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 #ifdef CONFIG_CGROUP_WRITEBACK
92 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
95 /* Whether legacy memory+swap accounting is active */
96 static bool do_memsw_account(void)
98 return !cgroup_subsys_on_dfl(memory_cgrp_subsys);
101 #define THRESHOLDS_EVENTS_TARGET 128
102 #define SOFTLIMIT_EVENTS_TARGET 1024
105 * Cgroups above their limits are maintained in a RB-Tree, independent of
106 * their hierarchy representation
109 struct mem_cgroup_tree_per_node {
110 struct rb_root rb_root;
111 struct rb_node *rb_rightmost;
115 struct mem_cgroup_tree {
116 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
119 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
122 struct mem_cgroup_eventfd_list {
123 struct list_head list;
124 struct eventfd_ctx *eventfd;
128 * cgroup_event represents events which userspace want to receive.
130 struct mem_cgroup_event {
132 * memcg which the event belongs to.
134 struct mem_cgroup *memcg;
136 * eventfd to signal userspace about the event.
138 struct eventfd_ctx *eventfd;
140 * Each of these stored in a list by the cgroup.
142 struct list_head list;
144 * register_event() callback will be used to add new userspace
145 * waiter for changes related to this event. Use eventfd_signal()
146 * on eventfd to send notification to userspace.
148 int (*register_event)(struct mem_cgroup *memcg,
149 struct eventfd_ctx *eventfd, const char *args);
151 * unregister_event() callback will be called when userspace closes
152 * the eventfd or on cgroup removing. This callback must be set,
153 * if you want provide notification functionality.
155 void (*unregister_event)(struct mem_cgroup *memcg,
156 struct eventfd_ctx *eventfd);
158 * All fields below needed to unregister event when
159 * userspace closes eventfd.
162 wait_queue_head_t *wqh;
163 wait_queue_entry_t wait;
164 struct work_struct remove;
167 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
168 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
170 /* Stuffs for move charges at task migration. */
172 * Types of charges to be moved.
174 #define MOVE_ANON 0x1U
175 #define MOVE_FILE 0x2U
176 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
178 /* "mc" and its members are protected by cgroup_mutex */
179 static struct move_charge_struct {
180 spinlock_t lock; /* for from, to */
181 struct mm_struct *mm;
182 struct mem_cgroup *from;
183 struct mem_cgroup *to;
185 unsigned long precharge;
186 unsigned long moved_charge;
187 unsigned long moved_swap;
188 struct task_struct *moving_task; /* a task moving charges */
189 wait_queue_head_t waitq; /* a waitq for other context */
191 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
192 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
196 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
197 * limit reclaim to prevent infinite loops, if they ever occur.
199 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
200 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
202 /* for encoding cft->private value on file */
210 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
211 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
212 #define MEMFILE_ATTR(val) ((val) & 0xffff)
215 * Iteration constructs for visiting all cgroups (under a tree). If
216 * loops are exited prematurely (break), mem_cgroup_iter_break() must
217 * be used for reference counting.
219 #define for_each_mem_cgroup_tree(iter, root) \
220 for (iter = mem_cgroup_iter(root, NULL, NULL); \
222 iter = mem_cgroup_iter(root, iter, NULL))
224 #define for_each_mem_cgroup(iter) \
225 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
227 iter = mem_cgroup_iter(NULL, iter, NULL))
229 static inline bool task_is_dying(void)
231 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
232 (current->flags & PF_EXITING);
235 /* Some nice accessors for the vmpressure. */
236 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
239 memcg = root_mem_cgroup;
240 return &memcg->vmpressure;
243 struct mem_cgroup *vmpressure_to_memcg(struct vmpressure *vmpr)
245 return container_of(vmpr, struct mem_cgroup, vmpressure);
248 #ifdef CONFIG_MEMCG_KMEM
249 static DEFINE_SPINLOCK(objcg_lock);
251 bool mem_cgroup_kmem_disabled(void)
253 return cgroup_memory_nokmem;
256 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
257 unsigned int nr_pages);
259 static void obj_cgroup_release(struct percpu_ref *ref)
261 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
262 unsigned int nr_bytes;
263 unsigned int nr_pages;
267 * At this point all allocated objects are freed, and
268 * objcg->nr_charged_bytes can't have an arbitrary byte value.
269 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
271 * The following sequence can lead to it:
272 * 1) CPU0: objcg == stock->cached_objcg
273 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
274 * PAGE_SIZE bytes are charged
275 * 3) CPU1: a process from another memcg is allocating something,
276 * the stock if flushed,
277 * objcg->nr_charged_bytes = PAGE_SIZE - 92
278 * 5) CPU0: we do release this object,
279 * 92 bytes are added to stock->nr_bytes
280 * 6) CPU0: stock is flushed,
281 * 92 bytes are added to objcg->nr_charged_bytes
283 * In the result, nr_charged_bytes == PAGE_SIZE.
284 * This page will be uncharged in obj_cgroup_release().
286 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
287 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
288 nr_pages = nr_bytes >> PAGE_SHIFT;
291 obj_cgroup_uncharge_pages(objcg, nr_pages);
293 spin_lock_irqsave(&objcg_lock, flags);
294 list_del(&objcg->list);
295 spin_unlock_irqrestore(&objcg_lock, flags);
297 percpu_ref_exit(ref);
298 kfree_rcu(objcg, rcu);
301 static struct obj_cgroup *obj_cgroup_alloc(void)
303 struct obj_cgroup *objcg;
306 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
310 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
316 INIT_LIST_HEAD(&objcg->list);
320 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
321 struct mem_cgroup *parent)
323 struct obj_cgroup *objcg, *iter;
325 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
327 spin_lock_irq(&objcg_lock);
329 /* 1) Ready to reparent active objcg. */
330 list_add(&objcg->list, &memcg->objcg_list);
331 /* 2) Reparent active objcg and already reparented objcgs to parent. */
332 list_for_each_entry(iter, &memcg->objcg_list, list)
333 WRITE_ONCE(iter->memcg, parent);
334 /* 3) Move already reparented objcgs to the parent's list */
335 list_splice(&memcg->objcg_list, &parent->objcg_list);
337 spin_unlock_irq(&objcg_lock);
339 percpu_ref_kill(&objcg->refcnt);
343 * A lot of the calls to the cache allocation functions are expected to be
344 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
345 * conditional to this static branch, we'll have to allow modules that does
346 * kmem_cache_alloc and the such to see this symbol as well
348 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
349 EXPORT_SYMBOL(memcg_kmem_enabled_key);
353 * mem_cgroup_css_from_page - css of the memcg associated with a page
354 * @page: page of interest
356 * If memcg is bound to the default hierarchy, css of the memcg associated
357 * with @page is returned. The returned css remains associated with @page
358 * until it is released.
360 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
363 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
365 struct mem_cgroup *memcg;
367 memcg = page_memcg(page);
369 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
370 memcg = root_mem_cgroup;
376 * page_cgroup_ino - return inode number of the memcg a page is charged to
379 * Look up the closest online ancestor of the memory cgroup @page is charged to
380 * and return its inode number or 0 if @page is not charged to any cgroup. It
381 * is safe to call this function without holding a reference to @page.
383 * Note, this function is inherently racy, because there is nothing to prevent
384 * the cgroup inode from getting torn down and potentially reallocated a moment
385 * after page_cgroup_ino() returns, so it only should be used by callers that
386 * do not care (such as procfs interfaces).
388 ino_t page_cgroup_ino(struct page *page)
390 struct mem_cgroup *memcg;
391 unsigned long ino = 0;
394 memcg = page_memcg_check(page);
396 while (memcg && !(memcg->css.flags & CSS_ONLINE))
397 memcg = parent_mem_cgroup(memcg);
399 ino = cgroup_ino(memcg->css.cgroup);
404 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
405 struct mem_cgroup_tree_per_node *mctz,
406 unsigned long new_usage_in_excess)
408 struct rb_node **p = &mctz->rb_root.rb_node;
409 struct rb_node *parent = NULL;
410 struct mem_cgroup_per_node *mz_node;
411 bool rightmost = true;
416 mz->usage_in_excess = new_usage_in_excess;
417 if (!mz->usage_in_excess)
421 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
423 if (mz->usage_in_excess < mz_node->usage_in_excess) {
432 mctz->rb_rightmost = &mz->tree_node;
434 rb_link_node(&mz->tree_node, parent, p);
435 rb_insert_color(&mz->tree_node, &mctz->rb_root);
439 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
440 struct mem_cgroup_tree_per_node *mctz)
445 if (&mz->tree_node == mctz->rb_rightmost)
446 mctz->rb_rightmost = rb_prev(&mz->tree_node);
448 rb_erase(&mz->tree_node, &mctz->rb_root);
452 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
453 struct mem_cgroup_tree_per_node *mctz)
457 spin_lock_irqsave(&mctz->lock, flags);
458 __mem_cgroup_remove_exceeded(mz, mctz);
459 spin_unlock_irqrestore(&mctz->lock, flags);
462 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
464 unsigned long nr_pages = page_counter_read(&memcg->memory);
465 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
466 unsigned long excess = 0;
468 if (nr_pages > soft_limit)
469 excess = nr_pages - soft_limit;
474 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, int nid)
476 unsigned long excess;
477 struct mem_cgroup_per_node *mz;
478 struct mem_cgroup_tree_per_node *mctz;
480 mctz = soft_limit_tree.rb_tree_per_node[nid];
484 * Necessary to update all ancestors when hierarchy is used.
485 * because their event counter is not touched.
487 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
488 mz = memcg->nodeinfo[nid];
489 excess = soft_limit_excess(memcg);
491 * We have to update the tree if mz is on RB-tree or
492 * mem is over its softlimit.
494 if (excess || mz->on_tree) {
497 spin_lock_irqsave(&mctz->lock, flags);
498 /* if on-tree, remove it */
500 __mem_cgroup_remove_exceeded(mz, mctz);
502 * Insert again. mz->usage_in_excess will be updated.
503 * If excess is 0, no tree ops.
505 __mem_cgroup_insert_exceeded(mz, mctz, excess);
506 spin_unlock_irqrestore(&mctz->lock, flags);
511 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
513 struct mem_cgroup_tree_per_node *mctz;
514 struct mem_cgroup_per_node *mz;
518 mz = memcg->nodeinfo[nid];
519 mctz = soft_limit_tree.rb_tree_per_node[nid];
521 mem_cgroup_remove_exceeded(mz, mctz);
525 static struct mem_cgroup_per_node *
526 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
528 struct mem_cgroup_per_node *mz;
532 if (!mctz->rb_rightmost)
533 goto done; /* Nothing to reclaim from */
535 mz = rb_entry(mctz->rb_rightmost,
536 struct mem_cgroup_per_node, tree_node);
538 * Remove the node now but someone else can add it back,
539 * we will to add it back at the end of reclaim to its correct
540 * position in the tree.
542 __mem_cgroup_remove_exceeded(mz, mctz);
543 if (!soft_limit_excess(mz->memcg) ||
544 !css_tryget(&mz->memcg->css))
550 static struct mem_cgroup_per_node *
551 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
553 struct mem_cgroup_per_node *mz;
555 spin_lock_irq(&mctz->lock);
556 mz = __mem_cgroup_largest_soft_limit_node(mctz);
557 spin_unlock_irq(&mctz->lock);
562 * memcg and lruvec stats flushing
564 * Many codepaths leading to stats update or read are performance sensitive and
565 * adding stats flushing in such codepaths is not desirable. So, to optimize the
566 * flushing the kernel does:
568 * 1) Periodically and asynchronously flush the stats every 2 seconds to not let
569 * rstat update tree grow unbounded.
571 * 2) Flush the stats synchronously on reader side only when there are more than
572 * (MEMCG_CHARGE_BATCH * nr_cpus) update events. Though this optimization
573 * will let stats be out of sync by atmost (MEMCG_CHARGE_BATCH * nr_cpus) but
574 * only for 2 seconds due to (1).
576 static void flush_memcg_stats_dwork(struct work_struct *w);
577 static DECLARE_DEFERRABLE_WORK(stats_flush_dwork, flush_memcg_stats_dwork);
578 static DEFINE_SPINLOCK(stats_flush_lock);
579 static DEFINE_PER_CPU(unsigned int, stats_updates);
580 static atomic_t stats_flush_threshold = ATOMIC_INIT(0);
581 static u64 flush_next_time;
583 #define FLUSH_TIME (2UL*HZ)
586 * Accessors to ensure that preemption is disabled on PREEMPT_RT because it can
587 * not rely on this as part of an acquired spinlock_t lock. These functions are
588 * never used in hardirq context on PREEMPT_RT and therefore disabling preemtion
591 static void memcg_stats_lock(void)
593 preempt_disable_nested();
594 VM_WARN_ON_IRQS_ENABLED();
597 static void __memcg_stats_lock(void)
599 preempt_disable_nested();
602 static void memcg_stats_unlock(void)
604 preempt_enable_nested();
607 static inline void memcg_rstat_updated(struct mem_cgroup *memcg, int val)
611 cgroup_rstat_updated(memcg->css.cgroup, smp_processor_id());
613 x = __this_cpu_add_return(stats_updates, abs(val));
614 if (x > MEMCG_CHARGE_BATCH) {
616 * If stats_flush_threshold exceeds the threshold
617 * (>num_online_cpus()), cgroup stats update will be triggered
618 * in __mem_cgroup_flush_stats(). Increasing this var further
619 * is redundant and simply adds overhead in atomic update.
621 if (atomic_read(&stats_flush_threshold) <= num_online_cpus())
622 atomic_add(x / MEMCG_CHARGE_BATCH, &stats_flush_threshold);
623 __this_cpu_write(stats_updates, 0);
627 static void __mem_cgroup_flush_stats(void)
631 if (!spin_trylock_irqsave(&stats_flush_lock, flag))
634 flush_next_time = jiffies_64 + 2*FLUSH_TIME;
635 cgroup_rstat_flush_irqsafe(root_mem_cgroup->css.cgroup);
636 atomic_set(&stats_flush_threshold, 0);
637 spin_unlock_irqrestore(&stats_flush_lock, flag);
640 void mem_cgroup_flush_stats(void)
642 if (atomic_read(&stats_flush_threshold) > num_online_cpus())
643 __mem_cgroup_flush_stats();
646 void mem_cgroup_flush_stats_delayed(void)
648 if (time_after64(jiffies_64, flush_next_time))
649 mem_cgroup_flush_stats();
652 static void flush_memcg_stats_dwork(struct work_struct *w)
654 __mem_cgroup_flush_stats();
655 queue_delayed_work(system_unbound_wq, &stats_flush_dwork, FLUSH_TIME);
658 /* Subset of vm_event_item to report for memcg event stats */
659 static const unsigned int memcg_vm_event_stat[] = {
673 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
677 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
683 #define NR_MEMCG_EVENTS ARRAY_SIZE(memcg_vm_event_stat)
684 static int mem_cgroup_events_index[NR_VM_EVENT_ITEMS] __read_mostly;
686 static void init_memcg_events(void)
690 for (i = 0; i < NR_MEMCG_EVENTS; ++i)
691 mem_cgroup_events_index[memcg_vm_event_stat[i]] = i + 1;
694 static inline int memcg_events_index(enum vm_event_item idx)
696 return mem_cgroup_events_index[idx] - 1;
699 struct memcg_vmstats_percpu {
700 /* Local (CPU and cgroup) page state & events */
701 long state[MEMCG_NR_STAT];
702 unsigned long events[NR_MEMCG_EVENTS];
704 /* Delta calculation for lockless upward propagation */
705 long state_prev[MEMCG_NR_STAT];
706 unsigned long events_prev[NR_MEMCG_EVENTS];
708 /* Cgroup1: threshold notifications & softlimit tree updates */
709 unsigned long nr_page_events;
710 unsigned long targets[MEM_CGROUP_NTARGETS];
713 struct memcg_vmstats {
714 /* Aggregated (CPU and subtree) page state & events */
715 long state[MEMCG_NR_STAT];
716 unsigned long events[NR_MEMCG_EVENTS];
718 /* Pending child counts during tree propagation */
719 long state_pending[MEMCG_NR_STAT];
720 unsigned long events_pending[NR_MEMCG_EVENTS];
723 unsigned long memcg_page_state(struct mem_cgroup *memcg, int idx)
725 long x = READ_ONCE(memcg->vmstats->state[idx]);
734 * __mod_memcg_state - update cgroup memory statistics
735 * @memcg: the memory cgroup
736 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
737 * @val: delta to add to the counter, can be negative
739 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
741 if (mem_cgroup_disabled())
744 __this_cpu_add(memcg->vmstats_percpu->state[idx], val);
745 memcg_rstat_updated(memcg, val);
748 /* idx can be of type enum memcg_stat_item or node_stat_item. */
749 static unsigned long memcg_page_state_local(struct mem_cgroup *memcg, int idx)
754 for_each_possible_cpu(cpu)
755 x += per_cpu(memcg->vmstats_percpu->state[idx], cpu);
763 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
766 struct mem_cgroup_per_node *pn;
767 struct mem_cgroup *memcg;
769 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
773 * The caller from rmap relay on disabled preemption becase they never
774 * update their counter from in-interrupt context. For these two
775 * counters we check that the update is never performed from an
776 * interrupt context while other caller need to have disabled interrupt.
778 __memcg_stats_lock();
779 if (IS_ENABLED(CONFIG_DEBUG_VM)) {
784 case NR_SHMEM_PMDMAPPED:
785 case NR_FILE_PMDMAPPED:
786 WARN_ON_ONCE(!in_task());
789 VM_WARN_ON_IRQS_ENABLED();
794 __this_cpu_add(memcg->vmstats_percpu->state[idx], val);
797 __this_cpu_add(pn->lruvec_stats_percpu->state[idx], val);
799 memcg_rstat_updated(memcg, val);
800 memcg_stats_unlock();
804 * __mod_lruvec_state - update lruvec memory statistics
805 * @lruvec: the lruvec
806 * @idx: the stat item
807 * @val: delta to add to the counter, can be negative
809 * The lruvec is the intersection of the NUMA node and a cgroup. This
810 * function updates the all three counters that are affected by a
811 * change of state at this level: per-node, per-cgroup, per-lruvec.
813 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
817 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
819 /* Update memcg and lruvec */
820 if (!mem_cgroup_disabled())
821 __mod_memcg_lruvec_state(lruvec, idx, val);
824 void __mod_lruvec_page_state(struct page *page, enum node_stat_item idx,
827 struct page *head = compound_head(page); /* rmap on tail pages */
828 struct mem_cgroup *memcg;
829 pg_data_t *pgdat = page_pgdat(page);
830 struct lruvec *lruvec;
833 memcg = page_memcg(head);
834 /* Untracked pages have no memcg, no lruvec. Update only the node */
837 __mod_node_page_state(pgdat, idx, val);
841 lruvec = mem_cgroup_lruvec(memcg, pgdat);
842 __mod_lruvec_state(lruvec, idx, val);
845 EXPORT_SYMBOL(__mod_lruvec_page_state);
847 void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val)
849 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
850 struct mem_cgroup *memcg;
851 struct lruvec *lruvec;
854 memcg = mem_cgroup_from_slab_obj(p);
857 * Untracked pages have no memcg, no lruvec. Update only the
858 * node. If we reparent the slab objects to the root memcg,
859 * when we free the slab object, we need to update the per-memcg
860 * vmstats to keep it correct for the root memcg.
863 __mod_node_page_state(pgdat, idx, val);
865 lruvec = mem_cgroup_lruvec(memcg, pgdat);
866 __mod_lruvec_state(lruvec, idx, val);
872 * __count_memcg_events - account VM events in a cgroup
873 * @memcg: the memory cgroup
874 * @idx: the event item
875 * @count: the number of events that occurred
877 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
880 int index = memcg_events_index(idx);
882 if (mem_cgroup_disabled() || index < 0)
886 __this_cpu_add(memcg->vmstats_percpu->events[index], count);
887 memcg_rstat_updated(memcg, count);
888 memcg_stats_unlock();
891 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
893 int index = memcg_events_index(event);
897 return READ_ONCE(memcg->vmstats->events[index]);
900 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
904 int index = memcg_events_index(event);
909 for_each_possible_cpu(cpu)
910 x += per_cpu(memcg->vmstats_percpu->events[index], cpu);
914 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
917 /* pagein of a big page is an event. So, ignore page size */
919 __count_memcg_events(memcg, PGPGIN, 1);
921 __count_memcg_events(memcg, PGPGOUT, 1);
922 nr_pages = -nr_pages; /* for event */
925 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
928 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
929 enum mem_cgroup_events_target target)
931 unsigned long val, next;
933 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
934 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
935 /* from time_after() in jiffies.h */
936 if ((long)(next - val) < 0) {
938 case MEM_CGROUP_TARGET_THRESH:
939 next = val + THRESHOLDS_EVENTS_TARGET;
941 case MEM_CGROUP_TARGET_SOFTLIMIT:
942 next = val + SOFTLIMIT_EVENTS_TARGET;
947 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
954 * Check events in order.
957 static void memcg_check_events(struct mem_cgroup *memcg, int nid)
959 if (IS_ENABLED(CONFIG_PREEMPT_RT))
962 /* threshold event is triggered in finer grain than soft limit */
963 if (unlikely(mem_cgroup_event_ratelimit(memcg,
964 MEM_CGROUP_TARGET_THRESH))) {
967 do_softlimit = mem_cgroup_event_ratelimit(memcg,
968 MEM_CGROUP_TARGET_SOFTLIMIT);
969 mem_cgroup_threshold(memcg);
970 if (unlikely(do_softlimit))
971 mem_cgroup_update_tree(memcg, nid);
975 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
978 * mm_update_next_owner() may clear mm->owner to NULL
979 * if it races with swapoff, page migration, etc.
980 * So this can be called with p == NULL.
985 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
987 EXPORT_SYMBOL(mem_cgroup_from_task);
989 static __always_inline struct mem_cgroup *active_memcg(void)
992 return this_cpu_read(int_active_memcg);
994 return current->active_memcg;
998 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
999 * @mm: mm from which memcg should be extracted. It can be NULL.
1001 * Obtain a reference on mm->memcg and returns it if successful. If mm
1002 * is NULL, then the memcg is chosen as follows:
1003 * 1) The active memcg, if set.
1004 * 2) current->mm->memcg, if available
1006 * If mem_cgroup is disabled, NULL is returned.
1008 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1010 struct mem_cgroup *memcg;
1012 if (mem_cgroup_disabled())
1016 * Page cache insertions can happen without an
1017 * actual mm context, e.g. during disk probing
1018 * on boot, loopback IO, acct() writes etc.
