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 /* Whether the swap controller is active */
92 #ifdef CONFIG_MEMCG_SWAP
93 static bool cgroup_memory_noswap __ro_after_init;
95 #define cgroup_memory_noswap 1
98 #ifdef CONFIG_CGROUP_WRITEBACK
99 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
102 /* Whether legacy memory+swap accounting is active */
103 static bool do_memsw_account(void)
105 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_noswap;
108 #define THRESHOLDS_EVENTS_TARGET 128
109 #define SOFTLIMIT_EVENTS_TARGET 1024
112 * Cgroups above their limits are maintained in a RB-Tree, independent of
113 * their hierarchy representation
116 struct mem_cgroup_tree_per_node {
117 struct rb_root rb_root;
118 struct rb_node *rb_rightmost;
122 struct mem_cgroup_tree {
123 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
126 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
129 struct mem_cgroup_eventfd_list {
130 struct list_head list;
131 struct eventfd_ctx *eventfd;
135 * cgroup_event represents events which userspace want to receive.
137 struct mem_cgroup_event {
139 * memcg which the event belongs to.
141 struct mem_cgroup *memcg;
143 * eventfd to signal userspace about the event.
145 struct eventfd_ctx *eventfd;
147 * Each of these stored in a list by the cgroup.
149 struct list_head list;
151 * register_event() callback will be used to add new userspace
152 * waiter for changes related to this event. Use eventfd_signal()
153 * on eventfd to send notification to userspace.
155 int (*register_event)(struct mem_cgroup *memcg,
156 struct eventfd_ctx *eventfd, const char *args);
158 * unregister_event() callback will be called when userspace closes
159 * the eventfd or on cgroup removing. This callback must be set,
160 * if you want provide notification functionality.
162 void (*unregister_event)(struct mem_cgroup *memcg,
163 struct eventfd_ctx *eventfd);
165 * All fields below needed to unregister event when
166 * userspace closes eventfd.
169 wait_queue_head_t *wqh;
170 wait_queue_entry_t wait;
171 struct work_struct remove;
174 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
175 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
177 /* Stuffs for move charges at task migration. */
179 * Types of charges to be moved.
181 #define MOVE_ANON 0x1U
182 #define MOVE_FILE 0x2U
183 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
185 /* "mc" and its members are protected by cgroup_mutex */
186 static struct move_charge_struct {
187 spinlock_t lock; /* for from, to */
188 struct mm_struct *mm;
189 struct mem_cgroup *from;
190 struct mem_cgroup *to;
192 unsigned long precharge;
193 unsigned long moved_charge;
194 unsigned long moved_swap;
195 struct task_struct *moving_task; /* a task moving charges */
196 wait_queue_head_t waitq; /* a waitq for other context */
198 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
199 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
203 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
204 * limit reclaim to prevent infinite loops, if they ever occur.
206 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
207 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
209 /* for encoding cft->private value on file */
217 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
218 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
219 #define MEMFILE_ATTR(val) ((val) & 0xffff)
222 * Iteration constructs for visiting all cgroups (under a tree). If
223 * loops are exited prematurely (break), mem_cgroup_iter_break() must
224 * be used for reference counting.
226 #define for_each_mem_cgroup_tree(iter, root) \
227 for (iter = mem_cgroup_iter(root, NULL, NULL); \
229 iter = mem_cgroup_iter(root, iter, NULL))
231 #define for_each_mem_cgroup(iter) \
232 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
234 iter = mem_cgroup_iter(NULL, iter, NULL))
236 static inline bool task_is_dying(void)
238 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
239 (current->flags & PF_EXITING);
242 /* Some nice accessors for the vmpressure. */
243 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
246 memcg = root_mem_cgroup;
247 return &memcg->vmpressure;
250 struct mem_cgroup *vmpressure_to_memcg(struct vmpressure *vmpr)
252 return container_of(vmpr, struct mem_cgroup, vmpressure);
255 #ifdef CONFIG_MEMCG_KMEM
256 static DEFINE_SPINLOCK(objcg_lock);
258 bool mem_cgroup_kmem_disabled(void)
260 return cgroup_memory_nokmem;
263 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
264 unsigned int nr_pages);
266 static void obj_cgroup_release(struct percpu_ref *ref)
268 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
269 unsigned int nr_bytes;
270 unsigned int nr_pages;
274 * At this point all allocated objects are freed, and
275 * objcg->nr_charged_bytes can't have an arbitrary byte value.
276 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
278 * The following sequence can lead to it:
279 * 1) CPU0: objcg == stock->cached_objcg
280 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
281 * PAGE_SIZE bytes are charged
282 * 3) CPU1: a process from another memcg is allocating something,
283 * the stock if flushed,
284 * objcg->nr_charged_bytes = PAGE_SIZE - 92
285 * 5) CPU0: we do release this object,
286 * 92 bytes are added to stock->nr_bytes
287 * 6) CPU0: stock is flushed,
288 * 92 bytes are added to objcg->nr_charged_bytes
290 * In the result, nr_charged_bytes == PAGE_SIZE.
291 * This page will be uncharged in obj_cgroup_release().
293 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
294 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
295 nr_pages = nr_bytes >> PAGE_SHIFT;
298 obj_cgroup_uncharge_pages(objcg, nr_pages);
300 spin_lock_irqsave(&objcg_lock, flags);
301 list_del(&objcg->list);
302 spin_unlock_irqrestore(&objcg_lock, flags);
304 percpu_ref_exit(ref);
305 kfree_rcu(objcg, rcu);
308 static struct obj_cgroup *obj_cgroup_alloc(void)
310 struct obj_cgroup *objcg;
313 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
317 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
323 INIT_LIST_HEAD(&objcg->list);
327 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
328 struct mem_cgroup *parent)
330 struct obj_cgroup *objcg, *iter;
332 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
334 spin_lock_irq(&objcg_lock);
336 /* 1) Ready to reparent active objcg. */
337 list_add(&objcg->list, &memcg->objcg_list);
338 /* 2) Reparent active objcg and already reparented objcgs to parent. */
339 list_for_each_entry(iter, &memcg->objcg_list, list)
340 WRITE_ONCE(iter->memcg, parent);
341 /* 3) Move already reparented objcgs to the parent's list */
342 list_splice(&memcg->objcg_list, &parent->objcg_list);
344 spin_unlock_irq(&objcg_lock);
346 percpu_ref_kill(&objcg->refcnt);
350 * A lot of the calls to the cache allocation functions are expected to be
351 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
352 * conditional to this static branch, we'll have to allow modules that does
353 * kmem_cache_alloc and the such to see this symbol as well
355 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
356 EXPORT_SYMBOL(memcg_kmem_enabled_key);
360 * mem_cgroup_css_from_page - css of the memcg associated with a page
361 * @page: page of interest
363 * If memcg is bound to the default hierarchy, css of the memcg associated
364 * with @page is returned. The returned css remains associated with @page
365 * until it is released.
367 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
370 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
372 struct mem_cgroup *memcg;
374 memcg = page_memcg(page);
376 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
377 memcg = root_mem_cgroup;
383 * page_cgroup_ino - return inode number of the memcg a page is charged to
386 * Look up the closest online ancestor of the memory cgroup @page is charged to
387 * and return its inode number or 0 if @page is not charged to any cgroup. It
388 * is safe to call this function without holding a reference to @page.
390 * Note, this function is inherently racy, because there is nothing to prevent
391 * the cgroup inode from getting torn down and potentially reallocated a moment
392 * after page_cgroup_ino() returns, so it only should be used by callers that
393 * do not care (such as procfs interfaces).
395 ino_t page_cgroup_ino(struct page *page)
397 struct mem_cgroup *memcg;
398 unsigned long ino = 0;
401 memcg = page_memcg_check(page);
403 while (memcg && !(memcg->css.flags & CSS_ONLINE))
404 memcg = parent_mem_cgroup(memcg);
406 ino = cgroup_ino(memcg->css.cgroup);
411 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
412 struct mem_cgroup_tree_per_node *mctz,
413 unsigned long new_usage_in_excess)
415 struct rb_node **p = &mctz->rb_root.rb_node;
416 struct rb_node *parent = NULL;
417 struct mem_cgroup_per_node *mz_node;
418 bool rightmost = true;
423 mz->usage_in_excess = new_usage_in_excess;
424 if (!mz->usage_in_excess)
428 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
430 if (mz->usage_in_excess < mz_node->usage_in_excess) {
439 mctz->rb_rightmost = &mz->tree_node;
441 rb_link_node(&mz->tree_node, parent, p);
442 rb_insert_color(&mz->tree_node, &mctz->rb_root);
446 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
447 struct mem_cgroup_tree_per_node *mctz)
452 if (&mz->tree_node == mctz->rb_rightmost)
453 mctz->rb_rightmost = rb_prev(&mz->tree_node);
455 rb_erase(&mz->tree_node, &mctz->rb_root);
459 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
460 struct mem_cgroup_tree_per_node *mctz)
464 spin_lock_irqsave(&mctz->lock, flags);
465 __mem_cgroup_remove_exceeded(mz, mctz);
466 spin_unlock_irqrestore(&mctz->lock, flags);
469 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
471 unsigned long nr_pages = page_counter_read(&memcg->memory);
472 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
473 unsigned long excess = 0;
475 if (nr_pages > soft_limit)
476 excess = nr_pages - soft_limit;
481 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, int nid)
483 unsigned long excess;
484 struct mem_cgroup_per_node *mz;
485 struct mem_cgroup_tree_per_node *mctz;
487 mctz = soft_limit_tree.rb_tree_per_node[nid];
491 * Necessary to update all ancestors when hierarchy is used.
492 * because their event counter is not touched.
494 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
495 mz = memcg->nodeinfo[nid];
496 excess = soft_limit_excess(memcg);
498 * We have to update the tree if mz is on RB-tree or
499 * mem is over its softlimit.
501 if (excess || mz->on_tree) {
504 spin_lock_irqsave(&mctz->lock, flags);
505 /* if on-tree, remove it */
507 __mem_cgroup_remove_exceeded(mz, mctz);
509 * Insert again. mz->usage_in_excess will be updated.
510 * If excess is 0, no tree ops.
512 __mem_cgroup_insert_exceeded(mz, mctz, excess);
513 spin_unlock_irqrestore(&mctz->lock, flags);
518 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
520 struct mem_cgroup_tree_per_node *mctz;
521 struct mem_cgroup_per_node *mz;
525 mz = memcg->nodeinfo[nid];
526 mctz = soft_limit_tree.rb_tree_per_node[nid];
528 mem_cgroup_remove_exceeded(mz, mctz);
532 static struct mem_cgroup_per_node *
533 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
535 struct mem_cgroup_per_node *mz;
539 if (!mctz->rb_rightmost)
540 goto done; /* Nothing to reclaim from */
542 mz = rb_entry(mctz->rb_rightmost,
543 struct mem_cgroup_per_node, tree_node);
545 * Remove the node now but someone else can add it back,
546 * we will to add it back at the end of reclaim to its correct
547 * position in the tree.
549 __mem_cgroup_remove_exceeded(mz, mctz);
550 if (!soft_limit_excess(mz->memcg) ||
551 !css_tryget(&mz->memcg->css))
557 static struct mem_cgroup_per_node *
558 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
560 struct mem_cgroup_per_node *mz;
562 spin_lock_irq(&mctz->lock);
563 mz = __mem_cgroup_largest_soft_limit_node(mctz);
564 spin_unlock_irq(&mctz->lock);
569 * memcg and lruvec stats flushing
571 * Many codepaths leading to stats update or read are performance sensitive and
572 * adding stats flushing in such codepaths is not desirable. So, to optimize the
573 * flushing the kernel does:
575 * 1) Periodically and asynchronously flush the stats every 2 seconds to not let
576 * rstat update tree grow unbounded.
578 * 2) Flush the stats synchronously on reader side only when there are more than
579 * (MEMCG_CHARGE_BATCH * nr_cpus) update events. Though this optimization
580 * will let stats be out of sync by atmost (MEMCG_CHARGE_BATCH * nr_cpus) but
581 * only for 2 seconds due to (1).
583 static void flush_memcg_stats_dwork(struct work_struct *w);
584 static DECLARE_DEFERRABLE_WORK(stats_flush_dwork, flush_memcg_stats_dwork);
585 static DEFINE_SPINLOCK(stats_flush_lock);
586 static DEFINE_PER_CPU(unsigned int, stats_updates);
587 static atomic_t stats_flush_threshold = ATOMIC_INIT(0);
588 static u64 flush_next_time;
590 #define FLUSH_TIME (2UL*HZ)
593 * Accessors to ensure that preemption is disabled on PREEMPT_RT because it can
594 * not rely on this as part of an acquired spinlock_t lock. These functions are
595 * never used in hardirq context on PREEMPT_RT and therefore disabling preemtion
598 static void memcg_stats_lock(void)
600 #ifdef CONFIG_PREEMPT_RT
603 VM_BUG_ON(!irqs_disabled());
607 static void __memcg_stats_lock(void)
609 #ifdef CONFIG_PREEMPT_RT
614 static void memcg_stats_unlock(void)
616 #ifdef CONFIG_PREEMPT_RT
621 static inline void memcg_rstat_updated(struct mem_cgroup *memcg, int val)
625 cgroup_rstat_updated(memcg->css.cgroup, smp_processor_id());
627 x = __this_cpu_add_return(stats_updates, abs(val));
628 if (x > MEMCG_CHARGE_BATCH) {
629 atomic_add(x / MEMCG_CHARGE_BATCH, &stats_flush_threshold);
630 __this_cpu_write(stats_updates, 0);
634 static void __mem_cgroup_flush_stats(void)
638 if (!spin_trylock_irqsave(&stats_flush_lock, flag))
641 flush_next_time = jiffies_64 + 2*FLUSH_TIME;
642 cgroup_rstat_flush_irqsafe(root_mem_cgroup->css.cgroup);
643 atomic_set(&stats_flush_threshold, 0);
644 spin_unlock_irqrestore(&stats_flush_lock, flag);
647 void mem_cgroup_flush_stats(void)
649 if (atomic_read(&stats_flush_threshold) > num_online_cpus())
650 __mem_cgroup_flush_stats();
653 void mem_cgroup_flush_stats_delayed(void)
655 if (time_after64(jiffies_64, flush_next_time))
656 mem_cgroup_flush_stats();
659 static void flush_memcg_stats_dwork(struct work_struct *w)
661 __mem_cgroup_flush_stats();
662 queue_delayed_work(system_unbound_wq, &stats_flush_dwork, FLUSH_TIME);
666 * __mod_memcg_state - update cgroup memory statistics
667 * @memcg: the memory cgroup
668 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
669 * @val: delta to add to the counter, can be negative
671 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
673 if (mem_cgroup_disabled())
676 __this_cpu_add(memcg->vmstats_percpu->state[idx], val);
677 memcg_rstat_updated(memcg, val);
680 /* idx can be of type enum memcg_stat_item or node_stat_item. */
681 static unsigned long memcg_page_state_local(struct mem_cgroup *memcg, int idx)
686 for_each_possible_cpu(cpu)
687 x += per_cpu(memcg->vmstats_percpu->state[idx], cpu);
695 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
698 struct mem_cgroup_per_node *pn;
699 struct mem_cgroup *memcg;
701 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
705 * The caller from rmap relay on disabled preemption becase they never
706 * update their counter from in-interrupt context. For these two
707 * counters we check that the update is never performed from an
708 * interrupt context while other caller need to have disabled interrupt.
710 __memcg_stats_lock();
711 if (IS_ENABLED(CONFIG_DEBUG_VM) && !IS_ENABLED(CONFIG_PREEMPT_RT)) {
716 case NR_SHMEM_PMDMAPPED:
717 case NR_FILE_PMDMAPPED:
718 WARN_ON_ONCE(!in_task());
721 WARN_ON_ONCE(!irqs_disabled());
726 __this_cpu_add(memcg->vmstats_percpu->state[idx], val);
729 __this_cpu_add(pn->lruvec_stats_percpu->state[idx], val);
731 memcg_rstat_updated(memcg, val);
732 memcg_stats_unlock();
736 * __mod_lruvec_state - update lruvec memory statistics
737 * @lruvec: the lruvec
738 * @idx: the stat item
739 * @val: delta to add to the counter, can be negative
741 * The lruvec is the intersection of the NUMA node and a cgroup. This
742 * function updates the all three counters that are affected by a
743 * change of state at this level: per-node, per-cgroup, per-lruvec.
745 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
749 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
751 /* Update memcg and lruvec */
752 if (!mem_cgroup_disabled())
753 __mod_memcg_lruvec_state(lruvec, idx, val);
756 void __mod_lruvec_page_state(struct page *page, enum node_stat_item idx,
759 struct page *head = compound_head(page); /* rmap on tail pages */
760 struct mem_cgroup *memcg;
761 pg_data_t *pgdat = page_pgdat(page);
762 struct lruvec *lruvec;
765 memcg = page_memcg(head);
766 /* Untracked pages have no memcg, no lruvec. Update only the node */
769 __mod_node_page_state(pgdat, idx, val);
773 lruvec = mem_cgroup_lruvec(memcg, pgdat);
774 __mod_lruvec_state(lruvec, idx, val);
777 EXPORT_SYMBOL(__mod_lruvec_page_state);
779 void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val)
781 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
782 struct mem_cgroup *memcg;
783 struct lruvec *lruvec;
786 memcg = mem_cgroup_from_obj(p);
789 * Untracked pages have no memcg, no lruvec. Update only the
790 * node. If we reparent the slab objects to the root memcg,
791 * when we free the slab object, we need to update the per-memcg
792 * vmstats to keep it correct for the root memcg.
795 __mod_node_page_state(pgdat, idx, val);
797 lruvec = mem_cgroup_lruvec(memcg, pgdat);
798 __mod_lruvec_state(lruvec, idx, val);
804 * __count_memcg_events - account VM events in a cgroup
805 * @memcg: the memory cgroup
806 * @idx: the event item
807 * @count: the number of events that occurred
809 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
812 if (mem_cgroup_disabled())
816 __this_cpu_add(memcg->vmstats_percpu->events[idx], count);
817 memcg_rstat_updated(memcg, count);
818 memcg_stats_unlock();
821 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
823 return READ_ONCE(memcg->vmstats.events[event]);
826 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
831 for_each_possible_cpu(cpu)
832 x += per_cpu(memcg->vmstats_percpu->events[event], cpu);
836 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
839 /* pagein of a big page is an event. So, ignore page size */
841 __count_memcg_events(memcg, PGPGIN, 1);
843 __count_memcg_events(memcg, PGPGOUT, 1);
844 nr_pages = -nr_pages; /* for event */
847 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
850 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
851 enum mem_cgroup_events_target target)
853 unsigned long val, next;
855 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
856 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
857 /* from time_after() in jiffies.h */
858 if ((long)(next - val) < 0) {
860 case MEM_CGROUP_TARGET_THRESH:
861 next = val + THRESHOLDS_EVENTS_TARGET;
863 case MEM_CGROUP_TARGET_SOFTLIMIT:
864 next = val + SOFTLIMIT_EVENTS_TARGET;
869 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
876 * Check events in order.
879 static void memcg_check_events(struct mem_cgroup *memcg, int nid)
881 if (IS_ENABLED(CONFIG_PREEMPT_RT))
884 /* threshold event is triggered in finer grain than soft limit */
885 if (unlikely(mem_cgroup_event_ratelimit(memcg,
886 MEM_CGROUP_TARGET_THRESH))) {
889 do_softlimit = mem_cgroup_event_ratelimit(memcg,
890 MEM_CGROUP_TARGET_SOFTLIMIT);
891 mem_cgroup_threshold(memcg);
892 if (unlikely(do_softlimit))
893 mem_cgroup_update_tree(memcg, nid);
897 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
900 * mm_update_next_owner() may clear mm->owner to NULL
901 * if it races with swapoff, page migration, etc.
902 * So this can be called with p == NULL.
907 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
909 EXPORT_SYMBOL(mem_cgroup_from_task);
911 static __always_inline struct mem_cgroup *active_memcg(void)
914 return this_cpu_read(int_active_memcg);
916 return current->active_memcg;
920 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
921 * @mm: mm from which memcg should be extracted. It can be NULL.
923 * Obtain a reference on mm->memcg and returns it if successful. If mm
924 * is NULL, then the memcg is chosen as follows:
925 * 1) The active memcg, if set.
926 * 2) current->mm->memcg, if available
928 * If mem_cgroup is disabled, NULL is returned.
930 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
932 struct mem_cgroup *memcg;
934 if (mem_cgroup_disabled())
938 * Page cache insertions can happen without an
939 * actual mm context, e.g. during disk probing
940 * on boot, loopback IO, acct() writes etc.
942 * No need to css_get on root memcg as the reference
943 * counting is disabled on the root level in the
944 * cgroup core. See CSS_NO_REF.
947 memcg = active_memcg();
948 if (unlikely(memcg)) {
949 /* remote memcg must hold a ref */
950 css_get(&memcg->css);
955 return root_mem_cgroup;
960 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
961 if (unlikely(!memcg))
962 memcg = root_mem_cgroup;
963 } while (!css_tryget(&memcg->css));
967 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
969 static __always_inline bool memcg_kmem_bypass(void)
971 /* Allow remote memcg charging from any context. */
972 if (unlikely(active_memcg()))
975 /* Memcg to charge can't be determined. */
976 if (!in_task() || !current->mm || (current->flags & PF_KTHREAD))
983 * mem_cgroup_iter - iterate over memory cgroup hierarchy
984 * @root: hierarchy root
985 * @prev: previously returned memcg, NULL on first invocation
986 * @reclaim: cookie for shared reclaim walks, NULL for full walks
988 * Returns references to children of the hierarchy below @root, or
989 * @root itself, or %NULL after a full round-trip.
