1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* memcontrol.c - Memory Controller
4 * Copyright IBM Corporation, 2007
5 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
7 * Copyright 2007 OpenVZ SWsoft Inc
8 * Author: Pavel Emelianov <xemul@openvz.org>
11 * Copyright (C) 2009 Nokia Corporation
12 * Author: Kirill A. Shutemov
14 * Kernel Memory Controller
15 * Copyright (C) 2012 Parallels Inc. and Google Inc.
16 * Authors: Glauber Costa and Suleiman Souhlal
19 * Charge lifetime sanitation
20 * Lockless page tracking & accounting
21 * Unified hierarchy configuration model
22 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
24 * Per memcg lru locking
25 * Copyright (C) 2020 Alibaba, Inc, Alex Shi
28 #include <linux/page_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
31 #include <linux/pagewalk.h>
32 #include <linux/sched/mm.h>
33 #include <linux/shmem_fs.h>
34 #include <linux/hugetlb.h>
35 #include <linux/pagemap.h>
36 #include <linux/vm_event_item.h>
37 #include <linux/smp.h>
38 #include <linux/page-flags.h>
39 #include <linux/backing-dev.h>
40 #include <linux/bit_spinlock.h>
41 #include <linux/rcupdate.h>
42 #include <linux/limits.h>
43 #include <linux/export.h>
44 #include <linux/mutex.h>
45 #include <linux/rbtree.h>
46 #include <linux/slab.h>
47 #include <linux/swap.h>
48 #include <linux/swapops.h>
49 #include <linux/spinlock.h>
50 #include <linux/eventfd.h>
51 #include <linux/poll.h>
52 #include <linux/sort.h>
54 #include <linux/seq_file.h>
55 #include <linux/vmpressure.h>
56 #include <linux/mm_inline.h>
57 #include <linux/swap_cgroup.h>
58 #include <linux/cpu.h>
59 #include <linux/oom.h>
60 #include <linux/lockdep.h>
61 #include <linux/file.h>
62 #include <linux/tracehook.h>
63 #include <linux/psi.h>
64 #include <linux/seq_buf.h>
70 #include <linux/uaccess.h>
72 #include <trace/events/vmscan.h>
74 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
75 EXPORT_SYMBOL(memory_cgrp_subsys);
77 struct mem_cgroup *root_mem_cgroup __read_mostly;
79 /* Active memory cgroup to use from an interrupt context */
80 DEFINE_PER_CPU(struct mem_cgroup *, int_active_memcg);
81 EXPORT_PER_CPU_SYMBOL_GPL(int_active_memcg);
83 /* Socket memory accounting disabled? */
84 static bool cgroup_memory_nosocket __ro_after_init;
86 /* Kernel memory accounting disabled? */
87 bool cgroup_memory_nokmem __ro_after_init;
89 /* Whether the swap controller is active */
90 #ifdef CONFIG_MEMCG_SWAP
91 bool cgroup_memory_noswap __ro_after_init;
93 #define cgroup_memory_noswap 1
96 #ifdef CONFIG_CGROUP_WRITEBACK
97 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
100 /* Whether legacy memory+swap accounting is active */
101 static bool do_memsw_account(void)
103 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_noswap;
106 #define THRESHOLDS_EVENTS_TARGET 128
107 #define SOFTLIMIT_EVENTS_TARGET 1024
110 * Cgroups above their limits are maintained in a RB-Tree, independent of
111 * their hierarchy representation
114 struct mem_cgroup_tree_per_node {
115 struct rb_root rb_root;
116 struct rb_node *rb_rightmost;
120 struct mem_cgroup_tree {
121 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
124 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
127 struct mem_cgroup_eventfd_list {
128 struct list_head list;
129 struct eventfd_ctx *eventfd;
133 * cgroup_event represents events which userspace want to receive.
135 struct mem_cgroup_event {
137 * memcg which the event belongs to.
139 struct mem_cgroup *memcg;
141 * eventfd to signal userspace about the event.
143 struct eventfd_ctx *eventfd;
145 * Each of these stored in a list by the cgroup.
147 struct list_head list;
149 * register_event() callback will be used to add new userspace
150 * waiter for changes related to this event. Use eventfd_signal()
151 * on eventfd to send notification to userspace.
153 int (*register_event)(struct mem_cgroup *memcg,
154 struct eventfd_ctx *eventfd, const char *args);
156 * unregister_event() callback will be called when userspace closes
157 * the eventfd or on cgroup removing. This callback must be set,
158 * if you want provide notification functionality.
160 void (*unregister_event)(struct mem_cgroup *memcg,
161 struct eventfd_ctx *eventfd);
163 * All fields below needed to unregister event when
164 * userspace closes eventfd.
167 wait_queue_head_t *wqh;
168 wait_queue_entry_t wait;
169 struct work_struct remove;
172 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
173 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
175 /* Stuffs for move charges at task migration. */
177 * Types of charges to be moved.
179 #define MOVE_ANON 0x1U
180 #define MOVE_FILE 0x2U
181 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
183 /* "mc" and its members are protected by cgroup_mutex */
184 static struct move_charge_struct {
185 spinlock_t lock; /* for from, to */
186 struct mm_struct *mm;
187 struct mem_cgroup *from;
188 struct mem_cgroup *to;
190 unsigned long precharge;
191 unsigned long moved_charge;
192 unsigned long moved_swap;
193 struct task_struct *moving_task; /* a task moving charges */
194 wait_queue_head_t waitq; /* a waitq for other context */
196 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
197 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
201 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
202 * limit reclaim to prevent infinite loops, if they ever occur.
204 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
205 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
207 /* for encoding cft->private value on file */
216 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
217 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
218 #define MEMFILE_ATTR(val) ((val) & 0xffff)
219 /* Used for OOM notifier */
220 #define OOM_CONTROL (0)
223 * Iteration constructs for visiting all cgroups (under a tree). If
224 * loops are exited prematurely (break), mem_cgroup_iter_break() must
225 * be used for reference counting.
227 #define for_each_mem_cgroup_tree(iter, root) \
228 for (iter = mem_cgroup_iter(root, NULL, NULL); \
230 iter = mem_cgroup_iter(root, iter, NULL))
232 #define for_each_mem_cgroup(iter) \
233 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
235 iter = mem_cgroup_iter(NULL, iter, NULL))
237 static inline bool task_is_dying(void)
239 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
240 (current->flags & PF_EXITING);
243 /* Some nice accessors for the vmpressure. */
244 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
247 memcg = root_mem_cgroup;
248 return &memcg->vmpressure;
251 struct mem_cgroup *vmpressure_to_memcg(struct vmpressure *vmpr)
253 return container_of(vmpr, struct mem_cgroup, vmpressure);
256 #ifdef CONFIG_MEMCG_KMEM
257 static DEFINE_SPINLOCK(objcg_lock);
259 bool mem_cgroup_kmem_disabled(void)
261 return cgroup_memory_nokmem;
264 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
265 unsigned int nr_pages);
267 static void obj_cgroup_release(struct percpu_ref *ref)
269 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
270 unsigned int nr_bytes;
271 unsigned int nr_pages;
275 * At this point all allocated objects are freed, and
276 * objcg->nr_charged_bytes can't have an arbitrary byte value.
277 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
279 * The following sequence can lead to it:
280 * 1) CPU0: objcg == stock->cached_objcg
281 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
282 * PAGE_SIZE bytes are charged
283 * 3) CPU1: a process from another memcg is allocating something,
284 * the stock if flushed,
285 * objcg->nr_charged_bytes = PAGE_SIZE - 92
286 * 5) CPU0: we do release this object,
287 * 92 bytes are added to stock->nr_bytes
288 * 6) CPU0: stock is flushed,
289 * 92 bytes are added to objcg->nr_charged_bytes
291 * In the result, nr_charged_bytes == PAGE_SIZE.
292 * This page will be uncharged in obj_cgroup_release().
294 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
295 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
296 nr_pages = nr_bytes >> PAGE_SHIFT;
299 obj_cgroup_uncharge_pages(objcg, nr_pages);
301 spin_lock_irqsave(&objcg_lock, flags);
302 list_del(&objcg->list);
303 spin_unlock_irqrestore(&objcg_lock, flags);
305 percpu_ref_exit(ref);
306 kfree_rcu(objcg, rcu);
309 static struct obj_cgroup *obj_cgroup_alloc(void)
311 struct obj_cgroup *objcg;
314 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
318 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
324 INIT_LIST_HEAD(&objcg->list);
328 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
329 struct mem_cgroup *parent)
331 struct obj_cgroup *objcg, *iter;
333 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
335 spin_lock_irq(&objcg_lock);
337 /* 1) Ready to reparent active objcg. */
338 list_add(&objcg->list, &memcg->objcg_list);
339 /* 2) Reparent active objcg and already reparented objcgs to parent. */
340 list_for_each_entry(iter, &memcg->objcg_list, list)
341 WRITE_ONCE(iter->memcg, parent);
342 /* 3) Move already reparented objcgs to the parent's list */
343 list_splice(&memcg->objcg_list, &parent->objcg_list);
345 spin_unlock_irq(&objcg_lock);
347 percpu_ref_kill(&objcg->refcnt);
351 * This will be used as a shrinker list's index.
352 * The main reason for not using cgroup id for this:
353 * this works better in sparse environments, where we have a lot of memcgs,
354 * but only a few kmem-limited. Or also, if we have, for instance, 200
355 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
356 * 200 entry array for that.
358 * The current size of the caches array is stored in memcg_nr_cache_ids. It
359 * will double each time we have to increase it.
361 static DEFINE_IDA(memcg_cache_ida);
362 int memcg_nr_cache_ids;
364 /* Protects memcg_nr_cache_ids */
365 static DECLARE_RWSEM(memcg_cache_ids_sem);
367 void memcg_get_cache_ids(void)
369 down_read(&memcg_cache_ids_sem);
372 void memcg_put_cache_ids(void)
374 up_read(&memcg_cache_ids_sem);
378 * MIN_SIZE is different than 1, because we would like to avoid going through
379 * the alloc/free process all the time. In a small machine, 4 kmem-limited
380 * cgroups is a reasonable guess. In the future, it could be a parameter or
381 * tunable, but that is strictly not necessary.
383 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
384 * this constant directly from cgroup, but it is understandable that this is
385 * better kept as an internal representation in cgroup.c. In any case, the
386 * cgrp_id space is not getting any smaller, and we don't have to necessarily
387 * increase ours as well if it increases.
389 #define MEMCG_CACHES_MIN_SIZE 4
390 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
393 * A lot of the calls to the cache allocation functions are expected to be
394 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
395 * conditional to this static branch, we'll have to allow modules that does
396 * kmem_cache_alloc and the such to see this symbol as well
398 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
399 EXPORT_SYMBOL(memcg_kmem_enabled_key);
403 * mem_cgroup_css_from_page - css of the memcg associated with a page
404 * @page: page of interest
406 * If memcg is bound to the default hierarchy, css of the memcg associated
407 * with @page is returned. The returned css remains associated with @page
408 * until it is released.
410 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
413 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
415 struct mem_cgroup *memcg;
417 memcg = page_memcg(page);
419 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
420 memcg = root_mem_cgroup;
426 * page_cgroup_ino - return inode number of the memcg a page is charged to
429 * Look up the closest online ancestor of the memory cgroup @page is charged to
430 * and return its inode number or 0 if @page is not charged to any cgroup. It
431 * is safe to call this function without holding a reference to @page.
433 * Note, this function is inherently racy, because there is nothing to prevent
434 * the cgroup inode from getting torn down and potentially reallocated a moment
435 * after page_cgroup_ino() returns, so it only should be used by callers that
436 * do not care (such as procfs interfaces).
438 ino_t page_cgroup_ino(struct page *page)
440 struct mem_cgroup *memcg;
441 unsigned long ino = 0;
444 memcg = page_memcg_check(page);
446 while (memcg && !(memcg->css.flags & CSS_ONLINE))
447 memcg = parent_mem_cgroup(memcg);
449 ino = cgroup_ino(memcg->css.cgroup);
454 static struct mem_cgroup_per_node *
455 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
457 int nid = page_to_nid(page);
459 return memcg->nodeinfo[nid];
462 static struct mem_cgroup_tree_per_node *
463 soft_limit_tree_node(int nid)
465 return soft_limit_tree.rb_tree_per_node[nid];
468 static struct mem_cgroup_tree_per_node *
469 soft_limit_tree_from_page(struct page *page)
471 int nid = page_to_nid(page);
473 return soft_limit_tree.rb_tree_per_node[nid];
476 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
477 struct mem_cgroup_tree_per_node *mctz,
478 unsigned long new_usage_in_excess)
480 struct rb_node **p = &mctz->rb_root.rb_node;
481 struct rb_node *parent = NULL;
482 struct mem_cgroup_per_node *mz_node;
483 bool rightmost = true;
488 mz->usage_in_excess = new_usage_in_excess;
489 if (!mz->usage_in_excess)
493 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
495 if (mz->usage_in_excess < mz_node->usage_in_excess) {
504 mctz->rb_rightmost = &mz->tree_node;
506 rb_link_node(&mz->tree_node, parent, p);
507 rb_insert_color(&mz->tree_node, &mctz->rb_root);
511 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
512 struct mem_cgroup_tree_per_node *mctz)
517 if (&mz->tree_node == mctz->rb_rightmost)
518 mctz->rb_rightmost = rb_prev(&mz->tree_node);
520 rb_erase(&mz->tree_node, &mctz->rb_root);
524 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
525 struct mem_cgroup_tree_per_node *mctz)
529 spin_lock_irqsave(&mctz->lock, flags);
530 __mem_cgroup_remove_exceeded(mz, mctz);
531 spin_unlock_irqrestore(&mctz->lock, flags);
534 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
536 unsigned long nr_pages = page_counter_read(&memcg->memory);
537 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
538 unsigned long excess = 0;
540 if (nr_pages > soft_limit)
541 excess = nr_pages - soft_limit;
546 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
548 unsigned long excess;
549 struct mem_cgroup_per_node *mz;
550 struct mem_cgroup_tree_per_node *mctz;
552 mctz = soft_limit_tree_from_page(page);
556 * Necessary to update all ancestors when hierarchy is used.
557 * because their event counter is not touched.
559 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
560 mz = mem_cgroup_page_nodeinfo(memcg, page);
561 excess = soft_limit_excess(memcg);
563 * We have to update the tree if mz is on RB-tree or
564 * mem is over its softlimit.
566 if (excess || mz->on_tree) {
569 spin_lock_irqsave(&mctz->lock, flags);
570 /* if on-tree, remove it */
572 __mem_cgroup_remove_exceeded(mz, mctz);
574 * Insert again. mz->usage_in_excess will be updated.
575 * If excess is 0, no tree ops.
577 __mem_cgroup_insert_exceeded(mz, mctz, excess);
578 spin_unlock_irqrestore(&mctz->lock, flags);
583 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
585 struct mem_cgroup_tree_per_node *mctz;
586 struct mem_cgroup_per_node *mz;
590 mz = memcg->nodeinfo[nid];
591 mctz = soft_limit_tree_node(nid);
593 mem_cgroup_remove_exceeded(mz, mctz);
597 static struct mem_cgroup_per_node *
598 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
600 struct mem_cgroup_per_node *mz;
604 if (!mctz->rb_rightmost)
605 goto done; /* Nothing to reclaim from */
607 mz = rb_entry(mctz->rb_rightmost,
608 struct mem_cgroup_per_node, tree_node);
610 * Remove the node now but someone else can add it back,
611 * we will to add it back at the end of reclaim to its correct
612 * position in the tree.
614 __mem_cgroup_remove_exceeded(mz, mctz);
615 if (!soft_limit_excess(mz->memcg) ||
616 !css_tryget(&mz->memcg->css))
622 static struct mem_cgroup_per_node *
623 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
625 struct mem_cgroup_per_node *mz;
627 spin_lock_irq(&mctz->lock);
628 mz = __mem_cgroup_largest_soft_limit_node(mctz);
629 spin_unlock_irq(&mctz->lock);
634 * memcg and lruvec stats flushing
636 * Many codepaths leading to stats update or read are performance sensitive and
637 * adding stats flushing in such codepaths is not desirable. So, to optimize the
638 * flushing the kernel does:
640 * 1) Periodically and asynchronously flush the stats every 2 seconds to not let
641 * rstat update tree grow unbounded.
643 * 2) Flush the stats synchronously on reader side only when there are more than
644 * (MEMCG_CHARGE_BATCH * nr_cpus) update events. Though this optimization
645 * will let stats be out of sync by atmost (MEMCG_CHARGE_BATCH * nr_cpus) but
646 * only for 2 seconds due to (1).
648 static void flush_memcg_stats_dwork(struct work_struct *w);
649 static DECLARE_DEFERRABLE_WORK(stats_flush_dwork, flush_memcg_stats_dwork);
650 static DEFINE_SPINLOCK(stats_flush_lock);
651 static DEFINE_PER_CPU(unsigned int, stats_updates);
652 static atomic_t stats_flush_threshold = ATOMIC_INIT(0);
653 static u64 flush_next_time;
655 #define FLUSH_TIME (2UL*HZ)
658 * Accessors to ensure that preemption is disabled on PREEMPT_RT because it can
659 * not rely on this as part of an acquired spinlock_t lock. These functions are
660 * never used in hardirq context on PREEMPT_RT and therefore disabling preemtion
663 static void memcg_stats_lock(void)
665 #ifdef CONFIG_PREEMPT_RT
668 VM_BUG_ON(!irqs_disabled());
672 static void __memcg_stats_lock(void)
674 #ifdef CONFIG_PREEMPT_RT
679 static void memcg_stats_unlock(void)
681 #ifdef CONFIG_PREEMPT_RT
686 static inline void memcg_rstat_updated(struct mem_cgroup *memcg, int val)
690 cgroup_rstat_updated(memcg->css.cgroup, smp_processor_id());
692 x = __this_cpu_add_return(stats_updates, abs(val));
693 if (x > MEMCG_CHARGE_BATCH) {
694 atomic_add(x / MEMCG_CHARGE_BATCH, &stats_flush_threshold);
695 __this_cpu_write(stats_updates, 0);
699 static void __mem_cgroup_flush_stats(void)
703 if (!spin_trylock_irqsave(&stats_flush_lock, flag))
706 flush_next_time = jiffies_64 + 2*FLUSH_TIME;
707 cgroup_rstat_flush_irqsafe(root_mem_cgroup->css.cgroup);
708 atomic_set(&stats_flush_threshold, 0);
709 spin_unlock_irqrestore(&stats_flush_lock, flag);
712 void mem_cgroup_flush_stats(void)
714 if (atomic_read(&stats_flush_threshold) > num_online_cpus())
715 __mem_cgroup_flush_stats();
718 void mem_cgroup_flush_stats_delayed(void)
720 if (time_after64(jiffies_64, flush_next_time))
721 mem_cgroup_flush_stats();
724 static void flush_memcg_stats_dwork(struct work_struct *w)
726 __mem_cgroup_flush_stats();
727 queue_delayed_work(system_unbound_wq, &stats_flush_dwork, FLUSH_TIME);
731 * __mod_memcg_state - update cgroup memory statistics
732 * @memcg: the memory cgroup
733 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
734 * @val: delta to add to the counter, can be negative
736 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
738 if (mem_cgroup_disabled())
741 __this_cpu_add(memcg->vmstats_percpu->state[idx], val);
742 memcg_rstat_updated(memcg, val);
745 /* idx can be of type enum memcg_stat_item or node_stat_item. */
746 static unsigned long memcg_page_state_local(struct mem_cgroup *memcg, int idx)
751 for_each_possible_cpu(cpu)
752 x += per_cpu(memcg->vmstats_percpu->state[idx], cpu);
760 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
763 struct mem_cgroup_per_node *pn;
764 struct mem_cgroup *memcg;
766 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
770 * The caller from rmap relay on disabled preemption becase they never
771 * update their counter from in-interrupt context. For these two
772 * counters we check that the update is never performed from an
773 * interrupt context while other caller need to have disabled interrupt.
775 __memcg_stats_lock();
776 if (IS_ENABLED(CONFIG_DEBUG_VM) && !IS_ENABLED(CONFIG_PREEMPT_RT)) {
781 case NR_SHMEM_PMDMAPPED:
782 case NR_FILE_PMDMAPPED:
783 WARN_ON_ONCE(!in_task());
786 WARN_ON_ONCE(!irqs_disabled());
791 __this_cpu_add(memcg->vmstats_percpu->state[idx], val);
794 __this_cpu_add(pn->lruvec_stats_percpu->state[idx], val);
796 memcg_rstat_updated(memcg, val);
797 memcg_stats_unlock();
801 * __mod_lruvec_state - update lruvec memory statistics
802 * @lruvec: the lruvec
803 * @idx: the stat item
804 * @val: delta to add to the counter, can be negative
806 * The lruvec is the intersection of the NUMA node and a cgroup. This
807 * function updates the all three counters that are affected by a
808 * change of state at this level: per-node, per-cgroup, per-lruvec.
810 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
814 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
816 /* Update memcg and lruvec */
817 if (!mem_cgroup_disabled())
818 __mod_memcg_lruvec_state(lruvec, idx, val);
821 void __mod_lruvec_page_state(struct page *page, enum node_stat_item idx,
824 struct page *head = compound_head(page); /* rmap on tail pages */
825 struct mem_cgroup *memcg;
826 pg_data_t *pgdat = page_pgdat(page);
827 struct lruvec *lruvec;
830 memcg = page_memcg(head);
831 /* Untracked pages have no memcg, no lruvec. Update only the node */
834 __mod_node_page_state(pgdat, idx, val);
838 lruvec = mem_cgroup_lruvec(memcg, pgdat);
839 __mod_lruvec_state(lruvec, idx, val);
842 EXPORT_SYMBOL(__mod_lruvec_page_state);
844 void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val)
846 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
847 struct mem_cgroup *memcg;
848 struct lruvec *lruvec;
851 memcg = mem_cgroup_from_obj(p);
854 * Untracked pages have no memcg, no lruvec. Update only the
855 * node. If we reparent the slab objects to the root memcg,
856 * when we free the slab object, we need to update the per-memcg
857 * vmstats to keep it correct for the root memcg.