1020 * No need to css_get on root memcg as the reference
1021 * counting is disabled on the root level in the
1022 * cgroup core. See CSS_NO_REF.
1024 if (unlikely(!mm)) {
1025 memcg = active_memcg();
1026 if (unlikely(memcg)) {
1027 /* remote memcg must hold a ref */
1028 css_get(&memcg->css);
1033 return root_mem_cgroup;
1038 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1039 if (unlikely(!memcg))
1040 memcg = root_mem_cgroup;
1041 } while (!css_tryget(&memcg->css));
1045 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1047 static __always_inline bool memcg_kmem_bypass(void)
1049 /* Allow remote memcg charging from any context. */
1050 if (unlikely(active_memcg()))
1053 /* Memcg to charge can't be determined. */
1054 if (!in_task() || !current->mm || (current->flags & PF_KTHREAD))
1061 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1062 * @root: hierarchy root
1063 * @prev: previously returned memcg, NULL on first invocation
1064 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1066 * Returns references to children of the hierarchy below @root, or
1067 * @root itself, or %NULL after a full round-trip.
1069 * Caller must pass the return value in @prev on subsequent
1070 * invocations for reference counting, or use mem_cgroup_iter_break()
1071 * to cancel a hierarchy walk before the round-trip is complete.
1073 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1074 * in the hierarchy among all concurrent reclaimers operating on the
1077 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1078 struct mem_cgroup *prev,
1079 struct mem_cgroup_reclaim_cookie *reclaim)
1081 struct mem_cgroup_reclaim_iter *iter;
1082 struct cgroup_subsys_state *css = NULL;
1083 struct mem_cgroup *memcg = NULL;
1084 struct mem_cgroup *pos = NULL;
1086 if (mem_cgroup_disabled())
1090 root = root_mem_cgroup;
1095 struct mem_cgroup_per_node *mz;
1097 mz = root->nodeinfo[reclaim->pgdat->node_id];
1101 * On start, join the current reclaim iteration cycle.
1102 * Exit when a concurrent walker completes it.
1105 reclaim->generation = iter->generation;
1106 else if (reclaim->generation != iter->generation)
1110 pos = READ_ONCE(iter->position);
1111 if (!pos || css_tryget(&pos->css))
1114 * css reference reached zero, so iter->position will
1115 * be cleared by ->css_released. However, we should not
1116 * rely on this happening soon, because ->css_released
1117 * is called from a work queue, and by busy-waiting we
1118 * might block it. So we clear iter->position right
1121 (void)cmpxchg(&iter->position, pos, NULL);
1131 css = css_next_descendant_pre(css, &root->css);
1134 * Reclaimers share the hierarchy walk, and a
1135 * new one might jump in right at the end of
1136 * the hierarchy - make sure they see at least
1137 * one group and restart from the beginning.
1145 * Verify the css and acquire a reference. The root
1146 * is provided by the caller, so we know it's alive
1147 * and kicking, and don't take an extra reference.
1149 if (css == &root->css || css_tryget(css)) {
1150 memcg = mem_cgroup_from_css(css);
1157 * The position could have already been updated by a competing
1158 * thread, so check that the value hasn't changed since we read
1159 * it to avoid reclaiming from the same cgroup twice.
1161 (void)cmpxchg(&iter->position, pos, memcg);
1172 if (prev && prev != root)
1173 css_put(&prev->css);
1179 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1180 * @root: hierarchy root
1181 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1183 void mem_cgroup_iter_break(struct mem_cgroup *root,
1184 struct mem_cgroup *prev)
1187 root = root_mem_cgroup;
1188 if (prev && prev != root)
1189 css_put(&prev->css);
1192 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1193 struct mem_cgroup *dead_memcg)
1195 struct mem_cgroup_reclaim_iter *iter;
1196 struct mem_cgroup_per_node *mz;
1199 for_each_node(nid) {
1200 mz = from->nodeinfo[nid];
1202 cmpxchg(&iter->position, dead_memcg, NULL);
1206 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1208 struct mem_cgroup *memcg = dead_memcg;
1209 struct mem_cgroup *last;
1212 __invalidate_reclaim_iterators(memcg, dead_memcg);
1214 } while ((memcg = parent_mem_cgroup(memcg)));
1217 * When cgroup1 non-hierarchy mode is used,
1218 * parent_mem_cgroup() does not walk all the way up to the
1219 * cgroup root (root_mem_cgroup). So we have to handle
1220 * dead_memcg from cgroup root separately.
1222 if (last != root_mem_cgroup)
1223 __invalidate_reclaim_iterators(root_mem_cgroup,
1228 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1229 * @memcg: hierarchy root
1230 * @fn: function to call for each task
1231 * @arg: argument passed to @fn
1233 * This function iterates over tasks attached to @memcg or to any of its
1234 * descendants and calls @fn for each task. If @fn returns a non-zero
1235 * value, the function breaks the iteration loop and returns the value.
1236 * Otherwise, it will iterate over all tasks and return 0.
1238 * This function must not be called for the root memory cgroup.
1240 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1241 int (*fn)(struct task_struct *, void *), void *arg)
1243 struct mem_cgroup *iter;
1246 BUG_ON(memcg == root_mem_cgroup);
1248 for_each_mem_cgroup_tree(iter, memcg) {
1249 struct css_task_iter it;
1250 struct task_struct *task;
1252 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1253 while (!ret && (task = css_task_iter_next(&it)))
1254 ret = fn(task, arg);
1255 css_task_iter_end(&it);
1257 mem_cgroup_iter_break(memcg, iter);
1264 #ifdef CONFIG_DEBUG_VM
1265 void lruvec_memcg_debug(struct lruvec *lruvec, struct folio *folio)
1267 struct mem_cgroup *memcg;
1269 if (mem_cgroup_disabled())
1272 memcg = folio_memcg(folio);
1275 VM_BUG_ON_FOLIO(lruvec_memcg(lruvec) != root_mem_cgroup, folio);
1277 VM_BUG_ON_FOLIO(lruvec_memcg(lruvec) != memcg, folio);
1282 * folio_lruvec_lock - Lock the lruvec for a folio.
1283 * @folio: Pointer to the folio.
1285 * These functions are safe to use under any of the following conditions:
1287 * - folio_test_lru false
1288 * - folio_memcg_lock()
1289 * - folio frozen (refcount of 0)
1291 * Return: The lruvec this folio is on with its lock held.
1293 struct lruvec *folio_lruvec_lock(struct folio *folio)
1295 struct lruvec *lruvec = folio_lruvec(folio);
1297 spin_lock(&lruvec->lru_lock);
1298 lruvec_memcg_debug(lruvec, folio);
1304 * folio_lruvec_lock_irq - Lock the lruvec for a folio.
1305 * @folio: Pointer to the folio.
1307 * These functions are safe to use under any of the following conditions:
1309 * - folio_test_lru false
1310 * - folio_memcg_lock()
1311 * - folio frozen (refcount of 0)
1313 * Return: The lruvec this folio is on with its lock held and interrupts
1316 struct lruvec *folio_lruvec_lock_irq(struct folio *folio)
1318 struct lruvec *lruvec = folio_lruvec(folio);
1320 spin_lock_irq(&lruvec->lru_lock);
1321 lruvec_memcg_debug(lruvec, folio);
1327 * folio_lruvec_lock_irqsave - Lock the lruvec for a folio.
1328 * @folio: Pointer to the folio.
1329 * @flags: Pointer to irqsave flags.
1331 * These functions are safe to use under any of the following conditions:
1333 * - folio_test_lru false
1334 * - folio_memcg_lock()
1335 * - folio frozen (refcount of 0)
1337 * Return: The lruvec this folio is on with its lock held and interrupts
1340 struct lruvec *folio_lruvec_lock_irqsave(struct folio *folio,
1341 unsigned long *flags)
1343 struct lruvec *lruvec = folio_lruvec(folio);
1345 spin_lock_irqsave(&lruvec->lru_lock, *flags);
1346 lruvec_memcg_debug(lruvec, folio);
1352 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1353 * @lruvec: mem_cgroup per zone lru vector
1354 * @lru: index of lru list the page is sitting on
1355 * @zid: zone id of the accounted pages
1356 * @nr_pages: positive when adding or negative when removing
1358 * This function must be called under lru_lock, just before a page is added
1359 * to or just after a page is removed from an lru list.
1361 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1362 int zid, int nr_pages)
1364 struct mem_cgroup_per_node *mz;
1365 unsigned long *lru_size;
1368 if (mem_cgroup_disabled())
1371 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1372 lru_size = &mz->lru_zone_size[zid][lru];
1375 *lru_size += nr_pages;
1378 if (WARN_ONCE(size < 0,
1379 "%s(%p, %d, %d): lru_size %ld\n",
1380 __func__, lruvec, lru, nr_pages, size)) {
1386 *lru_size += nr_pages;
1390 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1391 * @memcg: the memory cgroup
1393 * Returns the maximum amount of memory @mem can be charged with, in
1396 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1398 unsigned long margin = 0;
1399 unsigned long count;
1400 unsigned long limit;
1402 count = page_counter_read(&memcg->memory);
1403 limit = READ_ONCE(memcg->memory.max);
1405 margin = limit - count;
1407 if (do_memsw_account()) {
1408 count = page_counter_read(&memcg->memsw);
1409 limit = READ_ONCE(memcg->memsw.max);
1411 margin = min(margin, limit - count);
1420 * A routine for checking "mem" is under move_account() or not.
1422 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1423 * moving cgroups. This is for waiting at high-memory pressure
1426 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1428 struct mem_cgroup *from;
1429 struct mem_cgroup *to;
1432 * Unlike task_move routines, we access mc.to, mc.from not under
1433 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1435 spin_lock(&mc.lock);
1441 ret = mem_cgroup_is_descendant(from, memcg) ||
1442 mem_cgroup_is_descendant(to, memcg);
1444 spin_unlock(&mc.lock);
1448 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1450 if (mc.moving_task && current != mc.moving_task) {
1451 if (mem_cgroup_under_move(memcg)) {
1453 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1454 /* moving charge context might have finished. */
1457 finish_wait(&mc.waitq, &wait);
1464 struct memory_stat {
1469 static const struct memory_stat memory_stats[] = {
1470 { "anon", NR_ANON_MAPPED },
1471 { "file", NR_FILE_PAGES },
1472 { "kernel", MEMCG_KMEM },
1473 { "kernel_stack", NR_KERNEL_STACK_KB },
1474 { "pagetables", NR_PAGETABLE },
1475 { "sec_pagetables", NR_SECONDARY_PAGETABLE },
1476 { "percpu", MEMCG_PERCPU_B },
1477 { "sock", MEMCG_SOCK },
1478 { "vmalloc", MEMCG_VMALLOC },
1479 { "shmem", NR_SHMEM },
1480 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
1481 { "zswap", MEMCG_ZSWAP_B },
1482 { "zswapped", MEMCG_ZSWAPPED },
1484 { "file_mapped", NR_FILE_MAPPED },
1485 { "file_dirty", NR_FILE_DIRTY },
1486 { "file_writeback", NR_WRITEBACK },
1488 { "swapcached", NR_SWAPCACHE },
1490 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1491 { "anon_thp", NR_ANON_THPS },
1492 { "file_thp", NR_FILE_THPS },
1493 { "shmem_thp", NR_SHMEM_THPS },
1495 { "inactive_anon", NR_INACTIVE_ANON },
1496 { "active_anon", NR_ACTIVE_ANON },
1497 { "inactive_file", NR_INACTIVE_FILE },
1498 { "active_file", NR_ACTIVE_FILE },
1499 { "unevictable", NR_UNEVICTABLE },
1500 { "slab_reclaimable", NR_SLAB_RECLAIMABLE_B },
1501 { "slab_unreclaimable", NR_SLAB_UNRECLAIMABLE_B },
1503 /* The memory events */
1504 { "workingset_refault_anon", WORKINGSET_REFAULT_ANON },
1505 { "workingset_refault_file", WORKINGSET_REFAULT_FILE },
1506 { "workingset_activate_anon", WORKINGSET_ACTIVATE_ANON },
1507 { "workingset_activate_file", WORKINGSET_ACTIVATE_FILE },
1508 { "workingset_restore_anon", WORKINGSET_RESTORE_ANON },
1509 { "workingset_restore_file", WORKINGSET_RESTORE_FILE },
1510 { "workingset_nodereclaim", WORKINGSET_NODERECLAIM },
1513 /* Translate stat items to the correct unit for memory.stat output */
1514 static int memcg_page_state_unit(int item)
1517 case MEMCG_PERCPU_B:
1519 case NR_SLAB_RECLAIMABLE_B:
1520 case NR_SLAB_UNRECLAIMABLE_B:
1521 case WORKINGSET_REFAULT_ANON:
1522 case WORKINGSET_REFAULT_FILE:
1523 case WORKINGSET_ACTIVATE_ANON:
1524 case WORKINGSET_ACTIVATE_FILE:
1525 case WORKINGSET_RESTORE_ANON:
1526 case WORKINGSET_RESTORE_FILE:
1527 case WORKINGSET_NODERECLAIM:
1529 case NR_KERNEL_STACK_KB:
1536 static inline unsigned long memcg_page_state_output(struct mem_cgroup *memcg,
1539 return memcg_page_state(memcg, item) * memcg_page_state_unit(item);
1542 static void memory_stat_format(struct mem_cgroup *memcg, char *buf, int bufsize)
1547 seq_buf_init(&s, buf, bufsize);
1550 * Provide statistics on the state of the memory subsystem as
1551 * well as cumulative event counters that show past behavior.
1553 * This list is ordered following a combination of these gradients:
1554 * 1) generic big picture -> specifics and details
1555 * 2) reflecting userspace activity -> reflecting kernel heuristics
1557 * Current memory state:
1559 mem_cgroup_flush_stats();
1561 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1564 size = memcg_page_state_output(memcg, memory_stats[i].idx);
1565 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1567 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1568 size += memcg_page_state_output(memcg,
1569 NR_SLAB_RECLAIMABLE_B);
1570 seq_buf_printf(&s, "slab %llu\n", size);
1574 /* Accumulated memory events */
1575 seq_buf_printf(&s, "pgscan %lu\n",
1576 memcg_events(memcg, PGSCAN_KSWAPD) +
1577 memcg_events(memcg, PGSCAN_DIRECT));
1578 seq_buf_printf(&s, "pgsteal %lu\n",
1579 memcg_events(memcg, PGSTEAL_KSWAPD) +
1580 memcg_events(memcg, PGSTEAL_DIRECT));
1582 for (i = 0; i < ARRAY_SIZE(memcg_vm_event_stat); i++) {
1583 if (memcg_vm_event_stat[i] == PGPGIN ||
1584 memcg_vm_event_stat[i] == PGPGOUT)
1587 seq_buf_printf(&s, "%s %lu\n",
1588 vm_event_name(memcg_vm_event_stat[i]),
1589 memcg_events(memcg, memcg_vm_event_stat[i]));
1592 /* The above should easily fit into one page */
1593 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1596 #define K(x) ((x) << (PAGE_SHIFT-10))
1598 * mem_cgroup_print_oom_context: Print OOM information relevant to
1599 * memory controller.
1600 * @memcg: The memory cgroup that went over limit
1601 * @p: Task that is going to be killed
1603 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1606 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1611 pr_cont(",oom_memcg=");
1612 pr_cont_cgroup_path(memcg->css.cgroup);
1614 pr_cont(",global_oom");
1616 pr_cont(",task_memcg=");
1617 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1623 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1624 * memory controller.
1625 * @memcg: The memory cgroup that went over limit
1627 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1629 /* Use static buffer, for the caller is holding oom_lock. */
1630 static char buf[PAGE_SIZE];
1632 lockdep_assert_held(&oom_lock);
1634 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1635 K((u64)page_counter_read(&memcg->memory)),
1636 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1637 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1638 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1639 K((u64)page_counter_read(&memcg->swap)),
1640 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1642 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1643 K((u64)page_counter_read(&memcg->memsw)),
1644 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1645 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1646 K((u64)page_counter_read(&memcg->kmem)),
1647 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1650 pr_info("Memory cgroup stats for ");
1651 pr_cont_cgroup_path(memcg->css.cgroup);
1653 memory_stat_format(memcg, buf, sizeof(buf));
1658 * Return the memory (and swap, if configured) limit for a memcg.
1660 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1662 unsigned long max = READ_ONCE(memcg->memory.max);
1664 if (do_memsw_account()) {
1665 if (mem_cgroup_swappiness(memcg)) {
1666 /* Calculate swap excess capacity from memsw limit */
1667 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1669 max += min(swap, (unsigned long)total_swap_pages);
1672 if (mem_cgroup_swappiness(memcg))
1673 max += min(READ_ONCE(memcg->swap.max),
1674 (unsigned long)total_swap_pages);
1679 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1681 return page_counter_read(&memcg->memory);
1684 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1687 struct oom_control oc = {
1691 .gfp_mask = gfp_mask,
1696 if (mutex_lock_killable(&oom_lock))
1699 if (mem_cgroup_margin(memcg) >= (1 << order))
1703 * A few threads which were not waiting at mutex_lock_killable() can
1704 * fail to bail out. Therefore, check again after holding oom_lock.
1706 ret = task_is_dying() || out_of_memory(&oc);
1709 mutex_unlock(&oom_lock);
1713 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1716 unsigned long *total_scanned)
1718 struct mem_cgroup *victim = NULL;
1721 unsigned long excess;
1722 unsigned long nr_scanned;
1723 struct mem_cgroup_reclaim_cookie reclaim = {
1727 excess = soft_limit_excess(root_memcg);
1730 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1735 * If we have not been able to reclaim
1736 * anything, it might because there are
1737 * no reclaimable pages under this hierarchy
1742 * We want to do more targeted reclaim.
1743 * excess >> 2 is not to excessive so as to
1744 * reclaim too much, nor too less that we keep
1745 * coming back to reclaim from this cgroup
1747 if (total >= (excess >> 2) ||
1748 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1753 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1754 pgdat, &nr_scanned);
1755 *total_scanned += nr_scanned;
1756 if (!soft_limit_excess(root_memcg))
1759 mem_cgroup_iter_break(root_memcg, victim);
1763 #ifdef CONFIG_LOCKDEP
1764 static struct lockdep_map memcg_oom_lock_dep_map = {
1765 .name = "memcg_oom_lock",
1769 static DEFINE_SPINLOCK(memcg_oom_lock);
1772 * Check OOM-Killer is already running under our hierarchy.
1773 * If someone is running, return false.
1775 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1777 struct mem_cgroup *iter, *failed = NULL;
1779 spin_lock(&memcg_oom_lock);
1781 for_each_mem_cgroup_tree(iter, memcg) {
1782 if (iter->oom_lock) {
1784 * this subtree of our hierarchy is already locked
1785 * so we cannot give a lock.
1788 mem_cgroup_iter_break(memcg, iter);
1791 iter->oom_lock = true;
1796 * OK, we failed to lock the whole subtree so we have
1797 * to clean up what we set up to the failing subtree
1799 for_each_mem_cgroup_tree(iter, memcg) {
1800 if (iter == failed) {
1801 mem_cgroup_iter_break(memcg, iter);
1804 iter->oom_lock = false;
1807 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1809 spin_unlock(&memcg_oom_lock);
1814 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1816 struct mem_cgroup *iter;
1818 spin_lock(&memcg_oom_lock);
1819 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1820 for_each_mem_cgroup_tree(iter, memcg)
1821 iter->oom_lock = false;
1822 spin_unlock(&memcg_oom_lock);
1825 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1827 struct mem_cgroup *iter;
1829 spin_lock(&memcg_oom_lock);
1830 for_each_mem_cgroup_tree(iter, memcg)
1832 spin_unlock(&memcg_oom_lock);
1835 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1837 struct mem_cgroup *iter;
1840 * Be careful about under_oom underflows because a child memcg
1841 * could have been added after mem_cgroup_mark_under_oom.
1843 spin_lock(&memcg_oom_lock);
1844 for_each_mem_cgroup_tree(iter, memcg)
1845 if (iter->under_oom > 0)
1847 spin_unlock(&memcg_oom_lock);
1850 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1852 struct oom_wait_info {
1853 struct mem_cgroup *memcg;
1854 wait_queue_entry_t wait;
1857 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1858 unsigned mode, int sync, void *arg)
1860 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1861 struct mem_cgroup *oom_wait_memcg;
1862 struct oom_wait_info *oom_wait_info;
1864 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1865 oom_wait_memcg = oom_wait_info->memcg;
1867 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1868 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1870 return autoremove_wake_function(wait, mode, sync, arg);
1873 static void memcg_oom_recover(struct mem_cgroup *memcg)
1876 * For the following lockless ->under_oom test, the only required
1877 * guarantee is that it must see the state asserted by an OOM when
1878 * this function is called as a result of userland actions
1879 * triggered by the notification of the OOM. This is trivially
1880 * achieved by invoking mem_cgroup_mark_under_oom() before
1881 * triggering notification.
1883 if (memcg && memcg->under_oom)
1884 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1888 * Returns true if successfully killed one or more processes. Though in some
1889 * corner cases it can return true even without killing any process.
1891 static bool mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1895 if (order > PAGE_ALLOC_COSTLY_ORDER)
1898 memcg_memory_event(memcg, MEMCG_OOM);
1901 * We are in the middle of the charge context here, so we
1902 * don't want to block when potentially sitting on a callstack
1903 * that holds all kinds of filesystem and mm locks.
1905 * cgroup1 allows disabling the OOM killer and waiting for outside
1906 * handling until the charge can succeed; remember the context and put
1907 * the task to sleep at the end of the page fault when all locks are
1910 * On the other hand, in-kernel OOM killer allows for an async victim
1911 * memory reclaim (oom_reaper) and that means that we are not solely
1912 * relying on the oom victim to make a forward progress and we can
1913 * invoke the oom killer here.
1915 * Please note that mem_cgroup_out_of_memory might fail to find a
1916 * victim and then we have to bail out from the charge path.
1918 if (memcg->oom_kill_disable) {
1919 if (current->in_user_fault) {
1920 css_get(&memcg->css);
1921 current->memcg_in_oom = memcg;
1922 current->memcg_oom_gfp_mask = mask;
1923 current->memcg_oom_order = order;
1928 mem_cgroup_mark_under_oom(memcg);
1930 locked = mem_cgroup_oom_trylock(memcg);
1933 mem_cgroup_oom_notify(memcg);
1935 mem_cgroup_unmark_under_oom(memcg);
1936 ret = mem_cgroup_out_of_memory(memcg, mask, order);
1939 mem_cgroup_oom_unlock(memcg);
1945 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1946 * @handle: actually kill/wait or just clean up the OOM state
1948 * This has to be called at the end of a page fault if the memcg OOM
1949 * handler was enabled.
