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/memremap.h>
57 #include <linux/mm_inline.h>
58 #include <linux/swap_cgroup.h>
59 #include <linux/cpu.h>
60 #include <linux/oom.h>
61 #include <linux/lockdep.h>
62 #include <linux/file.h>
63 #include <linux/resume_user_mode.h>
64 #include <linux/psi.h>
65 #include <linux/seq_buf.h>
72 #include <linux/uaccess.h>
74 #include <trace/events/vmscan.h>
76 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
77 EXPORT_SYMBOL(memory_cgrp_subsys);
79 struct mem_cgroup *root_mem_cgroup __read_mostly;
81 /* Active memory cgroup to use from an interrupt context */
82 DEFINE_PER_CPU(struct mem_cgroup *, int_active_memcg);
83 EXPORT_PER_CPU_SYMBOL_GPL(int_active_memcg);
85 /* Socket memory accounting disabled? */
86 static bool cgroup_memory_nosocket __ro_after_init;
88 /* Kernel memory accounting disabled? */
89 static bool cgroup_memory_nokmem __ro_after_init;
91 /* BPF memory accounting disabled? */
92 static bool cgroup_memory_nobpf __ro_after_init;
94 #ifdef CONFIG_CGROUP_WRITEBACK
95 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
98 /* Whether legacy memory+swap accounting is active */
99 static bool do_memsw_account(void)
101 return !cgroup_subsys_on_dfl(memory_cgrp_subsys);
104 #define THRESHOLDS_EVENTS_TARGET 128
105 #define SOFTLIMIT_EVENTS_TARGET 1024
108 * Cgroups above their limits are maintained in a RB-Tree, independent of
109 * their hierarchy representation
112 struct mem_cgroup_tree_per_node {
113 struct rb_root rb_root;
114 struct rb_node *rb_rightmost;
118 struct mem_cgroup_tree {
119 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
122 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
125 struct mem_cgroup_eventfd_list {
126 struct list_head list;
127 struct eventfd_ctx *eventfd;
131 * cgroup_event represents events which userspace want to receive.
133 struct mem_cgroup_event {
135 * memcg which the event belongs to.
137 struct mem_cgroup *memcg;
139 * eventfd to signal userspace about the event.
141 struct eventfd_ctx *eventfd;
143 * Each of these stored in a list by the cgroup.
145 struct list_head list;
147 * register_event() callback will be used to add new userspace
148 * waiter for changes related to this event. Use eventfd_signal()
149 * on eventfd to send notification to userspace.
151 int (*register_event)(struct mem_cgroup *memcg,
152 struct eventfd_ctx *eventfd, const char *args);
154 * unregister_event() callback will be called when userspace closes
155 * the eventfd or on cgroup removing. This callback must be set,
156 * if you want provide notification functionality.
158 void (*unregister_event)(struct mem_cgroup *memcg,
159 struct eventfd_ctx *eventfd);
161 * All fields below needed to unregister event when
162 * userspace closes eventfd.
165 wait_queue_head_t *wqh;
166 wait_queue_entry_t wait;
167 struct work_struct remove;
170 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
171 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
173 /* Stuffs for move charges at task migration. */
175 * Types of charges to be moved.
177 #define MOVE_ANON 0x1U
178 #define MOVE_FILE 0x2U
179 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
181 /* "mc" and its members are protected by cgroup_mutex */
182 static struct move_charge_struct {
183 spinlock_t lock; /* for from, to */
184 struct mm_struct *mm;
185 struct mem_cgroup *from;
186 struct mem_cgroup *to;
188 unsigned long precharge;
189 unsigned long moved_charge;
190 unsigned long moved_swap;
191 struct task_struct *moving_task; /* a task moving charges */
192 wait_queue_head_t waitq; /* a waitq for other context */
194 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
195 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
199 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
200 * limit reclaim to prevent infinite loops, if they ever occur.
202 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
203 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
205 /* for encoding cft->private value on file */
213 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
214 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
215 #define MEMFILE_ATTR(val) ((val) & 0xffff)
218 * Iteration constructs for visiting all cgroups (under a tree). If
219 * loops are exited prematurely (break), mem_cgroup_iter_break() must
220 * be used for reference counting.
222 #define for_each_mem_cgroup_tree(iter, root) \
223 for (iter = mem_cgroup_iter(root, NULL, NULL); \
225 iter = mem_cgroup_iter(root, iter, NULL))
227 #define for_each_mem_cgroup(iter) \
228 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
230 iter = mem_cgroup_iter(NULL, iter, NULL))
232 static inline bool task_is_dying(void)
234 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
235 (current->flags & PF_EXITING);
238 /* Some nice accessors for the vmpressure. */
239 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
242 memcg = root_mem_cgroup;
243 return &memcg->vmpressure;
246 struct mem_cgroup *vmpressure_to_memcg(struct vmpressure *vmpr)
248 return container_of(vmpr, struct mem_cgroup, vmpressure);
251 #ifdef CONFIG_MEMCG_KMEM
252 static DEFINE_SPINLOCK(objcg_lock);
254 bool mem_cgroup_kmem_disabled(void)
256 return cgroup_memory_nokmem;
259 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
260 unsigned int nr_pages);
262 static void obj_cgroup_release(struct percpu_ref *ref)
264 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
265 unsigned int nr_bytes;
266 unsigned int nr_pages;
270 * At this point all allocated objects are freed, and
271 * objcg->nr_charged_bytes can't have an arbitrary byte value.
272 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
274 * The following sequence can lead to it:
275 * 1) CPU0: objcg == stock->cached_objcg
276 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
277 * PAGE_SIZE bytes are charged
278 * 3) CPU1: a process from another memcg is allocating something,
279 * the stock if flushed,
280 * objcg->nr_charged_bytes = PAGE_SIZE - 92
281 * 5) CPU0: we do release this object,
282 * 92 bytes are added to stock->nr_bytes
283 * 6) CPU0: stock is flushed,
284 * 92 bytes are added to objcg->nr_charged_bytes
286 * In the result, nr_charged_bytes == PAGE_SIZE.
287 * This page will be uncharged in obj_cgroup_release().
289 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
290 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
291 nr_pages = nr_bytes >> PAGE_SHIFT;
294 obj_cgroup_uncharge_pages(objcg, nr_pages);
296 spin_lock_irqsave(&objcg_lock, flags);
297 list_del(&objcg->list);
298 spin_unlock_irqrestore(&objcg_lock, flags);
300 percpu_ref_exit(ref);
301 kfree_rcu(objcg, rcu);
304 static struct obj_cgroup *obj_cgroup_alloc(void)
306 struct obj_cgroup *objcg;
309 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
313 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
319 INIT_LIST_HEAD(&objcg->list);
323 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
324 struct mem_cgroup *parent)
326 struct obj_cgroup *objcg, *iter;
328 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
330 spin_lock_irq(&objcg_lock);
332 /* 1) Ready to reparent active objcg. */
333 list_add(&objcg->list, &memcg->objcg_list);
334 /* 2) Reparent active objcg and already reparented objcgs to parent. */
335 list_for_each_entry(iter, &memcg->objcg_list, list)
336 WRITE_ONCE(iter->memcg, parent);
337 /* 3) Move already reparented objcgs to the parent's list */
338 list_splice(&memcg->objcg_list, &parent->objcg_list);
340 spin_unlock_irq(&objcg_lock);
342 percpu_ref_kill(&objcg->refcnt);
346 * A lot of the calls to the cache allocation functions are expected to be
347 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
348 * conditional to this static branch, we'll have to allow modules that does
349 * kmem_cache_alloc and the such to see this symbol as well
351 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
352 EXPORT_SYMBOL(memcg_kmem_enabled_key);
354 DEFINE_STATIC_KEY_FALSE(memcg_bpf_enabled_key);
355 EXPORT_SYMBOL(memcg_bpf_enabled_key);
359 * mem_cgroup_css_from_page - css of the memcg associated with a page
360 * @page: page of interest
362 * If memcg is bound to the default hierarchy, css of the memcg associated
363 * with @page is returned. The returned css remains associated with @page
364 * until it is released.
366 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
369 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
371 struct mem_cgroup *memcg;
373 memcg = page_memcg(page);
375 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
376 memcg = root_mem_cgroup;
382 * page_cgroup_ino - return inode number of the memcg a page is charged to
385 * Look up the closest online ancestor of the memory cgroup @page is charged to
386 * and return its inode number or 0 if @page is not charged to any cgroup. It
387 * is safe to call this function without holding a reference to @page.
389 * Note, this function is inherently racy, because there is nothing to prevent
390 * the cgroup inode from getting torn down and potentially reallocated a moment
391 * after page_cgroup_ino() returns, so it only should be used by callers that
392 * do not care (such as procfs interfaces).
394 ino_t page_cgroup_ino(struct page *page)
396 struct mem_cgroup *memcg;
397 unsigned long ino = 0;
400 memcg = page_memcg_check(page);
402 while (memcg && !(memcg->css.flags & CSS_ONLINE))
403 memcg = parent_mem_cgroup(memcg);
405 ino = cgroup_ino(memcg->css.cgroup);
410 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
411 struct mem_cgroup_tree_per_node *mctz,
412 unsigned long new_usage_in_excess)
414 struct rb_node **p = &mctz->rb_root.rb_node;
415 struct rb_node *parent = NULL;
416 struct mem_cgroup_per_node *mz_node;
417 bool rightmost = true;
422 mz->usage_in_excess = new_usage_in_excess;
423 if (!mz->usage_in_excess)
427 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
429 if (mz->usage_in_excess < mz_node->usage_in_excess) {
438 mctz->rb_rightmost = &mz->tree_node;
440 rb_link_node(&mz->tree_node, parent, p);
441 rb_insert_color(&mz->tree_node, &mctz->rb_root);
445 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
446 struct mem_cgroup_tree_per_node *mctz)
451 if (&mz->tree_node == mctz->rb_rightmost)
452 mctz->rb_rightmost = rb_prev(&mz->tree_node);
454 rb_erase(&mz->tree_node, &mctz->rb_root);
458 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
459 struct mem_cgroup_tree_per_node *mctz)
463 spin_lock_irqsave(&mctz->lock, flags);
464 __mem_cgroup_remove_exceeded(mz, mctz);
465 spin_unlock_irqrestore(&mctz->lock, flags);
468 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
470 unsigned long nr_pages = page_counter_read(&memcg->memory);
471 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
472 unsigned long excess = 0;
474 if (nr_pages > soft_limit)
475 excess = nr_pages - soft_limit;
480 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, int nid)
482 unsigned long excess;
483 struct mem_cgroup_per_node *mz;
484 struct mem_cgroup_tree_per_node *mctz;
486 mctz = soft_limit_tree.rb_tree_per_node[nid];
490 * Necessary to update all ancestors when hierarchy is used.
491 * because their event counter is not touched.
493 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
494 mz = memcg->nodeinfo[nid];
495 excess = soft_limit_excess(memcg);
497 * We have to update the tree if mz is on RB-tree or
498 * mem is over its softlimit.
500 if (excess || mz->on_tree) {
503 spin_lock_irqsave(&mctz->lock, flags);
504 /* if on-tree, remove it */
506 __mem_cgroup_remove_exceeded(mz, mctz);
508 * Insert again. mz->usage_in_excess will be updated.
509 * If excess is 0, no tree ops.
511 __mem_cgroup_insert_exceeded(mz, mctz, excess);
512 spin_unlock_irqrestore(&mctz->lock, flags);
517 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
519 struct mem_cgroup_tree_per_node *mctz;
520 struct mem_cgroup_per_node *mz;
524 mz = memcg->nodeinfo[nid];
525 mctz = soft_limit_tree.rb_tree_per_node[nid];
527 mem_cgroup_remove_exceeded(mz, mctz);
531 static struct mem_cgroup_per_node *
532 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
534 struct mem_cgroup_per_node *mz;
538 if (!mctz->rb_rightmost)
539 goto done; /* Nothing to reclaim from */
541 mz = rb_entry(mctz->rb_rightmost,
542 struct mem_cgroup_per_node, tree_node);
544 * Remove the node now but someone else can add it back,
545 * we will to add it back at the end of reclaim to its correct
546 * position in the tree.
548 __mem_cgroup_remove_exceeded(mz, mctz);
549 if (!soft_limit_excess(mz->memcg) ||
550 !css_tryget(&mz->memcg->css))
556 static struct mem_cgroup_per_node *
557 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
559 struct mem_cgroup_per_node *mz;
561 spin_lock_irq(&mctz->lock);
562 mz = __mem_cgroup_largest_soft_limit_node(mctz);
563 spin_unlock_irq(&mctz->lock);
568 * memcg and lruvec stats flushing
570 * Many codepaths leading to stats update or read are performance sensitive and
571 * adding stats flushing in such codepaths is not desirable. So, to optimize the
572 * flushing the kernel does:
574 * 1) Periodically and asynchronously flush the stats every 2 seconds to not let
575 * rstat update tree grow unbounded.
577 * 2) Flush the stats synchronously on reader side only when there are more than
578 * (MEMCG_CHARGE_BATCH * nr_cpus) update events. Though this optimization
579 * will let stats be out of sync by atmost (MEMCG_CHARGE_BATCH * nr_cpus) but
580 * only for 2 seconds due to (1).
582 static void flush_memcg_stats_dwork(struct work_struct *w);
583 static DECLARE_DEFERRABLE_WORK(stats_flush_dwork, flush_memcg_stats_dwork);
584 static DEFINE_SPINLOCK(stats_flush_lock);
585 static DEFINE_PER_CPU(unsigned int, stats_updates);
586 static atomic_t stats_flush_threshold = ATOMIC_INIT(0);
587 static u64 flush_next_time;
589 #define FLUSH_TIME (2UL*HZ)
592 * Accessors to ensure that preemption is disabled on PREEMPT_RT because it can
593 * not rely on this as part of an acquired spinlock_t lock. These functions are
594 * never used in hardirq context on PREEMPT_RT and therefore disabling preemtion
597 static void memcg_stats_lock(void)
599 preempt_disable_nested();
600 VM_WARN_ON_IRQS_ENABLED();
603 static void __memcg_stats_lock(void)
605 preempt_disable_nested();
608 static void memcg_stats_unlock(void)
610 preempt_enable_nested();
613 static inline void memcg_rstat_updated(struct mem_cgroup *memcg, int val)
617 cgroup_rstat_updated(memcg->css.cgroup, smp_processor_id());
619 x = __this_cpu_add_return(stats_updates, abs(val));
620 if (x > MEMCG_CHARGE_BATCH) {
622 * If stats_flush_threshold exceeds the threshold
623 * (>num_online_cpus()), cgroup stats update will be triggered
624 * in __mem_cgroup_flush_stats(). Increasing this var further
625 * is redundant and simply adds overhead in atomic update.
627 if (atomic_read(&stats_flush_threshold) <= num_online_cpus())
628 atomic_add(x / MEMCG_CHARGE_BATCH, &stats_flush_threshold);
629 __this_cpu_write(stats_updates, 0);
633 static void __mem_cgroup_flush_stats(void)
637 if (!spin_trylock_irqsave(&stats_flush_lock, flag))
640 flush_next_time = jiffies_64 + 2*FLUSH_TIME;
641 cgroup_rstat_flush_irqsafe(root_mem_cgroup->css.cgroup);
642 atomic_set(&stats_flush_threshold, 0);
643 spin_unlock_irqrestore(&stats_flush_lock, flag);
646 void mem_cgroup_flush_stats(void)
648 if (atomic_read(&stats_flush_threshold) > num_online_cpus())
649 __mem_cgroup_flush_stats();
652 void mem_cgroup_flush_stats_delayed(void)
654 if (time_after64(jiffies_64, flush_next_time))
655 mem_cgroup_flush_stats();
658 static void flush_memcg_stats_dwork(struct work_struct *w)
660 __mem_cgroup_flush_stats();
661 queue_delayed_work(system_unbound_wq, &stats_flush_dwork, FLUSH_TIME);
664 /* Subset of vm_event_item to report for memcg event stats */
665 static const unsigned int memcg_vm_event_stat[] = {
681 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
685 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
691 #define NR_MEMCG_EVENTS ARRAY_SIZE(memcg_vm_event_stat)
692 static int mem_cgroup_events_index[NR_VM_EVENT_ITEMS] __read_mostly;
694 static void init_memcg_events(void)
698 for (i = 0; i < NR_MEMCG_EVENTS; ++i)
699 mem_cgroup_events_index[memcg_vm_event_stat[i]] = i + 1;
702 static inline int memcg_events_index(enum vm_event_item idx)
704 return mem_cgroup_events_index[idx] - 1;
707 struct memcg_vmstats_percpu {
708 /* Local (CPU and cgroup) page state & events */
709 long state[MEMCG_NR_STAT];
710 unsigned long events[NR_MEMCG_EVENTS];
712 /* Delta calculation for lockless upward propagation */
713 long state_prev[MEMCG_NR_STAT];
714 unsigned long events_prev[NR_MEMCG_EVENTS];
716 /* Cgroup1: threshold notifications & softlimit tree updates */
717 unsigned long nr_page_events;
718 unsigned long targets[MEM_CGROUP_NTARGETS];
721 struct memcg_vmstats {
722 /* Aggregated (CPU and subtree) page state & events */
723 long state[MEMCG_NR_STAT];
724 unsigned long events[NR_MEMCG_EVENTS];
726 /* Pending child counts during tree propagation */
727 long state_pending[MEMCG_NR_STAT];
728 unsigned long events_pending[NR_MEMCG_EVENTS];
731 unsigned long memcg_page_state(struct mem_cgroup *memcg, int idx)
733 long x = READ_ONCE(memcg->vmstats->state[idx]);
742 * __mod_memcg_state - update cgroup memory statistics
743 * @memcg: the memory cgroup
744 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
745 * @val: delta to add to the counter, can be negative
747 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
749 if (mem_cgroup_disabled())
752 __this_cpu_add(memcg->vmstats_percpu->state[idx], val);
753 memcg_rstat_updated(memcg, val);
756 /* idx can be of type enum memcg_stat_item or node_stat_item. */
757 static unsigned long memcg_page_state_local(struct mem_cgroup *memcg, int idx)
762 for_each_possible_cpu(cpu)
763 x += per_cpu(memcg->vmstats_percpu->state[idx], cpu);
771 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
774 struct mem_cgroup_per_node *pn;
775 struct mem_cgroup *memcg;
777 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
781 * The caller from rmap relay on disabled preemption becase they never
782 * update their counter from in-interrupt context. For these two
783 * counters we check that the update is never performed from an
784 * interrupt context while other caller need to have disabled interrupt.
786 __memcg_stats_lock();
787 if (IS_ENABLED(CONFIG_DEBUG_VM)) {
792 case NR_SHMEM_PMDMAPPED:
793 case NR_FILE_PMDMAPPED:
794 WARN_ON_ONCE(!in_task());
797 VM_WARN_ON_IRQS_ENABLED();
802 __this_cpu_add(memcg->vmstats_percpu->state[idx], val);
805 __this_cpu_add(pn->lruvec_stats_percpu->state[idx], val);
807 memcg_rstat_updated(memcg, val);
808 memcg_stats_unlock();
812 * __mod_lruvec_state - update lruvec memory statistics
813 * @lruvec: the lruvec
814 * @idx: the stat item
815 * @val: delta to add to the counter, can be negative
817 * The lruvec is the intersection of the NUMA node and a cgroup. This
818 * function updates the all three counters that are affected by a
819 * change of state at this level: per-node, per-cgroup, per-lruvec.
821 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
825 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
827 /* Update memcg and lruvec */
828 if (!mem_cgroup_disabled())
829 __mod_memcg_lruvec_state(lruvec, idx, val);
832 void __mod_lruvec_page_state(struct page *page, enum node_stat_item idx,
835 struct page *head = compound_head(page); /* rmap on tail pages */
836 struct mem_cgroup *memcg;
837 pg_data_t *pgdat = page_pgdat(page);
838 struct lruvec *lruvec;
841 memcg = page_memcg(head);
842 /* Untracked pages have no memcg, no lruvec. Update only the node */
845 __mod_node_page_state(pgdat, idx, val);
849 lruvec = mem_cgroup_lruvec(memcg, pgdat);
850 __mod_lruvec_state(lruvec, idx, val);
853 EXPORT_SYMBOL(__mod_lruvec_page_state);
855 void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val)
857 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
858 struct mem_cgroup *memcg;
859 struct lruvec *lruvec;
862 memcg = mem_cgroup_from_slab_obj(p);
865 * Untracked pages have no memcg, no lruvec. Update only the
866 * node. If we reparent the slab objects to the root memcg,
867 * when we free the slab object, we need to update the per-memcg
868 * vmstats to keep it correct for the root memcg.
871 __mod_node_page_state(pgdat, idx, val);
873 lruvec = mem_cgroup_lruvec(memcg, pgdat);
874 __mod_lruvec_state(lruvec, idx, val);
880 * __count_memcg_events - account VM events in a cgroup
881 * @memcg: the memory cgroup
882 * @idx: the event item
883 * @count: the number of events that occurred
885 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
888 int index = memcg_events_index(idx);
890 if (mem_cgroup_disabled() || index < 0)
894 __this_cpu_add(memcg->vmstats_percpu->events[index], count);
895 memcg_rstat_updated(memcg, count);
896 memcg_stats_unlock();
899 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
901 int index = memcg_events_index(event);
905 return READ_ONCE(memcg->vmstats->events[index]);
908 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
912 int index = memcg_events_index(event);
917 for_each_possible_cpu(cpu)
918 x += per_cpu(memcg->vmstats_percpu->events[index], cpu);
922 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
925 /* pagein of a big page is an event. So, ignore page size */
927 __count_memcg_events(memcg, PGPGIN, 1);
929 __count_memcg_events(memcg, PGPGOUT, 1);
930 nr_pages = -nr_pages; /* for event */
933 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
936 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
937 enum mem_cgroup_events_target target)
939 unsigned long val, next;
941 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
942 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
943 /* from time_after() in jiffies.h */
944 if ((long)(next - val) < 0) {
946 case MEM_CGROUP_TARGET_THRESH:
947 next = val + THRESHOLDS_EVENTS_TARGET;
949 case MEM_CGROUP_TARGET_SOFTLIMIT:
950 next = val + SOFTLIMIT_EVENTS_TARGET;
955 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
962 * Check events in order.
965 static void memcg_check_events(struct mem_cgroup *memcg, int nid)
967 if (IS_ENABLED(CONFIG_PREEMPT_RT))
970 /* threshold event is triggered in finer grain than soft limit */
971 if (unlikely(mem_cgroup_event_ratelimit(memcg,
972 MEM_CGROUP_TARGET_THRESH))) {
975 do_softlimit = mem_cgroup_event_ratelimit(memcg,
976 MEM_CGROUP_TARGET_SOFTLIMIT);
977 mem_cgroup_threshold(memcg);
978 if (unlikely(do_softlimit))
979 mem_cgroup_update_tree(memcg, nid);
983 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
986 * mm_update_next_owner() may clear mm->owner to NULL
987 * if it races with swapoff, page migration, etc.
