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);
82 /* Socket memory accounting disabled? */
83 static bool cgroup_memory_nosocket;
85 /* Kernel memory accounting disabled? */
86 static bool cgroup_memory_nokmem;
88 /* Whether the swap controller is active */
89 #ifdef CONFIG_MEMCG_SWAP
90 bool cgroup_memory_noswap __read_mostly;
92 #define cgroup_memory_noswap 1
95 #ifdef CONFIG_CGROUP_WRITEBACK
96 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
99 /* Whether legacy memory+swap accounting is active */
100 static bool do_memsw_account(void)
102 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_noswap;
105 #define THRESHOLDS_EVENTS_TARGET 128
106 #define SOFTLIMIT_EVENTS_TARGET 1024
109 * Cgroups above their limits are maintained in a RB-Tree, independent of
110 * their hierarchy representation
113 struct mem_cgroup_tree_per_node {
114 struct rb_root rb_root;
115 struct rb_node *rb_rightmost;
119 struct mem_cgroup_tree {
120 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
123 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
126 struct mem_cgroup_eventfd_list {
127 struct list_head list;
128 struct eventfd_ctx *eventfd;
132 * cgroup_event represents events which userspace want to receive.
134 struct mem_cgroup_event {
136 * memcg which the event belongs to.
138 struct mem_cgroup *memcg;
140 * eventfd to signal userspace about the event.
142 struct eventfd_ctx *eventfd;
144 * Each of these stored in a list by the cgroup.
146 struct list_head list;
148 * register_event() callback will be used to add new userspace
149 * waiter for changes related to this event. Use eventfd_signal()
150 * on eventfd to send notification to userspace.
152 int (*register_event)(struct mem_cgroup *memcg,
153 struct eventfd_ctx *eventfd, const char *args);
155 * unregister_event() callback will be called when userspace closes
156 * the eventfd or on cgroup removing. This callback must be set,
157 * if you want provide notification functionality.
159 void (*unregister_event)(struct mem_cgroup *memcg,
160 struct eventfd_ctx *eventfd);
162 * All fields below needed to unregister event when
163 * userspace closes eventfd.
166 wait_queue_head_t *wqh;
167 wait_queue_entry_t wait;
168 struct work_struct remove;
171 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
172 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
174 /* Stuffs for move charges at task migration. */
176 * Types of charges to be moved.
178 #define MOVE_ANON 0x1U
179 #define MOVE_FILE 0x2U
180 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
182 /* "mc" and its members are protected by cgroup_mutex */
183 static struct move_charge_struct {
184 spinlock_t lock; /* for from, to */
185 struct mm_struct *mm;
186 struct mem_cgroup *from;
187 struct mem_cgroup *to;
189 unsigned long precharge;
190 unsigned long moved_charge;
191 unsigned long moved_swap;
192 struct task_struct *moving_task; /* a task moving charges */
193 wait_queue_head_t waitq; /* a waitq for other context */
195 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
196 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
200 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
201 * limit reclaim to prevent infinite loops, if they ever occur.
203 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
204 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
206 /* for encoding cft->private value on file */
215 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
216 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
217 #define MEMFILE_ATTR(val) ((val) & 0xffff)
218 /* Used for OOM nofiier */
219 #define OOM_CONTROL (0)
222 * Iteration constructs for visiting all cgroups (under a tree). If
223 * loops are exited prematurely (break), mem_cgroup_iter_break() must
224 * be used for reference counting.
226 #define for_each_mem_cgroup_tree(iter, root) \
227 for (iter = mem_cgroup_iter(root, NULL, NULL); \
229 iter = mem_cgroup_iter(root, iter, NULL))
231 #define for_each_mem_cgroup(iter) \
232 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
234 iter = mem_cgroup_iter(NULL, iter, NULL))
236 static inline bool should_force_charge(void)
238 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
239 (current->flags & PF_EXITING);
242 /* Some nice accessors for the vmpressure. */
243 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
246 memcg = root_mem_cgroup;
247 return &memcg->vmpressure;
250 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
252 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
255 #ifdef CONFIG_MEMCG_KMEM
256 extern spinlock_t css_set_lock;
258 static void obj_cgroup_release(struct percpu_ref *ref)
260 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
261 struct mem_cgroup *memcg;
262 unsigned int nr_bytes;
263 unsigned int nr_pages;
267 * At this point all allocated objects are freed, and
268 * objcg->nr_charged_bytes can't have an arbitrary byte value.
269 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
271 * The following sequence can lead to it:
272 * 1) CPU0: objcg == stock->cached_objcg
273 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
274 * PAGE_SIZE bytes are charged
275 * 3) CPU1: a process from another memcg is allocating something,
276 * the stock if flushed,
277 * objcg->nr_charged_bytes = PAGE_SIZE - 92
278 * 5) CPU0: we do release this object,
279 * 92 bytes are added to stock->nr_bytes
280 * 6) CPU0: stock is flushed,
281 * 92 bytes are added to objcg->nr_charged_bytes
283 * In the result, nr_charged_bytes == PAGE_SIZE.
284 * This page will be uncharged in obj_cgroup_release().
286 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
287 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
288 nr_pages = nr_bytes >> PAGE_SHIFT;
290 spin_lock_irqsave(&css_set_lock, flags);
291 memcg = obj_cgroup_memcg(objcg);
293 __memcg_kmem_uncharge(memcg, nr_pages);
294 list_del(&objcg->list);
295 mem_cgroup_put(memcg);
296 spin_unlock_irqrestore(&css_set_lock, flags);
298 percpu_ref_exit(ref);
299 kfree_rcu(objcg, rcu);
302 static struct obj_cgroup *obj_cgroup_alloc(void)
304 struct obj_cgroup *objcg;
307 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
311 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
317 INIT_LIST_HEAD(&objcg->list);
321 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
322 struct mem_cgroup *parent)
324 struct obj_cgroup *objcg, *iter;
326 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
328 spin_lock_irq(&css_set_lock);
330 /* Move active objcg to the parent's list */
331 xchg(&objcg->memcg, parent);
332 css_get(&parent->css);
333 list_add(&objcg->list, &parent->objcg_list);
335 /* Move already reparented objcgs to the parent's list */
336 list_for_each_entry(iter, &memcg->objcg_list, list) {
337 css_get(&parent->css);
338 xchg(&iter->memcg, parent);
339 css_put(&memcg->css);
341 list_splice(&memcg->objcg_list, &parent->objcg_list);
343 spin_unlock_irq(&css_set_lock);
345 percpu_ref_kill(&objcg->refcnt);
349 * This will be used as a shrinker list's index.
