 struct mem_cgroup *from,
5547 struct mem_cgroup *to)
5549 struct folio *folio = page_folio(page);
5550 struct lruvec *from_vec, *to_vec;
5551 struct pglist_data *pgdat;
5552 unsigned int nr_pages = compound ? folio_nr_pages(folio) : 1;
5555 VM_BUG_ON(from == to);
5556 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
5557 VM_BUG_ON(compound && !folio_test_large(folio));
5560 * Prevent mem_cgroup_migrate() from looking at
5561 * page's memory cgroup of its source page while we change it.
5564 if (!folio_trylock(folio))
5568 if (folio_memcg(folio) != from)
5571 pgdat = folio_pgdat(folio);
5572 from_vec = mem_cgroup_lruvec(from, pgdat);
5573 to_vec = mem_cgroup_lruvec(to, pgdat);
5575 folio_memcg_lock(folio);
5577 if (folio_test_anon(folio)) {
5578 if (folio_mapped(folio)) {
5579 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5580 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5581 if (folio_test_transhuge(folio)) {
5582 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5584 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5589 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5590 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5592 if (folio_test_swapbacked(folio)) {
5593 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5594 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5597 if (folio_mapped(folio)) {
5598 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5599 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5602 if (folio_test_dirty(folio)) {
5603 struct address_space *mapping = folio_mapping(folio);
5605 if (mapping_can_writeback(mapping)) {
5606 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5608 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5614 if (folio_test_writeback(folio)) {
5615 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5616 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5620 * All state has been migrated, let's switch to the new memcg.
5622 * It is safe to change page's memcg here because the page
5623 * is referenced, charged, isolated, and locked: we can't race
5624 * with (un)charging, migration, LRU putback, or anything else
5625 * that would rely on a stable page's memory cgroup.
5627 * Note that lock_page_memcg is a memcg lock, not a page lock,
5628 * to save space. As soon as we switch page's memory cgroup to a
5629 * new memcg that isn't locked, the above state can change
5630 * concurrently again. Make sure we're truly done with it.
5635 css_put(&from->css);
5637 folio->memcg_data = (unsigned long)to;
5639 __folio_memcg_unlock(from);
5642 nid = folio_nid(folio);
5644 local_irq_disable();
5645 mem_cgroup_charge_statistics(to, nr_pages);
5646 memcg_check_events(to, nid);
5647 mem_cgroup_charge_statistics(from, -nr_pages);
5648 memcg_check_events(from, nid);
5651 folio_unlock(folio);
5657 * get_mctgt_type - get target type of moving charge
5658 * @vma: the vma the pte to be checked belongs
5659 * @addr: the address corresponding to the pte to be checked
5660 * @ptent: the pte to be checked
5661 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5664 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5665 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5666 * move charge. if @target is not NULL, the page is stored in target->page
5667 * with extra refcnt got(Callers should handle it).
5668 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5669 * target for charge migration. if @target is not NULL, the entry is stored
5671 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5672 * (so ZONE_DEVICE page and thus not on the lru).
5673 * For now we such page is charge like a regular page would be as for all
5674 * intent and purposes it is just special memory taking the place of a
5677 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5679 * Called with pte lock held.
5682 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5683 unsigned long addr, pte_t ptent, union mc_target *target)
5685 struct page *page = NULL;
5686 enum mc_target_type ret = MC_TARGET_NONE;
5687 swp_entry_t ent = { .val = 0 };
5689 if (pte_present(ptent))
5690 page = mc_handle_present_pte(vma, addr, ptent);
5691 else if (is_swap_pte(ptent))
5692 page = mc_handle_swap_pte(vma, ptent, &ent);
5693 else if (pte_none(ptent))
5694 page = mc_handle_file_pte(vma, addr, ptent);
5696 if (!page && !ent.val)
5700 * Do only loose check w/o serialization.
5701 * mem_cgroup_move_account() checks the page is valid or
5702 * not under LRU exclusion.
5704 if (page_memcg(page) == mc.from) {
5705 ret = MC_TARGET_PAGE;
5706 if (is_device_private_page(page))
5707 ret = MC_TARGET_DEVICE;
5709 target->page = page;
5711 if (!ret || !target)
5715 * There is a swap entry and a page doesn't exist or isn't charged.
5716 * But we cannot move a tail-page in a THP.
5718 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5719 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5720 ret = MC_TARGET_SWAP;
5727 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5729 * We don't consider PMD mapped swapping or file mapped pages because THP does
5730 * not support them for now.
5731 * Caller should make sure that pmd_trans_huge(pmd) is true.
5733 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5734 unsigned long addr, pmd_t pmd, union mc_target *target)
5736 struct page *page = NULL;
5737 enum mc_target_type ret = MC_TARGET_NONE;
5739 if (unlikely(is_swap_pmd(pmd))) {
5740 VM_BUG_ON(thp_migration_supported() &&
5741 !is_pmd_migration_entry(pmd));
5744 page = pmd_page(pmd);
5745 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5746 if (!(mc.flags & MOVE_ANON))
5748 if (page_memcg(page) == mc.from) {
5749 ret = MC_TARGET_PAGE;
5752 target->page = page;
5758 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5759 unsigned long addr, pmd_t pmd, union mc_target *target)
5761 return MC_TARGET_NONE;
5765 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5766 unsigned long addr, unsigned long end,
5767 struct mm_walk *walk)
5769 struct vm_area_struct *vma = walk->vma;
5773 ptl = pmd_trans_huge_lock(pmd, vma);
5776 * Note their can not be MC_TARGET_DEVICE for now as we do not
5777 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5778 * this might change.