988 * So this can be called with p == NULL.
993 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
995 EXPORT_SYMBOL(mem_cgroup_from_task);
997 static __always_inline struct mem_cgroup *active_memcg(void)
1000 return this_cpu_read(int_active_memcg);
1002 return current->active_memcg;
1006 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1007 * @mm: mm from which memcg should be extracted. It can be NULL.
1009 * Obtain a reference on mm->memcg and returns it if successful. If mm
1010 * is NULL, then the memcg is chosen as follows:
1011 * 1) The active memcg, if set.
1012 * 2) current->mm->memcg, if available
1014 * If mem_cgroup is disabled, NULL is returned.
1016 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1018 struct mem_cgroup *memcg;
1020 if (mem_cgroup_disabled())
1024 * Page cache insertions can happen without an
1025 * actual mm context, e.g. during disk probing
1026 * on boot, loopback IO, acct() writes etc.
1028 * No need to css_get on root memcg as the reference
1029 * counting is disabled on the root level in the
1030 * cgroup core. See CSS_NO_REF.
1032 if (unlikely(!mm)) {
1033 memcg = active_memcg();
1034 if (unlikely(memcg)) {
1035 /* remote memcg must hold a ref */
1036 css_get(&memcg->css);
1041 return root_mem_cgroup;
1046 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1047 if (unlikely(!memcg))
1048 memcg = root_mem_cgroup;
1049 } while (!css_tryget(&memcg->css));
1053 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1055 static __always_inline bool memcg_kmem_bypass(void)
1057 /* Allow remote memcg charging from any context. */
1058 if (unlikely(active_memcg()))
1061 /* Memcg to charge can't be determined. */
1062 if (!in_task() || !current->mm || (current->flags & PF_KTHREAD))
1069 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1070 * @root: hierarchy root
1071 * @prev: previously returned memcg, NULL on first invocation
1072 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1074 * Returns references to children of the hierarchy below @root, or
1075 * @root itself, or %NULL after a full round-trip.
1077 * Caller must pass the return value in @prev on subsequent
1078 * invocations for reference counting, or use mem_cgroup_iter_break()
1079 * to cancel a hierarchy walk before the round-trip is complete.
1081 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1082 * in the hierarchy among all concurrent reclaimers operating on the
1085 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1086 struct mem_cgroup *prev,
1087 struct mem_cgroup_reclaim_cookie *reclaim)
1089 struct mem_cgroup_reclaim_iter *iter;
1090 struct cgroup_subsys_state *css = NULL;
1091 struct mem_cgroup *memcg = NULL;
1092 struct mem_cgroup *pos = NULL;
1094 if (mem_cgroup_disabled())
1098 root = root_mem_cgroup;
1103 struct mem_cgroup_per_node *mz;
1105 mz = root->nodeinfo[reclaim->pgdat->node_id];
1109 * On start, join the current reclaim iteration cycle.
1110 * Exit when a concurrent walker completes it.
1113 reclaim->generation = iter->generation;
1114 else if (reclaim->generation != iter->generation)
1118 pos = READ_ONCE(iter->position);
1119 if (!pos || css_tryget(&pos->css))
1122 * css reference reached zero, so iter->position will
1123 * be cleared by ->css_released. However, we should not
1124 * rely on this happening soon, because ->css_released
1125 * is called from a work queue, and by busy-waiting we
1126 * might block it. So we clear iter->position right
1129 (void)cmpxchg(&iter->position, pos, NULL);
1139 css = css_next_descendant_pre(css, &root->css);
1142 * Reclaimers share the hierarchy walk, and a
1143 * new one might jump in right at the end of
1144 * the hierarchy - make sure they see at least
1145 * one group and restart from the beginning.
1153 * Verify the css and acquire a reference. The root
1154 * is provided by the caller, so we know it's alive
1155 * and kicking, and don't take an extra reference.
1157 if (css == &root->css || css_tryget(css)) {
1158 memcg = mem_cgroup_from_css(css);
1165 * The position could have already been updated by a competing
1166 * thread, so check that the value hasn't changed since we read
1167 * it to avoid reclaiming from the same cgroup twice.
1169 (void)cmpxchg(&iter->position, pos, memcg);
1180 if (prev && prev != root)
1181 css_put(&prev->css);
1187 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1188 * @root: hierarchy root
1189 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1191 void mem_cgroup_iter_break(struct mem_cgroup *root,
1192 struct mem_cgroup *prev)
1195 root = root_mem_cgroup;
1196 if (prev && prev != root)
1197 css_put(&prev->css);
1200 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1201 struct mem_cgroup *dead_memcg)
1203 struct mem_cgroup_reclaim_iter *iter;
1204 struct mem_cgroup_per_node *mz;
1207 for_each_node(nid) {
1208 mz = from->nodeinfo[nid];
1210 cmpxchg(&iter->position, dead_memcg, NULL);
1214 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1216 struct mem_cgroup *memcg = dead_memcg;
1217 struct mem_cgroup *last;
1220 __invalidate_reclaim_iterators(memcg, dead_memcg);
1222 } while ((memcg = parent_mem_cgroup(memcg)));
1225 * When cgroup1 non-hierarchy mode is used,
1226 * parent_mem_cgroup() does not walk all the way up to the
1227 * cgroup root (root_mem_cgroup). So we have to handle
1228 * dead_memcg from cgroup root separately.
1230 if (!mem_cgroup_is_root(last))
1231 __invalidate_reclaim_iterators(root_mem_cgroup,
1236 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1237 * @memcg: hierarchy root
1238 * @fn: function to call for each task
1239 * @arg: argument passed to @fn
1241 * This function iterates over tasks attached to @memcg or to any of its
1242 * descendants and calls @fn for each task. If @fn returns a non-zero
1243 * value, the function breaks the iteration loop and returns the value.
1244 * Otherwise, it will iterate over all tasks and return 0.
1246 * This function must not be called for the root memory cgroup.
1248 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1249 int (*fn)(struct task_struct *, void *), void *arg)
1251 struct mem_cgroup *iter;
1254 BUG_ON(mem_cgroup_is_root(memcg));
1256 for_each_mem_cgroup_tree(iter, memcg) {
1257 struct css_task_iter it;
1258 struct task_struct *task;
1260 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1261 while (!ret && (task = css_task_iter_next(&it)))
1262 ret = fn(task, arg);
1263 css_task_iter_end(&it);
1265 mem_cgroup_iter_break(memcg, iter);
1272 #ifdef CONFIG_DEBUG_VM
1273 void lruvec_memcg_debug(struct lruvec *lruvec, struct folio *folio)
1275 struct mem_cgroup *memcg;
1277 if (mem_cgroup_disabled())
1280 memcg = folio_memcg(folio);
1283 VM_BUG_ON_FOLIO(!mem_cgroup_is_root(lruvec_memcg(lruvec)), folio);
1285 VM_BUG_ON_FOLIO(lruvec_memcg(lruvec) != memcg, folio);
1290 * folio_lruvec_lock - Lock the lruvec for a folio.
1291 * @folio: Pointer to the folio.
1293 * These functions are safe to use under any of the following conditions:
1295 * - folio_test_lru false
1296 * - folio_memcg_lock()
1297 * - folio frozen (refcount of 0)
1299 * Return: The lruvec this folio is on with its lock held.
1301 struct lruvec *folio_lruvec_lock(struct folio *folio)
1303 struct lruvec *lruvec = folio_lruvec(folio);
1305 spin_lock(&lruvec->lru_lock);
1306 lruvec_memcg_debug(lruvec, folio);
1312 * folio_lruvec_lock_irq - Lock the lruvec for a folio.
1313 * @folio: Pointer to the folio.
1315 * These functions are safe to use under any of the following conditions:
1317 * - folio_test_lru false
1318 * - folio_memcg_lock()
1319 * - folio frozen (refcount of 0)
1321 * Return: The lruvec this folio is on with its lock held and interrupts
1324 struct lruvec *folio_lruvec_lock_irq(struct folio *folio)
1326 struct lruvec *lruvec = folio_lruvec(folio);
1328 spin_lock_irq(&lruvec->lru_lock);
1329 lruvec_memcg_debug(lruvec, folio);
1335 * folio_lruvec_lock_irqsave - Lock the lruvec for a folio.
1336 * @folio: Pointer to the folio.
1337 * @flags: Pointer to irqsave flags.
1339 * These functions are safe to use under any of the following conditions:
1341 * - folio_test_lru false
1342 * - folio_memcg_lock()
1343 * - folio frozen (refcount of 0)
1345 * Return: The lruvec this folio is on with its lock held and interrupts
1348 struct lruvec *folio_lruvec_lock_irqsave(struct folio *folio,
1349 unsigned long *flags)
1351 struct lruvec *lruvec = folio_lruvec(folio);
1353 spin_lock_irqsave(&lruvec->lru_lock, *flags);
1354 lruvec_memcg_debug(lruvec, folio);
1360 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1361 * @lruvec: mem_cgroup per zone lru vector
1362 * @lru: index of lru list the page is sitting on
1363 * @zid: zone id of the accounted pages
1364 * @nr_pages: positive when adding or negative when removing
1366 * This function must be called under lru_lock, just before a page is added
1367 * to or just after a page is removed from an lru list.
1369 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1370 int zid, int nr_pages)
1372 struct mem_cgroup_per_node *mz;
1373 unsigned long *lru_size;
1376 if (mem_cgroup_disabled())
1379 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1380 lru_size = &mz->lru_zone_size[zid][lru];
1383 *lru_size += nr_pages;
1386 if (WARN_ONCE(size < 0,
1387 "%s(%p, %d, %d): lru_size %ld\n",
1388 __func__, lruvec, lru, nr_pages, size)) {
1394 *lru_size += nr_pages;
1398 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1399 * @memcg: the memory cgroup
1401 * Returns the maximum amount of memory @mem can be charged with, in
1404 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1406 unsigned long margin = 0;
1407 unsigned long count;
1408 unsigned long limit;
1410 count = page_counter_read(&memcg->memory);
1411 limit = READ_ONCE(memcg->memory.max);
1413 margin = limit - count;
1415 if (do_memsw_account()) {
1416 count = page_counter_read(&memcg->memsw);
1417 limit = READ_ONCE(memcg->memsw.max);
1419 margin = min(margin, limit - count);
1428 * A routine for checking "mem" is under move_account() or not.
1430 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1431 * moving cgroups. This is for waiting at high-memory pressure
1434 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1436 struct mem_cgroup *from;
1437 struct mem_cgroup *to;
1440 * Unlike task_move routines, we access mc.to, mc.from not under
1441 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1443 spin_lock(&mc.lock);
1449 ret = mem_cgroup_is_descendant(from, memcg) ||
1450 mem_cgroup_is_descendant(to, memcg);
1452 spin_unlock(&mc.lock);
1456 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1458 if (mc.moving_task && current != mc.moving_task) {
1459 if (mem_cgroup_under_move(memcg)) {
1461 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1462 /* moving charge context might have finished. */
1465 finish_wait(&mc.waitq, &wait);
1472 struct memory_stat {
1477 static const struct memory_stat memory_stats[] = {
1478 { "anon", NR_ANON_MAPPED },
1479 { "file", NR_FILE_PAGES },
1480 { "kernel", MEMCG_KMEM },
1481 { "kernel_stack", NR_KERNEL_STACK_KB },
1482 { "pagetables", NR_PAGETABLE },
1483 { "sec_pagetables", NR_SECONDARY_PAGETABLE },
1484 { "percpu", MEMCG_PERCPU_B },
1485 { "sock", MEMCG_SOCK },
1486 { "vmalloc", MEMCG_VMALLOC },
1487 { "shmem", NR_SHMEM },
1488 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
1489 { "zswap", MEMCG_ZSWAP_B },
1490 { "zswapped", MEMCG_ZSWAPPED },
1492 { "file_mapped", NR_FILE_MAPPED },
1493 { "file_dirty", NR_FILE_DIRTY },
1494 { "file_writeback", NR_WRITEBACK },
1496 { "swapcached", NR_SWAPCACHE },
1498 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1499 { "anon_thp", NR_ANON_THPS },
1500 { "file_thp", NR_FILE_THPS },
1501 { "shmem_thp", NR_SHMEM_THPS },
1503 { "inactive_anon", NR_INACTIVE_ANON },
1504 { "active_anon", NR_ACTIVE_ANON },
1505 { "inactive_file", NR_INACTIVE_FILE },
1506 { "active_file", NR_ACTIVE_FILE },
1507 { "unevictable", NR_UNEVICTABLE },
1508 { "slab_reclaimable", NR_SLAB_RECLAIMABLE_B },
1509 { "slab_unreclaimable", NR_SLAB_UNRECLAIMABLE_B },
1511 /* The memory events */
1512 { "workingset_refault_anon", WORKINGSET_REFAULT_ANON },
1513 { "workingset_refault_file", WORKINGSET_REFAULT_FILE },
1514 { "workingset_activate_anon", WORKINGSET_ACTIVATE_ANON },
1515 { "workingset_activate_file", WORKINGSET_ACTIVATE_FILE },
1516 { "workingset_restore_anon", WORKINGSET_RESTORE_ANON },
1517 { "workingset_restore_file", WORKINGSET_RESTORE_FILE },
1518 { "workingset_nodereclaim", WORKINGSET_NODERECLAIM },
1521 /* Translate stat items to the correct unit for memory.stat output */
1522 static int memcg_page_state_unit(int item)
1525 case MEMCG_PERCPU_B:
1527 case NR_SLAB_RECLAIMABLE_B:
1528 case NR_SLAB_UNRECLAIMABLE_B:
1529 case WORKINGSET_REFAULT_ANON:
1530 case WORKINGSET_REFAULT_FILE:
1531 case WORKINGSET_ACTIVATE_ANON:
1532 case WORKINGSET_ACTIVATE_FILE:
1533 case WORKINGSET_RESTORE_ANON:
1534 case WORKINGSET_RESTORE_FILE:
1535 case WORKINGSET_NODERECLAIM:
1537 case NR_KERNEL_STACK_KB:
1544 static inline unsigned long memcg_page_state_output(struct mem_cgroup *memcg,
1547 return memcg_page_state(memcg, item) * memcg_page_state_unit(item);
1550 static void memory_stat_format(struct mem_cgroup *memcg, char *buf, int bufsize)
1555 seq_buf_init(&s, buf, bufsize);
1558 * Provide statistics on the state of the memory subsystem as
1559 * well as cumulative event counters that show past behavior.
1561 * This list is ordered following a combination of these gradients:
1562 * 1) generic big picture -> specifics and details
1563 * 2) reflecting userspace activity -> reflecting kernel heuristics
1565 * Current memory state:
1567 mem_cgroup_flush_stats();
1569 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1572 size = memcg_page_state_output(memcg, memory_stats[i].idx);
1573 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1575 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1576 size += memcg_page_state_output(memcg,
1577 NR_SLAB_RECLAIMABLE_B);
1578 seq_buf_printf(&s, "slab %llu\n", size);
1582 /* Accumulated memory events */
1583 seq_buf_printf(&s, "pgscan %lu\n",
1584 memcg_events(memcg, PGSCAN_KSWAPD) +
1585 memcg_events(memcg, PGSCAN_DIRECT) +
1586 memcg_events(memcg, PGSCAN_KHUGEPAGED));
1587 seq_buf_printf(&s, "pgsteal %lu\n",
1588 memcg_events(memcg, PGSTEAL_KSWAPD) +
1589 memcg_events(memcg, PGSTEAL_DIRECT) +
1590 memcg_events(memcg, PGSTEAL_KHUGEPAGED));
1592 for (i = 0; i < ARRAY_SIZE(memcg_vm_event_stat); i++) {
1593 if (memcg_vm_event_stat[i] == PGPGIN ||
1594 memcg_vm_event_stat[i] == PGPGOUT)
1597 seq_buf_printf(&s, "%s %lu\n",
1598 vm_event_name(memcg_vm_event_stat[i]),
1599 memcg_events(memcg, memcg_vm_event_stat[i]));
1602 /* The above should easily fit into one page */
1603 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1606 #define K(x) ((x) << (PAGE_SHIFT-10))
1608 * mem_cgroup_print_oom_context: Print OOM information relevant to
1609 * memory controller.
1610 * @memcg: The memory cgroup that went over limit
1611 * @p: Task that is going to be killed
1613 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1616 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1621 pr_cont(",oom_memcg=");
1622 pr_cont_cgroup_path(memcg->css.cgroup);
1624 pr_cont(",global_oom");
1626 pr_cont(",task_memcg=");
1627 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1633 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1634 * memory controller.
1635 * @memcg: The memory cgroup that went over limit
1637 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1639 /* Use static buffer, for the caller is holding oom_lock. */
1640 static char buf[PAGE_SIZE];
1642 lockdep_assert_held(&oom_lock);
1644 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1645 K((u64)page_counter_read(&memcg->memory)),
1646 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1647 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1648 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1649 K((u64)page_counter_read(&memcg->swap)),
1650 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1652 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1653 K((u64)page_counter_read(&memcg->memsw)),
1654 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1655 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1656 K((u64)page_counter_read(&memcg->kmem)),
1657 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1660 pr_info("Memory cgroup stats for ");
1661 pr_cont_cgroup_path(memcg->css.cgroup);
1663 memory_stat_format(memcg, buf, sizeof(buf));
1668 * Return the memory (and swap, if configured) limit for a memcg.
1670 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1672 unsigned long max = READ_ONCE(memcg->memory.max);
1674 if (do_memsw_account()) {
1675 if (mem_cgroup_swappiness(memcg)) {
1676 /* Calculate swap excess capacity from memsw limit */
1677 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1679 max += min(swap, (unsigned long)total_swap_pages);
1682 if (mem_cgroup_swappiness(memcg))
1683 max += min(READ_ONCE(memcg->swap.max),
1684 (unsigned long)total_swap_pages);
1689 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1691 return page_counter_read(&memcg->memory);
1694 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1697 struct oom_control oc = {
1701 .gfp_mask = gfp_mask,
1706 if (mutex_lock_killable(&oom_lock))
1709 if (mem_cgroup_margin(memcg) >= (1 << order))
1713 * A few threads which were not waiting at mutex_lock_killable() can
1714 * fail to bail out. Therefore, check again after holding oom_lock.
1716 ret = task_is_dying() || out_of_memory(&oc);
1719 mutex_unlock(&oom_lock);
1723 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1726 unsigned long *total_scanned)
1728 struct mem_cgroup *victim = NULL;
1731 unsigned long excess;
1732 unsigned long nr_scanned;
1733 struct mem_cgroup_reclaim_cookie reclaim = {
1737 excess = soft_limit_excess(root_memcg);
1740 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1745 * If we have not been able to reclaim
1746 * anything, it might because there are
1747 * no reclaimable pages under this hierarchy
1752 * We want to do more targeted reclaim.
1753 * excess >> 2 is not to excessive so as to
1754 * reclaim too much, nor too less that we keep
1755 * coming back to reclaim from this cgroup
1757 if (total >= (excess >> 2) ||
1758 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1763 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1764 pgdat, &nr_scanned);
1765 *total_scanned += nr_scanned;
1766 if (!soft_limit_excess(root_memcg))
1769 mem_cgroup_iter_break(root_memcg, victim);
1773 #ifdef CONFIG_LOCKDEP
1774 static struct lockdep_map memcg_oom_lock_dep_map = {
1775 .name = "memcg_oom_lock",
1779 static DEFINE_SPINLOCK(memcg_oom_lock);
1782 * Check OOM-Killer is already running under our hierarchy.
1783 * If someone is running, return false.
1785 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1787 struct mem_cgroup *iter, *failed = NULL;
1789 spin_lock(&memcg_oom_lock);
1791 for_each_mem_cgroup_tree(iter, memcg) {
1792 if (iter->oom_lock) {
1794 * this subtree of our hierarchy is already locked
1795 * so we cannot give a lock.
1798 mem_cgroup_iter_break(memcg, iter);
1801 iter->oom_lock = true;
1806 * OK, we failed to lock the whole subtree so we have
1807 * to clean up what we set up to the failing subtree
1809 for_each_mem_cgroup_tree(iter, memcg) {
1810 if (iter == failed) {
1811 mem_cgroup_iter_break(memcg, iter);
1814 iter->oom_lock = false;
1817 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1819 spin_unlock(&memcg_oom_lock);
1824 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1826 struct mem_cgroup *iter;
1828 spin_lock(&memcg_oom_lock);
1829 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1830 for_each_mem_cgroup_tree(iter, memcg)
1831 iter->oom_lock = false;
1832 spin_unlock(&memcg_oom_lock);
1835 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1837 struct mem_cgroup *iter;
1839 spin_lock(&memcg_oom_lock);
1840 for_each_mem_cgroup_tree(iter, memcg)
1842 spin_unlock(&memcg_oom_lock);
1845 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1847 struct mem_cgroup *iter;
1850 * Be careful about under_oom underflows because a child memcg
1851 * could have been added after mem_cgroup_mark_under_oom.
1853 spin_lock(&memcg_oom_lock);
1854 for_each_mem_cgroup_tree(iter, memcg)
1855 if (iter->under_oom > 0)
1857 spin_unlock(&memcg_oom_lock);
1860 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1862 struct oom_wait_info {
1863 struct mem_cgroup *memcg;
1864 wait_queue_entry_t wait;
1867 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1868 unsigned mode, int sync, void *arg)
1870 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1871 struct mem_cgroup *oom_wait_memcg;
1872 struct oom_wait_info *oom_wait_info;
1874 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1875 oom_wait_memcg = oom_wait_info->memcg;
1877 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1878 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1880 return autoremove_wake_function(wait, mode, sync, arg);
1883 static void memcg_oom_recover(struct mem_cgroup *memcg)
1886 * For the following lockless ->under_oom test, the only required
1887 * guarantee is that it must see the state asserted by an OOM when
1888 * this function is called as a result of userland actions
1889 * triggered by the notification of the OOM. This is trivially
1890 * achieved by invoking mem_cgroup_mark_under_oom() before
1891 * triggering notification.
1893 if (memcg && memcg->under_oom)
1894 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1898 * Returns true if successfully killed one or more processes. Though in some
1899 * corner cases it can return true even without killing any process.
1901 static bool mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1905 if (order > PAGE_ALLOC_COSTLY_ORDER)
1908 memcg_memory_event(memcg, MEMCG_OOM);
1911 * We are in the middle of the charge context here, so we
1912 * don't want to block when potentially sitting on a callstack
1913 * that holds all kinds of filesystem and mm locks.
1915 * cgroup1 allows disabling the OOM killer and waiting for outside
1916 * handling until the charge can succeed; remember the context and put
1917 * the task to sleep at the end of the page fault when all locks are
1920 * On the other hand, in-kernel OOM killer allows for an async victim
1921 * memory reclaim (oom_reaper) and that means that we are not solely
1922 * relying on the oom victim to make a forward progress and we can
1923 * invoke the oom killer here.