860 __mod_node_page_state(pgdat, idx, val);
862 lruvec = mem_cgroup_lruvec(memcg, pgdat);
863 __mod_lruvec_state(lruvec, idx, val);
869 * mod_objcg_mlstate() may be called with irq enabled, so
870 * mod_memcg_lruvec_state() should be used.
872 static inline void mod_objcg_mlstate(struct obj_cgroup *objcg,
873 struct pglist_data *pgdat,
874 enum node_stat_item idx, int nr)
876 struct mem_cgroup *memcg;
877 struct lruvec *lruvec;
880 memcg = obj_cgroup_memcg(objcg);
881 lruvec = mem_cgroup_lruvec(memcg, pgdat);
882 mod_memcg_lruvec_state(lruvec, idx, nr);
887 * __count_memcg_events - account VM events in a cgroup
888 * @memcg: the memory cgroup
889 * @idx: the event item
890 * @count: the number of events that occurred
892 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
895 if (mem_cgroup_disabled())
899 __this_cpu_add(memcg->vmstats_percpu->events[idx], count);
900 memcg_rstat_updated(memcg, count);
901 memcg_stats_unlock();
904 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
906 return READ_ONCE(memcg->vmstats.events[event]);
909 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
914 for_each_possible_cpu(cpu)
915 x += per_cpu(memcg->vmstats_percpu->events[event], cpu);
919 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
923 /* pagein of a big page is an event. So, ignore page size */
925 __count_memcg_events(memcg, PGPGIN, 1);
927 __count_memcg_events(memcg, PGPGOUT, 1);
928 nr_pages = -nr_pages; /* for event */
931 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
934 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
935 enum mem_cgroup_events_target target)
937 unsigned long val, next;
939 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
940 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
941 /* from time_after() in jiffies.h */
942 if ((long)(next - val) < 0) {
944 case MEM_CGROUP_TARGET_THRESH:
945 next = val + THRESHOLDS_EVENTS_TARGET;
947 case MEM_CGROUP_TARGET_SOFTLIMIT:
948 next = val + SOFTLIMIT_EVENTS_TARGET;
953 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
960 * Check events in order.
963 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
965 if (IS_ENABLED(CONFIG_PREEMPT_RT))
968 /* threshold event is triggered in finer grain than soft limit */
969 if (unlikely(mem_cgroup_event_ratelimit(memcg,
970 MEM_CGROUP_TARGET_THRESH))) {
973 do_softlimit = mem_cgroup_event_ratelimit(memcg,
974 MEM_CGROUP_TARGET_SOFTLIMIT);
975 mem_cgroup_threshold(memcg);
976 if (unlikely(do_softlimit))
977 mem_cgroup_update_tree(memcg, page);
981 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
984 * mm_update_next_owner() may clear mm->owner to NULL
985 * if it races with swapoff, page migration, etc.
986 * So this can be called with p == NULL.
991 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
993 EXPORT_SYMBOL(mem_cgroup_from_task);
995 static __always_inline struct mem_cgroup *active_memcg(void)
998 return this_cpu_read(int_active_memcg);
1000 return current->active_memcg;
1004 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1005 * @mm: mm from which memcg should be extracted. It can be NULL.
1007 * Obtain a reference on mm->memcg and returns it if successful. If mm
1008 * is NULL, then the memcg is chosen as follows:
1009 * 1) The active memcg, if set.
1010 * 2) current->mm->memcg, if available
1012 * If mem_cgroup is disabled, NULL is returned.
1014 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1016 struct mem_cgroup *memcg;
1018 if (mem_cgroup_disabled())
1022 * Page cache insertions can happen without an
1023 * actual mm context, e.g. during disk probing
1024 * on boot, loopback IO, acct() writes etc.
1026 * No need to css_get on root memcg as the reference
1027 * counting is disabled on the root level in the
1028 * cgroup core. See CSS_NO_REF.
1030 if (unlikely(!mm)) {
1031 memcg = active_memcg();
1032 if (unlikely(memcg)) {
1033 /* remote memcg must hold a ref */
1034 css_get(&memcg->css);
1039 return root_mem_cgroup;
1044 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1045 if (unlikely(!memcg))
1046 memcg = root_mem_cgroup;
1047 } while (!css_tryget(&memcg->css));
1051 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1053 static __always_inline bool memcg_kmem_bypass(void)
1055 /* Allow remote memcg charging from any context. */
1056 if (unlikely(active_memcg()))
1059 /* Memcg to charge can't be determined. */
1060 if (!in_task() || !current->mm || (current->flags & PF_KTHREAD))
1067 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1068 * @root: hierarchy root
1069 * @prev: previously returned memcg, NULL on first invocation
1070 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1072 * Returns references to children of the hierarchy below @root, or
1073 * @root itself, or %NULL after a full round-trip.
1075 * Caller must pass the return value in @prev on subsequent
1076 * invocations for reference counting, or use mem_cgroup_iter_break()
1077 * to cancel a hierarchy walk before the round-trip is complete.
1079 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1080 * in the hierarchy among all concurrent reclaimers operating on the
1083 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1084 struct mem_cgroup *prev,
1085 struct mem_cgroup_reclaim_cookie *reclaim)
1087 struct mem_cgroup_reclaim_iter *iter;
1088 struct cgroup_subsys_state *css = NULL;
1089 struct mem_cgroup *memcg = NULL;
1090 struct mem_cgroup *pos = NULL;
1092 if (mem_cgroup_disabled())
1096 root = root_mem_cgroup;
1098 if (prev && !reclaim)
1104 struct mem_cgroup_per_node *mz;
1106 mz = root->nodeinfo[reclaim->pgdat->node_id];
1109 if (prev && reclaim->generation != iter->generation)
1113 pos = READ_ONCE(iter->position);
1114 if (!pos || css_tryget(&pos->css))
1117 * css reference reached zero, so iter->position will
1118 * be cleared by ->css_released. However, we should not
1119 * rely on this happening soon, because ->css_released
1120 * is called from a work queue, and by busy-waiting we
1121 * might block it. So we clear iter->position right
1124 (void)cmpxchg(&iter->position, pos, NULL);
1132 css = css_next_descendant_pre(css, &root->css);
1135 * Reclaimers share the hierarchy walk, and a
1136 * new one might jump in right at the end of
1137 * the hierarchy - make sure they see at least
1138 * one group and restart from the beginning.
1146 * Verify the css and acquire a reference. The root
1147 * is provided by the caller, so we know it's alive
1148 * and kicking, and don't take an extra reference.
1150 memcg = mem_cgroup_from_css(css);
1152 if (css == &root->css)
1155 if (css_tryget(css))
1163 * The position could have already been updated by a competing
1164 * thread, so check that the value hasn't changed since we read
1165 * it to avoid reclaiming from the same cgroup twice.
1167 (void)cmpxchg(&iter->position, pos, memcg);
1175 reclaim->generation = iter->generation;
1180 if (prev && prev != root)
1181 css_put(&prev->css);
1187 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1188 * @root: hierarchy root
1189 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1191 void mem_cgroup_iter_break(struct mem_cgroup *root,
1192 struct mem_cgroup *prev)
1195 root = root_mem_cgroup;
1196 if (prev && prev != root)
1197 css_put(&prev->css);
1200 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1201 struct mem_cgroup *dead_memcg)
1203 struct mem_cgroup_reclaim_iter *iter;
1204 struct mem_cgroup_per_node *mz;
1207 for_each_node(nid) {
1208 mz = from->nodeinfo[nid];
1210 cmpxchg(&iter->position, dead_memcg, NULL);
1214 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1216 struct mem_cgroup *memcg = dead_memcg;
1217 struct mem_cgroup *last;
1220 __invalidate_reclaim_iterators(memcg, dead_memcg);
1222 } while ((memcg = parent_mem_cgroup(memcg)));
1225 * When cgruop1 non-hierarchy mode is used,
1226 * parent_mem_cgroup() does not walk all the way up to the
1227 * cgroup root (root_mem_cgroup). So we have to handle
1228 * dead_memcg from cgroup root separately.
1230 if (last != root_mem_cgroup)
1231 __invalidate_reclaim_iterators(root_mem_cgroup,
1236 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1237 * @memcg: hierarchy root
1238 * @fn: function to call for each task
1239 * @arg: argument passed to @fn
1241 * This function iterates over tasks attached to @memcg or to any of its
1242 * descendants and calls @fn for each task. If @fn returns a non-zero
1243 * value, the function breaks the iteration loop and returns the value.
1244 * Otherwise, it will iterate over all tasks and return 0.
1246 * This function must not be called for the root memory cgroup.
1248 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1249 int (*fn)(struct task_struct *, void *), void *arg)
1251 struct mem_cgroup *iter;
1254 BUG_ON(memcg == root_mem_cgroup);
1256 for_each_mem_cgroup_tree(iter, memcg) {
1257 struct css_task_iter it;
1258 struct task_struct *task;
1260 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1261 while (!ret && (task = css_task_iter_next(&it)))
1262 ret = fn(task, arg);
1263 css_task_iter_end(&it);
1265 mem_cgroup_iter_break(memcg, iter);
1272 #ifdef CONFIG_DEBUG_VM
1273 void lruvec_memcg_debug(struct lruvec *lruvec, struct page *page)
1275 struct mem_cgroup *memcg;
1277 if (mem_cgroup_disabled())
1280 memcg = page_memcg(page);
1283 VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != root_mem_cgroup, page);
1285 VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != memcg, page);
1290 * lock_page_lruvec - lock and return lruvec for a given page.
1293 * These functions are safe to use under any of the following conditions:
1296 * - lock_page_memcg()
1297 * - page->_refcount is zero
1299 struct lruvec *lock_page_lruvec(struct page *page)
1301 struct lruvec *lruvec;
1303 lruvec = mem_cgroup_page_lruvec(page);
1304 spin_lock(&lruvec->lru_lock);
1306 lruvec_memcg_debug(lruvec, page);
1311 struct lruvec *lock_page_lruvec_irq(struct page *page)
1313 struct lruvec *lruvec;
1315 lruvec = mem_cgroup_page_lruvec(page);
1316 spin_lock_irq(&lruvec->lru_lock);
1318 lruvec_memcg_debug(lruvec, page);
1323 struct lruvec *lock_page_lruvec_irqsave(struct page *page, unsigned long *flags)
1325 struct lruvec *lruvec;
1327 lruvec = mem_cgroup_page_lruvec(page);
1328 spin_lock_irqsave(&lruvec->lru_lock, *flags);
1330 lruvec_memcg_debug(lruvec, page);
1336 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1337 * @lruvec: mem_cgroup per zone lru vector
1338 * @lru: index of lru list the page is sitting on
1339 * @zid: zone id of the accounted pages
1340 * @nr_pages: positive when adding or negative when removing
1342 * This function must be called under lru_lock, just before a page is added
1343 * to or just after a page is removed from an lru list (that ordering being
1344 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1346 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1347 int zid, int nr_pages)
1349 struct mem_cgroup_per_node *mz;
1350 unsigned long *lru_size;
1353 if (mem_cgroup_disabled())
1356 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1357 lru_size = &mz->lru_zone_size[zid][lru];
1360 *lru_size += nr_pages;
1363 if (WARN_ONCE(size < 0,
1364 "%s(%p, %d, %d): lru_size %ld\n",
1365 __func__, lruvec, lru, nr_pages, size)) {
1371 *lru_size += nr_pages;
1375 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1376 * @memcg: the memory cgroup
1378 * Returns the maximum amount of memory @mem can be charged with, in
1381 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1383 unsigned long margin = 0;
1384 unsigned long count;
1385 unsigned long limit;
1387 count = page_counter_read(&memcg->memory);
1388 limit = READ_ONCE(memcg->memory.max);
1390 margin = limit - count;
1392 if (do_memsw_account()) {
1393 count = page_counter_read(&memcg->memsw);
1394 limit = READ_ONCE(memcg->memsw.max);
1396 margin = min(margin, limit - count);
1405 * A routine for checking "mem" is under move_account() or not.
1407 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1408 * moving cgroups. This is for waiting at high-memory pressure
1411 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1413 struct mem_cgroup *from;
1414 struct mem_cgroup *to;
1417 * Unlike task_move routines, we access mc.to, mc.from not under
1418 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1420 spin_lock(&mc.lock);
1426 ret = mem_cgroup_is_descendant(from, memcg) ||
1427 mem_cgroup_is_descendant(to, memcg);
1429 spin_unlock(&mc.lock);
1433 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1435 if (mc.moving_task && current != mc.moving_task) {
1436 if (mem_cgroup_under_move(memcg)) {
1438 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1439 /* moving charge context might have finished. */
1442 finish_wait(&mc.waitq, &wait);
1449 struct memory_stat {
1454 static const struct memory_stat memory_stats[] = {
1455 { "anon", NR_ANON_MAPPED },
1456 { "file", NR_FILE_PAGES },
1457 { "kernel_stack", NR_KERNEL_STACK_KB },
1458 { "pagetables", NR_PAGETABLE },
1459 { "percpu", MEMCG_PERCPU_B },
1460 { "sock", MEMCG_SOCK },
1461 { "shmem", NR_SHMEM },
1462 { "file_mapped", NR_FILE_MAPPED },
1463 { "file_dirty", NR_FILE_DIRTY },
1464 { "file_writeback", NR_WRITEBACK },
1466 { "swapcached", NR_SWAPCACHE },
1468 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1469 { "anon_thp", NR_ANON_THPS },
1470 { "file_thp", NR_FILE_THPS },
1471 { "shmem_thp", NR_SHMEM_THPS },
1473 { "inactive_anon", NR_INACTIVE_ANON },
1474 { "active_anon", NR_ACTIVE_ANON },
1475 { "inactive_file", NR_INACTIVE_FILE },
1476 { "active_file", NR_ACTIVE_FILE },
1477 { "unevictable", NR_UNEVICTABLE },
1478 { "slab_reclaimable", NR_SLAB_RECLAIMABLE_B },
1479 { "slab_unreclaimable", NR_SLAB_UNRECLAIMABLE_B },
1481 /* The memory events */
1482 { "workingset_refault_anon", WORKINGSET_REFAULT_ANON },
1483 { "workingset_refault_file", WORKINGSET_REFAULT_FILE },
1484 { "workingset_activate_anon", WORKINGSET_ACTIVATE_ANON },
1485 { "workingset_activate_file", WORKINGSET_ACTIVATE_FILE },
1486 { "workingset_restore_anon", WORKINGSET_RESTORE_ANON },
1487 { "workingset_restore_file", WORKINGSET_RESTORE_FILE },
1488 { "workingset_nodereclaim", WORKINGSET_NODERECLAIM },
1491 /* Translate stat items to the correct unit for memory.stat output */
1492 static int memcg_page_state_unit(int item)
1495 case MEMCG_PERCPU_B:
1496 case NR_SLAB_RECLAIMABLE_B:
1497 case NR_SLAB_UNRECLAIMABLE_B:
1498 case WORKINGSET_REFAULT_ANON:
1499 case WORKINGSET_REFAULT_FILE:
1500 case WORKINGSET_ACTIVATE_ANON:
1501 case WORKINGSET_ACTIVATE_FILE:
1502 case WORKINGSET_RESTORE_ANON:
1503 case WORKINGSET_RESTORE_FILE:
1504 case WORKINGSET_NODERECLAIM:
1506 case NR_KERNEL_STACK_KB:
1513 static inline unsigned long memcg_page_state_output(struct mem_cgroup *memcg,
1516 return memcg_page_state(memcg, item) * memcg_page_state_unit(item);
1519 static char *memory_stat_format(struct mem_cgroup *memcg)
1524 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1529 * Provide statistics on the state of the memory subsystem as
1530 * well as cumulative event counters that show past behavior.
1532 * This list is ordered following a combination of these gradients:
1533 * 1) generic big picture -> specifics and details
1534 * 2) reflecting userspace activity -> reflecting kernel heuristics
1536 * Current memory state:
1538 mem_cgroup_flush_stats();
1540 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1543 size = memcg_page_state_output(memcg, memory_stats[i].idx);
1544 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1546 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1547 size += memcg_page_state_output(memcg,
1548 NR_SLAB_RECLAIMABLE_B);
1549 seq_buf_printf(&s, "slab %llu\n", size);
1553 /* Accumulated memory events */
1555 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1556 memcg_events(memcg, PGFAULT));
1557 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1558 memcg_events(memcg, PGMAJFAULT));
1559 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1560 memcg_events(memcg, PGREFILL));
1561 seq_buf_printf(&s, "pgscan %lu\n",
1562 memcg_events(memcg, PGSCAN_KSWAPD) +
1563 memcg_events(memcg, PGSCAN_DIRECT));
1564 seq_buf_printf(&s, "pgsteal %lu\n",
1565 memcg_events(memcg, PGSTEAL_KSWAPD) +
1566 memcg_events(memcg, PGSTEAL_DIRECT));
1567 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1568 memcg_events(memcg, PGACTIVATE));
1569 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1570 memcg_events(memcg, PGDEACTIVATE));
1571 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1572 memcg_events(memcg, PGLAZYFREE));
1573 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1574 memcg_events(memcg, PGLAZYFREED));
1576 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1577 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1578 memcg_events(memcg, THP_FAULT_ALLOC));
1579 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1580 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1581 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1583 /* The above should easily fit into one page */
1584 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1589 #define K(x) ((x) << (PAGE_SHIFT-10))
1591 * mem_cgroup_print_oom_context: Print OOM information relevant to
1592 * memory controller.
1593 * @memcg: The memory cgroup that went over limit
1594 * @p: Task that is going to be killed
1596 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1599 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1604 pr_cont(",oom_memcg=");
1605 pr_cont_cgroup_path(memcg->css.cgroup);
1607 pr_cont(",global_oom");
1609 pr_cont(",task_memcg=");
1610 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1616 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1617 * memory controller.
1618 * @memcg: The memory cgroup that went over limit
1620 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1624 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1625 K((u64)page_counter_read(&memcg->memory)),
1626 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1627 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1628 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1629 K((u64)page_counter_read(&memcg->swap)),
1630 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1632 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1633 K((u64)page_counter_read(&memcg->memsw)),
1634 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1635 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1636 K((u64)page_counter_read(&memcg->kmem)),
1637 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1640 pr_info("Memory cgroup stats for ");
1641 pr_cont_cgroup_path(memcg->css.cgroup);
1643 buf = memory_stat_format(memcg);
1651 * Return the memory (and swap, if configured) limit for a memcg.
1653 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1655 unsigned long max = READ_ONCE(memcg->memory.max);
1657 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1658 if (mem_cgroup_swappiness(memcg))
1659 max += min(READ_ONCE(memcg->swap.max),
1660 (unsigned long)total_swap_pages);
1662 if (mem_cgroup_swappiness(memcg)) {
1663 /* Calculate swap excess capacity from memsw limit */
1664 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1666 max += min(swap, (unsigned long)total_swap_pages);
1672 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1674 return page_counter_read(&memcg->memory);
1677 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1680 struct oom_control oc = {
1684 .gfp_mask = gfp_mask,
1689 if (mutex_lock_killable(&oom_lock))
1692 if (mem_cgroup_margin(memcg) >= (1 << order))
1696 * A few threads which were not waiting at mutex_lock_killable() can
1697 * fail to bail out. Therefore, check again after holding oom_lock.
1699 ret = task_is_dying() || out_of_memory(&oc);
1702 mutex_unlock(&oom_lock);
1706 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1709 unsigned long *total_scanned)
1711 struct mem_cgroup *victim = NULL;
1714 unsigned long excess;
1715 unsigned long nr_scanned;
1716 struct mem_cgroup_reclaim_cookie reclaim = {
1720 excess = soft_limit_excess(root_memcg);
1723 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1728 * If we have not been able to reclaim
1729 * anything, it might because there are
1730 * no reclaimable pages under this hierarchy
1735 * We want to do more targeted reclaim.
1736 * excess >> 2 is not to excessive so as to
1737 * reclaim too much, nor too less that we keep
1738 * coming back to reclaim from this cgroup
1740 if (total >= (excess >> 2) ||
1741 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1746 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1747 pgdat, &nr_scanned);
1748 *total_scanned += nr_scanned;
1749 if (!soft_limit_excess(root_memcg))
1752 mem_cgroup_iter_break(root_memcg, victim);
1756 #ifdef CONFIG_LOCKDEP
1757 static struct lockdep_map memcg_oom_lock_dep_map = {
1758 .name = "memcg_oom_lock",
1762 static DEFINE_SPINLOCK(memcg_oom_lock);
1765 * Check OOM-Killer is already running under our hierarchy.
1766 * If someone is running, return false.
1768 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1770 struct mem_cgroup *iter, *failed = NULL;
1772 spin_lock(&memcg_oom_lock);
1774 for_each_mem_cgroup_tree(iter, memcg) {
1775 if (iter->oom_lock) {
1777 * this subtree of our hierarchy is already locked
1778 * so we cannot give a lock.
1781 mem_cgroup_iter_break(memcg, iter);
1784 iter->oom_lock = true;
1789 * OK, we failed to lock the whole subtree so we have
1790 * to clean up what we set up to the failing subtree
1792 for_each_mem_cgroup_tree(iter, memcg) {
1793 if (iter == failed) {
1794 mem_cgroup_iter_break(memcg, iter);
1797 iter->oom_lock = false;
1800 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1802 spin_unlock(&memcg_oom_lock);
1807 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1809 struct mem_cgroup *iter;
1811 spin_lock(&memcg_oom_lock);
1812 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1813 for_each_mem_cgroup_tree(iter, memcg)
1814 iter->oom_lock = false;
1815 spin_unlock(&memcg_oom_lock);
1818 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1820 struct mem_cgroup *iter;
1822 spin_lock(&memcg_oom_lock);
1823 for_each_mem_cgroup_tree(iter, memcg)
1825 spin_unlock(&memcg_oom_lock);
1828 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1830 struct mem_cgroup *iter;
1833 * Be careful about under_oom underflows because a child memcg
1834 * could have been added after mem_cgroup_mark_under_oom.