991 * Caller must pass the return value in @prev on subsequent
992 * invocations for reference counting, or use mem_cgroup_iter_break()
993 * to cancel a hierarchy walk before the round-trip is complete.
995 * Reclaimers can specify a node in @reclaim to divide up the memcgs
996 * in the hierarchy among all concurrent reclaimers operating on the
999 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1000 struct mem_cgroup *prev,
1001 struct mem_cgroup_reclaim_cookie *reclaim)
1003 struct mem_cgroup_reclaim_iter *iter;
1004 struct cgroup_subsys_state *css = NULL;
1005 struct mem_cgroup *memcg = NULL;
1006 struct mem_cgroup *pos = NULL;
1008 if (mem_cgroup_disabled())
1012 root = root_mem_cgroup;
1017 struct mem_cgroup_per_node *mz;
1019 mz = root->nodeinfo[reclaim->pgdat->node_id];
1023 * On start, join the current reclaim iteration cycle.
1024 * Exit when a concurrent walker completes it.
1027 reclaim->generation = iter->generation;
1028 else if (reclaim->generation != iter->generation)
1032 pos = READ_ONCE(iter->position);
1033 if (!pos || css_tryget(&pos->css))
1036 * css reference reached zero, so iter->position will
1037 * be cleared by ->css_released. However, we should not
1038 * rely on this happening soon, because ->css_released
1039 * is called from a work queue, and by busy-waiting we
1040 * might block it. So we clear iter->position right
1043 (void)cmpxchg(&iter->position, pos, NULL);
1053 css = css_next_descendant_pre(css, &root->css);
1056 * Reclaimers share the hierarchy walk, and a
1057 * new one might jump in right at the end of
1058 * the hierarchy - make sure they see at least
1059 * one group and restart from the beginning.
1067 * Verify the css and acquire a reference. The root
1068 * is provided by the caller, so we know it's alive
1069 * and kicking, and don't take an extra reference.
1071 if (css == &root->css || css_tryget(css)) {
1072 memcg = mem_cgroup_from_css(css);
1079 * The position could have already been updated by a competing
1080 * thread, so check that the value hasn't changed since we read
1081 * it to avoid reclaiming from the same cgroup twice.
1083 (void)cmpxchg(&iter->position, pos, memcg);
1094 if (prev && prev != root)
1095 css_put(&prev->css);
1101 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1102 * @root: hierarchy root
1103 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1105 void mem_cgroup_iter_break(struct mem_cgroup *root,
1106 struct mem_cgroup *prev)
1109 root = root_mem_cgroup;
1110 if (prev && prev != root)
1111 css_put(&prev->css);
1114 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1115 struct mem_cgroup *dead_memcg)
1117 struct mem_cgroup_reclaim_iter *iter;
1118 struct mem_cgroup_per_node *mz;
1121 for_each_node(nid) {
1122 mz = from->nodeinfo[nid];
1124 cmpxchg(&iter->position, dead_memcg, NULL);
1128 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1130 struct mem_cgroup *memcg = dead_memcg;
1131 struct mem_cgroup *last;
1134 __invalidate_reclaim_iterators(memcg, dead_memcg);
1136 } while ((memcg = parent_mem_cgroup(memcg)));
1139 * When cgruop1 non-hierarchy mode is used,
1140 * parent_mem_cgroup() does not walk all the way up to the
1141 * cgroup root (root_mem_cgroup). So we have to handle
1142 * dead_memcg from cgroup root separately.
1144 if (last != root_mem_cgroup)
1145 __invalidate_reclaim_iterators(root_mem_cgroup,
1150 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1151 * @memcg: hierarchy root
1152 * @fn: function to call for each task
1153 * @arg: argument passed to @fn
1155 * This function iterates over tasks attached to @memcg or to any of its
1156 * descendants and calls @fn for each task. If @fn returns a non-zero
1157 * value, the function breaks the iteration loop and returns the value.
1158 * Otherwise, it will iterate over all tasks and return 0.
1160 * This function must not be called for the root memory cgroup.
1162 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1163 int (*fn)(struct task_struct *, void *), void *arg)
1165 struct mem_cgroup *iter;
1168 BUG_ON(memcg == root_mem_cgroup);
1170 for_each_mem_cgroup_tree(iter, memcg) {
1171 struct css_task_iter it;
1172 struct task_struct *task;
1174 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1175 while (!ret && (task = css_task_iter_next(&it)))
1176 ret = fn(task, arg);
1177 css_task_iter_end(&it);
1179 mem_cgroup_iter_break(memcg, iter);
1186 #ifdef CONFIG_DEBUG_VM
1187 void lruvec_memcg_debug(struct lruvec *lruvec, struct folio *folio)
1189 struct mem_cgroup *memcg;
1191 if (mem_cgroup_disabled())
1194 memcg = folio_memcg(folio);
1197 VM_BUG_ON_FOLIO(lruvec_memcg(lruvec) != root_mem_cgroup, folio);
1199 VM_BUG_ON_FOLIO(lruvec_memcg(lruvec) != memcg, folio);
1204 * folio_lruvec_lock - Lock the lruvec for a folio.
1205 * @folio: Pointer to the folio.
1207 * These functions are safe to use under any of the following conditions:
1209 * - folio_test_lru false
1210 * - folio_memcg_lock()
1211 * - folio frozen (refcount of 0)
1213 * Return: The lruvec this folio is on with its lock held.
1215 struct lruvec *folio_lruvec_lock(struct folio *folio)
1217 struct lruvec *lruvec = folio_lruvec(folio);
1219 spin_lock(&lruvec->lru_lock);
1220 lruvec_memcg_debug(lruvec, folio);
1226 * folio_lruvec_lock_irq - Lock the lruvec for a folio.
1227 * @folio: Pointer to the folio.
1229 * These functions are safe to use under any of the following conditions:
1231 * - folio_test_lru false
1232 * - folio_memcg_lock()
1233 * - folio frozen (refcount of 0)
1235 * Return: The lruvec this folio is on with its lock held and interrupts
1238 struct lruvec *folio_lruvec_lock_irq(struct folio *folio)
1240 struct lruvec *lruvec = folio_lruvec(folio);
1242 spin_lock_irq(&lruvec->lru_lock);
1243 lruvec_memcg_debug(lruvec, folio);
1249 * folio_lruvec_lock_irqsave - Lock the lruvec for a folio.
1250 * @folio: Pointer to the folio.
1251 * @flags: Pointer to irqsave flags.
1253 * These functions are safe to use under any of the following conditions:
1255 * - folio_test_lru false
1256 * - folio_memcg_lock()
1257 * - folio frozen (refcount of 0)
1259 * Return: The lruvec this folio is on with its lock held and interrupts
1262 struct lruvec *folio_lruvec_lock_irqsave(struct folio *folio,
1263 unsigned long *flags)
1265 struct lruvec *lruvec = folio_lruvec(folio);
1267 spin_lock_irqsave(&lruvec->lru_lock, *flags);
1268 lruvec_memcg_debug(lruvec, folio);
1274 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1275 * @lruvec: mem_cgroup per zone lru vector
1276 * @lru: index of lru list the page is sitting on
1277 * @zid: zone id of the accounted pages
1278 * @nr_pages: positive when adding or negative when removing
1280 * This function must be called under lru_lock, just before a page is added
1281 * to or just after a page is removed from an lru list.
1283 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1284 int zid, int nr_pages)
1286 struct mem_cgroup_per_node *mz;
1287 unsigned long *lru_size;
1290 if (mem_cgroup_disabled())
1293 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1294 lru_size = &mz->lru_zone_size[zid][lru];
1297 *lru_size += nr_pages;
1300 if (WARN_ONCE(size < 0,
1301 "%s(%p, %d, %d): lru_size %ld\n",
1302 __func__, lruvec, lru, nr_pages, size)) {
1308 *lru_size += nr_pages;
1312 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1313 * @memcg: the memory cgroup
1315 * Returns the maximum amount of memory @mem can be charged with, in
1318 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1320 unsigned long margin = 0;
1321 unsigned long count;
1322 unsigned long limit;
1324 count = page_counter_read(&memcg->memory);
1325 limit = READ_ONCE(memcg->memory.max);
1327 margin = limit - count;
1329 if (do_memsw_account()) {
1330 count = page_counter_read(&memcg->memsw);
1331 limit = READ_ONCE(memcg->memsw.max);
1333 margin = min(margin, limit - count);
1342 * A routine for checking "mem" is under move_account() or not.
1344 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1345 * moving cgroups. This is for waiting at high-memory pressure
1348 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1350 struct mem_cgroup *from;
1351 struct mem_cgroup *to;
1354 * Unlike task_move routines, we access mc.to, mc.from not under
1355 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1357 spin_lock(&mc.lock);
1363 ret = mem_cgroup_is_descendant(from, memcg) ||
1364 mem_cgroup_is_descendant(to, memcg);
1366 spin_unlock(&mc.lock);
1370 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1372 if (mc.moving_task && current != mc.moving_task) {
1373 if (mem_cgroup_under_move(memcg)) {
1375 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1376 /* moving charge context might have finished. */
1379 finish_wait(&mc.waitq, &wait);
1386 struct memory_stat {
1391 static const struct memory_stat memory_stats[] = {
1392 { "anon", NR_ANON_MAPPED },
1393 { "file", NR_FILE_PAGES },
1394 { "kernel", MEMCG_KMEM },
1395 { "kernel_stack", NR_KERNEL_STACK_KB },
1396 { "pagetables", NR_PAGETABLE },
1397 { "percpu", MEMCG_PERCPU_B },
1398 { "sock", MEMCG_SOCK },
1399 { "vmalloc", MEMCG_VMALLOC },
1400 { "shmem", NR_SHMEM },
1401 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
1402 { "zswap", MEMCG_ZSWAP_B },
1403 { "zswapped", MEMCG_ZSWAPPED },
1405 { "file_mapped", NR_FILE_MAPPED },
1406 { "file_dirty", NR_FILE_DIRTY },
1407 { "file_writeback", NR_WRITEBACK },
1409 { "swapcached", NR_SWAPCACHE },
1411 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1412 { "anon_thp", NR_ANON_THPS },
1413 { "file_thp", NR_FILE_THPS },
1414 { "shmem_thp", NR_SHMEM_THPS },
1416 { "inactive_anon", NR_INACTIVE_ANON },
1417 { "active_anon", NR_ACTIVE_ANON },
1418 { "inactive_file", NR_INACTIVE_FILE },
1419 { "active_file", NR_ACTIVE_FILE },
1420 { "unevictable", NR_UNEVICTABLE },
1421 { "slab_reclaimable", NR_SLAB_RECLAIMABLE_B },
1422 { "slab_unreclaimable", NR_SLAB_UNRECLAIMABLE_B },
1424 /* The memory events */
1425 { "workingset_refault_anon", WORKINGSET_REFAULT_ANON },
1426 { "workingset_refault_file", WORKINGSET_REFAULT_FILE },
1427 { "workingset_activate_anon", WORKINGSET_ACTIVATE_ANON },
1428 { "workingset_activate_file", WORKINGSET_ACTIVATE_FILE },
1429 { "workingset_restore_anon", WORKINGSET_RESTORE_ANON },
1430 { "workingset_restore_file", WORKINGSET_RESTORE_FILE },
1431 { "workingset_nodereclaim", WORKINGSET_NODERECLAIM },
1434 /* Translate stat items to the correct unit for memory.stat output */
1435 static int memcg_page_state_unit(int item)
1438 case MEMCG_PERCPU_B:
1440 case NR_SLAB_RECLAIMABLE_B:
1441 case NR_SLAB_UNRECLAIMABLE_B:
1442 case WORKINGSET_REFAULT_ANON:
1443 case WORKINGSET_REFAULT_FILE:
1444 case WORKINGSET_ACTIVATE_ANON:
1445 case WORKINGSET_ACTIVATE_FILE:
1446 case WORKINGSET_RESTORE_ANON:
1447 case WORKINGSET_RESTORE_FILE:
1448 case WORKINGSET_NODERECLAIM:
1450 case NR_KERNEL_STACK_KB:
1457 static inline unsigned long memcg_page_state_output(struct mem_cgroup *memcg,
1460 return memcg_page_state(memcg, item) * memcg_page_state_unit(item);
1463 /* Subset of vm_event_item to report for memcg event stats */
1464 static const unsigned int memcg_vm_event_stat[] = {
1476 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
1480 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1486 static char *memory_stat_format(struct mem_cgroup *memcg)
1491 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1496 * Provide statistics on the state of the memory subsystem as
1497 * well as cumulative event counters that show past behavior.
1499 * This list is ordered following a combination of these gradients:
1500 * 1) generic big picture -> specifics and details
1501 * 2) reflecting userspace activity -> reflecting kernel heuristics
1503 * Current memory state:
1505 mem_cgroup_flush_stats();
1507 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1510 size = memcg_page_state_output(memcg, memory_stats[i].idx);
1511 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1513 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1514 size += memcg_page_state_output(memcg,
1515 NR_SLAB_RECLAIMABLE_B);
1516 seq_buf_printf(&s, "slab %llu\n", size);
1520 /* Accumulated memory events */
1521 seq_buf_printf(&s, "pgscan %lu\n",
1522 memcg_events(memcg, PGSCAN_KSWAPD) +
1523 memcg_events(memcg, PGSCAN_DIRECT));
1524 seq_buf_printf(&s, "pgsteal %lu\n",
1525 memcg_events(memcg, PGSTEAL_KSWAPD) +
1526 memcg_events(memcg, PGSTEAL_DIRECT));
1528 for (i = 0; i < ARRAY_SIZE(memcg_vm_event_stat); i++)
1529 seq_buf_printf(&s, "%s %lu\n",
1530 vm_event_name(memcg_vm_event_stat[i]),
1531 memcg_events(memcg, memcg_vm_event_stat[i]));
1533 /* The above should easily fit into one page */
1534 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1539 #define K(x) ((x) << (PAGE_SHIFT-10))
1541 * mem_cgroup_print_oom_context: Print OOM information relevant to
1542 * memory controller.
1543 * @memcg: The memory cgroup that went over limit
1544 * @p: Task that is going to be killed
1546 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1549 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1554 pr_cont(",oom_memcg=");
1555 pr_cont_cgroup_path(memcg->css.cgroup);
1557 pr_cont(",global_oom");
1559 pr_cont(",task_memcg=");
1560 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1566 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1567 * memory controller.
1568 * @memcg: The memory cgroup that went over limit
1570 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1574 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1575 K((u64)page_counter_read(&memcg->memory)),
1576 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1577 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1578 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1579 K((u64)page_counter_read(&memcg->swap)),
1580 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1582 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1583 K((u64)page_counter_read(&memcg->memsw)),
1584 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1585 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1586 K((u64)page_counter_read(&memcg->kmem)),
1587 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1590 pr_info("Memory cgroup stats for ");
1591 pr_cont_cgroup_path(memcg->css.cgroup);
1593 buf = memory_stat_format(memcg);
1601 * Return the memory (and swap, if configured) limit for a memcg.
1603 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1605 unsigned long max = READ_ONCE(memcg->memory.max);
1607 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1608 if (mem_cgroup_swappiness(memcg))
1609 max += min(READ_ONCE(memcg->swap.max),
1610 (unsigned long)total_swap_pages);
1612 if (mem_cgroup_swappiness(memcg)) {
1613 /* Calculate swap excess capacity from memsw limit */
1614 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1616 max += min(swap, (unsigned long)total_swap_pages);
1622 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1624 return page_counter_read(&memcg->memory);
1627 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1630 struct oom_control oc = {
1634 .gfp_mask = gfp_mask,
1639 if (mutex_lock_killable(&oom_lock))
1642 if (mem_cgroup_margin(memcg) >= (1 << order))
1646 * A few threads which were not waiting at mutex_lock_killable() can
1647 * fail to bail out. Therefore, check again after holding oom_lock.
1649 ret = task_is_dying() || out_of_memory(&oc);
1652 mutex_unlock(&oom_lock);
1656 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1659 unsigned long *total_scanned)
1661 struct mem_cgroup *victim = NULL;
1664 unsigned long excess;
1665 unsigned long nr_scanned;
1666 struct mem_cgroup_reclaim_cookie reclaim = {
1670 excess = soft_limit_excess(root_memcg);
1673 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1678 * If we have not been able to reclaim
1679 * anything, it might because there are
1680 * no reclaimable pages under this hierarchy
1685 * We want to do more targeted reclaim.
1686 * excess >> 2 is not to excessive so as to
1687 * reclaim too much, nor too less that we keep
1688 * coming back to reclaim from this cgroup
1690 if (total >= (excess >> 2) ||
1691 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1696 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1697 pgdat, &nr_scanned);
1698 *total_scanned += nr_scanned;
1699 if (!soft_limit_excess(root_memcg))
1702 mem_cgroup_iter_break(root_memcg, victim);
1706 #ifdef CONFIG_LOCKDEP
1707 static struct lockdep_map memcg_oom_lock_dep_map = {
1708 .name = "memcg_oom_lock",
1712 static DEFINE_SPINLOCK(memcg_oom_lock);
1715 * Check OOM-Killer is already running under our hierarchy.
1716 * If someone is running, return false.
1718 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1720 struct mem_cgroup *iter, *failed = NULL;
1722 spin_lock(&memcg_oom_lock);
1724 for_each_mem_cgroup_tree(iter, memcg) {
1725 if (iter->oom_lock) {
1727 * this subtree of our hierarchy is already locked
1728 * so we cannot give a lock.
1731 mem_cgroup_iter_break(memcg, iter);
1734 iter->oom_lock = true;
1739 * OK, we failed to lock the whole subtree so we have
1740 * to clean up what we set up to the failing subtree
1742 for_each_mem_cgroup_tree(iter, memcg) {
1743 if (iter == failed) {
1744 mem_cgroup_iter_break(memcg, iter);
1747 iter->oom_lock = false;
1750 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1752 spin_unlock(&memcg_oom_lock);
1757 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1759 struct mem_cgroup *iter;
1761 spin_lock(&memcg_oom_lock);
1762 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1763 for_each_mem_cgroup_tree(iter, memcg)
1764 iter->oom_lock = false;
1765 spin_unlock(&memcg_oom_lock);
1768 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1770 struct mem_cgroup *iter;
1772 spin_lock(&memcg_oom_lock);
1773 for_each_mem_cgroup_tree(iter, memcg)
1775 spin_unlock(&memcg_oom_lock);
1778 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1780 struct mem_cgroup *iter;
1783 * Be careful about under_oom underflows because a child memcg
1784 * could have been added after mem_cgroup_mark_under_oom.
1786 spin_lock(&memcg_oom_lock);
1787 for_each_mem_cgroup_tree(iter, memcg)
1788 if (iter->under_oom > 0)
1790 spin_unlock(&memcg_oom_lock);
1793 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1795 struct oom_wait_info {
1796 struct mem_cgroup *memcg;
1797 wait_queue_entry_t wait;
1800 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1801 unsigned mode, int sync, void *arg)
1803 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1804 struct mem_cgroup *oom_wait_memcg;
1805 struct oom_wait_info *oom_wait_info;
1807 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1808 oom_wait_memcg = oom_wait_info->memcg;
1810 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1811 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1813 return autoremove_wake_function(wait, mode, sync, arg);
1816 static void memcg_oom_recover(struct mem_cgroup *memcg)
1819 * For the following lockless ->under_oom test, the only required
1820 * guarantee is that it must see the state asserted by an OOM when
1821 * this function is called as a result of userland actions
1822 * triggered by the notification of the OOM. This is trivially
1823 * achieved by invoking mem_cgroup_mark_under_oom() before
1824 * triggering notification.
1826 if (memcg && memcg->under_oom)
1827 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1831 * Returns true if successfully killed one or more processes. Though in some
1832 * corner cases it can return true even without killing any process.
1834 static bool mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1838 if (order > PAGE_ALLOC_COSTLY_ORDER)
1841 memcg_memory_event(memcg, MEMCG_OOM);
1844 * We are in the middle of the charge context here, so we
1845 * don't want to block when potentially sitting on a callstack
1846 * that holds all kinds of filesystem and mm locks.
1848 * cgroup1 allows disabling the OOM killer and waiting for outside
1849 * handling until the charge can succeed; remember the context and put
1850 * the task to sleep at the end of the page fault when all locks are
1853 * On the other hand, in-kernel OOM killer allows for an async victim
1854 * memory reclaim (oom_reaper) and that means that we are not solely
1855 * relying on the oom victim to make a forward progress and we can
1856 * invoke the oom killer here.
1858 * Please note that mem_cgroup_out_of_memory might fail to find a
1859 * victim and then we have to bail out from the charge path.
1861 if (memcg->oom_kill_disable) {
1862 if (current->in_user_fault) {
1863 css_get(&memcg->css);
1864 current->memcg_in_oom = memcg;
1865 current->memcg_oom_gfp_mask = mask;
1866 current->memcg_oom_order = order;
1871 mem_cgroup_mark_under_oom(memcg);
1873 locked = mem_cgroup_oom_trylock(memcg);
1876 mem_cgroup_oom_notify(memcg);
1878 mem_cgroup_unmark_under_oom(memcg);
1879 ret = mem_cgroup_out_of_memory(memcg, mask, order);
1882 mem_cgroup_oom_unlock(memcg);
1888 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1889 * @handle: actually kill/wait or just clean up the OOM state
1891 * This has to be called at the end of a page fault if the memcg OOM
1892 * handler was enabled.
1894 * Memcg supports userspace OOM handling where failed allocations must
1895 * sleep on a waitqueue until the userspace task resolves the
1896 * situation. Sleeping directly in the charge context with all kinds
1897 * of locks held is not a good idea, instead we remember an OOM state
1898 * in the task and mem_cgroup_oom_synchronize() has to be called at
1899 * the end of the page fault to complete the OOM handling.
1901 * Returns %true if an ongoing memcg OOM situation was detected and
1902 * completed, %false otherwise.