1925 * Please note that mem_cgroup_out_of_memory might fail to find a
1926 * victim and then we have to bail out from the charge path.
1928 if (memcg->oom_kill_disable) {
1929 if (current->in_user_fault) {
1930 css_get(&memcg->css);
1931 current->memcg_in_oom = memcg;
1932 current->memcg_oom_gfp_mask = mask;
1933 current->memcg_oom_order = order;
1938 mem_cgroup_mark_under_oom(memcg);
1940 locked = mem_cgroup_oom_trylock(memcg);
1943 mem_cgroup_oom_notify(memcg);
1945 mem_cgroup_unmark_under_oom(memcg);
1946 ret = mem_cgroup_out_of_memory(memcg, mask, order);
1949 mem_cgroup_oom_unlock(memcg);
1955 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1956 * @handle: actually kill/wait or just clean up the OOM state
1958 * This has to be called at the end of a page fault if the memcg OOM
1959 * handler was enabled.
1961 * Memcg supports userspace OOM handling where failed allocations must
1962 * sleep on a waitqueue until the userspace task resolves the
1963 * situation. Sleeping directly in the charge context with all kinds
1964 * of locks held is not a good idea, instead we remember an OOM state
1965 * in the task and mem_cgroup_oom_synchronize() has to be called at
1966 * the end of the page fault to complete the OOM handling.
1968 * Returns %true if an ongoing memcg OOM situation was detected and
1969 * completed, %false otherwise.
1971 bool mem_cgroup_oom_synchronize(bool handle)
1973 struct mem_cgroup *memcg = current->memcg_in_oom;
1974 struct oom_wait_info owait;
1977 /* OOM is global, do not handle */
1984 owait.memcg = memcg;
1985 owait.wait.flags = 0;
1986 owait.wait.func = memcg_oom_wake_function;
1987 owait.wait.private = current;
1988 INIT_LIST_HEAD(&owait.wait.entry);
1990 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1991 mem_cgroup_mark_under_oom(memcg);
1993 locked = mem_cgroup_oom_trylock(memcg);
1996 mem_cgroup_oom_notify(memcg);
1998 if (locked && !memcg->oom_kill_disable) {
1999 mem_cgroup_unmark_under_oom(memcg);
2000 finish_wait(&memcg_oom_waitq, &owait.wait);
2001 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
2002 current->memcg_oom_order);
2005 mem_cgroup_unmark_under_oom(memcg);
2006 finish_wait(&memcg_oom_waitq, &owait.wait);
2010 mem_cgroup_oom_unlock(memcg);
2012 * There is no guarantee that an OOM-lock contender
2013 * sees the wakeups triggered by the OOM kill
2014 * uncharges. Wake any sleepers explicitly.
2016 memcg_oom_recover(memcg);
2019 current->memcg_in_oom = NULL;
2020 css_put(&memcg->css);
2025 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2026 * @victim: task to be killed by the OOM killer
2027 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2029 * Returns a pointer to a memory cgroup, which has to be cleaned up
2030 * by killing all belonging OOM-killable tasks.
2032 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2034 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2035 struct mem_cgroup *oom_domain)
2037 struct mem_cgroup *oom_group = NULL;
2038 struct mem_cgroup *memcg;
2040 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2044 oom_domain = root_mem_cgroup;
2048 memcg = mem_cgroup_from_task(victim);
2049 if (mem_cgroup_is_root(memcg))
2053 * If the victim task has been asynchronously moved to a different
2054 * memory cgroup, we might end up killing tasks outside oom_domain.
2055 * In this case it's better to ignore memory.group.oom.
2057 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2061 * Traverse the memory cgroup hierarchy from the victim task's
2062 * cgroup up to the OOMing cgroup (or root) to find the
2063 * highest-level memory cgroup with oom.group set.
2065 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2066 if (memcg->oom_group)
2069 if (memcg == oom_domain)
2074 css_get(&oom_group->css);
2081 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2083 pr_info("Tasks in ");
2084 pr_cont_cgroup_path(memcg->css.cgroup);
2085 pr_cont(" are going to be killed due to memory.oom.group set\n");
2089 * folio_memcg_lock - Bind a folio to its memcg.
2090 * @folio: The folio.
2092 * This function prevents unlocked LRU folios from being moved to
2095 * It ensures lifetime of the bound memcg. The caller is responsible
2096 * for the lifetime of the folio.
2098 void folio_memcg_lock(struct folio *folio)
2100 struct mem_cgroup *memcg;
2101 unsigned long flags;
2104 * The RCU lock is held throughout the transaction. The fast
2105 * path can get away without acquiring the memcg->move_lock
2106 * because page moving starts with an RCU grace period.
2110 if (mem_cgroup_disabled())
2113 memcg = folio_memcg(folio);
2114 if (unlikely(!memcg))
2117 #ifdef CONFIG_PROVE_LOCKING
2118 local_irq_save(flags);
2119 might_lock(&memcg->move_lock);
2120 local_irq_restore(flags);
2123 if (atomic_read(&memcg->moving_account) <= 0)
2126 spin_lock_irqsave(&memcg->move_lock, flags);
2127 if (memcg != folio_memcg(folio)) {
2128 spin_unlock_irqrestore(&memcg->move_lock, flags);
2133 * When charge migration first begins, we can have multiple
2134 * critical sections holding the fast-path RCU lock and one
2135 * holding the slowpath move_lock. Track the task who has the
2136 * move_lock for unlock_page_memcg().
2138 memcg->move_lock_task = current;
2139 memcg->move_lock_flags = flags;
2142 void lock_page_memcg(struct page *page)
2144 folio_memcg_lock(page_folio(page));
2147 static void __folio_memcg_unlock(struct mem_cgroup *memcg)
2149 if (memcg && memcg->move_lock_task == current) {
2150 unsigned long flags = memcg->move_lock_flags;
2152 memcg->move_lock_task = NULL;
2153 memcg->move_lock_flags = 0;
2155 spin_unlock_irqrestore(&memcg->move_lock, flags);
2162 * folio_memcg_unlock - Release the binding between a folio and its memcg.
2163 * @folio: The folio.
2165 * This releases the binding created by folio_memcg_lock(). This does
2166 * not change the accounting of this folio to its memcg, but it does
2167 * permit others to change it.
2169 void folio_memcg_unlock(struct folio *folio)
2171 __folio_memcg_unlock(folio_memcg(folio));
2174 void unlock_page_memcg(struct page *page)
2176 folio_memcg_unlock(page_folio(page));
2179 struct memcg_stock_pcp {
2180 local_lock_t stock_lock;
2181 struct mem_cgroup *cached; /* this never be root cgroup */
2182 unsigned int nr_pages;
2184 #ifdef CONFIG_MEMCG_KMEM
2185 struct obj_cgroup *cached_objcg;
2186 struct pglist_data *cached_pgdat;
2187 unsigned int nr_bytes;
2188 int nr_slab_reclaimable_b;
2189 int nr_slab_unreclaimable_b;
2192 struct work_struct work;
2193 unsigned long flags;
2194 #define FLUSHING_CACHED_CHARGE 0
2196 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock) = {
2197 .stock_lock = INIT_LOCAL_LOCK(stock_lock),
2199 static DEFINE_MUTEX(percpu_charge_mutex);
2201 #ifdef CONFIG_MEMCG_KMEM
2202 static struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock);
2203 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2204 struct mem_cgroup *root_memcg);
2205 static void memcg_account_kmem(struct mem_cgroup *memcg, int nr_pages);
2208 static inline struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock)
2212 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2213 struct mem_cgroup *root_memcg)
2217 static void memcg_account_kmem(struct mem_cgroup *memcg, int nr_pages)
2223 * consume_stock: Try to consume stocked charge on this cpu.
2224 * @memcg: memcg to consume from.
2225 * @nr_pages: how many pages to charge.
2227 * The charges will only happen if @memcg matches the current cpu's memcg
2228 * stock, and at least @nr_pages are available in that stock. Failure to
2229 * service an allocation will refill the stock.
2231 * returns true if successful, false otherwise.
2233 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2235 struct memcg_stock_pcp *stock;
2236 unsigned long flags;
2239 if (nr_pages > MEMCG_CHARGE_BATCH)
2242 local_lock_irqsave(&memcg_stock.stock_lock, flags);
2244 stock = this_cpu_ptr(&memcg_stock);
2245 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2246 stock->nr_pages -= nr_pages;
2250 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2256 * Returns stocks cached in percpu and reset cached information.
2258 static void drain_stock(struct memcg_stock_pcp *stock)
2260 struct mem_cgroup *old = stock->cached;
2265 if (stock->nr_pages) {
2266 page_counter_uncharge(&old->memory, stock->nr_pages);
2267 if (do_memsw_account())
2268 page_counter_uncharge(&old->memsw, stock->nr_pages);
2269 stock->nr_pages = 0;
2273 stock->cached = NULL;
2276 static void drain_local_stock(struct work_struct *dummy)
2278 struct memcg_stock_pcp *stock;
2279 struct obj_cgroup *old = NULL;
2280 unsigned long flags;
2283 * The only protection from cpu hotplug (memcg_hotplug_cpu_dead) vs.
2284 * drain_stock races is that we always operate on local CPU stock
2285 * here with IRQ disabled
2287 local_lock_irqsave(&memcg_stock.stock_lock, flags);
2289 stock = this_cpu_ptr(&memcg_stock);
2290 old = drain_obj_stock(stock);
2292 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2294 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2296 obj_cgroup_put(old);
2300 * Cache charges(val) to local per_cpu area.
2301 * This will be consumed by consume_stock() function, later.
2303 static void __refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2305 struct memcg_stock_pcp *stock;
2307 stock = this_cpu_ptr(&memcg_stock);
2308 if (stock->cached != memcg) { /* reset if necessary */
2310 css_get(&memcg->css);
2311 stock->cached = memcg;
2313 stock->nr_pages += nr_pages;
2315 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2319 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2321 unsigned long flags;
2323 local_lock_irqsave(&memcg_stock.stock_lock, flags);
2324 __refill_stock(memcg, nr_pages);
2325 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2329 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2330 * of the hierarchy under it.
2332 static void drain_all_stock(struct mem_cgroup *root_memcg)
2336 /* If someone's already draining, avoid adding running more workers. */
2337 if (!mutex_trylock(&percpu_charge_mutex))
2340 * Notify other cpus that system-wide "drain" is running
2341 * We do not care about races with the cpu hotplug because cpu down
2342 * as well as workers from this path always operate on the local
2343 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2346 curcpu = smp_processor_id();
2347 for_each_online_cpu(cpu) {
2348 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2349 struct mem_cgroup *memcg;
2353 memcg = stock->cached;
2354 if (memcg && stock->nr_pages &&
2355 mem_cgroup_is_descendant(memcg, root_memcg))
2357 else if (obj_stock_flush_required(stock, root_memcg))
2362 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2364 drain_local_stock(&stock->work);
2366 schedule_work_on(cpu, &stock->work);
2370 mutex_unlock(&percpu_charge_mutex);
2373 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2375 struct memcg_stock_pcp *stock;
2377 stock = &per_cpu(memcg_stock, cpu);
2383 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2384 unsigned int nr_pages,
2387 unsigned long nr_reclaimed = 0;
2390 unsigned long pflags;
2392 if (page_counter_read(&memcg->memory) <=
2393 READ_ONCE(memcg->memory.high))
2396 memcg_memory_event(memcg, MEMCG_HIGH);
2398 psi_memstall_enter(&pflags);
2399 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2401 MEMCG_RECLAIM_MAY_SWAP);
2402 psi_memstall_leave(&pflags);
2403 } while ((memcg = parent_mem_cgroup(memcg)) &&
2404 !mem_cgroup_is_root(memcg));
2406 return nr_reclaimed;
2409 static void high_work_func(struct work_struct *work)
2411 struct mem_cgroup *memcg;
2413 memcg = container_of(work, struct mem_cgroup, high_work);
2414 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2418 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2419 * enough to still cause a significant slowdown in most cases, while still
2420 * allowing diagnostics and tracing to proceed without becoming stuck.
2422 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2425 * When calculating the delay, we use these either side of the exponentiation to
2426 * maintain precision and scale to a reasonable number of jiffies (see the table
2429 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2430 * overage ratio to a delay.
2431 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2432 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2433 * to produce a reasonable delay curve.
2435 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2436 * reasonable delay curve compared to precision-adjusted overage, not
2437 * penalising heavily at first, but still making sure that growth beyond the
2438 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2439 * example, with a high of 100 megabytes:
2441 * +-------+------------------------+
2442 * | usage | time to allocate in ms |
2443 * +-------+------------------------+
2465 * +-------+------------------------+
2467 #define MEMCG_DELAY_PRECISION_SHIFT 20
2468 #define MEMCG_DELAY_SCALING_SHIFT 14
2470 static u64 calculate_overage(unsigned long usage, unsigned long high)
2478 * Prevent division by 0 in overage calculation by acting as if
2479 * it was a threshold of 1 page
2481 high = max(high, 1UL);
2483 overage = usage - high;
2484 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2485 return div64_u64(overage, high);
2488 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2490 u64 overage, max_overage = 0;
2493 overage = calculate_overage(page_counter_read(&memcg->memory),
2494 READ_ONCE(memcg->memory.high));
2495 max_overage = max(overage, max_overage);
2496 } while ((memcg = parent_mem_cgroup(memcg)) &&
2497 !mem_cgroup_is_root(memcg));
2502 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2504 u64 overage, max_overage = 0;
2507 overage = calculate_overage(page_counter_read(&memcg->swap),
2508 READ_ONCE(memcg->swap.high));
2510 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2511 max_overage = max(overage, max_overage);
2512 } while ((memcg = parent_mem_cgroup(memcg)) &&
2513 !mem_cgroup_is_root(memcg));
2519 * Get the number of jiffies that we should penalise a mischievous cgroup which
2520 * is exceeding its memory.high by checking both it and its ancestors.
2522 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2523 unsigned int nr_pages,
2526 unsigned long penalty_jiffies;
2532 * We use overage compared to memory.high to calculate the number of
2533 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2534 * fairly lenient on small overages, and increasingly harsh when the
2535 * memcg in question makes it clear that it has no intention of stopping
2536 * its crazy behaviour, so we exponentially increase the delay based on
2539 penalty_jiffies = max_overage * max_overage * HZ;
2540 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2541 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2544 * Factor in the task's own contribution to the overage, such that four
2545 * N-sized allocations are throttled approximately the same as one
2546 * 4N-sized allocation.
2548 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2549 * larger the current charge patch is than that.
2551 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2555 * Scheduled by try_charge() to be executed from the userland return path
2556 * and reclaims memory over the high limit.
2558 void mem_cgroup_handle_over_high(void)
2560 unsigned long penalty_jiffies;
2561 unsigned long pflags;
2562 unsigned long nr_reclaimed;
2563 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2564 int nr_retries = MAX_RECLAIM_RETRIES;
2565 struct mem_cgroup *memcg;
2566 bool in_retry = false;
2568 if (likely(!nr_pages))
2571 memcg = get_mem_cgroup_from_mm(current->mm);
2572 current->memcg_nr_pages_over_high = 0;
2576 * The allocating task should reclaim at least the batch size, but for
2577 * subsequent retries we only want to do what's necessary to prevent oom
2578 * or breaching resource isolation.
2580 * This is distinct from memory.max or page allocator behaviour because
2581 * memory.high is currently batched, whereas memory.max and the page
2582 * allocator run every time an allocation is made.
2584 nr_reclaimed = reclaim_high(memcg,
2585 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2589 * memory.high is breached and reclaim is unable to keep up. Throttle
2590 * allocators proactively to slow down excessive growth.
2592 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2593 mem_find_max_overage(memcg));
2595 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2596 swap_find_max_overage(memcg));
2599 * Clamp the max delay per usermode return so as to still keep the
2600 * application moving forwards and also permit diagnostics, albeit
2603 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2606 * Don't sleep if the amount of jiffies this memcg owes us is so low
2607 * that it's not even worth doing, in an attempt to be nice to those who
2608 * go only a small amount over their memory.high value and maybe haven't
2609 * been aggressively reclaimed enough yet.
2611 if (penalty_jiffies <= HZ / 100)
2615 * If reclaim is making forward progress but we're still over
2616 * memory.high, we want to encourage that rather than doing allocator
2619 if (nr_reclaimed || nr_retries--) {
2625 * If we exit early, we're guaranteed to die (since
2626 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2627 * need to account for any ill-begotten jiffies to pay them off later.
2629 psi_memstall_enter(&pflags);
2630 schedule_timeout_killable(penalty_jiffies);
2631 psi_memstall_leave(&pflags);
2634 css_put(&memcg->css);
2637 static int try_charge_memcg(struct mem_cgroup *memcg, gfp_t gfp_mask,
2638 unsigned int nr_pages)
2640 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2641 int nr_retries = MAX_RECLAIM_RETRIES;
2642 struct mem_cgroup *mem_over_limit;
2643 struct page_counter *counter;
2644 unsigned long nr_reclaimed;
2645 bool passed_oom = false;
2646 unsigned int reclaim_options = MEMCG_RECLAIM_MAY_SWAP;
2647 bool drained = false;
2648 bool raised_max_event = false;
2649 unsigned long pflags;
2652 if (consume_stock(memcg, nr_pages))
2655 if (!do_memsw_account() ||
2656 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2657 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2659 if (do_memsw_account())
2660 page_counter_uncharge(&memcg->memsw, batch);
2661 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2663 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2664 reclaim_options &= ~MEMCG_RECLAIM_MAY_SWAP;
2667 if (batch > nr_pages) {
2673 * Prevent unbounded recursion when reclaim operations need to
2674 * allocate memory. This might exceed the limits temporarily,
2675 * but we prefer facilitating memory reclaim and getting back
2676 * under the limit over triggering OOM kills in these cases.
2678 if (unlikely(current->flags & PF_MEMALLOC))
2681 if (unlikely(task_in_memcg_oom(current)))
2684 if (!gfpflags_allow_blocking(gfp_mask))
2687 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2688 raised_max_event = true;
2690 psi_memstall_enter(&pflags);
2691 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2692 gfp_mask, reclaim_options);
2693 psi_memstall_leave(&pflags);
2695 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2699 drain_all_stock(mem_over_limit);
2704 if (gfp_mask & __GFP_NORETRY)
2707 * Even though the limit is exceeded at this point, reclaim
2708 * may have been able to free some pages. Retry the charge
2709 * before killing the task.
2711 * Only for regular pages, though: huge pages are rather
2712 * unlikely to succeed so close to the limit, and we fall back
2713 * to regular pages anyway in case of failure.
2715 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2718 * At task move, charge accounts can be doubly counted. So, it's
2719 * better to wait until the end of task_move if something is going on.
2721 if (mem_cgroup_wait_acct_move(mem_over_limit))
2727 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2730 /* Avoid endless loop for tasks bypassed by the oom killer */
2731 if (passed_oom && task_is_dying())
2735 * keep retrying as long as the memcg oom killer is able to make
2736 * a forward progress or bypass the charge if the oom killer
2737 * couldn't make any progress.
2739 if (mem_cgroup_oom(mem_over_limit, gfp_mask,
2740 get_order(nr_pages * PAGE_SIZE))) {
2742 nr_retries = MAX_RECLAIM_RETRIES;
2747 * Memcg doesn't have a dedicated reserve for atomic
2748 * allocations. But like the global atomic pool, we need to
2749 * put the burden of reclaim on regular allocation requests
2750 * and let these go through as privileged allocations.
2752 if (!(gfp_mask & (__GFP_NOFAIL | __GFP_HIGH)))
2756 * If the allocation has to be enforced, don't forget to raise
2757 * a MEMCG_MAX event.
2759 if (!raised_max_event)
2760 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2763 * The allocation either can't fail or will lead to more memory
2764 * being freed very soon. Allow memory usage go over the limit
2765 * temporarily by force charging it.
2767 page_counter_charge(&memcg->memory, nr_pages);
2768 if (do_memsw_account())
2769 page_counter_charge(&memcg->memsw, nr_pages);
2774 if (batch > nr_pages)
2775 refill_stock(memcg, batch - nr_pages);
2778 * If the hierarchy is above the normal consumption range, schedule
2779 * reclaim on returning to userland. We can perform reclaim here
2780 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2781 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2782 * not recorded as it most likely matches current's and won't
2783 * change in the meantime. As high limit is checked again before
2784 * reclaim, the cost of mismatch is negligible.
2787 bool mem_high, swap_high;
2789 mem_high = page_counter_read(&memcg->memory) >
2790 READ_ONCE(memcg->memory.high);
2791 swap_high = page_counter_read(&memcg->swap) >
2792 READ_ONCE(memcg->swap.high);
2794 /* Don't bother a random interrupted task */
2797 schedule_work(&memcg->high_work);
2803 if (mem_high || swap_high) {
2805 * The allocating tasks in this cgroup will need to do
2806 * reclaim or be throttled to prevent further growth
2807 * of the memory or swap footprints.
2809 * Target some best-effort fairness between the tasks,
2810 * and distribute reclaim work and delay penalties
2811 * based on how much each task is actually allocating.
2813 current->memcg_nr_pages_over_high += batch;
2814 set_notify_resume(current);
2817 } while ((memcg = parent_mem_cgroup(memcg)));
2819 if (current->memcg_nr_pages_over_high > MEMCG_CHARGE_BATCH &&
2820 !(current->flags & PF_MEMALLOC) &&
2821 gfpflags_allow_blocking(gfp_mask)) {
2822 mem_cgroup_handle_over_high();
2827 static inline int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2828 unsigned int nr_pages)
2830 if (mem_cgroup_is_root(memcg))
2833 return try_charge_memcg(memcg, gfp_mask, nr_pages);
2836 static inline void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2838 if (mem_cgroup_is_root(memcg))
2841 page_counter_uncharge(&memcg->memory, nr_pages);
2842 if (do_memsw_account())
2843 page_counter_uncharge(&memcg->memsw, nr_pages);
2846 static void commit_charge(struct folio *folio, struct mem_cgroup *memcg)
2848 VM_BUG_ON_FOLIO(folio_memcg(folio), folio);
2850 * Any of the following ensures page's memcg stability:
2854 * - lock_page_memcg()
2855 * - exclusive reference
2856 * - mem_cgroup_trylock_pages()
2858 folio->memcg_data = (unsigned long)memcg;
2861 #ifdef CONFIG_MEMCG_KMEM
2863 * The allocated objcg pointers array is not accounted directly.
2864 * Moreover, it should not come from DMA buffer and is not readily
2865 * reclaimable. So those GFP bits should be masked off.
2867 #define OBJCGS_CLEAR_MASK (__GFP_DMA | __GFP_RECLAIMABLE | __GFP_ACCOUNT)
2870 * mod_objcg_mlstate() may be called with irq enabled, so
2871 * mod_memcg_lruvec_state() should be used.