1836 spin_lock(&memcg_oom_lock);
1837 for_each_mem_cgroup_tree(iter, memcg)
1838 if (iter->under_oom > 0)
1840 spin_unlock(&memcg_oom_lock);
1843 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1845 struct oom_wait_info {
1846 struct mem_cgroup *memcg;
1847 wait_queue_entry_t wait;
1850 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1851 unsigned mode, int sync, void *arg)
1853 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1854 struct mem_cgroup *oom_wait_memcg;
1855 struct oom_wait_info *oom_wait_info;
1857 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1858 oom_wait_memcg = oom_wait_info->memcg;
1860 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1861 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1863 return autoremove_wake_function(wait, mode, sync, arg);
1866 static void memcg_oom_recover(struct mem_cgroup *memcg)
1869 * For the following lockless ->under_oom test, the only required
1870 * guarantee is that it must see the state asserted by an OOM when
1871 * this function is called as a result of userland actions
1872 * triggered by the notification of the OOM. This is trivially
1873 * achieved by invoking mem_cgroup_mark_under_oom() before
1874 * triggering notification.
1876 if (memcg && memcg->under_oom)
1877 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1887 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1889 enum oom_status ret;
1892 if (order > PAGE_ALLOC_COSTLY_ORDER)
1895 memcg_memory_event(memcg, MEMCG_OOM);
1898 * We are in the middle of the charge context here, so we
1899 * don't want to block when potentially sitting on a callstack
1900 * that holds all kinds of filesystem and mm locks.
1902 * cgroup1 allows disabling the OOM killer and waiting for outside
1903 * handling until the charge can succeed; remember the context and put
1904 * the task to sleep at the end of the page fault when all locks are
1907 * On the other hand, in-kernel OOM killer allows for an async victim
1908 * memory reclaim (oom_reaper) and that means that we are not solely
1909 * relying on the oom victim to make a forward progress and we can
1910 * invoke the oom killer here.
1912 * Please note that mem_cgroup_out_of_memory might fail to find a
1913 * victim and then we have to bail out from the charge path.
1915 if (memcg->oom_kill_disable) {
1916 if (!current->in_user_fault)
1918 css_get(&memcg->css);
1919 current->memcg_in_oom = memcg;
1920 current->memcg_oom_gfp_mask = mask;
1921 current->memcg_oom_order = order;
1926 mem_cgroup_mark_under_oom(memcg);
1928 locked = mem_cgroup_oom_trylock(memcg);
1931 mem_cgroup_oom_notify(memcg);
1933 mem_cgroup_unmark_under_oom(memcg);
1934 if (mem_cgroup_out_of_memory(memcg, mask, order))
1940 mem_cgroup_oom_unlock(memcg);
1946 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1947 * @handle: actually kill/wait or just clean up the OOM state
1949 * This has to be called at the end of a page fault if the memcg OOM
1950 * handler was enabled.
1952 * Memcg supports userspace OOM handling where failed allocations must
1953 * sleep on a waitqueue until the userspace task resolves the
1954 * situation. Sleeping directly in the charge context with all kinds
1955 * of locks held is not a good idea, instead we remember an OOM state
1956 * in the task and mem_cgroup_oom_synchronize() has to be called at
1957 * the end of the page fault to complete the OOM handling.
1959 * Returns %true if an ongoing memcg OOM situation was detected and
1960 * completed, %false otherwise.
1962 bool mem_cgroup_oom_synchronize(bool handle)
1964 struct mem_cgroup *memcg = current->memcg_in_oom;
1965 struct oom_wait_info owait;
1968 /* OOM is global, do not handle */
1975 owait.memcg = memcg;
1976 owait.wait.flags = 0;
1977 owait.wait.func = memcg_oom_wake_function;
1978 owait.wait.private = current;
1979 INIT_LIST_HEAD(&owait.wait.entry);
1981 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1982 mem_cgroup_mark_under_oom(memcg);
1984 locked = mem_cgroup_oom_trylock(memcg);
1987 mem_cgroup_oom_notify(memcg);
1989 if (locked && !memcg->oom_kill_disable) {
1990 mem_cgroup_unmark_under_oom(memcg);
1991 finish_wait(&memcg_oom_waitq, &owait.wait);
1992 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1993 current->memcg_oom_order);
1996 mem_cgroup_unmark_under_oom(memcg);
1997 finish_wait(&memcg_oom_waitq, &owait.wait);
2001 mem_cgroup_oom_unlock(memcg);
2003 * There is no guarantee that an OOM-lock contender
2004 * sees the wakeups triggered by the OOM kill
2005 * uncharges. Wake any sleepers explicitly.
2007 memcg_oom_recover(memcg);
2010 current->memcg_in_oom = NULL;
2011 css_put(&memcg->css);
2016 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2017 * @victim: task to be killed by the OOM killer
2018 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2020 * Returns a pointer to a memory cgroup, which has to be cleaned up
2021 * by killing all belonging OOM-killable tasks.
2023 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2025 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2026 struct mem_cgroup *oom_domain)
2028 struct mem_cgroup *oom_group = NULL;
2029 struct mem_cgroup *memcg;
2031 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2035 oom_domain = root_mem_cgroup;
2039 memcg = mem_cgroup_from_task(victim);
2040 if (memcg == root_mem_cgroup)
2044 * If the victim task has been asynchronously moved to a different
2045 * memory cgroup, we might end up killing tasks outside oom_domain.
2046 * In this case it's better to ignore memory.group.oom.
2048 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2052 * Traverse the memory cgroup hierarchy from the victim task's
2053 * cgroup up to the OOMing cgroup (or root) to find the
2054 * highest-level memory cgroup with oom.group set.
2056 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2057 if (memcg->oom_group)
2060 if (memcg == oom_domain)
2065 css_get(&oom_group->css);
2072 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2074 pr_info("Tasks in ");
2075 pr_cont_cgroup_path(memcg->css.cgroup);
2076 pr_cont(" are going to be killed due to memory.oom.group set\n");
2080 * lock_page_memcg - lock a page and memcg binding
2083 * This function protects unlocked LRU pages from being moved to
2086 * It ensures lifetime of the locked memcg. Caller is responsible
2087 * for the lifetime of the page.
2089 void lock_page_memcg(struct page *page)
2091 struct page *head = compound_head(page); /* rmap on tail pages */
2092 struct mem_cgroup *memcg;
2093 unsigned long flags;
2096 * The RCU lock is held throughout the transaction. The fast
2097 * path can get away without acquiring the memcg->move_lock
2098 * because page moving starts with an RCU grace period.
2102 if (mem_cgroup_disabled())
2105 memcg = page_memcg(head);
2106 if (unlikely(!memcg))
2109 #ifdef CONFIG_PROVE_LOCKING
2110 local_irq_save(flags);
2111 might_lock(&memcg->move_lock);
2112 local_irq_restore(flags);
2115 if (atomic_read(&memcg->moving_account) <= 0)
2118 spin_lock_irqsave(&memcg->move_lock, flags);
2119 if (memcg != page_memcg(head)) {
2120 spin_unlock_irqrestore(&memcg->move_lock, flags);
2125 * When charge migration first begins, we can have multiple
2126 * critical sections holding the fast-path RCU lock and one
2127 * holding the slowpath move_lock. Track the task who has the
2128 * move_lock for unlock_page_memcg().
2130 memcg->move_lock_task = current;
2131 memcg->move_lock_flags = flags;
2133 EXPORT_SYMBOL(lock_page_memcg);
2135 static void __unlock_page_memcg(struct mem_cgroup *memcg)
2137 if (memcg && memcg->move_lock_task == current) {
2138 unsigned long flags = memcg->move_lock_flags;
2140 memcg->move_lock_task = NULL;
2141 memcg->move_lock_flags = 0;
2143 spin_unlock_irqrestore(&memcg->move_lock, flags);
2150 * unlock_page_memcg - unlock a page and memcg binding
2153 void unlock_page_memcg(struct page *page)
2155 struct page *head = compound_head(page);
2157 __unlock_page_memcg(page_memcg(head));
2159 EXPORT_SYMBOL(unlock_page_memcg);
2161 struct memcg_stock_pcp {
2162 local_lock_t stock_lock;
2163 struct mem_cgroup *cached; /* this never be root cgroup */
2164 unsigned int nr_pages;
2166 #ifdef CONFIG_MEMCG_KMEM
2167 struct obj_cgroup *cached_objcg;
2168 struct pglist_data *cached_pgdat;
2169 unsigned int nr_bytes;
2170 int nr_slab_reclaimable_b;
2171 int nr_slab_unreclaimable_b;
2174 struct work_struct work;
2175 unsigned long flags;
2176 #define FLUSHING_CACHED_CHARGE 0
2178 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock) = {
2179 .stock_lock = INIT_LOCAL_LOCK(stock_lock),
2181 static DEFINE_MUTEX(percpu_charge_mutex);
2183 #ifdef CONFIG_MEMCG_KMEM
2184 static struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock);
2185 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2186 struct mem_cgroup *root_memcg);
2189 static inline struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock)
2193 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2194 struct mem_cgroup *root_memcg)
2201 * consume_stock: Try to consume stocked charge on this cpu.
2202 * @memcg: memcg to consume from.
2203 * @nr_pages: how many pages to charge.
2205 * The charges will only happen if @memcg matches the current cpu's memcg
2206 * stock, and at least @nr_pages are available in that stock. Failure to
2207 * service an allocation will refill the stock.
2209 * returns true if successful, false otherwise.
2211 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2213 struct memcg_stock_pcp *stock;
2214 unsigned long flags;
2217 if (nr_pages > MEMCG_CHARGE_BATCH)
2220 local_lock_irqsave(&memcg_stock.stock_lock, flags);
2222 stock = this_cpu_ptr(&memcg_stock);
2223 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2224 stock->nr_pages -= nr_pages;
2228 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2234 * Returns stocks cached in percpu and reset cached information.
2236 static void drain_stock(struct memcg_stock_pcp *stock)
2238 struct mem_cgroup *old = stock->cached;
2243 if (stock->nr_pages) {
2244 page_counter_uncharge(&old->memory, stock->nr_pages);
2245 if (do_memsw_account())
2246 page_counter_uncharge(&old->memsw, stock->nr_pages);
2247 stock->nr_pages = 0;
2251 stock->cached = NULL;
2254 static void drain_local_stock(struct work_struct *dummy)
2256 struct memcg_stock_pcp *stock;
2257 struct obj_cgroup *old = NULL;
2258 unsigned long flags;
2261 * The only protection from cpu hotplug (memcg_hotplug_cpu_dead) vs.
2262 * drain_stock races is that we always operate on local CPU stock
2263 * here with IRQ disabled
2265 local_lock_irqsave(&memcg_stock.stock_lock, flags);
2267 stock = this_cpu_ptr(&memcg_stock);
2268 old = drain_obj_stock(stock);
2270 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2272 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2274 obj_cgroup_put(old);
2278 * Cache charges(val) to local per_cpu area.
2279 * This will be consumed by consume_stock() function, later.
2281 static void __refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2283 struct memcg_stock_pcp *stock;
2285 stock = this_cpu_ptr(&memcg_stock);
2286 if (stock->cached != memcg) { /* reset if necessary */
2288 css_get(&memcg->css);
2289 stock->cached = memcg;
2291 stock->nr_pages += nr_pages;
2293 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2297 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2299 unsigned long flags;
2301 local_lock_irqsave(&memcg_stock.stock_lock, flags);
2302 __refill_stock(memcg, nr_pages);
2303 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2307 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2308 * of the hierarchy under it.
2310 static void drain_all_stock(struct mem_cgroup *root_memcg)
2314 /* If someone's already draining, avoid adding running more workers. */
2315 if (!mutex_trylock(&percpu_charge_mutex))
2318 * Notify other cpus that system-wide "drain" is running
2319 * We do not care about races with the cpu hotplug because cpu down
2320 * as well as workers from this path always operate on the local
2321 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2324 curcpu = smp_processor_id();
2325 for_each_online_cpu(cpu) {
2326 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2327 struct mem_cgroup *memcg;
2331 memcg = stock->cached;
2332 if (memcg && stock->nr_pages &&
2333 mem_cgroup_is_descendant(memcg, root_memcg))
2335 else if (obj_stock_flush_required(stock, root_memcg))
2340 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2342 drain_local_stock(&stock->work);
2344 schedule_work_on(cpu, &stock->work);
2348 mutex_unlock(&percpu_charge_mutex);
2351 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2353 struct memcg_stock_pcp *stock;
2355 stock = &per_cpu(memcg_stock, cpu);
2361 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2362 unsigned int nr_pages,
2365 unsigned long nr_reclaimed = 0;
2368 unsigned long pflags;
2370 if (page_counter_read(&memcg->memory) <=
2371 READ_ONCE(memcg->memory.high))
2374 memcg_memory_event(memcg, MEMCG_HIGH);
2376 psi_memstall_enter(&pflags);
2377 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2379 psi_memstall_leave(&pflags);
2380 } while ((memcg = parent_mem_cgroup(memcg)) &&
2381 !mem_cgroup_is_root(memcg));
2383 return nr_reclaimed;
2386 static void high_work_func(struct work_struct *work)
2388 struct mem_cgroup *memcg;
2390 memcg = container_of(work, struct mem_cgroup, high_work);
2391 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2395 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2396 * enough to still cause a significant slowdown in most cases, while still
2397 * allowing diagnostics and tracing to proceed without becoming stuck.
2399 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2402 * When calculating the delay, we use these either side of the exponentiation to
2403 * maintain precision and scale to a reasonable number of jiffies (see the table
2406 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2407 * overage ratio to a delay.
2408 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2409 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2410 * to produce a reasonable delay curve.
2412 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2413 * reasonable delay curve compared to precision-adjusted overage, not
2414 * penalising heavily at first, but still making sure that growth beyond the
2415 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2416 * example, with a high of 100 megabytes:
2418 * +-------+------------------------+
2419 * | usage | time to allocate in ms |
2420 * +-------+------------------------+
2442 * +-------+------------------------+
2444 #define MEMCG_DELAY_PRECISION_SHIFT 20
2445 #define MEMCG_DELAY_SCALING_SHIFT 14
2447 static u64 calculate_overage(unsigned long usage, unsigned long high)
2455 * Prevent division by 0 in overage calculation by acting as if
2456 * it was a threshold of 1 page
2458 high = max(high, 1UL);
2460 overage = usage - high;
2461 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2462 return div64_u64(overage, high);
2465 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2467 u64 overage, max_overage = 0;
2470 overage = calculate_overage(page_counter_read(&memcg->memory),
2471 READ_ONCE(memcg->memory.high));
2472 max_overage = max(overage, max_overage);
2473 } while ((memcg = parent_mem_cgroup(memcg)) &&
2474 !mem_cgroup_is_root(memcg));
2479 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2481 u64 overage, max_overage = 0;
2484 overage = calculate_overage(page_counter_read(&memcg->swap),
2485 READ_ONCE(memcg->swap.high));
2487 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2488 max_overage = max(overage, max_overage);
2489 } while ((memcg = parent_mem_cgroup(memcg)) &&
2490 !mem_cgroup_is_root(memcg));
2496 * Get the number of jiffies that we should penalise a mischievous cgroup which
2497 * is exceeding its memory.high by checking both it and its ancestors.
2499 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2500 unsigned int nr_pages,
2503 unsigned long penalty_jiffies;
2509 * We use overage compared to memory.high to calculate the number of
2510 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2511 * fairly lenient on small overages, and increasingly harsh when the
2512 * memcg in question makes it clear that it has no intention of stopping
2513 * its crazy behaviour, so we exponentially increase the delay based on
2516 penalty_jiffies = max_overage * max_overage * HZ;
2517 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2518 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2521 * Factor in the task's own contribution to the overage, such that four
2522 * N-sized allocations are throttled approximately the same as one
2523 * 4N-sized allocation.
2525 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2526 * larger the current charge patch is than that.
2528 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2532 * Scheduled by try_charge() to be executed from the userland return path
2533 * and reclaims memory over the high limit.
2535 void mem_cgroup_handle_over_high(void)
2537 unsigned long penalty_jiffies;
2538 unsigned long pflags;
2539 unsigned long nr_reclaimed;
2540 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2541 int nr_retries = MAX_RECLAIM_RETRIES;
2542 struct mem_cgroup *memcg;
2543 bool in_retry = false;
2545 if (likely(!nr_pages))
2548 memcg = get_mem_cgroup_from_mm(current->mm);
2549 current->memcg_nr_pages_over_high = 0;
2553 * The allocating task should reclaim at least the batch size, but for
2554 * subsequent retries we only want to do what's necessary to prevent oom
2555 * or breaching resource isolation.
2557 * This is distinct from memory.max or page allocator behaviour because
2558 * memory.high is currently batched, whereas memory.max and the page
2559 * allocator run every time an allocation is made.
2561 nr_reclaimed = reclaim_high(memcg,
2562 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2566 * memory.high is breached and reclaim is unable to keep up. Throttle
2567 * allocators proactively to slow down excessive growth.
2569 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2570 mem_find_max_overage(memcg));
2572 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2573 swap_find_max_overage(memcg));
2576 * Clamp the max delay per usermode return so as to still keep the
2577 * application moving forwards and also permit diagnostics, albeit
2580 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2583 * Don't sleep if the amount of jiffies this memcg owes us is so low
2584 * that it's not even worth doing, in an attempt to be nice to those who
2585 * go only a small amount over their memory.high value and maybe haven't
2586 * been aggressively reclaimed enough yet.
2588 if (penalty_jiffies <= HZ / 100)
2592 * If reclaim is making forward progress but we're still over
2593 * memory.high, we want to encourage that rather than doing allocator
2596 if (nr_reclaimed || nr_retries--) {
2602 * If we exit early, we're guaranteed to die (since
2603 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2604 * need to account for any ill-begotten jiffies to pay them off later.
2606 psi_memstall_enter(&pflags);
2607 schedule_timeout_killable(penalty_jiffies);
2608 psi_memstall_leave(&pflags);
2611 css_put(&memcg->css);
2614 static int try_charge_memcg(struct mem_cgroup *memcg, gfp_t gfp_mask,
2615 unsigned int nr_pages)
2617 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2618 int nr_retries = MAX_RECLAIM_RETRIES;
2619 struct mem_cgroup *mem_over_limit;
2620 struct page_counter *counter;
2621 enum oom_status oom_status;
2622 unsigned long nr_reclaimed;
2623 bool passed_oom = false;
2624 bool may_swap = true;
2625 bool drained = false;
2626 unsigned long pflags;
2629 if (consume_stock(memcg, nr_pages))
2632 if (!do_memsw_account() ||
2633 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2634 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2636 if (do_memsw_account())
2637 page_counter_uncharge(&memcg->memsw, batch);
2638 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2640 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2644 if (batch > nr_pages) {
2650 * Memcg doesn't have a dedicated reserve for atomic
2651 * allocations. But like the global atomic pool, we need to
2652 * put the burden of reclaim on regular allocation requests
2653 * and let these go through as privileged allocations.
2655 if (gfp_mask & __GFP_ATOMIC)
2659 * Prevent unbounded recursion when reclaim operations need to
2660 * allocate memory. This might exceed the limits temporarily,
2661 * but we prefer facilitating memory reclaim and getting back
2662 * under the limit over triggering OOM kills in these cases.
2664 if (unlikely(current->flags & PF_MEMALLOC))
2667 if (unlikely(task_in_memcg_oom(current)))
2670 if (!gfpflags_allow_blocking(gfp_mask))
2673 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2675 psi_memstall_enter(&pflags);
2676 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2677 gfp_mask, may_swap);
2678 psi_memstall_leave(&pflags);
2680 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2684 drain_all_stock(mem_over_limit);
2689 if (gfp_mask & __GFP_NORETRY)
2692 * Even though the limit is exceeded at this point, reclaim
2693 * may have been able to free some pages. Retry the charge
2694 * before killing the task.
2696 * Only for regular pages, though: huge pages are rather
2697 * unlikely to succeed so close to the limit, and we fall back
2698 * to regular pages anyway in case of failure.
2700 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2703 * At task move, charge accounts can be doubly counted. So, it's
2704 * better to wait until the end of task_move if something is going on.
2706 if (mem_cgroup_wait_acct_move(mem_over_limit))
2712 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2715 /* Avoid endless loop for tasks bypassed by the oom killer */
2716 if (passed_oom && task_is_dying())
2720 * keep retrying as long as the memcg oom killer is able to make
2721 * a forward progress or bypass the charge if the oom killer
2722 * couldn't make any progress.
2724 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2725 get_order(nr_pages * PAGE_SIZE));
2726 if (oom_status == OOM_SUCCESS) {
2728 nr_retries = MAX_RECLAIM_RETRIES;
2732 if (!(gfp_mask & __GFP_NOFAIL))
2736 * The allocation either can't fail or will lead to more memory
2737 * being freed very soon. Allow memory usage go over the limit
2738 * temporarily by force charging it.
2740 page_counter_charge(&memcg->memory, nr_pages);
2741 if (do_memsw_account())
2742 page_counter_charge(&memcg->memsw, nr_pages);
2747 if (batch > nr_pages)
2748 refill_stock(memcg, batch - nr_pages);
2751 * If the hierarchy is above the normal consumption range, schedule
2752 * reclaim on returning to userland. We can perform reclaim here
2753 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2754 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2755 * not recorded as it most likely matches current's and won't
2756 * change in the meantime. As high limit is checked again before
2757 * reclaim, the cost of mismatch is negligible.