1904 bool mem_cgroup_oom_synchronize(bool handle)
1906 struct mem_cgroup *memcg = current->memcg_in_oom;
1907 struct oom_wait_info owait;
1910 /* OOM is global, do not handle */
1917 owait.memcg = memcg;
1918 owait.wait.flags = 0;
1919 owait.wait.func = memcg_oom_wake_function;
1920 owait.wait.private = current;
1921 INIT_LIST_HEAD(&owait.wait.entry);
1923 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1924 mem_cgroup_mark_under_oom(memcg);
1926 locked = mem_cgroup_oom_trylock(memcg);
1929 mem_cgroup_oom_notify(memcg);
1931 if (locked && !memcg->oom_kill_disable) {
1932 mem_cgroup_unmark_under_oom(memcg);
1933 finish_wait(&memcg_oom_waitq, &owait.wait);
1934 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1935 current->memcg_oom_order);
1938 mem_cgroup_unmark_under_oom(memcg);
1939 finish_wait(&memcg_oom_waitq, &owait.wait);
1943 mem_cgroup_oom_unlock(memcg);
1945 * There is no guarantee that an OOM-lock contender
1946 * sees the wakeups triggered by the OOM kill
1947 * uncharges. Wake any sleepers explicitly.
1949 memcg_oom_recover(memcg);
1952 current->memcg_in_oom = NULL;
1953 css_put(&memcg->css);
1958 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1959 * @victim: task to be killed by the OOM killer
1960 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1962 * Returns a pointer to a memory cgroup, which has to be cleaned up
1963 * by killing all belonging OOM-killable tasks.
1965 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1967 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1968 struct mem_cgroup *oom_domain)
1970 struct mem_cgroup *oom_group = NULL;
1971 struct mem_cgroup *memcg;
1973 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
1977 oom_domain = root_mem_cgroup;
1981 memcg = mem_cgroup_from_task(victim);
1982 if (memcg == root_mem_cgroup)
1986 * If the victim task has been asynchronously moved to a different
1987 * memory cgroup, we might end up killing tasks outside oom_domain.
1988 * In this case it's better to ignore memory.group.oom.
1990 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
1994 * Traverse the memory cgroup hierarchy from the victim task's
1995 * cgroup up to the OOMing cgroup (or root) to find the
1996 * highest-level memory cgroup with oom.group set.
1998 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1999 if (memcg->oom_group)
2002 if (memcg == oom_domain)
2007 css_get(&oom_group->css);
2014 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2016 pr_info("Tasks in ");
2017 pr_cont_cgroup_path(memcg->css.cgroup);
2018 pr_cont(" are going to be killed due to memory.oom.group set\n");
2022 * folio_memcg_lock - Bind a folio to its memcg.
2023 * @folio: The folio.
2025 * This function prevents unlocked LRU folios from being moved to
2028 * It ensures lifetime of the bound memcg. The caller is responsible
2029 * for the lifetime of the folio.
2031 void folio_memcg_lock(struct folio *folio)
2033 struct mem_cgroup *memcg;
2034 unsigned long flags;
2037 * The RCU lock is held throughout the transaction. The fast
2038 * path can get away without acquiring the memcg->move_lock
2039 * because page moving starts with an RCU grace period.
2043 if (mem_cgroup_disabled())
2046 memcg = folio_memcg(folio);
2047 if (unlikely(!memcg))
2050 #ifdef CONFIG_PROVE_LOCKING
2051 local_irq_save(flags);
2052 might_lock(&memcg->move_lock);
2053 local_irq_restore(flags);
2056 if (atomic_read(&memcg->moving_account) <= 0)
2059 spin_lock_irqsave(&memcg->move_lock, flags);
2060 if (memcg != folio_memcg(folio)) {
2061 spin_unlock_irqrestore(&memcg->move_lock, flags);
2066 * When charge migration first begins, we can have multiple
2067 * critical sections holding the fast-path RCU lock and one
2068 * holding the slowpath move_lock. Track the task who has the
2069 * move_lock for unlock_page_memcg().
2071 memcg->move_lock_task = current;
2072 memcg->move_lock_flags = flags;
2075 void lock_page_memcg(struct page *page)
2077 folio_memcg_lock(page_folio(page));
2080 static void __folio_memcg_unlock(struct mem_cgroup *memcg)
2082 if (memcg && memcg->move_lock_task == current) {
2083 unsigned long flags = memcg->move_lock_flags;
2085 memcg->move_lock_task = NULL;
2086 memcg->move_lock_flags = 0;
2088 spin_unlock_irqrestore(&memcg->move_lock, flags);
2095 * folio_memcg_unlock - Release the binding between a folio and its memcg.
2096 * @folio: The folio.
2098 * This releases the binding created by folio_memcg_lock(). This does
2099 * not change the accounting of this folio to its memcg, but it does
2100 * permit others to change it.
2102 void folio_memcg_unlock(struct folio *folio)
2104 __folio_memcg_unlock(folio_memcg(folio));
2107 void unlock_page_memcg(struct page *page)
2109 folio_memcg_unlock(page_folio(page));
2112 struct memcg_stock_pcp {
2113 local_lock_t stock_lock;
2114 struct mem_cgroup *cached; /* this never be root cgroup */
2115 unsigned int nr_pages;
2117 #ifdef CONFIG_MEMCG_KMEM
2118 struct obj_cgroup *cached_objcg;
2119 struct pglist_data *cached_pgdat;
2120 unsigned int nr_bytes;
2121 int nr_slab_reclaimable_b;
2122 int nr_slab_unreclaimable_b;
2125 struct work_struct work;
2126 unsigned long flags;
2127 #define FLUSHING_CACHED_CHARGE 0
2129 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock) = {
2130 .stock_lock = INIT_LOCAL_LOCK(stock_lock),
2132 static DEFINE_MUTEX(percpu_charge_mutex);
2134 #ifdef CONFIG_MEMCG_KMEM
2135 static struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock);
2136 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2137 struct mem_cgroup *root_memcg);
2138 static void memcg_account_kmem(struct mem_cgroup *memcg, int nr_pages);
2141 static inline struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock)
2145 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2146 struct mem_cgroup *root_memcg)
2150 static void memcg_account_kmem(struct mem_cgroup *memcg, int nr_pages)
2156 * consume_stock: Try to consume stocked charge on this cpu.
2157 * @memcg: memcg to consume from.
2158 * @nr_pages: how many pages to charge.
2160 * The charges will only happen if @memcg matches the current cpu's memcg
2161 * stock, and at least @nr_pages are available in that stock. Failure to
2162 * service an allocation will refill the stock.
2164 * returns true if successful, false otherwise.
2166 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2168 struct memcg_stock_pcp *stock;
2169 unsigned long flags;
2172 if (nr_pages > MEMCG_CHARGE_BATCH)
2175 local_lock_irqsave(&memcg_stock.stock_lock, flags);
2177 stock = this_cpu_ptr(&memcg_stock);
2178 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2179 stock->nr_pages -= nr_pages;
2183 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2189 * Returns stocks cached in percpu and reset cached information.
2191 static void drain_stock(struct memcg_stock_pcp *stock)
2193 struct mem_cgroup *old = stock->cached;
2198 if (stock->nr_pages) {
2199 page_counter_uncharge(&old->memory, stock->nr_pages);
2200 if (do_memsw_account())
2201 page_counter_uncharge(&old->memsw, stock->nr_pages);
2202 stock->nr_pages = 0;
2206 stock->cached = NULL;
2209 static void drain_local_stock(struct work_struct *dummy)
2211 struct memcg_stock_pcp *stock;
2212 struct obj_cgroup *old = NULL;
2213 unsigned long flags;
2216 * The only protection from cpu hotplug (memcg_hotplug_cpu_dead) vs.
2217 * drain_stock races is that we always operate on local CPU stock
2218 * here with IRQ disabled
2220 local_lock_irqsave(&memcg_stock.stock_lock, flags);
2222 stock = this_cpu_ptr(&memcg_stock);
2223 old = drain_obj_stock(stock);
2225 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2227 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2229 obj_cgroup_put(old);
2233 * Cache charges(val) to local per_cpu area.
2234 * This will be consumed by consume_stock() function, later.
2236 static void __refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2238 struct memcg_stock_pcp *stock;
2240 stock = this_cpu_ptr(&memcg_stock);
2241 if (stock->cached != memcg) { /* reset if necessary */
2243 css_get(&memcg->css);
2244 stock->cached = memcg;
2246 stock->nr_pages += nr_pages;
2248 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2252 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2254 unsigned long flags;
2256 local_lock_irqsave(&memcg_stock.stock_lock, flags);
2257 __refill_stock(memcg, nr_pages);
2258 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2262 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2263 * of the hierarchy under it.
2265 static void drain_all_stock(struct mem_cgroup *root_memcg)
2269 /* If someone's already draining, avoid adding running more workers. */
2270 if (!mutex_trylock(&percpu_charge_mutex))
2273 * Notify other cpus that system-wide "drain" is running
2274 * We do not care about races with the cpu hotplug because cpu down
2275 * as well as workers from this path always operate on the local
2276 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2279 curcpu = smp_processor_id();
2280 for_each_online_cpu(cpu) {
2281 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2282 struct mem_cgroup *memcg;
2286 memcg = stock->cached;
2287 if (memcg && stock->nr_pages &&
2288 mem_cgroup_is_descendant(memcg, root_memcg))
2290 else if (obj_stock_flush_required(stock, root_memcg))
2295 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2297 drain_local_stock(&stock->work);
2299 schedule_work_on(cpu, &stock->work);
2303 mutex_unlock(&percpu_charge_mutex);
2306 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2308 struct memcg_stock_pcp *stock;
2310 stock = &per_cpu(memcg_stock, cpu);
2316 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2317 unsigned int nr_pages,
2320 unsigned long nr_reclaimed = 0;
2323 unsigned long pflags;
2325 if (page_counter_read(&memcg->memory) <=
2326 READ_ONCE(memcg->memory.high))
2329 memcg_memory_event(memcg, MEMCG_HIGH);
2331 psi_memstall_enter(&pflags);
2332 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2334 psi_memstall_leave(&pflags);
2335 } while ((memcg = parent_mem_cgroup(memcg)) &&
2336 !mem_cgroup_is_root(memcg));
2338 return nr_reclaimed;
2341 static void high_work_func(struct work_struct *work)
2343 struct mem_cgroup *memcg;
2345 memcg = container_of(work, struct mem_cgroup, high_work);
2346 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2350 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2351 * enough to still cause a significant slowdown in most cases, while still
2352 * allowing diagnostics and tracing to proceed without becoming stuck.
2354 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2357 * When calculating the delay, we use these either side of the exponentiation to
2358 * maintain precision and scale to a reasonable number of jiffies (see the table
2361 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2362 * overage ratio to a delay.
2363 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2364 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2365 * to produce a reasonable delay curve.
2367 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2368 * reasonable delay curve compared to precision-adjusted overage, not
2369 * penalising heavily at first, but still making sure that growth beyond the
2370 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2371 * example, with a high of 100 megabytes:
2373 * +-------+------------------------+
2374 * | usage | time to allocate in ms |
2375 * +-------+------------------------+
2397 * +-------+------------------------+
2399 #define MEMCG_DELAY_PRECISION_SHIFT 20
2400 #define MEMCG_DELAY_SCALING_SHIFT 14
2402 static u64 calculate_overage(unsigned long usage, unsigned long high)
2410 * Prevent division by 0 in overage calculation by acting as if
2411 * it was a threshold of 1 page
2413 high = max(high, 1UL);
2415 overage = usage - high;
2416 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2417 return div64_u64(overage, high);
2420 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2422 u64 overage, max_overage = 0;
2425 overage = calculate_overage(page_counter_read(&memcg->memory),
2426 READ_ONCE(memcg->memory.high));
2427 max_overage = max(overage, max_overage);
2428 } while ((memcg = parent_mem_cgroup(memcg)) &&
2429 !mem_cgroup_is_root(memcg));
2434 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2436 u64 overage, max_overage = 0;
2439 overage = calculate_overage(page_counter_read(&memcg->swap),
2440 READ_ONCE(memcg->swap.high));
2442 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2443 max_overage = max(overage, max_overage);
2444 } while ((memcg = parent_mem_cgroup(memcg)) &&
2445 !mem_cgroup_is_root(memcg));
2451 * Get the number of jiffies that we should penalise a mischievous cgroup which
2452 * is exceeding its memory.high by checking both it and its ancestors.
2454 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2455 unsigned int nr_pages,
2458 unsigned long penalty_jiffies;
2464 * We use overage compared to memory.high to calculate the number of
2465 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2466 * fairly lenient on small overages, and increasingly harsh when the
2467 * memcg in question makes it clear that it has no intention of stopping
2468 * its crazy behaviour, so we exponentially increase the delay based on
2471 penalty_jiffies = max_overage * max_overage * HZ;
2472 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2473 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2476 * Factor in the task's own contribution to the overage, such that four
2477 * N-sized allocations are throttled approximately the same as one
2478 * 4N-sized allocation.
2480 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2481 * larger the current charge patch is than that.
2483 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2487 * Scheduled by try_charge() to be executed from the userland return path
2488 * and reclaims memory over the high limit.
2490 void mem_cgroup_handle_over_high(void)
2492 unsigned long penalty_jiffies;
2493 unsigned long pflags;
2494 unsigned long nr_reclaimed;
2495 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2496 int nr_retries = MAX_RECLAIM_RETRIES;
2497 struct mem_cgroup *memcg;
2498 bool in_retry = false;
2500 if (likely(!nr_pages))
2503 memcg = get_mem_cgroup_from_mm(current->mm);
2504 current->memcg_nr_pages_over_high = 0;
2508 * The allocating task should reclaim at least the batch size, but for
2509 * subsequent retries we only want to do what's necessary to prevent oom
2510 * or breaching resource isolation.
2512 * This is distinct from memory.max or page allocator behaviour because
2513 * memory.high is currently batched, whereas memory.max and the page
2514 * allocator run every time an allocation is made.
2516 nr_reclaimed = reclaim_high(memcg,
2517 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2521 * memory.high is breached and reclaim is unable to keep up. Throttle
2522 * allocators proactively to slow down excessive growth.
2524 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2525 mem_find_max_overage(memcg));
2527 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2528 swap_find_max_overage(memcg));
2531 * Clamp the max delay per usermode return so as to still keep the
2532 * application moving forwards and also permit diagnostics, albeit
2535 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2538 * Don't sleep if the amount of jiffies this memcg owes us is so low
2539 * that it's not even worth doing, in an attempt to be nice to those who
2540 * go only a small amount over their memory.high value and maybe haven't
2541 * been aggressively reclaimed enough yet.
2543 if (penalty_jiffies <= HZ / 100)
2547 * If reclaim is making forward progress but we're still over
2548 * memory.high, we want to encourage that rather than doing allocator
2551 if (nr_reclaimed || nr_retries--) {
2557 * If we exit early, we're guaranteed to die (since
2558 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2559 * need to account for any ill-begotten jiffies to pay them off later.
2561 psi_memstall_enter(&pflags);
2562 schedule_timeout_killable(penalty_jiffies);
2563 psi_memstall_leave(&pflags);
2566 css_put(&memcg->css);
2569 static int try_charge_memcg(struct mem_cgroup *memcg, gfp_t gfp_mask,
2570 unsigned int nr_pages)
2572 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2573 int nr_retries = MAX_RECLAIM_RETRIES;
2574 struct mem_cgroup *mem_over_limit;
2575 struct page_counter *counter;
2576 unsigned long nr_reclaimed;
2577 bool passed_oom = false;
2578 bool may_swap = true;
2579 bool drained = false;
2580 unsigned long pflags;
2583 if (consume_stock(memcg, nr_pages))
2586 if (!do_memsw_account() ||
2587 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2588 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2590 if (do_memsw_account())
2591 page_counter_uncharge(&memcg->memsw, batch);
2592 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2594 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2598 if (batch > nr_pages) {
2604 * Prevent unbounded recursion when reclaim operations need to
2605 * allocate memory. This might exceed the limits temporarily,
2606 * but we prefer facilitating memory reclaim and getting back
2607 * under the limit over triggering OOM kills in these cases.
2609 if (unlikely(current->flags & PF_MEMALLOC))
2612 if (unlikely(task_in_memcg_oom(current)))
2615 if (!gfpflags_allow_blocking(gfp_mask))
2618 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2620 psi_memstall_enter(&pflags);
2621 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2622 gfp_mask, may_swap);
2623 psi_memstall_leave(&pflags);
2625 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2629 drain_all_stock(mem_over_limit);
2634 if (gfp_mask & __GFP_NORETRY)
2637 * Even though the limit is exceeded at this point, reclaim
2638 * may have been able to free some pages. Retry the charge
2639 * before killing the task.
2641 * Only for regular pages, though: huge pages are rather
2642 * unlikely to succeed so close to the limit, and we fall back
2643 * to regular pages anyway in case of failure.
2645 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2648 * At task move, charge accounts can be doubly counted. So, it's
2649 * better to wait until the end of task_move if something is going on.
2651 if (mem_cgroup_wait_acct_move(mem_over_limit))
2657 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2660 /* Avoid endless loop for tasks bypassed by the oom killer */
2661 if (passed_oom && task_is_dying())
2665 * keep retrying as long as the memcg oom killer is able to make
2666 * a forward progress or bypass the charge if the oom killer
2667 * couldn't make any progress.
2669 if (mem_cgroup_oom(mem_over_limit, gfp_mask,
2670 get_order(nr_pages * PAGE_SIZE))) {
2672 nr_retries = MAX_RECLAIM_RETRIES;
2677 * Memcg doesn't have a dedicated reserve for atomic
2678 * allocations. But like the global atomic pool, we need to
2679 * put the burden of reclaim on regular allocation requests
2680 * and let these go through as privileged allocations.
2682 if (!(gfp_mask & (__GFP_NOFAIL | __GFP_HIGH)))
2686 * The allocation either can't fail or will lead to more memory
2687 * being freed very soon. Allow memory usage go over the limit
2688 * temporarily by force charging it.
2690 page_counter_charge(&memcg->memory, nr_pages);
2691 if (do_memsw_account())
2692 page_counter_charge(&memcg->memsw, nr_pages);
2697 if (batch > nr_pages)
2698 refill_stock(memcg, batch - nr_pages);
2701 * If the hierarchy is above the normal consumption range, schedule
2702 * reclaim on returning to userland. We can perform reclaim here
2703 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2704 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2705 * not recorded as it most likely matches current's and won't
2706 * change in the meantime. As high limit is checked again before
2707 * reclaim, the cost of mismatch is negligible.
2710 bool mem_high, swap_high;
2712 mem_high = page_counter_read(&memcg->memory) >
2713 READ_ONCE(memcg->memory.high);
2714 swap_high = page_counter_read(&memcg->swap) >
2715 READ_ONCE(memcg->swap.high);
2717 /* Don't bother a random interrupted task */
2720 schedule_work(&memcg->high_work);
2726 if (mem_high || swap_high) {
2728 * The allocating tasks in this cgroup will need to do
2729 * reclaim or be throttled to prevent further growth
2730 * of the memory or swap footprints.
2732 * Target some best-effort fairness between the tasks,
2733 * and distribute reclaim work and delay penalties
2734 * based on how much each task is actually allocating.
2736 current->memcg_nr_pages_over_high += batch;
2737 set_notify_resume(current);
2740 } while ((memcg = parent_mem_cgroup(memcg)));
2742 if (current->memcg_nr_pages_over_high > MEMCG_CHARGE_BATCH &&
2743 !(current->flags & PF_MEMALLOC) &&
2744 gfpflags_allow_blocking(gfp_mask)) {
2745 mem_cgroup_handle_over_high();
2750 static inline int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2751 unsigned int nr_pages)
2753 if (mem_cgroup_is_root(memcg))
2756 return try_charge_memcg(memcg, gfp_mask, nr_pages);
2759 static inline void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2761 if (mem_cgroup_is_root(memcg))
2764 page_counter_uncharge(&memcg->memory, nr_pages);
2765 if (do_memsw_account())
2766 page_counter_uncharge(&memcg->memsw, nr_pages);
2769 static void commit_charge(struct folio *folio, struct mem_cgroup *memcg)
2771 VM_BUG_ON_FOLIO(folio_memcg(folio), folio);
2773 * Any of the following ensures page's memcg stability:
2777 * - lock_page_memcg()
2778 * - exclusive reference
2780 folio->memcg_data = (unsigned long)memcg;
2783 #ifdef CONFIG_MEMCG_KMEM
2785 * The allocated objcg pointers array is not accounted directly.
2786 * Moreover, it should not come from DMA buffer and is not readily
2787 * reclaimable. So those GFP bits should be masked off.
2789 #define OBJCGS_CLEAR_MASK (__GFP_DMA | __GFP_RECLAIMABLE | __GFP_ACCOUNT)
2792 * mod_objcg_mlstate() may be called with irq enabled, so
2793 * mod_memcg_lruvec_state() should be used.
2795 static inline void mod_objcg_mlstate(struct obj_cgroup *objcg,
2796 struct pglist_data *pgdat,
2797 enum node_stat_item idx, int nr)
2799 struct mem_cgroup *memcg;
2800 struct lruvec *lruvec;
2803 memcg = obj_cgroup_memcg(objcg);
2804 lruvec = mem_cgroup_lruvec(memcg, pgdat);
2805 mod_memcg_lruvec_state(lruvec, idx, nr);
2809 int memcg_alloc_slab_cgroups(struct slab *slab, struct kmem_cache *s,
2810 gfp_t gfp, bool new_slab)
2812 unsigned int objects = objs_per_slab(s, slab);
2813 unsigned long memcg_data;
2816 gfp &= ~OBJCGS_CLEAR_MASK;
2817 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2822 memcg_data = (unsigned long) vec | MEMCG_DATA_OBJCGS;
2825 * If the slab is brand new and nobody can yet access its
2826 * memcg_data, no synchronization is required and memcg_data can
2827 * be simply assigned.