5780 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5781 mc.precharge += HPAGE_PMD_NR;
5786 if (pmd_trans_unstable(pmd))
5788 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5789 for (; addr != end; pte++, addr += PAGE_SIZE)
5790 if (get_mctgt_type(vma, addr, *pte, NULL))
5791 mc.precharge++; /* increment precharge temporarily */
5792 pte_unmap_unlock(pte - 1, ptl);
5798 static const struct mm_walk_ops precharge_walk_ops = {
5799 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5802 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5804 unsigned long precharge;
5807 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5808 mmap_read_unlock(mm);
5810 precharge = mc.precharge;
5816 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5818 unsigned long precharge = mem_cgroup_count_precharge(mm);
5820 VM_BUG_ON(mc.moving_task);
5821 mc.moving_task = current;
5822 return mem_cgroup_do_precharge(precharge);
5825 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5826 static void __mem_cgroup_clear_mc(void)
5828 struct mem_cgroup *from = mc.from;
5829 struct mem_cgroup *to = mc.to;
5831 /* we must uncharge all the leftover precharges from mc.to */
5833 cancel_charge(mc.to, mc.precharge);
5837 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5838 * we must uncharge here.
5840 if (mc.moved_charge) {
5841 cancel_charge(mc.from, mc.moved_charge);
5842 mc.moved_charge = 0;
5844 /* we must fixup refcnts and charges */
5845 if (mc.moved_swap) {
5846 /* uncharge swap account from the old cgroup */
5847 if (!mem_cgroup_is_root(mc.from))
5848 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5850 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5853 * we charged both to->memory and to->memsw, so we
5854 * should uncharge to->memory.
5856 if (!mem_cgroup_is_root(mc.to))
5857 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5861 memcg_oom_recover(from);
5862 memcg_oom_recover(to);
5863 wake_up_all(&mc.waitq);
5866 static void mem_cgroup_clear_mc(void)
5868 struct mm_struct *mm = mc.mm;
5871 * we must clear moving_task before waking up waiters at the end of
5874 mc.moving_task = NULL;
5875 __mem_cgroup_clear_mc();
5876 spin_lock(&mc.lock);
5880 spin_unlock(&mc.lock);
5885 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5887 struct cgroup_subsys_state *css;
5888 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5889 struct mem_cgroup *from;
5890 struct task_struct *leader, *p;
5891 struct mm_struct *mm;
5892 unsigned long move_flags;
5895 /* charge immigration isn't supported on the default hierarchy */
5896 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5900 * Multi-process migrations only happen on the default hierarchy
5901 * where charge immigration is not used. Perform charge
5902 * immigration if @tset contains a leader and whine if there are
5906 cgroup_taskset_for_each_leader(leader, css, tset) {
5909 memcg = mem_cgroup_from_css(css);
5915 * We are now committed to this value whatever it is. Changes in this
5916 * tunable will only affect upcoming migrations, not the current one.
5917 * So we need to save it, and keep it going.
5919 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5923 from = mem_cgroup_from_task(p);
5925 VM_BUG_ON(from == memcg);
5927 mm = get_task_mm(p);
5930 /* We move charges only when we move a owner of the mm */
5931 if (mm->owner == p) {
5934 VM_BUG_ON(mc.precharge);
5935 VM_BUG_ON(mc.moved_charge);
5936 VM_BUG_ON(mc.moved_swap);
5938 spin_lock(&mc.lock);
5942 mc.flags = move_flags;
5943 spin_unlock(&mc.lock);
5944 /* We set mc.moving_task later */
5946 ret = mem_cgroup_precharge_mc(mm);
5948 mem_cgroup_clear_mc();
5955 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5958 mem_cgroup_clear_mc();
5961 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5962 unsigned long addr, unsigned long end,
5963 struct mm_walk *walk)
5966 struct vm_area_struct *vma = walk->vma;
5969 enum mc_target_type target_type;
5970 union mc_target target;
5973 ptl = pmd_trans_huge_lock(pmd, vma);
5975 if (mc.precharge < HPAGE_PMD_NR) {
5979 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5980 if (target_type == MC_TARGET_PAGE) {
5982 if (!isolate_lru_page(page)) {
5983 if (!mem_cgroup_move_account(page, true,
5985 mc.precharge -= HPAGE_PMD_NR;
5986 mc.moved_charge += HPAGE_PMD_NR;
5988 putback_lru_page(page);
5991 } else if (target_type == MC_TARGET_DEVICE) {
5993 if (!mem_cgroup_move_account(page, true,
5995 mc.precharge -= HPAGE_PMD_NR;
5996 mc.moved_charge += HPAGE_PMD_NR;
6004 if (pmd_trans_unstable(pmd))
6007 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6008 for (; addr != end; addr += PAGE_SIZE) {
6009 pte_t ptent = *(pte++);
6010 bool device = false;
6016 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6017 case MC_TARGET_DEVICE:
6020 case MC_TARGET_PAGE:
6023 * We can have a part of the split pmd here. Moving it
6024 * can be done but it would be too convoluted so simply
6025 * ignore such a partial THP and keep it in original
6026 * memcg. There should be somebody mapping the head.