2760 bool mem_high, swap_high;
2762 mem_high = page_counter_read(&memcg->memory) >
2763 READ_ONCE(memcg->memory.high);
2764 swap_high = page_counter_read(&memcg->swap) >
2765 READ_ONCE(memcg->swap.high);
2767 /* Don't bother a random interrupted task */
2768 if (in_interrupt()) {
2770 schedule_work(&memcg->high_work);
2776 if (mem_high || swap_high) {
2778 * The allocating tasks in this cgroup will need to do
2779 * reclaim or be throttled to prevent further growth
2780 * of the memory or swap footprints.
2782 * Target some best-effort fairness between the tasks,
2783 * and distribute reclaim work and delay penalties
2784 * based on how much each task is actually allocating.
2786 current->memcg_nr_pages_over_high += batch;
2787 set_notify_resume(current);
2790 } while ((memcg = parent_mem_cgroup(memcg)));
2795 static inline int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2796 unsigned int nr_pages)
2798 if (mem_cgroup_is_root(memcg))
2801 return try_charge_memcg(memcg, gfp_mask, nr_pages);
2804 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2805 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2807 if (mem_cgroup_is_root(memcg))
2810 page_counter_uncharge(&memcg->memory, nr_pages);
2811 if (do_memsw_account())
2812 page_counter_uncharge(&memcg->memsw, nr_pages);
2816 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2818 VM_BUG_ON_PAGE(page_memcg(page), page);
2820 * Any of the following ensures page's memcg stability:
2824 * - lock_page_memcg()
2825 * - exclusive reference
2827 page->memcg_data = (unsigned long)memcg;
2830 static struct mem_cgroup *get_mem_cgroup_from_objcg(struct obj_cgroup *objcg)
2832 struct mem_cgroup *memcg;
2836 memcg = obj_cgroup_memcg(objcg);
2837 if (unlikely(!css_tryget(&memcg->css)))
2844 #ifdef CONFIG_MEMCG_KMEM
2846 * The allocated objcg pointers array is not accounted directly.
2847 * Moreover, it should not come from DMA buffer and is not readily
2848 * reclaimable. So those GFP bits should be masked off.
2850 #define OBJCGS_CLEAR_MASK (__GFP_DMA | __GFP_RECLAIMABLE | __GFP_ACCOUNT)
2852 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2853 gfp_t gfp, bool new_page)
2855 unsigned int objects = objs_per_slab_page(s, page);
2856 unsigned long memcg_data;
2859 gfp &= ~OBJCGS_CLEAR_MASK;
2860 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2865 memcg_data = (unsigned long) vec | MEMCG_DATA_OBJCGS;
2868 * If the slab page is brand new and nobody can yet access
2869 * it's memcg_data, no synchronization is required and
2870 * memcg_data can be simply assigned.
2872 page->memcg_data = memcg_data;
2873 } else if (cmpxchg(&page->memcg_data, 0, memcg_data)) {
2875 * If the slab page is already in use, somebody can allocate
2876 * and assign obj_cgroups in parallel. In this case the existing
2877 * objcg vector should be reused.
2883 kmemleak_not_leak(vec);
2888 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2890 * A passed kernel object can be a slab object or a generic kernel page, so
2891 * different mechanisms for getting the memory cgroup pointer should be used.
2892 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2893 * can not know for sure how the kernel object is implemented.
2894 * mem_cgroup_from_obj() can be safely used in such cases.
2896 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2897 * cgroup_mutex, etc.
2899 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2903 if (mem_cgroup_disabled())
2906 page = virt_to_head_page(p);
2909 * Slab objects are accounted individually, not per-page.
2910 * Memcg membership data for each individual object is saved in
2911 * the page->obj_cgroups.
2913 if (page_objcgs_check(page)) {
2914 struct obj_cgroup *objcg;
2917 off = obj_to_index(page->slab_cache, page, p);
2918 objcg = page_objcgs(page)[off];
2920 return obj_cgroup_memcg(objcg);
2926 * page_memcg_check() is used here, because page_has_obj_cgroups()
2927 * check above could fail because the object cgroups vector wasn't set
2928 * at that moment, but it can be set concurrently.
2929 * page_memcg_check(page) will guarantee that a proper memory
2930 * cgroup pointer or NULL will be returned.
2932 return page_memcg_check(page);
2935 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2937 struct obj_cgroup *objcg = NULL;
2938 struct mem_cgroup *memcg;
2940 if (memcg_kmem_bypass())
2944 if (unlikely(active_memcg()))
2945 memcg = active_memcg();
2947 memcg = mem_cgroup_from_task(current);
2949 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2950 objcg = rcu_dereference(memcg->objcg);
2951 if (objcg && obj_cgroup_tryget(objcg))
2960 static int memcg_alloc_cache_id(void)
2965 id = ida_simple_get(&memcg_cache_ida,
2966 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2970 if (id < memcg_nr_cache_ids)
2974 * There's no space for the new id in memcg_caches arrays,
2975 * so we have to grow them.
2977 down_write(&memcg_cache_ids_sem);
2979 size = 2 * (id + 1);
2980 if (size < MEMCG_CACHES_MIN_SIZE)
2981 size = MEMCG_CACHES_MIN_SIZE;
2982 else if (size > MEMCG_CACHES_MAX_SIZE)
2983 size = MEMCG_CACHES_MAX_SIZE;
2985 err = memcg_update_all_list_lrus(size);
2987 memcg_nr_cache_ids = size;
2989 up_write(&memcg_cache_ids_sem);
2992 ida_simple_remove(&memcg_cache_ida, id);
2998 static void memcg_free_cache_id(int id)
3000 ida_simple_remove(&memcg_cache_ida, id);
3004 * obj_cgroup_uncharge_pages: uncharge a number of kernel pages from a objcg
3005 * @objcg: object cgroup to uncharge
3006 * @nr_pages: number of pages to uncharge
3008 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
3009 unsigned int nr_pages)
3011 struct mem_cgroup *memcg;
3013 memcg = get_mem_cgroup_from_objcg(objcg);
3015 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3016 page_counter_uncharge(&memcg->kmem, nr_pages);
3017 refill_stock(memcg, nr_pages);
3019 css_put(&memcg->css);
3023 * obj_cgroup_charge_pages: charge a number of kernel pages to a objcg
3024 * @objcg: object cgroup to charge
3025 * @gfp: reclaim mode
3026 * @nr_pages: number of pages to charge
3028 * Returns 0 on success, an error code on failure.
3030 static int obj_cgroup_charge_pages(struct obj_cgroup *objcg, gfp_t gfp,
3031 unsigned int nr_pages)
3033 struct page_counter *counter;
3034 struct mem_cgroup *memcg;
3037 memcg = get_mem_cgroup_from_objcg(objcg);
3039 ret = try_charge_memcg(memcg, gfp, nr_pages);
3043 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3044 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3047 * Enforce __GFP_NOFAIL allocation because callers are not
3048 * prepared to see failures and likely do not have any failure
3051 if (gfp & __GFP_NOFAIL) {
3052 page_counter_charge(&memcg->kmem, nr_pages);
3055 cancel_charge(memcg, nr_pages);
3059 css_put(&memcg->css);
3065 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3066 * @page: page to charge
3067 * @gfp: reclaim mode
3068 * @order: allocation order
3070 * Returns 0 on success, an error code on failure.
3072 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3074 struct obj_cgroup *objcg;
3077 objcg = get_obj_cgroup_from_current();
3079 ret = obj_cgroup_charge_pages(objcg, gfp, 1 << order);
3081 page->memcg_data = (unsigned long)objcg |
3085 obj_cgroup_put(objcg);
3091 * __memcg_kmem_uncharge_page: uncharge a kmem page
3092 * @page: page to uncharge
3093 * @order: allocation order
3095 void __memcg_kmem_uncharge_page(struct page *page, int order)
3097 struct obj_cgroup *objcg;
3098 unsigned int nr_pages = 1 << order;
3100 if (!PageMemcgKmem(page))
3103 objcg = __page_objcg(page);
3104 obj_cgroup_uncharge_pages(objcg, nr_pages);
3105 page->memcg_data = 0;
3106 obj_cgroup_put(objcg);
3109 void mod_objcg_state(struct obj_cgroup *objcg, struct pglist_data *pgdat,
3110 enum node_stat_item idx, int nr)
3112 struct memcg_stock_pcp *stock;
3113 struct obj_cgroup *old = NULL;
3114 unsigned long flags;
3117 local_lock_irqsave(&memcg_stock.stock_lock, flags);
3118 stock = this_cpu_ptr(&memcg_stock);
3121 * Save vmstat data in stock and skip vmstat array update unless
3122 * accumulating over a page of vmstat data or when pgdat or idx
3125 if (stock->cached_objcg != objcg) {
3126 old = drain_obj_stock(stock);
3127 obj_cgroup_get(objcg);
3128 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3129 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3130 stock->cached_objcg = objcg;
3131 stock->cached_pgdat = pgdat;
3132 } else if (stock->cached_pgdat != pgdat) {
3133 /* Flush the existing cached vmstat data */
3134 struct pglist_data *oldpg = stock->cached_pgdat;
3136 if (stock->nr_slab_reclaimable_b) {
3137 mod_objcg_mlstate(objcg, oldpg, NR_SLAB_RECLAIMABLE_B,
3138 stock->nr_slab_reclaimable_b);
3139 stock->nr_slab_reclaimable_b = 0;
3141 if (stock->nr_slab_unreclaimable_b) {
3142 mod_objcg_mlstate(objcg, oldpg, NR_SLAB_UNRECLAIMABLE_B,
3143 stock->nr_slab_unreclaimable_b);
3144 stock->nr_slab_unreclaimable_b = 0;
3146 stock->cached_pgdat = pgdat;
3149 bytes = (idx == NR_SLAB_RECLAIMABLE_B) ? &stock->nr_slab_reclaimable_b
3150 : &stock->nr_slab_unreclaimable_b;
3152 * Even for large object >= PAGE_SIZE, the vmstat data will still be
3153 * cached locally at least once before pushing it out.
3160 if (abs(*bytes) > PAGE_SIZE) {
3168 mod_objcg_mlstate(objcg, pgdat, idx, nr);
3170 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3172 obj_cgroup_put(old);
3175 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3177 struct memcg_stock_pcp *stock;
3178 unsigned long flags;
3181 local_lock_irqsave(&memcg_stock.stock_lock, flags);
3183 stock = this_cpu_ptr(&memcg_stock);
3184 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3185 stock->nr_bytes -= nr_bytes;
3189 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3194 static struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock)
3196 struct obj_cgroup *old = stock->cached_objcg;
3201 if (stock->nr_bytes) {
3202 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3203 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3206 struct mem_cgroup *memcg;
3208 memcg = get_mem_cgroup_from_objcg(old);
3210 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3211 page_counter_uncharge(&memcg->kmem, nr_pages);
3213 __refill_stock(memcg, nr_pages);
3215 css_put(&memcg->css);
3219 * The leftover is flushed to the centralized per-memcg value.
3220 * On the next attempt to refill obj stock it will be moved
3221 * to a per-cpu stock (probably, on an other CPU), see
3222 * refill_obj_stock().
3224 * How often it's flushed is a trade-off between the memory
3225 * limit enforcement accuracy and potential CPU contention,
3226 * so it might be changed in the future.
3228 atomic_add(nr_bytes, &old->nr_charged_bytes);
3229 stock->nr_bytes = 0;
3233 * Flush the vmstat data in current stock
3235 if (stock->nr_slab_reclaimable_b || stock->nr_slab_unreclaimable_b) {
3236 if (stock->nr_slab_reclaimable_b) {
3237 mod_objcg_mlstate(old, stock->cached_pgdat,
3238 NR_SLAB_RECLAIMABLE_B,
3239 stock->nr_slab_reclaimable_b);
3240 stock->nr_slab_reclaimable_b = 0;
3242 if (stock->nr_slab_unreclaimable_b) {
3243 mod_objcg_mlstate(old, stock->cached_pgdat,
3244 NR_SLAB_UNRECLAIMABLE_B,
3245 stock->nr_slab_unreclaimable_b);
3246 stock->nr_slab_unreclaimable_b = 0;
3248 stock->cached_pgdat = NULL;
3251 stock->cached_objcg = NULL;
3253 * The `old' objects needs to be released by the caller via
3254 * obj_cgroup_put() outside of memcg_stock_pcp::stock_lock.
3259 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3260 struct mem_cgroup *root_memcg)
3262 struct mem_cgroup *memcg;
3264 if (stock->cached_objcg) {
3265 memcg = obj_cgroup_memcg(stock->cached_objcg);
3266 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3273 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes,
3274 bool allow_uncharge)
3276 struct memcg_stock_pcp *stock;
3277 struct obj_cgroup *old = NULL;
3278 unsigned long flags;
3279 unsigned int nr_pages = 0;
3281 local_lock_irqsave(&memcg_stock.stock_lock, flags);
3283 stock = this_cpu_ptr(&memcg_stock);
3284 if (stock->cached_objcg != objcg) { /* reset if necessary */
3285 old = drain_obj_stock(stock);
3286 obj_cgroup_get(objcg);
3287 stock->cached_objcg = objcg;
3288 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3289 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3290 allow_uncharge = true; /* Allow uncharge when objcg changes */
3292 stock->nr_bytes += nr_bytes;
3294 if (allow_uncharge && (stock->nr_bytes > PAGE_SIZE)) {
3295 nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3296 stock->nr_bytes &= (PAGE_SIZE - 1);
3299 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3301 obj_cgroup_put(old);
3304 obj_cgroup_uncharge_pages(objcg, nr_pages);
3307 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3309 unsigned int nr_pages, nr_bytes;
3312 if (consume_obj_stock(objcg, size))
3316 * In theory, objcg->nr_charged_bytes can have enough
3317 * pre-charged bytes to satisfy the allocation. However,
3318 * flushing objcg->nr_charged_bytes requires two atomic
3319 * operations, and objcg->nr_charged_bytes can't be big.
3320 * The shared objcg->nr_charged_bytes can also become a
3321 * performance bottleneck if all tasks of the same memcg are
3322 * trying to update it. So it's better to ignore it and try
3323 * grab some new pages. The stock's nr_bytes will be flushed to
3324 * objcg->nr_charged_bytes later on when objcg changes.
3326 * The stock's nr_bytes may contain enough pre-charged bytes
3327 * to allow one less page from being charged, but we can't rely
3328 * on the pre-charged bytes not being changed outside of
3329 * consume_obj_stock() or refill_obj_stock(). So ignore those
3330 * pre-charged bytes as well when charging pages. To avoid a
3331 * page uncharge right after a page charge, we set the
3332 * allow_uncharge flag to false when calling refill_obj_stock()
3333 * to temporarily allow the pre-charged bytes to exceed the page
3334 * size limit. The maximum reachable value of the pre-charged
3335 * bytes is (sizeof(object) + PAGE_SIZE - 2) if there is no data
3338 nr_pages = size >> PAGE_SHIFT;
3339 nr_bytes = size & (PAGE_SIZE - 1);
3344 ret = obj_cgroup_charge_pages(objcg, gfp, nr_pages);
3345 if (!ret && nr_bytes)
3346 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes, false);
3351 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3353 refill_obj_stock(objcg, size, true);
3356 #endif /* CONFIG_MEMCG_KMEM */
3359 * Because page_memcg(head) is not set on tails, set it now.
3361 void split_page_memcg(struct page *head, unsigned int nr)
3363 struct mem_cgroup *memcg = page_memcg(head);
3366 if (mem_cgroup_disabled() || !memcg)
3369 for (i = 1; i < nr; i++)
3370 head[i].memcg_data = head->memcg_data;
3372 if (PageMemcgKmem(head))
3373 obj_cgroup_get_many(__page_objcg(head), nr - 1);
3375 css_get_many(&memcg->css, nr - 1);
3378 #ifdef CONFIG_MEMCG_SWAP
3380 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3381 * @entry: swap entry to be moved
3382 * @from: mem_cgroup which the entry is moved from
3383 * @to: mem_cgroup which the entry is moved to
3385 * It succeeds only when the swap_cgroup's record for this entry is the same
3386 * as the mem_cgroup's id of @from.
3388 * Returns 0 on success, -EINVAL on failure.
3390 * The caller must have charged to @to, IOW, called page_counter_charge() about
3391 * both res and memsw, and called css_get().
3393 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3394 struct mem_cgroup *from, struct mem_cgroup *to)
3396 unsigned short old_id, new_id;
3398 old_id = mem_cgroup_id(from);
3399 new_id = mem_cgroup_id(to);
3401 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3402 mod_memcg_state(from, MEMCG_SWAP, -1);
3403 mod_memcg_state(to, MEMCG_SWAP, 1);
3409 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3410 struct mem_cgroup *from, struct mem_cgroup *to)
3416 static DEFINE_MUTEX(memcg_max_mutex);
3418 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3419 unsigned long max, bool memsw)
3421 bool enlarge = false;
3422 bool drained = false;
3424 bool limits_invariant;
3425 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3428 if (signal_pending(current)) {
3433 mutex_lock(&memcg_max_mutex);
3435 * Make sure that the new limit (memsw or memory limit) doesn't
3436 * break our basic invariant rule memory.max <= memsw.max.
3438 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3439 max <= memcg->memsw.max;
3440 if (!limits_invariant) {
3441 mutex_unlock(&memcg_max_mutex);
3445 if (max > counter->max)
3447 ret = page_counter_set_max(counter, max);
3448 mutex_unlock(&memcg_max_mutex);
3454 drain_all_stock(memcg);
3459 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3460 GFP_KERNEL, !memsw)) {
3466 if (!ret && enlarge)
3467 memcg_oom_recover(memcg);
3472 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3474 unsigned long *total_scanned)
3476 unsigned long nr_reclaimed = 0;
3477 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3478 unsigned long reclaimed;
3480 struct mem_cgroup_tree_per_node *mctz;
3481 unsigned long excess;
3482 unsigned long nr_scanned;
3487 mctz = soft_limit_tree_node(pgdat->node_id);
3490 * Do not even bother to check the largest node if the root
3491 * is empty. Do it lockless to prevent lock bouncing. Races
3492 * are acceptable as soft limit is best effort anyway.
3494 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3498 * This loop can run a while, specially if mem_cgroup's continuously
3499 * keep exceeding their soft limit and putting the system under
3506 mz = mem_cgroup_largest_soft_limit_node(mctz);
3511 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3512 gfp_mask, &nr_scanned);
3513 nr_reclaimed += reclaimed;
3514 *total_scanned += nr_scanned;
3515 spin_lock_irq(&mctz->lock);
3516 __mem_cgroup_remove_exceeded(mz, mctz);
3519 * If we failed to reclaim anything from this memory cgroup
3520 * it is time to move on to the next cgroup
3524 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3526 excess = soft_limit_excess(mz->memcg);
3528 * One school of thought says that we should not add
3529 * back the node to the tree if reclaim returns 0.
3530 * But our reclaim could return 0, simply because due
3531 * to priority we are exposing a smaller subset of
3532 * memory to reclaim from. Consider this as a longer
3535 /* If excess == 0, no tree ops */
3536 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3537 spin_unlock_irq(&mctz->lock);
3538 css_put(&mz->memcg->css);
3541 * Could not reclaim anything and there are no more
3542 * mem cgroups to try or we seem to be looping without
3543 * reclaiming anything.
3545 if (!nr_reclaimed &&
3547 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3549 } while (!nr_reclaimed);
3551 css_put(&next_mz->memcg->css);
3552 return nr_reclaimed;
3556 * Reclaims as many pages from the given memcg as possible.
3558 * Caller is responsible for holding css reference for memcg.
3560 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3562 int nr_retries = MAX_RECLAIM_RETRIES;
3564 /* we call try-to-free pages for make this cgroup empty */
3565 lru_add_drain_all();
3567 drain_all_stock(memcg);
3569 /* try to free all pages in this cgroup */
3570 while (nr_retries && page_counter_read(&memcg->memory)) {
3573 if (signal_pending(current))
3576 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3580 /* maybe some writeback is necessary */
3581 congestion_wait(BLK_RW_ASYNC, HZ/10);
3589 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3590 char *buf, size_t nbytes,
3593 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3595 if (mem_cgroup_is_root(memcg))
3597 return mem_cgroup_force_empty(memcg) ?: nbytes;
3600 #ifdef CONFIG_MEMCG_SWAP
3601 static int mem_cgroup_force_reclaim(struct cgroup_subsys_state *css,
3602 struct cftype *cft, u64 val)
3604 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3605 unsigned long nr_to_reclaim = val;
3606 unsigned long total = 0;
3609 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
3610 total += try_to_free_mem_cgroup_pages(memcg, nr_to_reclaim,
3614 * If nothing was reclaimed after two attempts, there
3615 * may be no reclaimable pages in this hierarchy.
3616 * If more than nr_to_reclaim pages were already reclaimed,
3617 * finish force reclaim.