2873 static inline void mod_objcg_mlstate(struct obj_cgroup *objcg,
2874 struct pglist_data *pgdat,
2875 enum node_stat_item idx, int nr)
2877 struct mem_cgroup *memcg;
2878 struct lruvec *lruvec;
2881 memcg = obj_cgroup_memcg(objcg);
2882 lruvec = mem_cgroup_lruvec(memcg, pgdat);
2883 mod_memcg_lruvec_state(lruvec, idx, nr);
2887 int memcg_alloc_slab_cgroups(struct slab *slab, struct kmem_cache *s,
2888 gfp_t gfp, bool new_slab)
2890 unsigned int objects = objs_per_slab(s, slab);
2891 unsigned long memcg_data;
2894 gfp &= ~OBJCGS_CLEAR_MASK;
2895 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2900 memcg_data = (unsigned long) vec | MEMCG_DATA_OBJCGS;
2903 * If the slab is brand new and nobody can yet access its
2904 * memcg_data, no synchronization is required and memcg_data can
2905 * be simply assigned.
2907 slab->memcg_data = memcg_data;
2908 } else if (cmpxchg(&slab->memcg_data, 0, memcg_data)) {
2910 * If the slab is already in use, somebody can allocate and
2911 * assign obj_cgroups in parallel. In this case the existing
2912 * objcg vector should be reused.
2918 kmemleak_not_leak(vec);
2922 static __always_inline
2923 struct mem_cgroup *mem_cgroup_from_obj_folio(struct folio *folio, void *p)
2926 * Slab objects are accounted individually, not per-page.
2927 * Memcg membership data for each individual object is saved in
2930 if (folio_test_slab(folio)) {
2931 struct obj_cgroup **objcgs;
2935 slab = folio_slab(folio);
2936 objcgs = slab_objcgs(slab);
2940 off = obj_to_index(slab->slab_cache, slab, p);
2942 return obj_cgroup_memcg(objcgs[off]);
2948 * page_memcg_check() is used here, because in theory we can encounter
2949 * a folio where the slab flag has been cleared already, but
2950 * slab->memcg_data has not been freed yet
2951 * page_memcg_check(page) will guarantee that a proper memory
2952 * cgroup pointer or NULL will be returned.
2954 return page_memcg_check(folio_page(folio, 0));
2958 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2960 * A passed kernel object can be a slab object, vmalloc object or a generic
2961 * kernel page, so different mechanisms for getting the memory cgroup pointer
2964 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2965 * can not know for sure how the kernel object is implemented.
2966 * mem_cgroup_from_obj() can be safely used in such cases.
2968 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2969 * cgroup_mutex, etc.
2971 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2973 struct folio *folio;
2975 if (mem_cgroup_disabled())
2978 if (unlikely(is_vmalloc_addr(p)))
2979 folio = page_folio(vmalloc_to_page(p));
2981 folio = virt_to_folio(p);
2983 return mem_cgroup_from_obj_folio(folio, p);
2987 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2988 * Similar to mem_cgroup_from_obj(), but faster and not suitable for objects,
2989 * allocated using vmalloc().
2991 * A passed kernel object must be a slab object or a generic kernel page.
2993 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2994 * cgroup_mutex, etc.
2996 struct mem_cgroup *mem_cgroup_from_slab_obj(void *p)
2998 if (mem_cgroup_disabled())
3001 return mem_cgroup_from_obj_folio(virt_to_folio(p), p);
3004 static struct obj_cgroup *__get_obj_cgroup_from_memcg(struct mem_cgroup *memcg)
3006 struct obj_cgroup *objcg = NULL;
3008 for (; !mem_cgroup_is_root(memcg); memcg = parent_mem_cgroup(memcg)) {
3009 objcg = rcu_dereference(memcg->objcg);
3010 if (objcg && obj_cgroup_tryget(objcg))
3017 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
3019 struct obj_cgroup *objcg = NULL;
3020 struct mem_cgroup *memcg;
3022 if (memcg_kmem_bypass())
3026 if (unlikely(active_memcg()))
3027 memcg = active_memcg();
3029 memcg = mem_cgroup_from_task(current);
3030 objcg = __get_obj_cgroup_from_memcg(memcg);
3035 struct obj_cgroup *get_obj_cgroup_from_page(struct page *page)
3037 struct obj_cgroup *objcg;
3039 if (!memcg_kmem_enabled())
3042 if (PageMemcgKmem(page)) {
3043 objcg = __folio_objcg(page_folio(page));
3044 obj_cgroup_get(objcg);
3046 struct mem_cgroup *memcg;
3049 memcg = __folio_memcg(page_folio(page));
3051 objcg = __get_obj_cgroup_from_memcg(memcg);
3059 static void memcg_account_kmem(struct mem_cgroup *memcg, int nr_pages)
3061 mod_memcg_state(memcg, MEMCG_KMEM, nr_pages);
3062 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
3064 page_counter_charge(&memcg->kmem, nr_pages);
3066 page_counter_uncharge(&memcg->kmem, -nr_pages);
3072 * obj_cgroup_uncharge_pages: uncharge a number of kernel pages from a objcg
3073 * @objcg: object cgroup to uncharge
3074 * @nr_pages: number of pages to uncharge
3076 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
3077 unsigned int nr_pages)
3079 struct mem_cgroup *memcg;
3081 memcg = get_mem_cgroup_from_objcg(objcg);
3083 memcg_account_kmem(memcg, -nr_pages);
3084 refill_stock(memcg, nr_pages);
3086 css_put(&memcg->css);
3090 * obj_cgroup_charge_pages: charge a number of kernel pages to a objcg
3091 * @objcg: object cgroup to charge
3092 * @gfp: reclaim mode
3093 * @nr_pages: number of pages to charge
3095 * Returns 0 on success, an error code on failure.
3097 static int obj_cgroup_charge_pages(struct obj_cgroup *objcg, gfp_t gfp,
3098 unsigned int nr_pages)
3100 struct mem_cgroup *memcg;
3103 memcg = get_mem_cgroup_from_objcg(objcg);
3105 ret = try_charge_memcg(memcg, gfp, nr_pages);
3109 memcg_account_kmem(memcg, nr_pages);
3111 css_put(&memcg->css);
3117 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3118 * @page: page to charge
3119 * @gfp: reclaim mode
3120 * @order: allocation order
3122 * Returns 0 on success, an error code on failure.
3124 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3126 struct obj_cgroup *objcg;
3129 objcg = get_obj_cgroup_from_current();
3131 ret = obj_cgroup_charge_pages(objcg, gfp, 1 << order);
3133 page->memcg_data = (unsigned long)objcg |
3137 obj_cgroup_put(objcg);
3143 * __memcg_kmem_uncharge_page: uncharge a kmem page
3144 * @page: page to uncharge
3145 * @order: allocation order
3147 void __memcg_kmem_uncharge_page(struct page *page, int order)
3149 struct folio *folio = page_folio(page);
3150 struct obj_cgroup *objcg;
3151 unsigned int nr_pages = 1 << order;
3153 if (!folio_memcg_kmem(folio))
3156 objcg = __folio_objcg(folio);
3157 obj_cgroup_uncharge_pages(objcg, nr_pages);
3158 folio->memcg_data = 0;
3159 obj_cgroup_put(objcg);
3162 void mod_objcg_state(struct obj_cgroup *objcg, struct pglist_data *pgdat,
3163 enum node_stat_item idx, int nr)
3165 struct memcg_stock_pcp *stock;
3166 struct obj_cgroup *old = NULL;
3167 unsigned long flags;
3170 local_lock_irqsave(&memcg_stock.stock_lock, flags);
3171 stock = this_cpu_ptr(&memcg_stock);
3174 * Save vmstat data in stock and skip vmstat array update unless
3175 * accumulating over a page of vmstat data or when pgdat or idx
3178 if (stock->cached_objcg != objcg) {
3179 old = drain_obj_stock(stock);
3180 obj_cgroup_get(objcg);
3181 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3182 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3183 stock->cached_objcg = objcg;
3184 stock->cached_pgdat = pgdat;
3185 } else if (stock->cached_pgdat != pgdat) {
3186 /* Flush the existing cached vmstat data */
3187 struct pglist_data *oldpg = stock->cached_pgdat;
3189 if (stock->nr_slab_reclaimable_b) {
3190 mod_objcg_mlstate(objcg, oldpg, NR_SLAB_RECLAIMABLE_B,
3191 stock->nr_slab_reclaimable_b);
3192 stock->nr_slab_reclaimable_b = 0;
3194 if (stock->nr_slab_unreclaimable_b) {
3195 mod_objcg_mlstate(objcg, oldpg, NR_SLAB_UNRECLAIMABLE_B,
3196 stock->nr_slab_unreclaimable_b);
3197 stock->nr_slab_unreclaimable_b = 0;
3199 stock->cached_pgdat = pgdat;
3202 bytes = (idx == NR_SLAB_RECLAIMABLE_B) ? &stock->nr_slab_reclaimable_b
3203 : &stock->nr_slab_unreclaimable_b;
3205 * Even for large object >= PAGE_SIZE, the vmstat data will still be
3206 * cached locally at least once before pushing it out.
3213 if (abs(*bytes) > PAGE_SIZE) {
3221 mod_objcg_mlstate(objcg, pgdat, idx, nr);
3223 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3225 obj_cgroup_put(old);
3228 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3230 struct memcg_stock_pcp *stock;
3231 unsigned long flags;
3234 local_lock_irqsave(&memcg_stock.stock_lock, flags);
3236 stock = this_cpu_ptr(&memcg_stock);
3237 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3238 stock->nr_bytes -= nr_bytes;
3242 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3247 static struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock)
3249 struct obj_cgroup *old = stock->cached_objcg;
3254 if (stock->nr_bytes) {
3255 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3256 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3259 struct mem_cgroup *memcg;
3261 memcg = get_mem_cgroup_from_objcg(old);
3263 memcg_account_kmem(memcg, -nr_pages);
3264 __refill_stock(memcg, nr_pages);
3266 css_put(&memcg->css);
3270 * The leftover is flushed to the centralized per-memcg value.
3271 * On the next attempt to refill obj stock it will be moved
3272 * to a per-cpu stock (probably, on an other CPU), see
3273 * refill_obj_stock().
3275 * How often it's flushed is a trade-off between the memory
3276 * limit enforcement accuracy and potential CPU contention,
3277 * so it might be changed in the future.
3279 atomic_add(nr_bytes, &old->nr_charged_bytes);
3280 stock->nr_bytes = 0;
3284 * Flush the vmstat data in current stock
3286 if (stock->nr_slab_reclaimable_b || stock->nr_slab_unreclaimable_b) {
3287 if (stock->nr_slab_reclaimable_b) {
3288 mod_objcg_mlstate(old, stock->cached_pgdat,
3289 NR_SLAB_RECLAIMABLE_B,
3290 stock->nr_slab_reclaimable_b);
3291 stock->nr_slab_reclaimable_b = 0;
3293 if (stock->nr_slab_unreclaimable_b) {
3294 mod_objcg_mlstate(old, stock->cached_pgdat,
3295 NR_SLAB_UNRECLAIMABLE_B,
3296 stock->nr_slab_unreclaimable_b);
3297 stock->nr_slab_unreclaimable_b = 0;
3299 stock->cached_pgdat = NULL;
3302 stock->cached_objcg = NULL;
3304 * The `old' objects needs to be released by the caller via
3305 * obj_cgroup_put() outside of memcg_stock_pcp::stock_lock.
3310 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3311 struct mem_cgroup *root_memcg)
3313 struct mem_cgroup *memcg;
3315 if (stock->cached_objcg) {
3316 memcg = obj_cgroup_memcg(stock->cached_objcg);
3317 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3324 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes,
3325 bool allow_uncharge)
3327 struct memcg_stock_pcp *stock;
3328 struct obj_cgroup *old = NULL;
3329 unsigned long flags;
3330 unsigned int nr_pages = 0;
3332 local_lock_irqsave(&memcg_stock.stock_lock, flags);
3334 stock = this_cpu_ptr(&memcg_stock);
3335 if (stock->cached_objcg != objcg) { /* reset if necessary */
3336 old = drain_obj_stock(stock);
3337 obj_cgroup_get(objcg);
3338 stock->cached_objcg = objcg;
3339 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3340 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3341 allow_uncharge = true; /* Allow uncharge when objcg changes */
3343 stock->nr_bytes += nr_bytes;
3345 if (allow_uncharge && (stock->nr_bytes > PAGE_SIZE)) {
3346 nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3347 stock->nr_bytes &= (PAGE_SIZE - 1);
3350 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3352 obj_cgroup_put(old);
3355 obj_cgroup_uncharge_pages(objcg, nr_pages);
3358 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3360 unsigned int nr_pages, nr_bytes;
3363 if (consume_obj_stock(objcg, size))
3367 * In theory, objcg->nr_charged_bytes can have enough
3368 * pre-charged bytes to satisfy the allocation. However,
3369 * flushing objcg->nr_charged_bytes requires two atomic
3370 * operations, and objcg->nr_charged_bytes can't be big.
3371 * The shared objcg->nr_charged_bytes can also become a
3372 * performance bottleneck if all tasks of the same memcg are
3373 * trying to update it. So it's better to ignore it and try
3374 * grab some new pages. The stock's nr_bytes will be flushed to
3375 * objcg->nr_charged_bytes later on when objcg changes.
3377 * The stock's nr_bytes may contain enough pre-charged bytes
3378 * to allow one less page from being charged, but we can't rely
3379 * on the pre-charged bytes not being changed outside of
3380 * consume_obj_stock() or refill_obj_stock(). So ignore those
3381 * pre-charged bytes as well when charging pages. To avoid a
3382 * page uncharge right after a page charge, we set the
3383 * allow_uncharge flag to false when calling refill_obj_stock()
3384 * to temporarily allow the pre-charged bytes to exceed the page
3385 * size limit. The maximum reachable value of the pre-charged
3386 * bytes is (sizeof(object) + PAGE_SIZE - 2) if there is no data
3389 nr_pages = size >> PAGE_SHIFT;
3390 nr_bytes = size & (PAGE_SIZE - 1);
3395 ret = obj_cgroup_charge_pages(objcg, gfp, nr_pages);
3396 if (!ret && nr_bytes)
3397 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes, false);
3402 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3404 refill_obj_stock(objcg, size, true);
3407 #endif /* CONFIG_MEMCG_KMEM */
3410 * Because page_memcg(head) is not set on tails, set it now.
3412 void split_page_memcg(struct page *head, unsigned int nr)
3414 struct folio *folio = page_folio(head);
3415 struct mem_cgroup *memcg = folio_memcg(folio);
3418 if (mem_cgroup_disabled() || !memcg)
3421 for (i = 1; i < nr; i++)
3422 folio_page(folio, i)->memcg_data = folio->memcg_data;
3424 if (folio_memcg_kmem(folio))
3425 obj_cgroup_get_many(__folio_objcg(folio), nr - 1);
3427 css_get_many(&memcg->css, nr - 1);
3432 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3433 * @entry: swap entry to be moved
3434 * @from: mem_cgroup which the entry is moved from
3435 * @to: mem_cgroup which the entry is moved to
3437 * It succeeds only when the swap_cgroup's record for this entry is the same
3438 * as the mem_cgroup's id of @from.
3440 * Returns 0 on success, -EINVAL on failure.
3442 * The caller must have charged to @to, IOW, called page_counter_charge() about
3443 * both res and memsw, and called css_get().
3445 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3446 struct mem_cgroup *from, struct mem_cgroup *to)
3448 unsigned short old_id, new_id;
3450 old_id = mem_cgroup_id(from);
3451 new_id = mem_cgroup_id(to);
3453 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3454 mod_memcg_state(from, MEMCG_SWAP, -1);
3455 mod_memcg_state(to, MEMCG_SWAP, 1);
3461 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3462 struct mem_cgroup *from, struct mem_cgroup *to)
3468 static DEFINE_MUTEX(memcg_max_mutex);
3470 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3471 unsigned long max, bool memsw)
3473 bool enlarge = false;
3474 bool drained = false;
3476 bool limits_invariant;
3477 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3480 if (signal_pending(current)) {
3485 mutex_lock(&memcg_max_mutex);
3487 * Make sure that the new limit (memsw or memory limit) doesn't
3488 * break our basic invariant rule memory.max <= memsw.max.
3490 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3491 max <= memcg->memsw.max;
3492 if (!limits_invariant) {
3493 mutex_unlock(&memcg_max_mutex);
3497 if (max > counter->max)
3499 ret = page_counter_set_max(counter, max);
3500 mutex_unlock(&memcg_max_mutex);
3506 drain_all_stock(memcg);
3511 if (!try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL,
3512 memsw ? 0 : MEMCG_RECLAIM_MAY_SWAP)) {
3518 if (!ret && enlarge)
3519 memcg_oom_recover(memcg);
3524 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3526 unsigned long *total_scanned)
3528 unsigned long nr_reclaimed = 0;
3529 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3530 unsigned long reclaimed;
3532 struct mem_cgroup_tree_per_node *mctz;
3533 unsigned long excess;
3538 mctz = soft_limit_tree.rb_tree_per_node[pgdat->node_id];
3541 * Do not even bother to check the largest node if the root
3542 * is empty. Do it lockless to prevent lock bouncing. Races
3543 * are acceptable as soft limit is best effort anyway.
3545 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3549 * This loop can run a while, specially if mem_cgroup's continuously
3550 * keep exceeding their soft limit and putting the system under
3557 mz = mem_cgroup_largest_soft_limit_node(mctz);
3561 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3562 gfp_mask, total_scanned);
3563 nr_reclaimed += reclaimed;
3564 spin_lock_irq(&mctz->lock);
3567 * If we failed to reclaim anything from this memory cgroup
3568 * it is time to move on to the next cgroup
3572 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3574 excess = soft_limit_excess(mz->memcg);
3576 * One school of thought says that we should not add
3577 * back the node to the tree if reclaim returns 0.
3578 * But our reclaim could return 0, simply because due
3579 * to priority we are exposing a smaller subset of
3580 * memory to reclaim from. Consider this as a longer
3583 /* If excess == 0, no tree ops */
3584 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3585 spin_unlock_irq(&mctz->lock);
3586 css_put(&mz->memcg->css);
3589 * Could not reclaim anything and there are no more
3590 * mem cgroups to try or we seem to be looping without
3591 * reclaiming anything.
3593 if (!nr_reclaimed &&
3595 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3597 } while (!nr_reclaimed);
3599 css_put(&next_mz->memcg->css);
3600 return nr_reclaimed;
3604 * Reclaims as many pages from the given memcg as possible.
3606 * Caller is responsible for holding css reference for memcg.
3608 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3610 int nr_retries = MAX_RECLAIM_RETRIES;
3612 /* we call try-to-free pages for make this cgroup empty */
3613 lru_add_drain_all();
3615 drain_all_stock(memcg);
3617 /* try to free all pages in this cgroup */
3618 while (nr_retries && page_counter_read(&memcg->memory)) {
3619 if (signal_pending(current))
3622 if (!try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL,
3623 MEMCG_RECLAIM_MAY_SWAP))
3630 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3631 char *buf, size_t nbytes,
3634 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3636 if (mem_cgroup_is_root(memcg))
3638 return mem_cgroup_force_empty(memcg) ?: nbytes;
3641 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3647 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3648 struct cftype *cft, u64 val)
3653 pr_warn_once("Non-hierarchical mode is deprecated. "
3654 "Please report your usecase to linux-mm@kvack.org if you "
3655 "depend on this functionality.\n");
3660 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3664 if (mem_cgroup_is_root(memcg)) {
3665 mem_cgroup_flush_stats();
3666 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3667 memcg_page_state(memcg, NR_ANON_MAPPED);
3669 val += memcg_page_state(memcg, MEMCG_SWAP);
3672 val = page_counter_read(&memcg->memory);
3674 val = page_counter_read(&memcg->memsw);
3687 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3690 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3691 struct page_counter *counter;
3693 switch (MEMFILE_TYPE(cft->private)) {
3695 counter = &memcg->memory;
3698 counter = &memcg->memsw;
3701 counter = &memcg->kmem;
3704 counter = &memcg->tcpmem;
3710 switch (MEMFILE_ATTR(cft->private)) {
3712 if (counter == &memcg->memory)
3713 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3714 if (counter == &memcg->memsw)
3715 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3716 return (u64)page_counter_read(counter) * PAGE_SIZE;
3718 return (u64)counter->max * PAGE_SIZE;
3720 return (u64)counter->watermark * PAGE_SIZE;
3722 return counter->failcnt;
3723 case RES_SOFT_LIMIT:
3724 return (u64)memcg->soft_limit * PAGE_SIZE;
3730 #ifdef CONFIG_MEMCG_KMEM
3731 static int memcg_online_kmem(struct mem_cgroup *memcg)
3733 struct obj_cgroup *objcg;
3735 if (mem_cgroup_kmem_disabled())
3738 if (unlikely(mem_cgroup_is_root(memcg)))
3741 objcg = obj_cgroup_alloc();
3745 objcg->memcg = memcg;
3746 rcu_assign_pointer(memcg->objcg, objcg);
3748 static_branch_enable(&memcg_kmem_enabled_key);
3750 memcg->kmemcg_id = memcg->id.id;
3755 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3757 struct mem_cgroup *parent;
3759 if (mem_cgroup_kmem_disabled())
3762 if (unlikely(mem_cgroup_is_root(memcg)))
3765 parent = parent_mem_cgroup(memcg);
3767 parent = root_mem_cgroup;
3769 memcg_reparent_objcgs(memcg, parent);
3772 * After we have finished memcg_reparent_objcgs(), all list_lrus
3773 * corresponding to this cgroup are guaranteed to remain empty.
3774 * The ordering is imposed by list_lru_node->lock taken by
3775 * memcg_reparent_list_lrus().
3777 memcg_reparent_list_lrus(memcg, parent);
3780 static int memcg_online_kmem(struct mem_cgroup *memcg)
3784 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3787 #endif /* CONFIG_MEMCG_KMEM */
3789 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3793 mutex_lock(&memcg_max_mutex);
3795 ret = page_counter_set_max(&memcg->tcpmem, max);
3799 if (!memcg->tcpmem_active) {
3801 * The active flag needs to be written after the static_key
3802 * update. This is what guarantees that the socket activation
3803 * function is the last one to run. See mem_cgroup_sk_alloc()
3804 * for details, and note that we don't mark any socket as
3805 * belonging to this memcg until that flag is up.
3807 * We need to do this, because static_keys will span multiple
3808 * sites, but we can't control their order. If we mark a socket
3809 * as accounted, but the accounting functions are not patched in
3810 * yet, we'll lose accounting.
3812 * We never race with the readers in mem_cgroup_sk_alloc(),
3813 * because when this value change, the code to process it is not
3816 static_branch_inc(&memcg_sockets_enabled_key);
3817 memcg->tcpmem_active = true;
3820 mutex_unlock(&memcg_max_mutex);
3825 * The user of this function is...