6028 if (PageTransCompound(page))
6030 if (!device && isolate_lru_page(page))
6032 if (!mem_cgroup_move_account(page, false,
6035 /* we uncharge from mc.from later. */
6039 putback_lru_page(page);
6040 put: /* get_mctgt_type() gets the page */
6043 case MC_TARGET_SWAP:
6045 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6047 mem_cgroup_id_get_many(mc.to, 1);
6048 /* we fixup other refcnts and charges later. */
6056 pte_unmap_unlock(pte - 1, ptl);
6061 * We have consumed all precharges we got in can_attach().
6062 * We try charge one by one, but don't do any additional
6063 * charges to mc.to if we have failed in charge once in attach()
6066 ret = mem_cgroup_do_precharge(1);
6074 static const struct mm_walk_ops charge_walk_ops = {
6075 .pmd_entry = mem_cgroup_move_charge_pte_range,
6078 static void mem_cgroup_move_charge(void)
6080 lru_add_drain_all();
6082 * Signal lock_page_memcg() to take the memcg's move_lock
6083 * while we're moving its pages to another memcg. Then wait
6084 * for already started RCU-only updates to finish.
6086 atomic_inc(&mc.from->moving_account);
6089 if (unlikely(!mmap_read_trylock(mc.mm))) {
6091 * Someone who are holding the mmap_lock might be waiting in
6092 * waitq. So we cancel all extra charges, wake up all waiters,
6093 * and retry. Because we cancel precharges, we might not be able
6094 * to move enough charges, but moving charge is a best-effort
6095 * feature anyway, so it wouldn't be a big problem.
6097 __mem_cgroup_clear_mc();
6102 * When we have consumed all precharges and failed in doing
6103 * additional charge, the page walk just aborts.
6105 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6108 mmap_read_unlock(mc.mm);
6109 atomic_dec(&mc.from->moving_account);
6112 static void mem_cgroup_move_task(void)
6115 mem_cgroup_move_charge();
6116 mem_cgroup_clear_mc();
6119 #else /* !CONFIG_MMU */
6120 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6124 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6127 static void mem_cgroup_move_task(void)
6132 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6134 if (value == PAGE_COUNTER_MAX)
6135 seq_puts(m, "max\n");
6137 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6142 static u64 memory_current_read(struct cgroup_subsys_state *css,
6145 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6147 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6150 static int memory_min_show(struct seq_file *m, void *v)
6152 return seq_puts_memcg_tunable(m,
6153 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6156 static ssize_t memory_min_write(struct kernfs_open_file *of,
6157 char *buf, size_t nbytes, loff_t off)
6159 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6163 buf = strstrip(buf);
6164 err = page_counter_memparse(buf, "max", &min);
6168 page_counter_set_min(&memcg->memory, min);
6173 static int memory_low_show(struct seq_file *m, void *v)
6175 return seq_puts_memcg_tunable(m,
6176 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6179 static ssize_t memory_low_write(struct kernfs_open_file *of,
6180 char *buf, size_t nbytes, loff_t off)
6182 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6186 buf = strstrip(buf);
6187 err = page_counter_memparse(buf, "max", &low);
6191 page_counter_set_low(&memcg->memory, low);
6196 static int memory_high_show(struct seq_file *m, void *v)
6198 return seq_puts_memcg_tunable(m,
6199 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6202 static ssize_t memory_high_write(struct kernfs_open_file *of,
6203 char *buf, size_t nbytes, loff_t off)
6205 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6206 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6207 bool drained = false;
6211 buf = strstrip(buf);
6212 err = page_counter_memparse(buf, "max", &high);
6216 page_counter_set_high(&memcg->memory, high);
6219 unsigned long nr_pages = page_counter_read(&memcg->memory);
6220 unsigned long reclaimed;
6222 if (nr_pages <= high)
6225 if (signal_pending(current))
6229 drain_all_stock(memcg);
6234 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6237 if (!reclaimed && !nr_retries--)
6241 memcg_wb_domain_size_changed(memcg);
6245 static int memory_max_show(struct seq_file *m, void *v)
6247 return seq_puts_memcg_tunable(m,
6248 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6251 static ssize_t memory_max_write(struct kernfs_open_file *of,
6252 char *buf, size_t nbytes, loff_t off)
6254 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6255 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6256 bool drained = false;
6260 buf = strstrip(buf);
6261 err = page_counter_memparse(buf, "max", &max);
6265 xchg(&memcg->memory.max, max);
6268 unsigned long nr_pages = page_counter_read(&memcg->memory);
6270 if (nr_pages <= max)
6273 if (signal_pending(current))
6277 drain_all_stock(memcg);
6283 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6289 memcg_memory_event(memcg, MEMCG_OOM);
6290 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6294 memcg_wb_domain_size_changed(memcg);
6298 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6300 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6301 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6302 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6303 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6304 seq_printf(m, "oom_kill %lu\n",
6305 atomic_long_read(&events[MEMCG_OOM_KILL]));
6306 seq_printf(m, "oom_group_kill %lu\n",
6307 atomic_long_read(&events[MEMCG_OOM_GROUP_KILL]));
6310 static int memory_events_show(struct seq_file *m, void *v)
6312 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6314 __memory_events_show(m, memcg->memory_events);
6318 static int memory_events_local_show(struct seq_file *m, void *v)
6320 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6322 __memory_events_show(m, memcg->memory_events_local);