3619 if (loop && (!total || total > nr_to_reclaim))
3623 pr_info("%s: [Mem_reclaim] Loop: %d - Total_reclaimed: %lu - nr_to_reclaim: %lu\n",
3624 __func__, loop, total, nr_to_reclaim);
3630 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3636 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3637 struct cftype *cft, u64 val)
3642 pr_warn_once("Non-hierarchical mode is deprecated. "
3643 "Please report your usecase to linux-mm@kvack.org if you "
3644 "depend on this functionality.\n");
3649 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3653 if (mem_cgroup_is_root(memcg)) {
3654 mem_cgroup_flush_stats();
3655 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3656 memcg_page_state(memcg, NR_ANON_MAPPED);
3658 val += memcg_page_state(memcg, MEMCG_SWAP);
3661 val = page_counter_read(&memcg->memory);
3663 val = page_counter_read(&memcg->memsw);
3676 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3679 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3680 struct page_counter *counter;
3682 switch (MEMFILE_TYPE(cft->private)) {
3684 counter = &memcg->memory;
3687 counter = &memcg->memsw;
3690 counter = &memcg->kmem;
3693 counter = &memcg->tcpmem;
3699 switch (MEMFILE_ATTR(cft->private)) {
3701 if (counter == &memcg->memory)
3702 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3703 if (counter == &memcg->memsw)
3704 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3705 return (u64)page_counter_read(counter) * PAGE_SIZE;
3707 return (u64)counter->max * PAGE_SIZE;
3709 return (u64)counter->watermark * PAGE_SIZE;
3711 return counter->failcnt;
3712 case RES_SOFT_LIMIT:
3713 return (u64)memcg->soft_limit * PAGE_SIZE;
3719 #ifdef CONFIG_MEMCG_KMEM
3720 static int memcg_online_kmem(struct mem_cgroup *memcg)
3722 struct obj_cgroup *objcg;
3725 if (cgroup_memory_nokmem)
3728 BUG_ON(memcg->kmemcg_id >= 0);
3729 BUG_ON(memcg->kmem_state);
3731 memcg_id = memcg_alloc_cache_id();
3735 objcg = obj_cgroup_alloc();
3737 memcg_free_cache_id(memcg_id);
3740 objcg->memcg = memcg;
3741 rcu_assign_pointer(memcg->objcg, objcg);
3743 static_branch_enable(&memcg_kmem_enabled_key);
3745 memcg->kmemcg_id = memcg_id;
3746 memcg->kmem_state = KMEM_ONLINE;
3751 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3753 struct cgroup_subsys_state *css;
3754 struct mem_cgroup *parent, *child;
3757 if (memcg->kmem_state != KMEM_ONLINE)
3760 memcg->kmem_state = KMEM_ALLOCATED;
3762 parent = parent_mem_cgroup(memcg);
3764 parent = root_mem_cgroup;
3766 memcg_reparent_objcgs(memcg, parent);
3768 kmemcg_id = memcg->kmemcg_id;
3769 BUG_ON(kmemcg_id < 0);
3772 * Change kmemcg_id of this cgroup and all its descendants to the
3773 * parent's id, and then move all entries from this cgroup's list_lrus
3774 * to ones of the parent. After we have finished, all list_lrus
3775 * corresponding to this cgroup are guaranteed to remain empty. The
3776 * ordering is imposed by list_lru_node->lock taken by
3777 * memcg_drain_all_list_lrus().
3779 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3780 css_for_each_descendant_pre(css, &memcg->css) {
3781 child = mem_cgroup_from_css(css);
3782 BUG_ON(child->kmemcg_id != kmemcg_id);
3783 child->kmemcg_id = parent->kmemcg_id;
3787 memcg_drain_all_list_lrus(kmemcg_id, parent);
3789 memcg_free_cache_id(kmemcg_id);
3792 static void memcg_free_kmem(struct mem_cgroup *memcg)
3794 /* css_alloc() failed, offlining didn't happen */
3795 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3796 memcg_offline_kmem(memcg);
3799 static int memcg_online_kmem(struct mem_cgroup *memcg)
3803 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3806 static void memcg_free_kmem(struct mem_cgroup *memcg)
3809 #endif /* CONFIG_MEMCG_KMEM */
3811 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3816 mutex_lock(&memcg_max_mutex);
3817 ret = page_counter_set_max(&memcg->kmem, max);
3818 mutex_unlock(&memcg_max_mutex);
3822 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3826 mutex_lock(&memcg_max_mutex);
3828 ret = page_counter_set_max(&memcg->tcpmem, max);
3832 if (!memcg->tcpmem_active) {
3834 * The active flag needs to be written after the static_key
3835 * update. This is what guarantees that the socket activation
3836 * function is the last one to run. See mem_cgroup_sk_alloc()
3837 * for details, and note that we don't mark any socket as
3838 * belonging to this memcg until that flag is up.
3840 * We need to do this, because static_keys will span multiple
3841 * sites, but we can't control their order. If we mark a socket
3842 * as accounted, but the accounting functions are not patched in
3843 * yet, we'll lose accounting.
3845 * We never race with the readers in mem_cgroup_sk_alloc(),
3846 * because when this value change, the code to process it is not
3849 static_branch_inc(&memcg_sockets_enabled_key);
3850 memcg->tcpmem_active = true;
3853 mutex_unlock(&memcg_max_mutex);
3858 * The user of this function is...
3861 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3862 char *buf, size_t nbytes, loff_t off)
3864 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3865 unsigned long nr_pages;
3868 buf = strstrip(buf);
3869 ret = page_counter_memparse(buf, "-1", &nr_pages);
3873 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3875 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3879 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3881 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3884 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3887 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3888 "Please report your usecase to linux-mm@kvack.org if you "
3889 "depend on this functionality.\n");
3890 ret = memcg_update_kmem_max(memcg, nr_pages);
3893 ret = memcg_update_tcp_max(memcg, nr_pages);
3897 case RES_SOFT_LIMIT:
3898 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
3901 memcg->soft_limit = nr_pages;
3906 return ret ?: nbytes;
3909 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3910 size_t nbytes, loff_t off)
3912 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3913 struct page_counter *counter;
3915 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3917 counter = &memcg->memory;
3920 counter = &memcg->memsw;
3923 counter = &memcg->kmem;
3926 counter = &memcg->tcpmem;
3932 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3934 page_counter_reset_watermark(counter);
3937 counter->failcnt = 0;
3946 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3949 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3953 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3954 struct cftype *cft, u64 val)
3956 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3958 if (val & ~MOVE_MASK)
3962 * No kind of locking is needed in here, because ->can_attach() will
3963 * check this value once in the beginning of the process, and then carry
3964 * on with stale data. This means that changes to this value will only
3965 * affect task migrations starting after the change.
3967 memcg->move_charge_at_immigrate = val;
3971 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3972 struct cftype *cft, u64 val)
3980 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3981 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3982 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3984 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3985 int nid, unsigned int lru_mask, bool tree)
3987 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3988 unsigned long nr = 0;
3991 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3994 if (!(BIT(lru) & lru_mask))
3997 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3999 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
4004 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
4005 unsigned int lru_mask,
4008 unsigned long nr = 0;
4012 if (!(BIT(lru) & lru_mask))
4015 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
4017 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
4022 static int memcg_numa_stat_show(struct seq_file *m, void *v)
4026 unsigned int lru_mask;
4029 static const struct numa_stat stats[] = {
4030 { "total", LRU_ALL },
4031 { "file", LRU_ALL_FILE },
4032 { "anon", LRU_ALL_ANON },
4033 { "unevictable", BIT(LRU_UNEVICTABLE) },
4035 const struct numa_stat *stat;
4037 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4039 mem_cgroup_flush_stats();
4041 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4042 seq_printf(m, "%s=%lu", stat->name,
4043 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4045 for_each_node_state(nid, N_MEMORY)
4046 seq_printf(m, " N%d=%lu", nid,
4047 mem_cgroup_node_nr_lru_pages(memcg, nid,
4048 stat->lru_mask, false));
4052 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4054 seq_printf(m, "hierarchical_%s=%lu", stat->name,
4055 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4057 for_each_node_state(nid, N_MEMORY)
4058 seq_printf(m, " N%d=%lu", nid,
4059 mem_cgroup_node_nr_lru_pages(memcg, nid,
4060 stat->lru_mask, true));
4066 #endif /* CONFIG_NUMA */
4068 static const unsigned int memcg1_stats[] = {
4071 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4081 static const char *const memcg1_stat_names[] = {
4084 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4094 /* Universal VM events cgroup1 shows, original sort order */
4095 static const unsigned int memcg1_events[] = {
4102 static int memcg_stat_show(struct seq_file *m, void *v)
4104 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4105 unsigned long memory, memsw;
4106 struct mem_cgroup *mi;
4109 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4111 mem_cgroup_flush_stats();
4113 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4116 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4118 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4119 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4122 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4123 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4124 memcg_events_local(memcg, memcg1_events[i]));
4126 for (i = 0; i < NR_LRU_LISTS; i++)
4127 seq_printf(m, "%s %lu\n", lru_list_name(i),
4128 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4131 /* Hierarchical information */
4132 memory = memsw = PAGE_COUNTER_MAX;
4133 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4134 memory = min(memory, READ_ONCE(mi->memory.max));
4135 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4137 seq_printf(m, "hierarchical_memory_limit %llu\n",
4138 (u64)memory * PAGE_SIZE);
4139 if (do_memsw_account())
4140 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4141 (u64)memsw * PAGE_SIZE);
4143 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4146 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4148 nr = memcg_page_state(memcg, memcg1_stats[i]);
4149 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4150 (u64)nr * PAGE_SIZE);
4153 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4154 seq_printf(m, "total_%s %llu\n",
4155 vm_event_name(memcg1_events[i]),
4156 (u64)memcg_events(memcg, memcg1_events[i]));
4158 for (i = 0; i < NR_LRU_LISTS; i++)
4159 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4160 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4163 #ifdef CONFIG_DEBUG_VM
4166 struct mem_cgroup_per_node *mz;
4167 unsigned long anon_cost = 0;
4168 unsigned long file_cost = 0;
4170 for_each_online_pgdat(pgdat) {
4171 mz = memcg->nodeinfo[pgdat->node_id];
4173 anon_cost += mz->lruvec.anon_cost;
4174 file_cost += mz->lruvec.file_cost;
4176 seq_printf(m, "anon_cost %lu\n", anon_cost);
4177 seq_printf(m, "file_cost %lu\n", file_cost);
4184 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4187 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4189 return mem_cgroup_swappiness(memcg);
4192 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4193 struct cftype *cft, u64 val)
4195 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4200 if (!mem_cgroup_is_root(memcg))
4201 memcg->swappiness = val;
4203 vm_swappiness = val;
4208 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4210 struct mem_cgroup_threshold_ary *t;
4211 unsigned long usage;
4216 t = rcu_dereference(memcg->thresholds.primary);
4218 t = rcu_dereference(memcg->memsw_thresholds.primary);
4223 usage = mem_cgroup_usage(memcg, swap);
4226 * current_threshold points to threshold just below or equal to usage.
4227 * If it's not true, a threshold was crossed after last
4228 * call of __mem_cgroup_threshold().
4230 i = t->current_threshold;
4233 * Iterate backward over array of thresholds starting from
4234 * current_threshold and check if a threshold is crossed.
4235 * If none of thresholds below usage is crossed, we read
4236 * only one element of the array here.
4238 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4239 eventfd_signal(t->entries[i].eventfd, 1);
4241 /* i = current_threshold + 1 */
4245 * Iterate forward over array of thresholds starting from
4246 * current_threshold+1 and check if a threshold is crossed.
4247 * If none of thresholds above usage is crossed, we read
4248 * only one element of the array here.
4250 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4251 eventfd_signal(t->entries[i].eventfd, 1);
4253 /* Update current_threshold */
4254 t->current_threshold = i - 1;
4259 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4262 __mem_cgroup_threshold(memcg, false);
4263 if (do_memsw_account())
4264 __mem_cgroup_threshold(memcg, true);
4266 memcg = parent_mem_cgroup(memcg);
4270 static int compare_thresholds(const void *a, const void *b)
4272 const struct mem_cgroup_threshold *_a = a;
4273 const struct mem_cgroup_threshold *_b = b;
4275 if (_a->threshold > _b->threshold)
4278 if (_a->threshold < _b->threshold)
4284 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4286 struct mem_cgroup_eventfd_list *ev;
4288 spin_lock(&memcg_oom_lock);
4290 list_for_each_entry(ev, &memcg->oom_notify, list)
4291 eventfd_signal(ev->eventfd, 1);
4293 spin_unlock(&memcg_oom_lock);
4297 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4299 struct mem_cgroup *iter;
4301 for_each_mem_cgroup_tree(iter, memcg)
4302 mem_cgroup_oom_notify_cb(iter);
4305 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4306 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4308 struct mem_cgroup_thresholds *thresholds;
4309 struct mem_cgroup_threshold_ary *new;
4310 unsigned long threshold;
4311 unsigned long usage;
4314 ret = page_counter_memparse(args, "-1", &threshold);
4318 mutex_lock(&memcg->thresholds_lock);
4321 thresholds = &memcg->thresholds;
4322 usage = mem_cgroup_usage(memcg, false);
4323 } else if (type == _MEMSWAP) {
4324 thresholds = &memcg->memsw_thresholds;
4325 usage = mem_cgroup_usage(memcg, true);
4329 /* Check if a threshold crossed before adding a new one */
4330 if (thresholds->primary)
4331 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4333 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4335 /* Allocate memory for new array of thresholds */
4336 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4343 /* Copy thresholds (if any) to new array */
4344 if (thresholds->primary)
4345 memcpy(new->entries, thresholds->primary->entries,
4346 flex_array_size(new, entries, size - 1));
4348 /* Add new threshold */
4349 new->entries[size - 1].eventfd = eventfd;
4350 new->entries[size - 1].threshold = threshold;
4352 /* Sort thresholds. Registering of new threshold isn't time-critical */
4353 sort(new->entries, size, sizeof(*new->entries),
4354 compare_thresholds, NULL);
4356 /* Find current threshold */
4357 new->current_threshold = -1;
4358 for (i = 0; i < size; i++) {
4359 if (new->entries[i].threshold <= usage) {
4361 * new->current_threshold will not be used until
4362 * rcu_assign_pointer(), so it's safe to increment
4365 ++new->current_threshold;
4370 /* Free old spare buffer and save old primary buffer as spare */
4371 kfree(thresholds->spare);
4372 thresholds->spare = thresholds->primary;
4374 rcu_assign_pointer(thresholds->primary, new);
4376 /* To be sure that nobody uses thresholds */
4380 mutex_unlock(&memcg->thresholds_lock);
4385 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4386 struct eventfd_ctx *eventfd, const char *args)
4388 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4391 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4392 struct eventfd_ctx *eventfd, const char *args)
4394 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4397 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4398 struct eventfd_ctx *eventfd, enum res_type type)
4400 struct mem_cgroup_thresholds *thresholds;
4401 struct mem_cgroup_threshold_ary *new;
4402 unsigned long usage;
4403 int i, j, size, entries;
4405 mutex_lock(&memcg->thresholds_lock);
4408 thresholds = &memcg->thresholds;
4409 usage = mem_cgroup_usage(memcg, false);
4410 } else if (type == _MEMSWAP) {
4411 thresholds = &memcg->memsw_thresholds;
4412 usage = mem_cgroup_usage(memcg, true);
4416 if (!thresholds->primary)
4419 /* Check if a threshold crossed before removing */
4420 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4422 /* Calculate new number of threshold */
4424 for (i = 0; i < thresholds->primary->size; i++) {
4425 if (thresholds->primary->entries[i].eventfd != eventfd)
4431 new = thresholds->spare;
4433 /* If no items related to eventfd have been cleared, nothing to do */
4437 /* Set thresholds array to NULL if we don't have thresholds */
4446 /* Copy thresholds and find current threshold */
4447 new->current_threshold = -1;
4448 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4449 if (thresholds->primary->entries[i].eventfd == eventfd)
4452 new->entries[j] = thresholds->primary->entries[i];
4453 if (new->entries[j].threshold <= usage) {
4455 * new->current_threshold will not be used
4456 * until rcu_assign_pointer(), so it's safe to increment
4459 ++new->current_threshold;
4465 /* Swap primary and spare array */
4466 thresholds->spare = thresholds->primary;
4468 rcu_assign_pointer(thresholds->primary, new);
4470 /* To be sure that nobody uses thresholds */
4473 /* If all events are unregistered, free the spare array */
4475 kfree(thresholds->spare);
4476 thresholds->spare = NULL;
4479 mutex_unlock(&memcg->thresholds_lock);
4482 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4483 struct eventfd_ctx *eventfd)
4485 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4488 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4489 struct eventfd_ctx *eventfd)
4491 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4494 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4495 struct eventfd_ctx *eventfd, const char *args)
4497 struct mem_cgroup_eventfd_list *event;
4499 event = kmalloc(sizeof(*event), GFP_KERNEL);
4503 spin_lock(&memcg_oom_lock);
4505 event->eventfd = eventfd;
4506 list_add(&event->list, &memcg->oom_notify);
4508 /* already in OOM ? */
4509 if (memcg->under_oom)
4510 eventfd_signal(eventfd, 1);
4511 spin_unlock(&memcg_oom_lock);
4516 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4517 struct eventfd_ctx *eventfd)
4519 struct mem_cgroup_eventfd_list *ev, *tmp;
4521 spin_lock(&memcg_oom_lock);
4523 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4524 if (ev->eventfd == eventfd) {
4525 list_del(&ev->list);
4530 spin_unlock(&memcg_oom_lock);
4533 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4535 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4537 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4538 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4539 seq_printf(sf, "oom_kill %lu\n",
4540 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4544 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4545 struct cftype *cft, u64 val)
4547 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4549 /* cannot set to root cgroup and only 0 and 1 are allowed */
4550 if (mem_cgroup_is_root(memcg) || !((val == 0) || (val == 1)))
4553 memcg->oom_kill_disable = val;
4555 memcg_oom_recover(memcg);
4560 #ifdef CONFIG_CGROUP_WRITEBACK
4562 #include <trace/events/writeback.h>
4564 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4566 return wb_domain_init(&memcg->cgwb_domain, gfp);
4569 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4571 wb_domain_exit(&memcg->cgwb_domain);
4574 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4576 wb_domain_size_changed(&memcg->cgwb_domain);
4579 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4581 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4583 if (!memcg->css.parent)
4586 return &memcg->cgwb_domain;
4590 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4591 * @wb: bdi_writeback in question
4592 * @pfilepages: out parameter for number of file pages
4593 * @pheadroom: out parameter for number of allocatable pages according to memcg
4594 * @pdirty: out parameter for number of dirty pages
4595 * @pwriteback: out parameter for number of pages under writeback
4597 * Determine the numbers of file, headroom, dirty, and writeback pages in
4598 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4599 * is a bit more involved.
4601 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4602 * headroom is calculated as the lowest headroom of itself and the
4603 * ancestors. Note that this doesn't consider the actual amount of
4604 * available memory in the system. The caller should further cap
4605 * *@pheadroom accordingly.
4607 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4608 unsigned long *pheadroom, unsigned long *pdirty,
4609 unsigned long *pwriteback)
4611 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4612 struct mem_cgroup *parent;
4614 mem_cgroup_flush_stats();
4616 *pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
4617 *pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
4618 *pfilepages = memcg_page_state(memcg, NR_INACTIVE_FILE) +
4619 memcg_page_state(memcg, NR_ACTIVE_FILE);
4621 *pheadroom = PAGE_COUNTER_MAX;
4622 while ((parent = parent_mem_cgroup(memcg))) {
4623 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4624 READ_ONCE(memcg->memory.high));
4625 unsigned long used = page_counter_read(&memcg->memory);
4627 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4633 * Foreign dirty flushing
4635 * There's an inherent mismatch between memcg and writeback. The former
4636 * tracks ownership per-page while the latter per-inode. This was a
4637 * deliberate design decision because honoring per-page ownership in the
4638 * writeback path is complicated, may lead to higher CPU and IO overheads
4639 * and deemed unnecessary given that write-sharing an inode across
4640 * different cgroups isn't a common use-case.
4642 * Combined with inode majority-writer ownership switching, this works well
4643 * enough in most cases but there are some pathological cases. For
4644 * example, let's say there are two cgroups A and B which keep writing to
4645 * different but confined parts of the same inode. B owns the inode and
4646 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4647 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4648 * triggering background writeback. A will be slowed down without a way to
4649 * make writeback of the dirty pages happen.
4651 * Conditions like the above can lead to a cgroup getting repeatedly and
4652 * severely throttled after making some progress after each
4653 * dirty_expire_interval while the underlying IO device is almost
4656 * Solving this problem completely requires matching the ownership tracking
4657 * granularities between memcg and writeback in either direction. However,
4658 * the more egregious behaviors can be avoided by simply remembering the
4659 * most recent foreign dirtying events and initiating remote flushes on
4660 * them when local writeback isn't enough to keep the memory clean enough.
4662 * The following two functions implement such mechanism. When a foreign
4663 * page - a page whose memcg and writeback ownerships don't match - is
4664 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4665 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4666 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4667 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4668 * foreign bdi_writebacks which haven't expired. Both the numbers of
4669 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4670 * limited to MEMCG_CGWB_FRN_CNT.
4672 * The mechanism only remembers IDs and doesn't hold any object references.
4673 * As being wrong occasionally doesn't matter, updates and accesses to the
4674 * records are lockless and racy.
4676 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4677 struct bdi_writeback *wb)
4679 struct mem_cgroup *memcg = page_memcg(page);
4680 struct memcg_cgwb_frn *frn;
4681 u64 now = get_jiffies_64();
4682 u64 oldest_at = now;
4686 trace_track_foreign_dirty(page, wb);
4689 * Pick the slot to use. If there is already a slot for @wb, keep
4690 * using it. If not replace the oldest one which isn't being
4693 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4694 frn = &memcg->cgwb_frn[i];
4695 if (frn->bdi_id == wb->bdi->id &&
4696 frn->memcg_id == wb->memcg_css->id)
4698 if (time_before64(frn->at, oldest_at) &&
4699 atomic_read(&frn->done.cnt) == 1) {
4701 oldest_at = frn->at;
4705 if (i < MEMCG_CGWB_FRN_CNT) {
4707 * Re-using an existing one. Update timestamp lazily to
4708 * avoid making the cacheline hot. We want them to be
4709 * reasonably up-to-date and significantly shorter than
4710 * dirty_expire_interval as that's what expires the record.
4711 * Use the shorter of 1s and dirty_expire_interval / 8.