1951 * Memcg supports userspace OOM handling where failed allocations must
1952 * sleep on a waitqueue until the userspace task resolves the
1953 * situation. Sleeping directly in the charge context with all kinds
1954 * of locks held is not a good idea, instead we remember an OOM state
1955 * in the task and mem_cgroup_oom_synchronize() has to be called at
1956 * the end of the page fault to complete the OOM handling.
1958 * Returns %true if an ongoing memcg OOM situation was detected and
1959 * completed, %false otherwise.
1961 bool mem_cgroup_oom_synchronize(bool handle)
1963 struct mem_cgroup *memcg = current->memcg_in_oom;
1964 struct oom_wait_info owait;
1967 /* OOM is global, do not handle */
1974 owait.memcg = memcg;
1975 owait.wait.flags = 0;
1976 owait.wait.func = memcg_oom_wake_function;
1977 owait.wait.private = current;
1978 INIT_LIST_HEAD(&owait.wait.entry);
1980 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1981 mem_cgroup_mark_under_oom(memcg);
1983 locked = mem_cgroup_oom_trylock(memcg);
1986 mem_cgroup_oom_notify(memcg);
1988 if (locked && !memcg->oom_kill_disable) {
1989 mem_cgroup_unmark_under_oom(memcg);
1990 finish_wait(&memcg_oom_waitq, &owait.wait);
1991 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1992 current->memcg_oom_order);
1995 mem_cgroup_unmark_under_oom(memcg);
1996 finish_wait(&memcg_oom_waitq, &owait.wait);
2000 mem_cgroup_oom_unlock(memcg);
2002 * There is no guarantee that an OOM-lock contender
2003 * sees the wakeups triggered by the OOM kill
2004 * uncharges. Wake any sleepers explicitly.
2006 memcg_oom_recover(memcg);
2009 current->memcg_in_oom = NULL;
2010 css_put(&memcg->css);
2015 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2016 * @victim: task to be killed by the OOM killer
2017 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2019 * Returns a pointer to a memory cgroup, which has to be cleaned up
2020 * by killing all belonging OOM-killable tasks.
2022 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2024 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2025 struct mem_cgroup *oom_domain)
2027 struct mem_cgroup *oom_group = NULL;
2028 struct mem_cgroup *memcg;
2030 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2034 oom_domain = root_mem_cgroup;
2038 memcg = mem_cgroup_from_task(victim);
2039 if (memcg == root_mem_cgroup)
2043 * If the victim task has been asynchronously moved to a different
2044 * memory cgroup, we might end up killing tasks outside oom_domain.
2045 * In this case it's better to ignore memory.group.oom.
2047 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2051 * Traverse the memory cgroup hierarchy from the victim task's
2052 * cgroup up to the OOMing cgroup (or root) to find the
2053 * highest-level memory cgroup with oom.group set.
2055 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2056 if (memcg->oom_group)
2059 if (memcg == oom_domain)
2064 css_get(&oom_group->css);
2071 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2073 pr_info("Tasks in ");
2074 pr_cont_cgroup_path(memcg->css.cgroup);
2075 pr_cont(" are going to be killed due to memory.oom.group set\n");
2079 * folio_memcg_lock - Bind a folio to its memcg.
2080 * @folio: The folio.
2082 * This function prevents unlocked LRU folios from being moved to
2085 * It ensures lifetime of the bound memcg. The caller is responsible
2086 * for the lifetime of the folio.
2088 void folio_memcg_lock(struct folio *folio)
2090 struct mem_cgroup *memcg;
2091 unsigned long flags;
2094 * The RCU lock is held throughout the transaction. The fast
2095 * path can get away without acquiring the memcg->move_lock
2096 * because page moving starts with an RCU grace period.
2100 if (mem_cgroup_disabled())
2103 memcg = folio_memcg(folio);
2104 if (unlikely(!memcg))
2107 #ifdef CONFIG_PROVE_LOCKING
2108 local_irq_save(flags);
2109 might_lock(&memcg->move_lock);
2110 local_irq_restore(flags);
2113 if (atomic_read(&memcg->moving_account) <= 0)
2116 spin_lock_irqsave(&memcg->move_lock, flags);
2117 if (memcg != folio_memcg(folio)) {
2118 spin_unlock_irqrestore(&memcg->move_lock, flags);
2123 * When charge migration first begins, we can have multiple
2124 * critical sections holding the fast-path RCU lock and one
2125 * holding the slowpath move_lock. Track the task who has the
2126 * move_lock for unlock_page_memcg().
2128 memcg->move_lock_task = current;
2129 memcg->move_lock_flags = flags;
2132 void lock_page_memcg(struct page *page)
2134 folio_memcg_lock(page_folio(page));
2137 static void __folio_memcg_unlock(struct mem_cgroup *memcg)
2139 if (memcg && memcg->move_lock_task == current) {
2140 unsigned long flags = memcg->move_lock_flags;
2142 memcg->move_lock_task = NULL;
2143 memcg->move_lock_flags = 0;
2145 spin_unlock_irqrestore(&memcg->move_lock, flags);
2152 * folio_memcg_unlock - Release the binding between a folio and its memcg.
2153 * @folio: The folio.
2155 * This releases the binding created by folio_memcg_lock(). This does
2156 * not change the accounting of this folio to its memcg, but it does
2157 * permit others to change it.
2159 void folio_memcg_unlock(struct folio *folio)
2161 __folio_memcg_unlock(folio_memcg(folio));
2164 void unlock_page_memcg(struct page *page)
2166 folio_memcg_unlock(page_folio(page));
2169 struct memcg_stock_pcp {
2170 local_lock_t stock_lock;
2171 struct mem_cgroup *cached; /* this never be root cgroup */
2172 unsigned int nr_pages;
2174 #ifdef CONFIG_MEMCG_KMEM
2175 struct obj_cgroup *cached_objcg;
2176 struct pglist_data *cached_pgdat;
2177 unsigned int nr_bytes;
2178 int nr_slab_reclaimable_b;
2179 int nr_slab_unreclaimable_b;
2182 struct work_struct work;
2183 unsigned long flags;
2184 #define FLUSHING_CACHED_CHARGE 0
2186 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock) = {
2187 .stock_lock = INIT_LOCAL_LOCK(stock_lock),
2189 static DEFINE_MUTEX(percpu_charge_mutex);
2191 #ifdef CONFIG_MEMCG_KMEM
2192 static struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock);
2193 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2194 struct mem_cgroup *root_memcg);
2195 static void memcg_account_kmem(struct mem_cgroup *memcg, int nr_pages);
2198 static inline struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock)
2202 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2203 struct mem_cgroup *root_memcg)
2207 static void memcg_account_kmem(struct mem_cgroup *memcg, int nr_pages)
2213 * consume_stock: Try to consume stocked charge on this cpu.
2214 * @memcg: memcg to consume from.
2215 * @nr_pages: how many pages to charge.
2217 * The charges will only happen if @memcg matches the current cpu's memcg
2218 * stock, and at least @nr_pages are available in that stock. Failure to
2219 * service an allocation will refill the stock.
2221 * returns true if successful, false otherwise.
2223 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2225 struct memcg_stock_pcp *stock;
2226 unsigned long flags;
2229 if (nr_pages > MEMCG_CHARGE_BATCH)
2232 local_lock_irqsave(&memcg_stock.stock_lock, flags);
2234 stock = this_cpu_ptr(&memcg_stock);
2235 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2236 stock->nr_pages -= nr_pages;
2240 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2246 * Returns stocks cached in percpu and reset cached information.
2248 static void drain_stock(struct memcg_stock_pcp *stock)
2250 struct mem_cgroup *old = stock->cached;
2255 if (stock->nr_pages) {
2256 page_counter_uncharge(&old->memory, stock->nr_pages);
2257 if (do_memsw_account())
2258 page_counter_uncharge(&old->memsw, stock->nr_pages);
2259 stock->nr_pages = 0;
2263 stock->cached = NULL;
2266 static void drain_local_stock(struct work_struct *dummy)
2268 struct memcg_stock_pcp *stock;
2269 struct obj_cgroup *old = NULL;
2270 unsigned long flags;
2273 * The only protection from cpu hotplug (memcg_hotplug_cpu_dead) vs.
2274 * drain_stock races is that we always operate on local CPU stock
2275 * here with IRQ disabled
2277 local_lock_irqsave(&memcg_stock.stock_lock, flags);
2279 stock = this_cpu_ptr(&memcg_stock);
2280 old = drain_obj_stock(stock);
2282 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2284 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2286 obj_cgroup_put(old);
2290 * Cache charges(val) to local per_cpu area.
2291 * This will be consumed by consume_stock() function, later.
2293 static void __refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2295 struct memcg_stock_pcp *stock;
2297 stock = this_cpu_ptr(&memcg_stock);
2298 if (stock->cached != memcg) { /* reset if necessary */
2300 css_get(&memcg->css);
2301 stock->cached = memcg;
2303 stock->nr_pages += nr_pages;
2305 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2309 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2311 unsigned long flags;
2313 local_lock_irqsave(&memcg_stock.stock_lock, flags);
2314 __refill_stock(memcg, nr_pages);
2315 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2319 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2320 * of the hierarchy under it.
2322 static void drain_all_stock(struct mem_cgroup *root_memcg)
2326 /* If someone's already draining, avoid adding running more workers. */
2327 if (!mutex_trylock(&percpu_charge_mutex))
2330 * Notify other cpus that system-wide "drain" is running
2331 * We do not care about races with the cpu hotplug because cpu down
2332 * as well as workers from this path always operate on the local
2333 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2336 curcpu = smp_processor_id();
2337 for_each_online_cpu(cpu) {
2338 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2339 struct mem_cgroup *memcg;
2343 memcg = stock->cached;
2344 if (memcg && stock->nr_pages &&
2345 mem_cgroup_is_descendant(memcg, root_memcg))
2347 else if (obj_stock_flush_required(stock, root_memcg))
2352 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2354 drain_local_stock(&stock->work);
2356 schedule_work_on(cpu, &stock->work);
2360 mutex_unlock(&percpu_charge_mutex);
2363 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2365 struct memcg_stock_pcp *stock;
2367 stock = &per_cpu(memcg_stock, cpu);
2373 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2374 unsigned int nr_pages,
2377 unsigned long nr_reclaimed = 0;
2380 unsigned long pflags;
2382 if (page_counter_read(&memcg->memory) <=
2383 READ_ONCE(memcg->memory.high))
2386 memcg_memory_event(memcg, MEMCG_HIGH);
2388 psi_memstall_enter(&pflags);
2389 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2391 MEMCG_RECLAIM_MAY_SWAP);
2392 psi_memstall_leave(&pflags);
2393 } while ((memcg = parent_mem_cgroup(memcg)) &&
2394 !mem_cgroup_is_root(memcg));
2396 return nr_reclaimed;
2399 static void high_work_func(struct work_struct *work)
2401 struct mem_cgroup *memcg;
2403 memcg = container_of(work, struct mem_cgroup, high_work);
2404 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2408 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2409 * enough to still cause a significant slowdown in most cases, while still
2410 * allowing diagnostics and tracing to proceed without becoming stuck.
2412 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2415 * When calculating the delay, we use these either side of the exponentiation to
2416 * maintain precision and scale to a reasonable number of jiffies (see the table
2419 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2420 * overage ratio to a delay.
2421 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2422 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2423 * to produce a reasonable delay curve.
2425 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2426 * reasonable delay curve compared to precision-adjusted overage, not
2427 * penalising heavily at first, but still making sure that growth beyond the
2428 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2429 * example, with a high of 100 megabytes:
2431 * +-------+------------------------+
2432 * | usage | time to allocate in ms |
2433 * +-------+------------------------+
2455 * +-------+------------------------+
2457 #define MEMCG_DELAY_PRECISION_SHIFT 20
2458 #define MEMCG_DELAY_SCALING_SHIFT 14
2460 static u64 calculate_overage(unsigned long usage, unsigned long high)
2468 * Prevent division by 0 in overage calculation by acting as if
2469 * it was a threshold of 1 page
2471 high = max(high, 1UL);
2473 overage = usage - high;
2474 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2475 return div64_u64(overage, high);
2478 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2480 u64 overage, max_overage = 0;
2483 overage = calculate_overage(page_counter_read(&memcg->memory),
2484 READ_ONCE(memcg->memory.high));
2485 max_overage = max(overage, max_overage);
2486 } while ((memcg = parent_mem_cgroup(memcg)) &&
2487 !mem_cgroup_is_root(memcg));
2492 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2494 u64 overage, max_overage = 0;
2497 overage = calculate_overage(page_counter_read(&memcg->swap),
2498 READ_ONCE(memcg->swap.high));
2500 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2501 max_overage = max(overage, max_overage);
2502 } while ((memcg = parent_mem_cgroup(memcg)) &&
2503 !mem_cgroup_is_root(memcg));
2509 * Get the number of jiffies that we should penalise a mischievous cgroup which
2510 * is exceeding its memory.high by checking both it and its ancestors.
2512 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2513 unsigned int nr_pages,
2516 unsigned long penalty_jiffies;
2522 * We use overage compared to memory.high to calculate the number of
2523 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2524 * fairly lenient on small overages, and increasingly harsh when the
2525 * memcg in question makes it clear that it has no intention of stopping
2526 * its crazy behaviour, so we exponentially increase the delay based on
2529 penalty_jiffies = max_overage * max_overage * HZ;
2530 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2531 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2534 * Factor in the task's own contribution to the overage, such that four
2535 * N-sized allocations are throttled approximately the same as one
2536 * 4N-sized allocation.
2538 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2539 * larger the current charge patch is than that.
2541 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2545 * Scheduled by try_charge() to be executed from the userland return path
2546 * and reclaims memory over the high limit.
2548 void mem_cgroup_handle_over_high(gfp_t gfp_mask)
2550 unsigned long penalty_jiffies;
2551 unsigned long pflags;
2552 unsigned long nr_reclaimed;
2553 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2554 int nr_retries = MAX_RECLAIM_RETRIES;
2555 struct mem_cgroup *memcg;
2556 bool in_retry = false;
2558 if (likely(!nr_pages))
2561 memcg = get_mem_cgroup_from_mm(current->mm);
2562 current->memcg_nr_pages_over_high = 0;
2566 * The allocating task should reclaim at least the batch size, but for
2567 * subsequent retries we only want to do what's necessary to prevent oom
2568 * or breaching resource isolation.
2570 * This is distinct from memory.max or page allocator behaviour because
2571 * memory.high is currently batched, whereas memory.max and the page
2572 * allocator run every time an allocation is made.
2574 nr_reclaimed = reclaim_high(memcg,
2575 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2579 * memory.high is breached and reclaim is unable to keep up. Throttle
2580 * allocators proactively to slow down excessive growth.
2582 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2583 mem_find_max_overage(memcg));
2585 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2586 swap_find_max_overage(memcg));
2589 * Clamp the max delay per usermode return so as to still keep the
2590 * application moving forwards and also permit diagnostics, albeit
2593 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2596 * Don't sleep if the amount of jiffies this memcg owes us is so low
2597 * that it's not even worth doing, in an attempt to be nice to those who
2598 * go only a small amount over their memory.high value and maybe haven't
2599 * been aggressively reclaimed enough yet.
2601 if (penalty_jiffies <= HZ / 100)
2605 * If reclaim is making forward progress but we're still over
2606 * memory.high, we want to encourage that rather than doing allocator
2609 if (nr_reclaimed || nr_retries--) {
2615 * If we exit early, we're guaranteed to die (since
2616 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2617 * need to account for any ill-begotten jiffies to pay them off later.
2619 psi_memstall_enter(&pflags);
2620 schedule_timeout_killable(penalty_jiffies);
2621 psi_memstall_leave(&pflags);
2624 css_put(&memcg->css);
2627 static int try_charge_memcg(struct mem_cgroup *memcg, gfp_t gfp_mask,
2628 unsigned int nr_pages)
2630 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2631 int nr_retries = MAX_RECLAIM_RETRIES;
2632 struct mem_cgroup *mem_over_limit;
2633 struct page_counter *counter;
2634 unsigned long nr_reclaimed;
2635 bool passed_oom = false;
2636 unsigned int reclaim_options = MEMCG_RECLAIM_MAY_SWAP;
2637 bool drained = false;
2638 bool raised_max_event = false;
2639 unsigned long pflags;
2642 if (consume_stock(memcg, nr_pages))
2645 if (!do_memsw_account() ||
2646 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2647 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2649 if (do_memsw_account())
2650 page_counter_uncharge(&memcg->memsw, batch);
2651 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2653 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2654 reclaim_options &= ~MEMCG_RECLAIM_MAY_SWAP;
2657 if (batch > nr_pages) {
2663 * Prevent unbounded recursion when reclaim operations need to
2664 * allocate memory. This might exceed the limits temporarily,
2665 * but we prefer facilitating memory reclaim and getting back
2666 * under the limit over triggering OOM kills in these cases.
2668 if (unlikely(current->flags & PF_MEMALLOC))
2671 if (unlikely(task_in_memcg_oom(current)))
2674 if (!gfpflags_allow_blocking(gfp_mask))
2677 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2678 raised_max_event = true;
2680 psi_memstall_enter(&pflags);
2681 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2682 gfp_mask, reclaim_options);
2683 psi_memstall_leave(&pflags);
2685 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2689 drain_all_stock(mem_over_limit);
2694 if (gfp_mask & __GFP_NORETRY)
2697 * Even though the limit is exceeded at this point, reclaim
2698 * may have been able to free some pages. Retry the charge
2699 * before killing the task.
2701 * Only for regular pages, though: huge pages are rather
2702 * unlikely to succeed so close to the limit, and we fall back
2703 * to regular pages anyway in case of failure.
2705 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2708 * At task move, charge accounts can be doubly counted. So, it's
2709 * better to wait until the end of task_move if something is going on.
2711 if (mem_cgroup_wait_acct_move(mem_over_limit))
2717 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2720 /* Avoid endless loop for tasks bypassed by the oom killer */
2721 if (passed_oom && task_is_dying())
2725 * keep retrying as long as the memcg oom killer is able to make
2726 * a forward progress or bypass the charge if the oom killer
2727 * couldn't make any progress.
2729 if (mem_cgroup_oom(mem_over_limit, gfp_mask,
2730 get_order(nr_pages * PAGE_SIZE))) {
2732 nr_retries = MAX_RECLAIM_RETRIES;
2737 * Memcg doesn't have a dedicated reserve for atomic
2738 * allocations. But like the global atomic pool, we need to
2739 * put the burden of reclaim on regular allocation requests
2740 * and let these go through as privileged allocations.
2742 if (!(gfp_mask & (__GFP_NOFAIL | __GFP_HIGH)))
2746 * If the allocation has to be enforced, don't forget to raise
2747 * a MEMCG_MAX event.
2749 if (!raised_max_event)
2750 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2753 * The allocation either can't fail or will lead to more memory
2754 * being freed very soon. Allow memory usage go over the limit
2755 * temporarily by force charging it.
2757 page_counter_charge(&memcg->memory, nr_pages);
2758 if (do_memsw_account())
2759 page_counter_charge(&memcg->memsw, nr_pages);
2764 if (batch > nr_pages)
2765 refill_stock(memcg, batch - nr_pages);
2768 * If the hierarchy is above the normal consumption range, schedule
2769 * reclaim on returning to userland. We can perform reclaim here
2770 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2771 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2772 * not recorded as it most likely matches current's and won't
2773 * change in the meantime. As high limit is checked again before
2774 * reclaim, the cost of mismatch is negligible.
2777 bool mem_high, swap_high;
2779 mem_high = page_counter_read(&memcg->memory) >
2780 READ_ONCE(memcg->memory.high);
2781 swap_high = page_counter_read(&memcg->swap) >
2782 READ_ONCE(memcg->swap.high);
2784 /* Don't bother a random interrupted task */
2787 schedule_work(&memcg->high_work);
2793 if (mem_high || swap_high) {
2795 * The allocating tasks in this cgroup will need to do
2796 * reclaim or be throttled to prevent further growth
2797 * of the memory or swap footprints.
2799 * Target some best-effort fairness between the tasks,
2800 * and distribute reclaim work and delay penalties
2801 * based on how much each task is actually allocating.
2803 current->memcg_nr_pages_over_high += batch;
2804 set_notify_resume(current);
2807 } while ((memcg = parent_mem_cgroup(memcg)));
2809 if (current->memcg_nr_pages_over_high > MEMCG_CHARGE_BATCH &&
2810 !(current->flags & PF_MEMALLOC) &&
2811 gfpflags_allow_blocking(gfp_mask)) {
2812 mem_cgroup_handle_over_high(gfp_mask);
2817 static inline int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2818 unsigned int nr_pages)
2820 if (mem_cgroup_is_root(memcg))
2823 return try_charge_memcg(memcg, gfp_mask, nr_pages);
2826 static inline void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2828 if (mem_cgroup_is_root(memcg))
2831 page_counter_uncharge(&memcg->memory, nr_pages);
2832 if (do_memsw_account())
2833 page_counter_uncharge(&memcg->memsw, nr_pages);
2836 static void commit_charge(struct folio *folio, struct mem_cgroup *memcg)
2838 VM_BUG_ON_FOLIO(folio_memcg(folio), folio);
2840 * Any of the following ensures page's memcg stability:
2844 * - lock_page_memcg()
2845 * - exclusive reference
2846 * - mem_cgroup_trylock_pages()
2848 folio->memcg_data = (unsigned long)memcg;
2851 #ifdef CONFIG_MEMCG_KMEM
2853 * The allocated objcg pointers array is not accounted directly.
2854 * Moreover, it should not come from DMA buffer and is not readily
2855 * reclaimable. So those GFP bits should be masked off.
2857 #define OBJCGS_CLEAR_MASK (__GFP_DMA | __GFP_RECLAIMABLE | \
2858 __GFP_ACCOUNT | __GFP_NOFAIL)
2861 * mod_objcg_mlstate() may be called with irq enabled, so
2862 * mod_memcg_lruvec_state() should be used.
2864 static inline void mod_objcg_mlstate(struct obj_cgroup *objcg,
2865 struct pglist_data *pgdat,
2866 enum node_stat_item idx, int nr)
2868 struct mem_cgroup *memcg;
2869 struct lruvec *lruvec;
2872 memcg = obj_cgroup_memcg(objcg);
2873 lruvec = mem_cgroup_lruvec(memcg, pgdat);
2874 mod_memcg_lruvec_state(lruvec, idx, nr);
2878 int memcg_alloc_slab_cgroups(struct slab *slab, struct kmem_cache *s,
2879 gfp_t gfp, bool new_slab)
2881 unsigned int objects = objs_per_slab(s, slab);
2882 unsigned long memcg_data;
2885 gfp &= ~OBJCGS_CLEAR_MASK;
2886 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2891 memcg_data = (unsigned long) vec | MEMCG_DATA_OBJCGS;
2894 * If the slab is brand new and nobody can yet access its
2895 * memcg_data, no synchronization is required and memcg_data can
2896 * be simply assigned.