6326 static int memory_stat_show(struct seq_file *m, void *v)
6328 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6331 buf = memory_stat_format(memcg);
6340 static inline unsigned long lruvec_page_state_output(struct lruvec *lruvec,
6343 return lruvec_page_state(lruvec, item) * memcg_page_state_unit(item);
6346 static int memory_numa_stat_show(struct seq_file *m, void *v)
6349 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6351 mem_cgroup_flush_stats();
6353 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6356 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6359 seq_printf(m, "%s", memory_stats[i].name);
6360 for_each_node_state(nid, N_MEMORY) {
6362 struct lruvec *lruvec;
6364 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6365 size = lruvec_page_state_output(lruvec,
6366 memory_stats[i].idx);
6367 seq_printf(m, " N%d=%llu", nid, size);
6376 static int memory_oom_group_show(struct seq_file *m, void *v)
6378 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6380 seq_printf(m, "%d\n", memcg->oom_group);
6385 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6386 char *buf, size_t nbytes, loff_t off)
6388 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6391 buf = strstrip(buf);
6395 ret = kstrtoint(buf, 0, &oom_group);
6399 if (oom_group != 0 && oom_group != 1)
6402 memcg->oom_group = oom_group;
6407 static struct cftype memory_files[] = {
6410 .flags = CFTYPE_NOT_ON_ROOT,
6411 .read_u64 = memory_current_read,
6415 .flags = CFTYPE_NOT_ON_ROOT,
6416 .seq_show = memory_min_show,
6417 .write = memory_min_write,
6421 .flags = CFTYPE_NOT_ON_ROOT,
6422 .seq_show = memory_low_show,
6423 .write = memory_low_write,
6427 .flags = CFTYPE_NOT_ON_ROOT,
6428 .seq_show = memory_high_show,
6429 .write = memory_high_write,
6433 .flags = CFTYPE_NOT_ON_ROOT,
6434 .seq_show = memory_max_show,
6435 .write = memory_max_write,
6439 .flags = CFTYPE_NOT_ON_ROOT,
6440 .file_offset = offsetof(struct mem_cgroup, events_file),
6441 .seq_show = memory_events_show,
6444 .name = "events.local",
6445 .flags = CFTYPE_NOT_ON_ROOT,
6446 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6447 .seq_show = memory_events_local_show,
6451 .seq_show = memory_stat_show,
6455 .name = "numa_stat",
6456 .seq_show = memory_numa_stat_show,
6460 .name = "oom.group",
6461 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6462 .seq_show = memory_oom_group_show,
6463 .write = memory_oom_group_write,
6468 struct cgroup_subsys memory_cgrp_subsys = {
6469 .css_alloc = mem_cgroup_css_alloc,
6470 .css_online = mem_cgroup_css_online,
6471 .css_offline = mem_cgroup_css_offline,
6472 .css_released = mem_cgroup_css_released,
6473 .css_free = mem_cgroup_css_free,
6474 .css_reset = mem_cgroup_css_reset,
6475 .css_rstat_flush = mem_cgroup_css_rstat_flush,
6476 .can_attach = mem_cgroup_can_attach,
6477 .cancel_attach = mem_cgroup_cancel_attach,
6478 .post_attach = mem_cgroup_move_task,
6479 .dfl_cftypes = memory_files,
6480 .legacy_cftypes = mem_cgroup_legacy_files,
6485 * This function calculates an individual cgroup's effective
6486 * protection which is derived from its own memory.min/low, its
6487 * parent's and siblings' settings, as well as the actual memory
6488 * distribution in the tree.
6490 * The following rules apply to the effective protection values:
6492 * 1. At the first level of reclaim, effective protection is equal to
6493 * the declared protection in memory.min and memory.low.
6495 * 2. To enable safe delegation of the protection configuration, at
6496 * subsequent levels the effective protection is capped to the
6497 * parent's effective protection.
6499 * 3. To make complex and dynamic subtrees easier to configure, the
6500 * user is allowed to overcommit the declared protection at a given
6501 * level. If that is the case, the parent's effective protection is
6502 * distributed to the children in proportion to how much protection
6503 * they have declared and how much of it they are utilizing.
6505 * This makes distribution proportional, but also work-conserving:
6506 * if one cgroup claims much more protection than it uses memory,
6507 * the unused remainder is available to its siblings.
6509 * 4. Conversely, when the declared protection is undercommitted at a
6510 * given level, the distribution of the larger parental protection
6511 * budget is NOT proportional. A cgroup's protection from a sibling
6512 * is capped to its own memory.min/low setting.
6514 * 5. However, to allow protecting recursive subtrees from each other
6515 * without having to declare each individual cgroup's fixed share
6516 * of the ancestor's claim to protection, any unutilized -
6517 * "floating" - protection from up the tree is distributed in
6518 * proportion to each cgroup's *usage*. This makes the protection
6519 * neutral wrt sibling cgroups and lets them compete freely over
6520 * the shared parental protection budget, but it protects the
6521 * subtree as a whole from neighboring subtrees.
6523 * Note that 4. and 5. are not in conflict: 4. is about protecting
6524 * against immediate siblings whereas 5. is about protecting against
6525 * neighboring subtrees.
6527 static unsigned long effective_protection(unsigned long usage,
6528 unsigned long parent_usage,
6529 unsigned long setting,
6530 unsigned long parent_effective,
6531 unsigned long siblings_protected)
6533 unsigned long protected;
6536 protected = min(usage, setting);
6538 * If all cgroups at this level combined claim and use more
6539 * protection then what the parent affords them, distribute
6540 * shares in proportion to utilization.
6542 * We are using actual utilization rather than the statically
6543 * claimed protection in order to be work-conserving: claimed
6544 * but unused protection is available to siblings that would
6545 * otherwise get a smaller chunk than what they claimed.