3828 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3829 char *buf, size_t nbytes, loff_t off)
3831 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3832 unsigned long nr_pages;
3835 buf = strstrip(buf);
3836 ret = page_counter_memparse(buf, "-1", &nr_pages);
3840 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3842 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3846 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3848 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3851 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3854 /* kmem.limit_in_bytes is deprecated. */
3858 ret = memcg_update_tcp_max(memcg, nr_pages);
3862 case RES_SOFT_LIMIT:
3863 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
3866 memcg->soft_limit = nr_pages;
3871 return ret ?: nbytes;
3874 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3875 size_t nbytes, loff_t off)
3877 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3878 struct page_counter *counter;
3880 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3882 counter = &memcg->memory;
3885 counter = &memcg->memsw;
3888 counter = &memcg->kmem;
3891 counter = &memcg->tcpmem;
3897 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3899 page_counter_reset_watermark(counter);
3902 counter->failcnt = 0;
3911 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3914 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3918 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3919 struct cftype *cft, u64 val)
3921 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3923 if (val & ~MOVE_MASK)
3927 * No kind of locking is needed in here, because ->can_attach() will
3928 * check this value once in the beginning of the process, and then carry
3929 * on with stale data. This means that changes to this value will only
3930 * affect task migrations starting after the change.
3932 memcg->move_charge_at_immigrate = val;
3936 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3937 struct cftype *cft, u64 val)
3945 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3946 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3947 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3949 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3950 int nid, unsigned int lru_mask, bool tree)
3952 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3953 unsigned long nr = 0;
3956 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3959 if (!(BIT(lru) & lru_mask))
3962 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3964 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3969 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3970 unsigned int lru_mask,
3973 unsigned long nr = 0;
3977 if (!(BIT(lru) & lru_mask))
3980 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3982 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3987 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3991 unsigned int lru_mask;
3994 static const struct numa_stat stats[] = {
3995 { "total", LRU_ALL },
3996 { "file", LRU_ALL_FILE },
3997 { "anon", LRU_ALL_ANON },
3998 { "unevictable", BIT(LRU_UNEVICTABLE) },
4000 const struct numa_stat *stat;
4002 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4004 mem_cgroup_flush_stats();
4006 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4007 seq_printf(m, "%s=%lu", stat->name,
4008 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4010 for_each_node_state(nid, N_MEMORY)
4011 seq_printf(m, " N%d=%lu", nid,
4012 mem_cgroup_node_nr_lru_pages(memcg, nid,
4013 stat->lru_mask, false));
4017 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4019 seq_printf(m, "hierarchical_%s=%lu", stat->name,
4020 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4022 for_each_node_state(nid, N_MEMORY)
4023 seq_printf(m, " N%d=%lu", nid,
4024 mem_cgroup_node_nr_lru_pages(memcg, nid,
4025 stat->lru_mask, true));
4031 #endif /* CONFIG_NUMA */
4033 static const unsigned int memcg1_stats[] = {
4036 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4043 WORKINGSET_REFAULT_ANON,
4044 WORKINGSET_REFAULT_FILE,
4048 static const char *const memcg1_stat_names[] = {
4051 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4058 "workingset_refault_anon",
4059 "workingset_refault_file",
4063 /* Universal VM events cgroup1 shows, original sort order */
4064 static const unsigned int memcg1_events[] = {
4071 static int memcg_stat_show(struct seq_file *m, void *v)
4073 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4074 unsigned long memory, memsw;
4075 struct mem_cgroup *mi;
4078 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4080 mem_cgroup_flush_stats();
4082 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4085 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4087 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4088 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
4089 nr * memcg_page_state_unit(memcg1_stats[i]));
4092 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4093 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4094 memcg_events_local(memcg, memcg1_events[i]));
4096 for (i = 0; i < NR_LRU_LISTS; i++)
4097 seq_printf(m, "%s %lu\n", lru_list_name(i),
4098 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4101 /* Hierarchical information */
4102 memory = memsw = PAGE_COUNTER_MAX;
4103 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4104 memory = min(memory, READ_ONCE(mi->memory.max));
4105 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4107 seq_printf(m, "hierarchical_memory_limit %llu\n",
4108 (u64)memory * PAGE_SIZE);
4109 if (do_memsw_account())
4110 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4111 (u64)memsw * PAGE_SIZE);
4113 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4116 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4118 nr = memcg_page_state(memcg, memcg1_stats[i]);
4119 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4120 (u64)nr * memcg_page_state_unit(memcg1_stats[i]));
4123 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4124 seq_printf(m, "total_%s %llu\n",
4125 vm_event_name(memcg1_events[i]),
4126 (u64)memcg_events(memcg, memcg1_events[i]));
4128 for (i = 0; i < NR_LRU_LISTS; i++)
4129 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4130 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4133 #ifdef CONFIG_DEBUG_VM
4136 struct mem_cgroup_per_node *mz;
4137 unsigned long anon_cost = 0;
4138 unsigned long file_cost = 0;
4140 for_each_online_pgdat(pgdat) {
4141 mz = memcg->nodeinfo[pgdat->node_id];
4143 anon_cost += mz->lruvec.anon_cost;
4144 file_cost += mz->lruvec.file_cost;
4146 seq_printf(m, "anon_cost %lu\n", anon_cost);
4147 seq_printf(m, "file_cost %lu\n", file_cost);
4154 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4157 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4159 return mem_cgroup_swappiness(memcg);
4162 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4163 struct cftype *cft, u64 val)
4165 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4170 if (!mem_cgroup_is_root(memcg))
4171 memcg->swappiness = val;
4173 vm_swappiness = val;
4178 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4180 struct mem_cgroup_threshold_ary *t;
4181 unsigned long usage;
4186 t = rcu_dereference(memcg->thresholds.primary);
4188 t = rcu_dereference(memcg->memsw_thresholds.primary);
4193 usage = mem_cgroup_usage(memcg, swap);
4196 * current_threshold points to threshold just below or equal to usage.
4197 * If it's not true, a threshold was crossed after last
4198 * call of __mem_cgroup_threshold().
4200 i = t->current_threshold;
4203 * Iterate backward over array of thresholds starting from
4204 * current_threshold and check if a threshold is crossed.
4205 * If none of thresholds below usage is crossed, we read
4206 * only one element of the array here.
4208 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4209 eventfd_signal(t->entries[i].eventfd, 1);
4211 /* i = current_threshold + 1 */
4215 * Iterate forward over array of thresholds starting from
4216 * current_threshold+1 and check if a threshold is crossed.
4217 * If none of thresholds above usage is crossed, we read
4218 * only one element of the array here.
4220 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4221 eventfd_signal(t->entries[i].eventfd, 1);
4223 /* Update current_threshold */
4224 t->current_threshold = i - 1;
4229 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4232 __mem_cgroup_threshold(memcg, false);
4233 if (do_memsw_account())
4234 __mem_cgroup_threshold(memcg, true);
4236 memcg = parent_mem_cgroup(memcg);
4240 static int compare_thresholds(const void *a, const void *b)
4242 const struct mem_cgroup_threshold *_a = a;
4243 const struct mem_cgroup_threshold *_b = b;
4245 if (_a->threshold > _b->threshold)
4248 if (_a->threshold < _b->threshold)
4254 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4256 struct mem_cgroup_eventfd_list *ev;
4258 spin_lock(&memcg_oom_lock);
4260 list_for_each_entry(ev, &memcg->oom_notify, list)
4261 eventfd_signal(ev->eventfd, 1);
4263 spin_unlock(&memcg_oom_lock);
4267 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4269 struct mem_cgroup *iter;
4271 for_each_mem_cgroup_tree(iter, memcg)
4272 mem_cgroup_oom_notify_cb(iter);
4275 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4276 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4278 struct mem_cgroup_thresholds *thresholds;
4279 struct mem_cgroup_threshold_ary *new;
4280 unsigned long threshold;
4281 unsigned long usage;
4284 ret = page_counter_memparse(args, "-1", &threshold);
4288 mutex_lock(&memcg->thresholds_lock);
4291 thresholds = &memcg->thresholds;
4292 usage = mem_cgroup_usage(memcg, false);
4293 } else if (type == _MEMSWAP) {
4294 thresholds = &memcg->memsw_thresholds;
4295 usage = mem_cgroup_usage(memcg, true);
4299 /* Check if a threshold crossed before adding a new one */
4300 if (thresholds->primary)
4301 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4303 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4305 /* Allocate memory for new array of thresholds */
4306 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4313 /* Copy thresholds (if any) to new array */
4314 if (thresholds->primary)
4315 memcpy(new->entries, thresholds->primary->entries,
4316 flex_array_size(new, entries, size - 1));
4318 /* Add new threshold */
4319 new->entries[size - 1].eventfd = eventfd;
4320 new->entries[size - 1].threshold = threshold;
4322 /* Sort thresholds. Registering of new threshold isn't time-critical */
4323 sort(new->entries, size, sizeof(*new->entries),
4324 compare_thresholds, NULL);
4326 /* Find current threshold */
4327 new->current_threshold = -1;
4328 for (i = 0; i < size; i++) {
4329 if (new->entries[i].threshold <= usage) {
4331 * new->current_threshold will not be used until
4332 * rcu_assign_pointer(), so it's safe to increment
4335 ++new->current_threshold;
4340 /* Free old spare buffer and save old primary buffer as spare */
4341 kfree(thresholds->spare);
4342 thresholds->spare = thresholds->primary;
4344 rcu_assign_pointer(thresholds->primary, new);
4346 /* To be sure that nobody uses thresholds */
4350 mutex_unlock(&memcg->thresholds_lock);
4355 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4356 struct eventfd_ctx *eventfd, const char *args)
4358 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4361 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4362 struct eventfd_ctx *eventfd, const char *args)
4364 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4367 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4368 struct eventfd_ctx *eventfd, enum res_type type)
4370 struct mem_cgroup_thresholds *thresholds;
4371 struct mem_cgroup_threshold_ary *new;
4372 unsigned long usage;
4373 int i, j, size, entries;
4375 mutex_lock(&memcg->thresholds_lock);
4378 thresholds = &memcg->thresholds;
4379 usage = mem_cgroup_usage(memcg, false);
4380 } else if (type == _MEMSWAP) {
4381 thresholds = &memcg->memsw_thresholds;
4382 usage = mem_cgroup_usage(memcg, true);
4386 if (!thresholds->primary)
4389 /* Check if a threshold crossed before removing */
4390 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4392 /* Calculate new number of threshold */
4394 for (i = 0; i < thresholds->primary->size; i++) {
4395 if (thresholds->primary->entries[i].eventfd != eventfd)
4401 new = thresholds->spare;
4403 /* If no items related to eventfd have been cleared, nothing to do */
4407 /* Set thresholds array to NULL if we don't have thresholds */
4416 /* Copy thresholds and find current threshold */
4417 new->current_threshold = -1;
4418 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4419 if (thresholds->primary->entries[i].eventfd == eventfd)
4422 new->entries[j] = thresholds->primary->entries[i];
4423 if (new->entries[j].threshold <= usage) {
4425 * new->current_threshold will not be used
4426 * until rcu_assign_pointer(), so it's safe to increment
4429 ++new->current_threshold;
4435 /* Swap primary and spare array */
4436 thresholds->spare = thresholds->primary;
4438 rcu_assign_pointer(thresholds->primary, new);
4440 /* To be sure that nobody uses thresholds */
4443 /* If all events are unregistered, free the spare array */
4445 kfree(thresholds->spare);
4446 thresholds->spare = NULL;
4449 mutex_unlock(&memcg->thresholds_lock);
4452 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4453 struct eventfd_ctx *eventfd)
4455 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4458 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4459 struct eventfd_ctx *eventfd)
4461 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4464 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4465 struct eventfd_ctx *eventfd, const char *args)
4467 struct mem_cgroup_eventfd_list *event;
4469 event = kmalloc(sizeof(*event), GFP_KERNEL);
4473 spin_lock(&memcg_oom_lock);
4475 event->eventfd = eventfd;
4476 list_add(&event->list, &memcg->oom_notify);
4478 /* already in OOM ? */
4479 if (memcg->under_oom)
4480 eventfd_signal(eventfd, 1);
4481 spin_unlock(&memcg_oom_lock);
4486 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4487 struct eventfd_ctx *eventfd)
4489 struct mem_cgroup_eventfd_list *ev, *tmp;
4491 spin_lock(&memcg_oom_lock);
4493 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4494 if (ev->eventfd == eventfd) {
4495 list_del(&ev->list);
4500 spin_unlock(&memcg_oom_lock);
4503 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4505 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4507 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4508 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4509 seq_printf(sf, "oom_kill %lu\n",
4510 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4514 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4515 struct cftype *cft, u64 val)
4517 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4519 /* cannot set to root cgroup and only 0 and 1 are allowed */
4520 if (mem_cgroup_is_root(memcg) || !((val == 0) || (val == 1)))
4523 memcg->oom_kill_disable = val;
4525 memcg_oom_recover(memcg);
4530 #ifdef CONFIG_CGROUP_WRITEBACK
4532 #include <trace/events/writeback.h>
4534 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4536 return wb_domain_init(&memcg->cgwb_domain, gfp);
4539 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4541 wb_domain_exit(&memcg->cgwb_domain);
4544 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4546 wb_domain_size_changed(&memcg->cgwb_domain);
4549 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4551 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4553 if (!memcg->css.parent)
4556 return &memcg->cgwb_domain;
4560 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4561 * @wb: bdi_writeback in question
4562 * @pfilepages: out parameter for number of file pages
4563 * @pheadroom: out parameter for number of allocatable pages according to memcg
4564 * @pdirty: out parameter for number of dirty pages
4565 * @pwriteback: out parameter for number of pages under writeback
4567 * Determine the numbers of file, headroom, dirty, and writeback pages in
4568 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4569 * is a bit more involved.
4571 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4572 * headroom is calculated as the lowest headroom of itself and the
4573 * ancestors. Note that this doesn't consider the actual amount of
4574 * available memory in the system. The caller should further cap
4575 * *@pheadroom accordingly.
4577 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4578 unsigned long *pheadroom, unsigned long *pdirty,
4579 unsigned long *pwriteback)
4581 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4582 struct mem_cgroup *parent;
4584 mem_cgroup_flush_stats();
4586 *pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
4587 *pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
4588 *pfilepages = memcg_page_state(memcg, NR_INACTIVE_FILE) +
4589 memcg_page_state(memcg, NR_ACTIVE_FILE);
4591 *pheadroom = PAGE_COUNTER_MAX;
4592 while ((parent = parent_mem_cgroup(memcg))) {
4593 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4594 READ_ONCE(memcg->memory.high));
4595 unsigned long used = page_counter_read(&memcg->memory);
4597 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4603 * Foreign dirty flushing
4605 * There's an inherent mismatch between memcg and writeback. The former
4606 * tracks ownership per-page while the latter per-inode. This was a
4607 * deliberate design decision because honoring per-page ownership in the
4608 * writeback path is complicated, may lead to higher CPU and IO overheads
4609 * and deemed unnecessary given that write-sharing an inode across
4610 * different cgroups isn't a common use-case.
4612 * Combined with inode majority-writer ownership switching, this works well
4613 * enough in most cases but there are some pathological cases. For
4614 * example, let's say there are two cgroups A and B which keep writing to
4615 * different but confined parts of the same inode. B owns the inode and
4616 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4617 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4618 * triggering background writeback. A will be slowed down without a way to
4619 * make writeback of the dirty pages happen.
4621 * Conditions like the above can lead to a cgroup getting repeatedly and
4622 * severely throttled after making some progress after each
4623 * dirty_expire_interval while the underlying IO device is almost
4626 * Solving this problem completely requires matching the ownership tracking
4627 * granularities between memcg and writeback in either direction. However,
4628 * the more egregious behaviors can be avoided by simply remembering the
4629 * most recent foreign dirtying events and initiating remote flushes on
4630 * them when local writeback isn't enough to keep the memory clean enough.
4632 * The following two functions implement such mechanism. When a foreign
4633 * page - a page whose memcg and writeback ownerships don't match - is
4634 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4635 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4636 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4637 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4638 * foreign bdi_writebacks which haven't expired. Both the numbers of
4639 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4640 * limited to MEMCG_CGWB_FRN_CNT.
4642 * The mechanism only remembers IDs and doesn't hold any object references.
4643 * As being wrong occasionally doesn't matter, updates and accesses to the
4644 * records are lockless and racy.
4646 void mem_cgroup_track_foreign_dirty_slowpath(struct folio *folio,
4647 struct bdi_writeback *wb)
4649 struct mem_cgroup *memcg = folio_memcg(folio);
4650 struct memcg_cgwb_frn *frn;
4651 u64 now = get_jiffies_64();
4652 u64 oldest_at = now;
4656 trace_track_foreign_dirty(folio, wb);
4659 * Pick the slot to use. If there is already a slot for @wb, keep
4660 * using it. If not replace the oldest one which isn't being
4663 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4664 frn = &memcg->cgwb_frn[i];
4665 if (frn->bdi_id == wb->bdi->id &&
4666 frn->memcg_id == wb->memcg_css->id)
4668 if (time_before64(frn->at, oldest_at) &&
4669 atomic_read(&frn->done.cnt) == 1) {
4671 oldest_at = frn->at;
4675 if (i < MEMCG_CGWB_FRN_CNT) {
4677 * Re-using an existing one. Update timestamp lazily to
4678 * avoid making the cacheline hot. We want them to be
4679 * reasonably up-to-date and significantly shorter than
4680 * dirty_expire_interval as that's what expires the record.
4681 * Use the shorter of 1s and dirty_expire_interval / 8.
4683 unsigned long update_intv =
4684 min_t(unsigned long, HZ,
4685 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4687 if (time_before64(frn->at, now - update_intv))
4689 } else if (oldest >= 0) {
4690 /* replace the oldest free one */
4691 frn = &memcg->cgwb_frn[oldest];
4692 frn->bdi_id = wb->bdi->id;
4693 frn->memcg_id = wb->memcg_css->id;
4698 /* issue foreign writeback flushes for recorded foreign dirtying events */
4699 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4701 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4702 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4703 u64 now = jiffies_64;
4706 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4707 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4710 * If the record is older than dirty_expire_interval,
4711 * writeback on it has already started. No need to kick it
4712 * off again. Also, don't start a new one if there's
4713 * already one in flight.
4715 if (time_after64(frn->at, now - intv) &&
4716 atomic_read(&frn->done.cnt) == 1) {
4718 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4719 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id,
4720 WB_REASON_FOREIGN_FLUSH,
4726 #else /* CONFIG_CGROUP_WRITEBACK */
4728 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4733 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4737 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4741 #endif /* CONFIG_CGROUP_WRITEBACK */
4744 * DO NOT USE IN NEW FILES.
4746 * "cgroup.event_control" implementation.
4748 * This is way over-engineered. It tries to support fully configurable
4749 * events for each user. Such level of flexibility is completely
4750 * unnecessary especially in the light of the planned unified hierarchy.
4752 * Please deprecate this and replace with something simpler if at all
4757 * Unregister event and free resources.
4759 * Gets called from workqueue.
4761 static void memcg_event_remove(struct work_struct *work)
4763 struct mem_cgroup_event *event =
4764 container_of(work, struct mem_cgroup_event, remove);
4765 struct mem_cgroup *memcg = event->memcg;
4767 remove_wait_queue(event->wqh, &event->wait);
4769 event->unregister_event(memcg, event->eventfd);
4771 /* Notify userspace the event is going away. */
4772 eventfd_signal(event->eventfd, 1);
4774 eventfd_ctx_put(event->eventfd);
4776 css_put(&memcg->css);
4780 * Gets called on EPOLLHUP on eventfd when user closes it.
4782 * Called with wqh->lock held and interrupts disabled.
4784 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4785 int sync, void *key)
4787 struct mem_cgroup_event *event =
4788 container_of(wait, struct mem_cgroup_event, wait);
4789 struct mem_cgroup *memcg = event->memcg;
4790 __poll_t flags = key_to_poll(key);
4792 if (flags & EPOLLHUP) {
4794 * If the event has been detached at cgroup removal, we
4795 * can simply return knowing the other side will cleanup
4798 * We can't race against event freeing since the other
4799 * side will require wqh->lock via remove_wait_queue(),
4802 spin_lock(&memcg->event_list_lock);
4803 if (!list_empty(&event->list)) {
4804 list_del_init(&event->list);
4806 * We are in atomic context, but cgroup_event_remove()
4807 * may sleep, so we have to call it in workqueue.
4809 schedule_work(&event->remove);
4811 spin_unlock(&memcg->event_list_lock);
4817 static void memcg_event_ptable_queue_proc(struct file *file,
4818 wait_queue_head_t *wqh, poll_table *pt)
4820 struct mem_cgroup_event *event =
4821 container_of(pt, struct mem_cgroup_event, pt);
4824 add_wait_queue(wqh, &event->wait);
4828 * DO NOT USE IN NEW FILES.
4830 * Parse input and register new cgroup event handler.
4832 * Input must be in format '<event_fd> <control_fd> <args>'.
4833 * Interpretation of args is defined by control file implementation.
4835 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4836 char *buf, size_t nbytes, loff_t off)
4838 struct cgroup_subsys_state *css = of_css(of);
4839 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4840 struct mem_cgroup_event *event;
4841 struct cgroup_subsys_state *cfile_css;
4842 unsigned int efd, cfd;
4845 struct dentry *cdentry;
4850 if (IS_ENABLED(CONFIG_PREEMPT_RT))
4853 buf = strstrip(buf);
4855 efd = simple_strtoul(buf, &endp, 10);
4860 cfd = simple_strtoul(buf, &endp, 10);
4861 if ((*endp != ' ') && (*endp != '\0'))
4865 event = kzalloc(sizeof(*event), GFP_KERNEL);
4869 event->memcg = memcg;
4870 INIT_LIST_HEAD(&event->list);
4871 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4872 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4873 INIT_WORK(&event->remove, memcg_event_remove);
4881 event->eventfd = eventfd_ctx_fileget(efile.file);
4882 if (IS_ERR(event->eventfd)) {
4883 ret = PTR_ERR(event->eventfd);
4890 goto out_put_eventfd;
4893 /* the process need read permission on control file */
4894 /* AV: shouldn't we check that it's been opened for read instead? */
4895 ret = file_permission(cfile.file, MAY_READ);
4900 * The control file must be a regular cgroup1 file. As a regular cgroup
4901 * file can't be renamed, it's safe to access its name afterwards.
4903 cdentry = cfile.file->f_path.dentry;
4904 if (cdentry->d_sb->s_type != &cgroup_fs_type || !d_is_reg(cdentry)) {
4910 * Determine the event callbacks and set them in @event. This used
4911 * to be done via struct cftype but cgroup core no longer knows
4912 * about these events. The following is crude but the whole thing
4913 * is for compatibility anyway.
4915 * DO NOT ADD NEW FILES.