2898 slab->memcg_data = memcg_data;
2899 } else if (cmpxchg(&slab->memcg_data, 0, memcg_data)) {
2901 * If the slab is already in use, somebody can allocate and
2902 * assign obj_cgroups in parallel. In this case the existing
2903 * objcg vector should be reused.
2909 kmemleak_not_leak(vec);
2913 static __always_inline
2914 struct mem_cgroup *mem_cgroup_from_obj_folio(struct folio *folio, void *p)
2917 * Slab objects are accounted individually, not per-page.
2918 * Memcg membership data for each individual object is saved in
2921 if (folio_test_slab(folio)) {
2922 struct obj_cgroup **objcgs;
2926 slab = folio_slab(folio);
2927 objcgs = slab_objcgs(slab);
2931 off = obj_to_index(slab->slab_cache, slab, p);
2933 return obj_cgroup_memcg(objcgs[off]);
2939 * page_memcg_check() is used here, because in theory we can encounter
2940 * a folio where the slab flag has been cleared already, but
2941 * slab->memcg_data has not been freed yet
2942 * page_memcg_check(page) will guarantee that a proper memory
2943 * cgroup pointer or NULL will be returned.
2945 return page_memcg_check(folio_page(folio, 0));
2949 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2951 * A passed kernel object can be a slab object, vmalloc object or a generic
2952 * kernel page, so different mechanisms for getting the memory cgroup pointer
2955 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2956 * can not know for sure how the kernel object is implemented.
2957 * mem_cgroup_from_obj() can be safely used in such cases.
2959 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2960 * cgroup_mutex, etc.
2962 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2964 struct folio *folio;
2966 if (mem_cgroup_disabled())
2969 if (unlikely(is_vmalloc_addr(p)))
2970 folio = page_folio(vmalloc_to_page(p));
2972 folio = virt_to_folio(p);
2974 return mem_cgroup_from_obj_folio(folio, p);
2978 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2979 * Similar to mem_cgroup_from_obj(), but faster and not suitable for objects,
2980 * allocated using vmalloc().
2982 * A passed kernel object must be a slab object or a generic kernel page.
2984 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2985 * cgroup_mutex, etc.
2987 struct mem_cgroup *mem_cgroup_from_slab_obj(void *p)
2989 if (mem_cgroup_disabled())
2992 return mem_cgroup_from_obj_folio(virt_to_folio(p), p);
2995 static struct obj_cgroup *__get_obj_cgroup_from_memcg(struct mem_cgroup *memcg)
2997 struct obj_cgroup *objcg = NULL;
2999 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
3000 objcg = rcu_dereference(memcg->objcg);
3001 if (objcg && obj_cgroup_tryget(objcg))
3008 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
3010 struct obj_cgroup *objcg = NULL;
3011 struct mem_cgroup *memcg;
3013 if (memcg_kmem_bypass())
3017 if (unlikely(active_memcg()))
3018 memcg = active_memcg();
3020 memcg = mem_cgroup_from_task(current);
3021 objcg = __get_obj_cgroup_from_memcg(memcg);
3026 struct obj_cgroup *get_obj_cgroup_from_page(struct page *page)
3028 struct obj_cgroup *objcg;
3030 if (!memcg_kmem_enabled())
3033 if (PageMemcgKmem(page)) {
3034 objcg = __folio_objcg(page_folio(page));
3035 obj_cgroup_get(objcg);
3037 struct mem_cgroup *memcg;
3040 memcg = __folio_memcg(page_folio(page));
3042 objcg = __get_obj_cgroup_from_memcg(memcg);
3050 static void memcg_account_kmem(struct mem_cgroup *memcg, int nr_pages)
3052 mod_memcg_state(memcg, MEMCG_KMEM, nr_pages);
3053 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
3055 page_counter_charge(&memcg->kmem, nr_pages);
3057 page_counter_uncharge(&memcg->kmem, -nr_pages);
3063 * obj_cgroup_uncharge_pages: uncharge a number of kernel pages from a objcg
3064 * @objcg: object cgroup to uncharge
3065 * @nr_pages: number of pages to uncharge
3067 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
3068 unsigned int nr_pages)
3070 struct mem_cgroup *memcg;
3072 memcg = get_mem_cgroup_from_objcg(objcg);
3074 memcg_account_kmem(memcg, -nr_pages);
3075 refill_stock(memcg, nr_pages);
3077 css_put(&memcg->css);
3081 * obj_cgroup_charge_pages: charge a number of kernel pages to a objcg
3082 * @objcg: object cgroup to charge
3083 * @gfp: reclaim mode
3084 * @nr_pages: number of pages to charge
3086 * Returns 0 on success, an error code on failure.
3088 static int obj_cgroup_charge_pages(struct obj_cgroup *objcg, gfp_t gfp,
3089 unsigned int nr_pages)
3091 struct mem_cgroup *memcg;
3094 memcg = get_mem_cgroup_from_objcg(objcg);
3096 ret = try_charge_memcg(memcg, gfp, nr_pages);
3100 memcg_account_kmem(memcg, nr_pages);
3102 css_put(&memcg->css);
3108 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3109 * @page: page to charge
3110 * @gfp: reclaim mode
3111 * @order: allocation order
3113 * Returns 0 on success, an error code on failure.
3115 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3117 struct obj_cgroup *objcg;
3120 objcg = get_obj_cgroup_from_current();
3122 ret = obj_cgroup_charge_pages(objcg, gfp, 1 << order);
3124 page->memcg_data = (unsigned long)objcg |
3128 obj_cgroup_put(objcg);
3134 * __memcg_kmem_uncharge_page: uncharge a kmem page
3135 * @page: page to uncharge
3136 * @order: allocation order
3138 void __memcg_kmem_uncharge_page(struct page *page, int order)
3140 struct folio *folio = page_folio(page);
3141 struct obj_cgroup *objcg;
3142 unsigned int nr_pages = 1 << order;
3144 if (!folio_memcg_kmem(folio))
3147 objcg = __folio_objcg(folio);
3148 obj_cgroup_uncharge_pages(objcg, nr_pages);
3149 folio->memcg_data = 0;
3150 obj_cgroup_put(objcg);
3153 void mod_objcg_state(struct obj_cgroup *objcg, struct pglist_data *pgdat,
3154 enum node_stat_item idx, int nr)
3156 struct memcg_stock_pcp *stock;
3157 struct obj_cgroup *old = NULL;
3158 unsigned long flags;
3161 local_lock_irqsave(&memcg_stock.stock_lock, flags);
3162 stock = this_cpu_ptr(&memcg_stock);
3165 * Save vmstat data in stock and skip vmstat array update unless
3166 * accumulating over a page of vmstat data or when pgdat or idx
3169 if (READ_ONCE(stock->cached_objcg) != objcg) {
3170 old = drain_obj_stock(stock);
3171 obj_cgroup_get(objcg);
3172 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3173 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3174 WRITE_ONCE(stock->cached_objcg, objcg);
3175 stock->cached_pgdat = pgdat;
3176 } else if (stock->cached_pgdat != pgdat) {
3177 /* Flush the existing cached vmstat data */
3178 struct pglist_data *oldpg = stock->cached_pgdat;
3180 if (stock->nr_slab_reclaimable_b) {
3181 mod_objcg_mlstate(objcg, oldpg, NR_SLAB_RECLAIMABLE_B,
3182 stock->nr_slab_reclaimable_b);
3183 stock->nr_slab_reclaimable_b = 0;
3185 if (stock->nr_slab_unreclaimable_b) {
3186 mod_objcg_mlstate(objcg, oldpg, NR_SLAB_UNRECLAIMABLE_B,
3187 stock->nr_slab_unreclaimable_b);
3188 stock->nr_slab_unreclaimable_b = 0;
3190 stock->cached_pgdat = pgdat;
3193 bytes = (idx == NR_SLAB_RECLAIMABLE_B) ? &stock->nr_slab_reclaimable_b
3194 : &stock->nr_slab_unreclaimable_b;
3196 * Even for large object >= PAGE_SIZE, the vmstat data will still be
3197 * cached locally at least once before pushing it out.
3204 if (abs(*bytes) > PAGE_SIZE) {
3212 mod_objcg_mlstate(objcg, pgdat, idx, nr);
3214 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3216 obj_cgroup_put(old);
3219 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3221 struct memcg_stock_pcp *stock;
3222 unsigned long flags;
3225 local_lock_irqsave(&memcg_stock.stock_lock, flags);
3227 stock = this_cpu_ptr(&memcg_stock);
3228 if (objcg == READ_ONCE(stock->cached_objcg) && stock->nr_bytes >= nr_bytes) {
3229 stock->nr_bytes -= nr_bytes;
3233 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3238 static struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock)
3240 struct obj_cgroup *old = READ_ONCE(stock->cached_objcg);
3245 if (stock->nr_bytes) {
3246 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3247 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3250 struct mem_cgroup *memcg;
3252 memcg = get_mem_cgroup_from_objcg(old);
3254 memcg_account_kmem(memcg, -nr_pages);
3255 __refill_stock(memcg, nr_pages);
3257 css_put(&memcg->css);
3261 * The leftover is flushed to the centralized per-memcg value.
3262 * On the next attempt to refill obj stock it will be moved
3263 * to a per-cpu stock (probably, on an other CPU), see
3264 * refill_obj_stock().
3266 * How often it's flushed is a trade-off between the memory
3267 * limit enforcement accuracy and potential CPU contention,
3268 * so it might be changed in the future.
3270 atomic_add(nr_bytes, &old->nr_charged_bytes);
3271 stock->nr_bytes = 0;
3275 * Flush the vmstat data in current stock
3277 if (stock->nr_slab_reclaimable_b || stock->nr_slab_unreclaimable_b) {
3278 if (stock->nr_slab_reclaimable_b) {
3279 mod_objcg_mlstate(old, stock->cached_pgdat,
3280 NR_SLAB_RECLAIMABLE_B,
3281 stock->nr_slab_reclaimable_b);
3282 stock->nr_slab_reclaimable_b = 0;
3284 if (stock->nr_slab_unreclaimable_b) {
3285 mod_objcg_mlstate(old, stock->cached_pgdat,
3286 NR_SLAB_UNRECLAIMABLE_B,
3287 stock->nr_slab_unreclaimable_b);
3288 stock->nr_slab_unreclaimable_b = 0;
3290 stock->cached_pgdat = NULL;
3293 WRITE_ONCE(stock->cached_objcg, NULL);
3295 * The `old' objects needs to be released by the caller via
3296 * obj_cgroup_put() outside of memcg_stock_pcp::stock_lock.
3301 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3302 struct mem_cgroup *root_memcg)
3304 struct obj_cgroup *objcg = READ_ONCE(stock->cached_objcg);
3305 struct mem_cgroup *memcg;
3308 memcg = obj_cgroup_memcg(objcg);
3309 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3316 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes,
3317 bool allow_uncharge)
3319 struct memcg_stock_pcp *stock;
3320 struct obj_cgroup *old = NULL;
3321 unsigned long flags;
3322 unsigned int nr_pages = 0;
3324 local_lock_irqsave(&memcg_stock.stock_lock, flags);
3326 stock = this_cpu_ptr(&memcg_stock);
3327 if (READ_ONCE(stock->cached_objcg) != objcg) { /* reset if necessary */
3328 old = drain_obj_stock(stock);
3329 obj_cgroup_get(objcg);
3330 WRITE_ONCE(stock->cached_objcg, objcg);
3331 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3332 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3333 allow_uncharge = true; /* Allow uncharge when objcg changes */
3335 stock->nr_bytes += nr_bytes;
3337 if (allow_uncharge && (stock->nr_bytes > PAGE_SIZE)) {
3338 nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3339 stock->nr_bytes &= (PAGE_SIZE - 1);
3342 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3344 obj_cgroup_put(old);
3347 obj_cgroup_uncharge_pages(objcg, nr_pages);
3350 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3352 unsigned int nr_pages, nr_bytes;
3355 if (consume_obj_stock(objcg, size))
3359 * In theory, objcg->nr_charged_bytes can have enough
3360 * pre-charged bytes to satisfy the allocation. However,
3361 * flushing objcg->nr_charged_bytes requires two atomic
3362 * operations, and objcg->nr_charged_bytes can't be big.
3363 * The shared objcg->nr_charged_bytes can also become a
3364 * performance bottleneck if all tasks of the same memcg are
3365 * trying to update it. So it's better to ignore it and try
3366 * grab some new pages. The stock's nr_bytes will be flushed to
3367 * objcg->nr_charged_bytes later on when objcg changes.
3369 * The stock's nr_bytes may contain enough pre-charged bytes
3370 * to allow one less page from being charged, but we can't rely
3371 * on the pre-charged bytes not being changed outside of
3372 * consume_obj_stock() or refill_obj_stock(). So ignore those
3373 * pre-charged bytes as well when charging pages. To avoid a
3374 * page uncharge right after a page charge, we set the
3375 * allow_uncharge flag to false when calling refill_obj_stock()
3376 * to temporarily allow the pre-charged bytes to exceed the page
3377 * size limit. The maximum reachable value of the pre-charged
3378 * bytes is (sizeof(object) + PAGE_SIZE - 2) if there is no data
3381 nr_pages = size >> PAGE_SHIFT;
3382 nr_bytes = size & (PAGE_SIZE - 1);
3387 ret = obj_cgroup_charge_pages(objcg, gfp, nr_pages);
3388 if (!ret && nr_bytes)
3389 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes, false);
3394 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3396 refill_obj_stock(objcg, size, true);
3399 #endif /* CONFIG_MEMCG_KMEM */
3402 * Because page_memcg(head) is not set on tails, set it now.
3404 void split_page_memcg(struct page *head, unsigned int nr)
3406 struct folio *folio = page_folio(head);
3407 struct mem_cgroup *memcg = folio_memcg(folio);
3410 if (mem_cgroup_disabled() || !memcg)
3413 for (i = 1; i < nr; i++)
3414 folio_page(folio, i)->memcg_data = folio->memcg_data;
3416 if (folio_memcg_kmem(folio))
3417 obj_cgroup_get_many(__folio_objcg(folio), nr - 1);
3419 css_get_many(&memcg->css, nr - 1);
3424 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3425 * @entry: swap entry to be moved
3426 * @from: mem_cgroup which the entry is moved from
3427 * @to: mem_cgroup which the entry is moved to
3429 * It succeeds only when the swap_cgroup's record for this entry is the same
3430 * as the mem_cgroup's id of @from.
3432 * Returns 0 on success, -EINVAL on failure.
3434 * The caller must have charged to @to, IOW, called page_counter_charge() about
3435 * both res and memsw, and called css_get().
3437 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3438 struct mem_cgroup *from, struct mem_cgroup *to)
3440 unsigned short old_id, new_id;
3442 old_id = mem_cgroup_id(from);
3443 new_id = mem_cgroup_id(to);
3445 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3446 mod_memcg_state(from, MEMCG_SWAP, -1);
3447 mod_memcg_state(to, MEMCG_SWAP, 1);
3453 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3454 struct mem_cgroup *from, struct mem_cgroup *to)
3460 static DEFINE_MUTEX(memcg_max_mutex);
3462 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3463 unsigned long max, bool memsw)
3465 bool enlarge = false;
3466 bool drained = false;
3468 bool limits_invariant;
3469 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3472 if (signal_pending(current)) {
3477 mutex_lock(&memcg_max_mutex);
3479 * Make sure that the new limit (memsw or memory limit) doesn't
3480 * break our basic invariant rule memory.max <= memsw.max.
3482 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3483 max <= memcg->memsw.max;
3484 if (!limits_invariant) {
3485 mutex_unlock(&memcg_max_mutex);
3489 if (max > counter->max)
3491 ret = page_counter_set_max(counter, max);
3492 mutex_unlock(&memcg_max_mutex);
3498 drain_all_stock(memcg);
3503 if (!try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL,
3504 memsw ? 0 : MEMCG_RECLAIM_MAY_SWAP)) {
3510 if (!ret && enlarge)
3511 memcg_oom_recover(memcg);
3516 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3518 unsigned long *total_scanned)
3520 unsigned long nr_reclaimed = 0;
3521 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3522 unsigned long reclaimed;
3524 struct mem_cgroup_tree_per_node *mctz;
3525 unsigned long excess;
3530 mctz = soft_limit_tree.rb_tree_per_node[pgdat->node_id];
3533 * Do not even bother to check the largest node if the root
3534 * is empty. Do it lockless to prevent lock bouncing. Races
3535 * are acceptable as soft limit is best effort anyway.
3537 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3541 * This loop can run a while, specially if mem_cgroup's continuously
3542 * keep exceeding their soft limit and putting the system under
3549 mz = mem_cgroup_largest_soft_limit_node(mctz);
3553 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3554 gfp_mask, total_scanned);
3555 nr_reclaimed += reclaimed;
3556 spin_lock_irq(&mctz->lock);
3559 * If we failed to reclaim anything from this memory cgroup
3560 * it is time to move on to the next cgroup
3564 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3566 excess = soft_limit_excess(mz->memcg);
3568 * One school of thought says that we should not add
3569 * back the node to the tree if reclaim returns 0.
3570 * But our reclaim could return 0, simply because due
3571 * to priority we are exposing a smaller subset of
3572 * memory to reclaim from. Consider this as a longer
3575 /* If excess == 0, no tree ops */
3576 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3577 spin_unlock_irq(&mctz->lock);
3578 css_put(&mz->memcg->css);
3581 * Could not reclaim anything and there are no more
3582 * mem cgroups to try or we seem to be looping without
3583 * reclaiming anything.
3585 if (!nr_reclaimed &&
3587 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3589 } while (!nr_reclaimed);
3591 css_put(&next_mz->memcg->css);
3592 return nr_reclaimed;
3596 * Reclaims as many pages from the given memcg as possible.
3598 * Caller is responsible for holding css reference for memcg.
3600 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3602 int nr_retries = MAX_RECLAIM_RETRIES;
3604 /* we call try-to-free pages for make this cgroup empty */
3605 lru_add_drain_all();
3607 drain_all_stock(memcg);
3609 /* try to free all pages in this cgroup */
3610 while (nr_retries && page_counter_read(&memcg->memory)) {
3611 if (signal_pending(current))
3614 if (!try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL,
3615 MEMCG_RECLAIM_MAY_SWAP))
3622 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3623 char *buf, size_t nbytes,
3626 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3628 if (mem_cgroup_is_root(memcg))
3630 return mem_cgroup_force_empty(memcg) ?: nbytes;
3633 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3639 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3640 struct cftype *cft, u64 val)
3645 pr_warn_once("Non-hierarchical mode is deprecated. "
3646 "Please report your usecase to linux-mm@kvack.org if you "
3647 "depend on this functionality.\n");
3652 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3656 if (mem_cgroup_is_root(memcg)) {
3657 mem_cgroup_flush_stats();
3658 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3659 memcg_page_state(memcg, NR_ANON_MAPPED);
3661 val += memcg_page_state(memcg, MEMCG_SWAP);
3664 val = page_counter_read(&memcg->memory);
3666 val = page_counter_read(&memcg->memsw);
3679 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3682 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3683 struct page_counter *counter;
3685 switch (MEMFILE_TYPE(cft->private)) {
3687 counter = &memcg->memory;
3690 counter = &memcg->memsw;
3693 counter = &memcg->kmem;
3696 counter = &memcg->tcpmem;
3702 switch (MEMFILE_ATTR(cft->private)) {
3704 if (counter == &memcg->memory)
3705 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3706 if (counter == &memcg->memsw)
3707 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3708 return (u64)page_counter_read(counter) * PAGE_SIZE;
3710 return (u64)counter->max * PAGE_SIZE;
3712 return (u64)counter->watermark * PAGE_SIZE;
3714 return counter->failcnt;
3715 case RES_SOFT_LIMIT:
3716 return (u64)memcg->soft_limit * PAGE_SIZE;
3722 #ifdef CONFIG_MEMCG_KMEM
3723 static int memcg_online_kmem(struct mem_cgroup *memcg)
3725 struct obj_cgroup *objcg;
3727 if (mem_cgroup_kmem_disabled())
3730 if (unlikely(mem_cgroup_is_root(memcg)))
3733 objcg = obj_cgroup_alloc();
3737 objcg->memcg = memcg;
3738 rcu_assign_pointer(memcg->objcg, objcg);
3740 static_branch_enable(&memcg_kmem_enabled_key);
3742 memcg->kmemcg_id = memcg->id.id;
3747 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3749 struct mem_cgroup *parent;
3751 if (mem_cgroup_kmem_disabled())
3754 if (unlikely(mem_cgroup_is_root(memcg)))
3757 parent = parent_mem_cgroup(memcg);
3759 parent = root_mem_cgroup;
3761 memcg_reparent_objcgs(memcg, parent);
3764 * After we have finished memcg_reparent_objcgs(), all list_lrus
3765 * corresponding to this cgroup are guaranteed to remain empty.
3766 * The ordering is imposed by list_lru_node->lock taken by
3767 * memcg_reparent_list_lrus().
3769 memcg_reparent_list_lrus(memcg, parent);
3772 static int memcg_online_kmem(struct mem_cgroup *memcg)
3776 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3779 #endif /* CONFIG_MEMCG_KMEM */
3781 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3785 mutex_lock(&memcg_max_mutex);
3787 ret = page_counter_set_max(&memcg->tcpmem, max);
3791 if (!memcg->tcpmem_active) {
3793 * The active flag needs to be written after the static_key
3794 * update. This is what guarantees that the socket activation
3795 * function is the last one to run. See mem_cgroup_sk_alloc()
3796 * for details, and note that we don't mark any socket as
3797 * belonging to this memcg until that flag is up.
3799 * We need to do this, because static_keys will span multiple
3800 * sites, but we can't control their order. If we mark a socket
3801 * as accounted, but the accounting functions are not patched in
3802 * yet, we'll lose accounting.
3804 * We never race with the readers in mem_cgroup_sk_alloc(),
3805 * because when this value change, the code to process it is not
3808 static_branch_inc(&memcg_sockets_enabled_key);
3809 memcg->tcpmem_active = true;
3812 mutex_unlock(&memcg_max_mutex);
3817 * The user of this function is...