2829 slab->memcg_data = memcg_data;
2830 } else if (cmpxchg(&slab->memcg_data, 0, memcg_data)) {
2832 * If the slab is already in use, somebody can allocate and
2833 * assign obj_cgroups in parallel. In this case the existing
2834 * objcg vector should be reused.
2840 kmemleak_not_leak(vec);
2845 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2847 * A passed kernel object can be a slab object or a generic kernel page, so
2848 * different mechanisms for getting the memory cgroup pointer should be used.
2849 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2850 * can not know for sure how the kernel object is implemented.
2851 * mem_cgroup_from_obj() can be safely used in such cases.
2853 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2854 * cgroup_mutex, etc.
2856 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2858 struct folio *folio;
2860 if (mem_cgroup_disabled())
2863 folio = virt_to_folio(p);
2866 * Slab objects are accounted individually, not per-page.
2867 * Memcg membership data for each individual object is saved in
2870 if (folio_test_slab(folio)) {
2871 struct obj_cgroup **objcgs;
2875 slab = folio_slab(folio);
2876 objcgs = slab_objcgs(slab);
2880 off = obj_to_index(slab->slab_cache, slab, p);
2882 return obj_cgroup_memcg(objcgs[off]);
2888 * page_memcg_check() is used here, because in theory we can encounter
2889 * a folio where the slab flag has been cleared already, but
2890 * slab->memcg_data has not been freed yet
2891 * page_memcg_check(page) will guarantee that a proper memory
2892 * cgroup pointer or NULL will be returned.
2894 return page_memcg_check(folio_page(folio, 0));
2897 static struct obj_cgroup *__get_obj_cgroup_from_memcg(struct mem_cgroup *memcg)
2899 struct obj_cgroup *objcg = NULL;
2901 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2902 objcg = rcu_dereference(memcg->objcg);
2903 if (objcg && obj_cgroup_tryget(objcg))
2910 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2912 struct obj_cgroup *objcg = NULL;
2913 struct mem_cgroup *memcg;
2915 if (memcg_kmem_bypass())
2919 if (unlikely(active_memcg()))
2920 memcg = active_memcg();
2922 memcg = mem_cgroup_from_task(current);
2923 objcg = __get_obj_cgroup_from_memcg(memcg);
2928 struct obj_cgroup *get_obj_cgroup_from_page(struct page *page)
2930 struct obj_cgroup *objcg;
2932 if (!memcg_kmem_enabled() || memcg_kmem_bypass())
2935 if (PageMemcgKmem(page)) {
2936 objcg = __folio_objcg(page_folio(page));
2937 obj_cgroup_get(objcg);
2939 struct mem_cgroup *memcg;
2942 memcg = __folio_memcg(page_folio(page));
2944 objcg = __get_obj_cgroup_from_memcg(memcg);
2952 static void memcg_account_kmem(struct mem_cgroup *memcg, int nr_pages)
2954 mod_memcg_state(memcg, MEMCG_KMEM, nr_pages);
2955 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
2957 page_counter_charge(&memcg->kmem, nr_pages);
2959 page_counter_uncharge(&memcg->kmem, -nr_pages);
2965 * obj_cgroup_uncharge_pages: uncharge a number of kernel pages from a objcg
2966 * @objcg: object cgroup to uncharge
2967 * @nr_pages: number of pages to uncharge
2969 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
2970 unsigned int nr_pages)
2972 struct mem_cgroup *memcg;
2974 memcg = get_mem_cgroup_from_objcg(objcg);
2976 memcg_account_kmem(memcg, -nr_pages);
2977 refill_stock(memcg, nr_pages);
2979 css_put(&memcg->css);
2983 * obj_cgroup_charge_pages: charge a number of kernel pages to a objcg
2984 * @objcg: object cgroup to charge
2985 * @gfp: reclaim mode
2986 * @nr_pages: number of pages to charge
2988 * Returns 0 on success, an error code on failure.
2990 static int obj_cgroup_charge_pages(struct obj_cgroup *objcg, gfp_t gfp,
2991 unsigned int nr_pages)
2993 struct mem_cgroup *memcg;
2996 memcg = get_mem_cgroup_from_objcg(objcg);
2998 ret = try_charge_memcg(memcg, gfp, nr_pages);
3002 memcg_account_kmem(memcg, nr_pages);
3004 css_put(&memcg->css);
3010 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3011 * @page: page to charge
3012 * @gfp: reclaim mode
3013 * @order: allocation order
3015 * Returns 0 on success, an error code on failure.
3017 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3019 struct obj_cgroup *objcg;
3022 objcg = get_obj_cgroup_from_current();
3024 ret = obj_cgroup_charge_pages(objcg, gfp, 1 << order);
3026 page->memcg_data = (unsigned long)objcg |
3030 obj_cgroup_put(objcg);
3036 * __memcg_kmem_uncharge_page: uncharge a kmem page
3037 * @page: page to uncharge
3038 * @order: allocation order
3040 void __memcg_kmem_uncharge_page(struct page *page, int order)
3042 struct folio *folio = page_folio(page);
3043 struct obj_cgroup *objcg;
3044 unsigned int nr_pages = 1 << order;
3046 if (!folio_memcg_kmem(folio))
3049 objcg = __folio_objcg(folio);
3050 obj_cgroup_uncharge_pages(objcg, nr_pages);
3051 folio->memcg_data = 0;
3052 obj_cgroup_put(objcg);
3055 void mod_objcg_state(struct obj_cgroup *objcg, struct pglist_data *pgdat,
3056 enum node_stat_item idx, int nr)
3058 struct memcg_stock_pcp *stock;
3059 struct obj_cgroup *old = NULL;
3060 unsigned long flags;
3063 local_lock_irqsave(&memcg_stock.stock_lock, flags);
3064 stock = this_cpu_ptr(&memcg_stock);
3067 * Save vmstat data in stock and skip vmstat array update unless
3068 * accumulating over a page of vmstat data or when pgdat or idx
3071 if (stock->cached_objcg != objcg) {
3072 old = drain_obj_stock(stock);
3073 obj_cgroup_get(objcg);
3074 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3075 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3076 stock->cached_objcg = objcg;
3077 stock->cached_pgdat = pgdat;
3078 } else if (stock->cached_pgdat != pgdat) {
3079 /* Flush the existing cached vmstat data */
3080 struct pglist_data *oldpg = stock->cached_pgdat;
3082 if (stock->nr_slab_reclaimable_b) {
3083 mod_objcg_mlstate(objcg, oldpg, NR_SLAB_RECLAIMABLE_B,
3084 stock->nr_slab_reclaimable_b);
3085 stock->nr_slab_reclaimable_b = 0;
3087 if (stock->nr_slab_unreclaimable_b) {
3088 mod_objcg_mlstate(objcg, oldpg, NR_SLAB_UNRECLAIMABLE_B,
3089 stock->nr_slab_unreclaimable_b);
3090 stock->nr_slab_unreclaimable_b = 0;
3092 stock->cached_pgdat = pgdat;
3095 bytes = (idx == NR_SLAB_RECLAIMABLE_B) ? &stock->nr_slab_reclaimable_b
3096 : &stock->nr_slab_unreclaimable_b;
3098 * Even for large object >= PAGE_SIZE, the vmstat data will still be
3099 * cached locally at least once before pushing it out.
3106 if (abs(*bytes) > PAGE_SIZE) {
3114 mod_objcg_mlstate(objcg, pgdat, idx, nr);
3116 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3118 obj_cgroup_put(old);
3121 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3123 struct memcg_stock_pcp *stock;
3124 unsigned long flags;
3127 local_lock_irqsave(&memcg_stock.stock_lock, flags);
3129 stock = this_cpu_ptr(&memcg_stock);
3130 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3131 stock->nr_bytes -= nr_bytes;
3135 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3140 static struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock)
3142 struct obj_cgroup *old = stock->cached_objcg;
3147 if (stock->nr_bytes) {
3148 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3149 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3152 struct mem_cgroup *memcg;
3154 memcg = get_mem_cgroup_from_objcg(old);
3156 memcg_account_kmem(memcg, -nr_pages);
3157 __refill_stock(memcg, nr_pages);
3159 css_put(&memcg->css);
3163 * The leftover is flushed to the centralized per-memcg value.
3164 * On the next attempt to refill obj stock it will be moved
3165 * to a per-cpu stock (probably, on an other CPU), see
3166 * refill_obj_stock().
3168 * How often it's flushed is a trade-off between the memory
3169 * limit enforcement accuracy and potential CPU contention,
3170 * so it might be changed in the future.
3172 atomic_add(nr_bytes, &old->nr_charged_bytes);
3173 stock->nr_bytes = 0;
3177 * Flush the vmstat data in current stock
3179 if (stock->nr_slab_reclaimable_b || stock->nr_slab_unreclaimable_b) {
3180 if (stock->nr_slab_reclaimable_b) {
3181 mod_objcg_mlstate(old, stock->cached_pgdat,
3182 NR_SLAB_RECLAIMABLE_B,
3183 stock->nr_slab_reclaimable_b);
3184 stock->nr_slab_reclaimable_b = 0;
3186 if (stock->nr_slab_unreclaimable_b) {
3187 mod_objcg_mlstate(old, stock->cached_pgdat,
3188 NR_SLAB_UNRECLAIMABLE_B,
3189 stock->nr_slab_unreclaimable_b);
3190 stock->nr_slab_unreclaimable_b = 0;
3192 stock->cached_pgdat = NULL;
3195 stock->cached_objcg = NULL;
3197 * The `old' objects needs to be released by the caller via
3198 * obj_cgroup_put() outside of memcg_stock_pcp::stock_lock.
3203 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3204 struct mem_cgroup *root_memcg)
3206 struct mem_cgroup *memcg;
3208 if (stock->cached_objcg) {
3209 memcg = obj_cgroup_memcg(stock->cached_objcg);
3210 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3217 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes,
3218 bool allow_uncharge)
3220 struct memcg_stock_pcp *stock;
3221 struct obj_cgroup *old = NULL;
3222 unsigned long flags;
3223 unsigned int nr_pages = 0;
3225 local_lock_irqsave(&memcg_stock.stock_lock, flags);
3227 stock = this_cpu_ptr(&memcg_stock);
3228 if (stock->cached_objcg != objcg) { /* reset if necessary */
3229 old = drain_obj_stock(stock);
3230 obj_cgroup_get(objcg);
3231 stock->cached_objcg = objcg;
3232 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3233 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3234 allow_uncharge = true; /* Allow uncharge when objcg changes */
3236 stock->nr_bytes += nr_bytes;
3238 if (allow_uncharge && (stock->nr_bytes > PAGE_SIZE)) {
3239 nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3240 stock->nr_bytes &= (PAGE_SIZE - 1);
3243 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3245 obj_cgroup_put(old);
3248 obj_cgroup_uncharge_pages(objcg, nr_pages);
3251 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3253 unsigned int nr_pages, nr_bytes;
3256 if (consume_obj_stock(objcg, size))
3260 * In theory, objcg->nr_charged_bytes can have enough
3261 * pre-charged bytes to satisfy the allocation. However,
3262 * flushing objcg->nr_charged_bytes requires two atomic
3263 * operations, and objcg->nr_charged_bytes can't be big.
3264 * The shared objcg->nr_charged_bytes can also become a
3265 * performance bottleneck if all tasks of the same memcg are
3266 * trying to update it. So it's better to ignore it and try
3267 * grab some new pages. The stock's nr_bytes will be flushed to
3268 * objcg->nr_charged_bytes later on when objcg changes.
3270 * The stock's nr_bytes may contain enough pre-charged bytes
3271 * to allow one less page from being charged, but we can't rely
3272 * on the pre-charged bytes not being changed outside of
3273 * consume_obj_stock() or refill_obj_stock(). So ignore those
3274 * pre-charged bytes as well when charging pages. To avoid a
3275 * page uncharge right after a page charge, we set the
3276 * allow_uncharge flag to false when calling refill_obj_stock()
3277 * to temporarily allow the pre-charged bytes to exceed the page
3278 * size limit. The maximum reachable value of the pre-charged
3279 * bytes is (sizeof(object) + PAGE_SIZE - 2) if there is no data
3282 nr_pages = size >> PAGE_SHIFT;
3283 nr_bytes = size & (PAGE_SIZE - 1);
3288 ret = obj_cgroup_charge_pages(objcg, gfp, nr_pages);
3289 if (!ret && nr_bytes)
3290 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes, false);
3295 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3297 refill_obj_stock(objcg, size, true);
3300 #endif /* CONFIG_MEMCG_KMEM */
3303 * Because page_memcg(head) is not set on tails, set it now.
3305 void split_page_memcg(struct page *head, unsigned int nr)
3307 struct folio *folio = page_folio(head);
3308 struct mem_cgroup *memcg = folio_memcg(folio);
3311 if (mem_cgroup_disabled() || !memcg)
3314 for (i = 1; i < nr; i++)
3315 folio_page(folio, i)->memcg_data = folio->memcg_data;
3317 if (folio_memcg_kmem(folio))
3318 obj_cgroup_get_many(__folio_objcg(folio), nr - 1);
3320 css_get_many(&memcg->css, nr - 1);
3323 #ifdef CONFIG_MEMCG_SWAP
3325 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3326 * @entry: swap entry to be moved
3327 * @from: mem_cgroup which the entry is moved from
3328 * @to: mem_cgroup which the entry is moved to
3330 * It succeeds only when the swap_cgroup's record for this entry is the same
3331 * as the mem_cgroup's id of @from.
3333 * Returns 0 on success, -EINVAL on failure.
3335 * The caller must have charged to @to, IOW, called page_counter_charge() about
3336 * both res and memsw, and called css_get().
3338 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3339 struct mem_cgroup *from, struct mem_cgroup *to)
3341 unsigned short old_id, new_id;
3343 old_id = mem_cgroup_id(from);
3344 new_id = mem_cgroup_id(to);
3346 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3347 mod_memcg_state(from, MEMCG_SWAP, -1);
3348 mod_memcg_state(to, MEMCG_SWAP, 1);
3354 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3355 struct mem_cgroup *from, struct mem_cgroup *to)
3361 static DEFINE_MUTEX(memcg_max_mutex);
3363 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3364 unsigned long max, bool memsw)
3366 bool enlarge = false;
3367 bool drained = false;
3369 bool limits_invariant;
3370 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3373 if (signal_pending(current)) {
3378 mutex_lock(&memcg_max_mutex);
3380 * Make sure that the new limit (memsw or memory limit) doesn't
3381 * break our basic invariant rule memory.max <= memsw.max.
3383 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3384 max <= memcg->memsw.max;
3385 if (!limits_invariant) {
3386 mutex_unlock(&memcg_max_mutex);
3390 if (max > counter->max)
3392 ret = page_counter_set_max(counter, max);
3393 mutex_unlock(&memcg_max_mutex);
3399 drain_all_stock(memcg);
3404 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3405 GFP_KERNEL, !memsw)) {
3411 if (!ret && enlarge)
3412 memcg_oom_recover(memcg);
3417 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3419 unsigned long *total_scanned)
3421 unsigned long nr_reclaimed = 0;
3422 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3423 unsigned long reclaimed;
3425 struct mem_cgroup_tree_per_node *mctz;
3426 unsigned long excess;
3431 mctz = soft_limit_tree.rb_tree_per_node[pgdat->node_id];
3434 * Do not even bother to check the largest node if the root
3435 * is empty. Do it lockless to prevent lock bouncing. Races
3436 * are acceptable as soft limit is best effort anyway.
3438 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3442 * This loop can run a while, specially if mem_cgroup's continuously
3443 * keep exceeding their soft limit and putting the system under
3450 mz = mem_cgroup_largest_soft_limit_node(mctz);
3454 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3455 gfp_mask, total_scanned);
3456 nr_reclaimed += reclaimed;
3457 spin_lock_irq(&mctz->lock);
3460 * If we failed to reclaim anything from this memory cgroup
3461 * it is time to move on to the next cgroup
3465 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3467 excess = soft_limit_excess(mz->memcg);
3469 * One school of thought says that we should not add
3470 * back the node to the tree if reclaim returns 0.
3471 * But our reclaim could return 0, simply because due
3472 * to priority we are exposing a smaller subset of
3473 * memory to reclaim from. Consider this as a longer
3476 /* If excess == 0, no tree ops */
3477 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3478 spin_unlock_irq(&mctz->lock);
3479 css_put(&mz->memcg->css);
3482 * Could not reclaim anything and there are no more
3483 * mem cgroups to try or we seem to be looping without
3484 * reclaiming anything.
3486 if (!nr_reclaimed &&
3488 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3490 } while (!nr_reclaimed);
3492 css_put(&next_mz->memcg->css);
3493 return nr_reclaimed;
3497 * Reclaims as many pages from the given memcg as possible.
3499 * Caller is responsible for holding css reference for memcg.
3501 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3503 int nr_retries = MAX_RECLAIM_RETRIES;
3505 /* we call try-to-free pages for make this cgroup empty */
3506 lru_add_drain_all();
3508 drain_all_stock(memcg);
3510 /* try to free all pages in this cgroup */
3511 while (nr_retries && page_counter_read(&memcg->memory)) {
3512 if (signal_pending(current))
3515 if (!try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true))
3522 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3523 char *buf, size_t nbytes,
3526 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3528 if (mem_cgroup_is_root(memcg))
3530 return mem_cgroup_force_empty(memcg) ?: nbytes;
3533 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3539 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3540 struct cftype *cft, u64 val)
3545 pr_warn_once("Non-hierarchical mode is deprecated. "
3546 "Please report your usecase to linux-mm@kvack.org if you "
3547 "depend on this functionality.\n");
3552 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3556 if (mem_cgroup_is_root(memcg)) {
3557 mem_cgroup_flush_stats();
3558 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3559 memcg_page_state(memcg, NR_ANON_MAPPED);
3561 val += memcg_page_state(memcg, MEMCG_SWAP);
3564 val = page_counter_read(&memcg->memory);
3566 val = page_counter_read(&memcg->memsw);
3579 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3582 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3583 struct page_counter *counter;
3585 switch (MEMFILE_TYPE(cft->private)) {
3587 counter = &memcg->memory;
3590 counter = &memcg->memsw;
3593 counter = &memcg->kmem;
3596 counter = &memcg->tcpmem;
3602 switch (MEMFILE_ATTR(cft->private)) {
3604 if (counter == &memcg->memory)
3605 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3606 if (counter == &memcg->memsw)
3607 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3608 return (u64)page_counter_read(counter) * PAGE_SIZE;
3610 return (u64)counter->max * PAGE_SIZE;
3612 return (u64)counter->watermark * PAGE_SIZE;
3614 return counter->failcnt;
3615 case RES_SOFT_LIMIT:
3616 return (u64)memcg->soft_limit * PAGE_SIZE;
3622 #ifdef CONFIG_MEMCG_KMEM
3623 static int memcg_online_kmem(struct mem_cgroup *memcg)
3625 struct obj_cgroup *objcg;
3627 if (cgroup_memory_nokmem)
3630 if (unlikely(mem_cgroup_is_root(memcg)))
3633 objcg = obj_cgroup_alloc();
3637 objcg->memcg = memcg;
3638 rcu_assign_pointer(memcg->objcg, objcg);
3640 static_branch_enable(&memcg_kmem_enabled_key);
3642 memcg->kmemcg_id = memcg->id.id;
3647 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3649 struct mem_cgroup *parent;
3651 if (cgroup_memory_nokmem)
3654 if (unlikely(mem_cgroup_is_root(memcg)))
3657 parent = parent_mem_cgroup(memcg);
3659 parent = root_mem_cgroup;
3661 memcg_reparent_objcgs(memcg, parent);
3664 * After we have finished memcg_reparent_objcgs(), all list_lrus
3665 * corresponding to this cgroup are guaranteed to remain empty.
3666 * The ordering is imposed by list_lru_node->lock taken by
3667 * memcg_reparent_list_lrus().
3669 memcg_reparent_list_lrus(memcg, parent);
3672 static int memcg_online_kmem(struct mem_cgroup *memcg)
3676 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3679 #endif /* CONFIG_MEMCG_KMEM */
3681 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3685 mutex_lock(&memcg_max_mutex);
3687 ret = page_counter_set_max(&memcg->tcpmem, max);
3691 if (!memcg->tcpmem_active) {
3693 * The active flag needs to be written after the static_key
3694 * update. This is what guarantees that the socket activation
3695 * function is the last one to run. See mem_cgroup_sk_alloc()
3696 * for details, and note that we don't mark any socket as
3697 * belonging to this memcg until that flag is up.
3699 * We need to do this, because static_keys will span multiple
3700 * sites, but we can't control their order. If we mark a socket
3701 * as accounted, but the accounting functions are not patched in
3702 * yet, we'll lose accounting.
3704 * We never race with the readers in mem_cgroup_sk_alloc(),
3705 * because when this value change, the code to process it is not
3708 static_branch_inc(&memcg_sockets_enabled_key);
3709 memcg->tcpmem_active = true;
3712 mutex_unlock(&memcg_max_mutex);
3717 * The user of this function is...
3720 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3721 char *buf, size_t nbytes, loff_t off)
3723 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3724 unsigned long nr_pages;
3727 buf = strstrip(buf);
3728 ret = page_counter_memparse(buf, "-1", &nr_pages);
3732 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3734 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3738 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3740 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3743 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3746 /* kmem.limit_in_bytes is deprecated. */
3750 ret = memcg_update_tcp_max(memcg, nr_pages);
3754 case RES_SOFT_LIMIT:
3755 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
3758 memcg->soft_limit = nr_pages;
3763 return ret ?: nbytes;
3766 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3767 size_t nbytes, loff_t off)
3769 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3770 struct page_counter *counter;
3772 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3774 counter = &memcg->memory;
3777 counter = &memcg->memsw;
3780 counter = &memcg->kmem;
3783 counter = &memcg->tcpmem;
3789 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3791 page_counter_reset_watermark(counter);
3794 counter->failcnt = 0;
3803 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3806 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3810 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3811 struct cftype *cft, u64 val)
3813 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3815 if (val & ~MOVE_MASK)
3819 * No kind of locking is needed in here, because ->can_attach() will
3820 * check this value once in the beginning of the process, and then carry
3821 * on with stale data. This means that changes to this value will only
3822 * affect task migrations starting after the change.