6547 if (siblings_protected > parent_effective)
6548 return protected * parent_effective / siblings_protected;
6551 * Ok, utilized protection of all children is within what the
6552 * parent affords them, so we know whatever this child claims
6553 * and utilizes is effectively protected.
6555 * If there is unprotected usage beyond this value, reclaim
6556 * will apply pressure in proportion to that amount.
6558 * If there is unutilized protection, the cgroup will be fully
6559 * shielded from reclaim, but we do return a smaller value for
6560 * protection than what the group could enjoy in theory. This
6561 * is okay. With the overcommit distribution above, effective
6562 * protection is always dependent on how memory is actually
6563 * consumed among the siblings anyway.
6568 * If the children aren't claiming (all of) the protection
6569 * afforded to them by the parent, distribute the remainder in
6570 * proportion to the (unprotected) memory of each cgroup. That
6571 * way, cgroups that aren't explicitly prioritized wrt each
6572 * other compete freely over the allowance, but they are
6573 * collectively protected from neighboring trees.
6575 * We're using unprotected memory for the weight so that if
6576 * some cgroups DO claim explicit protection, we don't protect
6577 * the same bytes twice.
6579 * Check both usage and parent_usage against the respective
6580 * protected values. One should imply the other, but they
6581 * aren't read atomically - make sure the division is sane.
6583 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6585 if (parent_effective > siblings_protected &&
6586 parent_usage > siblings_protected &&
6587 usage > protected) {
6588 unsigned long unclaimed;
6590 unclaimed = parent_effective - siblings_protected;
6591 unclaimed *= usage - protected;
6592 unclaimed /= parent_usage - siblings_protected;
6601 * mem_cgroup_calculate_protection - check if memory consumption is in the normal range
6602 * @root: the top ancestor of the sub-tree being checked
6603 * @memcg: the memory cgroup to check
6605 * WARNING: This function is not stateless! It can only be used as part
6606 * of a top-down tree iteration, not for isolated queries.
6608 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6609 struct mem_cgroup *memcg)
6611 unsigned long usage, parent_usage;
6612 struct mem_cgroup *parent;
6614 if (mem_cgroup_disabled())
6618 root = root_mem_cgroup;
6621 * Effective values of the reclaim targets are ignored so they
6622 * can be stale. Have a look at mem_cgroup_protection for more
6624 * TODO: calculation should be more robust so that we do not need
6625 * that special casing.
6630 usage = page_counter_read(&memcg->memory);
6634 parent = parent_mem_cgroup(memcg);
6635 /* No parent means a non-hierarchical mode on v1 memcg */
6639 if (parent == root) {
6640 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6641 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6645 parent_usage = page_counter_read(&parent->memory);
6647 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6648 READ_ONCE(memcg->memory.min),
6649 READ_ONCE(parent->memory.emin),
6650 atomic_long_read(&parent->memory.children_min_usage)));
6652 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6653 READ_ONCE(memcg->memory.low),
6654 READ_ONCE(parent->memory.elow),
6655 atomic_long_read(&parent->memory.children_low_usage)));
6658 static int charge_memcg(struct folio *folio, struct mem_cgroup *memcg,
6661 long nr_pages = folio_nr_pages(folio);
6664 ret = try_charge(memcg, gfp, nr_pages);
6668 css_get(&memcg->css);
6669 commit_charge(folio, memcg);
6671 local_irq_disable();
6672 mem_cgroup_charge_statistics(memcg, nr_pages);
6673 memcg_check_events(memcg, folio_nid(folio));
6679 int __mem_cgroup_charge(struct folio *folio, struct mm_struct *mm, gfp_t gfp)
6681 struct mem_cgroup *memcg;
6684 memcg = get_mem_cgroup_from_mm(mm);
6685 ret = charge_memcg(folio, memcg, gfp);
6686 css_put(&memcg->css);
6692 * mem_cgroup_swapin_charge_page - charge a newly allocated page for swapin
6693 * @page: page to charge
6694 * @mm: mm context of the victim
6695 * @gfp: reclaim mode
6696 * @entry: swap entry for which the page is allocated
6698 * This function charges a page allocated for swapin. Please call this before
6699 * adding the page to the swapcache.
6701 * Returns 0 on success. Otherwise, an error code is returned.
6703 int mem_cgroup_swapin_charge_page(struct page *page, struct mm_struct *mm,
6704 gfp_t gfp, swp_entry_t entry)
6706 struct folio *folio = page_folio(page);
6707 struct mem_cgroup *memcg;
6711 if (mem_cgroup_disabled())
6714 id = lookup_swap_cgroup_id(entry);
6716 memcg = mem_cgroup_from_id(id);
6717 if (!memcg || !css_tryget_online(&memcg->css))
6718 memcg = get_mem_cgroup_from_mm(mm);
6721 ret = charge_memcg(folio, memcg, gfp);
6723 css_put(&memcg->css);
6728 * mem_cgroup_swapin_uncharge_swap - uncharge swap slot
6729 * @entry: swap entry for which the page is charged
6731 * Call this function after successfully adding the charged page to swapcache.