350 * The main reason for not using cgroup id for this:
351 * this works better in sparse environments, where we have a lot of memcgs,
352 * but only a few kmem-limited. Or also, if we have, for instance, 200
353 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
354 * 200 entry array for that.
356 * The current size of the caches array is stored in memcg_nr_cache_ids. It
357 * will double each time we have to increase it.
359 static DEFINE_IDA(memcg_cache_ida);
360 int memcg_nr_cache_ids;
362 /* Protects memcg_nr_cache_ids */
363 static DECLARE_RWSEM(memcg_cache_ids_sem);
365 void memcg_get_cache_ids(void)
367 down_read(&memcg_cache_ids_sem);
370 void memcg_put_cache_ids(void)
372 up_read(&memcg_cache_ids_sem);
376 * MIN_SIZE is different than 1, because we would like to avoid going through
377 * the alloc/free process all the time. In a small machine, 4 kmem-limited
378 * cgroups is a reasonable guess. In the future, it could be a parameter or
379 * tunable, but that is strictly not necessary.
381 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
382 * this constant directly from cgroup, but it is understandable that this is
383 * better kept as an internal representation in cgroup.c. In any case, the
384 * cgrp_id space is not getting any smaller, and we don't have to necessarily
385 * increase ours as well if it increases.
387 #define MEMCG_CACHES_MIN_SIZE 4
388 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
391 * A lot of the calls to the cache allocation functions are expected to be
392 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
393 * conditional to this static branch, we'll have to allow modules that does
394 * kmem_cache_alloc and the such to see this symbol as well
396 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
397 EXPORT_SYMBOL(memcg_kmem_enabled_key);
400 static int memcg_shrinker_map_size;
401 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
403 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
405 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
408 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
409 int size, int old_size)
411 struct memcg_shrinker_map *new, *old;
414 lockdep_assert_held(&memcg_shrinker_map_mutex);
417 old = rcu_dereference_protected(
418 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
419 /* Not yet online memcg */
423 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
427 /* Set all old bits, clear all new bits */
428 memset(new->map, (int)0xff, old_size);
429 memset((void *)new->map + old_size, 0, size - old_size);
431 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
432 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
438 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
440 struct mem_cgroup_per_node *pn;
441 struct memcg_shrinker_map *map;
444 if (mem_cgroup_is_root(memcg))
448 pn = mem_cgroup_nodeinfo(memcg, nid);
449 map = rcu_dereference_protected(pn->shrinker_map, true);
452 rcu_assign_pointer(pn->shrinker_map, NULL);
456 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
458 struct memcg_shrinker_map *map;
459 int nid, size, ret = 0;
461 if (mem_cgroup_is_root(memcg))
464 mutex_lock(&memcg_shrinker_map_mutex);
465 size = memcg_shrinker_map_size;
467 map = kvzalloc_node(sizeof(*map) + size, GFP_KERNEL, nid);
469 memcg_free_shrinker_maps(memcg);
473 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
475 mutex_unlock(&memcg_shrinker_map_mutex);
480 int memcg_expand_shrinker_maps(int new_id)
482 int size, old_size, ret = 0;
483 struct mem_cgroup *memcg;
485 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
486 old_size = memcg_shrinker_map_size;
487 if (size <= old_size)
490 mutex_lock(&memcg_shrinker_map_mutex);
491 if (!root_mem_cgroup)
494 for_each_mem_cgroup(memcg) {
495 if (mem_cgroup_is_root(memcg))
497 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
499 mem_cgroup_iter_break(NULL, memcg);
505 memcg_shrinker_map_size = size;
506 mutex_unlock(&memcg_shrinker_map_mutex);
510 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
512 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
513 struct memcg_shrinker_map *map;
516 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
517 /* Pairs with smp mb in shrink_slab() */
518 smp_mb__before_atomic();
519 set_bit(shrinker_id, map->map);
525 * mem_cgroup_css_from_page - css of the memcg associated with a page
526 * @page: page of interest
528 * If memcg is bound to the default hierarchy, css of the memcg associated
529 * with @page is returned. The returned css remains associated with @page
530 * until it is released.
532 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
535 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
537 struct mem_cgroup *memcg;
539 memcg = page_memcg(page);
541 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
542 memcg = root_mem_cgroup;
548 * page_cgroup_ino - return inode number of the memcg a page is charged to
551 * Look up the closest online ancestor of the memory cgroup @page is charged to
552 * and return its inode number or 0 if @page is not charged to any cgroup. It
553 * is safe to call this function without holding a reference to @page.
555 * Note, this function is inherently racy, because there is nothing to prevent
556 * the cgroup inode from getting torn down and potentially reallocated a moment
557 * after page_cgroup_ino() returns, so it only should be used by callers that
558 * do not care (such as procfs interfaces).