3824 memcg->move_charge_at_immigrate = val;
3828 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3829 struct cftype *cft, u64 val)
3837 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3838 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3839 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3841 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3842 int nid, unsigned int lru_mask, bool tree)
3844 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3845 unsigned long nr = 0;
3848 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3851 if (!(BIT(lru) & lru_mask))
3854 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3856 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3861 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3862 unsigned int lru_mask,
3865 unsigned long nr = 0;
3869 if (!(BIT(lru) & lru_mask))
3872 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3874 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3879 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3883 unsigned int lru_mask;
3886 static const struct numa_stat stats[] = {
3887 { "total", LRU_ALL },
3888 { "file", LRU_ALL_FILE },
3889 { "anon", LRU_ALL_ANON },
3890 { "unevictable", BIT(LRU_UNEVICTABLE) },
3892 const struct numa_stat *stat;
3894 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3896 mem_cgroup_flush_stats();
3898 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3899 seq_printf(m, "%s=%lu", stat->name,
3900 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3902 for_each_node_state(nid, N_MEMORY)
3903 seq_printf(m, " N%d=%lu", nid,
3904 mem_cgroup_node_nr_lru_pages(memcg, nid,
3905 stat->lru_mask, false));
3909 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3911 seq_printf(m, "hierarchical_%s=%lu", stat->name,
3912 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3914 for_each_node_state(nid, N_MEMORY)
3915 seq_printf(m, " N%d=%lu", nid,
3916 mem_cgroup_node_nr_lru_pages(memcg, nid,
3917 stat->lru_mask, true));
3923 #endif /* CONFIG_NUMA */
3925 static const unsigned int memcg1_stats[] = {
3928 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3938 static const char *const memcg1_stat_names[] = {
3941 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3951 /* Universal VM events cgroup1 shows, original sort order */
3952 static const unsigned int memcg1_events[] = {
3959 static int memcg_stat_show(struct seq_file *m, void *v)
3961 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3962 unsigned long memory, memsw;
3963 struct mem_cgroup *mi;
3966 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3968 mem_cgroup_flush_stats();
3970 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3973 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3975 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
3976 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
3979 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3980 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
3981 memcg_events_local(memcg, memcg1_events[i]));
3983 for (i = 0; i < NR_LRU_LISTS; i++)
3984 seq_printf(m, "%s %lu\n", lru_list_name(i),
3985 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3988 /* Hierarchical information */
3989 memory = memsw = PAGE_COUNTER_MAX;
3990 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3991 memory = min(memory, READ_ONCE(mi->memory.max));
3992 memsw = min(memsw, READ_ONCE(mi->memsw.max));
3994 seq_printf(m, "hierarchical_memory_limit %llu\n",
3995 (u64)memory * PAGE_SIZE);
3996 if (do_memsw_account())
3997 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3998 (u64)memsw * PAGE_SIZE);
4000 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4003 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4005 nr = memcg_page_state(memcg, memcg1_stats[i]);
4006 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4007 (u64)nr * PAGE_SIZE);
4010 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4011 seq_printf(m, "total_%s %llu\n",
4012 vm_event_name(memcg1_events[i]),
4013 (u64)memcg_events(memcg, memcg1_events[i]));
4015 for (i = 0; i < NR_LRU_LISTS; i++)
4016 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4017 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4020 #ifdef CONFIG_DEBUG_VM
4023 struct mem_cgroup_per_node *mz;
4024 unsigned long anon_cost = 0;
4025 unsigned long file_cost = 0;
4027 for_each_online_pgdat(pgdat) {
4028 mz = memcg->nodeinfo[pgdat->node_id];
4030 anon_cost += mz->lruvec.anon_cost;
4031 file_cost += mz->lruvec.file_cost;
4033 seq_printf(m, "anon_cost %lu\n", anon_cost);
4034 seq_printf(m, "file_cost %lu\n", file_cost);
4041 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4044 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4046 return mem_cgroup_swappiness(memcg);
4049 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4050 struct cftype *cft, u64 val)
4052 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4057 if (!mem_cgroup_is_root(memcg))
4058 memcg->swappiness = val;
4060 vm_swappiness = val;
4065 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4067 struct mem_cgroup_threshold_ary *t;
4068 unsigned long usage;
4073 t = rcu_dereference(memcg->thresholds.primary);
4075 t = rcu_dereference(memcg->memsw_thresholds.primary);
4080 usage = mem_cgroup_usage(memcg, swap);
4083 * current_threshold points to threshold just below or equal to usage.
4084 * If it's not true, a threshold was crossed after last
4085 * call of __mem_cgroup_threshold().
4087 i = t->current_threshold;
4090 * Iterate backward over array of thresholds starting from
4091 * current_threshold and check if a threshold is crossed.
4092 * If none of thresholds below usage is crossed, we read
4093 * only one element of the array here.
4095 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4096 eventfd_signal(t->entries[i].eventfd, 1);
4098 /* i = current_threshold + 1 */
4102 * Iterate forward over array of thresholds starting from
4103 * current_threshold+1 and check if a threshold is crossed.
4104 * If none of thresholds above usage is crossed, we read
4105 * only one element of the array here.
4107 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4108 eventfd_signal(t->entries[i].eventfd, 1);
4110 /* Update current_threshold */
4111 t->current_threshold = i - 1;
4116 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4119 __mem_cgroup_threshold(memcg, false);
4120 if (do_memsw_account())
4121 __mem_cgroup_threshold(memcg, true);
4123 memcg = parent_mem_cgroup(memcg);
4127 static int compare_thresholds(const void *a, const void *b)
4129 const struct mem_cgroup_threshold *_a = a;
4130 const struct mem_cgroup_threshold *_b = b;
4132 if (_a->threshold > _b->threshold)
4135 if (_a->threshold < _b->threshold)
4141 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4143 struct mem_cgroup_eventfd_list *ev;
4145 spin_lock(&memcg_oom_lock);
4147 list_for_each_entry(ev, &memcg->oom_notify, list)
4148 eventfd_signal(ev->eventfd, 1);
4150 spin_unlock(&memcg_oom_lock);
4154 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4156 struct mem_cgroup *iter;
4158 for_each_mem_cgroup_tree(iter, memcg)
4159 mem_cgroup_oom_notify_cb(iter);
4162 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4163 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4165 struct mem_cgroup_thresholds *thresholds;
4166 struct mem_cgroup_threshold_ary *new;
4167 unsigned long threshold;
4168 unsigned long usage;
4171 ret = page_counter_memparse(args, "-1", &threshold);
4175 mutex_lock(&memcg->thresholds_lock);
4178 thresholds = &memcg->thresholds;
4179 usage = mem_cgroup_usage(memcg, false);
4180 } else if (type == _MEMSWAP) {
4181 thresholds = &memcg->memsw_thresholds;
4182 usage = mem_cgroup_usage(memcg, true);
4186 /* Check if a threshold crossed before adding a new one */
4187 if (thresholds->primary)
4188 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4190 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4192 /* Allocate memory for new array of thresholds */
4193 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4200 /* Copy thresholds (if any) to new array */
4201 if (thresholds->primary)
4202 memcpy(new->entries, thresholds->primary->entries,
4203 flex_array_size(new, entries, size - 1));
4205 /* Add new threshold */
4206 new->entries[size - 1].eventfd = eventfd;
4207 new->entries[size - 1].threshold = threshold;
4209 /* Sort thresholds. Registering of new threshold isn't time-critical */
4210 sort(new->entries, size, sizeof(*new->entries),
4211 compare_thresholds, NULL);
4213 /* Find current threshold */
4214 new->current_threshold = -1;
4215 for (i = 0; i < size; i++) {
4216 if (new->entries[i].threshold <= usage) {
4218 * new->current_threshold will not be used until
4219 * rcu_assign_pointer(), so it's safe to increment
4222 ++new->current_threshold;
4227 /* Free old spare buffer and save old primary buffer as spare */
4228 kfree(thresholds->spare);
4229 thresholds->spare = thresholds->primary;
4231 rcu_assign_pointer(thresholds->primary, new);
4233 /* To be sure that nobody uses thresholds */
4237 mutex_unlock(&memcg->thresholds_lock);
4242 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4243 struct eventfd_ctx *eventfd, const char *args)
4245 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4248 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4249 struct eventfd_ctx *eventfd, const char *args)
4251 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4254 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4255 struct eventfd_ctx *eventfd, enum res_type type)
4257 struct mem_cgroup_thresholds *thresholds;
4258 struct mem_cgroup_threshold_ary *new;
4259 unsigned long usage;
4260 int i, j, size, entries;
4262 mutex_lock(&memcg->thresholds_lock);
4265 thresholds = &memcg->thresholds;
4266 usage = mem_cgroup_usage(memcg, false);
4267 } else if (type == _MEMSWAP) {
4268 thresholds = &memcg->memsw_thresholds;
4269 usage = mem_cgroup_usage(memcg, true);
4273 if (!thresholds->primary)
4276 /* Check if a threshold crossed before removing */
4277 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4279 /* Calculate new number of threshold */
4281 for (i = 0; i < thresholds->primary->size; i++) {
4282 if (thresholds->primary->entries[i].eventfd != eventfd)
4288 new = thresholds->spare;
4290 /* If no items related to eventfd have been cleared, nothing to do */
4294 /* Set thresholds array to NULL if we don't have thresholds */
4303 /* Copy thresholds and find current threshold */
4304 new->current_threshold = -1;
4305 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4306 if (thresholds->primary->entries[i].eventfd == eventfd)
4309 new->entries[j] = thresholds->primary->entries[i];
4310 if (new->entries[j].threshold <= usage) {
4312 * new->current_threshold will not be used
4313 * until rcu_assign_pointer(), so it's safe to increment
4316 ++new->current_threshold;
4322 /* Swap primary and spare array */
4323 thresholds->spare = thresholds->primary;
4325 rcu_assign_pointer(thresholds->primary, new);
4327 /* To be sure that nobody uses thresholds */
4330 /* If all events are unregistered, free the spare array */
4332 kfree(thresholds->spare);
4333 thresholds->spare = NULL;
4336 mutex_unlock(&memcg->thresholds_lock);
4339 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4340 struct eventfd_ctx *eventfd)
4342 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4345 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4346 struct eventfd_ctx *eventfd)
4348 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4351 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4352 struct eventfd_ctx *eventfd, const char *args)
4354 struct mem_cgroup_eventfd_list *event;
4356 event = kmalloc(sizeof(*event), GFP_KERNEL);
4360 spin_lock(&memcg_oom_lock);
4362 event->eventfd = eventfd;
4363 list_add(&event->list, &memcg->oom_notify);
4365 /* already in OOM ? */
4366 if (memcg->under_oom)
4367 eventfd_signal(eventfd, 1);
4368 spin_unlock(&memcg_oom_lock);
4373 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4374 struct eventfd_ctx *eventfd)
4376 struct mem_cgroup_eventfd_list *ev, *tmp;
4378 spin_lock(&memcg_oom_lock);
4380 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4381 if (ev->eventfd == eventfd) {
4382 list_del(&ev->list);
4387 spin_unlock(&memcg_oom_lock);
4390 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4392 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4394 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4395 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4396 seq_printf(sf, "oom_kill %lu\n",
4397 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4401 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4402 struct cftype *cft, u64 val)
4404 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4406 /* cannot set to root cgroup and only 0 and 1 are allowed */
4407 if (mem_cgroup_is_root(memcg) || !((val == 0) || (val == 1)))
4410 memcg->oom_kill_disable = val;
4412 memcg_oom_recover(memcg);
4417 #ifdef CONFIG_CGROUP_WRITEBACK
4419 #include <trace/events/writeback.h>
4421 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4423 return wb_domain_init(&memcg->cgwb_domain, gfp);
4426 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4428 wb_domain_exit(&memcg->cgwb_domain);
4431 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4433 wb_domain_size_changed(&memcg->cgwb_domain);
4436 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4438 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4440 if (!memcg->css.parent)
4443 return &memcg->cgwb_domain;
4447 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4448 * @wb: bdi_writeback in question
4449 * @pfilepages: out parameter for number of file pages
4450 * @pheadroom: out parameter for number of allocatable pages according to memcg
4451 * @pdirty: out parameter for number of dirty pages
4452 * @pwriteback: out parameter for number of pages under writeback
4454 * Determine the numbers of file, headroom, dirty, and writeback pages in
4455 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4456 * is a bit more involved.
4458 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4459 * headroom is calculated as the lowest headroom of itself and the
4460 * ancestors. Note that this doesn't consider the actual amount of
4461 * available memory in the system. The caller should further cap
4462 * *@pheadroom accordingly.
4464 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4465 unsigned long *pheadroom, unsigned long *pdirty,
4466 unsigned long *pwriteback)
4468 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4469 struct mem_cgroup *parent;
4471 mem_cgroup_flush_stats();
4473 *pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
4474 *pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
4475 *pfilepages = memcg_page_state(memcg, NR_INACTIVE_FILE) +
4476 memcg_page_state(memcg, NR_ACTIVE_FILE);
4478 *pheadroom = PAGE_COUNTER_MAX;
4479 while ((parent = parent_mem_cgroup(memcg))) {
4480 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4481 READ_ONCE(memcg->memory.high));
4482 unsigned long used = page_counter_read(&memcg->memory);
4484 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4490 * Foreign dirty flushing
4492 * There's an inherent mismatch between memcg and writeback. The former
4493 * tracks ownership per-page while the latter per-inode. This was a
4494 * deliberate design decision because honoring per-page ownership in the
4495 * writeback path is complicated, may lead to higher CPU and IO overheads
4496 * and deemed unnecessary given that write-sharing an inode across
4497 * different cgroups isn't a common use-case.
4499 * Combined with inode majority-writer ownership switching, this works well
4500 * enough in most cases but there are some pathological cases. For
4501 * example, let's say there are two cgroups A and B which keep writing to
4502 * different but confined parts of the same inode. B owns the inode and
4503 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4504 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4505 * triggering background writeback. A will be slowed down without a way to
4506 * make writeback of the dirty pages happen.
4508 * Conditions like the above can lead to a cgroup getting repeatedly and
4509 * severely throttled after making some progress after each
4510 * dirty_expire_interval while the underlying IO device is almost
4513 * Solving this problem completely requires matching the ownership tracking
4514 * granularities between memcg and writeback in either direction. However,
4515 * the more egregious behaviors can be avoided by simply remembering the
4516 * most recent foreign dirtying events and initiating remote flushes on
4517 * them when local writeback isn't enough to keep the memory clean enough.
4519 * The following two functions implement such mechanism. When a foreign
4520 * page - a page whose memcg and writeback ownerships don't match - is
4521 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4522 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4523 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4524 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4525 * foreign bdi_writebacks which haven't expired. Both the numbers of
4526 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4527 * limited to MEMCG_CGWB_FRN_CNT.
4529 * The mechanism only remembers IDs and doesn't hold any object references.
4530 * As being wrong occasionally doesn't matter, updates and accesses to the
4531 * records are lockless and racy.
4533 void mem_cgroup_track_foreign_dirty_slowpath(struct folio *folio,
4534 struct bdi_writeback *wb)
4536 struct mem_cgroup *memcg = folio_memcg(folio);
4537 struct memcg_cgwb_frn *frn;
4538 u64 now = get_jiffies_64();
4539 u64 oldest_at = now;
4543 trace_track_foreign_dirty(folio, wb);
4546 * Pick the slot to use. If there is already a slot for @wb, keep
4547 * using it. If not replace the oldest one which isn't being
4550 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4551 frn = &memcg->cgwb_frn[i];
4552 if (frn->bdi_id == wb->bdi->id &&
4553 frn->memcg_id == wb->memcg_css->id)
4555 if (time_before64(frn->at, oldest_at) &&
4556 atomic_read(&frn->done.cnt) == 1) {
4558 oldest_at = frn->at;
4562 if (i < MEMCG_CGWB_FRN_CNT) {
4564 * Re-using an existing one. Update timestamp lazily to
4565 * avoid making the cacheline hot. We want them to be
4566 * reasonably up-to-date and significantly shorter than
4567 * dirty_expire_interval as that's what expires the record.
4568 * Use the shorter of 1s and dirty_expire_interval / 8.
4570 unsigned long update_intv =
4571 min_t(unsigned long, HZ,
4572 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4574 if (time_before64(frn->at, now - update_intv))
4576 } else if (oldest >= 0) {
4577 /* replace the oldest free one */
4578 frn = &memcg->cgwb_frn[oldest];
4579 frn->bdi_id = wb->bdi->id;
4580 frn->memcg_id = wb->memcg_css->id;
4585 /* issue foreign writeback flushes for recorded foreign dirtying events */
4586 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4588 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4589 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4590 u64 now = jiffies_64;
4593 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4594 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4597 * If the record is older than dirty_expire_interval,
4598 * writeback on it has already started. No need to kick it
4599 * off again. Also, don't start a new one if there's
4600 * already one in flight.
4602 if (time_after64(frn->at, now - intv) &&
4603 atomic_read(&frn->done.cnt) == 1) {
4605 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4606 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id,
4607 WB_REASON_FOREIGN_FLUSH,
4613 #else /* CONFIG_CGROUP_WRITEBACK */
4615 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4620 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4624 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4628 #endif /* CONFIG_CGROUP_WRITEBACK */
4631 * DO NOT USE IN NEW FILES.
4633 * "cgroup.event_control" implementation.
4635 * This is way over-engineered. It tries to support fully configurable
4636 * events for each user. Such level of flexibility is completely
4637 * unnecessary especially in the light of the planned unified hierarchy.
4639 * Please deprecate this and replace with something simpler if at all
4644 * Unregister event and free resources.
4646 * Gets called from workqueue.
4648 static void memcg_event_remove(struct work_struct *work)
4650 struct mem_cgroup_event *event =
4651 container_of(work, struct mem_cgroup_event, remove);
4652 struct mem_cgroup *memcg = event->memcg;
4654 remove_wait_queue(event->wqh, &event->wait);
4656 event->unregister_event(memcg, event->eventfd);
4658 /* Notify userspace the event is going away. */
4659 eventfd_signal(event->eventfd, 1);
4661 eventfd_ctx_put(event->eventfd);
4663 css_put(&memcg->css);
4667 * Gets called on EPOLLHUP on eventfd when user closes it.
4669 * Called with wqh->lock held and interrupts disabled.
4671 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4672 int sync, void *key)
4674 struct mem_cgroup_event *event =
4675 container_of(wait, struct mem_cgroup_event, wait);
4676 struct mem_cgroup *memcg = event->memcg;
4677 __poll_t flags = key_to_poll(key);
4679 if (flags & EPOLLHUP) {
4681 * If the event has been detached at cgroup removal, we
4682 * can simply return knowing the other side will cleanup
4685 * We can't race against event freeing since the other
4686 * side will require wqh->lock via remove_wait_queue(),
4689 spin_lock(&memcg->event_list_lock);
4690 if (!list_empty(&event->list)) {
4691 list_del_init(&event->list);
4693 * We are in atomic context, but cgroup_event_remove()
4694 * may sleep, so we have to call it in workqueue.
4696 schedule_work(&event->remove);
4698 spin_unlock(&memcg->event_list_lock);
4704 static void memcg_event_ptable_queue_proc(struct file *file,
4705 wait_queue_head_t *wqh, poll_table *pt)
4707 struct mem_cgroup_event *event =
4708 container_of(pt, struct mem_cgroup_event, pt);
4711 add_wait_queue(wqh, &event->wait);
4715 * DO NOT USE IN NEW FILES.
4717 * Parse input and register new cgroup event handler.
4719 * Input must be in format '<event_fd> <control_fd> <args>'.
4720 * Interpretation of args is defined by control file implementation.
4722 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4723 char *buf, size_t nbytes, loff_t off)
4725 struct cgroup_subsys_state *css = of_css(of);
4726 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4727 struct mem_cgroup_event *event;
4728 struct cgroup_subsys_state *cfile_css;
4729 unsigned int efd, cfd;
4736 if (IS_ENABLED(CONFIG_PREEMPT_RT))
4739 buf = strstrip(buf);
4741 efd = simple_strtoul(buf, &endp, 10);
4746 cfd = simple_strtoul(buf, &endp, 10);
4747 if ((*endp != ' ') && (*endp != '\0'))
4751 event = kzalloc(sizeof(*event), GFP_KERNEL);
4755 event->memcg = memcg;
4756 INIT_LIST_HEAD(&event->list);
4757 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4758 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4759 INIT_WORK(&event->remove, memcg_event_remove);
4767 event->eventfd = eventfd_ctx_fileget(efile.file);
4768 if (IS_ERR(event->eventfd)) {
4769 ret = PTR_ERR(event->eventfd);
4776 goto out_put_eventfd;
4779 /* the process need read permission on control file */
4780 /* AV: shouldn't we check that it's been opened for read instead? */
4781 ret = file_permission(cfile.file, MAY_READ);
4786 * Determine the event callbacks and set them in @event. This used
4787 * to be done via struct cftype but cgroup core no longer knows
4788 * about these events. The following is crude but the whole thing
4789 * is for compatibility anyway.
4791 * DO NOT ADD NEW FILES.