6733 * Note: This function assumes the page for which swap slot is being uncharged
6736 void mem_cgroup_swapin_uncharge_swap(swp_entry_t entry)
6739 * Cgroup1's unified memory+swap counter has been charged with the
6740 * new swapcache page, finish the transfer by uncharging the swap
6741 * slot. The swap slot would also get uncharged when it dies, but
6742 * it can stick around indefinitely and we'd count the page twice
6745 * Cgroup2 has separate resource counters for memory and swap,
6746 * so this is a non-issue here. Memory and swap charge lifetimes
6747 * correspond 1:1 to page and swap slot lifetimes: we charge the
6748 * page to memory here, and uncharge swap when the slot is freed.
6750 if (!mem_cgroup_disabled() && do_memsw_account()) {
6752 * The swap entry might not get freed for a long time,
6753 * let's not wait for it. The page already received a
6754 * memory+swap charge, drop the swap entry duplicate.
6756 mem_cgroup_uncharge_swap(entry, 1);
6760 struct uncharge_gather {
6761 struct mem_cgroup *memcg;
6762 unsigned long nr_memory;
6763 unsigned long pgpgout;
6764 unsigned long nr_kmem;
6768 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6770 memset(ug, 0, sizeof(*ug));
6773 static void uncharge_batch(const struct uncharge_gather *ug)
6775 unsigned long flags;
6777 if (ug->nr_memory) {
6778 page_counter_uncharge(&ug->memcg->memory, ug->nr_memory);
6779 if (do_memsw_account())
6780 page_counter_uncharge(&ug->memcg->memsw, ug->nr_memory);
6782 memcg_account_kmem(ug->memcg, -ug->nr_kmem);
6783 memcg_oom_recover(ug->memcg);
6786 local_irq_save(flags);
6787 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6788 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_memory);
6789 memcg_check_events(ug->memcg, ug->nid);
6790 local_irq_restore(flags);
6792 /* drop reference from uncharge_folio */
6793 css_put(&ug->memcg->css);
6796 static void uncharge_folio(struct folio *folio, struct uncharge_gather *ug)
6799 struct mem_cgroup *memcg;
6800 struct obj_cgroup *objcg;
6802 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
6805 * Nobody should be changing or seriously looking at
6806 * folio memcg or objcg at this point, we have fully
6807 * exclusive access to the folio.
6809 if (folio_memcg_kmem(folio)) {
6810 objcg = __folio_objcg(folio);
6812 * This get matches the put at the end of the function and
6813 * kmem pages do not hold memcg references anymore.
6815 memcg = get_mem_cgroup_from_objcg(objcg);
6817 memcg = __folio_memcg(folio);
6823 if (ug->memcg != memcg) {
6826 uncharge_gather_clear(ug);
6829 ug->nid = folio_nid(folio);
6831 /* pairs with css_put in uncharge_batch */
6832 css_get(&memcg->css);
6835 nr_pages = folio_nr_pages(folio);
6837 if (folio_memcg_kmem(folio)) {
6838 ug->nr_memory += nr_pages;
6839 ug->nr_kmem += nr_pages;
6841 folio->memcg_data = 0;
6842 obj_cgroup_put(objcg);
6844 /* LRU pages aren't accounted at the root level */
6845 if (!mem_cgroup_is_root(memcg))
6846 ug->nr_memory += nr_pages;
6849 folio->memcg_data = 0;
6852 css_put(&memcg->css);
6855 void __mem_cgroup_uncharge(struct folio *folio)
6857 struct uncharge_gather ug;
6859 /* Don't touch folio->lru of any random page, pre-check: */
6860 if (!folio_memcg(folio))
6863 uncharge_gather_clear(&ug);
6864 uncharge_folio(folio, &ug);
6865 uncharge_batch(&ug);
6869 * __mem_cgroup_uncharge_list - uncharge a list of page
6870 * @page_list: list of pages to uncharge
6872 * Uncharge a list of pages previously charged with
6873 * __mem_cgroup_charge().
6875 void __mem_cgroup_uncharge_list(struct list_head *page_list)
6877 struct uncharge_gather ug;
6878 struct folio *folio;
6880 uncharge_gather_clear(&ug);
6881 list_for_each_entry(folio, page_list, lru)
6882 uncharge_folio(folio, &ug);
6884 uncharge_batch(&ug);
6888 * mem_cgroup_migrate - Charge a folio's replacement.
6889 * @old: Currently circulating folio.
6890 * @new: Replacement folio.
6892 * Charge @new as a replacement folio for @old. @old will
6893 * be uncharged upon free.
6895 * Both folios must be locked, @new->mapping must be set up.
6897 void mem_cgroup_migrate(struct folio *old, struct folio *new)
6899 struct mem_cgroup *memcg;
6900 long nr_pages = folio_nr_pages(new);
6901 unsigned long flags;
6903 VM_BUG_ON_FOLIO(!folio_test_locked(old), old);
6904 VM_BUG_ON_FOLIO(!folio_test_locked(new), new);
6905 VM_BUG_ON_FOLIO(folio_test_anon(old) != folio_test_anon(new), new);
6906 VM_BUG_ON_FOLIO(folio_nr_pages(old) != nr_pages, new);
6908 if (mem_cgroup_disabled())
6911 /* Page cache replacement: new folio already charged? */
6912 if (folio_memcg(new))
6915 memcg = folio_memcg(old);
6916 VM_WARN_ON_ONCE_FOLIO(!memcg, old);
6920 /* Force-charge the new page. The old one will be freed soon */
6921 if (!mem_cgroup_is_root(memcg)) {
6922 page_counter_charge(&memcg->memory, nr_pages);
6923 if (do_memsw_account())
6924 page_counter_charge(&memcg->memsw, nr_pages);
6927 css_get(&memcg->css);
6928 commit_charge(new, memcg);
6930 local_irq_save(flags);
6931 mem_cgroup_charge_statistics(memcg, nr_pages);
6932 memcg_check_events(memcg, folio_nid(new));
6933 local_irq_restore(flags);
6936 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6937 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6939 void mem_cgroup_sk_alloc(struct sock *sk)
6941 struct mem_cgroup *memcg;
6943 if (!mem_cgroup_sockets_enabled)
6946 /* Do not associate the sock with unrelated interrupted task's memcg. */
6951 memcg = mem_cgroup_from_task(current);
6952 if (memcg == root_mem_cgroup)
6954 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6956 if (css_tryget(&memcg->css))
6957 sk->sk_memcg = memcg;
6962 void mem_cgroup_sk_free(struct sock *sk)
6965 css_put(&sk->sk_memcg->css);
6969 * mem_cgroup_charge_skmem - charge socket memory
6970 * @memcg: memcg to charge
6971 * @nr_pages: number of pages to charge
6972 * @gfp_mask: reclaim mode
6974 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6975 * @memcg's configured limit, %false if it doesn't.