3820 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3821 char *buf, size_t nbytes, loff_t off)
3823 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3824 unsigned long nr_pages;
3827 buf = strstrip(buf);
3828 ret = page_counter_memparse(buf, "-1", &nr_pages);
3832 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3834 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3838 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3840 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3843 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3846 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3847 "Writing any value to this file has no effect. "
3848 "Please report your usecase to linux-mm@kvack.org if you "
3849 "depend on this functionality.\n");
3853 ret = memcg_update_tcp_max(memcg, nr_pages);
3857 case RES_SOFT_LIMIT:
3858 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
3861 memcg->soft_limit = nr_pages;
3866 return ret ?: nbytes;
3869 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3870 size_t nbytes, loff_t off)
3872 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3873 struct page_counter *counter;
3875 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3877 counter = &memcg->memory;
3880 counter = &memcg->memsw;
3883 counter = &memcg->kmem;
3886 counter = &memcg->tcpmem;
3892 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3894 page_counter_reset_watermark(counter);
3897 counter->failcnt = 0;
3906 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3909 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3913 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3914 struct cftype *cft, u64 val)
3916 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3918 pr_warn_once("Cgroup memory moving (move_charge_at_immigrate) is deprecated. "
3919 "Please report your usecase to linux-mm@kvack.org if you "
3920 "depend on this functionality.\n");
3922 if (val & ~MOVE_MASK)
3926 * No kind of locking is needed in here, because ->can_attach() will
3927 * check this value once in the beginning of the process, and then carry
3928 * on with stale data. This means that changes to this value will only
3929 * affect task migrations starting after the change.
3931 memcg->move_charge_at_immigrate = val;
3935 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3936 struct cftype *cft, u64 val)
3944 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3945 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3946 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3948 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3949 int nid, unsigned int lru_mask, bool tree)
3951 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3952 unsigned long nr = 0;
3955 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3958 if (!(BIT(lru) & lru_mask))
3961 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3963 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3968 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3969 unsigned int lru_mask,
3972 unsigned long nr = 0;
3976 if (!(BIT(lru) & lru_mask))
3979 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3981 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3986 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3990 unsigned int lru_mask;
3993 static const struct numa_stat stats[] = {
3994 { "total", LRU_ALL },
3995 { "file", LRU_ALL_FILE },
3996 { "anon", LRU_ALL_ANON },
3997 { "unevictable", BIT(LRU_UNEVICTABLE) },
3999 const struct numa_stat *stat;
4001 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4003 mem_cgroup_flush_stats();
4005 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4006 seq_printf(m, "%s=%lu", stat->name,
4007 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4009 for_each_node_state(nid, N_MEMORY)
4010 seq_printf(m, " N%d=%lu", nid,
4011 mem_cgroup_node_nr_lru_pages(memcg, nid,
4012 stat->lru_mask, false));
4016 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4018 seq_printf(m, "hierarchical_%s=%lu", stat->name,
4019 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4021 for_each_node_state(nid, N_MEMORY)
4022 seq_printf(m, " N%d=%lu", nid,
4023 mem_cgroup_node_nr_lru_pages(memcg, nid,
4024 stat->lru_mask, true));
4030 #endif /* CONFIG_NUMA */
4032 static const unsigned int memcg1_stats[] = {
4035 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4042 WORKINGSET_REFAULT_ANON,
4043 WORKINGSET_REFAULT_FILE,
4047 static const char *const memcg1_stat_names[] = {
4050 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4057 "workingset_refault_anon",
4058 "workingset_refault_file",
4062 /* Universal VM events cgroup1 shows, original sort order */
4063 static const unsigned int memcg1_events[] = {
4070 static int memcg_stat_show(struct seq_file *m, void *v)
4072 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4073 unsigned long memory, memsw;
4074 struct mem_cgroup *mi;
4077 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4079 mem_cgroup_flush_stats();
4081 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4084 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4086 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4087 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
4088 nr * memcg_page_state_unit(memcg1_stats[i]));
4091 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4092 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4093 memcg_events_local(memcg, memcg1_events[i]));
4095 for (i = 0; i < NR_LRU_LISTS; i++)
4096 seq_printf(m, "%s %lu\n", lru_list_name(i),
4097 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4100 /* Hierarchical information */
4101 memory = memsw = PAGE_COUNTER_MAX;
4102 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4103 memory = min(memory, READ_ONCE(mi->memory.max));
4104 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4106 seq_printf(m, "hierarchical_memory_limit %llu\n",
4107 (u64)memory * PAGE_SIZE);
4108 if (do_memsw_account())
4109 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4110 (u64)memsw * PAGE_SIZE);
4112 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4115 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4117 nr = memcg_page_state(memcg, memcg1_stats[i]);
4118 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4119 (u64)nr * memcg_page_state_unit(memcg1_stats[i]));
4122 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4123 seq_printf(m, "total_%s %llu\n",
4124 vm_event_name(memcg1_events[i]),
4125 (u64)memcg_events(memcg, memcg1_events[i]));
4127 for (i = 0; i < NR_LRU_LISTS; i++)
4128 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4129 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4132 #ifdef CONFIG_DEBUG_VM
4135 struct mem_cgroup_per_node *mz;
4136 unsigned long anon_cost = 0;
4137 unsigned long file_cost = 0;
4139 for_each_online_pgdat(pgdat) {
4140 mz = memcg->nodeinfo[pgdat->node_id];
4142 anon_cost += mz->lruvec.anon_cost;
4143 file_cost += mz->lruvec.file_cost;
4145 seq_printf(m, "anon_cost %lu\n", anon_cost);
4146 seq_printf(m, "file_cost %lu\n", file_cost);
4153 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4156 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4158 return mem_cgroup_swappiness(memcg);
4161 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4162 struct cftype *cft, u64 val)
4164 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4169 if (!mem_cgroup_is_root(memcg))
4170 memcg->swappiness = val;
4172 vm_swappiness = val;
4177 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4179 struct mem_cgroup_threshold_ary *t;
4180 unsigned long usage;
4185 t = rcu_dereference(memcg->thresholds.primary);
4187 t = rcu_dereference(memcg->memsw_thresholds.primary);
4192 usage = mem_cgroup_usage(memcg, swap);
4195 * current_threshold points to threshold just below or equal to usage.
4196 * If it's not true, a threshold was crossed after last
4197 * call of __mem_cgroup_threshold().
4199 i = t->current_threshold;
4202 * Iterate backward over array of thresholds starting from
4203 * current_threshold and check if a threshold is crossed.
4204 * If none of thresholds below usage is crossed, we read
4205 * only one element of the array here.
4207 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4208 eventfd_signal(t->entries[i].eventfd, 1);
4210 /* i = current_threshold + 1 */
4214 * Iterate forward over array of thresholds starting from
4215 * current_threshold+1 and check if a threshold is crossed.
4216 * If none of thresholds above usage is crossed, we read
4217 * only one element of the array here.
4219 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4220 eventfd_signal(t->entries[i].eventfd, 1);
4222 /* Update current_threshold */
4223 t->current_threshold = i - 1;
4228 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4231 __mem_cgroup_threshold(memcg, false);
4232 if (do_memsw_account())
4233 __mem_cgroup_threshold(memcg, true);
4235 memcg = parent_mem_cgroup(memcg);
4239 static int compare_thresholds(const void *a, const void *b)
4241 const struct mem_cgroup_threshold *_a = a;
4242 const struct mem_cgroup_threshold *_b = b;
4244 if (_a->threshold > _b->threshold)
4247 if (_a->threshold < _b->threshold)
4253 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4255 struct mem_cgroup_eventfd_list *ev;
4257 spin_lock(&memcg_oom_lock);
4259 list_for_each_entry(ev, &memcg->oom_notify, list)
4260 eventfd_signal(ev->eventfd, 1);
4262 spin_unlock(&memcg_oom_lock);
4266 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4268 struct mem_cgroup *iter;
4270 for_each_mem_cgroup_tree(iter, memcg)
4271 mem_cgroup_oom_notify_cb(iter);
4274 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4275 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4277 struct mem_cgroup_thresholds *thresholds;
4278 struct mem_cgroup_threshold_ary *new;
4279 unsigned long threshold;
4280 unsigned long usage;
4283 ret = page_counter_memparse(args, "-1", &threshold);
4287 mutex_lock(&memcg->thresholds_lock);
4290 thresholds = &memcg->thresholds;
4291 usage = mem_cgroup_usage(memcg, false);
4292 } else if (type == _MEMSWAP) {
4293 thresholds = &memcg->memsw_thresholds;
4294 usage = mem_cgroup_usage(memcg, true);
4298 /* Check if a threshold crossed before adding a new one */
4299 if (thresholds->primary)
4300 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4302 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4304 /* Allocate memory for new array of thresholds */
4305 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4312 /* Copy thresholds (if any) to new array */
4313 if (thresholds->primary)
4314 memcpy(new->entries, thresholds->primary->entries,
4315 flex_array_size(new, entries, size - 1));
4317 /* Add new threshold */
4318 new->entries[size - 1].eventfd = eventfd;
4319 new->entries[size - 1].threshold = threshold;
4321 /* Sort thresholds. Registering of new threshold isn't time-critical */
4322 sort(new->entries, size, sizeof(*new->entries),
4323 compare_thresholds, NULL);
4325 /* Find current threshold */
4326 new->current_threshold = -1;
4327 for (i = 0; i < size; i++) {
4328 if (new->entries[i].threshold <= usage) {
4330 * new->current_threshold will not be used until
4331 * rcu_assign_pointer(), so it's safe to increment
4334 ++new->current_threshold;
4339 /* Free old spare buffer and save old primary buffer as spare */
4340 kfree(thresholds->spare);
4341 thresholds->spare = thresholds->primary;
4343 rcu_assign_pointer(thresholds->primary, new);
4345 /* To be sure that nobody uses thresholds */
4349 mutex_unlock(&memcg->thresholds_lock);
4354 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4355 struct eventfd_ctx *eventfd, const char *args)
4357 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4360 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4361 struct eventfd_ctx *eventfd, const char *args)
4363 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4366 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4367 struct eventfd_ctx *eventfd, enum res_type type)
4369 struct mem_cgroup_thresholds *thresholds;
4370 struct mem_cgroup_threshold_ary *new;
4371 unsigned long usage;
4372 int i, j, size, entries;
4374 mutex_lock(&memcg->thresholds_lock);
4377 thresholds = &memcg->thresholds;
4378 usage = mem_cgroup_usage(memcg, false);
4379 } else if (type == _MEMSWAP) {
4380 thresholds = &memcg->memsw_thresholds;
4381 usage = mem_cgroup_usage(memcg, true);
4385 if (!thresholds->primary)
4388 /* Check if a threshold crossed before removing */
4389 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4391 /* Calculate new number of threshold */
4393 for (i = 0; i < thresholds->primary->size; i++) {
4394 if (thresholds->primary->entries[i].eventfd != eventfd)
4400 new = thresholds->spare;
4402 /* If no items related to eventfd have been cleared, nothing to do */
4406 /* Set thresholds array to NULL if we don't have thresholds */
4415 /* Copy thresholds and find current threshold */
4416 new->current_threshold = -1;
4417 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4418 if (thresholds->primary->entries[i].eventfd == eventfd)
4421 new->entries[j] = thresholds->primary->entries[i];
4422 if (new->entries[j].threshold <= usage) {
4424 * new->current_threshold will not be used
4425 * until rcu_assign_pointer(), so it's safe to increment
4428 ++new->current_threshold;
4434 /* Swap primary and spare array */
4435 thresholds->spare = thresholds->primary;
4437 rcu_assign_pointer(thresholds->primary, new);
4439 /* To be sure that nobody uses thresholds */
4442 /* If all events are unregistered, free the spare array */
4444 kfree(thresholds->spare);
4445 thresholds->spare = NULL;
4448 mutex_unlock(&memcg->thresholds_lock);
4451 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4452 struct eventfd_ctx *eventfd)
4454 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4457 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4458 struct eventfd_ctx *eventfd)
4460 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4463 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4464 struct eventfd_ctx *eventfd, const char *args)
4466 struct mem_cgroup_eventfd_list *event;
4468 event = kmalloc(sizeof(*event), GFP_KERNEL);
4472 spin_lock(&memcg_oom_lock);
4474 event->eventfd = eventfd;
4475 list_add(&event->list, &memcg->oom_notify);
4477 /* already in OOM ? */
4478 if (memcg->under_oom)
4479 eventfd_signal(eventfd, 1);
4480 spin_unlock(&memcg_oom_lock);
4485 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4486 struct eventfd_ctx *eventfd)
4488 struct mem_cgroup_eventfd_list *ev, *tmp;
4490 spin_lock(&memcg_oom_lock);
4492 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4493 if (ev->eventfd == eventfd) {
4494 list_del(&ev->list);
4499 spin_unlock(&memcg_oom_lock);
4502 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4504 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4506 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4507 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4508 seq_printf(sf, "oom_kill %lu\n",
4509 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4513 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4514 struct cftype *cft, u64 val)
4516 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4518 /* cannot set to root cgroup and only 0 and 1 are allowed */
4519 if (mem_cgroup_is_root(memcg) || !((val == 0) || (val == 1)))
4522 memcg->oom_kill_disable = val;
4524 memcg_oom_recover(memcg);
4529 #ifdef CONFIG_CGROUP_WRITEBACK
4531 #include <trace/events/writeback.h>
4533 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4535 return wb_domain_init(&memcg->cgwb_domain, gfp);
4538 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4540 wb_domain_exit(&memcg->cgwb_domain);
4543 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4545 wb_domain_size_changed(&memcg->cgwb_domain);
4548 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4550 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4552 if (!memcg->css.parent)
4555 return &memcg->cgwb_domain;
4559 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4560 * @wb: bdi_writeback in question
4561 * @pfilepages: out parameter for number of file pages
4562 * @pheadroom: out parameter for number of allocatable pages according to memcg
4563 * @pdirty: out parameter for number of dirty pages
4564 * @pwriteback: out parameter for number of pages under writeback
4566 * Determine the numbers of file, headroom, dirty, and writeback pages in
4567 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4568 * is a bit more involved.
4570 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4571 * headroom is calculated as the lowest headroom of itself and the
4572 * ancestors. Note that this doesn't consider the actual amount of
4573 * available memory in the system. The caller should further cap
4574 * *@pheadroom accordingly.
4576 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4577 unsigned long *pheadroom, unsigned long *pdirty,
4578 unsigned long *pwriteback)
4580 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4581 struct mem_cgroup *parent;
4583 mem_cgroup_flush_stats();
4585 *pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
4586 *pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
4587 *pfilepages = memcg_page_state(memcg, NR_INACTIVE_FILE) +
4588 memcg_page_state(memcg, NR_ACTIVE_FILE);
4590 *pheadroom = PAGE_COUNTER_MAX;
4591 while ((parent = parent_mem_cgroup(memcg))) {
4592 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4593 READ_ONCE(memcg->memory.high));
4594 unsigned long used = page_counter_read(&memcg->memory);
4596 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4602 * Foreign dirty flushing
4604 * There's an inherent mismatch between memcg and writeback. The former
4605 * tracks ownership per-page while the latter per-inode. This was a
4606 * deliberate design decision because honoring per-page ownership in the
4607 * writeback path is complicated, may lead to higher CPU and IO overheads
4608 * and deemed unnecessary given that write-sharing an inode across
4609 * different cgroups isn't a common use-case.
4611 * Combined with inode majority-writer ownership switching, this works well
4612 * enough in most cases but there are some pathological cases. For
4613 * example, let's say there are two cgroups A and B which keep writing to
4614 * different but confined parts of the same inode. B owns the inode and
4615 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4616 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4617 * triggering background writeback. A will be slowed down without a way to
4618 * make writeback of the dirty pages happen.
4620 * Conditions like the above can lead to a cgroup getting repeatedly and
4621 * severely throttled after making some progress after each
4622 * dirty_expire_interval while the underlying IO device is almost
4625 * Solving this problem completely requires matching the ownership tracking
4626 * granularities between memcg and writeback in either direction. However,
4627 * the more egregious behaviors can be avoided by simply remembering the
4628 * most recent foreign dirtying events and initiating remote flushes on
4629 * them when local writeback isn't enough to keep the memory clean enough.
4631 * The following two functions implement such mechanism. When a foreign
4632 * page - a page whose memcg and writeback ownerships don't match - is
4633 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4634 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4635 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4636 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4637 * foreign bdi_writebacks which haven't expired. Both the numbers of
4638 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4639 * limited to MEMCG_CGWB_FRN_CNT.
4641 * The mechanism only remembers IDs and doesn't hold any object references.
4642 * As being wrong occasionally doesn't matter, updates and accesses to the
4643 * records are lockless and racy.
4645 void mem_cgroup_track_foreign_dirty_slowpath(struct folio *folio,
4646 struct bdi_writeback *wb)
4648 struct mem_cgroup *memcg = folio_memcg(folio);
4649 struct memcg_cgwb_frn *frn;
4650 u64 now = get_jiffies_64();
4651 u64 oldest_at = now;
4655 trace_track_foreign_dirty(folio, wb);
4658 * Pick the slot to use. If there is already a slot for @wb, keep
4659 * using it. If not replace the oldest one which isn't being
4662 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4663 frn = &memcg->cgwb_frn[i];
4664 if (frn->bdi_id == wb->bdi->id &&
4665 frn->memcg_id == wb->memcg_css->id)
4667 if (time_before64(frn->at, oldest_at) &&
4668 atomic_read(&frn->done.cnt) == 1) {
4670 oldest_at = frn->at;
4674 if (i < MEMCG_CGWB_FRN_CNT) {
4676 * Re-using an existing one. Update timestamp lazily to
4677 * avoid making the cacheline hot. We want them to be
4678 * reasonably up-to-date and significantly shorter than
4679 * dirty_expire_interval as that's what expires the record.
4680 * Use the shorter of 1s and dirty_expire_interval / 8.
4682 unsigned long update_intv =
4683 min_t(unsigned long, HZ,
4684 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4686 if (time_before64(frn->at, now - update_intv))
4688 } else if (oldest >= 0) {
4689 /* replace the oldest free one */
4690 frn = &memcg->cgwb_frn[oldest];
4691 frn->bdi_id = wb->bdi->id;
4692 frn->memcg_id = wb->memcg_css->id;
4697 /* issue foreign writeback flushes for recorded foreign dirtying events */
4698 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4700 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4701 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4702 u64 now = jiffies_64;
4705 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4706 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4709 * If the record is older than dirty_expire_interval,
4710 * writeback on it has already started. No need to kick it
4711 * off again. Also, don't start a new one if there's
4712 * already one in flight.
4714 if (time_after64(frn->at, now - intv) &&
4715 atomic_read(&frn->done.cnt) == 1) {
4717 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4718 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id,
4719 WB_REASON_FOREIGN_FLUSH,
4725 #else /* CONFIG_CGROUP_WRITEBACK */
4727 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4732 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4736 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4740 #endif /* CONFIG_CGROUP_WRITEBACK */
4743 * DO NOT USE IN NEW FILES.
4745 * "cgroup.event_control" implementation.
4747 * This is way over-engineered. It tries to support fully configurable
4748 * events for each user. Such level of flexibility is completely
4749 * unnecessary especially in the light of the planned unified hierarchy.
4751 * Please deprecate this and replace with something simpler if at all
4756 * Unregister event and free resources.
4758 * Gets called from workqueue.
4760 static void memcg_event_remove(struct work_struct *work)
4762 struct mem_cgroup_event *event =
4763 container_of(work, struct mem_cgroup_event, remove);
4764 struct mem_cgroup *memcg = event->memcg;
4766 remove_wait_queue(event->wqh, &event->wait);
4768 event->unregister_event(memcg, event->eventfd);
4770 /* Notify userspace the event is going away. */
4771 eventfd_signal(event->eventfd, 1);
4773 eventfd_ctx_put(event->eventfd);
4775 css_put(&memcg->css);
4779 * Gets called on EPOLLHUP on eventfd when user closes it.
4781 * Called with wqh->lock held and interrupts disabled.
4783 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4784 int sync, void *key)
4786 struct mem_cgroup_event *event =
4787 container_of(wait, struct mem_cgroup_event, wait);
4788 struct mem_cgroup *memcg = event->memcg;
4789 __poll_t flags = key_to_poll(key);
4791 if (flags & EPOLLHUP) {
4793 * If the event has been detached at cgroup removal, we
4794 * can simply return knowing the other side will cleanup
4797 * We can't race against event freeing since the other
4798 * side will require wqh->lock via remove_wait_queue(),
4801 spin_lock(&memcg->event_list_lock);
4802 if (!list_empty(&event->list)) {
4803 list_del_init(&event->list);
4805 * We are in atomic context, but cgroup_event_remove()
4806 * may sleep, so we have to call it in workqueue.
4808 schedule_work(&event->remove);
4810 spin_unlock(&memcg->event_list_lock);
4816 static void memcg_event_ptable_queue_proc(struct file *file,
4817 wait_queue_head_t *wqh, poll_table *pt)
4819 struct mem_cgroup_event *event =
4820 container_of(pt, struct mem_cgroup_event, pt);
4823 add_wait_queue(wqh, &event->wait);
4827 * DO NOT USE IN NEW FILES.
4829 * Parse input and register new cgroup event handler.
4831 * Input must be in format '<event_fd> <control_fd> <args>'.
4832 * Interpretation of args is defined by control file implementation.
4834 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4835 char *buf, size_t nbytes, loff_t off)
4837 struct cgroup_subsys_state *css = of_css(of);
4838 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4839 struct mem_cgroup_event *event;
4840 struct cgroup_subsys_state *cfile_css;
4841 unsigned int efd, cfd;
4844 struct dentry *cdentry;
4849 if (IS_ENABLED(CONFIG_PREEMPT_RT))
4852 buf = strstrip(buf);
4854 efd = simple_strtoul(buf, &endp, 10);
4859 cfd = simple_strtoul(buf, &endp, 10);
4860 if ((*endp != ' ') && (*endp != '\0'))
4864 event = kzalloc(sizeof(*event), GFP_KERNEL);
4868 event->memcg = memcg;
4869 INIT_LIST_HEAD(&event->list);
4870 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4871 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4872 INIT_WORK(&event->remove, memcg_event_remove);
4880 event->eventfd = eventfd_ctx_fileget(efile.file);
4881 if (IS_ERR(event->eventfd)) {
4882 ret = PTR_ERR(event->eventfd);
4889 goto out_put_eventfd;
4892 /* the process need read permission on control file */
4893 /* AV: shouldn't we check that it's been opened for read instead? */
4894 ret = file_permission(cfile.file, MAY_READ);
4899 * The control file must be a regular cgroup1 file. As a regular cgroup
4900 * file can't be renamed, it's safe to access its name afterwards.