4793 name = cfile.file->f_path.dentry->d_name.name;
4795 if (!strcmp(name, "memory.usage_in_bytes")) {
4796 event->register_event = mem_cgroup_usage_register_event;
4797 event->unregister_event = mem_cgroup_usage_unregister_event;
4798 } else if (!strcmp(name, "memory.oom_control")) {
4799 event->register_event = mem_cgroup_oom_register_event;
4800 event->unregister_event = mem_cgroup_oom_unregister_event;
4801 } else if (!strcmp(name, "memory.pressure_level")) {
4802 event->register_event = vmpressure_register_event;
4803 event->unregister_event = vmpressure_unregister_event;
4804 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4805 event->register_event = memsw_cgroup_usage_register_event;
4806 event->unregister_event = memsw_cgroup_usage_unregister_event;
4813 * Verify @cfile should belong to @css. Also, remaining events are
4814 * automatically removed on cgroup destruction but the removal is
4815 * asynchronous, so take an extra ref on @css.
4817 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4818 &memory_cgrp_subsys);
4820 if (IS_ERR(cfile_css))
4822 if (cfile_css != css) {
4827 ret = event->register_event(memcg, event->eventfd, buf);
4831 vfs_poll(efile.file, &event->pt);
4833 spin_lock_irq(&memcg->event_list_lock);
4834 list_add(&event->list, &memcg->event_list);
4835 spin_unlock_irq(&memcg->event_list_lock);
4847 eventfd_ctx_put(event->eventfd);
4856 #if defined(CONFIG_MEMCG_KMEM) && (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
4857 static int mem_cgroup_slab_show(struct seq_file *m, void *p)
4861 * Please, take a look at tools/cgroup/slabinfo.py .
4867 static struct cftype mem_cgroup_legacy_files[] = {
4869 .name = "usage_in_bytes",
4870 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4871 .read_u64 = mem_cgroup_read_u64,
4874 .name = "max_usage_in_bytes",
4875 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4876 .write = mem_cgroup_reset,
4877 .read_u64 = mem_cgroup_read_u64,
4880 .name = "limit_in_bytes",
4881 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4882 .write = mem_cgroup_write,
4883 .read_u64 = mem_cgroup_read_u64,
4886 .name = "soft_limit_in_bytes",
4887 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4888 .write = mem_cgroup_write,
4889 .read_u64 = mem_cgroup_read_u64,
4893 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4894 .write = mem_cgroup_reset,
4895 .read_u64 = mem_cgroup_read_u64,
4899 .seq_show = memcg_stat_show,
4902 .name = "force_empty",
4903 .write = mem_cgroup_force_empty_write,
4906 .name = "use_hierarchy",
4907 .write_u64 = mem_cgroup_hierarchy_write,
4908 .read_u64 = mem_cgroup_hierarchy_read,
4911 .name = "cgroup.event_control", /* XXX: for compat */
4912 .write = memcg_write_event_control,
4913 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4916 .name = "swappiness",
4917 .read_u64 = mem_cgroup_swappiness_read,
4918 .write_u64 = mem_cgroup_swappiness_write,
4921 .name = "move_charge_at_immigrate",
4922 .read_u64 = mem_cgroup_move_charge_read,
4923 .write_u64 = mem_cgroup_move_charge_write,
4926 .name = "oom_control",
4927 .seq_show = mem_cgroup_oom_control_read,
4928 .write_u64 = mem_cgroup_oom_control_write,
4931 .name = "pressure_level",
4935 .name = "numa_stat",
4936 .seq_show = memcg_numa_stat_show,
4940 .name = "kmem.limit_in_bytes",
4941 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4942 .write = mem_cgroup_write,
4943 .read_u64 = mem_cgroup_read_u64,
4946 .name = "kmem.usage_in_bytes",
4947 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4948 .read_u64 = mem_cgroup_read_u64,
4951 .name = "kmem.failcnt",
4952 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4953 .write = mem_cgroup_reset,
4954 .read_u64 = mem_cgroup_read_u64,
4957 .name = "kmem.max_usage_in_bytes",
4958 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4959 .write = mem_cgroup_reset,
4960 .read_u64 = mem_cgroup_read_u64,
4962 #if defined(CONFIG_MEMCG_KMEM) && \
4963 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
4965 .name = "kmem.slabinfo",
4966 .seq_show = mem_cgroup_slab_show,
4970 .name = "kmem.tcp.limit_in_bytes",
4971 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4972 .write = mem_cgroup_write,
4973 .read_u64 = mem_cgroup_read_u64,
4976 .name = "kmem.tcp.usage_in_bytes",
4977 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4978 .read_u64 = mem_cgroup_read_u64,
4981 .name = "kmem.tcp.failcnt",
4982 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4983 .write = mem_cgroup_reset,
4984 .read_u64 = mem_cgroup_read_u64,
4987 .name = "kmem.tcp.max_usage_in_bytes",
4988 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4989 .write = mem_cgroup_reset,
4990 .read_u64 = mem_cgroup_read_u64,
4992 { }, /* terminate */
4996 * Private memory cgroup IDR
4998 * Swap-out records and page cache shadow entries need to store memcg
4999 * references in constrained space, so we maintain an ID space that is
5000 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5001 * memory-controlled cgroups to 64k.
5003 * However, there usually are many references to the offline CSS after
5004 * the cgroup has been destroyed, such as page cache or reclaimable
5005 * slab objects, that don't need to hang on to the ID. We want to keep
5006 * those dead CSS from occupying IDs, or we might quickly exhaust the
5007 * relatively small ID space and prevent the creation of new cgroups
5008 * even when there are much fewer than 64k cgroups - possibly none.
5010 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5011 * be freed and recycled when it's no longer needed, which is usually
5012 * when the CSS is offlined.
5014 * The only exception to that are records of swapped out tmpfs/shmem
5015 * pages that need to be attributed to live ancestors on swapin. But
5016 * those references are manageable from userspace.
5019 static DEFINE_IDR(mem_cgroup_idr);
5021 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5023 if (memcg->id.id > 0) {
5024 idr_remove(&mem_cgroup_idr, memcg->id.id);
5029 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5032 refcount_add(n, &memcg->id.ref);
5035 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5037 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5038 mem_cgroup_id_remove(memcg);
5040 /* Memcg ID pins CSS */
5041 css_put(&memcg->css);
5045 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5047 mem_cgroup_id_put_many(memcg, 1);
5051 * mem_cgroup_from_id - look up a memcg from a memcg id
5052 * @id: the memcg id to look up
5054 * Caller must hold rcu_read_lock().
5056 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5058 WARN_ON_ONCE(!rcu_read_lock_held());
5059 return idr_find(&mem_cgroup_idr, id);
5062 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5064 struct mem_cgroup_per_node *pn;
5066 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, node);
5070 pn->lruvec_stats_percpu = alloc_percpu_gfp(struct lruvec_stats_percpu,
5071 GFP_KERNEL_ACCOUNT);
5072 if (!pn->lruvec_stats_percpu) {
5077 lruvec_init(&pn->lruvec);
5080 memcg->nodeinfo[node] = pn;
5084 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5086 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5091 free_percpu(pn->lruvec_stats_percpu);
5095 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5100 free_mem_cgroup_per_node_info(memcg, node);
5101 free_percpu(memcg->vmstats_percpu);
5105 static void mem_cgroup_free(struct mem_cgroup *memcg)
5107 memcg_wb_domain_exit(memcg);
5108 __mem_cgroup_free(memcg);
5111 static struct mem_cgroup *mem_cgroup_alloc(void)
5113 struct mem_cgroup *memcg;
5115 int __maybe_unused i;
5116 long error = -ENOMEM;
5118 memcg = kzalloc(struct_size(memcg, nodeinfo, nr_node_ids), GFP_KERNEL);
5120 return ERR_PTR(error);
5122 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5123 1, MEM_CGROUP_ID_MAX + 1, GFP_KERNEL);
5124 if (memcg->id.id < 0) {
5125 error = memcg->id.id;
5129 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5130 GFP_KERNEL_ACCOUNT);
5131 if (!memcg->vmstats_percpu)
5135 if (alloc_mem_cgroup_per_node_info(memcg, node))
5138 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5141 INIT_WORK(&memcg->high_work, high_work_func);
5142 INIT_LIST_HEAD(&memcg->oom_notify);
5143 mutex_init(&memcg->thresholds_lock);
5144 spin_lock_init(&memcg->move_lock);
5145 vmpressure_init(&memcg->vmpressure);
5146 INIT_LIST_HEAD(&memcg->event_list);
5147 spin_lock_init(&memcg->event_list_lock);
5148 memcg->socket_pressure = jiffies;
5149 #ifdef CONFIG_MEMCG_KMEM
5150 memcg->kmemcg_id = -1;
5151 INIT_LIST_HEAD(&memcg->objcg_list);
5153 #ifdef CONFIG_CGROUP_WRITEBACK
5154 INIT_LIST_HEAD(&memcg->cgwb_list);
5155 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5156 memcg->cgwb_frn[i].done =
5157 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5159 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5160 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5161 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5162 memcg->deferred_split_queue.split_queue_len = 0;
5164 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5167 mem_cgroup_id_remove(memcg);
5168 __mem_cgroup_free(memcg);
5169 return ERR_PTR(error);
5172 static struct cgroup_subsys_state * __ref
5173 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5175 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5176 struct mem_cgroup *memcg, *old_memcg;
5178 old_memcg = set_active_memcg(parent);
5179 memcg = mem_cgroup_alloc();
5180 set_active_memcg(old_memcg);
5182 return ERR_CAST(memcg);
5184 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5185 memcg->soft_limit = PAGE_COUNTER_MAX;
5186 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
5187 memcg->zswap_max = PAGE_COUNTER_MAX;
5189 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5191 memcg->swappiness = mem_cgroup_swappiness(parent);
5192 memcg->oom_kill_disable = parent->oom_kill_disable;
5194 page_counter_init(&memcg->memory, &parent->memory);
5195 page_counter_init(&memcg->swap, &parent->swap);
5196 page_counter_init(&memcg->kmem, &parent->kmem);
5197 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5199 page_counter_init(&memcg->memory, NULL);
5200 page_counter_init(&memcg->swap, NULL);
5201 page_counter_init(&memcg->kmem, NULL);
5202 page_counter_init(&memcg->tcpmem, NULL);
5204 root_mem_cgroup = memcg;
5208 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5209 static_branch_inc(&memcg_sockets_enabled_key);
5214 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5216 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5218 if (memcg_online_kmem(memcg))
5222 * A memcg must be visible for expand_shrinker_info()
5223 * by the time the maps are allocated. So, we allocate maps
5224 * here, when for_each_mem_cgroup() can't skip it.
5226 if (alloc_shrinker_info(memcg))
5229 /* Online state pins memcg ID, memcg ID pins CSS */
5230 refcount_set(&memcg->id.ref, 1);
5233 if (unlikely(mem_cgroup_is_root(memcg)))
5234 queue_delayed_work(system_unbound_wq, &stats_flush_dwork,
5238 memcg_offline_kmem(memcg);
5240 mem_cgroup_id_remove(memcg);
5244 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5246 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5247 struct mem_cgroup_event *event, *tmp;
5250 * Unregister events and notify userspace.
5251 * Notify userspace about cgroup removing only after rmdir of cgroup
5252 * directory to avoid race between userspace and kernelspace.
5254 spin_lock_irq(&memcg->event_list_lock);
5255 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5256 list_del_init(&event->list);
5257 schedule_work(&event->remove);
5259 spin_unlock_irq(&memcg->event_list_lock);
5261 page_counter_set_min(&memcg->memory, 0);
5262 page_counter_set_low(&memcg->memory, 0);
5264 memcg_offline_kmem(memcg);
5265 reparent_shrinker_deferred(memcg);
5266 wb_memcg_offline(memcg);
5268 drain_all_stock(memcg);
5270 mem_cgroup_id_put(memcg);
5273 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5275 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5277 invalidate_reclaim_iterators(memcg);
5280 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5282 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5283 int __maybe_unused i;
5285 #ifdef CONFIG_CGROUP_WRITEBACK
5286 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5287 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5289 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5290 static_branch_dec(&memcg_sockets_enabled_key);
5292 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5293 static_branch_dec(&memcg_sockets_enabled_key);
5295 vmpressure_cleanup(&memcg->vmpressure);
5296 cancel_work_sync(&memcg->high_work);
5297 mem_cgroup_remove_from_trees(memcg);
5298 free_shrinker_info(memcg);
5299 mem_cgroup_free(memcg);
5303 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5304 * @css: the target css
5306 * Reset the states of the mem_cgroup associated with @css. This is
5307 * invoked when the userland requests disabling on the default hierarchy
5308 * but the memcg is pinned through dependency. The memcg should stop
5309 * applying policies and should revert to the vanilla state as it may be
5310 * made visible again.
5312 * The current implementation only resets the essential configurations.
5313 * This needs to be expanded to cover all the visible parts.
5315 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5317 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5319 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5320 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5321 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5322 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5323 page_counter_set_min(&memcg->memory, 0);
5324 page_counter_set_low(&memcg->memory, 0);
5325 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5326 memcg->soft_limit = PAGE_COUNTER_MAX;
5327 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5328 memcg_wb_domain_size_changed(memcg);
5331 static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state *css, int cpu)
5333 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5334 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5335 struct memcg_vmstats_percpu *statc;
5339 statc = per_cpu_ptr(memcg->vmstats_percpu, cpu);
5341 for (i = 0; i < MEMCG_NR_STAT; i++) {
5343 * Collect the aggregated propagation counts of groups
5344 * below us. We're in a per-cpu loop here and this is
5345 * a global counter, so the first cycle will get them.
5347 delta = memcg->vmstats.state_pending[i];
5349 memcg->vmstats.state_pending[i] = 0;
5351 /* Add CPU changes on this level since the last flush */
5352 v = READ_ONCE(statc->state[i]);
5353 if (v != statc->state_prev[i]) {
5354 delta += v - statc->state_prev[i];
5355 statc->state_prev[i] = v;
5361 /* Aggregate counts on this level and propagate upwards */
5362 memcg->vmstats.state[i] += delta;
5364 parent->vmstats.state_pending[i] += delta;
5367 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
5368 delta = memcg->vmstats.events_pending[i];
5370 memcg->vmstats.events_pending[i] = 0;
5372 v = READ_ONCE(statc->events[i]);
5373 if (v != statc->events_prev[i]) {
5374 delta += v - statc->events_prev[i];
5375 statc->events_prev[i] = v;
5381 memcg->vmstats.events[i] += delta;
5383 parent->vmstats.events_pending[i] += delta;
5386 for_each_node_state(nid, N_MEMORY) {
5387 struct mem_cgroup_per_node *pn = memcg->nodeinfo[nid];
5388 struct mem_cgroup_per_node *ppn = NULL;
5389 struct lruvec_stats_percpu *lstatc;
5392 ppn = parent->nodeinfo[nid];
5394 lstatc = per_cpu_ptr(pn->lruvec_stats_percpu, cpu);
5396 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
5397 delta = pn->lruvec_stats.state_pending[i];
5399 pn->lruvec_stats.state_pending[i] = 0;
5401 v = READ_ONCE(lstatc->state[i]);
5402 if (v != lstatc->state_prev[i]) {
5403 delta += v - lstatc->state_prev[i];
5404 lstatc->state_prev[i] = v;
5410 pn->lruvec_stats.state[i] += delta;
5412 ppn->lruvec_stats.state_pending[i] += delta;
5418 /* Handlers for move charge at task migration. */
5419 static int mem_cgroup_do_precharge(unsigned long count)
5423 /* Try a single bulk charge without reclaim first, kswapd may wake */
5424 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5426 mc.precharge += count;
5430 /* Try charges one by one with reclaim, but do not retry */
5432 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5446 enum mc_target_type {
5453 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5454 unsigned long addr, pte_t ptent)
5456 struct page *page = vm_normal_page(vma, addr, ptent);
5458 if (!page || !page_mapped(page))
5460 if (PageAnon(page)) {
5461 if (!(mc.flags & MOVE_ANON))
5464 if (!(mc.flags & MOVE_FILE))
5467 if (!get_page_unless_zero(page))
5473 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5474 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5475 pte_t ptent, swp_entry_t *entry)
5477 struct page *page = NULL;
5478 swp_entry_t ent = pte_to_swp_entry(ptent);
5480 if (!(mc.flags & MOVE_ANON))
5484 * Handle device private pages that are not accessible by the CPU, but
5485 * stored as special swap entries in the page table.
5487 if (is_device_private_entry(ent)) {
5488 page = pfn_swap_entry_to_page(ent);
5489 if (!get_page_unless_zero(page))
5494 if (non_swap_entry(ent))
5498 * Because lookup_swap_cache() updates some statistics counter,
5499 * we call find_get_page() with swapper_space directly.
5501 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5502 entry->val = ent.val;
5507 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5508 pte_t ptent, swp_entry_t *entry)
5514 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5515 unsigned long addr, pte_t ptent)
5517 if (!vma->vm_file) /* anonymous vma */
5519 if (!(mc.flags & MOVE_FILE))
5522 /* page is moved even if it's not RSS of this task(page-faulted). */
5523 /* shmem/tmpfs may report page out on swap: account for that too. */
5524 return find_get_incore_page(vma->vm_file->f_mapping,
5525 linear_page_index(vma, addr));
5529 * mem_cgroup_move_account - move account of the page
5531 * @compound: charge the page as compound or small page
5532 * @from: mem_cgroup which the page is moved from.
5533 * @to: mem_cgroup which the page is moved to. @from != @to.
5535 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5537 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5540 static int mem_cgroup_move_account(struct page *page,
5542 struct mem_cgroup *from,
5543 struct mem_cgroup *to)
5545 struct folio *folio = page_folio(page);
5546 struct lruvec *from_vec, *to_vec;
5547 struct pglist_data *pgdat;
5548 unsigned int nr_pages = compound ? folio_nr_pages(folio) : 1;
5551 VM_BUG_ON(from == to);
5552 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
5553 VM_BUG_ON(compound && !folio_test_large(folio));
5556 * Prevent mem_cgroup_migrate() from looking at
5557 * page's memory cgroup of its source page while we change it.
5560 if (!folio_trylock(folio))
5564 if (folio_memcg(folio) != from)
5567 pgdat = folio_pgdat(folio);
5568 from_vec = mem_cgroup_lruvec(from, pgdat);
5569 to_vec = mem_cgroup_lruvec(to, pgdat);
5571 folio_memcg_lock(folio);
5573 if (folio_test_anon(folio)) {
5574 if (folio_mapped(folio)) {
5575 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5576 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5577 if (folio_test_transhuge(folio)) {
5578 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5580 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5585 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5586 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5588 if (folio_test_swapbacked(folio)) {
5589 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5590 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5593 if (folio_mapped(folio)) {
5594 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5595 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5598 if (folio_test_dirty(folio)) {
5599 struct address_space *mapping = folio_mapping(folio);
5601 if (mapping_can_writeback(mapping)) {
5602 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5604 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5610 if (folio_test_writeback(folio)) {
5611 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5612 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5616 * All state has been migrated, let's switch to the new memcg.
5618 * It is safe to change page's memcg here because the page
5619 * is referenced, charged, isolated, and locked: we can't race
5620 * with (un)charging, migration, LRU putback, or anything else
5621 * that would rely on a stable page's memory cgroup.
5623 * Note that lock_page_memcg is a memcg lock, not a page lock,
5624 * to save space. As soon as we switch page's memory cgroup to a
5625 * new memcg that isn't locked, the above state can change
5626 * concurrently again. Make sure we're truly done with it.
5631 css_put(&from->css);
5633 folio->memcg_data = (unsigned long)to;
5635 __folio_memcg_unlock(from);
5638 nid = folio_nid(folio);
5640 local_irq_disable();
5641 mem_cgroup_charge_statistics(to, nr_pages);
5642 memcg_check_events(to, nid);
5643 mem_cgroup_charge_statistics(from, -nr_pages);
5644 memcg_check_events(from, nid);
5647 folio_unlock(folio);
5653 * get_mctgt_type - get target type of moving charge
5654 * @vma: the vma the pte to be checked belongs
5655 * @addr: the address corresponding to the pte to be checked
5656 * @ptent: the pte to be checked
5657 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5660 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5661 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5662 * move charge. if @target is not NULL, the page is stored in target->page
5663 * with extra refcnt got(Callers should handle it).
5664 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5665 * target for charge migration. if @target is not NULL, the entry is stored
5667 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5668 * (so ZONE_DEVICE page and thus not on the lru).
5669 * For now we such page is charge like a regular page would be as for all
5670 * intent and purposes it is just special memory taking the place of a
5673 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5675 * Called with pte lock held.
5678 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5679 unsigned long addr, pte_t ptent, union mc_target *target)
5681 struct page *page = NULL;
5682 enum mc_target_type ret = MC_TARGET_NONE;
5683 swp_entry_t ent = { .val = 0 };
5685 if (pte_present(ptent))
5686 page = mc_handle_present_pte(vma, addr, ptent);
5687 else if (pte_none_mostly(ptent))
5689 * PTE markers should be treated as a none pte here, separated
5690 * from other swap handling below.
5692 page = mc_handle_file_pte(vma, addr, ptent);
5693 else if (is_swap_pte(ptent))
5694 page = mc_handle_swap_pte(vma, ptent, &ent);
5696 if (!page && !ent.val)
5700 * Do only loose check w/o serialization.
5701 * mem_cgroup_move_account() checks the page is valid or
5702 * not under LRU exclusion.
5704 if (page_memcg(page) == mc.from) {
5705 ret = MC_TARGET_PAGE;
5706 if (is_device_private_page(page))
5707 ret = MC_TARGET_DEVICE;
5709 target->page = page;
5711 if (!ret || !target)
5715 * There is a swap entry and a page doesn't exist or isn't charged.
5716 * But we cannot move a tail-page in a THP.
5718 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5719 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5720 ret = MC_TARGET_SWAP;
5727 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5729 * We don't consider PMD mapped swapping or file mapped pages because THP does
5730 * not support them for now.