4917 name = cdentry->d_name.name;
4919 if (!strcmp(name, "memory.usage_in_bytes")) {
4920 event->register_event = mem_cgroup_usage_register_event;
4921 event->unregister_event = mem_cgroup_usage_unregister_event;
4922 } else if (!strcmp(name, "memory.oom_control")) {
4923 event->register_event = mem_cgroup_oom_register_event;
4924 event->unregister_event = mem_cgroup_oom_unregister_event;
4925 } else if (!strcmp(name, "memory.pressure_level")) {
4926 event->register_event = vmpressure_register_event;
4927 event->unregister_event = vmpressure_unregister_event;
4928 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4929 event->register_event = memsw_cgroup_usage_register_event;
4930 event->unregister_event = memsw_cgroup_usage_unregister_event;
4937 * Verify @cfile should belong to @css. Also, remaining events are
4938 * automatically removed on cgroup destruction but the removal is
4939 * asynchronous, so take an extra ref on @css.
4941 cfile_css = css_tryget_online_from_dir(cdentry->d_parent,
4942 &memory_cgrp_subsys);
4944 if (IS_ERR(cfile_css))
4946 if (cfile_css != css) {
4951 ret = event->register_event(memcg, event->eventfd, buf);
4955 vfs_poll(efile.file, &event->pt);
4957 spin_lock_irq(&memcg->event_list_lock);
4958 list_add(&event->list, &memcg->event_list);
4959 spin_unlock_irq(&memcg->event_list_lock);
4971 eventfd_ctx_put(event->eventfd);
4980 #if defined(CONFIG_MEMCG_KMEM) && (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
4981 static int mem_cgroup_slab_show(struct seq_file *m, void *p)
4985 * Please, take a look at tools/cgroup/memcg_slabinfo.py .
4991 static struct cftype mem_cgroup_legacy_files[] = {
4993 .name = "usage_in_bytes",
4994 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4995 .read_u64 = mem_cgroup_read_u64,
4998 .name = "max_usage_in_bytes",
4999 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5000 .write = mem_cgroup_reset,
5001 .read_u64 = mem_cgroup_read_u64,
5004 .name = "limit_in_bytes",
5005 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5006 .write = mem_cgroup_write,
5007 .read_u64 = mem_cgroup_read_u64,
5010 .name = "soft_limit_in_bytes",
5011 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5012 .write = mem_cgroup_write,
5013 .read_u64 = mem_cgroup_read_u64,
5017 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5018 .write = mem_cgroup_reset,
5019 .read_u64 = mem_cgroup_read_u64,
5023 .seq_show = memcg_stat_show,
5026 .name = "force_empty",
5027 .write = mem_cgroup_force_empty_write,
5030 .name = "use_hierarchy",
5031 .write_u64 = mem_cgroup_hierarchy_write,
5032 .read_u64 = mem_cgroup_hierarchy_read,
5035 .name = "cgroup.event_control", /* XXX: for compat */
5036 .write = memcg_write_event_control,
5037 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
5040 .name = "swappiness",
5041 .read_u64 = mem_cgroup_swappiness_read,
5042 .write_u64 = mem_cgroup_swappiness_write,
5045 .name = "move_charge_at_immigrate",
5046 .read_u64 = mem_cgroup_move_charge_read,
5047 .write_u64 = mem_cgroup_move_charge_write,
5050 .name = "oom_control",
5051 .seq_show = mem_cgroup_oom_control_read,
5052 .write_u64 = mem_cgroup_oom_control_write,
5055 .name = "pressure_level",
5059 .name = "numa_stat",
5060 .seq_show = memcg_numa_stat_show,
5064 .name = "kmem.limit_in_bytes",
5065 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5066 .write = mem_cgroup_write,
5067 .read_u64 = mem_cgroup_read_u64,
5070 .name = "kmem.usage_in_bytes",
5071 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5072 .read_u64 = mem_cgroup_read_u64,
5075 .name = "kmem.failcnt",
5076 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5077 .write = mem_cgroup_reset,
5078 .read_u64 = mem_cgroup_read_u64,
5081 .name = "kmem.max_usage_in_bytes",
5082 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5083 .write = mem_cgroup_reset,
5084 .read_u64 = mem_cgroup_read_u64,
5086 #if defined(CONFIG_MEMCG_KMEM) && \
5087 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5089 .name = "kmem.slabinfo",
5090 .seq_show = mem_cgroup_slab_show,
5094 .name = "kmem.tcp.limit_in_bytes",
5095 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5096 .write = mem_cgroup_write,
5097 .read_u64 = mem_cgroup_read_u64,
5100 .name = "kmem.tcp.usage_in_bytes",
5101 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5102 .read_u64 = mem_cgroup_read_u64,
5105 .name = "kmem.tcp.failcnt",
5106 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5107 .write = mem_cgroup_reset,
5108 .read_u64 = mem_cgroup_read_u64,
5111 .name = "kmem.tcp.max_usage_in_bytes",
5112 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5113 .write = mem_cgroup_reset,
5114 .read_u64 = mem_cgroup_read_u64,
5116 { }, /* terminate */
5120 * Private memory cgroup IDR
5122 * Swap-out records and page cache shadow entries need to store memcg
5123 * references in constrained space, so we maintain an ID space that is
5124 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5125 * memory-controlled cgroups to 64k.
5127 * However, there usually are many references to the offline CSS after
5128 * the cgroup has been destroyed, such as page cache or reclaimable
5129 * slab objects, that don't need to hang on to the ID. We want to keep
5130 * those dead CSS from occupying IDs, or we might quickly exhaust the
5131 * relatively small ID space and prevent the creation of new cgroups
5132 * even when there are much fewer than 64k cgroups - possibly none.
5134 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5135 * be freed and recycled when it's no longer needed, which is usually
5136 * when the CSS is offlined.
5138 * The only exception to that are records of swapped out tmpfs/shmem
5139 * pages that need to be attributed to live ancestors on swapin. But
5140 * those references are manageable from userspace.
5143 static DEFINE_IDR(mem_cgroup_idr);
5145 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5147 if (memcg->id.id > 0) {
5148 idr_remove(&mem_cgroup_idr, memcg->id.id);
5153 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5156 refcount_add(n, &memcg->id.ref);
5159 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5161 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5162 mem_cgroup_id_remove(memcg);
5164 /* Memcg ID pins CSS */
5165 css_put(&memcg->css);
5169 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5171 mem_cgroup_id_put_many(memcg, 1);
5175 * mem_cgroup_from_id - look up a memcg from a memcg id
5176 * @id: the memcg id to look up
5178 * Caller must hold rcu_read_lock().
5180 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5182 WARN_ON_ONCE(!rcu_read_lock_held());
5183 return idr_find(&mem_cgroup_idr, id);
5186 #ifdef CONFIG_SHRINKER_DEBUG
5187 struct mem_cgroup *mem_cgroup_get_from_ino(unsigned long ino)
5189 struct cgroup *cgrp;
5190 struct cgroup_subsys_state *css;
5191 struct mem_cgroup *memcg;
5193 cgrp = cgroup_get_from_id(ino);
5195 return ERR_CAST(cgrp);
5197 css = cgroup_get_e_css(cgrp, &memory_cgrp_subsys);
5199 memcg = container_of(css, struct mem_cgroup, css);
5201 memcg = ERR_PTR(-ENOENT);
5209 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5211 struct mem_cgroup_per_node *pn;
5213 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, node);
5217 pn->lruvec_stats_percpu = alloc_percpu_gfp(struct lruvec_stats_percpu,
5218 GFP_KERNEL_ACCOUNT);
5219 if (!pn->lruvec_stats_percpu) {
5224 lruvec_init(&pn->lruvec);
5227 memcg->nodeinfo[node] = pn;
5231 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5233 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5238 free_percpu(pn->lruvec_stats_percpu);
5242 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5247 free_mem_cgroup_per_node_info(memcg, node);
5248 kfree(memcg->vmstats);
5249 free_percpu(memcg->vmstats_percpu);
5253 static void mem_cgroup_free(struct mem_cgroup *memcg)
5255 lru_gen_exit_memcg(memcg);
5256 memcg_wb_domain_exit(memcg);
5257 __mem_cgroup_free(memcg);
5260 static struct mem_cgroup *mem_cgroup_alloc(void)
5262 struct mem_cgroup *memcg;
5264 int __maybe_unused i;
5265 long error = -ENOMEM;
5267 memcg = kzalloc(struct_size(memcg, nodeinfo, nr_node_ids), GFP_KERNEL);
5269 return ERR_PTR(error);
5271 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5272 1, MEM_CGROUP_ID_MAX + 1, GFP_KERNEL);
5273 if (memcg->id.id < 0) {
5274 error = memcg->id.id;
5278 memcg->vmstats = kzalloc(sizeof(struct memcg_vmstats), GFP_KERNEL);
5279 if (!memcg->vmstats)
5282 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5283 GFP_KERNEL_ACCOUNT);
5284 if (!memcg->vmstats_percpu)
5288 if (alloc_mem_cgroup_per_node_info(memcg, node))
5291 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5294 INIT_WORK(&memcg->high_work, high_work_func);
5295 INIT_LIST_HEAD(&memcg->oom_notify);
5296 mutex_init(&memcg->thresholds_lock);
5297 spin_lock_init(&memcg->move_lock);
5298 vmpressure_init(&memcg->vmpressure);
5299 INIT_LIST_HEAD(&memcg->event_list);
5300 spin_lock_init(&memcg->event_list_lock);
5301 memcg->socket_pressure = jiffies;
5302 #ifdef CONFIG_MEMCG_KMEM
5303 memcg->kmemcg_id = -1;
5304 INIT_LIST_HEAD(&memcg->objcg_list);
5306 #ifdef CONFIG_CGROUP_WRITEBACK
5307 INIT_LIST_HEAD(&memcg->cgwb_list);
5308 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5309 memcg->cgwb_frn[i].done =
5310 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5312 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5313 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5314 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5315 memcg->deferred_split_queue.split_queue_len = 0;
5317 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5318 lru_gen_init_memcg(memcg);
5321 mem_cgroup_id_remove(memcg);
5322 __mem_cgroup_free(memcg);
5323 return ERR_PTR(error);
5326 static struct cgroup_subsys_state * __ref
5327 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5329 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5330 struct mem_cgroup *memcg, *old_memcg;
5332 old_memcg = set_active_memcg(parent);
5333 memcg = mem_cgroup_alloc();
5334 set_active_memcg(old_memcg);
5336 return ERR_CAST(memcg);
5338 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5339 memcg->soft_limit = PAGE_COUNTER_MAX;
5340 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
5341 memcg->zswap_max = PAGE_COUNTER_MAX;
5343 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5345 memcg->swappiness = mem_cgroup_swappiness(parent);
5346 memcg->oom_kill_disable = parent->oom_kill_disable;
5348 page_counter_init(&memcg->memory, &parent->memory);
5349 page_counter_init(&memcg->swap, &parent->swap);
5350 page_counter_init(&memcg->kmem, &parent->kmem);
5351 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5353 init_memcg_events();
5354 page_counter_init(&memcg->memory, NULL);
5355 page_counter_init(&memcg->swap, NULL);
5356 page_counter_init(&memcg->kmem, NULL);
5357 page_counter_init(&memcg->tcpmem, NULL);
5359 root_mem_cgroup = memcg;
5363 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5364 static_branch_inc(&memcg_sockets_enabled_key);
5366 #if defined(CONFIG_MEMCG_KMEM)
5367 if (!cgroup_memory_nobpf)
5368 static_branch_inc(&memcg_bpf_enabled_key);
5374 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5376 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5378 if (memcg_online_kmem(memcg))
5382 * A memcg must be visible for expand_shrinker_info()
5383 * by the time the maps are allocated. So, we allocate maps
5384 * here, when for_each_mem_cgroup() can't skip it.
5386 if (alloc_shrinker_info(memcg))
5389 /* Online state pins memcg ID, memcg ID pins CSS */
5390 refcount_set(&memcg->id.ref, 1);
5393 if (unlikely(mem_cgroup_is_root(memcg)))
5394 queue_delayed_work(system_unbound_wq, &stats_flush_dwork,
5398 memcg_offline_kmem(memcg);
5400 mem_cgroup_id_remove(memcg);
5404 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5406 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5407 struct mem_cgroup_event *event, *tmp;
5410 * Unregister events and notify userspace.
5411 * Notify userspace about cgroup removing only after rmdir of cgroup
5412 * directory to avoid race between userspace and kernelspace.
5414 spin_lock_irq(&memcg->event_list_lock);
5415 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5416 list_del_init(&event->list);
5417 schedule_work(&event->remove);
5419 spin_unlock_irq(&memcg->event_list_lock);
5421 page_counter_set_min(&memcg->memory, 0);
5422 page_counter_set_low(&memcg->memory, 0);
5424 memcg_offline_kmem(memcg);
5425 reparent_shrinker_deferred(memcg);
5426 wb_memcg_offline(memcg);
5428 drain_all_stock(memcg);
5430 mem_cgroup_id_put(memcg);
5433 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5435 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5437 invalidate_reclaim_iterators(memcg);
5440 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5442 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5443 int __maybe_unused i;
5445 #ifdef CONFIG_CGROUP_WRITEBACK
5446 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5447 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5449 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5450 static_branch_dec(&memcg_sockets_enabled_key);
5452 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5453 static_branch_dec(&memcg_sockets_enabled_key);
5455 #if defined(CONFIG_MEMCG_KMEM)
5456 if (!cgroup_memory_nobpf)
5457 static_branch_dec(&memcg_bpf_enabled_key);
5460 vmpressure_cleanup(&memcg->vmpressure);
5461 cancel_work_sync(&memcg->high_work);
5462 mem_cgroup_remove_from_trees(memcg);
5463 free_shrinker_info(memcg);
5464 mem_cgroup_free(memcg);
5468 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5469 * @css: the target css
5471 * Reset the states of the mem_cgroup associated with @css. This is
5472 * invoked when the userland requests disabling on the default hierarchy
5473 * but the memcg is pinned through dependency. The memcg should stop
5474 * applying policies and should revert to the vanilla state as it may be
5475 * made visible again.
5477 * The current implementation only resets the essential configurations.
5478 * This needs to be expanded to cover all the visible parts.
5480 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5482 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5484 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5485 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5486 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5487 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5488 page_counter_set_min(&memcg->memory, 0);
5489 page_counter_set_low(&memcg->memory, 0);
5490 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5491 memcg->soft_limit = PAGE_COUNTER_MAX;
5492 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5493 memcg_wb_domain_size_changed(memcg);
5496 static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state *css, int cpu)
5498 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5499 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5500 struct memcg_vmstats_percpu *statc;
5504 statc = per_cpu_ptr(memcg->vmstats_percpu, cpu);
5506 for (i = 0; i < MEMCG_NR_STAT; i++) {
5508 * Collect the aggregated propagation counts of groups
5509 * below us. We're in a per-cpu loop here and this is
5510 * a global counter, so the first cycle will get them.
5512 delta = memcg->vmstats->state_pending[i];
5514 memcg->vmstats->state_pending[i] = 0;
5516 /* Add CPU changes on this level since the last flush */
5517 v = READ_ONCE(statc->state[i]);
5518 if (v != statc->state_prev[i]) {
5519 delta += v - statc->state_prev[i];
5520 statc->state_prev[i] = v;
5526 /* Aggregate counts on this level and propagate upwards */
5527 memcg->vmstats->state[i] += delta;
5529 parent->vmstats->state_pending[i] += delta;
5532 for (i = 0; i < NR_MEMCG_EVENTS; i++) {
5533 delta = memcg->vmstats->events_pending[i];
5535 memcg->vmstats->events_pending[i] = 0;
5537 v = READ_ONCE(statc->events[i]);
5538 if (v != statc->events_prev[i]) {
5539 delta += v - statc->events_prev[i];
5540 statc->events_prev[i] = v;
5546 memcg->vmstats->events[i] += delta;
5548 parent->vmstats->events_pending[i] += delta;
5551 for_each_node_state(nid, N_MEMORY) {
5552 struct mem_cgroup_per_node *pn = memcg->nodeinfo[nid];
5553 struct mem_cgroup_per_node *ppn = NULL;
5554 struct lruvec_stats_percpu *lstatc;
5557 ppn = parent->nodeinfo[nid];
5559 lstatc = per_cpu_ptr(pn->lruvec_stats_percpu, cpu);
5561 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
5562 delta = pn->lruvec_stats.state_pending[i];
5564 pn->lruvec_stats.state_pending[i] = 0;
5566 v = READ_ONCE(lstatc->state[i]);
5567 if (v != lstatc->state_prev[i]) {
5568 delta += v - lstatc->state_prev[i];
5569 lstatc->state_prev[i] = v;
5575 pn->lruvec_stats.state[i] += delta;
5577 ppn->lruvec_stats.state_pending[i] += delta;
5583 /* Handlers for move charge at task migration. */
5584 static int mem_cgroup_do_precharge(unsigned long count)
5588 /* Try a single bulk charge without reclaim first, kswapd may wake */
5589 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5591 mc.precharge += count;
5595 /* Try charges one by one with reclaim, but do not retry */
5597 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5611 enum mc_target_type {
5618 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5619 unsigned long addr, pte_t ptent)
5621 struct page *page = vm_normal_page(vma, addr, ptent);
5623 if (!page || !page_mapped(page))
5625 if (PageAnon(page)) {
5626 if (!(mc.flags & MOVE_ANON))
5629 if (!(mc.flags & MOVE_FILE))
5632 if (!get_page_unless_zero(page))
5638 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5639 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5640 pte_t ptent, swp_entry_t *entry)
5642 struct page *page = NULL;
5643 swp_entry_t ent = pte_to_swp_entry(ptent);
5645 if (!(mc.flags & MOVE_ANON))
5649 * Handle device private pages that are not accessible by the CPU, but
5650 * stored as special swap entries in the page table.
5652 if (is_device_private_entry(ent)) {
5653 page = pfn_swap_entry_to_page(ent);
5654 if (!get_page_unless_zero(page))
5659 if (non_swap_entry(ent))
5663 * Because swap_cache_get_folio() updates some statistics counter,
5664 * we call find_get_page() with swapper_space directly.
5666 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5667 entry->val = ent.val;
5672 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5673 pte_t ptent, swp_entry_t *entry)
5679 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5680 unsigned long addr, pte_t ptent)
5682 unsigned long index;
5683 struct folio *folio;
5685 if (!vma->vm_file) /* anonymous vma */
5687 if (!(mc.flags & MOVE_FILE))
5690 /* folio is moved even if it's not RSS of this task(page-faulted). */
5691 /* shmem/tmpfs may report page out on swap: account for that too. */
5692 index = linear_page_index(vma, addr);
5693 folio = filemap_get_incore_folio(vma->vm_file->f_mapping, index);
5696 return folio_file_page(folio, index);
5700 * mem_cgroup_move_account - move account of the page
5702 * @compound: charge the page as compound or small page
5703 * @from: mem_cgroup which the page is moved from.
5704 * @to: mem_cgroup which the page is moved to. @from != @to.
5706 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5708 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5711 static int mem_cgroup_move_account(struct page *page,
5713 struct mem_cgroup *from,
5714 struct mem_cgroup *to)
5716 struct folio *folio = page_folio(page);
5717 struct lruvec *from_vec, *to_vec;
5718 struct pglist_data *pgdat;
5719 unsigned int nr_pages = compound ? folio_nr_pages(folio) : 1;
5722 VM_BUG_ON(from == to);
5723 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
5724 VM_BUG_ON(compound && !folio_test_large(folio));
5727 * Prevent mem_cgroup_migrate() from looking at
5728 * page's memory cgroup of its source page while we change it.
5731 if (!folio_trylock(folio))
5735 if (folio_memcg(folio) != from)
5738 pgdat = folio_pgdat(folio);
5739 from_vec = mem_cgroup_lruvec(from, pgdat);
5740 to_vec = mem_cgroup_lruvec(to, pgdat);
5742 folio_memcg_lock(folio);
5744 if (folio_test_anon(folio)) {
5745 if (folio_mapped(folio)) {
5746 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5747 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5748 if (folio_test_transhuge(folio)) {
5749 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5751 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5756 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5757 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5759 if (folio_test_swapbacked(folio)) {
5760 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5761 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5764 if (folio_mapped(folio)) {
5765 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5766 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5769 if (folio_test_dirty(folio)) {
5770 struct address_space *mapping = folio_mapping(folio);
5772 if (mapping_can_writeback(mapping)) {
5773 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5775 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5782 if (folio_test_swapcache(folio)) {
5783 __mod_lruvec_state(from_vec, NR_SWAPCACHE, -nr_pages);
5784 __mod_lruvec_state(to_vec, NR_SWAPCACHE, nr_pages);
5787 if (folio_test_writeback(folio)) {
5788 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5789 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5793 * All state has been migrated, let's switch to the new memcg.
5795 * It is safe to change page's memcg here because the page
5796 * is referenced, charged, isolated, and locked: we can't race
5797 * with (un)charging, migration, LRU putback, or anything else
5798 * that would rely on a stable page's memory cgroup.
5800 * Note that lock_page_memcg is a memcg lock, not a page lock,
5801 * to save space. As soon as we switch page's memory cgroup to a
5802 * new memcg that isn't locked, the above state can change
5803 * concurrently again. Make sure we're truly done with it.
5808 css_put(&from->css);
5810 folio->memcg_data = (unsigned long)to;
5812 __folio_memcg_unlock(from);
5815 nid = folio_nid(folio);
5817 local_irq_disable();
5818 mem_cgroup_charge_statistics(to, nr_pages);
5819 memcg_check_events(to, nid);
5820 mem_cgroup_charge_statistics(from, -nr_pages);
5821 memcg_check_events(from, nid);
5824 folio_unlock(folio);
5830 * get_mctgt_type - get target type of moving charge
5831 * @vma: the vma the pte to be checked belongs
5832 * @addr: the address corresponding to the pte to be checked
5833 * @ptent: the pte to be checked
5834 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5837 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5838 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5839 * move charge. if @target is not NULL, the page is stored in target->page
5840 * with extra refcnt got(Callers should handle it).
5841 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5842 * target for charge migration. if @target is not NULL, the entry is stored
5844 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is device memory and
5845 * thus not on the lru.
5846 * For now we such page is charge like a regular page would be as for all
5847 * intent and purposes it is just special memory taking the place of a
5850 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5852 * Called with pte lock held.
5855 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5856 unsigned long addr, pte_t ptent, union mc_target *target)
5858 struct page *page = NULL;
5859 enum mc_target_type ret = MC_TARGET_NONE;
5860 swp_entry_t ent = { .val = 0 };
5862 if (pte_present(ptent))
5863 page = mc_handle_present_pte(vma, addr, ptent);
5864 else if (pte_none_mostly(ptent))
5866 * PTE markers should be treated as a none pte here, separated
5867 * from other swap handling below.
5869 page = mc_handle_file_pte(vma, addr, ptent);
5870 else if (is_swap_pte(ptent))
5871 page = mc_handle_swap_pte(vma, ptent, &ent);
5873 if (!page && !ent.val)
5877 * Do only loose check w/o serialization.