4902 cdentry = cfile.file->f_path.dentry;
4903 if (cdentry->d_sb->s_type != &cgroup_fs_type || !d_is_reg(cdentry)) {
4909 * Determine the event callbacks and set them in @event. This used
4910 * to be done via struct cftype but cgroup core no longer knows
4911 * about these events. The following is crude but the whole thing
4912 * is for compatibility anyway.
4914 * DO NOT ADD NEW FILES.
4916 name = cdentry->d_name.name;
4918 if (!strcmp(name, "memory.usage_in_bytes")) {
4919 event->register_event = mem_cgroup_usage_register_event;
4920 event->unregister_event = mem_cgroup_usage_unregister_event;
4921 } else if (!strcmp(name, "memory.oom_control")) {
4922 event->register_event = mem_cgroup_oom_register_event;
4923 event->unregister_event = mem_cgroup_oom_unregister_event;
4924 } else if (!strcmp(name, "memory.pressure_level")) {
4925 event->register_event = vmpressure_register_event;
4926 event->unregister_event = vmpressure_unregister_event;
4927 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4928 event->register_event = memsw_cgroup_usage_register_event;
4929 event->unregister_event = memsw_cgroup_usage_unregister_event;
4936 * Verify @cfile should belong to @css. Also, remaining events are
4937 * automatically removed on cgroup destruction but the removal is
4938 * asynchronous, so take an extra ref on @css.
4940 cfile_css = css_tryget_online_from_dir(cdentry->d_parent,
4941 &memory_cgrp_subsys);
4943 if (IS_ERR(cfile_css))
4945 if (cfile_css != css) {
4950 ret = event->register_event(memcg, event->eventfd, buf);
4954 vfs_poll(efile.file, &event->pt);
4956 spin_lock_irq(&memcg->event_list_lock);
4957 list_add(&event->list, &memcg->event_list);
4958 spin_unlock_irq(&memcg->event_list_lock);
4970 eventfd_ctx_put(event->eventfd);
4979 #if defined(CONFIG_MEMCG_KMEM) && (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
4980 static int mem_cgroup_slab_show(struct seq_file *m, void *p)
4984 * Please, take a look at tools/cgroup/memcg_slabinfo.py .
4990 static struct cftype mem_cgroup_legacy_files[] = {
4992 .name = "usage_in_bytes",
4993 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4994 .read_u64 = mem_cgroup_read_u64,
4997 .name = "max_usage_in_bytes",
4998 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4999 .write = mem_cgroup_reset,
5000 .read_u64 = mem_cgroup_read_u64,
5003 .name = "limit_in_bytes",
5004 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5005 .write = mem_cgroup_write,
5006 .read_u64 = mem_cgroup_read_u64,
5009 .name = "soft_limit_in_bytes",
5010 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5011 .write = mem_cgroup_write,
5012 .read_u64 = mem_cgroup_read_u64,
5016 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5017 .write = mem_cgroup_reset,
5018 .read_u64 = mem_cgroup_read_u64,
5022 .seq_show = memcg_stat_show,
5025 .name = "force_empty",
5026 .write = mem_cgroup_force_empty_write,
5029 .name = "use_hierarchy",
5030 .write_u64 = mem_cgroup_hierarchy_write,
5031 .read_u64 = mem_cgroup_hierarchy_read,
5034 .name = "cgroup.event_control", /* XXX: for compat */
5035 .write = memcg_write_event_control,
5036 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
5039 .name = "swappiness",
5040 .read_u64 = mem_cgroup_swappiness_read,
5041 .write_u64 = mem_cgroup_swappiness_write,
5044 .name = "move_charge_at_immigrate",
5045 .read_u64 = mem_cgroup_move_charge_read,
5046 .write_u64 = mem_cgroup_move_charge_write,
5049 .name = "oom_control",
5050 .seq_show = mem_cgroup_oom_control_read,
5051 .write_u64 = mem_cgroup_oom_control_write,
5054 .name = "pressure_level",
5058 .name = "numa_stat",
5059 .seq_show = memcg_numa_stat_show,
5063 .name = "kmem.limit_in_bytes",
5064 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5065 .write = mem_cgroup_write,
5066 .read_u64 = mem_cgroup_read_u64,
5069 .name = "kmem.usage_in_bytes",
5070 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5071 .read_u64 = mem_cgroup_read_u64,
5074 .name = "kmem.failcnt",
5075 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5076 .write = mem_cgroup_reset,
5077 .read_u64 = mem_cgroup_read_u64,
5080 .name = "kmem.max_usage_in_bytes",
5081 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5082 .write = mem_cgroup_reset,
5083 .read_u64 = mem_cgroup_read_u64,
5085 #if defined(CONFIG_MEMCG_KMEM) && \
5086 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5088 .name = "kmem.slabinfo",
5089 .seq_show = mem_cgroup_slab_show,
5093 .name = "kmem.tcp.limit_in_bytes",
5094 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5095 .write = mem_cgroup_write,
5096 .read_u64 = mem_cgroup_read_u64,
5099 .name = "kmem.tcp.usage_in_bytes",
5100 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5101 .read_u64 = mem_cgroup_read_u64,
5104 .name = "kmem.tcp.failcnt",
5105 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5106 .write = mem_cgroup_reset,
5107 .read_u64 = mem_cgroup_read_u64,
5110 .name = "kmem.tcp.max_usage_in_bytes",
5111 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5112 .write = mem_cgroup_reset,
5113 .read_u64 = mem_cgroup_read_u64,
5115 { }, /* terminate */
5119 * Private memory cgroup IDR
5121 * Swap-out records and page cache shadow entries need to store memcg
5122 * references in constrained space, so we maintain an ID space that is
5123 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5124 * memory-controlled cgroups to 64k.
5126 * However, there usually are many references to the offline CSS after
5127 * the cgroup has been destroyed, such as page cache or reclaimable
5128 * slab objects, that don't need to hang on to the ID. We want to keep
5129 * those dead CSS from occupying IDs, or we might quickly exhaust the
5130 * relatively small ID space and prevent the creation of new cgroups
5131 * even when there are much fewer than 64k cgroups - possibly none.
5133 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5134 * be freed and recycled when it's no longer needed, which is usually
5135 * when the CSS is offlined.
5137 * The only exception to that are records of swapped out tmpfs/shmem
5138 * pages that need to be attributed to live ancestors on swapin. But
5139 * those references are manageable from userspace.
5142 static DEFINE_IDR(mem_cgroup_idr);
5144 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5146 if (memcg->id.id > 0) {
5147 idr_remove(&mem_cgroup_idr, memcg->id.id);
5152 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5155 refcount_add(n, &memcg->id.ref);
5158 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5160 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5161 mem_cgroup_id_remove(memcg);
5163 /* Memcg ID pins CSS */
5164 css_put(&memcg->css);
5168 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5170 mem_cgroup_id_put_many(memcg, 1);
5174 * mem_cgroup_from_id - look up a memcg from a memcg id
5175 * @id: the memcg id to look up
5177 * Caller must hold rcu_read_lock().
5179 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5181 WARN_ON_ONCE(!rcu_read_lock_held());
5182 return idr_find(&mem_cgroup_idr, id);
5185 #ifdef CONFIG_SHRINKER_DEBUG
5186 struct mem_cgroup *mem_cgroup_get_from_ino(unsigned long ino)
5188 struct cgroup *cgrp;
5189 struct cgroup_subsys_state *css;
5190 struct mem_cgroup *memcg;
5192 cgrp = cgroup_get_from_id(ino);
5194 return ERR_CAST(cgrp);
5196 css = cgroup_get_e_css(cgrp, &memory_cgrp_subsys);
5198 memcg = container_of(css, struct mem_cgroup, css);
5200 memcg = ERR_PTR(-ENOENT);
5208 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5210 struct mem_cgroup_per_node *pn;
5212 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, node);
5216 pn->lruvec_stats_percpu = alloc_percpu_gfp(struct lruvec_stats_percpu,
5217 GFP_KERNEL_ACCOUNT);
5218 if (!pn->lruvec_stats_percpu) {
5223 lruvec_init(&pn->lruvec);
5226 memcg->nodeinfo[node] = pn;
5230 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5232 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5237 free_percpu(pn->lruvec_stats_percpu);
5241 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5246 free_mem_cgroup_per_node_info(memcg, node);
5247 kfree(memcg->vmstats);
5248 free_percpu(memcg->vmstats_percpu);
5252 static void mem_cgroup_free(struct mem_cgroup *memcg)
5254 lru_gen_exit_memcg(memcg);
5255 memcg_wb_domain_exit(memcg);
5256 __mem_cgroup_free(memcg);
5259 static struct mem_cgroup *mem_cgroup_alloc(void)
5261 struct mem_cgroup *memcg;
5263 int __maybe_unused i;
5264 long error = -ENOMEM;
5266 memcg = kzalloc(struct_size(memcg, nodeinfo, nr_node_ids), GFP_KERNEL);
5268 return ERR_PTR(error);
5270 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5271 1, MEM_CGROUP_ID_MAX + 1, GFP_KERNEL);
5272 if (memcg->id.id < 0) {
5273 error = memcg->id.id;
5277 memcg->vmstats = kzalloc(sizeof(struct memcg_vmstats), GFP_KERNEL);
5278 if (!memcg->vmstats)
5281 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5282 GFP_KERNEL_ACCOUNT);
5283 if (!memcg->vmstats_percpu)
5287 if (alloc_mem_cgroup_per_node_info(memcg, node))
5290 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5293 INIT_WORK(&memcg->high_work, high_work_func);
5294 INIT_LIST_HEAD(&memcg->oom_notify);
5295 mutex_init(&memcg->thresholds_lock);
5296 spin_lock_init(&memcg->move_lock);
5297 vmpressure_init(&memcg->vmpressure);
5298 INIT_LIST_HEAD(&memcg->event_list);
5299 spin_lock_init(&memcg->event_list_lock);
5300 memcg->socket_pressure = jiffies;
5301 #ifdef CONFIG_MEMCG_KMEM
5302 memcg->kmemcg_id = -1;
5303 INIT_LIST_HEAD(&memcg->objcg_list);
5305 #ifdef CONFIG_CGROUP_WRITEBACK
5306 INIT_LIST_HEAD(&memcg->cgwb_list);
5307 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5308 memcg->cgwb_frn[i].done =
5309 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5311 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5312 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5313 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5314 memcg->deferred_split_queue.split_queue_len = 0;
5316 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5317 lru_gen_init_memcg(memcg);
5320 mem_cgroup_id_remove(memcg);
5321 __mem_cgroup_free(memcg);
5322 return ERR_PTR(error);
5325 static struct cgroup_subsys_state * __ref
5326 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5328 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5329 struct mem_cgroup *memcg, *old_memcg;
5331 old_memcg = set_active_memcg(parent);
5332 memcg = mem_cgroup_alloc();
5333 set_active_memcg(old_memcg);
5335 return ERR_CAST(memcg);
5337 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5338 memcg->soft_limit = PAGE_COUNTER_MAX;
5339 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
5340 memcg->zswap_max = PAGE_COUNTER_MAX;
5342 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5344 memcg->swappiness = mem_cgroup_swappiness(parent);
5345 memcg->oom_kill_disable = parent->oom_kill_disable;
5347 page_counter_init(&memcg->memory, &parent->memory);
5348 page_counter_init(&memcg->swap, &parent->swap);
5349 page_counter_init(&memcg->kmem, &parent->kmem);
5350 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5352 init_memcg_events();
5353 page_counter_init(&memcg->memory, NULL);
5354 page_counter_init(&memcg->swap, NULL);
5355 page_counter_init(&memcg->kmem, NULL);
5356 page_counter_init(&memcg->tcpmem, NULL);
5358 root_mem_cgroup = memcg;
5362 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5363 static_branch_inc(&memcg_sockets_enabled_key);
5368 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5370 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5372 if (memcg_online_kmem(memcg))
5376 * A memcg must be visible for expand_shrinker_info()
5377 * by the time the maps are allocated. So, we allocate maps
5378 * here, when for_each_mem_cgroup() can't skip it.
5380 if (alloc_shrinker_info(memcg))
5383 /* Online state pins memcg ID, memcg ID pins CSS */
5384 refcount_set(&memcg->id.ref, 1);
5387 if (unlikely(mem_cgroup_is_root(memcg)))
5388 queue_delayed_work(system_unbound_wq, &stats_flush_dwork,
5392 memcg_offline_kmem(memcg);
5394 mem_cgroup_id_remove(memcg);
5398 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5400 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5401 struct mem_cgroup_event *event, *tmp;
5404 * Unregister events and notify userspace.
5405 * Notify userspace about cgroup removing only after rmdir of cgroup
5406 * directory to avoid race between userspace and kernelspace.
5408 spin_lock_irq(&memcg->event_list_lock);
5409 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5410 list_del_init(&event->list);
5411 schedule_work(&event->remove);
5413 spin_unlock_irq(&memcg->event_list_lock);
5415 page_counter_set_min(&memcg->memory, 0);
5416 page_counter_set_low(&memcg->memory, 0);
5418 memcg_offline_kmem(memcg);
5419 reparent_shrinker_deferred(memcg);
5420 wb_memcg_offline(memcg);
5422 drain_all_stock(memcg);
5424 mem_cgroup_id_put(memcg);
5427 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5429 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5431 invalidate_reclaim_iterators(memcg);
5434 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5436 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5437 int __maybe_unused i;
5439 #ifdef CONFIG_CGROUP_WRITEBACK
5440 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5441 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5443 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5444 static_branch_dec(&memcg_sockets_enabled_key);
5446 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5447 static_branch_dec(&memcg_sockets_enabled_key);
5449 vmpressure_cleanup(&memcg->vmpressure);
5450 cancel_work_sync(&memcg->high_work);
5451 mem_cgroup_remove_from_trees(memcg);
5452 free_shrinker_info(memcg);
5453 mem_cgroup_free(memcg);
5457 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5458 * @css: the target css
5460 * Reset the states of the mem_cgroup associated with @css. This is
5461 * invoked when the userland requests disabling on the default hierarchy
5462 * but the memcg is pinned through dependency. The memcg should stop
5463 * applying policies and should revert to the vanilla state as it may be
5464 * made visible again.
5466 * The current implementation only resets the essential configurations.
5467 * This needs to be expanded to cover all the visible parts.
5469 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5471 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5473 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5474 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5475 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5476 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5477 page_counter_set_min(&memcg->memory, 0);
5478 page_counter_set_low(&memcg->memory, 0);
5479 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5480 memcg->soft_limit = PAGE_COUNTER_MAX;
5481 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5482 memcg_wb_domain_size_changed(memcg);
5485 static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state *css, int cpu)
5487 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5488 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5489 struct memcg_vmstats_percpu *statc;
5493 statc = per_cpu_ptr(memcg->vmstats_percpu, cpu);
5495 for (i = 0; i < MEMCG_NR_STAT; i++) {
5497 * Collect the aggregated propagation counts of groups
5498 * below us. We're in a per-cpu loop here and this is
5499 * a global counter, so the first cycle will get them.
5501 delta = memcg->vmstats->state_pending[i];
5503 memcg->vmstats->state_pending[i] = 0;
5505 /* Add CPU changes on this level since the last flush */
5506 v = READ_ONCE(statc->state[i]);
5507 if (v != statc->state_prev[i]) {
5508 delta += v - statc->state_prev[i];
5509 statc->state_prev[i] = v;
5515 /* Aggregate counts on this level and propagate upwards */
5516 memcg->vmstats->state[i] += delta;
5518 parent->vmstats->state_pending[i] += delta;
5521 for (i = 0; i < NR_MEMCG_EVENTS; i++) {
5522 delta = memcg->vmstats->events_pending[i];
5524 memcg->vmstats->events_pending[i] = 0;
5526 v = READ_ONCE(statc->events[i]);
5527 if (v != statc->events_prev[i]) {
5528 delta += v - statc->events_prev[i];
5529 statc->events_prev[i] = v;
5535 memcg->vmstats->events[i] += delta;
5537 parent->vmstats->events_pending[i] += delta;
5540 for_each_node_state(nid, N_MEMORY) {
5541 struct mem_cgroup_per_node *pn = memcg->nodeinfo[nid];
5542 struct mem_cgroup_per_node *ppn = NULL;
5543 struct lruvec_stats_percpu *lstatc;
5546 ppn = parent->nodeinfo[nid];
5548 lstatc = per_cpu_ptr(pn->lruvec_stats_percpu, cpu);
5550 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
5551 delta = pn->lruvec_stats.state_pending[i];
5553 pn->lruvec_stats.state_pending[i] = 0;
5555 v = READ_ONCE(lstatc->state[i]);
5556 if (v != lstatc->state_prev[i]) {
5557 delta += v - lstatc->state_prev[i];
5558 lstatc->state_prev[i] = v;
5564 pn->lruvec_stats.state[i] += delta;
5566 ppn->lruvec_stats.state_pending[i] += delta;
5572 /* Handlers for move charge at task migration. */
5573 static int mem_cgroup_do_precharge(unsigned long count)
5577 /* Try a single bulk charge without reclaim first, kswapd may wake */
5578 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5580 mc.precharge += count;
5584 /* Try charges one by one with reclaim, but do not retry */
5586 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5600 enum mc_target_type {
5607 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5608 unsigned long addr, pte_t ptent)
5610 struct page *page = vm_normal_page(vma, addr, ptent);
5612 if (!page || !page_mapped(page))
5614 if (PageAnon(page)) {
5615 if (!(mc.flags & MOVE_ANON))
5618 if (!(mc.flags & MOVE_FILE))
5621 if (!get_page_unless_zero(page))
5627 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5628 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5629 pte_t ptent, swp_entry_t *entry)
5631 struct page *page = NULL;
5632 swp_entry_t ent = pte_to_swp_entry(ptent);
5634 if (!(mc.flags & MOVE_ANON))
5638 * Handle device private pages that are not accessible by the CPU, but
5639 * stored as special swap entries in the page table.
5641 if (is_device_private_entry(ent)) {
5642 page = pfn_swap_entry_to_page(ent);
5643 if (!get_page_unless_zero(page))
5648 if (non_swap_entry(ent))
5652 * Because swap_cache_get_folio() updates some statistics counter,
5653 * we call find_get_page() with swapper_space directly.
5655 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5656 entry->val = ent.val;
5661 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5662 pte_t ptent, swp_entry_t *entry)
5668 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5669 unsigned long addr, pte_t ptent)
5671 if (!vma->vm_file) /* anonymous vma */
5673 if (!(mc.flags & MOVE_FILE))
5676 /* page is moved even if it's not RSS of this task(page-faulted). */
5677 /* shmem/tmpfs may report page out on swap: account for that too. */
5678 return find_get_incore_page(vma->vm_file->f_mapping,
5679 linear_page_index(vma, addr));
5683 * mem_cgroup_move_account - move account of the page
5685 * @compound: charge the page as compound or small page
5686 * @from: mem_cgroup which the page is moved from.
5687 * @to: mem_cgroup which the page is moved to. @from != @to.
5689 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5691 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5694 static int mem_cgroup_move_account(struct page *page,
5696 struct mem_cgroup *from,
5697 struct mem_cgroup *to)
5699 struct folio *folio = page_folio(page);
5700 struct lruvec *from_vec, *to_vec;
5701 struct pglist_data *pgdat;
5702 unsigned int nr_pages = compound ? folio_nr_pages(folio) : 1;
5705 VM_BUG_ON(from == to);
5706 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
5707 VM_BUG_ON(compound && !folio_test_large(folio));
5710 * Prevent mem_cgroup_migrate() from looking at
5711 * page's memory cgroup of its source page while we change it.
5714 if (!folio_trylock(folio))
5718 if (folio_memcg(folio) != from)
5721 pgdat = folio_pgdat(folio);
5722 from_vec = mem_cgroup_lruvec(from, pgdat);
5723 to_vec = mem_cgroup_lruvec(to, pgdat);
5725 folio_memcg_lock(folio);
5727 if (folio_test_anon(folio)) {
5728 if (folio_mapped(folio)) {
5729 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5730 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5731 if (folio_test_transhuge(folio)) {
5732 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5734 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5739 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5740 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5742 if (folio_test_swapbacked(folio)) {
5743 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5744 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5747 if (folio_mapped(folio)) {
5748 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5749 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5752 if (folio_test_dirty(folio)) {
5753 struct address_space *mapping = folio_mapping(folio);
5755 if (mapping_can_writeback(mapping)) {
5756 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5758 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5764 if (folio_test_writeback(folio)) {
5765 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5766 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5770 * All state has been migrated, let's switch to the new memcg.
5772 * It is safe to change page's memcg here because the page
5773 * is referenced, charged, isolated, and locked: we can't race
5774 * with (un)charging, migration, LRU putback, or anything else
5775 * that would rely on a stable page's memory cgroup.
5777 * Note that lock_page_memcg is a memcg lock, not a page lock,
5778 * to save space. As soon as we switch page's memory cgroup to a
5779 * new memcg that isn't locked, the above state can change
5780 * concurrently again. Make sure we're truly done with it.
5785 css_put(&from->css);
5787 folio->memcg_data = (unsigned long)to;
5789 __folio_memcg_unlock(from);
5792 nid = folio_nid(folio);
5794 local_irq_disable();
5795 mem_cgroup_charge_statistics(to, nr_pages);
5796 memcg_check_events(to, nid);
5797 mem_cgroup_charge_statistics(from, -nr_pages);
5798 memcg_check_events(from, nid);
5801 folio_unlock(folio);
5807 * get_mctgt_type - get target type of moving charge
5808 * @vma: the vma the pte to be checked belongs
5809 * @addr: the address corresponding to the pte to be checked
5810 * @ptent: the pte to be checked
5811 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5814 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5815 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5816 * move charge. if @target is not NULL, the page is stored in target->page
5817 * with extra refcnt got(Callers should handle it).
5818 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5819 * target for charge migration. if @target is not NULL, the entry is stored
5821 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is device memory and
5822 * thus not on the lru.
5823 * For now we such page is charge like a regular page would be as for all
5824 * intent and purposes it is just special memory taking the place of a
5827 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5829 * Called with pte lock held.
5832 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5833 unsigned long addr, pte_t ptent, union mc_target *target)
5835 struct page *page = NULL;
5836 enum mc_target_type ret = MC_TARGET_NONE;
5837 swp_entry_t ent = { .val = 0 };
5839 if (pte_present(ptent))
5840 page = mc_handle_present_pte(vma, addr, ptent);
5841 else if (pte_none_mostly(ptent))
5843 * PTE markers should be treated as a none pte here, separated
5844 * from other swap handling below.
5846 page = mc_handle_file_pte(vma, addr, ptent);
5847 else if (is_swap_pte(ptent))
5848 page = mc_handle_swap_pte(vma, ptent, &ent);
5850 if (!page && !ent.val)
5854 * Do only loose check w/o serialization.