5731 * Caller should make sure that pmd_trans_huge(pmd) is true.
5733 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5734 unsigned long addr, pmd_t pmd, union mc_target *target)
5736 struct page *page = NULL;
5737 enum mc_target_type ret = MC_TARGET_NONE;
5739 if (unlikely(is_swap_pmd(pmd))) {
5740 VM_BUG_ON(thp_migration_supported() &&
5741 !is_pmd_migration_entry(pmd));
5744 page = pmd_page(pmd);
5745 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5746 if (!(mc.flags & MOVE_ANON))
5748 if (page_memcg(page) == mc.from) {
5749 ret = MC_TARGET_PAGE;
5752 target->page = page;
5758 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5759 unsigned long addr, pmd_t pmd, union mc_target *target)
5761 return MC_TARGET_NONE;
5765 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5766 unsigned long addr, unsigned long end,
5767 struct mm_walk *walk)
5769 struct vm_area_struct *vma = walk->vma;
5773 ptl = pmd_trans_huge_lock(pmd, vma);
5776 * Note their can not be MC_TARGET_DEVICE for now as we do not
5777 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5778 * this might change.
5780 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5781 mc.precharge += HPAGE_PMD_NR;
5786 if (pmd_trans_unstable(pmd))
5788 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5789 for (; addr != end; pte++, addr += PAGE_SIZE)
5790 if (get_mctgt_type(vma, addr, *pte, NULL))
5791 mc.precharge++; /* increment precharge temporarily */
5792 pte_unmap_unlock(pte - 1, ptl);
5798 static const struct mm_walk_ops precharge_walk_ops = {
5799 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5802 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5804 unsigned long precharge;
5807 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5808 mmap_read_unlock(mm);
5810 precharge = mc.precharge;
5816 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5818 unsigned long precharge = mem_cgroup_count_precharge(mm);
5820 VM_BUG_ON(mc.moving_task);
5821 mc.moving_task = current;
5822 return mem_cgroup_do_precharge(precharge);
5825 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5826 static void __mem_cgroup_clear_mc(void)
5828 struct mem_cgroup *from = mc.from;
5829 struct mem_cgroup *to = mc.to;
5831 /* we must uncharge all the leftover precharges from mc.to */
5833 cancel_charge(mc.to, mc.precharge);
5837 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5838 * we must uncharge here.
5840 if (mc.moved_charge) {
5841 cancel_charge(mc.from, mc.moved_charge);
5842 mc.moved_charge = 0;
5844 /* we must fixup refcnts and charges */
5845 if (mc.moved_swap) {
5846 /* uncharge swap account from the old cgroup */
5847 if (!mem_cgroup_is_root(mc.from))
5848 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5850 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5853 * we charged both to->memory and to->memsw, so we
5854 * should uncharge to->memory.
5856 if (!mem_cgroup_is_root(mc.to))
5857 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5861 memcg_oom_recover(from);
5862 memcg_oom_recover(to);
5863 wake_up_all(&mc.waitq);
5866 static void mem_cgroup_clear_mc(void)
5868 struct mm_struct *mm = mc.mm;
5871 * we must clear moving_task before waking up waiters at the end of
5874 mc.moving_task = NULL;
5875 __mem_cgroup_clear_mc();
5876 spin_lock(&mc.lock);
5880 spin_unlock(&mc.lock);
5885 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5887 struct cgroup_subsys_state *css;
5888 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5889 struct mem_cgroup *from;
5890 struct task_struct *leader, *p;
5891 struct mm_struct *mm;
5892 unsigned long move_flags;
5895 /* charge immigration isn't supported on the default hierarchy */
5896 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5900 * Multi-process migrations only happen on the default hierarchy
5901 * where charge immigration is not used. Perform charge
5902 * immigration if @tset contains a leader and whine if there are
5906 cgroup_taskset_for_each_leader(leader, css, tset) {
5909 memcg = mem_cgroup_from_css(css);
5915 * We are now committed to this value whatever it is. Changes in this
5916 * tunable will only affect upcoming migrations, not the current one.
5917 * So we need to save it, and keep it going.
5919 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5923 from = mem_cgroup_from_task(p);
5925 VM_BUG_ON(from == memcg);
5927 mm = get_task_mm(p);
5930 /* We move charges only when we move a owner of the mm */
5931 if (mm->owner == p) {
5934 VM_BUG_ON(mc.precharge);
5935 VM_BUG_ON(mc.moved_charge);
5936 VM_BUG_ON(mc.moved_swap);
5938 spin_lock(&mc.lock);
5942 mc.flags = move_flags;
5943 spin_unlock(&mc.lock);
5944 /* We set mc.moving_task later */
5946 ret = mem_cgroup_precharge_mc(mm);
5948 mem_cgroup_clear_mc();
5955 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5958 mem_cgroup_clear_mc();
5961 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5962 unsigned long addr, unsigned long end,
5963 struct mm_walk *walk)
5966 struct vm_area_struct *vma = walk->vma;
5969 enum mc_target_type target_type;
5970 union mc_target target;
5973 ptl = pmd_trans_huge_lock(pmd, vma);
5975 if (mc.precharge < HPAGE_PMD_NR) {
5979 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5980 if (target_type == MC_TARGET_PAGE) {
5982 if (!isolate_lru_page(page)) {
5983 if (!mem_cgroup_move_account(page, true,
5985 mc.precharge -= HPAGE_PMD_NR;
5986 mc.moved_charge += HPAGE_PMD_NR;
5988 putback_lru_page(page);
5991 } else if (target_type == MC_TARGET_DEVICE) {
5993 if (!mem_cgroup_move_account(page, true,
5995 mc.precharge -= HPAGE_PMD_NR;
5996 mc.moved_charge += HPAGE_PMD_NR;
6004 if (pmd_trans_unstable(pmd))
6007 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6008 for (; addr != end; addr += PAGE_SIZE) {
6009 pte_t ptent = *(pte++);
6010 bool device = false;
6016 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6017 case MC_TARGET_DEVICE:
6020 case MC_TARGET_PAGE:
6023 * We can have a part of the split pmd here. Moving it
6024 * can be done but it would be too convoluted so simply
6025 * ignore such a partial THP and keep it in original
6026 * memcg. There should be somebody mapping the head.
6028 if (PageTransCompound(page))
6030 if (!device && isolate_lru_page(page))
6032 if (!mem_cgroup_move_account(page, false,
6035 /* we uncharge from mc.from later. */
6039 putback_lru_page(page);
6040 put: /* get_mctgt_type() gets the page */
6043 case MC_TARGET_SWAP:
6045 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6047 mem_cgroup_id_get_many(mc.to, 1);
6048 /* we fixup other refcnts and charges later. */
6056 pte_unmap_unlock(pte - 1, ptl);
6061 * We have consumed all precharges we got in can_attach().
6062 * We try charge one by one, but don't do any additional
6063 * charges to mc.to if we have failed in charge once in attach()
6066 ret = mem_cgroup_do_precharge(1);
6074 static const struct mm_walk_ops charge_walk_ops = {
6075 .pmd_entry = mem_cgroup_move_charge_pte_range,
6078 static void mem_cgroup_move_charge(void)
6080 lru_add_drain_all();
6082 * Signal lock_page_memcg() to take the memcg's move_lock
6083 * while we're moving its pages to another memcg. Then wait
6084 * for already started RCU-only updates to finish.
6086 atomic_inc(&mc.from->moving_account);
6089 if (unlikely(!mmap_read_trylock(mc.mm))) {
6091 * Someone who are holding the mmap_lock might be waiting in
6092 * waitq. So we cancel all extra charges, wake up all waiters,
6093 * and retry. Because we cancel precharges, we might not be able
6094 * to move enough charges, but moving charge is a best-effort
6095 * feature anyway, so it wouldn't be a big problem.
6097 __mem_cgroup_clear_mc();
6102 * When we have consumed all precharges and failed in doing
6103 * additional charge, the page walk just aborts.
6105 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6108 mmap_read_unlock(mc.mm);
6109 atomic_dec(&mc.from->moving_account);
6112 static void mem_cgroup_move_task(void)
6115 mem_cgroup_move_charge();
6116 mem_cgroup_clear_mc();
6119 #else /* !CONFIG_MMU */
6120 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6124 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6127 static void mem_cgroup_move_task(void)
6132 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6134 if (value == PAGE_COUNTER_MAX)
6135 seq_puts(m, "max\n");
6137 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6142 static u64 memory_current_read(struct cgroup_subsys_state *css,
6145 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6147 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6150 static u64 memory_peak_read(struct cgroup_subsys_state *css,
6153 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6155 return (u64)memcg->memory.watermark * PAGE_SIZE;
6158 static int memory_min_show(struct seq_file *m, void *v)
6160 return seq_puts_memcg_tunable(m,
6161 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6164 static ssize_t memory_min_write(struct kernfs_open_file *of,
6165 char *buf, size_t nbytes, loff_t off)
6167 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6171 buf = strstrip(buf);
6172 err = page_counter_memparse(buf, "max", &min);
6176 page_counter_set_min(&memcg->memory, min);
6181 static int memory_low_show(struct seq_file *m, void *v)
6183 return seq_puts_memcg_tunable(m,
6184 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6187 static ssize_t memory_low_write(struct kernfs_open_file *of,
6188 char *buf, size_t nbytes, loff_t off)
6190 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6194 buf = strstrip(buf);
6195 err = page_counter_memparse(buf, "max", &low);
6199 page_counter_set_low(&memcg->memory, low);
6204 static int memory_high_show(struct seq_file *m, void *v)
6206 return seq_puts_memcg_tunable(m,
6207 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6210 static ssize_t memory_high_write(struct kernfs_open_file *of,
6211 char *buf, size_t nbytes, loff_t off)
6213 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6214 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6215 bool drained = false;
6219 buf = strstrip(buf);
6220 err = page_counter_memparse(buf, "max", &high);
6224 page_counter_set_high(&memcg->memory, high);
6227 unsigned long nr_pages = page_counter_read(&memcg->memory);
6228 unsigned long reclaimed;
6230 if (nr_pages <= high)
6233 if (signal_pending(current))
6237 drain_all_stock(memcg);
6242 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6245 if (!reclaimed && !nr_retries--)
6249 memcg_wb_domain_size_changed(memcg);
6253 static int memory_max_show(struct seq_file *m, void *v)
6255 return seq_puts_memcg_tunable(m,
6256 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6259 static ssize_t memory_max_write(struct kernfs_open_file *of,
6260 char *buf, size_t nbytes, loff_t off)
6262 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6263 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6264 bool drained = false;
6268 buf = strstrip(buf);
6269 err = page_counter_memparse(buf, "max", &max);
6273 xchg(&memcg->memory.max, max);
6276 unsigned long nr_pages = page_counter_read(&memcg->memory);
6278 if (nr_pages <= max)
6281 if (signal_pending(current))
6285 drain_all_stock(memcg);
6291 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6297 memcg_memory_event(memcg, MEMCG_OOM);
6298 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6302 memcg_wb_domain_size_changed(memcg);
6306 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6308 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6309 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6310 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6311 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6312 seq_printf(m, "oom_kill %lu\n",
6313 atomic_long_read(&events[MEMCG_OOM_KILL]));
6314 seq_printf(m, "oom_group_kill %lu\n",
6315 atomic_long_read(&events[MEMCG_OOM_GROUP_KILL]));
6318 static int memory_events_show(struct seq_file *m, void *v)
6320 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6322 __memory_events_show(m, memcg->memory_events);
6326 static int memory_events_local_show(struct seq_file *m, void *v)
6328 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6330 __memory_events_show(m, memcg->memory_events_local);
6334 static int memory_stat_show(struct seq_file *m, void *v)
6336 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6339 buf = memory_stat_format(memcg);
6348 static inline unsigned long lruvec_page_state_output(struct lruvec *lruvec,
6351 return lruvec_page_state(lruvec, item) * memcg_page_state_unit(item);
6354 static int memory_numa_stat_show(struct seq_file *m, void *v)
6357 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6359 mem_cgroup_flush_stats();
6361 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6364 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6367 seq_printf(m, "%s", memory_stats[i].name);
6368 for_each_node_state(nid, N_MEMORY) {
6370 struct lruvec *lruvec;
6372 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6373 size = lruvec_page_state_output(lruvec,
6374 memory_stats[i].idx);
6375 seq_printf(m, " N%d=%llu", nid, size);
6384 static int memory_oom_group_show(struct seq_file *m, void *v)
6386 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6388 seq_printf(m, "%d\n", memcg->oom_group);
6393 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6394 char *buf, size_t nbytes, loff_t off)
6396 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6399 buf = strstrip(buf);
6403 ret = kstrtoint(buf, 0, &oom_group);
6407 if (oom_group != 0 && oom_group != 1)
6410 memcg->oom_group = oom_group;
6415 static ssize_t memory_reclaim(struct kernfs_open_file *of, char *buf,
6416 size_t nbytes, loff_t off)
6418 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6419 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6420 unsigned long nr_to_reclaim, nr_reclaimed = 0;
6423 buf = strstrip(buf);
6424 err = page_counter_memparse(buf, "", &nr_to_reclaim);
6428 while (nr_reclaimed < nr_to_reclaim) {
6429 unsigned long reclaimed;
6431 if (signal_pending(current))
6435 * This is the final attempt, drain percpu lru caches in the
6436 * hope of introducing more evictable pages for
6437 * try_to_free_mem_cgroup_pages().
6440 lru_add_drain_all();
6442 reclaimed = try_to_free_mem_cgroup_pages(memcg,
6443 nr_to_reclaim - nr_reclaimed,
6446 if (!reclaimed && !nr_retries--)
6449 nr_reclaimed += reclaimed;
6455 static struct cftype memory_files[] = {
6458 .flags = CFTYPE_NOT_ON_ROOT,
6459 .read_u64 = memory_current_read,
6463 .flags = CFTYPE_NOT_ON_ROOT,
6464 .read_u64 = memory_peak_read,
6468 .flags = CFTYPE_NOT_ON_ROOT,
6469 .seq_show = memory_min_show,
6470 .write = memory_min_write,
6474 .flags = CFTYPE_NOT_ON_ROOT,
6475 .seq_show = memory_low_show,
6476 .write = memory_low_write,
6480 .flags = CFTYPE_NOT_ON_ROOT,
6481 .seq_show = memory_high_show,
6482 .write = memory_high_write,
6486 .flags = CFTYPE_NOT_ON_ROOT,
6487 .seq_show = memory_max_show,
6488 .write = memory_max_write,
6492 .flags = CFTYPE_NOT_ON_ROOT,
6493 .file_offset = offsetof(struct mem_cgroup, events_file),
6494 .seq_show = memory_events_show,
6497 .name = "events.local",
6498 .flags = CFTYPE_NOT_ON_ROOT,
6499 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6500 .seq_show = memory_events_local_show,
6504 .seq_show = memory_stat_show,
6508 .name = "numa_stat",
6509 .seq_show = memory_numa_stat_show,
6513 .name = "oom.group",
6514 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6515 .seq_show = memory_oom_group_show,
6516 .write = memory_oom_group_write,
6520 .flags = CFTYPE_NS_DELEGATABLE,
6521 .write = memory_reclaim,
6526 struct cgroup_subsys memory_cgrp_subsys = {
6527 .css_alloc = mem_cgroup_css_alloc,
6528 .css_online = mem_cgroup_css_online,
6529 .css_offline = mem_cgroup_css_offline,
6530 .css_released = mem_cgroup_css_released,
6531 .css_free = mem_cgroup_css_free,
6532 .css_reset = mem_cgroup_css_reset,
6533 .css_rstat_flush = mem_cgroup_css_rstat_flush,
6534 .can_attach = mem_cgroup_can_attach,
6535 .cancel_attach = mem_cgroup_cancel_attach,
6536 .post_attach = mem_cgroup_move_task,
6537 .dfl_cftypes = memory_files,
6538 .legacy_cftypes = mem_cgroup_legacy_files,
6543 * This function calculates an individual cgroup's effective
6544 * protection which is derived from its own memory.min/low, its
6545 * parent's and siblings' settings, as well as the actual memory
6546 * distribution in the tree.
6548 * The following rules apply to the effective protection values:
6550 * 1. At the first level of reclaim, effective protection is equal to
6551 * the declared protection in memory.min and memory.low.
6553 * 2. To enable safe delegation of the protection configuration, at
6554 * subsequent levels the effective protection is capped to the
6555 * parent's effective protection.
6557 * 3. To make complex and dynamic subtrees easier to configure, the
6558 * user is allowed to overcommit the declared protection at a given
6559 * level. If that is the case, the parent's effective protection is
6560 * distributed to the children in proportion to how much protection
6561 * they have declared and how much of it they are utilizing.
6563 * This makes distribution proportional, but also work-conserving:
6564 * if one cgroup claims much more protection than it uses memory,
6565 * the unused remainder is available to its siblings.
6567 * 4. Conversely, when the declared protection is undercommitted at a
6568 * given level, the distribution of the larger parental protection
6569 * budget is NOT proportional. A cgroup's protection from a sibling
6570 * is capped to its own memory.min/low setting.
6572 * 5. However, to allow protecting recursive subtrees from each other
6573 * without having to declare each individual cgroup's fixed share
6574 * of the ancestor's claim to protection, any unutilized -
6575 * "floating" - protection from up the tree is distributed in
6576 * proportion to each cgroup's *usage*. This makes the protection
6577 * neutral wrt sibling cgroups and lets them compete freely over
6578 * the shared parental protection budget, but it protects the
6579 * subtree as a whole from neighboring subtrees.
6581 * Note that 4. and 5. are not in conflict: 4. is about protecting
6582 * against immediate siblings whereas 5. is about protecting against
6583 * neighboring subtrees.
6585 static unsigned long effective_protection(unsigned long usage,
6586 unsigned long parent_usage,
6587 unsigned long setting,
6588 unsigned long parent_effective,
6589 unsigned long siblings_protected)
6591 unsigned long protected;
6594 protected = min(usage, setting);
6596 * If all cgroups at this level combined claim and use more
6597 * protection then what the parent affords them, distribute
6598 * shares in proportion to utilization.
6600 * We are using actual utilization rather than the statically
6601 * claimed protection in order to be work-conserving: claimed
6602 * but unused protection is available to siblings that would
6603 * otherwise get a smaller chunk than what they claimed.
6605 if (siblings_protected > parent_effective)
6606 return protected * parent_effective / siblings_protected;
6609 * Ok, utilized protection of all children is within what the
6610 * parent affords them, so we know whatever this child claims
6611 * and utilizes is effectively protected.
6613 * If there is unprotected usage beyond this value, reclaim
6614 * will apply pressure in proportion to that amount.
6616 * If there is unutilized protection, the cgroup will be fully
6617 * shielded from reclaim, but we do return a smaller value for
6618 * protection than what the group could enjoy in theory. This
6619 * is okay. With the overcommit distribution above, effective
6620 * protection is always dependent on how memory is actually
6621 * consumed among the siblings anyway.
6626 * If the children aren't claiming (all of) the protection
6627 * afforded to them by the parent, distribute the remainder in
6628 * proportion to the (unprotected) memory of each cgroup. That
6629 * way, cgroups that aren't explicitly prioritized wrt each
6630 * other compete freely over the allowance, but they are
6631 * collectively protected from neighboring trees.
6633 * We're using unprotected memory for the weight so that if
6634 * some cgroups DO claim explicit protection, we don't protect
6635 * the same bytes twice.
6637 * Check both usage and parent_usage against the respective
6638 * protected values. One should imply the other, but they
6639 * aren't read atomically - make sure the division is sane.
6641 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6643 if (parent_effective > siblings_protected &&
6644 parent_usage > siblings_protected &&
6645 usage > protected) {
6646 unsigned long unclaimed;
6648 unclaimed = parent_effective - siblings_protected;
6649 unclaimed *= usage - protected;
6650 unclaimed /= parent_usage - siblings_protected;
6659 * mem_cgroup_calculate_protection - check if memory consumption is in the normal range
6660 * @root: the top ancestor of the sub-tree being checked
6661 * @memcg: the memory cgroup to check
6663 * WARNING: This function is not stateless! It can only be used as part
6664 * of a top-down tree iteration, not for isolated queries.
6666 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6667 struct mem_cgroup *memcg)
6669 unsigned long usage, parent_usage;
6670 struct mem_cgroup *parent;
6672 if (mem_cgroup_disabled())
6676 root = root_mem_cgroup;
6679 * Effective values of the reclaim targets are ignored so they
6680 * can be stale. Have a look at mem_cgroup_protection for more
6682 * TODO: calculation should be more robust so that we do not need
6683 * that special casing.
6688 usage = page_counter_read(&memcg->memory);
6692 parent = parent_mem_cgroup(memcg);
6694 if (parent == root) {
6695 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6696 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6700 parent_usage = page_counter_read(&parent->memory);
6702 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6703 READ_ONCE(memcg->memory.min),
6704 READ_ONCE(parent->memory.emin),
6705 atomic_long_read(&parent->memory.children_min_usage)));
6707 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6708 READ_ONCE(memcg->memory.low),
6709 READ_ONCE(parent->memory.elow),
6710 atomic_long_read(&parent->memory.children_low_usage)));
6713 static int charge_memcg(struct folio *folio, struct mem_cgroup *memcg,
6716 long nr_pages = folio_nr_pages(folio);
6719 ret = try_charge(memcg, gfp, nr_pages);
6723 css_get(&memcg->css);
6724 commit_charge(folio, memcg);
6726 local_irq_disable();
6727 mem_cgroup_charge_statistics(memcg, nr_pages);
6728 memcg_check_events(memcg, folio_nid(folio));
6734 int __mem_cgroup_charge(struct folio *folio, struct mm_struct *mm, gfp_t gfp)
6736 struct mem_cgroup *memcg;
6739 memcg = get_mem_cgroup_from_mm(mm);
6740 ret = charge_memcg(folio, memcg, gfp);
6741 css_put(&memcg->css);
6747 * mem_cgroup_swapin_charge_page - charge a newly allocated page for swapin
6748 * @page: page to charge
6749 * @mm: mm context of the victim
6750 * @gfp: reclaim mode
6751 * @entry: swap entry for which the page is allocated
6753 * This function charges a page allocated for swapin. Please call this before
6754 * adding the page to the swapcache.