560 ino_t page_cgroup_ino(struct page *page)
562 struct mem_cgroup *memcg;
563 unsigned long ino = 0;
566 memcg = page_memcg_check(page);
568 while (memcg && !(memcg->css.flags & CSS_ONLINE))
569 memcg = parent_mem_cgroup(memcg);
571 ino = cgroup_ino(memcg->css.cgroup);
576 static struct mem_cgroup_per_node *
577 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
579 int nid = page_to_nid(page);
581 return memcg->nodeinfo[nid];
584 static struct mem_cgroup_tree_per_node *
585 soft_limit_tree_node(int nid)
587 return soft_limit_tree.rb_tree_per_node[nid];
590 static struct mem_cgroup_tree_per_node *
591 soft_limit_tree_from_page(struct page *page)
593 int nid = page_to_nid(page);
595 return soft_limit_tree.rb_tree_per_node[nid];
598 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
599 struct mem_cgroup_tree_per_node *mctz,
600 unsigned long new_usage_in_excess)
602 struct rb_node **p = &mctz->rb_root.rb_node;
603 struct rb_node *parent = NULL;
604 struct mem_cgroup_per_node *mz_node;
605 bool rightmost = true;
610 mz->usage_in_excess = new_usage_in_excess;
611 if (!mz->usage_in_excess)
615 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
617 if (mz->usage_in_excess < mz_node->usage_in_excess) {
626 mctz->rb_rightmost = &mz->tree_node;
628 rb_link_node(&mz->tree_node, parent, p);
629 rb_insert_color(&mz->tree_node, &mctz->rb_root);
633 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
634 struct mem_cgroup_tree_per_node *mctz)
639 if (&mz->tree_node == mctz->rb_rightmost)
640 mctz->rb_rightmost = rb_prev(&mz->tree_node);
642 rb_erase(&mz->tree_node, &mctz->rb_root);
646 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
647 struct mem_cgroup_tree_per_node *mctz)
651 spin_lock_irqsave(&mctz->lock, flags);
652 __mem_cgroup_remove_exceeded(mz, mctz);
653 spin_unlock_irqrestore(&mctz->lock, flags);
656 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
658 unsigned long nr_pages = page_counter_read(&memcg->memory);
659 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
660 unsigned long excess = 0;
662 if (nr_pages > soft_limit)
663 excess = nr_pages - soft_limit;
668 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
670 unsigned long excess;
671 struct mem_cgroup_per_node *mz;
672 struct mem_cgroup_tree_per_node *mctz;
674 mctz = soft_limit_tree_from_page(page);
678 * Necessary to update all ancestors when hierarchy is used.
679 * because their event counter is not touched.
681 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
682 mz = mem_cgroup_page_nodeinfo(memcg, page);
683 excess = soft_limit_excess(memcg);
685 * We have to update the tree if mz is on RB-tree or
686 * mem is over its softlimit.
688 if (excess || mz->on_tree) {
691 spin_lock_irqsave(&mctz->lock, flags);
692 /* if on-tree, remove it */
694 __mem_cgroup_remove_exceeded(mz, mctz);
696 * Insert again. mz->usage_in_excess will be updated.
697 * If excess is 0, no tree ops.
699 __mem_cgroup_insert_exceeded(mz, mctz, excess);
700 spin_unlock_irqrestore(&mctz->lock, flags);
705 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
707 struct mem_cgroup_tree_per_node *mctz;
708 struct mem_cgroup_per_node *mz;
712 mz = mem_cgroup_nodeinfo(memcg, nid);
713 mctz = soft_limit_tree_node(nid);
715 mem_cgroup_remove_exceeded(mz, mctz);
719 static struct mem_cgroup_per_node *
720 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
722 struct mem_cgroup_per_node *mz;
726 if (!mctz->rb_rightmost)
727 goto done; /* Nothing to reclaim from */
729 mz = rb_entry(mctz->rb_rightmost,
730 struct mem_cgroup_per_node, tree_node);
732 * Remove the node now but someone else can add it back,
733 * we will to add it back at the end of reclaim to its correct
734 * position in the tree.
736 __mem_cgroup_remove_exceeded(mz, mctz);
737 if (!soft_limit_excess(mz->memcg) ||
738 !css_tryget(&mz->memcg->css))
744 static struct mem_cgroup_per_node *
745 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
747 struct mem_cgroup_per_node *mz;
749 spin_lock_irq(&mctz->lock);
750 mz = __mem_cgroup_largest_soft_limit_node(mctz);
751 spin_unlock_irq(&mctz->lock);
756 * __mod_memcg_state - update cgroup memory statistics
757 * @memcg: the memory cgroup
758 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
759 * @val: delta to add to the counter, can be negative
761 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
763 long x, threshold = MEMCG_CHARGE_BATCH;
765 if (mem_cgroup_disabled())
768 if (memcg_stat_item_in_bytes(idx))
769 threshold <<= PAGE_SHIFT;
771 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
772 if (unlikely(abs(x) > threshold)) {
773 struct mem_cgroup *mi;
776 * Batch local counters to keep them in sync with
777 * the hierarchical ones.
779 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
780 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
781 atomic_long_add(x, &mi->vmstats[idx]);
784 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
787 static struct mem_cgroup_per_node *
788 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
790 struct mem_cgroup *parent;
792 parent = parent_mem_cgroup(pn->memcg);
795 return mem_cgroup_nodeinfo(parent, nid);
798 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
801 struct mem_cgroup_per_node *pn;
802 struct mem_cgroup *memcg;
803 long x, threshold = MEMCG_CHARGE_BATCH;
805 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
809 __mod_memcg_state(memcg, idx, val);
812 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
814 if (vmstat_item_in_bytes(idx))
815 threshold <<= PAGE_SHIFT;
817 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
818 if (unlikely(abs(x) > threshold)) {
819 pg_data_t *pgdat = lruvec_pgdat(lruvec);
820 struct mem_cgroup_per_node *pi;
822 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
823 atomic_long_add(x, &pi->lruvec_stat[idx]);
826 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
830 * __mod_lruvec_state - update lruvec memory statistics
831 * @lruvec: the lruvec
832 * @idx: the stat item
833 * @val: delta to add to the counter, can be negative
835 * The lruvec is the intersection of the NUMA node and a cgroup. This
836 * function updates the all three counters that are affected by a
837 * change of state at this level: per-node, per-cgroup, per-lruvec.