6977 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages,
6980 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6981 struct page_counter *fail;
6983 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6984 memcg->tcpmem_pressure = 0;
6987 memcg->tcpmem_pressure = 1;
6988 if (gfp_mask & __GFP_NOFAIL) {
6989 page_counter_charge(&memcg->tcpmem, nr_pages);
6995 if (try_charge(memcg, gfp_mask, nr_pages) == 0) {
6996 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7004 * mem_cgroup_uncharge_skmem - uncharge socket memory
7005 * @memcg: memcg to uncharge
7006 * @nr_pages: number of pages to uncharge
7008 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7010 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7011 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7015 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7017 refill_stock(memcg, nr_pages);
7020 static int __init cgroup_memory(char *s)
7024 while ((token = strsep(&s, ",")) != NULL) {
7027 if (!strcmp(token, "nosocket"))
7028 cgroup_memory_nosocket = true;
7029 if (!strcmp(token, "nokmem"))
7030 cgroup_memory_nokmem = true;
7034 __setup("cgroup.memory=", cgroup_memory);
7037 * subsys_initcall() for memory controller.
7039 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7040 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7041 * basically everything that doesn't depend on a specific mem_cgroup structure
7042 * should be initialized from here.
7044 static int __init mem_cgroup_init(void)
7049 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7050 * used for per-memcg-per-cpu caching of per-node statistics. In order
7051 * to work fine, we should make sure that the overfill threshold can't
7052 * exceed S32_MAX / PAGE_SIZE.
7054 BUILD_BUG_ON(MEMCG_CHARGE_BATCH > S32_MAX / PAGE_SIZE);
7056 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7057 memcg_hotplug_cpu_dead);
7059 for_each_possible_cpu(cpu)
7060 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7063 for_each_node(node) {
7064 struct mem_cgroup_tree_per_node *rtpn;
7066 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7067 node_online(node) ? node : NUMA_NO_NODE);
7069 rtpn->rb_root = RB_ROOT;
7070 rtpn->rb_rightmost = NULL;
7071 spin_lock_init(&rtpn->lock);
7072 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7077 subsys_initcall(mem_cgroup_init);
7079 #ifdef CONFIG_MEMCG_SWAP
7080 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7082 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7084 * The root cgroup cannot be destroyed, so it's refcount must
7087 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7091 memcg = parent_mem_cgroup(memcg);
7093 memcg = root_mem_cgroup;
7099 * mem_cgroup_swapout - transfer a memsw charge to swap
7100 * @page: page whose memsw charge to transfer
7101 * @entry: swap entry to move the charge to
7103 * Transfer the memsw charge of @page to @entry.
7105 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7107 struct mem_cgroup *memcg, *swap_memcg;
7108 unsigned int nr_entries;
7109 unsigned short oldid;
7111 VM_BUG_ON_PAGE(PageLRU(page), page);
7112 VM_BUG_ON_PAGE(page_count(page), page);
7114 if (mem_cgroup_disabled())
7117 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7120 memcg = page_memcg(page);
7122 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7127 * In case the memcg owning these pages has been offlined and doesn't
7128 * have an ID allocated to it anymore, charge the closest online
7129 * ancestor for the swap instead and transfer the memory+swap charge.
7131 swap_memcg = mem_cgroup_id_get_online(memcg);
7132 nr_entries = thp_nr_pages(page);
7133 /* Get references for the tail pages, too */
7135 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7136 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7138 VM_BUG_ON_PAGE(oldid, page);
7139 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7141 page->memcg_data = 0;
7143 if (!mem_cgroup_is_root(memcg))
7144 page_counter_uncharge(&memcg->memory, nr_entries);
7146 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7147 if (!mem_cgroup_is_root(swap_memcg))
7148 page_counter_charge(&swap_memcg->memsw, nr_entries);
7149 page_counter_uncharge(&memcg->memsw, nr_entries);
7153 * Interrupts should be disabled here because the caller holds the
7154 * i_pages lock which is taken with interrupts-off. It is
7155 * important here to have the interrupts disabled because it is the
7156 * only synchronisation we have for updating the per-CPU variables.
7159 mem_cgroup_charge_statistics(memcg, -nr_entries);
7160 memcg_stats_unlock();
7161 memcg_check_events(memcg, page_to_nid(page));
7163 css_put(&memcg->css);
7167 * __mem_cgroup_try_charge_swap - try charging swap space for a page
7168 * @page: page being added to swap
7169 * @entry: swap entry to charge
7171 * Try to charge @page's memcg for the swap space at @entry.