4713 unsigned long update_intv =
4714 min_t(unsigned long, HZ,
4715 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4717 if (time_before64(frn->at, now - update_intv))
4719 } else if (oldest >= 0) {
4720 /* replace the oldest free one */
4721 frn = &memcg->cgwb_frn[oldest];
4722 frn->bdi_id = wb->bdi->id;
4723 frn->memcg_id = wb->memcg_css->id;
4728 /* issue foreign writeback flushes for recorded foreign dirtying events */
4729 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4731 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4732 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4733 u64 now = jiffies_64;
4736 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4737 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4740 * If the record is older than dirty_expire_interval,
4741 * writeback on it has already started. No need to kick it
4742 * off again. Also, don't start a new one if there's
4743 * already one in flight.
4745 if (time_after64(frn->at, now - intv) &&
4746 atomic_read(&frn->done.cnt) == 1) {
4748 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4749 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id,
4750 WB_REASON_FOREIGN_FLUSH,
4756 #else /* CONFIG_CGROUP_WRITEBACK */
4758 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4763 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4767 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4771 #endif /* CONFIG_CGROUP_WRITEBACK */
4774 * DO NOT USE IN NEW FILES.
4776 * "cgroup.event_control" implementation.
4778 * This is way over-engineered. It tries to support fully configurable
4779 * events for each user. Such level of flexibility is completely
4780 * unnecessary especially in the light of the planned unified hierarchy.
4782 * Please deprecate this and replace with something simpler if at all
4787 * Unregister event and free resources.
4789 * Gets called from workqueue.
4791 static void memcg_event_remove(struct work_struct *work)
4793 struct mem_cgroup_event *event =
4794 container_of(work, struct mem_cgroup_event, remove);
4795 struct mem_cgroup *memcg = event->memcg;
4797 remove_wait_queue(event->wqh, &event->wait);
4799 event->unregister_event(memcg, event->eventfd);
4801 /* Notify userspace the event is going away. */
4802 eventfd_signal(event->eventfd, 1);
4804 eventfd_ctx_put(event->eventfd);
4806 css_put(&memcg->css);
4810 * Gets called on EPOLLHUP on eventfd when user closes it.
4812 * Called with wqh->lock held and interrupts disabled.
4814 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4815 int sync, void *key)
4817 struct mem_cgroup_event *event =
4818 container_of(wait, struct mem_cgroup_event, wait);
4819 struct mem_cgroup *memcg = event->memcg;
4820 __poll_t flags = key_to_poll(key);
4822 if (flags & EPOLLHUP) {
4824 * If the event has been detached at cgroup removal, we
4825 * can simply return knowing the other side will cleanup
4828 * We can't race against event freeing since the other
4829 * side will require wqh->lock via remove_wait_queue(),
4832 spin_lock(&memcg->event_list_lock);
4833 if (!list_empty(&event->list)) {
4834 list_del_init(&event->list);
4836 * We are in atomic context, but cgroup_event_remove()
4837 * may sleep, so we have to call it in workqueue.
4839 schedule_work(&event->remove);
4841 spin_unlock(&memcg->event_list_lock);
4847 static void memcg_event_ptable_queue_proc(struct file *file,
4848 wait_queue_head_t *wqh, poll_table *pt)
4850 struct mem_cgroup_event *event =
4851 container_of(pt, struct mem_cgroup_event, pt);
4854 add_wait_queue(wqh, &event->wait);
4858 * DO NOT USE IN NEW FILES.
4860 * Parse input and register new cgroup event handler.
4862 * Input must be in format '<event_fd> <control_fd> <args>'.
4863 * Interpretation of args is defined by control file implementation.
4865 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4866 char *buf, size_t nbytes, loff_t off)
4868 struct cgroup_subsys_state *css = of_css(of);
4869 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4870 struct mem_cgroup_event *event;
4871 struct cgroup_subsys_state *cfile_css;
4872 unsigned int efd, cfd;
4875 struct dentry *cdentry;
4880 if (IS_ENABLED(CONFIG_PREEMPT_RT))
4883 buf = strstrip(buf);
4885 efd = simple_strtoul(buf, &endp, 10);
4890 cfd = simple_strtoul(buf, &endp, 10);
4891 if ((*endp != ' ') && (*endp != '\0'))
4895 event = kzalloc(sizeof(*event), GFP_KERNEL);
4899 event->memcg = memcg;
4900 INIT_LIST_HEAD(&event->list);
4901 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4902 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4903 INIT_WORK(&event->remove, memcg_event_remove);
4911 event->eventfd = eventfd_ctx_fileget(efile.file);
4912 if (IS_ERR(event->eventfd)) {
4913 ret = PTR_ERR(event->eventfd);
4920 goto out_put_eventfd;
4923 /* the process need read permission on control file */
4924 /* AV: shouldn't we check that it's been opened for read instead? */
4925 ret = file_permission(cfile.file, MAY_READ);
4930 * The control file must be a regular cgroup1 file. As a regular cgroup
4931 * file can't be renamed, it's safe to access its name afterwards.
4933 cdentry = cfile.file->f_path.dentry;
4934 if (cdentry->d_sb->s_type != &cgroup_fs_type || !d_is_reg(cdentry)) {
4940 * Determine the event callbacks and set them in @event. This used
4941 * to be done via struct cftype but cgroup core no longer knows
4942 * about these events. The following is crude but the whole thing
4943 * is for compatibility anyway.
4945 * DO NOT ADD NEW FILES.
4947 name = cdentry->d_name.name;
4949 if (!strcmp(name, "memory.usage_in_bytes")) {
4950 event->register_event = mem_cgroup_usage_register_event;
4951 event->unregister_event = mem_cgroup_usage_unregister_event;
4952 } else if (!strcmp(name, "memory.oom_control")) {
4953 event->register_event = mem_cgroup_oom_register_event;
4954 event->unregister_event = mem_cgroup_oom_unregister_event;
4955 } else if (!strcmp(name, "memory.pressure_level")) {
4956 event->register_event = vmpressure_register_event;
4957 event->unregister_event = vmpressure_unregister_event;
4958 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4959 event->register_event = memsw_cgroup_usage_register_event;
4960 event->unregister_event = memsw_cgroup_usage_unregister_event;
4967 * Verify @cfile should belong to @css. Also, remaining events are
4968 * automatically removed on cgroup destruction but the removal is
4969 * asynchronous, so take an extra ref on @css.
4971 cfile_css = css_tryget_online_from_dir(cdentry->d_parent,
4972 &memory_cgrp_subsys);
4974 if (IS_ERR(cfile_css))
4976 if (cfile_css != css) {
4981 ret = event->register_event(memcg, event->eventfd, buf);
4985 vfs_poll(efile.file, &event->pt);
4987 spin_lock_irq(&memcg->event_list_lock);
4988 list_add(&event->list, &memcg->event_list);
4989 spin_unlock_irq(&memcg->event_list_lock);
5001 eventfd_ctx_put(event->eventfd);
5010 static struct cftype mem_cgroup_legacy_files[] = {
5012 .name = "usage_in_bytes",
5013 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5014 .read_u64 = mem_cgroup_read_u64,
5017 .name = "max_usage_in_bytes",
5018 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5019 .write = mem_cgroup_reset,
5020 .read_u64 = mem_cgroup_read_u64,
5023 .name = "limit_in_bytes",
5024 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5025 .write = mem_cgroup_write,
5026 .read_u64 = mem_cgroup_read_u64,
5029 .name = "soft_limit_in_bytes",
5030 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5031 .write = mem_cgroup_write,
5032 .read_u64 = mem_cgroup_read_u64,
5036 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5037 .write = mem_cgroup_reset,
5038 .read_u64 = mem_cgroup_read_u64,
5042 .seq_show = memcg_stat_show,
5045 .name = "force_empty",
5046 .write = mem_cgroup_force_empty_write,
5049 .name = "use_hierarchy",
5050 .write_u64 = mem_cgroup_hierarchy_write,
5051 .read_u64 = mem_cgroup_hierarchy_read,
5054 .name = "cgroup.event_control", /* XXX: for compat */
5055 .write = memcg_write_event_control,
5056 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
5059 .name = "swappiness",
5060 .read_u64 = mem_cgroup_swappiness_read,
5061 .write_u64 = mem_cgroup_swappiness_write,
5064 .name = "move_charge_at_immigrate",
5065 .read_u64 = mem_cgroup_move_charge_read,
5066 .write_u64 = mem_cgroup_move_charge_write,
5069 .name = "oom_control",
5070 .seq_show = mem_cgroup_oom_control_read,
5071 .write_u64 = mem_cgroup_oom_control_write,
5072 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5075 .name = "pressure_level",
5079 .name = "numa_stat",
5080 .seq_show = memcg_numa_stat_show,
5084 .name = "kmem.limit_in_bytes",
5085 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5086 .write = mem_cgroup_write,
5087 .read_u64 = mem_cgroup_read_u64,
5090 .name = "kmem.usage_in_bytes",
5091 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5092 .read_u64 = mem_cgroup_read_u64,
5095 .name = "kmem.failcnt",
5096 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5097 .write = mem_cgroup_reset,
5098 .read_u64 = mem_cgroup_read_u64,
5101 .name = "kmem.max_usage_in_bytes",
5102 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5103 .write = mem_cgroup_reset,
5104 .read_u64 = mem_cgroup_read_u64,
5106 #if defined(CONFIG_MEMCG_KMEM) && \
5107 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5109 .name = "kmem.slabinfo",
5110 .seq_show = memcg_slab_show,
5114 .name = "kmem.tcp.limit_in_bytes",
5115 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5116 .write = mem_cgroup_write,
5117 .read_u64 = mem_cgroup_read_u64,
5120 .name = "kmem.tcp.usage_in_bytes",
5121 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5122 .read_u64 = mem_cgroup_read_u64,
5125 .name = "kmem.tcp.failcnt",
5126 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5127 .write = mem_cgroup_reset,
5128 .read_u64 = mem_cgroup_read_u64,
5131 .name = "kmem.tcp.max_usage_in_bytes",
5132 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5133 .write = mem_cgroup_reset,
5134 .read_u64 = mem_cgroup_read_u64,
5136 { }, /* terminate */
5140 * Private memory cgroup IDR
5142 * Swap-out records and page cache shadow entries need to store memcg
5143 * references in constrained space, so we maintain an ID space that is
5144 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5145 * memory-controlled cgroups to 64k.
5147 * However, there usually are many references to the offline CSS after
5148 * the cgroup has been destroyed, such as page cache or reclaimable
5149 * slab objects, that don't need to hang on to the ID. We want to keep
5150 * those dead CSS from occupying IDs, or we might quickly exhaust the
5151 * relatively small ID space and prevent the creation of new cgroups
5152 * even when there are much fewer than 64k cgroups - possibly none.
5154 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5155 * be freed and recycled when it's no longer needed, which is usually
5156 * when the CSS is offlined.
5158 * The only exception to that are records of swapped out tmpfs/shmem
5159 * pages that need to be attributed to live ancestors on swapin. But
5160 * those references are manageable from userspace.
5163 static DEFINE_IDR(mem_cgroup_idr);
5165 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5167 if (memcg->id.id > 0) {
5168 idr_remove(&mem_cgroup_idr, memcg->id.id);
5173 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5176 refcount_add(n, &memcg->id.ref);
5179 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5181 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5182 mem_cgroup_id_remove(memcg);
5184 /* Memcg ID pins CSS */
5185 css_put(&memcg->css);
5189 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5191 mem_cgroup_id_put_many(memcg, 1);
5195 * mem_cgroup_from_id - look up a memcg from a memcg id
5196 * @id: the memcg id to look up
5198 * Caller must hold rcu_read_lock().
5200 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5202 WARN_ON_ONCE(!rcu_read_lock_held());
5203 return idr_find(&mem_cgroup_idr, id);
5206 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5208 struct mem_cgroup_per_node *pn;
5211 * This routine is called against possible nodes.
5212 * But it's BUG to call kmalloc() against offline node.
5214 * TODO: this routine can waste much memory for nodes which will
5215 * never be onlined. It's better to use memory hotplug callback
5218 if (!node_state(node, N_NORMAL_MEMORY))
5220 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5224 pn->lruvec_stats_percpu = alloc_percpu_gfp(struct lruvec_stats_percpu,
5225 GFP_KERNEL_ACCOUNT);
5226 if (!pn->lruvec_stats_percpu) {
5231 lruvec_init(&pn->lruvec);
5232 pn->usage_in_excess = 0;
5233 pn->on_tree = false;
5236 memcg->nodeinfo[node] = pn;
5240 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5242 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5247 free_percpu(pn->lruvec_stats_percpu);
5251 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5256 free_mem_cgroup_per_node_info(memcg, node);
5257 free_percpu(memcg->vmstats_percpu);
5261 static void mem_cgroup_free(struct mem_cgroup *memcg)
5263 memcg_wb_domain_exit(memcg);
5264 __mem_cgroup_free(memcg);
5267 static struct mem_cgroup *mem_cgroup_alloc(void)
5269 struct mem_cgroup *memcg;
5272 int __maybe_unused i;
5273 long error = -ENOMEM;
5275 size = sizeof(struct mem_cgroup);
5276 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5278 memcg = kzalloc(size, GFP_KERNEL);
5280 return ERR_PTR(error);
5282 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5283 1, MEM_CGROUP_ID_MAX,
5285 if (memcg->id.id < 0) {
5286 error = memcg->id.id;
5290 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5291 GFP_KERNEL_ACCOUNT);
5292 if (!memcg->vmstats_percpu)
5296 if (alloc_mem_cgroup_per_node_info(memcg, node))
5299 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5302 INIT_WORK(&memcg->high_work, high_work_func);
5303 INIT_LIST_HEAD(&memcg->oom_notify);
5304 mutex_init(&memcg->thresholds_lock);
5305 spin_lock_init(&memcg->move_lock);
5306 vmpressure_init(&memcg->vmpressure);
5307 INIT_LIST_HEAD(&memcg->event_list);
5308 spin_lock_init(&memcg->event_list_lock);
5309 memcg->socket_pressure = jiffies;
5310 #ifdef CONFIG_MEMCG_KMEM
5311 memcg->kmemcg_id = -1;
5312 INIT_LIST_HEAD(&memcg->objcg_list);
5314 #ifdef CONFIG_CGROUP_WRITEBACK
5315 INIT_LIST_HEAD(&memcg->cgwb_list);
5316 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5317 memcg->cgwb_frn[i].done =
5318 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5320 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5321 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5322 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5323 memcg->deferred_split_queue.split_queue_len = 0;
5325 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5328 mem_cgroup_id_remove(memcg);
5329 __mem_cgroup_free(memcg);
5330 return ERR_PTR(error);
5333 static struct cgroup_subsys_state * __ref
5334 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5336 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5337 struct mem_cgroup *memcg, *old_memcg;
5338 long error = -ENOMEM;
5340 old_memcg = set_active_memcg(parent);
5341 memcg = mem_cgroup_alloc();
5342 set_active_memcg(old_memcg);
5344 return ERR_CAST(memcg);
5346 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5347 memcg->soft_limit = PAGE_COUNTER_MAX;
5348 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5350 memcg->swappiness = mem_cgroup_swappiness(parent);
5351 memcg->oom_kill_disable = parent->oom_kill_disable;
5353 page_counter_init(&memcg->memory, &parent->memory);
5354 page_counter_init(&memcg->swap, &parent->swap);
5355 page_counter_init(&memcg->kmem, &parent->kmem);
5356 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5358 page_counter_init(&memcg->memory, NULL);
5359 page_counter_init(&memcg->swap, NULL);
5360 page_counter_init(&memcg->kmem, NULL);
5361 page_counter_init(&memcg->tcpmem, NULL);
5363 root_mem_cgroup = memcg;
5367 /* The following stuff does not apply to the root */
5368 error = memcg_online_kmem(memcg);
5372 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5373 static_branch_inc(&memcg_sockets_enabled_key);
5377 mem_cgroup_id_remove(memcg);
5378 mem_cgroup_free(memcg);
5379 return ERR_PTR(error);
5382 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5384 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5387 * A memcg must be visible for expand_shrinker_info()
5388 * by the time the maps are allocated. So, we allocate maps
5389 * here, when for_each_mem_cgroup() can't skip it.
5391 if (alloc_shrinker_info(memcg)) {
5392 mem_cgroup_id_remove(memcg);
5396 /* Online state pins memcg ID, memcg ID pins CSS */
5397 refcount_set(&memcg->id.ref, 1);
5400 if (unlikely(mem_cgroup_is_root(memcg)))
5401 queue_delayed_work(system_unbound_wq, &stats_flush_dwork,
5406 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5408 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5409 struct mem_cgroup_event *event, *tmp;
5412 * Unregister events and notify userspace.
5413 * Notify userspace about cgroup removing only after rmdir of cgroup
5414 * directory to avoid race between userspace and kernelspace.
5416 spin_lock_irq(&memcg->event_list_lock);
5417 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5418 list_del_init(&event->list);
5419 schedule_work(&event->remove);
5421 spin_unlock_irq(&memcg->event_list_lock);
5423 page_counter_set_min(&memcg->memory, 0);
5424 page_counter_set_low(&memcg->memory, 0);
5426 memcg_offline_kmem(memcg);
5427 reparent_shrinker_deferred(memcg);
5428 wb_memcg_offline(memcg);
5430 drain_all_stock(memcg);
5432 mem_cgroup_id_put(memcg);
5435 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5437 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5439 invalidate_reclaim_iterators(memcg);
5442 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5444 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5445 int __maybe_unused i;
5447 #ifdef CONFIG_CGROUP_WRITEBACK
5448 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5449 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5451 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5452 static_branch_dec(&memcg_sockets_enabled_key);
5454 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5455 static_branch_dec(&memcg_sockets_enabled_key);
5457 vmpressure_cleanup(&memcg->vmpressure);
5458 cancel_work_sync(&memcg->high_work);
5459 mem_cgroup_remove_from_trees(memcg);
5460 free_shrinker_info(memcg);
5461 memcg_free_kmem(memcg);
5462 mem_cgroup_free(memcg);
5466 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5467 * @css: the target css
5469 * Reset the states of the mem_cgroup associated with @css. This is
5470 * invoked when the userland requests disabling on the default hierarchy
5471 * but the memcg is pinned through dependency. The memcg should stop
5472 * applying policies and should revert to the vanilla state as it may be
5473 * made visible again.
5475 * The current implementation only resets the essential configurations.
5476 * This needs to be expanded to cover all the visible parts.
5478 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5480 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5482 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5483 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5484 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5485 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5486 page_counter_set_min(&memcg->memory, 0);
5487 page_counter_set_low(&memcg->memory, 0);
5488 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5489 memcg->soft_limit = PAGE_COUNTER_MAX;
5490 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5491 memcg_wb_domain_size_changed(memcg);
5494 static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state *css, int cpu)
5496 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5497 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5498 struct memcg_vmstats_percpu *statc;
5502 statc = per_cpu_ptr(memcg->vmstats_percpu, cpu);
5504 for (i = 0; i < MEMCG_NR_STAT; i++) {
5506 * Collect the aggregated propagation counts of groups
5507 * below us. We're in a per-cpu loop here and this is
5508 * a global counter, so the first cycle will get them.
5510 delta = memcg->vmstats.state_pending[i];
5512 memcg->vmstats.state_pending[i] = 0;
5514 /* Add CPU changes on this level since the last flush */
5515 v = READ_ONCE(statc->state[i]);
5516 if (v != statc->state_prev[i]) {
5517 delta += v - statc->state_prev[i];
5518 statc->state_prev[i] = v;
5524 /* Aggregate counts on this level and propagate upwards */
5525 memcg->vmstats.state[i] += delta;
5527 parent->vmstats.state_pending[i] += delta;
5530 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
5531 delta = memcg->vmstats.events_pending[i];
5533 memcg->vmstats.events_pending[i] = 0;
5535 v = READ_ONCE(statc->events[i]);
5536 if (v != statc->events_prev[i]) {
5537 delta += v - statc->events_prev[i];
5538 statc->events_prev[i] = v;
5544 memcg->vmstats.events[i] += delta;
5546 parent->vmstats.events_pending[i] += delta;
5549 for_each_node_state(nid, N_MEMORY) {
5550 struct mem_cgroup_per_node *pn = memcg->nodeinfo[nid];
5551 struct mem_cgroup_per_node *ppn = NULL;
5552 struct lruvec_stats_percpu *lstatc;
5555 ppn = parent->nodeinfo[nid];
5557 lstatc = per_cpu_ptr(pn->lruvec_stats_percpu, cpu);
5559 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
5560 delta = pn->lruvec_stats.state_pending[i];
5562 pn->lruvec_stats.state_pending[i] = 0;
5564 v = READ_ONCE(lstatc->state[i]);
5565 if (v != lstatc->state_prev[i]) {
5566 delta += v - lstatc->state_prev[i];
5567 lstatc->state_prev[i] = v;
5573 pn->lruvec_stats.state[i] += delta;
5575 ppn->lruvec_stats.state_pending[i] += delta;
5581 /* Handlers for move charge at task migration. */
5582 static int mem_cgroup_do_precharge(unsigned long count)
5586 /* Try a single bulk charge without reclaim first, kswapd may wake */
5587 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5589 mc.precharge += count;
5593 /* Try charges one by one with reclaim, but do not retry */
5595 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5609 enum mc_target_type {
5616 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5617 unsigned long addr, pte_t ptent)
5619 struct page *page = vm_normal_page(vma, addr, ptent);
5621 if (!page || !page_mapped(page))
5623 if (PageAnon(page)) {
5624 if (!(mc.flags & MOVE_ANON))
5627 if (!(mc.flags & MOVE_FILE))
5630 if (!get_page_unless_zero(page))
5636 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5637 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5638 pte_t ptent, swp_entry_t *entry)
5640 struct page *page = NULL;
5641 swp_entry_t ent = pte_to_swp_entry(ptent);
5643 if (!(mc.flags & MOVE_ANON))
5647 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5648 * a device and because they are not accessible by CPU they are store
5649 * as special swap entry in the CPU page table.