839 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
843 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
845 /* Update memcg and lruvec */
846 if (!mem_cgroup_disabled())
847 __mod_memcg_lruvec_state(lruvec, idx, val);
850 void __mod_lruvec_page_state(struct page *page, enum node_stat_item idx,
853 struct page *head = compound_head(page); /* rmap on tail pages */
854 struct mem_cgroup *memcg = page_memcg(head);
855 pg_data_t *pgdat = page_pgdat(page);
856 struct lruvec *lruvec;
858 /* Untracked pages have no memcg, no lruvec. Update only the node */
860 __mod_node_page_state(pgdat, idx, val);
864 lruvec = mem_cgroup_lruvec(memcg, pgdat);
865 __mod_lruvec_state(lruvec, idx, val);
867 EXPORT_SYMBOL(__mod_lruvec_page_state);
869 void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val)
871 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
872 struct mem_cgroup *memcg;
873 struct lruvec *lruvec;
876 memcg = mem_cgroup_from_obj(p);
879 * Untracked pages have no memcg, no lruvec. Update only the
880 * node. If we reparent the slab objects to the root memcg,
881 * when we free the slab object, we need to update the per-memcg
882 * vmstats to keep it correct for the root memcg.
885 __mod_node_page_state(pgdat, idx, val);
887 lruvec = mem_cgroup_lruvec(memcg, pgdat);
888 __mod_lruvec_state(lruvec, idx, val);
894 * __count_memcg_events - account VM events in a cgroup
895 * @memcg: the memory cgroup
896 * @idx: the event item
897 * @count: the number of events that occured
899 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
904 if (mem_cgroup_disabled())
907 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
908 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
909 struct mem_cgroup *mi;
912 * Batch local counters to keep them in sync with
913 * the hierarchical ones.
915 __this_cpu_add(memcg->vmstats_local->events[idx], x);
916 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
917 atomic_long_add(x, &mi->vmevents[idx]);
920 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
923 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
925 return atomic_long_read(&memcg->vmevents[event]);
928 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
933 for_each_possible_cpu(cpu)
934 x += per_cpu(memcg->vmstats_local->events[event], cpu);
938 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
942 /* pagein of a big page is an event. So, ignore page size */
944 __count_memcg_events(memcg, PGPGIN, 1);
946 __count_memcg_events(memcg, PGPGOUT, 1);
947 nr_pages = -nr_pages; /* for event */
950 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
953 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
954 enum mem_cgroup_events_target target)
956 unsigned long val, next;
958 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
959 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
960 /* from time_after() in jiffies.h */
961 if ((long)(next - val) < 0) {
963 case MEM_CGROUP_TARGET_THRESH:
964 next = val + THRESHOLDS_EVENTS_TARGET;
966 case MEM_CGROUP_TARGET_SOFTLIMIT:
967 next = val + SOFTLIMIT_EVENTS_TARGET;
972 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
979 * Check events in order.
982 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
984 /* threshold event is triggered in finer grain than soft limit */
985 if (unlikely(mem_cgroup_event_ratelimit(memcg,
986 MEM_CGROUP_TARGET_THRESH))) {
989 do_softlimit = mem_cgroup_event_ratelimit(memcg,
990 MEM_CGROUP_TARGET_SOFTLIMIT);
991 mem_cgroup_threshold(memcg);
992 if (unlikely(do_softlimit))
993 mem_cgroup_update_tree(memcg, page);
997 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1000 * mm_update_next_owner() may clear mm->owner to NULL
1001 * if it races with swapoff, page migration, etc.
1002 * So this can be called with p == NULL.
1007 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1009 EXPORT_SYMBOL(mem_cgroup_from_task);
1012 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1013 * @mm: mm from which memcg should be extracted. It can be NULL.
1015 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
1016 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
1019 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1021 struct mem_cgroup *memcg;
1023 if (mem_cgroup_disabled())
1029 * Page cache insertions can happen withou an
1030 * actual mm context, e.g. during disk probing
1031 * on boot, loopback IO, acct() writes etc.
1034 memcg = root_mem_cgroup;
1036 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1037 if (unlikely(!memcg))
1038 memcg = root_mem_cgroup;
1040 } while (!css_tryget(&memcg->css));
1044 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1047 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
1048 * @page: page from which memcg should be extracted.
1050 * Obtain a reference on page->memcg and returns it if successful. Otherwise
1051 * root_mem_cgroup is returned.
1053 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
1055 struct mem_cgroup *memcg = page_memcg(page);
1057 if (mem_cgroup_disabled())
1061 /* Page should not get uncharged and freed memcg under us. */
1062 if (!memcg || WARN_ON_ONCE(!css_tryget(&memcg->css)))
1063 memcg = root_mem_cgroup;
1067 EXPORT_SYMBOL(get_mem_cgroup_from_page);
1069 static __always_inline struct mem_cgroup *active_memcg(void)
1072 return this_cpu_read(int_active_memcg);
1074 return current->active_memcg;
1077 static __always_inline struct mem_cgroup *get_active_memcg(void)
1079 struct mem_cgroup *memcg;
1082 memcg = active_memcg();
1084 /* current->active_memcg must hold a ref. */
1085 if (WARN_ON_ONCE(!css_tryget(&memcg->css)))
1086 memcg = root_mem_cgroup;
1088 memcg = current->active_memcg;
1095 static __always_inline bool memcg_kmem_bypass(void)
1097 /* Allow remote memcg charging from any context. */
1098 if (unlikely(active_memcg()))
1101 /* Memcg to charge can't be determined. */
1102 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
1109 * If active memcg is set, do not fallback to current->mm->memcg.
1111 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
1113 if (memcg_kmem_bypass())
1116 if (unlikely(active_memcg()))
1117 return get_active_memcg();
1119 return get_mem_cgroup_from_mm(current->mm);
1123 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1124 * @root: hierarchy root
1125 * @prev: previously returned memcg, NULL on first invocation
1126 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1128 * Returns references to children of the hierarchy below @root, or
1129 * @root itself, or %NULL after a full round-trip.
1131 * Caller must pass the return value in @prev on subsequent
1132 * invocations for reference counting, or use mem_cgroup_iter_break()
1133 * to cancel a hierarchy walk before the round-trip is complete.