5878 * mem_cgroup_move_account() checks the page is valid or
5879 * not under LRU exclusion.
5881 if (page_memcg(page) == mc.from) {
5882 ret = MC_TARGET_PAGE;
5883 if (is_device_private_page(page) ||
5884 is_device_coherent_page(page))
5885 ret = MC_TARGET_DEVICE;
5887 target->page = page;
5889 if (!ret || !target)
5893 * There is a swap entry and a page doesn't exist or isn't charged.
5894 * But we cannot move a tail-page in a THP.
5896 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5897 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5898 ret = MC_TARGET_SWAP;
5905 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5907 * We don't consider PMD mapped swapping or file mapped pages because THP does
5908 * not support them for now.
5909 * Caller should make sure that pmd_trans_huge(pmd) is true.
5911 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5912 unsigned long addr, pmd_t pmd, union mc_target *target)
5914 struct page *page = NULL;
5915 enum mc_target_type ret = MC_TARGET_NONE;
5917 if (unlikely(is_swap_pmd(pmd))) {
5918 VM_BUG_ON(thp_migration_supported() &&
5919 !is_pmd_migration_entry(pmd));
5922 page = pmd_page(pmd);
5923 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5924 if (!(mc.flags & MOVE_ANON))
5926 if (page_memcg(page) == mc.from) {
5927 ret = MC_TARGET_PAGE;
5930 target->page = page;
5936 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5937 unsigned long addr, pmd_t pmd, union mc_target *target)
5939 return MC_TARGET_NONE;
5943 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5944 unsigned long addr, unsigned long end,
5945 struct mm_walk *walk)
5947 struct vm_area_struct *vma = walk->vma;
5951 ptl = pmd_trans_huge_lock(pmd, vma);
5954 * Note their can not be MC_TARGET_DEVICE for now as we do not
5955 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5956 * this might change.
5958 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5959 mc.precharge += HPAGE_PMD_NR;
5964 if (pmd_trans_unstable(pmd))
5966 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5967 for (; addr != end; pte++, addr += PAGE_SIZE)
5968 if (get_mctgt_type(vma, addr, *pte, NULL))
5969 mc.precharge++; /* increment precharge temporarily */
5970 pte_unmap_unlock(pte - 1, ptl);
5976 static const struct mm_walk_ops precharge_walk_ops = {
5977 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5980 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5982 unsigned long precharge;
5985 walk_page_range(mm, 0, ULONG_MAX, &precharge_walk_ops, NULL);
5986 mmap_read_unlock(mm);
5988 precharge = mc.precharge;
5994 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5996 unsigned long precharge = mem_cgroup_count_precharge(mm);
5998 VM_BUG_ON(mc.moving_task);
5999 mc.moving_task = current;
6000 return mem_cgroup_do_precharge(precharge);
6003 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6004 static void __mem_cgroup_clear_mc(void)
6006 struct mem_cgroup *from = mc.from;
6007 struct mem_cgroup *to = mc.to;
6009 /* we must uncharge all the leftover precharges from mc.to */
6011 cancel_charge(mc.to, mc.precharge);
6015 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6016 * we must uncharge here.
6018 if (mc.moved_charge) {
6019 cancel_charge(mc.from, mc.moved_charge);
6020 mc.moved_charge = 0;
6022 /* we must fixup refcnts and charges */
6023 if (mc.moved_swap) {
6024 /* uncharge swap account from the old cgroup */
6025 if (!mem_cgroup_is_root(mc.from))
6026 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
6028 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
6031 * we charged both to->memory and to->memsw, so we
6032 * should uncharge to->memory.
6034 if (!mem_cgroup_is_root(mc.to))
6035 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
6039 memcg_oom_recover(from);
6040 memcg_oom_recover(to);
6041 wake_up_all(&mc.waitq);
6044 static void mem_cgroup_clear_mc(void)
6046 struct mm_struct *mm = mc.mm;
6049 * we must clear moving_task before waking up waiters at the end of
6052 mc.moving_task = NULL;
6053 __mem_cgroup_clear_mc();
6054 spin_lock(&mc.lock);
6058 spin_unlock(&mc.lock);
6063 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6065 struct cgroup_subsys_state *css;
6066 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
6067 struct mem_cgroup *from;
6068 struct task_struct *leader, *p;
6069 struct mm_struct *mm;
6070 unsigned long move_flags;
6073 /* charge immigration isn't supported on the default hierarchy */
6074 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6078 * Multi-process migrations only happen on the default hierarchy
6079 * where charge immigration is not used. Perform charge
6080 * immigration if @tset contains a leader and whine if there are
6084 cgroup_taskset_for_each_leader(leader, css, tset) {
6087 memcg = mem_cgroup_from_css(css);
6093 * We are now committed to this value whatever it is. Changes in this
6094 * tunable will only affect upcoming migrations, not the current one.
6095 * So we need to save it, and keep it going.
6097 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
6101 from = mem_cgroup_from_task(p);
6103 VM_BUG_ON(from == memcg);
6105 mm = get_task_mm(p);
6108 /* We move charges only when we move a owner of the mm */
6109 if (mm->owner == p) {
6112 VM_BUG_ON(mc.precharge);
6113 VM_BUG_ON(mc.moved_charge);
6114 VM_BUG_ON(mc.moved_swap);
6116 spin_lock(&mc.lock);
6120 mc.flags = move_flags;
6121 spin_unlock(&mc.lock);
6122 /* We set mc.moving_task later */
6124 ret = mem_cgroup_precharge_mc(mm);
6126 mem_cgroup_clear_mc();
6133 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6136 mem_cgroup_clear_mc();
6139 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6140 unsigned long addr, unsigned long end,
6141 struct mm_walk *walk)
6144 struct vm_area_struct *vma = walk->vma;
6147 enum mc_target_type target_type;
6148 union mc_target target;
6151 ptl = pmd_trans_huge_lock(pmd, vma);
6153 if (mc.precharge < HPAGE_PMD_NR) {
6157 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6158 if (target_type == MC_TARGET_PAGE) {
6160 if (!isolate_lru_page(page)) {
6161 if (!mem_cgroup_move_account(page, true,
6163 mc.precharge -= HPAGE_PMD_NR;
6164 mc.moved_charge += HPAGE_PMD_NR;
6166 putback_lru_page(page);
6169 } else if (target_type == MC_TARGET_DEVICE) {
6171 if (!mem_cgroup_move_account(page, true,
6173 mc.precharge -= HPAGE_PMD_NR;
6174 mc.moved_charge += HPAGE_PMD_NR;
6182 if (pmd_trans_unstable(pmd))
6185 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6186 for (; addr != end; addr += PAGE_SIZE) {
6187 pte_t ptent = *(pte++);
6188 bool device = false;
6194 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6195 case MC_TARGET_DEVICE:
6198 case MC_TARGET_PAGE:
6201 * We can have a part of the split pmd here. Moving it
6202 * can be done but it would be too convoluted so simply
6203 * ignore such a partial THP and keep it in original
6204 * memcg. There should be somebody mapping the head.
6206 if (PageTransCompound(page))
6208 if (!device && isolate_lru_page(page))
6210 if (!mem_cgroup_move_account(page, false,
6213 /* we uncharge from mc.from later. */
6217 putback_lru_page(page);
6218 put: /* get_mctgt_type() gets the page */
6221 case MC_TARGET_SWAP:
6223 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6225 mem_cgroup_id_get_many(mc.to, 1);
6226 /* we fixup other refcnts and charges later. */
6234 pte_unmap_unlock(pte - 1, ptl);
6239 * We have consumed all precharges we got in can_attach().
6240 * We try charge one by one, but don't do any additional
6241 * charges to mc.to if we have failed in charge once in attach()
6244 ret = mem_cgroup_do_precharge(1);
6252 static const struct mm_walk_ops charge_walk_ops = {
6253 .pmd_entry = mem_cgroup_move_charge_pte_range,
6256 static void mem_cgroup_move_charge(void)
6258 lru_add_drain_all();
6260 * Signal lock_page_memcg() to take the memcg's move_lock
6261 * while we're moving its pages to another memcg. Then wait
6262 * for already started RCU-only updates to finish.
6264 atomic_inc(&mc.from->moving_account);
6267 if (unlikely(!mmap_read_trylock(mc.mm))) {
6269 * Someone who are holding the mmap_lock might be waiting in
6270 * waitq. So we cancel all extra charges, wake up all waiters,
6271 * and retry. Because we cancel precharges, we might not be able
6272 * to move enough charges, but moving charge is a best-effort
6273 * feature anyway, so it wouldn't be a big problem.
6275 __mem_cgroup_clear_mc();
6280 * When we have consumed all precharges and failed in doing
6281 * additional charge, the page walk just aborts.
6283 walk_page_range(mc.mm, 0, ULONG_MAX, &charge_walk_ops, NULL);
6284 mmap_read_unlock(mc.mm);
6285 atomic_dec(&mc.from->moving_account);
6288 static void mem_cgroup_move_task(void)
6291 mem_cgroup_move_charge();
6292 mem_cgroup_clear_mc();
6295 #else /* !CONFIG_MMU */
6296 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6300 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6303 static void mem_cgroup_move_task(void)
6308 #ifdef CONFIG_LRU_GEN
6309 static void mem_cgroup_attach(struct cgroup_taskset *tset)
6311 struct task_struct *task;
6312 struct cgroup_subsys_state *css;
6314 /* find the first leader if there is any */
6315 cgroup_taskset_for_each_leader(task, css, tset)
6322 if (task->mm && READ_ONCE(task->mm->owner) == task)
6323 lru_gen_migrate_mm(task->mm);
6327 static void mem_cgroup_attach(struct cgroup_taskset *tset)
6330 #endif /* CONFIG_LRU_GEN */
6332 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6334 if (value == PAGE_COUNTER_MAX)
6335 seq_puts(m, "max\n");
6337 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6342 static u64 memory_current_read(struct cgroup_subsys_state *css,
6345 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6347 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6350 static u64 memory_peak_read(struct cgroup_subsys_state *css,
6353 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6355 return (u64)memcg->memory.watermark * PAGE_SIZE;
6358 static int memory_min_show(struct seq_file *m, void *v)
6360 return seq_puts_memcg_tunable(m,
6361 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6364 static ssize_t memory_min_write(struct kernfs_open_file *of,
6365 char *buf, size_t nbytes, loff_t off)
6367 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6371 buf = strstrip(buf);
6372 err = page_counter_memparse(buf, "max", &min);
6376 page_counter_set_min(&memcg->memory, min);
6381 static int memory_low_show(struct seq_file *m, void *v)
6383 return seq_puts_memcg_tunable(m,
6384 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6387 static ssize_t memory_low_write(struct kernfs_open_file *of,
6388 char *buf, size_t nbytes, loff_t off)
6390 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6394 buf = strstrip(buf);
6395 err = page_counter_memparse(buf, "max", &low);
6399 page_counter_set_low(&memcg->memory, low);
6404 static int memory_high_show(struct seq_file *m, void *v)
6406 return seq_puts_memcg_tunable(m,
6407 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6410 static ssize_t memory_high_write(struct kernfs_open_file *of,
6411 char *buf, size_t nbytes, loff_t off)
6413 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6414 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6415 bool drained = false;
6419 buf = strstrip(buf);
6420 err = page_counter_memparse(buf, "max", &high);
6424 page_counter_set_high(&memcg->memory, high);
6427 unsigned long nr_pages = page_counter_read(&memcg->memory);
6428 unsigned long reclaimed;
6430 if (nr_pages <= high)
6433 if (signal_pending(current))
6437 drain_all_stock(memcg);
6442 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6443 GFP_KERNEL, MEMCG_RECLAIM_MAY_SWAP);
6445 if (!reclaimed && !nr_retries--)
6449 memcg_wb_domain_size_changed(memcg);
6453 static int memory_max_show(struct seq_file *m, void *v)
6455 return seq_puts_memcg_tunable(m,
6456 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6459 static ssize_t memory_max_write(struct kernfs_open_file *of,
6460 char *buf, size_t nbytes, loff_t off)
6462 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6463 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6464 bool drained = false;
6468 buf = strstrip(buf);
6469 err = page_counter_memparse(buf, "max", &max);
6473 xchg(&memcg->memory.max, max);
6476 unsigned long nr_pages = page_counter_read(&memcg->memory);
6478 if (nr_pages <= max)
6481 if (signal_pending(current))
6485 drain_all_stock(memcg);
6491 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6492 GFP_KERNEL, MEMCG_RECLAIM_MAY_SWAP))
6497 memcg_memory_event(memcg, MEMCG_OOM);
6498 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6502 memcg_wb_domain_size_changed(memcg);
6506 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6508 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6509 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6510 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6511 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6512 seq_printf(m, "oom_kill %lu\n",
6513 atomic_long_read(&events[MEMCG_OOM_KILL]));
6514 seq_printf(m, "oom_group_kill %lu\n",
6515 atomic_long_read(&events[MEMCG_OOM_GROUP_KILL]));
6518 static int memory_events_show(struct seq_file *m, void *v)
6520 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6522 __memory_events_show(m, memcg->memory_events);
6526 static int memory_events_local_show(struct seq_file *m, void *v)
6528 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6530 __memory_events_show(m, memcg->memory_events_local);
6534 static int memory_stat_show(struct seq_file *m, void *v)
6536 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6537 char *buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
6541 memory_stat_format(memcg, buf, PAGE_SIZE);
6548 static inline unsigned long lruvec_page_state_output(struct lruvec *lruvec,
6551 return lruvec_page_state(lruvec, item) * memcg_page_state_unit(item);
6554 static int memory_numa_stat_show(struct seq_file *m, void *v)
6557 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6559 mem_cgroup_flush_stats();
6561 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6564 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6567 seq_printf(m, "%s", memory_stats[i].name);
6568 for_each_node_state(nid, N_MEMORY) {
6570 struct lruvec *lruvec;
6572 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6573 size = lruvec_page_state_output(lruvec,
6574 memory_stats[i].idx);
6575 seq_printf(m, " N%d=%llu", nid, size);
6584 static int memory_oom_group_show(struct seq_file *m, void *v)
6586 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6588 seq_printf(m, "%d\n", memcg->oom_group);
6593 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6594 char *buf, size_t nbytes, loff_t off)
6596 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6599 buf = strstrip(buf);
6603 ret = kstrtoint(buf, 0, &oom_group);
6607 if (oom_group != 0 && oom_group != 1)
6610 memcg->oom_group = oom_group;
6615 static ssize_t memory_reclaim(struct kernfs_open_file *of, char *buf,
6616 size_t nbytes, loff_t off)
6618 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6619 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6620 unsigned long nr_to_reclaim, nr_reclaimed = 0;
6621 unsigned int reclaim_options;
6624 buf = strstrip(buf);
6625 err = page_counter_memparse(buf, "", &nr_to_reclaim);
6629 reclaim_options = MEMCG_RECLAIM_MAY_SWAP | MEMCG_RECLAIM_PROACTIVE;
6630 while (nr_reclaimed < nr_to_reclaim) {
6631 unsigned long reclaimed;
6633 if (signal_pending(current))
6637 * This is the final attempt, drain percpu lru caches in the
6638 * hope of introducing more evictable pages for
6639 * try_to_free_mem_cgroup_pages().
6642 lru_add_drain_all();
6644 reclaimed = try_to_free_mem_cgroup_pages(memcg,
6645 nr_to_reclaim - nr_reclaimed,
6646 GFP_KERNEL, reclaim_options);
6648 if (!reclaimed && !nr_retries--)
6651 nr_reclaimed += reclaimed;
6657 static struct cftype memory_files[] = {
6660 .flags = CFTYPE_NOT_ON_ROOT,
6661 .read_u64 = memory_current_read,
6665 .flags = CFTYPE_NOT_ON_ROOT,
6666 .read_u64 = memory_peak_read,
6670 .flags = CFTYPE_NOT_ON_ROOT,
6671 .seq_show = memory_min_show,
6672 .write = memory_min_write,
6676 .flags = CFTYPE_NOT_ON_ROOT,
6677 .seq_show = memory_low_show,
6678 .write = memory_low_write,
6682 .flags = CFTYPE_NOT_ON_ROOT,
6683 .seq_show = memory_high_show,
6684 .write = memory_high_write,
6688 .flags = CFTYPE_NOT_ON_ROOT,
6689 .seq_show = memory_max_show,
6690 .write = memory_max_write,
6694 .flags = CFTYPE_NOT_ON_ROOT,
6695 .file_offset = offsetof(struct mem_cgroup, events_file),
6696 .seq_show = memory_events_show,
6699 .name = "events.local",
6700 .flags = CFTYPE_NOT_ON_ROOT,
6701 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6702 .seq_show = memory_events_local_show,
6706 .seq_show = memory_stat_show,
6710 .name = "numa_stat",
6711 .seq_show = memory_numa_stat_show,
6715 .name = "oom.group",
6716 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6717 .seq_show = memory_oom_group_show,
6718 .write = memory_oom_group_write,
6722 .flags = CFTYPE_NS_DELEGATABLE,
6723 .write = memory_reclaim,
6728 struct cgroup_subsys memory_cgrp_subsys = {
6729 .css_alloc = mem_cgroup_css_alloc,
6730 .css_online = mem_cgroup_css_online,
6731 .css_offline = mem_cgroup_css_offline,
6732 .css_released = mem_cgroup_css_released,
6733 .css_free = mem_cgroup_css_free,
6734 .css_reset = mem_cgroup_css_reset,
6735 .css_rstat_flush = mem_cgroup_css_rstat_flush,
6736 .can_attach = mem_cgroup_can_attach,
6737 .attach = mem_cgroup_attach,
6738 .cancel_attach = mem_cgroup_cancel_attach,
6739 .post_attach = mem_cgroup_move_task,
6740 .dfl_cftypes = memory_files,
6741 .legacy_cftypes = mem_cgroup_legacy_files,
6746 * This function calculates an individual cgroup's effective
6747 * protection which is derived from its own memory.min/low, its
6748 * parent's and siblings' settings, as well as the actual memory
6749 * distribution in the tree.
6751 * The following rules apply to the effective protection values:
6753 * 1. At the first level of reclaim, effective protection is equal to
6754 * the declared protection in memory.min and memory.low.
6756 * 2. To enable safe delegation of the protection configuration, at
6757 * subsequent levels the effective protection is capped to the
6758 * parent's effective protection.
6760 * 3. To make complex and dynamic subtrees easier to configure, the
6761 * user is allowed to overcommit the declared protection at a given
6762 * level. If that is the case, the parent's effective protection is
6763 * distributed to the children in proportion to how much protection
6764 * they have declared and how much of it they are utilizing.
6766 * This makes distribution proportional, but also work-conserving:
6767 * if one cgroup claims much more protection than it uses memory,
6768 * the unused remainder is available to its siblings.
6770 * 4. Conversely, when the declared protection is undercommitted at a
6771 * given level, the distribution of the larger parental protection
6772 * budget is NOT proportional. A cgroup's protection from a sibling
6773 * is capped to its own memory.min/low setting.
6775 * 5. However, to allow protecting recursive subtrees from each other
6776 * without having to declare each individual cgroup's fixed share
6777 * of the ancestor's claim to protection, any unutilized -
6778 * "floating" - protection from up the tree is distributed in
6779 * proportion to each cgroup's *usage*. This makes the protection
6780 * neutral wrt sibling cgroups and lets them compete freely over
6781 * the shared parental protection budget, but it protects the
6782 * subtree as a whole from neighboring subtrees.
6784 * Note that 4. and 5. are not in conflict: 4. is about protecting
6785 * against immediate siblings whereas 5. is about protecting against
6786 * neighboring subtrees.
6788 static unsigned long effective_protection(unsigned long usage,
6789 unsigned long parent_usage,
6790 unsigned long setting,
6791 unsigned long parent_effective,
6792 unsigned long siblings_protected)
6794 unsigned long protected;
6797 protected = min(usage, setting);
6799 * If all cgroups at this level combined claim and use more
6800 * protection then what the parent affords them, distribute
6801 * shares in proportion to utilization.
6803 * We are using actual utilization rather than the statically
6804 * claimed protection in order to be work-conserving: claimed
6805 * but unused protection is available to siblings that would
6806 * otherwise get a smaller chunk than what they claimed.
6808 if (siblings_protected > parent_effective)
6809 return protected * parent_effective / siblings_protected;
6812 * Ok, utilized protection of all children is within what the
6813 * parent affords them, so we know whatever this child claims
6814 * and utilizes is effectively protected.
6816 * If there is unprotected usage beyond this value, reclaim
6817 * will apply pressure in proportion to that amount.
6819 * If there is unutilized protection, the cgroup will be fully
6820 * shielded from reclaim, but we do return a smaller value for
6821 * protection than what the group could enjoy in theory. This
6822 * is okay. With the overcommit distribution above, effective
6823 * protection is always dependent on how memory is actually
6824 * consumed among the siblings anyway.
6829 * If the children aren't claiming (all of) the protection
6830 * afforded to them by the parent, distribute the remainder in
6831 * proportion to the (unprotected) memory of each cgroup. That
6832 * way, cgroups that aren't explicitly prioritized wrt each
6833 * other compete freely over the allowance, but they are
6834 * collectively protected from neighboring trees.
6836 * We're using unprotected memory for the weight so that if
6837 * some cgroups DO claim explicit protection, we don't protect
6838 * the same bytes twice.
6840 * Check both usage and parent_usage against the respective
6841 * protected values. One should imply the other, but they
6842 * aren't read atomically - make sure the division is sane.
6844 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6846 if (parent_effective > siblings_protected &&
6847 parent_usage > siblings_protected &&
6848 usage > protected) {
6849 unsigned long unclaimed;
6851 unclaimed = parent_effective - siblings_protected;
6852 unclaimed *= usage - protected;
6853 unclaimed /= parent_usage - siblings_protected;
6862 * mem_cgroup_calculate_protection - check if memory consumption is in the normal range
6863 * @root: the top ancestor of the sub-tree being checked
6864 * @memcg: the memory cgroup to check
6866 * WARNING: This function is not stateless! It can only be used as part
6867 * of a top-down tree iteration, not for isolated queries.
6869 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6870 struct mem_cgroup *memcg)
6872 unsigned long usage, parent_usage;
6873 struct mem_cgroup *parent;
6875 if (mem_cgroup_disabled())
6879 root = root_mem_cgroup;
6882 * Effective values of the reclaim targets are ignored so they
6883 * can be stale. Have a look at mem_cgroup_protection for more
6885 * TODO: calculation should be more robust so that we do not need
6886 * that special casing.