5651 if (is_device_private_entry(ent)) {
5652 page = pfn_swap_entry_to_page(ent);
5654 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5655 * a refcount of 1 when free (unlike normal page)
5657 if (!page_ref_add_unless(page, 1, 1))
5662 if (non_swap_entry(ent))
5666 * Because lookup_swap_cache() updates some statistics counter,
5667 * we call find_get_page() with swapper_space directly.
5669 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5670 entry->val = ent.val;
5675 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5676 pte_t ptent, swp_entry_t *entry)
5682 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5683 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5685 if (!vma->vm_file) /* anonymous vma */
5687 if (!(mc.flags & MOVE_FILE))
5690 /* page is moved even if it's not RSS of this task(page-faulted). */
5691 /* shmem/tmpfs may report page out on swap: account for that too. */
5692 return find_get_incore_page(vma->vm_file->f_mapping,
5693 linear_page_index(vma, addr));
5697 * mem_cgroup_move_account - move account of the page
5699 * @compound: charge the page as compound or small page
5700 * @from: mem_cgroup which the page is moved from.
5701 * @to: mem_cgroup which the page is moved to. @from != @to.
5703 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5705 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5708 static int mem_cgroup_move_account(struct page *page,
5710 struct mem_cgroup *from,
5711 struct mem_cgroup *to)
5713 struct lruvec *from_vec, *to_vec;
5714 struct pglist_data *pgdat;
5715 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5718 VM_BUG_ON(from == to);
5719 VM_BUG_ON_PAGE(PageLRU(page), page);
5720 VM_BUG_ON(compound && !PageTransHuge(page));
5723 * Prevent mem_cgroup_migrate() from looking at
5724 * page's memory cgroup of its source page while we change it.
5727 if (!trylock_page(page))
5731 if (page_memcg(page) != from)
5734 pgdat = page_pgdat(page);
5735 from_vec = mem_cgroup_lruvec(from, pgdat);
5736 to_vec = mem_cgroup_lruvec(to, pgdat);
5738 lock_page_memcg(page);
5740 if (PageAnon(page)) {
5741 if (page_mapped(page)) {
5742 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5743 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5744 if (PageTransHuge(page)) {
5745 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5747 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5752 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5753 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5755 if (PageSwapBacked(page)) {
5756 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5757 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5760 if (page_mapped(page)) {
5761 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5762 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5765 if (PageDirty(page)) {
5766 struct address_space *mapping = page_mapping(page);
5768 if (mapping_can_writeback(mapping)) {
5769 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5771 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5777 if (PageWriteback(page)) {
5778 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5779 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5783 * All state has been migrated, let's switch to the new memcg.
5785 * It is safe to change page's memcg here because the page
5786 * is referenced, charged, isolated, and locked: we can't race
5787 * with (un)charging, migration, LRU putback, or anything else
5788 * that would rely on a stable page's memory cgroup.
5790 * Note that lock_page_memcg is a memcg lock, not a page lock,
5791 * to save space. As soon as we switch page's memory cgroup to a
5792 * new memcg that isn't locked, the above state can change
5793 * concurrently again. Make sure we're truly done with it.
5798 css_put(&from->css);
5800 page->memcg_data = (unsigned long)to;
5802 __unlock_page_memcg(from);
5806 local_irq_disable();
5807 mem_cgroup_charge_statistics(to, page, nr_pages);
5808 memcg_check_events(to, page);
5809 mem_cgroup_charge_statistics(from, page, -nr_pages);
5810 memcg_check_events(from, page);
5819 * get_mctgt_type - get target type of moving charge
5820 * @vma: the vma the pte to be checked belongs
5821 * @addr: the address corresponding to the pte to be checked
5822 * @ptent: the pte to be checked
5823 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5826 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5827 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5828 * move charge. if @target is not NULL, the page is stored in target->page
5829 * with extra refcnt got(Callers should handle it).
5830 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5831 * target for charge migration. if @target is not NULL, the entry is stored
5833 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5834 * (so ZONE_DEVICE page and thus not on the lru).
5835 * For now we such page is charge like a regular page would be as for all
5836 * intent and purposes it is just special memory taking the place of a
5839 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5841 * Called with pte lock held.
5844 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5845 unsigned long addr, pte_t ptent, union mc_target *target)
5847 struct page *page = NULL;
5848 enum mc_target_type ret = MC_TARGET_NONE;
5849 swp_entry_t ent = { .val = 0 };
5851 if (pte_present(ptent))
5852 page = mc_handle_present_pte(vma, addr, ptent);
5853 else if (is_swap_pte(ptent))
5854 page = mc_handle_swap_pte(vma, ptent, &ent);
5855 else if (pte_none(ptent))
5856 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5858 if (!page && !ent.val)
5862 * Do only loose check w/o serialization.
5863 * mem_cgroup_move_account() checks the page is valid or
5864 * not under LRU exclusion.
5866 if (page_memcg(page) == mc.from) {
5867 ret = MC_TARGET_PAGE;
5868 if (is_device_private_page(page))
5869 ret = MC_TARGET_DEVICE;
5871 target->page = page;
5873 if (!ret || !target)
5877 * There is a swap entry and a page doesn't exist or isn't charged.
5878 * But we cannot move a tail-page in a THP.
5880 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5881 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5882 ret = MC_TARGET_SWAP;
5889 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5891 * We don't consider PMD mapped swapping or file mapped pages because THP does
5892 * not support them for now.
5893 * Caller should make sure that pmd_trans_huge(pmd) is true.
5895 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5896 unsigned long addr, pmd_t pmd, union mc_target *target)
5898 struct page *page = NULL;
5899 enum mc_target_type ret = MC_TARGET_NONE;
5901 if (unlikely(is_swap_pmd(pmd))) {
5902 VM_BUG_ON(thp_migration_supported() &&
5903 !is_pmd_migration_entry(pmd));
5906 page = pmd_page(pmd);
5907 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5908 if (!(mc.flags & MOVE_ANON))
5910 if (page_memcg(page) == mc.from) {
5911 ret = MC_TARGET_PAGE;
5914 target->page = page;
5920 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5921 unsigned long addr, pmd_t pmd, union mc_target *target)
5923 return MC_TARGET_NONE;
5927 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5928 unsigned long addr, unsigned long end,
5929 struct mm_walk *walk)
5931 struct vm_area_struct *vma = walk->vma;
5935 ptl = pmd_trans_huge_lock(pmd, vma);
5938 * Note their can not be MC_TARGET_DEVICE for now as we do not
5939 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5940 * this might change.
5942 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5943 mc.precharge += HPAGE_PMD_NR;
5948 if (pmd_trans_unstable(pmd))
5950 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5951 for (; addr != end; pte++, addr += PAGE_SIZE)
5952 if (get_mctgt_type(vma, addr, *pte, NULL))
5953 mc.precharge++; /* increment precharge temporarily */
5954 pte_unmap_unlock(pte - 1, ptl);
5960 static const struct mm_walk_ops precharge_walk_ops = {
5961 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5964 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5966 unsigned long precharge;
5969 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5970 mmap_read_unlock(mm);
5972 precharge = mc.precharge;
5978 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5980 unsigned long precharge = mem_cgroup_count_precharge(mm);
5982 VM_BUG_ON(mc.moving_task);
5983 mc.moving_task = current;
5984 return mem_cgroup_do_precharge(precharge);
5987 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5988 static void __mem_cgroup_clear_mc(void)
5990 struct mem_cgroup *from = mc.from;
5991 struct mem_cgroup *to = mc.to;
5993 /* we must uncharge all the leftover precharges from mc.to */
5995 cancel_charge(mc.to, mc.precharge);
5999 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6000 * we must uncharge here.
6002 if (mc.moved_charge) {
6003 cancel_charge(mc.from, mc.moved_charge);
6004 mc.moved_charge = 0;
6006 /* we must fixup refcnts and charges */
6007 if (mc.moved_swap) {
6008 /* uncharge swap account from the old cgroup */
6009 if (!mem_cgroup_is_root(mc.from))
6010 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
6012 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
6015 * we charged both to->memory and to->memsw, so we
6016 * should uncharge to->memory.
6018 if (!mem_cgroup_is_root(mc.to))
6019 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
6023 memcg_oom_recover(from);
6024 memcg_oom_recover(to);
6025 wake_up_all(&mc.waitq);
6028 static void mem_cgroup_clear_mc(void)
6030 struct mm_struct *mm = mc.mm;
6033 * we must clear moving_task before waking up waiters at the end of
6036 mc.moving_task = NULL;
6037 __mem_cgroup_clear_mc();
6038 spin_lock(&mc.lock);
6042 spin_unlock(&mc.lock);
6047 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6049 struct cgroup_subsys_state *css;
6050 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
6051 struct mem_cgroup *from;
6052 struct task_struct *leader, *p;
6053 struct mm_struct *mm;
6054 unsigned long move_flags;
6057 /* charge immigration isn't supported on the default hierarchy */
6058 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6062 * Multi-process migrations only happen on the default hierarchy
6063 * where charge immigration is not used. Perform charge
6064 * immigration if @tset contains a leader and whine if there are
6068 cgroup_taskset_for_each_leader(leader, css, tset) {
6071 memcg = mem_cgroup_from_css(css);
6077 * We are now committed to this value whatever it is. Changes in this
6078 * tunable will only affect upcoming migrations, not the current one.
6079 * So we need to save it, and keep it going.
6081 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
6085 from = mem_cgroup_from_task(p);
6087 VM_BUG_ON(from == memcg);
6089 mm = get_task_mm(p);
6092 /* We move charges only when we move a owner of the mm */
6093 if (mm->owner == p) {
6096 VM_BUG_ON(mc.precharge);
6097 VM_BUG_ON(mc.moved_charge);
6098 VM_BUG_ON(mc.moved_swap);
6100 spin_lock(&mc.lock);
6104 mc.flags = move_flags;
6105 spin_unlock(&mc.lock);
6106 /* We set mc.moving_task later */
6108 ret = mem_cgroup_precharge_mc(mm);
6110 mem_cgroup_clear_mc();
6117 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6120 mem_cgroup_clear_mc();
6123 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6124 unsigned long addr, unsigned long end,
6125 struct mm_walk *walk)
6128 struct vm_area_struct *vma = walk->vma;
6131 enum mc_target_type target_type;
6132 union mc_target target;
6135 ptl = pmd_trans_huge_lock(pmd, vma);
6137 if (mc.precharge < HPAGE_PMD_NR) {
6141 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6142 if (target_type == MC_TARGET_PAGE) {
6144 if (!isolate_lru_page(page)) {
6145 if (!mem_cgroup_move_account(page, true,
6147 mc.precharge -= HPAGE_PMD_NR;
6148 mc.moved_charge += HPAGE_PMD_NR;
6150 putback_lru_page(page);
6153 } else if (target_type == MC_TARGET_DEVICE) {
6155 if (!mem_cgroup_move_account(page, true,
6157 mc.precharge -= HPAGE_PMD_NR;
6158 mc.moved_charge += HPAGE_PMD_NR;
6166 if (pmd_trans_unstable(pmd))
6169 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6170 for (; addr != end; addr += PAGE_SIZE) {
6171 pte_t ptent = *(pte++);
6172 bool device = false;
6178 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6179 case MC_TARGET_DEVICE:
6182 case MC_TARGET_PAGE:
6185 * We can have a part of the split pmd here. Moving it
6186 * can be done but it would be too convoluted so simply
6187 * ignore such a partial THP and keep it in original
6188 * memcg. There should be somebody mapping the head.
6190 if (PageTransCompound(page))
6192 if (!device && isolate_lru_page(page))
6194 if (!mem_cgroup_move_account(page, false,
6197 /* we uncharge from mc.from later. */
6201 putback_lru_page(page);
6202 put: /* get_mctgt_type() gets the page */
6205 case MC_TARGET_SWAP:
6207 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6209 mem_cgroup_id_get_many(mc.to, 1);
6210 /* we fixup other refcnts and charges later. */
6218 pte_unmap_unlock(pte - 1, ptl);
6223 * We have consumed all precharges we got in can_attach().
6224 * We try charge one by one, but don't do any additional
6225 * charges to mc.to if we have failed in charge once in attach()
6228 ret = mem_cgroup_do_precharge(1);
6236 static const struct mm_walk_ops charge_walk_ops = {
6237 .pmd_entry = mem_cgroup_move_charge_pte_range,
6240 static void mem_cgroup_move_charge(void)
6242 lru_add_drain_all();
6244 * Signal lock_page_memcg() to take the memcg's move_lock
6245 * while we're moving its pages to another memcg. Then wait
6246 * for already started RCU-only updates to finish.
6248 atomic_inc(&mc.from->moving_account);
6251 if (unlikely(!mmap_read_trylock(mc.mm))) {
6253 * Someone who are holding the mmap_lock might be waiting in
6254 * waitq. So we cancel all extra charges, wake up all waiters,
6255 * and retry. Because we cancel precharges, we might not be able
6256 * to move enough charges, but moving charge is a best-effort
6257 * feature anyway, so it wouldn't be a big problem.
6259 __mem_cgroup_clear_mc();
6264 * When we have consumed all precharges and failed in doing
6265 * additional charge, the page walk just aborts.
6267 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6270 mmap_read_unlock(mc.mm);
6271 atomic_dec(&mc.from->moving_account);
6274 static void mem_cgroup_move_task(void)
6277 mem_cgroup_move_charge();
6278 mem_cgroup_clear_mc();
6281 #else /* !CONFIG_MMU */
6282 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6286 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6289 static void mem_cgroup_move_task(void)
6294 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6296 if (value == PAGE_COUNTER_MAX)
6297 seq_puts(m, "max\n");
6299 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6304 static u64 memory_current_read(struct cgroup_subsys_state *css,
6307 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6309 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6312 static int memory_min_show(struct seq_file *m, void *v)
6314 return seq_puts_memcg_tunable(m,
6315 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6318 static ssize_t memory_min_write(struct kernfs_open_file *of,
6319 char *buf, size_t nbytes, loff_t off)
6321 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6325 buf = strstrip(buf);
6326 err = page_counter_memparse(buf, "max", &min);
6330 page_counter_set_min(&memcg->memory, min);
6335 static int memory_low_show(struct seq_file *m, void *v)
6337 return seq_puts_memcg_tunable(m,
6338 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6341 static ssize_t memory_low_write(struct kernfs_open_file *of,
6342 char *buf, size_t nbytes, loff_t off)
6344 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6348 buf = strstrip(buf);
6349 err = page_counter_memparse(buf, "max", &low);
6353 page_counter_set_low(&memcg->memory, low);
6358 static int memory_high_show(struct seq_file *m, void *v)
6360 return seq_puts_memcg_tunable(m,
6361 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6364 static ssize_t memory_high_write(struct kernfs_open_file *of,
6365 char *buf, size_t nbytes, loff_t off)
6367 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6368 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6369 bool drained = false;
6373 buf = strstrip(buf);
6374 err = page_counter_memparse(buf, "max", &high);
6378 page_counter_set_high(&memcg->memory, high);
6381 unsigned long nr_pages = page_counter_read(&memcg->memory);
6382 unsigned long reclaimed;
6384 if (nr_pages <= high)
6387 if (signal_pending(current))
6391 drain_all_stock(memcg);
6396 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6399 if (!reclaimed && !nr_retries--)
6403 memcg_wb_domain_size_changed(memcg);
6407 static int memory_max_show(struct seq_file *m, void *v)
6409 return seq_puts_memcg_tunable(m,
6410 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6413 static ssize_t memory_max_write(struct kernfs_open_file *of,
6414 char *buf, size_t nbytes, loff_t off)
6416 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6417 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6418 bool drained = false;
6422 buf = strstrip(buf);
6423 err = page_counter_memparse(buf, "max", &max);
6427 xchg(&memcg->memory.max, max);
6430 unsigned long nr_pages = page_counter_read(&memcg->memory);
6432 if (nr_pages <= max)
6435 if (signal_pending(current))
6439 drain_all_stock(memcg);
6445 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6451 memcg_memory_event(memcg, MEMCG_OOM);
6452 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6456 memcg_wb_domain_size_changed(memcg);
6460 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6462 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6463 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6464 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6465 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6466 seq_printf(m, "oom_kill %lu\n",
6467 atomic_long_read(&events[MEMCG_OOM_KILL]));
6470 static int memory_events_show(struct seq_file *m, void *v)
6472 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6474 __memory_events_show(m, memcg->memory_events);
6478 static int memory_events_local_show(struct seq_file *m, void *v)
6480 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6482 __memory_events_show(m, memcg->memory_events_local);
6486 static int memory_stat_show(struct seq_file *m, void *v)
6488 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6491 buf = memory_stat_format(memcg);
6500 static inline unsigned long lruvec_page_state_output(struct lruvec *lruvec,
6503 return lruvec_page_state(lruvec, item) * memcg_page_state_unit(item);
6506 static int memory_numa_stat_show(struct seq_file *m, void *v)
6509 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6511 mem_cgroup_flush_stats();
6513 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6516 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6519 seq_printf(m, "%s", memory_stats[i].name);
6520 for_each_node_state(nid, N_MEMORY) {
6522 struct lruvec *lruvec;
6524 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6525 size = lruvec_page_state_output(lruvec,
6526 memory_stats[i].idx);
6527 seq_printf(m, " N%d=%llu", nid, size);
6536 static int memory_oom_group_show(struct seq_file *m, void *v)
6538 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6540 seq_printf(m, "%d\n", memcg->oom_group);
6545 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6546 char *buf, size_t nbytes, loff_t off)
6548 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6551 buf = strstrip(buf);
6555 ret = kstrtoint(buf, 0, &oom_group);
6559 if (oom_group != 0 && oom_group != 1)
6562 memcg->oom_group = oom_group;
6567 static struct cftype memory_files[] = {
6570 .flags = CFTYPE_NOT_ON_ROOT,
6571 .read_u64 = memory_current_read,
6575 .flags = CFTYPE_NOT_ON_ROOT,
6576 .seq_show = memory_min_show,
6577 .write = memory_min_write,
6581 .flags = CFTYPE_NOT_ON_ROOT,
6582 .seq_show = memory_low_show,
6583 .write = memory_low_write,
6587 .flags = CFTYPE_NOT_ON_ROOT,
6588 .seq_show = memory_high_show,
6589 .write = memory_high_write,
6593 .flags = CFTYPE_NOT_ON_ROOT,
6594 .seq_show = memory_max_show,
6595 .write = memory_max_write,
6599 .flags = CFTYPE_NOT_ON_ROOT,
6600 .file_offset = offsetof(struct mem_cgroup, events_file),
6601 .seq_show = memory_events_show,
6604 .name = "events.local",
6605 .flags = CFTYPE_NOT_ON_ROOT,
6606 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6607 .seq_show = memory_events_local_show,
6611 .seq_show = memory_stat_show,
6615 .name = "numa_stat",
6616 .seq_show = memory_numa_stat_show,
6620 .name = "oom.group",
6621 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6622 .seq_show = memory_oom_group_show,
6623 .write = memory_oom_group_write,
6628 struct cgroup_subsys memory_cgrp_subsys = {
6629 .css_alloc = mem_cgroup_css_alloc,
6630 .css_online = mem_cgroup_css_online,
6631 .css_offline = mem_cgroup_css_offline,
6632 .css_released = mem_cgroup_css_released,
6633 .css_free = mem_cgroup_css_free,
6634 .css_reset = mem_cgroup_css_reset,
6635 .css_rstat_flush = mem_cgroup_css_rstat_flush,
6636 .can_attach = mem_cgroup_can_attach,
6637 .cancel_attach = mem_cgroup_cancel_attach,
6638 .post_attach = mem_cgroup_move_task,
6639 .dfl_cftypes = memory_files,
6640 .legacy_cftypes = mem_cgroup_legacy_files,
6645 * This function calculates an individual cgroup's effective
6646 * protection which is derived from its own memory.min/low, its
6647 * parent's and siblings' settings, as well as the actual memory
6648 * distribution in the tree.
6650 * The following rules apply to the effective protection values:
6652 * 1. At the first level of reclaim, effective protection is equal to
6653 * the declared protection in memory.min and memory.low.
6655 * 2. To enable safe delegation of the protection configuration, at
6656 * subsequent levels the effective protection is capped to the
6657 * parent's effective protection.
6659 * 3. To make complex and dynamic subtrees easier to configure, the
6660 * user is allowed to overcommit the declared protection at a given
6661 * level. If that is the case, the parent's effective protection is
6662 * distributed to the children in proportion to how much protection
6663 * they have declared and how much of it they are utilizing.
6665 * This makes distribution proportional, but also work-conserving:
6666 * if one cgroup claims much more protection than it uses memory,
6667 * the unused remainder is available to its siblings.
6669 * 4. Conversely, when the declared protection is undercommitted at a
6670 * given level, the distribution of the larger parental protection
6671 * budget is NOT proportional. A cgroup's protection from a sibling
6672 * is capped to its own memory.min/low setting.
6674 * 5. However, to allow protecting recursive subtrees from each other
6675 * without having to declare each individual cgroup's fixed share
6676 * of the ancestor's claim to protection, any unutilized -
6677 * "floating" - protection from up the tree is distributed in
6678 * proportion to each cgroup's *usage*. This makes the protection
6679 * neutral wrt sibling cgroups and lets them compete freely over
6680 * the shared parental protection budget, but it protects the
6681 * subtree as a whole from neighboring subtrees.
6683 * Note that 4. and 5. are not in conflict: 4. is about protecting
6684 * against immediate siblings whereas 5. is about protecting against
6685 * neighboring subtrees.