1135 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1136 * in the hierarchy among all concurrent reclaimers operating on the
1139 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1140 struct mem_cgroup *prev,
1141 struct mem_cgroup_reclaim_cookie *reclaim)
1143 struct mem_cgroup_reclaim_iter *iter;
1144 struct cgroup_subsys_state *css = NULL;
1145 struct mem_cgroup *memcg = NULL;
1146 struct mem_cgroup *pos = NULL;
1148 if (mem_cgroup_disabled())
1152 root = root_mem_cgroup;
1154 if (prev && !reclaim)
1160 struct mem_cgroup_per_node *mz;
1162 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1165 if (prev && reclaim->generation != iter->generation)
1169 pos = READ_ONCE(iter->position);
1170 if (!pos || css_tryget(&pos->css))
1173 * css reference reached zero, so iter->position will
1174 * be cleared by ->css_released. However, we should not
1175 * rely on this happening soon, because ->css_released
1176 * is called from a work queue, and by busy-waiting we
1177 * might block it. So we clear iter->position right
1180 (void)cmpxchg(&iter->position, pos, NULL);
1188 css = css_next_descendant_pre(css, &root->css);
1191 * Reclaimers share the hierarchy walk, and a
1192 * new one might jump in right at the end of
1193 * the hierarchy - make sure they see at least
1194 * one group and restart from the beginning.
1202 * Verify the css and acquire a reference. The root
1203 * is provided by the caller, so we know it's alive
1204 * and kicking, and don't take an extra reference.
1206 memcg = mem_cgroup_from_css(css);
1208 if (css == &root->css)
1211 if (css_tryget(css))
1219 * The position could have already been updated by a competing
1220 * thread, so check that the value hasn't changed since we read
1221 * it to avoid reclaiming from the same cgroup twice.
1223 (void)cmpxchg(&iter->position, pos, memcg);
1231 reclaim->generation = iter->generation;
1236 if (prev && prev != root)
1237 css_put(&prev->css);
1243 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1244 * @root: hierarchy root
1245 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1247 void mem_cgroup_iter_break(struct mem_cgroup *root,
1248 struct mem_cgroup *prev)
1251 root = root_mem_cgroup;
1252 if (prev && prev != root)
1253 css_put(&prev->css);
1256 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1257 struct mem_cgroup *dead_memcg)
1259 struct mem_cgroup_reclaim_iter *iter;
1260 struct mem_cgroup_per_node *mz;
1263 for_each_node(nid) {
1264 mz = mem_cgroup_nodeinfo(from, nid);
1266 cmpxchg(&iter->position, dead_memcg, NULL);
1270 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1272 struct mem_cgroup *memcg = dead_memcg;
1273 struct mem_cgroup *last;
1276 __invalidate_reclaim_iterators(memcg, dead_memcg);
1278 } while ((memcg = parent_mem_cgroup(memcg)));
1281 * When cgruop1 non-hierarchy mode is used,
1282 * parent_mem_cgroup() does not walk all the way up to the
1283 * cgroup root (root_mem_cgroup). So we have to handle
1284 * dead_memcg from cgroup root separately.
1286 if (last != root_mem_cgroup)
1287 __invalidate_reclaim_iterators(root_mem_cgroup,
1292 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1293 * @memcg: hierarchy root
1294 * @fn: function to call for each task
1295 * @arg: argument passed to @fn
1297 * This function iterates over tasks attached to @memcg or to any of its
1298 * descendants and calls @fn for each task. If @fn returns a non-zero
1299 * value, the function breaks the iteration loop and returns the value.
1300 * Otherwise, it will iterate over all tasks and return 0.
1302 * This function must not be called for the root memory cgroup.
1304 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1305 int (*fn)(struct task_struct *, void *), void *arg)
1307 struct mem_cgroup *iter;
1310 BUG_ON(memcg == root_mem_cgroup);
1312 for_each_mem_cgroup_tree(iter, memcg) {
1313 struct css_task_iter it;
1314 struct task_struct *task;
1316 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1317 while (!ret && (task = css_task_iter_next(&it)))
1318 ret = fn(task, arg);
1319 css_task_iter_end(&it);
1321 mem_cgroup_iter_break(memcg, iter);
1328 #ifdef CONFIG_DEBUG_VM
1329 void lruvec_memcg_debug(struct lruvec *lruvec, struct page *page)
1331 struct mem_cgroup *memcg;
1333 if (mem_cgroup_disabled())
1336 memcg = page_memcg(page);
1339 VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != root_mem_cgroup, page);
1341 VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != memcg, page);
1346 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1348 * @pgdat: pgdat of the page
1350 * This function relies on page's memcg being stable - see the
1351 * access rules in commit_charge().
1353 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1355 struct mem_cgroup_per_node *mz;
1356 struct mem_cgroup *memcg;
1357 struct lruvec *lruvec;
1359 if (mem_cgroup_disabled()) {
1360 lruvec = &pgdat->__lruvec;
1364 memcg = page_memcg(page);
1365 VM_WARN_ON_ONCE_PAGE(!memcg, page);
1367 memcg = root_mem_cgroup;
1369 mz = mem_cgroup_page_nodeinfo(memcg, page);
1370 lruvec = &mz->lruvec;
1373 * Since a node can be onlined after the mem_cgroup was created,
1374 * we have to be prepared to initialize lruvec->zone here;
1375 * and if offlined then reonlined, we need to reinitialize it.
1377 if (unlikely(lruvec->pgdat != pgdat))
1378 lruvec->pgdat = pgdat;
1383 * lock_page_lruvec - lock and return lruvec for a given page.
1386 * This series functions should be used in either conditions:
1387 * PageLRU is cleared or unset
1388 * or page->_refcount is zero
1389 * or page is locked.