6891 usage = page_counter_read(&memcg->memory);
6895 parent = parent_mem_cgroup(memcg);
6897 if (parent == root) {
6898 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6899 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6903 parent_usage = page_counter_read(&parent->memory);
6905 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6906 READ_ONCE(memcg->memory.min),
6907 READ_ONCE(parent->memory.emin),
6908 atomic_long_read(&parent->memory.children_min_usage)));
6910 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6911 READ_ONCE(memcg->memory.low),
6912 READ_ONCE(parent->memory.elow),
6913 atomic_long_read(&parent->memory.children_low_usage)));
6916 static int charge_memcg(struct folio *folio, struct mem_cgroup *memcg,
6919 long nr_pages = folio_nr_pages(folio);
6922 ret = try_charge(memcg, gfp, nr_pages);
6926 css_get(&memcg->css);
6927 commit_charge(folio, memcg);
6929 local_irq_disable();
6930 mem_cgroup_charge_statistics(memcg, nr_pages);
6931 memcg_check_events(memcg, folio_nid(folio));
6937 int __mem_cgroup_charge(struct folio *folio, struct mm_struct *mm, gfp_t gfp)
6939 struct mem_cgroup *memcg;
6942 memcg = get_mem_cgroup_from_mm(mm);
6943 ret = charge_memcg(folio, memcg, gfp);
6944 css_put(&memcg->css);
6950 * mem_cgroup_swapin_charge_folio - Charge a newly allocated folio for swapin.
6951 * @folio: folio to charge.
6952 * @mm: mm context of the victim
6953 * @gfp: reclaim mode
6954 * @entry: swap entry for which the folio is allocated
6956 * This function charges a folio allocated for swapin. Please call this before
6957 * adding the folio to the swapcache.
6959 * Returns 0 on success. Otherwise, an error code is returned.
6961 int mem_cgroup_swapin_charge_folio(struct folio *folio, struct mm_struct *mm,
6962 gfp_t gfp, swp_entry_t entry)
6964 struct mem_cgroup *memcg;
6968 if (mem_cgroup_disabled())
6971 id = lookup_swap_cgroup_id(entry);
6973 memcg = mem_cgroup_from_id(id);
6974 if (!memcg || !css_tryget_online(&memcg->css))
6975 memcg = get_mem_cgroup_from_mm(mm);
6978 ret = charge_memcg(folio, memcg, gfp);
6980 css_put(&memcg->css);
6985 * mem_cgroup_swapin_uncharge_swap - uncharge swap slot
6986 * @entry: swap entry for which the page is charged
6988 * Call this function after successfully adding the charged page to swapcache.
6990 * Note: This function assumes the page for which swap slot is being uncharged
6993 void mem_cgroup_swapin_uncharge_swap(swp_entry_t entry)
6996 * Cgroup1's unified memory+swap counter has been charged with the
6997 * new swapcache page, finish the transfer by uncharging the swap
6998 * slot. The swap slot would also get uncharged when it dies, but
6999 * it can stick around indefinitely and we'd count the page twice
7002 * Cgroup2 has separate resource counters for memory and swap,
7003 * so this is a non-issue here. Memory and swap charge lifetimes
7004 * correspond 1:1 to page and swap slot lifetimes: we charge the
7005 * page to memory here, and uncharge swap when the slot is freed.
7007 if (!mem_cgroup_disabled() && do_memsw_account()) {
7009 * The swap entry might not get freed for a long time,
7010 * let's not wait for it. The page already received a
7011 * memory+swap charge, drop the swap entry duplicate.
7013 mem_cgroup_uncharge_swap(entry, 1);
7017 struct uncharge_gather {
7018 struct mem_cgroup *memcg;
7019 unsigned long nr_memory;
7020 unsigned long pgpgout;
7021 unsigned long nr_kmem;
7025 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
7027 memset(ug, 0, sizeof(*ug));
7030 static void uncharge_batch(const struct uncharge_gather *ug)
7032 unsigned long flags;
7034 if (ug->nr_memory) {
7035 page_counter_uncharge(&ug->memcg->memory, ug->nr_memory);
7036 if (do_memsw_account())
7037 page_counter_uncharge(&ug->memcg->memsw, ug->nr_memory);
7039 memcg_account_kmem(ug->memcg, -ug->nr_kmem);
7040 memcg_oom_recover(ug->memcg);
7043 local_irq_save(flags);
7044 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
7045 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_memory);
7046 memcg_check_events(ug->memcg, ug->nid);
7047 local_irq_restore(flags);
7049 /* drop reference from uncharge_folio */
7050 css_put(&ug->memcg->css);
7053 static void uncharge_folio(struct folio *folio, struct uncharge_gather *ug)
7056 struct mem_cgroup *memcg;
7057 struct obj_cgroup *objcg;
7059 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
7062 * Nobody should be changing or seriously looking at
7063 * folio memcg or objcg at this point, we have fully
7064 * exclusive access to the folio.
7066 if (folio_memcg_kmem(folio)) {
7067 objcg = __folio_objcg(folio);
7069 * This get matches the put at the end of the function and
7070 * kmem pages do not hold memcg references anymore.
7072 memcg = get_mem_cgroup_from_objcg(objcg);
7074 memcg = __folio_memcg(folio);
7080 if (ug->memcg != memcg) {
7083 uncharge_gather_clear(ug);
7086 ug->nid = folio_nid(folio);
7088 /* pairs with css_put in uncharge_batch */
7089 css_get(&memcg->css);
7092 nr_pages = folio_nr_pages(folio);
7094 if (folio_memcg_kmem(folio)) {
7095 ug->nr_memory += nr_pages;
7096 ug->nr_kmem += nr_pages;
7098 folio->memcg_data = 0;
7099 obj_cgroup_put(objcg);
7101 /* LRU pages aren't accounted at the root level */
7102 if (!mem_cgroup_is_root(memcg))
7103 ug->nr_memory += nr_pages;
7106 folio->memcg_data = 0;
7109 css_put(&memcg->css);
7112 void __mem_cgroup_uncharge(struct folio *folio)
7114 struct uncharge_gather ug;
7116 /* Don't touch folio->lru of any random page, pre-check: */
7117 if (!folio_memcg(folio))
7120 uncharge_gather_clear(&ug);
7121 uncharge_folio(folio, &ug);
7122 uncharge_batch(&ug);
7126 * __mem_cgroup_uncharge_list - uncharge a list of page
7127 * @page_list: list of pages to uncharge
7129 * Uncharge a list of pages previously charged with
7130 * __mem_cgroup_charge().
7132 void __mem_cgroup_uncharge_list(struct list_head *page_list)
7134 struct uncharge_gather ug;
7135 struct folio *folio;
7137 uncharge_gather_clear(&ug);
7138 list_for_each_entry(folio, page_list, lru)
7139 uncharge_folio(folio, &ug);
7141 uncharge_batch(&ug);
7145 * mem_cgroup_migrate - Charge a folio's replacement.
7146 * @old: Currently circulating folio.
7147 * @new: Replacement folio.
7149 * Charge @new as a replacement folio for @old. @old will
7150 * be uncharged upon free.
7152 * Both folios must be locked, @new->mapping must be set up.
7154 void mem_cgroup_migrate(struct folio *old, struct folio *new)
7156 struct mem_cgroup *memcg;
7157 long nr_pages = folio_nr_pages(new);
7158 unsigned long flags;
7160 VM_BUG_ON_FOLIO(!folio_test_locked(old), old);
7161 VM_BUG_ON_FOLIO(!folio_test_locked(new), new);
7162 VM_BUG_ON_FOLIO(folio_test_anon(old) != folio_test_anon(new), new);
7163 VM_BUG_ON_FOLIO(folio_nr_pages(old) != nr_pages, new);
7165 if (mem_cgroup_disabled())
7168 /* Page cache replacement: new folio already charged? */
7169 if (folio_memcg(new))
7172 memcg = folio_memcg(old);
7173 VM_WARN_ON_ONCE_FOLIO(!memcg, old);
7177 /* Force-charge the new page. The old one will be freed soon */
7178 if (!mem_cgroup_is_root(memcg)) {
7179 page_counter_charge(&memcg->memory, nr_pages);
7180 if (do_memsw_account())
7181 page_counter_charge(&memcg->memsw, nr_pages);
7184 css_get(&memcg->css);
7185 commit_charge(new, memcg);
7187 local_irq_save(flags);
7188 mem_cgroup_charge_statistics(memcg, nr_pages);
7189 memcg_check_events(memcg, folio_nid(new));
7190 local_irq_restore(flags);
7193 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
7194 EXPORT_SYMBOL(memcg_sockets_enabled_key);
7196 void mem_cgroup_sk_alloc(struct sock *sk)
7198 struct mem_cgroup *memcg;
7200 if (!mem_cgroup_sockets_enabled)
7203 /* Do not associate the sock with unrelated interrupted task's memcg. */
7208 memcg = mem_cgroup_from_task(current);
7209 if (mem_cgroup_is_root(memcg))
7211 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
7213 if (css_tryget(&memcg->css))
7214 sk->sk_memcg = memcg;
7219 void mem_cgroup_sk_free(struct sock *sk)
7222 css_put(&sk->sk_memcg->css);
7226 * mem_cgroup_charge_skmem - charge socket memory
7227 * @memcg: memcg to charge
7228 * @nr_pages: number of pages to charge
7229 * @gfp_mask: reclaim mode
7231 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7232 * @memcg's configured limit, %false if it doesn't.
7234 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages,
7237 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7238 struct page_counter *fail;
7240 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7241 memcg->tcpmem_pressure = 0;
7244 memcg->tcpmem_pressure = 1;
7245 if (gfp_mask & __GFP_NOFAIL) {
7246 page_counter_charge(&memcg->tcpmem, nr_pages);
7252 if (try_charge(memcg, gfp_mask, nr_pages) == 0) {
7253 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7261 * mem_cgroup_uncharge_skmem - uncharge socket memory
7262 * @memcg: memcg to uncharge
7263 * @nr_pages: number of pages to uncharge
7265 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7267 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7268 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7272 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7274 refill_stock(memcg, nr_pages);
7277 static int __init cgroup_memory(char *s)
7281 while ((token = strsep(&s, ",")) != NULL) {
7284 if (!strcmp(token, "nosocket"))
7285 cgroup_memory_nosocket = true;
7286 if (!strcmp(token, "nokmem"))
7287 cgroup_memory_nokmem = true;
7288 if (!strcmp(token, "nobpf"))
7289 cgroup_memory_nobpf = true;
7293 __setup("cgroup.memory=", cgroup_memory);
7296 * subsys_initcall() for memory controller.
7298 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7299 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7300 * basically everything that doesn't depend on a specific mem_cgroup structure
7301 * should be initialized from here.
7303 static int __init mem_cgroup_init(void)
7308 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7309 * used for per-memcg-per-cpu caching of per-node statistics. In order
7310 * to work fine, we should make sure that the overfill threshold can't
7311 * exceed S32_MAX / PAGE_SIZE.
7313 BUILD_BUG_ON(MEMCG_CHARGE_BATCH > S32_MAX / PAGE_SIZE);
7315 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7316 memcg_hotplug_cpu_dead);
7318 for_each_possible_cpu(cpu)
7319 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7322 for_each_node(node) {
7323 struct mem_cgroup_tree_per_node *rtpn;
7325 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7326 node_online(node) ? node : NUMA_NO_NODE);
7328 rtpn->rb_root = RB_ROOT;
7329 rtpn->rb_rightmost = NULL;
7330 spin_lock_init(&rtpn->lock);
7331 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7336 subsys_initcall(mem_cgroup_init);
7339 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7341 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7343 * The root cgroup cannot be destroyed, so it's refcount must
7346 if (WARN_ON_ONCE(mem_cgroup_is_root(memcg))) {
7350 memcg = parent_mem_cgroup(memcg);
7352 memcg = root_mem_cgroup;
7358 * mem_cgroup_swapout - transfer a memsw charge to swap
7359 * @folio: folio whose memsw charge to transfer
7360 * @entry: swap entry to move the charge to
7362 * Transfer the memsw charge of @folio to @entry.
7364 void mem_cgroup_swapout(struct folio *folio, swp_entry_t entry)
7366 struct mem_cgroup *memcg, *swap_memcg;
7367 unsigned int nr_entries;
7368 unsigned short oldid;
7370 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
7371 VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
7373 if (mem_cgroup_disabled())
7376 if (!do_memsw_account())
7379 memcg = folio_memcg(folio);
7381 VM_WARN_ON_ONCE_FOLIO(!memcg, folio);
7386 * In case the memcg owning these pages has been offlined and doesn't
7387 * have an ID allocated to it anymore, charge the closest online
7388 * ancestor for the swap instead and transfer the memory+swap charge.
7390 swap_memcg = mem_cgroup_id_get_online(memcg);
7391 nr_entries = folio_nr_pages(folio);
7392 /* Get references for the tail pages, too */
7394 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7395 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7397 VM_BUG_ON_FOLIO(oldid, folio);
7398 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7400 folio->memcg_data = 0;
7402 if (!mem_cgroup_is_root(memcg))
7403 page_counter_uncharge(&memcg->memory, nr_entries);
7405 if (memcg != swap_memcg) {
7406 if (!mem_cgroup_is_root(swap_memcg))
7407 page_counter_charge(&swap_memcg->memsw, nr_entries);
7408 page_counter_uncharge(&memcg->memsw, nr_entries);
7412 * Interrupts should be disabled here because the caller holds the
7413 * i_pages lock which is taken with interrupts-off. It is
7414 * important here to have the interrupts disabled because it is the
7415 * only synchronisation we have for updating the per-CPU variables.
7418 mem_cgroup_charge_statistics(memcg, -nr_entries);
7419 memcg_stats_unlock();
7420 memcg_check_events(memcg, folio_nid(folio));
7422 css_put(&memcg->css);
7426 * __mem_cgroup_try_charge_swap - try charging swap space for a folio
7427 * @folio: folio being added to swap
7428 * @entry: swap entry to charge
7430 * Try to charge @folio's memcg for the swap space at @entry.
7432 * Returns 0 on success, -ENOMEM on failure.
7434 int __mem_cgroup_try_charge_swap(struct folio *folio, swp_entry_t entry)
7436 unsigned int nr_pages = folio_nr_pages(folio);
7437 struct page_counter *counter;
7438 struct mem_cgroup *memcg;
7439 unsigned short oldid;
7441 if (do_memsw_account())
7444 memcg = folio_memcg(folio);
7446 VM_WARN_ON_ONCE_FOLIO(!memcg, folio);
7451 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7455 memcg = mem_cgroup_id_get_online(memcg);
7457 if (!mem_cgroup_is_root(memcg) &&
7458 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7459 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7460 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7461 mem_cgroup_id_put(memcg);
7465 /* Get references for the tail pages, too */
7467 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7468 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7469 VM_BUG_ON_FOLIO(oldid, folio);
7470 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7476 * __mem_cgroup_uncharge_swap - uncharge swap space
7477 * @entry: swap entry to uncharge
7478 * @nr_pages: the amount of swap space to uncharge
7480 void __mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7482 struct mem_cgroup *memcg;
7485 if (mem_cgroup_disabled())
7488 id = swap_cgroup_record(entry, 0, nr_pages);
7490 memcg = mem_cgroup_from_id(id);
7492 if (!mem_cgroup_is_root(memcg)) {
7493 if (do_memsw_account())
7494 page_counter_uncharge(&memcg->memsw, nr_pages);
7496 page_counter_uncharge(&memcg->swap, nr_pages);
7498 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7499 mem_cgroup_id_put_many(memcg, nr_pages);
7504 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7506 long nr_swap_pages = get_nr_swap_pages();
7508 if (mem_cgroup_disabled() || do_memsw_account())
7509 return nr_swap_pages;
7510 for (; !mem_cgroup_is_root(memcg); memcg = parent_mem_cgroup(memcg))
7511 nr_swap_pages = min_t(long, nr_swap_pages,
7512 READ_ONCE(memcg->swap.max) -
7513 page_counter_read(&memcg->swap));
7514 return nr_swap_pages;
7517 bool mem_cgroup_swap_full(struct folio *folio)
7519 struct mem_cgroup *memcg;
7521 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
7525 if (do_memsw_account())
7528 memcg = folio_memcg(folio);
7532 for (; !mem_cgroup_is_root(memcg); memcg = parent_mem_cgroup(memcg)) {
7533 unsigned long usage = page_counter_read(&memcg->swap);
7535 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7536 usage * 2 >= READ_ONCE(memcg->swap.max))
7543 static int __init setup_swap_account(char *s)
7545 pr_warn_once("The swapaccount= commandline option is deprecated. "
7546 "Please report your usecase to linux-mm@kvack.org if you "
7547 "depend on this functionality.\n");
7550 __setup("swapaccount=", setup_swap_account);
7552 static u64 swap_current_read(struct cgroup_subsys_state *css,
7555 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7557 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7560 static int swap_high_show(struct seq_file *m, void *v)
7562 return seq_puts_memcg_tunable(m,
7563 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7566 static ssize_t swap_high_write(struct kernfs_open_file *of,
7567 char *buf, size_t nbytes, loff_t off)
7569 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7573 buf = strstrip(buf);
7574 err = page_counter_memparse(buf, "max", &high);
7578 page_counter_set_high(&memcg->swap, high);
7583 static int swap_max_show(struct seq_file *m, void *v)
7585 return seq_puts_memcg_tunable(m,
7586 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7589 static ssize_t swap_max_write(struct kernfs_open_file *of,
7590 char *buf, size_t nbytes, loff_t off)
7592 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7596 buf = strstrip(buf);
7597 err = page_counter_memparse(buf, "max", &max);
7601 xchg(&memcg->swap.max, max);
7606 static int swap_events_show(struct seq_file *m, void *v)
7608 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7610 seq_printf(m, "high %lu\n",
7611 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7612 seq_printf(m, "max %lu\n",
7613 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7614 seq_printf(m, "fail %lu\n",
7615 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7620 static struct cftype swap_files[] = {
7622 .name = "swap.current",
7623 .flags = CFTYPE_NOT_ON_ROOT,
7624 .read_u64 = swap_current_read,
7627 .name = "swap.high",
7628 .flags = CFTYPE_NOT_ON_ROOT,
7629 .seq_show = swap_high_show,
7630 .write = swap_high_write,
7634 .flags = CFTYPE_NOT_ON_ROOT,
7635 .seq_show = swap_max_show,
7636 .write = swap_max_write,
7639 .name = "swap.events",
7640 .flags = CFTYPE_NOT_ON_ROOT,
7641 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7642 .seq_show = swap_events_show,
7647 static struct cftype memsw_files[] = {
7649 .name = "memsw.usage_in_bytes",
7650 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7651 .read_u64 = mem_cgroup_read_u64,
7654 .name = "memsw.max_usage_in_bytes",
7655 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7656 .write = mem_cgroup_reset,
7657 .read_u64 = mem_cgroup_read_u64,
7660 .name = "memsw.limit_in_bytes",
7661 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7662 .write = mem_cgroup_write,
7663 .read_u64 = mem_cgroup_read_u64,
7666 .name = "memsw.failcnt",
7667 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7668 .write = mem_cgroup_reset,
7669 .read_u64 = mem_cgroup_read_u64,
7671 { }, /* terminate */
7674 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
7676 * obj_cgroup_may_zswap - check if this cgroup can zswap
7677 * @objcg: the object cgroup
7679 * Check if the hierarchical zswap limit has been reached.
7681 * This doesn't check for specific headroom, and it is not atomic
7682 * either. But with zswap, the size of the allocation is only known
7683 * once compression has occured, and this optimistic pre-check avoids
7684 * spending cycles on compression when there is already no room left
7685 * or zswap is disabled altogether somewhere in the hierarchy.
7687 bool obj_cgroup_may_zswap(struct obj_cgroup *objcg)
7689 struct mem_cgroup *memcg, *original_memcg;
7692 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7695 original_memcg = get_mem_cgroup_from_objcg(objcg);
7696 for (memcg = original_memcg; !mem_cgroup_is_root(memcg);
7697 memcg = parent_mem_cgroup(memcg)) {
7698 unsigned long max = READ_ONCE(memcg->zswap_max);
7699 unsigned long pages;
7701 if (max == PAGE_COUNTER_MAX)
7708 cgroup_rstat_flush(memcg->css.cgroup);
7709 pages = memcg_page_state(memcg, MEMCG_ZSWAP_B) / PAGE_SIZE;
7715 mem_cgroup_put(original_memcg);
7720 * obj_cgroup_charge_zswap - charge compression backend memory
7721 * @objcg: the object cgroup
7722 * @size: size of compressed object
7724 * This forces the charge after obj_cgroup_may_swap() allowed
7725 * compression and storage in zwap for this cgroup to go ahead.
7727 void obj_cgroup_charge_zswap(struct obj_cgroup *objcg, size_t size)
7729 struct mem_cgroup *memcg;
7731 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7734 VM_WARN_ON_ONCE(!(current->flags & PF_MEMALLOC));
7736 /* PF_MEMALLOC context, charging must succeed */
7737 if (obj_cgroup_charge(objcg, GFP_KERNEL, size))
7741 memcg = obj_cgroup_memcg(objcg);
7742 mod_memcg_state(memcg, MEMCG_ZSWAP_B, size);
7743 mod_memcg_state(memcg, MEMCG_ZSWAPPED, 1);
7748 * obj_cgroup_uncharge_zswap - uncharge compression backend memory
7749 * @objcg: the object cgroup
7750 * @size: size of compressed object
7752 * Uncharges zswap memory on page in.
7754 void obj_cgroup_uncharge_zswap(struct obj_cgroup *objcg, size_t size)
7756 struct mem_cgroup *memcg;
7758 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7761 obj_cgroup_uncharge(objcg, size);
7764 memcg = obj_cgroup_memcg(objcg);
7765 mod_memcg_state(memcg, MEMCG_ZSWAP_B, -size);
7766 mod_memcg_state(memcg, MEMCG_ZSWAPPED, -1);
7770 static u64 zswap_current_read(struct cgroup_subsys_state *css,
7773 cgroup_rstat_flush(css->cgroup);
7774 return memcg_page_state(mem_cgroup_from_css(css), MEMCG_ZSWAP_B);
7777 static int zswap_max_show(struct seq_file *m, void *v)
7779 return seq_puts_memcg_tunable(m,
7780 READ_ONCE(mem_cgroup_from_seq(m)->zswap_max));
7783 static ssize_t zswap_max_write(struct kernfs_open_file *of,
7784 char *buf, size_t nbytes, loff_t off)
7786 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7790 buf = strstrip(buf);
7791 err = page_counter_memparse(buf, "max", &max);
7795 xchg(&memcg->zswap_max, max);
7800 static struct cftype zswap_files[] = {
7802 .name = "zswap.current",
7803 .flags = CFTYPE_NOT_ON_ROOT,
7804 .read_u64 = zswap_current_read,
7807 .name = "zswap.max",
7808 .flags = CFTYPE_NOT_ON_ROOT,
7809 .seq_show = zswap_max_show,
7810 .write = zswap_max_write,
7814 #endif /* CONFIG_MEMCG_KMEM && CONFIG_ZSWAP */
7816 static int __init mem_cgroup_swap_init(void)
7818 if (mem_cgroup_disabled())
7821 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7822 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7823 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
7824 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, zswap_files));
7828 subsys_initcall(mem_cgroup_swap_init);
7830 #endif /* CONFIG_SWAP */