6756 * Returns 0 on success. Otherwise, an error code is returned.
6758 int mem_cgroup_swapin_charge_page(struct page *page, struct mm_struct *mm,
6759 gfp_t gfp, swp_entry_t entry)
6761 struct folio *folio = page_folio(page);
6762 struct mem_cgroup *memcg;
6766 if (mem_cgroup_disabled())
6769 id = lookup_swap_cgroup_id(entry);
6771 memcg = mem_cgroup_from_id(id);
6772 if (!memcg || !css_tryget_online(&memcg->css))
6773 memcg = get_mem_cgroup_from_mm(mm);
6776 ret = charge_memcg(folio, memcg, gfp);
6778 css_put(&memcg->css);
6783 * mem_cgroup_swapin_uncharge_swap - uncharge swap slot
6784 * @entry: swap entry for which the page is charged
6786 * Call this function after successfully adding the charged page to swapcache.
6788 * Note: This function assumes the page for which swap slot is being uncharged
6791 void mem_cgroup_swapin_uncharge_swap(swp_entry_t entry)
6794 * Cgroup1's unified memory+swap counter has been charged with the
6795 * new swapcache page, finish the transfer by uncharging the swap
6796 * slot. The swap slot would also get uncharged when it dies, but
6797 * it can stick around indefinitely and we'd count the page twice
6800 * Cgroup2 has separate resource counters for memory and swap,
6801 * so this is a non-issue here. Memory and swap charge lifetimes
6802 * correspond 1:1 to page and swap slot lifetimes: we charge the
6803 * page to memory here, and uncharge swap when the slot is freed.
6805 if (!mem_cgroup_disabled() && do_memsw_account()) {
6807 * The swap entry might not get freed for a long time,
6808 * let's not wait for it. The page already received a
6809 * memory+swap charge, drop the swap entry duplicate.
6811 mem_cgroup_uncharge_swap(entry, 1);
6815 struct uncharge_gather {
6816 struct mem_cgroup *memcg;
6817 unsigned long nr_memory;
6818 unsigned long pgpgout;
6819 unsigned long nr_kmem;
6823 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6825 memset(ug, 0, sizeof(*ug));
6828 static void uncharge_batch(const struct uncharge_gather *ug)
6830 unsigned long flags;
6832 if (ug->nr_memory) {
6833 page_counter_uncharge(&ug->memcg->memory, ug->nr_memory);
6834 if (do_memsw_account())
6835 page_counter_uncharge(&ug->memcg->memsw, ug->nr_memory);
6837 memcg_account_kmem(ug->memcg, -ug->nr_kmem);
6838 memcg_oom_recover(ug->memcg);
6841 local_irq_save(flags);
6842 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6843 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_memory);
6844 memcg_check_events(ug->memcg, ug->nid);
6845 local_irq_restore(flags);
6847 /* drop reference from uncharge_folio */
6848 css_put(&ug->memcg->css);
6851 static void uncharge_folio(struct folio *folio, struct uncharge_gather *ug)
6854 struct mem_cgroup *memcg;
6855 struct obj_cgroup *objcg;
6857 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
6860 * Nobody should be changing or seriously looking at
6861 * folio memcg or objcg at this point, we have fully
6862 * exclusive access to the folio.
6864 if (folio_memcg_kmem(folio)) {
6865 objcg = __folio_objcg(folio);
6867 * This get matches the put at the end of the function and
6868 * kmem pages do not hold memcg references anymore.
6870 memcg = get_mem_cgroup_from_objcg(objcg);
6872 memcg = __folio_memcg(folio);
6878 if (ug->memcg != memcg) {
6881 uncharge_gather_clear(ug);
6884 ug->nid = folio_nid(folio);
6886 /* pairs with css_put in uncharge_batch */
6887 css_get(&memcg->css);
6890 nr_pages = folio_nr_pages(folio);
6892 if (folio_memcg_kmem(folio)) {
6893 ug->nr_memory += nr_pages;
6894 ug->nr_kmem += nr_pages;
6896 folio->memcg_data = 0;
6897 obj_cgroup_put(objcg);
6899 /* LRU pages aren't accounted at the root level */
6900 if (!mem_cgroup_is_root(memcg))
6901 ug->nr_memory += nr_pages;
6904 folio->memcg_data = 0;
6907 css_put(&memcg->css);
6910 void __mem_cgroup_uncharge(struct folio *folio)
6912 struct uncharge_gather ug;
6914 /* Don't touch folio->lru of any random page, pre-check: */
6915 if (!folio_memcg(folio))
6918 uncharge_gather_clear(&ug);
6919 uncharge_folio(folio, &ug);
6920 uncharge_batch(&ug);
6924 * __mem_cgroup_uncharge_list - uncharge a list of page
6925 * @page_list: list of pages to uncharge
6927 * Uncharge a list of pages previously charged with
6928 * __mem_cgroup_charge().
6930 void __mem_cgroup_uncharge_list(struct list_head *page_list)
6932 struct uncharge_gather ug;
6933 struct folio *folio;
6935 uncharge_gather_clear(&ug);
6936 list_for_each_entry(folio, page_list, lru)
6937 uncharge_folio(folio, &ug);
6939 uncharge_batch(&ug);
6943 * mem_cgroup_migrate - Charge a folio's replacement.
6944 * @old: Currently circulating folio.
6945 * @new: Replacement folio.
6947 * Charge @new as a replacement folio for @old. @old will
6948 * be uncharged upon free.
6950 * Both folios must be locked, @new->mapping must be set up.
6952 void mem_cgroup_migrate(struct folio *old, struct folio *new)
6954 struct mem_cgroup *memcg;
6955 long nr_pages = folio_nr_pages(new);
6956 unsigned long flags;
6958 VM_BUG_ON_FOLIO(!folio_test_locked(old), old);
6959 VM_BUG_ON_FOLIO(!folio_test_locked(new), new);
6960 VM_BUG_ON_FOLIO(folio_test_anon(old) != folio_test_anon(new), new);
6961 VM_BUG_ON_FOLIO(folio_nr_pages(old) != nr_pages, new);
6963 if (mem_cgroup_disabled())
6966 /* Page cache replacement: new folio already charged? */
6967 if (folio_memcg(new))
6970 memcg = folio_memcg(old);
6971 VM_WARN_ON_ONCE_FOLIO(!memcg, old);
6975 /* Force-charge the new page. The old one will be freed soon */
6976 if (!mem_cgroup_is_root(memcg)) {
6977 page_counter_charge(&memcg->memory, nr_pages);
6978 if (do_memsw_account())
6979 page_counter_charge(&memcg->memsw, nr_pages);
6982 css_get(&memcg->css);
6983 commit_charge(new, memcg);
6985 local_irq_save(flags);
6986 mem_cgroup_charge_statistics(memcg, nr_pages);
6987 memcg_check_events(memcg, folio_nid(new));
6988 local_irq_restore(flags);
6991 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6992 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6994 void mem_cgroup_sk_alloc(struct sock *sk)
6996 struct mem_cgroup *memcg;
6998 if (!mem_cgroup_sockets_enabled)
7001 /* Do not associate the sock with unrelated interrupted task's memcg. */
7006 memcg = mem_cgroup_from_task(current);
7007 if (memcg == root_mem_cgroup)
7009 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
7011 if (css_tryget(&memcg->css))
7012 sk->sk_memcg = memcg;
7017 void mem_cgroup_sk_free(struct sock *sk)
7020 css_put(&sk->sk_memcg->css);
7024 * mem_cgroup_charge_skmem - charge socket memory
7025 * @memcg: memcg to charge
7026 * @nr_pages: number of pages to charge
7027 * @gfp_mask: reclaim mode
7029 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7030 * @memcg's configured limit, %false if it doesn't.
7032 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages,
7035 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7036 struct page_counter *fail;
7038 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7039 memcg->tcpmem_pressure = 0;
7042 memcg->tcpmem_pressure = 1;
7043 if (gfp_mask & __GFP_NOFAIL) {
7044 page_counter_charge(&memcg->tcpmem, nr_pages);
7050 if (try_charge(memcg, gfp_mask, nr_pages) == 0) {
7051 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7059 * mem_cgroup_uncharge_skmem - uncharge socket memory
7060 * @memcg: memcg to uncharge
7061 * @nr_pages: number of pages to uncharge
7063 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7065 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7066 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7070 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7072 refill_stock(memcg, nr_pages);
7075 static int __init cgroup_memory(char *s)
7079 while ((token = strsep(&s, ",")) != NULL) {
7082 if (!strcmp(token, "nosocket"))
7083 cgroup_memory_nosocket = true;
7084 if (!strcmp(token, "nokmem"))
7085 cgroup_memory_nokmem = true;
7089 __setup("cgroup.memory=", cgroup_memory);
7092 * subsys_initcall() for memory controller.
7094 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7095 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7096 * basically everything that doesn't depend on a specific mem_cgroup structure
7097 * should be initialized from here.
7099 static int __init mem_cgroup_init(void)
7104 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7105 * used for per-memcg-per-cpu caching of per-node statistics. In order
7106 * to work fine, we should make sure that the overfill threshold can't
7107 * exceed S32_MAX / PAGE_SIZE.
7109 BUILD_BUG_ON(MEMCG_CHARGE_BATCH > S32_MAX / PAGE_SIZE);
7111 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7112 memcg_hotplug_cpu_dead);
7114 for_each_possible_cpu(cpu)
7115 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7118 for_each_node(node) {
7119 struct mem_cgroup_tree_per_node *rtpn;
7121 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7122 node_online(node) ? node : NUMA_NO_NODE);
7124 rtpn->rb_root = RB_ROOT;
7125 rtpn->rb_rightmost = NULL;
7126 spin_lock_init(&rtpn->lock);
7127 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7132 subsys_initcall(mem_cgroup_init);
7134 #ifdef CONFIG_MEMCG_SWAP
7135 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7137 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7139 * The root cgroup cannot be destroyed, so it's refcount must
7142 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7146 memcg = parent_mem_cgroup(memcg);
7148 memcg = root_mem_cgroup;
7154 * mem_cgroup_swapout - transfer a memsw charge to swap
7155 * @folio: folio whose memsw charge to transfer
7156 * @entry: swap entry to move the charge to
7158 * Transfer the memsw charge of @folio to @entry.
7160 void mem_cgroup_swapout(struct folio *folio, swp_entry_t entry)
7162 struct mem_cgroup *memcg, *swap_memcg;
7163 unsigned int nr_entries;
7164 unsigned short oldid;
7166 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
7167 VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
7169 if (mem_cgroup_disabled())
7172 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7175 memcg = folio_memcg(folio);
7177 VM_WARN_ON_ONCE_FOLIO(!memcg, folio);
7182 * In case the memcg owning these pages has been offlined and doesn't
7183 * have an ID allocated to it anymore, charge the closest online
7184 * ancestor for the swap instead and transfer the memory+swap charge.
7186 swap_memcg = mem_cgroup_id_get_online(memcg);
7187 nr_entries = folio_nr_pages(folio);
7188 /* Get references for the tail pages, too */
7190 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7191 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7193 VM_BUG_ON_FOLIO(oldid, folio);
7194 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7196 folio->memcg_data = 0;
7198 if (!mem_cgroup_is_root(memcg))
7199 page_counter_uncharge(&memcg->memory, nr_entries);
7201 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7202 if (!mem_cgroup_is_root(swap_memcg))
7203 page_counter_charge(&swap_memcg->memsw, nr_entries);
7204 page_counter_uncharge(&memcg->memsw, nr_entries);
7208 * Interrupts should be disabled here because the caller holds the
7209 * i_pages lock which is taken with interrupts-off. It is
7210 * important here to have the interrupts disabled because it is the
7211 * only synchronisation we have for updating the per-CPU variables.
7214 mem_cgroup_charge_statistics(memcg, -nr_entries);
7215 memcg_stats_unlock();
7216 memcg_check_events(memcg, folio_nid(folio));
7218 css_put(&memcg->css);
7222 * __mem_cgroup_try_charge_swap - try charging swap space for a folio
7223 * @folio: folio being added to swap
7224 * @entry: swap entry to charge
7226 * Try to charge @folio's memcg for the swap space at @entry.
7228 * Returns 0 on success, -ENOMEM on failure.
7230 int __mem_cgroup_try_charge_swap(struct folio *folio, swp_entry_t entry)
7232 unsigned int nr_pages = folio_nr_pages(folio);
7233 struct page_counter *counter;
7234 struct mem_cgroup *memcg;
7235 unsigned short oldid;
7237 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7240 memcg = folio_memcg(folio);
7242 VM_WARN_ON_ONCE_FOLIO(!memcg, folio);
7247 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7251 memcg = mem_cgroup_id_get_online(memcg);
7253 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7254 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7255 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7256 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7257 mem_cgroup_id_put(memcg);
7261 /* Get references for the tail pages, too */
7263 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7264 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7265 VM_BUG_ON_FOLIO(oldid, folio);
7266 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7272 * __mem_cgroup_uncharge_swap - uncharge swap space
7273 * @entry: swap entry to uncharge
7274 * @nr_pages: the amount of swap space to uncharge
7276 void __mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7278 struct mem_cgroup *memcg;
7281 id = swap_cgroup_record(entry, 0, nr_pages);
7283 memcg = mem_cgroup_from_id(id);
7285 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7286 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7287 page_counter_uncharge(&memcg->swap, nr_pages);
7289 page_counter_uncharge(&memcg->memsw, nr_pages);
7291 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7292 mem_cgroup_id_put_many(memcg, nr_pages);
7297 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7299 long nr_swap_pages = get_nr_swap_pages();
7301 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7302 return nr_swap_pages;
7303 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7304 nr_swap_pages = min_t(long, nr_swap_pages,
7305 READ_ONCE(memcg->swap.max) -
7306 page_counter_read(&memcg->swap));
7307 return nr_swap_pages;
7310 bool mem_cgroup_swap_full(struct page *page)
7312 struct mem_cgroup *memcg;
7314 VM_BUG_ON_PAGE(!PageLocked(page), page);
7318 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7321 memcg = page_memcg(page);
7325 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7326 unsigned long usage = page_counter_read(&memcg->swap);
7328 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7329 usage * 2 >= READ_ONCE(memcg->swap.max))
7336 static int __init setup_swap_account(char *s)
7338 if (!strcmp(s, "1"))
7339 cgroup_memory_noswap = false;
7340 else if (!strcmp(s, "0"))
7341 cgroup_memory_noswap = true;
7344 __setup("swapaccount=", setup_swap_account);
7346 static u64 swap_current_read(struct cgroup_subsys_state *css,
7349 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7351 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7354 static int swap_high_show(struct seq_file *m, void *v)
7356 return seq_puts_memcg_tunable(m,
7357 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7360 static ssize_t swap_high_write(struct kernfs_open_file *of,
7361 char *buf, size_t nbytes, loff_t off)
7363 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7367 buf = strstrip(buf);
7368 err = page_counter_memparse(buf, "max", &high);
7372 page_counter_set_high(&memcg->swap, high);
7377 static int swap_max_show(struct seq_file *m, void *v)
7379 return seq_puts_memcg_tunable(m,
7380 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7383 static ssize_t swap_max_write(struct kernfs_open_file *of,
7384 char *buf, size_t nbytes, loff_t off)
7386 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7390 buf = strstrip(buf);
7391 err = page_counter_memparse(buf, "max", &max);
7395 xchg(&memcg->swap.max, max);
7400 static int swap_events_show(struct seq_file *m, void *v)
7402 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7404 seq_printf(m, "high %lu\n",
7405 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7406 seq_printf(m, "max %lu\n",
7407 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7408 seq_printf(m, "fail %lu\n",
7409 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7414 static struct cftype swap_files[] = {
7416 .name = "swap.current",
7417 .flags = CFTYPE_NOT_ON_ROOT,
7418 .read_u64 = swap_current_read,
7421 .name = "swap.high",
7422 .flags = CFTYPE_NOT_ON_ROOT,
7423 .seq_show = swap_high_show,
7424 .write = swap_high_write,
7428 .flags = CFTYPE_NOT_ON_ROOT,
7429 .seq_show = swap_max_show,
7430 .write = swap_max_write,
7433 .name = "swap.events",
7434 .flags = CFTYPE_NOT_ON_ROOT,
7435 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7436 .seq_show = swap_events_show,
7441 static struct cftype memsw_files[] = {
7443 .name = "memsw.usage_in_bytes",
7444 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7445 .read_u64 = mem_cgroup_read_u64,
7448 .name = "memsw.max_usage_in_bytes",
7449 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7450 .write = mem_cgroup_reset,
7451 .read_u64 = mem_cgroup_read_u64,
7454 .name = "memsw.limit_in_bytes",
7455 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7456 .write = mem_cgroup_write,
7457 .read_u64 = mem_cgroup_read_u64,
7460 .name = "memsw.failcnt",
7461 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7462 .write = mem_cgroup_reset,
7463 .read_u64 = mem_cgroup_read_u64,
7465 { }, /* terminate */
7468 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
7470 * obj_cgroup_may_zswap - check if this cgroup can zswap
7471 * @objcg: the object cgroup
7473 * Check if the hierarchical zswap limit has been reached.
7475 * This doesn't check for specific headroom, and it is not atomic
7476 * either. But with zswap, the size of the allocation is only known
7477 * once compression has occured, and this optimistic pre-check avoids
7478 * spending cycles on compression when there is already no room left
7479 * or zswap is disabled altogether somewhere in the hierarchy.
7481 bool obj_cgroup_may_zswap(struct obj_cgroup *objcg)
7483 struct mem_cgroup *memcg, *original_memcg;
7486 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7489 original_memcg = get_mem_cgroup_from_objcg(objcg);
7490 for (memcg = original_memcg; memcg != root_mem_cgroup;
7491 memcg = parent_mem_cgroup(memcg)) {
7492 unsigned long max = READ_ONCE(memcg->zswap_max);
7493 unsigned long pages;
7495 if (max == PAGE_COUNTER_MAX)
7502 cgroup_rstat_flush(memcg->css.cgroup);
7503 pages = memcg_page_state(memcg, MEMCG_ZSWAP_B) / PAGE_SIZE;
7509 mem_cgroup_put(original_memcg);
7514 * obj_cgroup_charge_zswap - charge compression backend memory
7515 * @objcg: the object cgroup
7516 * @size: size of compressed object
7518 * This forces the charge after obj_cgroup_may_swap() allowed
7519 * compression and storage in zwap for this cgroup to go ahead.
7521 void obj_cgroup_charge_zswap(struct obj_cgroup *objcg, size_t size)
7523 struct mem_cgroup *memcg;
7525 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7528 VM_WARN_ON_ONCE(!(current->flags & PF_MEMALLOC));
7530 /* PF_MEMALLOC context, charging must succeed */
7531 if (obj_cgroup_charge(objcg, GFP_KERNEL, size))
7535 memcg = obj_cgroup_memcg(objcg);
7536 mod_memcg_state(memcg, MEMCG_ZSWAP_B, size);
7537 mod_memcg_state(memcg, MEMCG_ZSWAPPED, 1);
7542 * obj_cgroup_uncharge_zswap - uncharge compression backend memory
7543 * @objcg: the object cgroup
7544 * @size: size of compressed object
7546 * Uncharges zswap memory on page in.
7548 void obj_cgroup_uncharge_zswap(struct obj_cgroup *objcg, size_t size)
7550 struct mem_cgroup *memcg;
7552 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7555 obj_cgroup_uncharge(objcg, size);
7558 memcg = obj_cgroup_memcg(objcg);
7559 mod_memcg_state(memcg, MEMCG_ZSWAP_B, -size);
7560 mod_memcg_state(memcg, MEMCG_ZSWAPPED, -1);
7564 static u64 zswap_current_read(struct cgroup_subsys_state *css,
7567 cgroup_rstat_flush(css->cgroup);
7568 return memcg_page_state(mem_cgroup_from_css(css), MEMCG_ZSWAP_B);
7571 static int zswap_max_show(struct seq_file *m, void *v)
7573 return seq_puts_memcg_tunable(m,
7574 READ_ONCE(mem_cgroup_from_seq(m)->zswap_max));
7577 static ssize_t zswap_max_write(struct kernfs_open_file *of,
7578 char *buf, size_t nbytes, loff_t off)
7580 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7584 buf = strstrip(buf);
7585 err = page_counter_memparse(buf, "max", &max);
7589 xchg(&memcg->zswap_max, max);
7594 static struct cftype zswap_files[] = {
7596 .name = "zswap.current",
7597 .flags = CFTYPE_NOT_ON_ROOT,
7598 .read_u64 = zswap_current_read,
7601 .name = "zswap.max",
7602 .flags = CFTYPE_NOT_ON_ROOT,
7603 .seq_show = zswap_max_show,
7604 .write = zswap_max_write,
7608 #endif /* CONFIG_MEMCG_KMEM && CONFIG_ZSWAP */
7611 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7612 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7613 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7614 * boot parameter. This may result in premature OOPS inside
7615 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7617 static int __init mem_cgroup_swap_init(void)
7619 /* No memory control -> no swap control */
7620 if (mem_cgroup_disabled())
7621 cgroup_memory_noswap = true;
7623 if (cgroup_memory_noswap)
7626 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7627 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7628 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
7629 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, zswap_files));
7633 core_initcall(mem_cgroup_swap_init);
7635 #endif /* CONFIG_MEMCG_SWAP */