1391 struct lruvec *lock_page_lruvec(struct page *page)
1393 struct lruvec *lruvec;
1394 struct pglist_data *pgdat = page_pgdat(page);
1397 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1398 spin_lock(&lruvec->lru_lock);
1401 lruvec_memcg_debug(lruvec, page);
1406 struct lruvec *lock_page_lruvec_irq(struct page *page)
1408 struct lruvec *lruvec;
1409 struct pglist_data *pgdat = page_pgdat(page);
1412 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1413 spin_lock_irq(&lruvec->lru_lock);
1416 lruvec_memcg_debug(lruvec, page);
1421 struct lruvec *lock_page_lruvec_irqsave(struct page *page, unsigned long *flags)
1423 struct lruvec *lruvec;
1424 struct pglist_data *pgdat = page_pgdat(page);
1427 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1428 spin_lock_irqsave(&lruvec->lru_lock, *flags);
1431 lruvec_memcg_debug(lruvec, page);
1437 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1438 * @lruvec: mem_cgroup per zone lru vector
1439 * @lru: index of lru list the page is sitting on
1440 * @zid: zone id of the accounted pages
1441 * @nr_pages: positive when adding or negative when removing
1443 * This function must be called under lru_lock, just before a page is added
1444 * to or just after a page is removed from an lru list (that ordering being
1445 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1447 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1448 int zid, int nr_pages)
1450 struct mem_cgroup_per_node *mz;
1451 unsigned long *lru_size;
1454 if (mem_cgroup_disabled())
1457 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1458 lru_size = &mz->lru_zone_size[zid][lru];
1461 *lru_size += nr_pages;
1464 if (WARN_ONCE(size < 0,
1465 "%s(%p, %d, %d): lru_size %ld\n",
1466 __func__, lruvec, lru, nr_pages, size)) {
1472 *lru_size += nr_pages;
1476 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1477 * @memcg: the memory cgroup
1479 * Returns the maximum amount of memory @mem can be charged with, in
1482 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1484 unsigned long margin = 0;
1485 unsigned long count;
1486 unsigned long limit;
1488 count = page_counter_read(&memcg->memory);
1489 limit = READ_ONCE(memcg->memory.max);
1491 margin = limit - count;
1493 if (do_memsw_account()) {
1494 count = page_counter_read(&memcg->memsw);
1495 limit = READ_ONCE(memcg->memsw.max);
1497 margin = min(margin, limit - count);
1506 * A routine for checking "mem" is under move_account() or not.
1508 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1509 * moving cgroups. This is for waiting at high-memory pressure
1512 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1514 struct mem_cgroup *from;
1515 struct mem_cgroup *to;
1518 * Unlike task_move routines, we access mc.to, mc.from not under
1519 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1521 spin_lock(&mc.lock);
1527 ret = mem_cgroup_is_descendant(from, memcg) ||
1528 mem_cgroup_is_descendant(to, memcg);
1530 spin_unlock(&mc.lock);
1534 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1536 if (mc.moving_task && current != mc.moving_task) {
1537 if (mem_cgroup_under_move(memcg)) {
1539 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1540 /* moving charge context might have finished. */
1543 finish_wait(&mc.waitq, &wait);
1550 struct memory_stat {
1556 static struct memory_stat memory_stats[] = {
1557 { "anon", PAGE_SIZE, NR_ANON_MAPPED },
1558 { "file", PAGE_SIZE, NR_FILE_PAGES },
1559 { "kernel_stack", 1024, NR_KERNEL_STACK_KB },
1560 { "pagetables", PAGE_SIZE, NR_PAGETABLE },
1561 { "percpu", 1, MEMCG_PERCPU_B },
1562 { "sock", PAGE_SIZE, MEMCG_SOCK },
1563 { "shmem", PAGE_SIZE, NR_SHMEM },
1564 { "file_mapped", PAGE_SIZE, NR_FILE_MAPPED },
1565 { "file_dirty", PAGE_SIZE, NR_FILE_DIRTY },
1566 { "file_writeback", PAGE_SIZE, NR_WRITEBACK },
1567 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1569 * The ratio will be initialized in memory_stats_init(). Because
1570 * on some architectures, the macro of HPAGE_PMD_SIZE is not
1571 * constant(e.g. powerpc).
1573 { "anon_thp", 0, NR_ANON_THPS },
1574 { "file_thp", 0, NR_FILE_THPS },
1575 { "shmem_thp", 0, NR_SHMEM_THPS },
1577 { "inactive_anon", PAGE_SIZE, NR_INACTIVE_ANON },
1578 { "active_anon", PAGE_SIZE, NR_ACTIVE_ANON },
1579 { "inactive_file", PAGE_SIZE, NR_INACTIVE_FILE },
1580 { "active_file", PAGE_SIZE, NR_ACTIVE_FILE },
1581 { "unevictable", PAGE_SIZE, NR_UNEVICTABLE },
1584 * Note: The slab_reclaimable and slab_unreclaimable must be
1585 * together and slab_reclaimable must be in front.
1587 { "slab_reclaimable", 1, NR_SLAB_RECLAIMABLE_B },
1588 { "slab_unreclaimable", 1, NR_SLAB_UNRECLAIMABLE_B },
1590 /* The memory events */
1591 { "workingset_refault_anon", 1, WORKINGSET_REFAULT_ANON },
1592 { "workingset_refault_file", 1, WORKINGSET_REFAULT_FILE },
1593 { "workingset_activate_anon", 1, WORKINGSET_ACTIVATE_ANON },
1594 { "workingset_activate_file", 1, WORKINGSET_ACTIVATE_FILE },
1595 { "workingset_restore_anon", 1, WORKINGSET_RESTORE_ANON },
1596 { "workingset_restore_file", 1, WORKINGSET_RESTORE_FILE },
1597 { "workingset_nodereclaim", 1, WORKINGSET_NODERECLAIM },
1600 static int __init memory_stats_init(void)
1604 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1605 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1606 if (memory_stats[i].idx == NR_ANON_THPS ||
1607 memory_stats[i].idx == NR_FILE_THPS ||
1608 memory_stats[i].idx == NR_SHMEM_THPS)
1609 memory_stats[i].ratio = HPAGE_PMD_SIZE;
1611 VM_BUG_ON(!memory_stats[i].ratio);
1612 VM_BUG_ON(memory_stats[i].idx >= MEMCG_NR_STAT);
1617 pure_initcall(memory_stats_init);
1619 static char *memory_stat_format(struct mem_cgroup *memcg)
1624 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1629 * Provide statistics on the state of the memory subsystem as
1630 * well as cumulative event counters that show past behavior.