5855 * mem_cgroup_move_account() checks the page is valid or
5856 * not under LRU exclusion.
5858 if (page_memcg(page) == mc.from) {
5859 ret = MC_TARGET_PAGE;
5860 if (is_device_private_page(page) ||
5861 is_device_coherent_page(page))
5862 ret = MC_TARGET_DEVICE;
5864 target->page = page;
5866 if (!ret || !target)
5870 * There is a swap entry and a page doesn't exist or isn't charged.
5871 * But we cannot move a tail-page in a THP.
5873 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5874 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5875 ret = MC_TARGET_SWAP;
5882 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5884 * We don't consider PMD mapped swapping or file mapped pages because THP does
5885 * not support them for now.
5886 * Caller should make sure that pmd_trans_huge(pmd) is true.
5888 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5889 unsigned long addr, pmd_t pmd, union mc_target *target)
5891 struct page *page = NULL;
5892 enum mc_target_type ret = MC_TARGET_NONE;
5894 if (unlikely(is_swap_pmd(pmd))) {
5895 VM_BUG_ON(thp_migration_supported() &&
5896 !is_pmd_migration_entry(pmd));
5899 page = pmd_page(pmd);
5900 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5901 if (!(mc.flags & MOVE_ANON))
5903 if (page_memcg(page) == mc.from) {
5904 ret = MC_TARGET_PAGE;
5907 target->page = page;
5913 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5914 unsigned long addr, pmd_t pmd, union mc_target *target)
5916 return MC_TARGET_NONE;
5920 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5921 unsigned long addr, unsigned long end,
5922 struct mm_walk *walk)
5924 struct vm_area_struct *vma = walk->vma;
5928 ptl = pmd_trans_huge_lock(pmd, vma);
5931 * Note their can not be MC_TARGET_DEVICE for now as we do not
5932 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5933 * this might change.
5935 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5936 mc.precharge += HPAGE_PMD_NR;
5941 if (pmd_trans_unstable(pmd))
5943 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5944 for (; addr != end; pte++, addr += PAGE_SIZE)
5945 if (get_mctgt_type(vma, addr, *pte, NULL))
5946 mc.precharge++; /* increment precharge temporarily */
5947 pte_unmap_unlock(pte - 1, ptl);
5953 static const struct mm_walk_ops precharge_walk_ops = {
5954 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5957 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5959 unsigned long precharge;
5962 walk_page_range(mm, 0, ULONG_MAX, &precharge_walk_ops, NULL);
5963 mmap_read_unlock(mm);
5965 precharge = mc.precharge;
5971 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5973 unsigned long precharge = mem_cgroup_count_precharge(mm);
5975 VM_BUG_ON(mc.moving_task);
5976 mc.moving_task = current;
5977 return mem_cgroup_do_precharge(precharge);
5980 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5981 static void __mem_cgroup_clear_mc(void)
5983 struct mem_cgroup *from = mc.from;
5984 struct mem_cgroup *to = mc.to;
5986 /* we must uncharge all the leftover precharges from mc.to */
5988 cancel_charge(mc.to, mc.precharge);
5992 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5993 * we must uncharge here.
5995 if (mc.moved_charge) {
5996 cancel_charge(mc.from, mc.moved_charge);
5997 mc.moved_charge = 0;
5999 /* we must fixup refcnts and charges */
6000 if (mc.moved_swap) {
6001 /* uncharge swap account from the old cgroup */
6002 if (!mem_cgroup_is_root(mc.from))
6003 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
6005 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
6008 * we charged both to->memory and to->memsw, so we
6009 * should uncharge to->memory.
6011 if (!mem_cgroup_is_root(mc.to))
6012 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
6016 memcg_oom_recover(from);
6017 memcg_oom_recover(to);
6018 wake_up_all(&mc.waitq);
6021 static void mem_cgroup_clear_mc(void)
6023 struct mm_struct *mm = mc.mm;
6026 * we must clear moving_task before waking up waiters at the end of
6029 mc.moving_task = NULL;
6030 __mem_cgroup_clear_mc();
6031 spin_lock(&mc.lock);
6035 spin_unlock(&mc.lock);
6040 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6042 struct cgroup_subsys_state *css;
6043 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
6044 struct mem_cgroup *from;
6045 struct task_struct *leader, *p;
6046 struct mm_struct *mm;
6047 unsigned long move_flags;
6050 /* charge immigration isn't supported on the default hierarchy */
6051 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6055 * Multi-process migrations only happen on the default hierarchy
6056 * where charge immigration is not used. Perform charge
6057 * immigration if @tset contains a leader and whine if there are
6061 cgroup_taskset_for_each_leader(leader, css, tset) {
6064 memcg = mem_cgroup_from_css(css);
6070 * We are now committed to this value whatever it is. Changes in this
6071 * tunable will only affect upcoming migrations, not the current one.
6072 * So we need to save it, and keep it going.
6074 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
6078 from = mem_cgroup_from_task(p);
6080 VM_BUG_ON(from == memcg);
6082 mm = get_task_mm(p);
6085 /* We move charges only when we move a owner of the mm */
6086 if (mm->owner == p) {
6089 VM_BUG_ON(mc.precharge);
6090 VM_BUG_ON(mc.moved_charge);
6091 VM_BUG_ON(mc.moved_swap);
6093 spin_lock(&mc.lock);
6097 mc.flags = move_flags;
6098 spin_unlock(&mc.lock);
6099 /* We set mc.moving_task later */
6101 ret = mem_cgroup_precharge_mc(mm);
6103 mem_cgroup_clear_mc();
6110 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6113 mem_cgroup_clear_mc();
6116 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6117 unsigned long addr, unsigned long end,
6118 struct mm_walk *walk)
6121 struct vm_area_struct *vma = walk->vma;
6124 enum mc_target_type target_type;
6125 union mc_target target;
6128 ptl = pmd_trans_huge_lock(pmd, vma);
6130 if (mc.precharge < HPAGE_PMD_NR) {
6134 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6135 if (target_type == MC_TARGET_PAGE) {
6137 if (!isolate_lru_page(page)) {
6138 if (!mem_cgroup_move_account(page, true,
6140 mc.precharge -= HPAGE_PMD_NR;
6141 mc.moved_charge += HPAGE_PMD_NR;
6143 putback_lru_page(page);
6146 } else if (target_type == MC_TARGET_DEVICE) {
6148 if (!mem_cgroup_move_account(page, true,
6150 mc.precharge -= HPAGE_PMD_NR;
6151 mc.moved_charge += HPAGE_PMD_NR;
6159 if (pmd_trans_unstable(pmd))
6162 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6163 for (; addr != end; addr += PAGE_SIZE) {
6164 pte_t ptent = *(pte++);
6165 bool device = false;
6171 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6172 case MC_TARGET_DEVICE:
6175 case MC_TARGET_PAGE:
6178 * We can have a part of the split pmd here. Moving it
6179 * can be done but it would be too convoluted so simply
6180 * ignore such a partial THP and keep it in original
6181 * memcg. There should be somebody mapping the head.
6183 if (PageTransCompound(page))
6185 if (!device && isolate_lru_page(page))
6187 if (!mem_cgroup_move_account(page, false,
6190 /* we uncharge from mc.from later. */
6194 putback_lru_page(page);
6195 put: /* get_mctgt_type() gets the page */
6198 case MC_TARGET_SWAP:
6200 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6202 mem_cgroup_id_get_many(mc.to, 1);
6203 /* we fixup other refcnts and charges later. */
6211 pte_unmap_unlock(pte - 1, ptl);
6216 * We have consumed all precharges we got in can_attach().
6217 * We try charge one by one, but don't do any additional
6218 * charges to mc.to if we have failed in charge once in attach()
6221 ret = mem_cgroup_do_precharge(1);
6229 static const struct mm_walk_ops charge_walk_ops = {
6230 .pmd_entry = mem_cgroup_move_charge_pte_range,
6233 static void mem_cgroup_move_charge(void)
6235 lru_add_drain_all();
6237 * Signal lock_page_memcg() to take the memcg's move_lock
6238 * while we're moving its pages to another memcg. Then wait
6239 * for already started RCU-only updates to finish.
6241 atomic_inc(&mc.from->moving_account);
6244 if (unlikely(!mmap_read_trylock(mc.mm))) {
6246 * Someone who are holding the mmap_lock might be waiting in
6247 * waitq. So we cancel all extra charges, wake up all waiters,
6248 * and retry. Because we cancel precharges, we might not be able
6249 * to move enough charges, but moving charge is a best-effort
6250 * feature anyway, so it wouldn't be a big problem.
6252 __mem_cgroup_clear_mc();
6257 * When we have consumed all precharges and failed in doing
6258 * additional charge, the page walk just aborts.
6260 walk_page_range(mc.mm, 0, ULONG_MAX, &charge_walk_ops, NULL);
6261 mmap_read_unlock(mc.mm);
6262 atomic_dec(&mc.from->moving_account);
6265 static void mem_cgroup_move_task(void)
6268 mem_cgroup_move_charge();
6269 mem_cgroup_clear_mc();
6272 #else /* !CONFIG_MMU */
6273 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6277 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6280 static void mem_cgroup_move_task(void)
6285 #ifdef CONFIG_LRU_GEN
6286 static void mem_cgroup_attach(struct cgroup_taskset *tset)
6288 struct task_struct *task;
6289 struct cgroup_subsys_state *css;
6291 /* find the first leader if there is any */
6292 cgroup_taskset_for_each_leader(task, css, tset)
6299 if (task->mm && READ_ONCE(task->mm->owner) == task)
6300 lru_gen_migrate_mm(task->mm);
6304 static void mem_cgroup_attach(struct cgroup_taskset *tset)
6307 #endif /* CONFIG_LRU_GEN */
6309 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6311 if (value == PAGE_COUNTER_MAX)
6312 seq_puts(m, "max\n");
6314 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6319 static u64 memory_current_read(struct cgroup_subsys_state *css,
6322 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6324 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6327 static u64 memory_peak_read(struct cgroup_subsys_state *css,
6330 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6332 return (u64)memcg->memory.watermark * PAGE_SIZE;
6335 static int memory_min_show(struct seq_file *m, void *v)
6337 return seq_puts_memcg_tunable(m,
6338 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6341 static ssize_t memory_min_write(struct kernfs_open_file *of,
6342 char *buf, size_t nbytes, loff_t off)
6344 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6348 buf = strstrip(buf);
6349 err = page_counter_memparse(buf, "max", &min);
6353 page_counter_set_min(&memcg->memory, min);
6358 static int memory_low_show(struct seq_file *m, void *v)
6360 return seq_puts_memcg_tunable(m,
6361 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6364 static ssize_t memory_low_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", &low);
6376 page_counter_set_low(&memcg->memory, low);
6381 static int memory_high_show(struct seq_file *m, void *v)
6383 return seq_puts_memcg_tunable(m,
6384 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6387 static ssize_t memory_high_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));
6391 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6392 bool drained = false;
6396 buf = strstrip(buf);
6397 err = page_counter_memparse(buf, "max", &high);
6401 page_counter_set_high(&memcg->memory, high);
6404 unsigned long nr_pages = page_counter_read(&memcg->memory);
6405 unsigned long reclaimed;
6407 if (nr_pages <= high)
6410 if (signal_pending(current))
6414 drain_all_stock(memcg);
6419 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6420 GFP_KERNEL, MEMCG_RECLAIM_MAY_SWAP);
6422 if (!reclaimed && !nr_retries--)
6426 memcg_wb_domain_size_changed(memcg);
6430 static int memory_max_show(struct seq_file *m, void *v)
6432 return seq_puts_memcg_tunable(m,
6433 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6436 static ssize_t memory_max_write(struct kernfs_open_file *of,
6437 char *buf, size_t nbytes, loff_t off)
6439 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6440 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6441 bool drained = false;
6445 buf = strstrip(buf);
6446 err = page_counter_memparse(buf, "max", &max);
6450 xchg(&memcg->memory.max, max);
6453 unsigned long nr_pages = page_counter_read(&memcg->memory);
6455 if (nr_pages <= max)
6458 if (signal_pending(current))
6462 drain_all_stock(memcg);
6468 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6469 GFP_KERNEL, MEMCG_RECLAIM_MAY_SWAP))
6474 memcg_memory_event(memcg, MEMCG_OOM);
6475 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6479 memcg_wb_domain_size_changed(memcg);
6483 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6485 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6486 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6487 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6488 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6489 seq_printf(m, "oom_kill %lu\n",
6490 atomic_long_read(&events[MEMCG_OOM_KILL]));
6491 seq_printf(m, "oom_group_kill %lu\n",
6492 atomic_long_read(&events[MEMCG_OOM_GROUP_KILL]));
6495 static int memory_events_show(struct seq_file *m, void *v)
6497 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6499 __memory_events_show(m, memcg->memory_events);
6503 static int memory_events_local_show(struct seq_file *m, void *v)
6505 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6507 __memory_events_show(m, memcg->memory_events_local);
6511 static int memory_stat_show(struct seq_file *m, void *v)
6513 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6514 char *buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
6518 memory_stat_format(memcg, buf, PAGE_SIZE);
6525 static inline unsigned long lruvec_page_state_output(struct lruvec *lruvec,
6528 return lruvec_page_state(lruvec, item) * memcg_page_state_unit(item);
6531 static int memory_numa_stat_show(struct seq_file *m, void *v)
6534 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6536 mem_cgroup_flush_stats();
6538 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6541 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6544 seq_printf(m, "%s", memory_stats[i].name);
6545 for_each_node_state(nid, N_MEMORY) {
6547 struct lruvec *lruvec;
6549 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6550 size = lruvec_page_state_output(lruvec,
6551 memory_stats[i].idx);
6552 seq_printf(m, " N%d=%llu", nid, size);
6561 static int memory_oom_group_show(struct seq_file *m, void *v)
6563 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6565 seq_printf(m, "%d\n", memcg->oom_group);
6570 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6571 char *buf, size_t nbytes, loff_t off)
6573 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6576 buf = strstrip(buf);
6580 ret = kstrtoint(buf, 0, &oom_group);
6584 if (oom_group != 0 && oom_group != 1)
6587 memcg->oom_group = oom_group;
6592 static ssize_t memory_reclaim(struct kernfs_open_file *of, char *buf,
6593 size_t nbytes, loff_t off)
6595 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6596 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6597 unsigned long nr_to_reclaim, nr_reclaimed = 0;
6598 unsigned int reclaim_options;
6601 buf = strstrip(buf);
6602 err = page_counter_memparse(buf, "", &nr_to_reclaim);
6606 reclaim_options = MEMCG_RECLAIM_MAY_SWAP | MEMCG_RECLAIM_PROACTIVE;
6607 while (nr_reclaimed < nr_to_reclaim) {
6608 unsigned long reclaimed;
6610 if (signal_pending(current))
6614 * This is the final attempt, drain percpu lru caches in the
6615 * hope of introducing more evictable pages for
6616 * try_to_free_mem_cgroup_pages().
6619 lru_add_drain_all();
6621 reclaimed = try_to_free_mem_cgroup_pages(memcg,
6622 nr_to_reclaim - nr_reclaimed,
6623 GFP_KERNEL, reclaim_options);
6625 if (!reclaimed && !nr_retries--)
6628 nr_reclaimed += reclaimed;
6634 static struct cftype memory_files[] = {
6637 .flags = CFTYPE_NOT_ON_ROOT,
6638 .read_u64 = memory_current_read,
6642 .flags = CFTYPE_NOT_ON_ROOT,
6643 .read_u64 = memory_peak_read,
6647 .flags = CFTYPE_NOT_ON_ROOT,
6648 .seq_show = memory_min_show,
6649 .write = memory_min_write,
6653 .flags = CFTYPE_NOT_ON_ROOT,
6654 .seq_show = memory_low_show,
6655 .write = memory_low_write,
6659 .flags = CFTYPE_NOT_ON_ROOT,
6660 .seq_show = memory_high_show,
6661 .write = memory_high_write,
6665 .flags = CFTYPE_NOT_ON_ROOT,
6666 .seq_show = memory_max_show,
6667 .write = memory_max_write,
6671 .flags = CFTYPE_NOT_ON_ROOT,
6672 .file_offset = offsetof(struct mem_cgroup, events_file),
6673 .seq_show = memory_events_show,
6676 .name = "events.local",
6677 .flags = CFTYPE_NOT_ON_ROOT,
6678 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6679 .seq_show = memory_events_local_show,
6683 .seq_show = memory_stat_show,
6687 .name = "numa_stat",
6688 .seq_show = memory_numa_stat_show,
6692 .name = "oom.group",
6693 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6694 .seq_show = memory_oom_group_show,
6695 .write = memory_oom_group_write,
6699 .flags = CFTYPE_NS_DELEGATABLE,
6700 .write = memory_reclaim,
6705 struct cgroup_subsys memory_cgrp_subsys = {
6706 .css_alloc = mem_cgroup_css_alloc,
6707 .css_online = mem_cgroup_css_online,
6708 .css_offline = mem_cgroup_css_offline,
6709 .css_released = mem_cgroup_css_released,
6710 .css_free = mem_cgroup_css_free,
6711 .css_reset = mem_cgroup_css_reset,
6712 .css_rstat_flush = mem_cgroup_css_rstat_flush,
6713 .can_attach = mem_cgroup_can_attach,
6714 .attach = mem_cgroup_attach,
6715 .cancel_attach = mem_cgroup_cancel_attach,
6716 .post_attach = mem_cgroup_move_task,
6717 .dfl_cftypes = memory_files,
6718 .legacy_cftypes = mem_cgroup_legacy_files,
6723 * This function calculates an individual cgroup's effective
6724 * protection which is derived from its own memory.min/low, its
6725 * parent's and siblings' settings, as well as the actual memory
6726 * distribution in the tree.
6728 * The following rules apply to the effective protection values:
6730 * 1. At the first level of reclaim, effective protection is equal to
6731 * the declared protection in memory.min and memory.low.
6733 * 2. To enable safe delegation of the protection configuration, at
6734 * subsequent levels the effective protection is capped to the
6735 * parent's effective protection.
6737 * 3. To make complex and dynamic subtrees easier to configure, the
6738 * user is allowed to overcommit the declared protection at a given
6739 * level. If that is the case, the parent's effective protection is
6740 * distributed to the children in proportion to how much protection
6741 * they have declared and how much of it they are utilizing.
6743 * This makes distribution proportional, but also work-conserving:
6744 * if one cgroup claims much more protection than it uses memory,
6745 * the unused remainder is available to its siblings.
6747 * 4. Conversely, when the declared protection is undercommitted at a
6748 * given level, the distribution of the larger parental protection
6749 * budget is NOT proportional. A cgroup's protection from a sibling
6750 * is capped to its own memory.min/low setting.
6752 * 5. However, to allow protecting recursive subtrees from each other
6753 * without having to declare each individual cgroup's fixed share
6754 * of the ancestor's claim to protection, any unutilized -
6755 * "floating" - protection from up the tree is distributed in
6756 * proportion to each cgroup's *usage*. This makes the protection
6757 * neutral wrt sibling cgroups and lets them compete freely over
6758 * the shared parental protection budget, but it protects the
6759 * subtree as a whole from neighboring subtrees.
6761 * Note that 4. and 5. are not in conflict: 4. is about protecting
6762 * against immediate siblings whereas 5. is about protecting against
6763 * neighboring subtrees.
6765 static unsigned long effective_protection(unsigned long usage,
6766 unsigned long parent_usage,
6767 unsigned long setting,
6768 unsigned long parent_effective,
6769 unsigned long siblings_protected)
6771 unsigned long protected;
6774 protected = min(usage, setting);
6776 * If all cgroups at this level combined claim and use more
6777 * protection then what the parent affords them, distribute
6778 * shares in proportion to utilization.
6780 * We are using actual utilization rather than the statically
6781 * claimed protection in order to be work-conserving: claimed
6782 * but unused protection is available to siblings that would
6783 * otherwise get a smaller chunk than what they claimed.
6785 if (siblings_protected > parent_effective)
6786 return protected * parent_effective / siblings_protected;
6789 * Ok, utilized protection of all children is within what the
6790 * parent affords them, so we know whatever this child claims
6791 * and utilizes is effectively protected.
6793 * If there is unprotected usage beyond this value, reclaim
6794 * will apply pressure in proportion to that amount.
6796 * If there is unutilized protection, the cgroup will be fully
6797 * shielded from reclaim, but we do return a smaller value for
6798 * protection than what the group could enjoy in theory. This
6799 * is okay. With the overcommit distribution above, effective
6800 * protection is always dependent on how memory is actually
6801 * consumed among the siblings anyway.
6806 * If the children aren't claiming (all of) the protection
6807 * afforded to them by the parent, distribute the remainder in
6808 * proportion to the (unprotected) memory of each cgroup. That
6809 * way, cgroups that aren't explicitly prioritized wrt each
6810 * other compete freely over the allowance, but they are
6811 * collectively protected from neighboring trees.
6813 * We're using unprotected memory for the weight so that if
6814 * some cgroups DO claim explicit protection, we don't protect
6815 * the same bytes twice.
6817 * Check both usage and parent_usage against the respective
6818 * protected values. One should imply the other, but they
6819 * aren't read atomically - make sure the division is sane.
6821 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6823 if (parent_effective > siblings_protected &&
6824 parent_usage > siblings_protected &&
6825 usage > protected) {
6826 unsigned long unclaimed;
6828 unclaimed = parent_effective - siblings_protected;
6829 unclaimed *= usage - protected;
6830 unclaimed /= parent_usage - siblings_protected;
6839 * mem_cgroup_calculate_protection - check if memory consumption is in the normal range
6840 * @root: the top ancestor of the sub-tree being checked
6841 * @memcg: the memory cgroup to check
6843 * WARNING: This function is not stateless! It can only be used as part
6844 * of a top-down tree iteration, not for isolated queries.
6846 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6847 struct mem_cgroup *memcg)
6849 unsigned long usage, parent_usage;
6850 struct mem_cgroup *parent;
6852 if (mem_cgroup_disabled())
6856 root = root_mem_cgroup;
6859 * Effective values of the reclaim targets are ignored so they
6860 * can be stale. Have a look at mem_cgroup_protection for more
6862 * TODO: calculation should be more robust so that we do not need
6863 * that special casing.