6687 static unsigned long effective_protection(unsigned long usage,
6688 unsigned long parent_usage,
6689 unsigned long setting,
6690 unsigned long parent_effective,
6691 unsigned long siblings_protected)
6693 unsigned long protected;
6696 protected = min(usage, setting);
6698 * If all cgroups at this level combined claim and use more
6699 * protection then what the parent affords them, distribute
6700 * shares in proportion to utilization.
6702 * We are using actual utilization rather than the statically
6703 * claimed protection in order to be work-conserving: claimed
6704 * but unused protection is available to siblings that would
6705 * otherwise get a smaller chunk than what they claimed.
6707 if (siblings_protected > parent_effective)
6708 return protected * parent_effective / siblings_protected;
6711 * Ok, utilized protection of all children is within what the
6712 * parent affords them, so we know whatever this child claims
6713 * and utilizes is effectively protected.
6715 * If there is unprotected usage beyond this value, reclaim
6716 * will apply pressure in proportion to that amount.
6718 * If there is unutilized protection, the cgroup will be fully
6719 * shielded from reclaim, but we do return a smaller value for
6720 * protection than what the group could enjoy in theory. This
6721 * is okay. With the overcommit distribution above, effective
6722 * protection is always dependent on how memory is actually
6723 * consumed among the siblings anyway.
6728 * If the children aren't claiming (all of) the protection
6729 * afforded to them by the parent, distribute the remainder in
6730 * proportion to the (unprotected) memory of each cgroup. That
6731 * way, cgroups that aren't explicitly prioritized wrt each
6732 * other compete freely over the allowance, but they are
6733 * collectively protected from neighboring trees.
6735 * We're using unprotected memory for the weight so that if
6736 * some cgroups DO claim explicit protection, we don't protect
6737 * the same bytes twice.
6739 * Check both usage and parent_usage against the respective
6740 * protected values. One should imply the other, but they
6741 * aren't read atomically - make sure the division is sane.
6743 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6745 if (parent_effective > siblings_protected &&
6746 parent_usage > siblings_protected &&
6747 usage > protected) {
6748 unsigned long unclaimed;
6750 unclaimed = parent_effective - siblings_protected;
6751 unclaimed *= usage - protected;
6752 unclaimed /= parent_usage - siblings_protected;
6761 * mem_cgroup_calculate_protection - check if memory consumption is in the normal range
6762 * @root: the top ancestor of the sub-tree being checked
6763 * @memcg: the memory cgroup to check
6765 * WARNING: This function is not stateless! It can only be used as part
6766 * of a top-down tree iteration, not for isolated queries.
6768 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6769 struct mem_cgroup *memcg)
6771 unsigned long usage, parent_usage;
6772 struct mem_cgroup *parent;
6774 if (mem_cgroup_disabled())
6778 root = root_mem_cgroup;
6781 * Effective values of the reclaim targets are ignored so they
6782 * can be stale. Have a look at mem_cgroup_protection for more
6784 * TODO: calculation should be more robust so that we do not need
6785 * that special casing.
6790 usage = page_counter_read(&memcg->memory);
6794 parent = parent_mem_cgroup(memcg);
6795 /* No parent means a non-hierarchical mode on v1 memcg */
6799 if (parent == root) {
6800 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6801 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6805 parent_usage = page_counter_read(&parent->memory);
6807 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6808 READ_ONCE(memcg->memory.min),
6809 READ_ONCE(parent->memory.emin),
6810 atomic_long_read(&parent->memory.children_min_usage)));
6812 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6813 READ_ONCE(memcg->memory.low),
6814 READ_ONCE(parent->memory.elow),
6815 atomic_long_read(&parent->memory.children_low_usage)));
6818 static int charge_memcg(struct page *page, struct mem_cgroup *memcg, gfp_t gfp)
6820 unsigned int nr_pages = thp_nr_pages(page);
6823 ret = try_charge(memcg, gfp, nr_pages);
6827 css_get(&memcg->css);
6828 commit_charge(page, memcg);
6830 local_irq_disable();
6831 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6832 memcg_check_events(memcg, page);
6839 * __mem_cgroup_charge - charge a newly allocated page to a cgroup
6840 * @page: page to charge
6841 * @mm: mm context of the victim
6842 * @gfp_mask: reclaim mode
6844 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6845 * pages according to @gfp_mask if necessary. if @mm is NULL, try to
6846 * charge to the active memcg.
6848 * Do not use this for pages allocated for swapin.
6850 * Returns 0 on success. Otherwise, an error code is returned.
6852 int __mem_cgroup_charge(struct page *page, struct mm_struct *mm,
6855 struct mem_cgroup *memcg;
6858 memcg = get_mem_cgroup_from_mm(mm);
6859 ret = charge_memcg(page, memcg, gfp_mask);
6860 css_put(&memcg->css);
6866 * mem_cgroup_swapin_charge_page - charge a newly allocated page for swapin
6867 * @page: page to charge
6868 * @mm: mm context of the victim
6869 * @gfp: reclaim mode
6870 * @entry: swap entry for which the page is allocated
6872 * This function charges a page allocated for swapin. Please call this before
6873 * adding the page to the swapcache.
6875 * Returns 0 on success. Otherwise, an error code is returned.
6877 int mem_cgroup_swapin_charge_page(struct page *page, struct mm_struct *mm,
6878 gfp_t gfp, swp_entry_t entry)
6880 struct mem_cgroup *memcg;
6884 if (mem_cgroup_disabled())
6887 id = lookup_swap_cgroup_id(entry);
6889 memcg = mem_cgroup_from_id(id);
6890 if (!memcg || !css_tryget_online(&memcg->css))
6891 memcg = get_mem_cgroup_from_mm(mm);
6894 ret = charge_memcg(page, memcg, gfp);
6896 css_put(&memcg->css);
6901 * mem_cgroup_swapin_uncharge_swap - uncharge swap slot
6902 * @entry: swap entry for which the page is charged
6904 * Call this function after successfully adding the charged page to swapcache.
6906 * Note: This function assumes the page for which swap slot is being uncharged
6909 void mem_cgroup_swapin_uncharge_swap(swp_entry_t entry)
6912 * Cgroup1's unified memory+swap counter has been charged with the
6913 * new swapcache page, finish the transfer by uncharging the swap
6914 * slot. The swap slot would also get uncharged when it dies, but
6915 * it can stick around indefinitely and we'd count the page twice
6918 * Cgroup2 has separate resource counters for memory and swap,
6919 * so this is a non-issue here. Memory and swap charge lifetimes
6920 * correspond 1:1 to page and swap slot lifetimes: we charge the
6921 * page to memory here, and uncharge swap when the slot is freed.
6923 if (!mem_cgroup_disabled() && do_memsw_account()) {
6925 * The swap entry might not get freed for a long time,
6926 * let's not wait for it. The page already received a
6927 * memory+swap charge, drop the swap entry duplicate.
6929 mem_cgroup_uncharge_swap(entry, 1);
6933 struct uncharge_gather {
6934 struct mem_cgroup *memcg;
6935 unsigned long nr_memory;
6936 unsigned long pgpgout;
6937 unsigned long nr_kmem;
6938 struct page *dummy_page;
6941 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6943 memset(ug, 0, sizeof(*ug));
6946 static void uncharge_batch(const struct uncharge_gather *ug)
6948 unsigned long flags;
6950 if (ug->nr_memory) {
6951 page_counter_uncharge(&ug->memcg->memory, ug->nr_memory);
6952 if (do_memsw_account())
6953 page_counter_uncharge(&ug->memcg->memsw, ug->nr_memory);
6954 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6955 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6956 memcg_oom_recover(ug->memcg);
6959 local_irq_save(flags);
6960 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6961 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_memory);
6962 memcg_check_events(ug->memcg, ug->dummy_page);
6963 local_irq_restore(flags);
6965 /* drop reference from uncharge_page */
6966 css_put(&ug->memcg->css);
6969 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6971 unsigned long nr_pages;
6972 struct mem_cgroup *memcg;
6973 struct obj_cgroup *objcg;
6975 VM_BUG_ON_PAGE(PageLRU(page), page);
6978 * Nobody should be changing or seriously looking at
6979 * page memcg or objcg at this point, we have fully
6980 * exclusive access to the page.
6982 if (PageMemcgKmem(page)) {
6983 objcg = __page_objcg(page);
6985 * This get matches the put at the end of the function and
6986 * kmem pages do not hold memcg references anymore.
6988 memcg = get_mem_cgroup_from_objcg(objcg);
6990 memcg = __page_memcg(page);
6996 if (ug->memcg != memcg) {
6999 uncharge_gather_clear(ug);
7002 ug->dummy_page = page;
7004 /* pairs with css_put in uncharge_batch */
7005 css_get(&memcg->css);
7008 nr_pages = compound_nr(page);
7010 if (PageMemcgKmem(page)) {
7011 ug->nr_memory += nr_pages;
7012 ug->nr_kmem += nr_pages;
7014 page->memcg_data = 0;
7015 obj_cgroup_put(objcg);
7017 /* LRU pages aren't accounted at the root level */
7018 if (!mem_cgroup_is_root(memcg))
7019 ug->nr_memory += nr_pages;
7022 page->memcg_data = 0;
7025 css_put(&memcg->css);
7029 * __mem_cgroup_uncharge - uncharge a page
7030 * @page: page to uncharge
7032 * Uncharge a page previously charged with __mem_cgroup_charge().
7034 void __mem_cgroup_uncharge(struct page *page)
7036 struct uncharge_gather ug;
7038 /* Don't touch page->lru of any random page, pre-check: */
7039 if (!page_memcg(page))
7042 uncharge_gather_clear(&ug);
7043 uncharge_page(page, &ug);
7044 uncharge_batch(&ug);
7048 * __mem_cgroup_uncharge_list - uncharge a list of page
7049 * @page_list: list of pages to uncharge
7051 * Uncharge a list of pages previously charged with
7052 * __mem_cgroup_charge().
7054 void __mem_cgroup_uncharge_list(struct list_head *page_list)
7056 struct uncharge_gather ug;
7059 uncharge_gather_clear(&ug);
7060 list_for_each_entry(page, page_list, lru)
7061 uncharge_page(page, &ug);
7063 uncharge_batch(&ug);
7067 * mem_cgroup_migrate - charge a page's replacement
7068 * @oldpage: currently circulating page
7069 * @newpage: replacement page
7071 * Charge @newpage as a replacement page for @oldpage. @oldpage will
7072 * be uncharged upon free.
7074 * Both pages must be locked, @newpage->mapping must be set up.
7076 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
7078 struct mem_cgroup *memcg;
7079 unsigned int nr_pages;
7080 unsigned long flags;
7082 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
7083 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
7084 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
7085 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
7088 if (mem_cgroup_disabled())
7091 /* Page cache replacement: new page already charged? */
7092 if (page_memcg(newpage))
7095 memcg = page_memcg(oldpage);
7096 VM_WARN_ON_ONCE_PAGE(!memcg, oldpage);
7100 /* Force-charge the new page. The old one will be freed soon */
7101 nr_pages = thp_nr_pages(newpage);
7103 if (!mem_cgroup_is_root(memcg)) {
7104 page_counter_charge(&memcg->memory, nr_pages);
7105 if (do_memsw_account())
7106 page_counter_charge(&memcg->memsw, nr_pages);
7109 css_get(&memcg->css);
7110 commit_charge(newpage, memcg);
7112 local_irq_save(flags);
7113 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
7114 memcg_check_events(memcg, newpage);
7115 local_irq_restore(flags);
7118 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
7119 EXPORT_SYMBOL(memcg_sockets_enabled_key);
7121 void mem_cgroup_sk_alloc(struct sock *sk)
7123 struct mem_cgroup *memcg;
7125 if (!mem_cgroup_sockets_enabled)
7128 /* Do not associate the sock with unrelated interrupted task's memcg. */
7133 memcg = mem_cgroup_from_task(current);
7134 if (memcg == root_mem_cgroup)
7136 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
7138 if (css_tryget(&memcg->css))
7139 sk->sk_memcg = memcg;
7144 void mem_cgroup_sk_free(struct sock *sk)
7147 css_put(&sk->sk_memcg->css);
7151 * mem_cgroup_charge_skmem - charge socket memory
7152 * @memcg: memcg to charge
7153 * @nr_pages: number of pages to charge
7154 * @gfp_mask: reclaim mode
7156 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7157 * @memcg's configured limit, %false if it doesn't.
7159 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages,
7162 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7163 struct page_counter *fail;
7165 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7166 memcg->tcpmem_pressure = 0;
7169 memcg->tcpmem_pressure = 1;
7170 if (gfp_mask & __GFP_NOFAIL) {
7171 page_counter_charge(&memcg->tcpmem, nr_pages);
7177 if (try_charge(memcg, gfp_mask, nr_pages) == 0) {
7178 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7186 * mem_cgroup_uncharge_skmem - uncharge socket memory
7187 * @memcg: memcg to uncharge
7188 * @nr_pages: number of pages to uncharge
7190 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7192 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7193 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7197 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7199 refill_stock(memcg, nr_pages);
7202 static int __init cgroup_memory(char *s)
7206 while ((token = strsep(&s, ",")) != NULL) {
7209 if (!strcmp(token, "nosocket"))
7210 cgroup_memory_nosocket = true;
7211 if (!strcmp(token, "nokmem"))
7212 cgroup_memory_nokmem = true;
7216 __setup("cgroup.memory=", cgroup_memory);
7219 * subsys_initcall() for memory controller.
7221 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7222 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7223 * basically everything that doesn't depend on a specific mem_cgroup structure
7224 * should be initialized from here.
7226 static int __init mem_cgroup_init(void)
7231 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7232 * used for per-memcg-per-cpu caching of per-node statistics. In order
7233 * to work fine, we should make sure that the overfill threshold can't
7234 * exceed S32_MAX / PAGE_SIZE.
7236 BUILD_BUG_ON(MEMCG_CHARGE_BATCH > S32_MAX / PAGE_SIZE);
7238 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7239 memcg_hotplug_cpu_dead);
7241 for_each_possible_cpu(cpu)
7242 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7245 for_each_node(node) {
7246 struct mem_cgroup_tree_per_node *rtpn;
7248 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7249 node_online(node) ? node : NUMA_NO_NODE);
7251 rtpn->rb_root = RB_ROOT;
7252 rtpn->rb_rightmost = NULL;
7253 spin_lock_init(&rtpn->lock);
7254 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7259 subsys_initcall(mem_cgroup_init);
7261 #ifdef CONFIG_MEMCG_SWAP
7262 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7264 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7266 * The root cgroup cannot be destroyed, so it's refcount must
7269 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7273 memcg = parent_mem_cgroup(memcg);
7275 memcg = root_mem_cgroup;
7281 * mem_cgroup_swapout - transfer a memsw charge to swap
7282 * @page: page whose memsw charge to transfer
7283 * @entry: swap entry to move the charge to
7285 * Transfer the memsw charge of @page to @entry.
7287 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7289 struct mem_cgroup *memcg, *swap_memcg;
7290 unsigned int nr_entries;
7291 unsigned short oldid;
7293 VM_BUG_ON_PAGE(PageLRU(page), page);
7294 VM_BUG_ON_PAGE(page_count(page), page);
7296 if (mem_cgroup_disabled())
7299 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7302 memcg = page_memcg(page);
7304 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7309 * In case the memcg owning these pages has been offlined and doesn't
7310 * have an ID allocated to it anymore, charge the closest online
7311 * ancestor for the swap instead and transfer the memory+swap charge.
7313 swap_memcg = mem_cgroup_id_get_online(memcg);
7314 nr_entries = thp_nr_pages(page);
7315 /* Get references for the tail pages, too */
7317 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7318 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7320 VM_BUG_ON_PAGE(oldid, page);
7321 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7323 page->memcg_data = 0;
7325 if (!mem_cgroup_is_root(memcg))
7326 page_counter_uncharge(&memcg->memory, nr_entries);
7328 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7329 if (!mem_cgroup_is_root(swap_memcg))
7330 page_counter_charge(&swap_memcg->memsw, nr_entries);
7331 page_counter_uncharge(&memcg->memsw, nr_entries);
7335 * Interrupts should be disabled here because the caller holds the
7336 * i_pages lock which is taken with interrupts-off. It is
7337 * important here to have the interrupts disabled because it is the
7338 * only synchronisation we have for updating the per-CPU variables.
7341 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7342 memcg_stats_unlock();
7343 memcg_check_events(memcg, page);
7345 css_put(&memcg->css);
7349 * __mem_cgroup_try_charge_swap - try charging swap space for a page
7350 * @page: page being added to swap
7351 * @entry: swap entry to charge
7353 * Try to charge @page's memcg for the swap space at @entry.
7355 * Returns 0 on success, -ENOMEM on failure.
7357 int __mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7359 unsigned int nr_pages = thp_nr_pages(page);
7360 struct page_counter *counter;
7361 struct mem_cgroup *memcg;
7362 unsigned short oldid;
7364 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7367 memcg = page_memcg(page);
7369 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7374 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7378 memcg = mem_cgroup_id_get_online(memcg);
7380 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7381 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7382 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7383 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7384 mem_cgroup_id_put(memcg);
7388 /* Get references for the tail pages, too */
7390 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7391 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7392 VM_BUG_ON_PAGE(oldid, page);
7393 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7399 * __mem_cgroup_uncharge_swap - uncharge swap space
7400 * @entry: swap entry to uncharge
7401 * @nr_pages: the amount of swap space to uncharge
7403 void __mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7405 struct mem_cgroup *memcg;
7408 id = swap_cgroup_record(entry, 0, nr_pages);
7410 memcg = mem_cgroup_from_id(id);
7412 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7413 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7414 page_counter_uncharge(&memcg->swap, nr_pages);
7416 page_counter_uncharge(&memcg->memsw, nr_pages);
7418 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7419 mem_cgroup_id_put_many(memcg, nr_pages);
7424 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7426 long nr_swap_pages = get_nr_swap_pages();
7428 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7429 return nr_swap_pages;
7430 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7431 nr_swap_pages = min_t(long, nr_swap_pages,
7432 READ_ONCE(memcg->swap.max) -
7433 page_counter_read(&memcg->swap));
7434 return nr_swap_pages;
7437 bool mem_cgroup_swap_full(struct page *page)
7439 struct mem_cgroup *memcg;
7441 VM_BUG_ON_PAGE(!PageLocked(page), page);
7445 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7448 memcg = page_memcg(page);
7452 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7453 unsigned long usage = page_counter_read(&memcg->swap);
7455 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7456 usage * 2 >= READ_ONCE(memcg->swap.max))
7463 static int __init setup_swap_account(char *s)
7465 if (!strcmp(s, "1"))
7466 cgroup_memory_noswap = false;
7467 else if (!strcmp(s, "0"))
7468 cgroup_memory_noswap = true;
7471 __setup("swapaccount=", setup_swap_account);
7473 static u64 swap_current_read(struct cgroup_subsys_state *css,
7476 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7478 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7481 static int swap_high_show(struct seq_file *m, void *v)
7483 return seq_puts_memcg_tunable(m,
7484 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7487 static ssize_t swap_high_write(struct kernfs_open_file *of,
7488 char *buf, size_t nbytes, loff_t off)
7490 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7494 buf = strstrip(buf);
7495 err = page_counter_memparse(buf, "max", &high);
7499 page_counter_set_high(&memcg->swap, high);
7504 static int swap_max_show(struct seq_file *m, void *v)
7506 return seq_puts_memcg_tunable(m,
7507 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7510 static ssize_t swap_max_write(struct kernfs_open_file *of,
7511 char *buf, size_t nbytes, loff_t off)
7513 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7517 buf = strstrip(buf);
7518 err = page_counter_memparse(buf, "max", &max);
7522 xchg(&memcg->swap.max, max);
7527 static int swap_events_show(struct seq_file *m, void *v)
7529 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7531 seq_printf(m, "high %lu\n",
7532 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7533 seq_printf(m, "max %lu\n",
7534 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7535 seq_printf(m, "fail %lu\n",
7536 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7541 static struct cftype swap_files[] = {
7543 .name = "swap.current",
7544 .flags = CFTYPE_NOT_ON_ROOT,
7545 .read_u64 = swap_current_read,
7548 .name = "swap.high",
7549 .flags = CFTYPE_NOT_ON_ROOT,
7550 .seq_show = swap_high_show,
7551 .write = swap_high_write,
7555 .flags = CFTYPE_NOT_ON_ROOT,
7556 .seq_show = swap_max_show,
7557 .write = swap_max_write,
7560 .name = "swap.events",
7561 .flags = CFTYPE_NOT_ON_ROOT,
7562 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7563 .seq_show = swap_events_show,
7568 static struct cftype memsw_files[] = {
7570 .name = "memsw.usage_in_bytes",
7571 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7572 .read_u64 = mem_cgroup_read_u64,
7575 .name = "memsw.max_usage_in_bytes",
7576 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7577 .write = mem_cgroup_reset,
7578 .read_u64 = mem_cgroup_read_u64,
7581 .name = "memsw.limit_in_bytes",
7582 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7583 .write = mem_cgroup_write,
7584 .read_u64 = mem_cgroup_read_u64,
7587 .name = "memsw.failcnt",
7588 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7589 .write = mem_cgroup_reset,
7590 .read_u64 = mem_cgroup_read_u64,
7593 .name = "force_reclaim",
7594 .write_u64 = mem_cgroup_force_reclaim,
7596 { }, /* terminate */
7600 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7601 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7602 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7603 * boot parameter. This may result in premature OOPS inside
7604 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7606 static int __init mem_cgroup_swap_init(void)
7608 /* No memory control -> no swap control */
7609 if (mem_cgroup_disabled())
7610 cgroup_memory_noswap = true;
7612 if (cgroup_memory_noswap)
7615 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7616 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7620 core_initcall(mem_cgroup_swap_init);
7622 #endif /* CONFIG_MEMCG_SWAP */