1632 * This list is ordered following a combination of these gradients:
1633 * 1) generic big picture -> specifics and details
1634 * 2) reflecting userspace activity -> reflecting kernel heuristics
1636 * Current memory state:
1639 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1642 size = memcg_page_state(memcg, memory_stats[i].idx);
1643 size *= memory_stats[i].ratio;
1644 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1646 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1647 size = memcg_page_state(memcg, NR_SLAB_RECLAIMABLE_B) +
1648 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE_B);
1649 seq_buf_printf(&s, "slab %llu\n", size);
1653 /* Accumulated memory events */
1655 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1656 memcg_events(memcg, PGFAULT));
1657 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1658 memcg_events(memcg, PGMAJFAULT));
1659 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1660 memcg_events(memcg, PGREFILL));
1661 seq_buf_printf(&s, "pgscan %lu\n",
1662 memcg_events(memcg, PGSCAN_KSWAPD) +
1663 memcg_events(memcg, PGSCAN_DIRECT));
1664 seq_buf_printf(&s, "pgsteal %lu\n",
1665 memcg_events(memcg, PGSTEAL_KSWAPD) +
1666 memcg_events(memcg, PGSTEAL_DIRECT));
1667 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1668 memcg_events(memcg, PGACTIVATE));
1669 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1670 memcg_events(memcg, PGDEACTIVATE));
1671 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1672 memcg_events(memcg, PGLAZYFREE));
1673 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1674 memcg_events(memcg, PGLAZYFREED));
1676 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1677 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1678 memcg_events(memcg, THP_FAULT_ALLOC));
1679 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1680 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1681 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1683 /* The above should easily fit into one page */
1684 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1689 #define K(x) ((x) << (PAGE_SHIFT-10))
1691 * mem_cgroup_print_oom_context: Print OOM information relevant to
1692 * memory controller.
1693 * @memcg: The memory cgroup that went over limit
1694 * @p: Task that is going to be killed
1696 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1699 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1704 pr_cont(",oom_memcg=");
1705 pr_cont_cgroup_path(memcg->css.cgroup);
1707 pr_cont(",global_oom");
1709 pr_cont(",task_memcg=");
1710 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1716 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1717 * memory controller.
1718 * @memcg: The memory cgroup that went over limit
1720 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1724 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1725 K((u64)page_counter_read(&memcg->memory)),
1726 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1727 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1728 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1729 K((u64)page_counter_read(&memcg->swap)),
1730 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1732 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1733 K((u64)page_counter_read(&memcg->memsw)),
1734 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1735 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1736 K((u64)page_counter_read(&memcg->kmem)),
1737 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1740 pr_info("Memory cgroup stats for ");
1741 pr_cont_cgroup_path(memcg->css.cgroup);
1743 buf = memory_stat_format(memcg);
1751 * Return the memory (and swap, if configured) limit for a memcg.
1753 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1755 unsigned long max = READ_ONCE(memcg->memory.max);
1757 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1758 if (mem_cgroup_swappiness(memcg))
1759 max += min(READ_ONCE(memcg->swap.max),
1760 (unsigned long)total_swap_pages);
1762 if (mem_cgroup_swappiness(memcg)) {
1763 /* Calculate swap excess capacity from memsw limit */
1764 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1766 max += min(swap, (unsigned long)total_swap_pages);
1772 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1774 return page_counter_read(&memcg->memory);
1777 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1780 struct oom_control oc = {
1784 .gfp_mask = gfp_mask,
1789 if (mutex_lock_killable(&oom_lock))
1792 if (mem_cgroup_margin(memcg) >= (1 << order))
1796 * A few threads which were not waiting at mutex_lock_killable() can
1797 * fail to bail out. Therefore, check again after holding oom_lock.
1799 ret = should_force_charge() || out_of_memory(&oc);
1802 mutex_unlock(&oom_lock);
1806 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1809 unsigned long *total_scanned)
1811 struct mem_cgroup *victim = NULL;
1814 unsigned long excess;
1815 unsigned long nr_scanned;
1816 struct mem_cgroup_reclaim_cookie reclaim = {
1820 excess = soft_limit_excess(root_memcg);
1823 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1828 * If we have not been able to reclaim
1829 * anything, it might because there are
1830 * no reclaimable pages under this hierarchy
1835 * We want to do more targeted reclaim.
1836 * excess >> 2 is not to excessive so as to
1837 * reclaim too much, nor too less that we keep
1838 * coming back to reclaim from this cgroup
1840 if (total >= (excess >> 2) ||
1841 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1846 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1847 pgdat, &nr_scanned);
1848 *total_scanned += nr_scanned;
1849 if (!soft_limit_excess(root_memcg))
1852 mem_cgroup_iter_break(root_memcg, victim);
1856 #ifdef CONFIG_LOCKDEP
1857 static struct lockdep_map memcg_oom_lock_dep_map = {
1858 .name = "memcg_oom_lock",
1862 static DEFINE_SPINLOCK(memcg_oom_lock);
1865 * Check OOM-Killer is already running under our hierarchy.
1866 * If someone is running, return false.
1868 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1870 struct mem_cgroup *iter, *failed = NULL;
1872 spin_lock(&memcg_oom_lock);
1874 for_each_mem_cgroup_tree(iter, memcg) {
1875 if (iter->oom_lock) {
1877 * this subtree of our hierarchy is already locked
1878 * so we cannot give a lock.
1881 mem_cgroup_iter_break(memcg, iter);
1884 iter->oom_lock = true;
1889 * OK, we failed to lock the whole subtree so we have
1890 * to clean up what we set up to the failing subtree
1892 for_each_mem_cgroup_tree(iter, memcg) {
1893 if (iter == failed) {
1894 mem_cgroup_iter_break(memcg, iter);
1897 iter->oom_lock = false;
1900 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1902 spin_unlock(&memcg_oom_lock);
1907 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1909 struct mem_cgroup *iter;
1911 spin_lock(&memcg_oom_lock);
1912 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1913 for_each_mem_cgroup_tree(iter, memcg)
1914 iter->oom_lock = false;
1915 spin_unlock(&memcg_oom_lock);
1918 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1920 struct mem_cgroup *iter;
1922 spin_lock(&memcg_oom_lock);
1923 for_each_mem_cgroup_tree(iter, memcg)
1925 spin_unlock(&memcg_oom_lock);
1928 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1930 struct mem_cgroup *iter;
1933 * Be careful about under_oom underflows becase a child memcg
1934 * could have been added after mem_cgroup_mark_under_oom.