6868 usage = page_counter_read(&memcg->memory);
6872 parent = parent_mem_cgroup(memcg);
6874 if (parent == root) {
6875 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6876 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6880 parent_usage = page_counter_read(&parent->memory);
6882 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6883 READ_ONCE(memcg->memory.min),
6884 READ_ONCE(parent->memory.emin),
6885 atomic_long_read(&parent->memory.children_min_usage)));
6887 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6888 READ_ONCE(memcg->memory.low),
6889 READ_ONCE(parent->memory.elow),
6890 atomic_long_read(&parent->memory.children_low_usage)));
6893 static int charge_memcg(struct folio *folio, struct mem_cgroup *memcg,
6896 long nr_pages = folio_nr_pages(folio);
6899 ret = try_charge(memcg, gfp, nr_pages);
6903 css_get(&memcg->css);
6904 commit_charge(folio, memcg);
6906 local_irq_disable();
6907 mem_cgroup_charge_statistics(memcg, nr_pages);
6908 memcg_check_events(memcg, folio_nid(folio));
6914 int __mem_cgroup_charge(struct folio *folio, struct mm_struct *mm, gfp_t gfp)
6916 struct mem_cgroup *memcg;
6919 memcg = get_mem_cgroup_from_mm(mm);
6920 ret = charge_memcg(folio, memcg, gfp);
6921 css_put(&memcg->css);
6927 * mem_cgroup_swapin_charge_folio - Charge a newly allocated folio for swapin.
6928 * @folio: folio to charge.
6929 * @mm: mm context of the victim
6930 * @gfp: reclaim mode
6931 * @entry: swap entry for which the folio is allocated
6933 * This function charges a folio allocated for swapin. Please call this before
6934 * adding the folio to the swapcache.
6936 * Returns 0 on success. Otherwise, an error code is returned.
6938 int mem_cgroup_swapin_charge_folio(struct folio *folio, struct mm_struct *mm,
6939 gfp_t gfp, swp_entry_t entry)
6941 struct mem_cgroup *memcg;
6945 if (mem_cgroup_disabled())
6948 id = lookup_swap_cgroup_id(entry);
6950 memcg = mem_cgroup_from_id(id);
6951 if (!memcg || !css_tryget_online(&memcg->css))
6952 memcg = get_mem_cgroup_from_mm(mm);
6955 ret = charge_memcg(folio, memcg, gfp);
6957 css_put(&memcg->css);
6962 * mem_cgroup_swapin_uncharge_swap - uncharge swap slot
6963 * @entry: swap entry for which the page is charged
6965 * Call this function after successfully adding the charged page to swapcache.
6967 * Note: This function assumes the page for which swap slot is being uncharged
6970 void mem_cgroup_swapin_uncharge_swap(swp_entry_t entry)
6973 * Cgroup1's unified memory+swap counter has been charged with the
6974 * new swapcache page, finish the transfer by uncharging the swap
6975 * slot. The swap slot would also get uncharged when it dies, but
6976 * it can stick around indefinitely and we'd count the page twice
6979 * Cgroup2 has separate resource counters for memory and swap,
6980 * so this is a non-issue here. Memory and swap charge lifetimes
6981 * correspond 1:1 to page and swap slot lifetimes: we charge the
6982 * page to memory here, and uncharge swap when the slot is freed.
6984 if (!mem_cgroup_disabled() && do_memsw_account()) {
6986 * The swap entry might not get freed for a long time,
6987 * let's not wait for it. The page already received a
6988 * memory+swap charge, drop the swap entry duplicate.
6990 mem_cgroup_uncharge_swap(entry, 1);
6994 struct uncharge_gather {
6995 struct mem_cgroup *memcg;
6996 unsigned long nr_memory;
6997 unsigned long pgpgout;
6998 unsigned long nr_kmem;
7002 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
7004 memset(ug, 0, sizeof(*ug));
7007 static void uncharge_batch(const struct uncharge_gather *ug)
7009 unsigned long flags;
7011 if (ug->nr_memory) {
7012 page_counter_uncharge(&ug->memcg->memory, ug->nr_memory);
7013 if (do_memsw_account())
7014 page_counter_uncharge(&ug->memcg->memsw, ug->nr_memory);
7016 memcg_account_kmem(ug->memcg, -ug->nr_kmem);
7017 memcg_oom_recover(ug->memcg);
7020 local_irq_save(flags);
7021 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
7022 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_memory);
7023 memcg_check_events(ug->memcg, ug->nid);
7024 local_irq_restore(flags);
7026 /* drop reference from uncharge_folio */
7027 css_put(&ug->memcg->css);
7030 static void uncharge_folio(struct folio *folio, struct uncharge_gather *ug)
7033 struct mem_cgroup *memcg;
7034 struct obj_cgroup *objcg;
7036 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
7039 * Nobody should be changing or seriously looking at
7040 * folio memcg or objcg at this point, we have fully
7041 * exclusive access to the folio.
7043 if (folio_memcg_kmem(folio)) {
7044 objcg = __folio_objcg(folio);
7046 * This get matches the put at the end of the function and
7047 * kmem pages do not hold memcg references anymore.
7049 memcg = get_mem_cgroup_from_objcg(objcg);
7051 memcg = __folio_memcg(folio);
7057 if (ug->memcg != memcg) {
7060 uncharge_gather_clear(ug);
7063 ug->nid = folio_nid(folio);
7065 /* pairs with css_put in uncharge_batch */
7066 css_get(&memcg->css);
7069 nr_pages = folio_nr_pages(folio);
7071 if (folio_memcg_kmem(folio)) {
7072 ug->nr_memory += nr_pages;
7073 ug->nr_kmem += nr_pages;
7075 folio->memcg_data = 0;
7076 obj_cgroup_put(objcg);
7078 /* LRU pages aren't accounted at the root level */
7079 if (!mem_cgroup_is_root(memcg))
7080 ug->nr_memory += nr_pages;
7083 folio->memcg_data = 0;
7086 css_put(&memcg->css);
7089 void __mem_cgroup_uncharge(struct folio *folio)
7091 struct uncharge_gather ug;
7093 /* Don't touch folio->lru of any random page, pre-check: */
7094 if (!folio_memcg(folio))
7097 uncharge_gather_clear(&ug);
7098 uncharge_folio(folio, &ug);
7099 uncharge_batch(&ug);
7103 * __mem_cgroup_uncharge_list - uncharge a list of page
7104 * @page_list: list of pages to uncharge
7106 * Uncharge a list of pages previously charged with
7107 * __mem_cgroup_charge().
7109 void __mem_cgroup_uncharge_list(struct list_head *page_list)
7111 struct uncharge_gather ug;
7112 struct folio *folio;
7114 uncharge_gather_clear(&ug);
7115 list_for_each_entry(folio, page_list, lru)
7116 uncharge_folio(folio, &ug);
7118 uncharge_batch(&ug);
7122 * mem_cgroup_migrate - Charge a folio's replacement.
7123 * @old: Currently circulating folio.
7124 * @new: Replacement folio.
7126 * Charge @new as a replacement folio for @old. @old will
7127 * be uncharged upon free.
7129 * Both folios must be locked, @new->mapping must be set up.
7131 void mem_cgroup_migrate(struct folio *old, struct folio *new)
7133 struct mem_cgroup *memcg;
7134 long nr_pages = folio_nr_pages(new);
7135 unsigned long flags;
7137 VM_BUG_ON_FOLIO(!folio_test_locked(old), old);
7138 VM_BUG_ON_FOLIO(!folio_test_locked(new), new);
7139 VM_BUG_ON_FOLIO(folio_test_anon(old) != folio_test_anon(new), new);
7140 VM_BUG_ON_FOLIO(folio_nr_pages(old) != nr_pages, new);
7142 if (mem_cgroup_disabled())
7145 /* Page cache replacement: new folio already charged? */
7146 if (folio_memcg(new))
7149 memcg = folio_memcg(old);
7150 VM_WARN_ON_ONCE_FOLIO(!memcg, old);
7154 /* Force-charge the new page. The old one will be freed soon */
7155 if (!mem_cgroup_is_root(memcg)) {
7156 page_counter_charge(&memcg->memory, nr_pages);
7157 if (do_memsw_account())
7158 page_counter_charge(&memcg->memsw, nr_pages);
7161 css_get(&memcg->css);
7162 commit_charge(new, memcg);
7164 local_irq_save(flags);
7165 mem_cgroup_charge_statistics(memcg, nr_pages);
7166 memcg_check_events(memcg, folio_nid(new));
7167 local_irq_restore(flags);
7170 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
7171 EXPORT_SYMBOL(memcg_sockets_enabled_key);
7173 void mem_cgroup_sk_alloc(struct sock *sk)
7175 struct mem_cgroup *memcg;
7177 if (!mem_cgroup_sockets_enabled)
7180 /* Do not associate the sock with unrelated interrupted task's memcg. */
7185 memcg = mem_cgroup_from_task(current);
7186 if (memcg == root_mem_cgroup)
7188 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
7190 if (css_tryget(&memcg->css))
7191 sk->sk_memcg = memcg;
7196 void mem_cgroup_sk_free(struct sock *sk)
7199 css_put(&sk->sk_memcg->css);
7203 * mem_cgroup_charge_skmem - charge socket memory
7204 * @memcg: memcg to charge
7205 * @nr_pages: number of pages to charge
7206 * @gfp_mask: reclaim mode
7208 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7209 * @memcg's configured limit, %false if it doesn't.
7211 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages,
7214 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7215 struct page_counter *fail;
7217 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7218 memcg->tcpmem_pressure = 0;
7221 memcg->tcpmem_pressure = 1;
7222 if (gfp_mask & __GFP_NOFAIL) {
7223 page_counter_charge(&memcg->tcpmem, nr_pages);
7229 if (try_charge(memcg, gfp_mask, nr_pages) == 0) {
7230 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7238 * mem_cgroup_uncharge_skmem - uncharge socket memory
7239 * @memcg: memcg to uncharge
7240 * @nr_pages: number of pages to uncharge
7242 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7244 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7245 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7249 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7251 refill_stock(memcg, nr_pages);
7254 static int __init cgroup_memory(char *s)
7258 while ((token = strsep(&s, ",")) != NULL) {
7261 if (!strcmp(token, "nosocket"))
7262 cgroup_memory_nosocket = true;
7263 if (!strcmp(token, "nokmem"))
7264 cgroup_memory_nokmem = true;
7268 __setup("cgroup.memory=", cgroup_memory);
7271 * subsys_initcall() for memory controller.
7273 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7274 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7275 * basically everything that doesn't depend on a specific mem_cgroup structure
7276 * should be initialized from here.
7278 static int __init mem_cgroup_init(void)
7283 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7284 * used for per-memcg-per-cpu caching of per-node statistics. In order
7285 * to work fine, we should make sure that the overfill threshold can't
7286 * exceed S32_MAX / PAGE_SIZE.
7288 BUILD_BUG_ON(MEMCG_CHARGE_BATCH > S32_MAX / PAGE_SIZE);
7290 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7291 memcg_hotplug_cpu_dead);
7293 for_each_possible_cpu(cpu)
7294 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7297 for_each_node(node) {
7298 struct mem_cgroup_tree_per_node *rtpn;
7300 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7301 node_online(node) ? node : NUMA_NO_NODE);
7303 rtpn->rb_root = RB_ROOT;
7304 rtpn->rb_rightmost = NULL;
7305 spin_lock_init(&rtpn->lock);
7306 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7311 subsys_initcall(mem_cgroup_init);
7314 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7316 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7318 * The root cgroup cannot be destroyed, so it's refcount must
7321 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7325 memcg = parent_mem_cgroup(memcg);
7327 memcg = root_mem_cgroup;
7333 * mem_cgroup_swapout - transfer a memsw charge to swap
7334 * @folio: folio whose memsw charge to transfer
7335 * @entry: swap entry to move the charge to
7337 * Transfer the memsw charge of @folio to @entry.
7339 void mem_cgroup_swapout(struct folio *folio, swp_entry_t entry)
7341 struct mem_cgroup *memcg, *swap_memcg;
7342 unsigned int nr_entries;
7343 unsigned short oldid;
7345 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
7346 VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
7348 if (mem_cgroup_disabled())
7351 if (!do_memsw_account())
7354 memcg = folio_memcg(folio);
7356 VM_WARN_ON_ONCE_FOLIO(!memcg, folio);
7361 * In case the memcg owning these pages has been offlined and doesn't
7362 * have an ID allocated to it anymore, charge the closest online
7363 * ancestor for the swap instead and transfer the memory+swap charge.
7365 swap_memcg = mem_cgroup_id_get_online(memcg);
7366 nr_entries = folio_nr_pages(folio);
7367 /* Get references for the tail pages, too */
7369 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7370 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7372 VM_BUG_ON_FOLIO(oldid, folio);
7373 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7375 folio->memcg_data = 0;
7377 if (!mem_cgroup_is_root(memcg))
7378 page_counter_uncharge(&memcg->memory, nr_entries);
7380 if (memcg != swap_memcg) {
7381 if (!mem_cgroup_is_root(swap_memcg))
7382 page_counter_charge(&swap_memcg->memsw, nr_entries);
7383 page_counter_uncharge(&memcg->memsw, nr_entries);
7387 * Interrupts should be disabled here because the caller holds the
7388 * i_pages lock which is taken with interrupts-off. It is
7389 * important here to have the interrupts disabled because it is the
7390 * only synchronisation we have for updating the per-CPU variables.
7393 mem_cgroup_charge_statistics(memcg, -nr_entries);
7394 memcg_stats_unlock();
7395 memcg_check_events(memcg, folio_nid(folio));
7397 css_put(&memcg->css);
7401 * __mem_cgroup_try_charge_swap - try charging swap space for a folio
7402 * @folio: folio being added to swap
7403 * @entry: swap entry to charge
7405 * Try to charge @folio's memcg for the swap space at @entry.
7407 * Returns 0 on success, -ENOMEM on failure.
7409 int __mem_cgroup_try_charge_swap(struct folio *folio, swp_entry_t entry)
7411 unsigned int nr_pages = folio_nr_pages(folio);
7412 struct page_counter *counter;
7413 struct mem_cgroup *memcg;
7414 unsigned short oldid;
7416 if (do_memsw_account())
7419 memcg = folio_memcg(folio);
7421 VM_WARN_ON_ONCE_FOLIO(!memcg, folio);
7426 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7430 memcg = mem_cgroup_id_get_online(memcg);
7432 if (!mem_cgroup_is_root(memcg) &&
7433 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7434 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7435 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7436 mem_cgroup_id_put(memcg);
7440 /* Get references for the tail pages, too */
7442 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7443 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7444 VM_BUG_ON_FOLIO(oldid, folio);
7445 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7451 * __mem_cgroup_uncharge_swap - uncharge swap space
7452 * @entry: swap entry to uncharge
7453 * @nr_pages: the amount of swap space to uncharge
7455 void __mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7457 struct mem_cgroup *memcg;
7460 if (mem_cgroup_disabled())
7463 id = swap_cgroup_record(entry, 0, nr_pages);
7465 memcg = mem_cgroup_from_id(id);
7467 if (!mem_cgroup_is_root(memcg)) {
7468 if (do_memsw_account())
7469 page_counter_uncharge(&memcg->memsw, nr_pages);
7471 page_counter_uncharge(&memcg->swap, nr_pages);
7473 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7474 mem_cgroup_id_put_many(memcg, nr_pages);
7479 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7481 long nr_swap_pages = get_nr_swap_pages();
7483 if (mem_cgroup_disabled() || do_memsw_account())
7484 return nr_swap_pages;
7485 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7486 nr_swap_pages = min_t(long, nr_swap_pages,
7487 READ_ONCE(memcg->swap.max) -
7488 page_counter_read(&memcg->swap));
7489 return nr_swap_pages;
7492 bool mem_cgroup_swap_full(struct folio *folio)
7494 struct mem_cgroup *memcg;
7496 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
7500 if (do_memsw_account())
7503 memcg = folio_memcg(folio);
7507 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7508 unsigned long usage = page_counter_read(&memcg->swap);
7510 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7511 usage * 2 >= READ_ONCE(memcg->swap.max))
7518 static int __init setup_swap_account(char *s)
7520 pr_warn_once("The swapaccount= commandline option is deprecated. "
7521 "Please report your usecase to linux-mm@kvack.org if you "
7522 "depend on this functionality.\n");
7525 __setup("swapaccount=", setup_swap_account);
7527 static u64 swap_current_read(struct cgroup_subsys_state *css,
7530 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7532 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7535 static int swap_high_show(struct seq_file *m, void *v)
7537 return seq_puts_memcg_tunable(m,
7538 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7541 static ssize_t swap_high_write(struct kernfs_open_file *of,
7542 char *buf, size_t nbytes, loff_t off)
7544 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7548 buf = strstrip(buf);
7549 err = page_counter_memparse(buf, "max", &high);
7553 page_counter_set_high(&memcg->swap, high);
7558 static int swap_max_show(struct seq_file *m, void *v)
7560 return seq_puts_memcg_tunable(m,
7561 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7564 static ssize_t swap_max_write(struct kernfs_open_file *of,
7565 char *buf, size_t nbytes, loff_t off)
7567 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7571 buf = strstrip(buf);
7572 err = page_counter_memparse(buf, "max", &max);
7576 xchg(&memcg->swap.max, max);
7581 static int swap_events_show(struct seq_file *m, void *v)
7583 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7585 seq_printf(m, "high %lu\n",
7586 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7587 seq_printf(m, "max %lu\n",
7588 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7589 seq_printf(m, "fail %lu\n",
7590 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7595 static struct cftype swap_files[] = {
7597 .name = "swap.current",
7598 .flags = CFTYPE_NOT_ON_ROOT,
7599 .read_u64 = swap_current_read,
7602 .name = "swap.high",
7603 .flags = CFTYPE_NOT_ON_ROOT,
7604 .seq_show = swap_high_show,
7605 .write = swap_high_write,
7609 .flags = CFTYPE_NOT_ON_ROOT,
7610 .seq_show = swap_max_show,
7611 .write = swap_max_write,
7614 .name = "swap.events",
7615 .flags = CFTYPE_NOT_ON_ROOT,
7616 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7617 .seq_show = swap_events_show,
7622 static struct cftype memsw_files[] = {
7624 .name = "memsw.usage_in_bytes",
7625 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7626 .read_u64 = mem_cgroup_read_u64,
7629 .name = "memsw.max_usage_in_bytes",
7630 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7631 .write = mem_cgroup_reset,
7632 .read_u64 = mem_cgroup_read_u64,
7635 .name = "memsw.limit_in_bytes",
7636 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7637 .write = mem_cgroup_write,
7638 .read_u64 = mem_cgroup_read_u64,
7641 .name = "memsw.failcnt",
7642 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7643 .write = mem_cgroup_reset,
7644 .read_u64 = mem_cgroup_read_u64,
7646 { }, /* terminate */
7649 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
7651 * obj_cgroup_may_zswap - check if this cgroup can zswap
7652 * @objcg: the object cgroup
7654 * Check if the hierarchical zswap limit has been reached.
7656 * This doesn't check for specific headroom, and it is not atomic
7657 * either. But with zswap, the size of the allocation is only known
7658 * once compression has occured, and this optimistic pre-check avoids
7659 * spending cycles on compression when there is already no room left
7660 * or zswap is disabled altogether somewhere in the hierarchy.
7662 bool obj_cgroup_may_zswap(struct obj_cgroup *objcg)
7664 struct mem_cgroup *memcg, *original_memcg;
7667 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7670 original_memcg = get_mem_cgroup_from_objcg(objcg);
7671 for (memcg = original_memcg; memcg != root_mem_cgroup;
7672 memcg = parent_mem_cgroup(memcg)) {
7673 unsigned long max = READ_ONCE(memcg->zswap_max);
7674 unsigned long pages;
7676 if (max == PAGE_COUNTER_MAX)
7683 cgroup_rstat_flush(memcg->css.cgroup);
7684 pages = memcg_page_state(memcg, MEMCG_ZSWAP_B) / PAGE_SIZE;
7690 mem_cgroup_put(original_memcg);
7695 * obj_cgroup_charge_zswap - charge compression backend memory
7696 * @objcg: the object cgroup
7697 * @size: size of compressed object
7699 * This forces the charge after obj_cgroup_may_swap() allowed
7700 * compression and storage in zwap for this cgroup to go ahead.
7702 void obj_cgroup_charge_zswap(struct obj_cgroup *objcg, size_t size)
7704 struct mem_cgroup *memcg;
7706 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7709 VM_WARN_ON_ONCE(!(current->flags & PF_MEMALLOC));
7711 /* PF_MEMALLOC context, charging must succeed */
7712 if (obj_cgroup_charge(objcg, GFP_KERNEL, size))
7716 memcg = obj_cgroup_memcg(objcg);
7717 mod_memcg_state(memcg, MEMCG_ZSWAP_B, size);
7718 mod_memcg_state(memcg, MEMCG_ZSWAPPED, 1);
7723 * obj_cgroup_uncharge_zswap - uncharge compression backend memory
7724 * @objcg: the object cgroup
7725 * @size: size of compressed object
7727 * Uncharges zswap memory on page in.
7729 void obj_cgroup_uncharge_zswap(struct obj_cgroup *objcg, size_t size)
7731 struct mem_cgroup *memcg;
7733 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7736 obj_cgroup_uncharge(objcg, size);
7739 memcg = obj_cgroup_memcg(objcg);
7740 mod_memcg_state(memcg, MEMCG_ZSWAP_B, -size);
7741 mod_memcg_state(memcg, MEMCG_ZSWAPPED, -1);
7745 static u64 zswap_current_read(struct cgroup_subsys_state *css,
7748 cgroup_rstat_flush(css->cgroup);
7749 return memcg_page_state(mem_cgroup_from_css(css), MEMCG_ZSWAP_B);
7752 static int zswap_max_show(struct seq_file *m, void *v)
7754 return seq_puts_memcg_tunable(m,
7755 READ_ONCE(mem_cgroup_from_seq(m)->zswap_max));
7758 static ssize_t zswap_max_write(struct kernfs_open_file *of,
7759 char *buf, size_t nbytes, loff_t off)
7761 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7765 buf = strstrip(buf);
7766 err = page_counter_memparse(buf, "max", &max);
7770 xchg(&memcg->zswap_max, max);
7775 static struct cftype zswap_files[] = {
7777 .name = "zswap.current",
7778 .flags = CFTYPE_NOT_ON_ROOT,
7779 .read_u64 = zswap_current_read,
7782 .name = "zswap.max",
7783 .flags = CFTYPE_NOT_ON_ROOT,
7784 .seq_show = zswap_max_show,
7785 .write = zswap_max_write,
7789 #endif /* CONFIG_MEMCG_KMEM && CONFIG_ZSWAP */
7791 static int __init mem_cgroup_swap_init(void)
7793 if (mem_cgroup_disabled())
7796 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7797 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7798 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
7799 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, zswap_files));
7803 subsys_initcall(mem_cgroup_swap_init);
7805 #endif /* CONFIG_SWAP */