7173 * Returns 0 on success, -ENOMEM on failure.
7175 int __mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7177 unsigned int nr_pages = thp_nr_pages(page);
7178 struct page_counter *counter;
7179 struct mem_cgroup *memcg;
7180 unsigned short oldid;
7182 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7185 memcg = page_memcg(page);
7187 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7192 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7196 memcg = mem_cgroup_id_get_online(memcg);
7198 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7199 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7200 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7201 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7202 mem_cgroup_id_put(memcg);
7206 /* Get references for the tail pages, too */
7208 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7209 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7210 VM_BUG_ON_PAGE(oldid, page);
7211 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7217 * __mem_cgroup_uncharge_swap - uncharge swap space
7218 * @entry: swap entry to uncharge
7219 * @nr_pages: the amount of swap space to uncharge
7221 void __mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7223 struct mem_cgroup *memcg;
7226 id = swap_cgroup_record(entry, 0, nr_pages);
7228 memcg = mem_cgroup_from_id(id);
7230 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7231 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7232 page_counter_uncharge(&memcg->swap, nr_pages);
7234 page_counter_uncharge(&memcg->memsw, nr_pages);
7236 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7237 mem_cgroup_id_put_many(memcg, nr_pages);
7242 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7244 long nr_swap_pages = get_nr_swap_pages();
7246 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7247 return nr_swap_pages;
7248 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7249 nr_swap_pages = min_t(long, nr_swap_pages,
7250 READ_ONCE(memcg->swap.max) -
7251 page_counter_read(&memcg->swap));
7252 return nr_swap_pages;
7255 bool mem_cgroup_swap_full(struct page *page)
7257 struct mem_cgroup *memcg;
7259 VM_BUG_ON_PAGE(!PageLocked(page), page);
7263 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7266 memcg = page_memcg(page);
7270 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7271 unsigned long usage = page_counter_read(&memcg->swap);
7273 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7274 usage * 2 >= READ_ONCE(memcg->swap.max))
7281 static int __init setup_swap_account(char *s)
7283 if (!strcmp(s, "1"))
7284 cgroup_memory_noswap = false;
7285 else if (!strcmp(s, "0"))
7286 cgroup_memory_noswap = true;
7289 __setup("swapaccount=", setup_swap_account);
7291 static u64 swap_current_read(struct cgroup_subsys_state *css,
7294 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7296 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7299 static int swap_high_show(struct seq_file *m, void *v)
7301 return seq_puts_memcg_tunable(m,
7302 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7305 static ssize_t swap_high_write(struct kernfs_open_file *of,
7306 char *buf, size_t nbytes, loff_t off)
7308 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7312 buf = strstrip(buf);
7313 err = page_counter_memparse(buf, "max", &high);
7317 page_counter_set_high(&memcg->swap, high);
7322 static int swap_max_show(struct seq_file *m, void *v)
7324 return seq_puts_memcg_tunable(m,
7325 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7328 static ssize_t swap_max_write(struct kernfs_open_file *of,
7329 char *buf, size_t nbytes, loff_t off)
7331 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7335 buf = strstrip(buf);
7336 err = page_counter_memparse(buf, "max", &max);
7340 xchg(&memcg->swap.max, max);
7345 static int swap_events_show(struct seq_file *m, void *v)
7347 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7349 seq_printf(m, "high %lu\n",
7350 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7351 seq_printf(m, "max %lu\n",
7352 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7353 seq_printf(m, "fail %lu\n",
7354 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7359 static struct cftype swap_files[] = {
7361 .name = "swap.current",
7362 .flags = CFTYPE_NOT_ON_ROOT,
7363 .read_u64 = swap_current_read,
7366 .name = "swap.high",
7367 .flags = CFTYPE_NOT_ON_ROOT,
7368 .seq_show = swap_high_show,
7369 .write = swap_high_write,
7373 .flags = CFTYPE_NOT_ON_ROOT,
7374 .seq_show = swap_max_show,
7375 .write = swap_max_write,
7378 .name = "swap.events",
7379 .flags = CFTYPE_NOT_ON_ROOT,
7380 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7381 .seq_show = swap_events_show,
7386 static struct cftype memsw_files[] = {
7388 .name = "memsw.usage_in_bytes",
7389 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7390 .read_u64 = mem_cgroup_read_u64,
7393 .name = "memsw.max_usage_in_bytes",
7394 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7395 .write = mem_cgroup_reset,
7396 .read_u64 = mem_cgroup_read_u64,
7399 .name = "memsw.limit_in_bytes",
7400 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7401 .write = mem_cgroup_write,
7402 .read_u64 = mem_cgroup_read_u64,
7405 .name = "memsw.failcnt",
7406 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7407 .write = mem_cgroup_reset,
7408 .read_u64 = mem_cgroup_read_u64,
7410 { }, /* terminate */
7414 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7415 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7416 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7417 * boot parameter. This may result in premature OOPS inside
7418 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7420 static int __init mem_cgroup_swap_init(void)
7422 /* No memory control -> no swap control */
7423 if (mem_cgroup_disabled())
7424 cgroup_memory_noswap = true;
7426 if (cgroup_memory_noswap)
7429 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7430 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7434 core_initcall(mem_cgroup_swap_init);
7436 #endif /* CONFIG_MEMCG_SWAP */