1936 spin_lock(&memcg_oom_lock);
1937 for_each_mem_cgroup_tree(iter, memcg)
1938 if (iter->under_oom > 0)
1940 spin_unlock(&memcg_oom_lock);
1943 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1945 struct oom_wait_info {
1946 struct mem_cgroup *memcg;
1947 wait_queue_entry_t wait;
1950 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1951 unsigned mode, int sync, void *arg)
1953 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1954 struct mem_cgroup *oom_wait_memcg;
1955 struct oom_wait_info *oom_wait_info;
1957 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1958 oom_wait_memcg = oom_wait_info->memcg;
1960 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1961 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1963 return autoremove_wake_function(wait, mode, sync, arg);
1966 static void memcg_oom_recover(struct mem_cgroup *memcg)
1969 * For the following lockless ->under_oom test, the only required
1970 * guarantee is that it must see the state asserted by an OOM when
1971 * this function is called as a result of userland actions
1972 * triggered by the notification of the OOM. This is trivially
1973 * achieved by invoking mem_cgroup_mark_under_oom() before
1974 * triggering notification.
1976 if (memcg && memcg->under_oom)
1977 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1987 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1989 enum oom_status ret;
1992 if (order > PAGE_ALLOC_COSTLY_ORDER)
1995 memcg_memory_event(memcg, MEMCG_OOM);
1998 * We are in the middle of the charge context here, so we
1999 * don't want to block when potentially sitting on a callstack
2000 * that holds all kinds of filesystem and mm locks.
2002 * cgroup1 allows disabling the OOM killer and waiting for outside
2003 * handling until the charge can succeed; remember the context and put
2004 * the task to sleep at the end of the page fault when all locks are
2007 * On the other hand, in-kernel OOM killer allows for an async victim
2008 * memory reclaim (oom_reaper) and that means that we are not solely
2009 * relying on the oom victim to make a forward progress and we can
2010 * invoke the oom killer here.
2012 * Please note that mem_cgroup_out_of_memory might fail to find a
2013 * victim and then we have to bail out from the charge path.
2015 if (memcg->oom_kill_disable) {
2016 if (!current->in_user_fault)
2018 css_get(&memcg->css);
2019 current->memcg_in_oom = memcg;
2020 current->memcg_oom_gfp_mask = mask;
2021 current->memcg_oom_order = order;
2026 mem_cgroup_mark_under_oom(memcg);
2028 locked = mem_cgroup_oom_trylock(memcg);
2031 mem_cgroup_oom_notify(memcg);
2033 mem_cgroup_unmark_under_oom(memcg);
2034 if (mem_cgroup_out_of_memory(memcg, mask, order))
2040 mem_cgroup_oom_unlock(memcg);
2046 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2047 * @handle: actually kill/wait or just clean up the OOM state
2049 * This has to be called at the end of a page fault if the memcg OOM
2050 * handler was enabled.
2052 * Memcg supports userspace OOM handling where failed allocations must
2053 * sleep on a waitqueue until the userspace task resolves the
2054 * situation. Sleeping directly in the charge context with all kinds
2055 * of locks held is not a good idea, instead we remember an OOM state
2056 * in the task and mem_cgroup_oom_synchronize() has to be called at
2057 * the end of the page fault to complete the OOM handling.
2059 * Returns %true if an ongoing memcg OOM situation was detected and
2060 * completed, %false otherwise.
2062 bool mem_cgroup_oom_synchronize(bool handle)
2064 struct mem_cgroup *memcg = current->memcg_in_oom;
2065 struct oom_wait_info owait;
2068 /* OOM is global, do not handle */
2075 owait.memcg = memcg;
2076 owait.wait.flags = 0;
2077 owait.wait.func = memcg_oom_wake_function;
2078 owait.wait.private = current;
2079 INIT_LIST_HEAD(&owait.wait.entry);
2081 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2082 mem_cgroup_mark_under_oom(memcg);
2084 locked = mem_cgroup_oom_trylock(memcg);
2087 mem_cgroup_oom_notify(memcg);
2089 if (locked && !memcg->oom_kill_disable) {
2090 mem_cgroup_unmark_under_oom(memcg);
2091 finish_wait(&memcg_oom_waitq, &owait.wait);
2092 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
2093 current->memcg_oom_order);
2096 mem_cgroup_unmark_under_oom(memcg);
2097 finish_wait(&memcg_oom_waitq, &owait.wait);
2101 mem_cgroup_oom_unlock(memcg);
2103 * There is no guarantee that an OOM-lock contender
2104 * sees the wakeups triggered by the OOM kill
2105 * uncharges. Wake any sleepers explicitely.
2107 memcg_oom_recover(memcg);
2110 current->memcg_in_oom = NULL;
2111 css_put(&memcg->css);
2116 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2117 * @victim: task to be killed by the OOM killer
2118 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2120 * Returns a pointer to a memory cgroup, which has to be cleaned up
2121 * by killing all belonging OOM-killable tasks.
2123 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2125 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2126 struct mem_cgroup *oom_domain)
2128 struct mem_cgroup *oom_group = NULL;
2129 struct mem_cgroup *memcg;
2131 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2135 oom_domain = root_mem_cgroup;
2139 memcg = mem_cgroup_from_task(victim);
2140 if (memcg == root_mem_cgroup)
2144 * If the victim task has been asynchronously moved to a different
2145 * memory cgroup, we might end up killing tasks outside oom_domain.
2146 * In this case it's better to ignore memory.group.oom.
2148 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2152 * Traverse the memory cgroup hierarchy from the victim task's
2153 * cgroup up to the OOMing cgroup (or root) to find the
2154 * highest-level memory cgroup with oom.group set.
2156 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2157 if (memcg->oom_group)
2160 if (memcg == oom_domain)
2165 css_get(&oom_group->css);
2172 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2174 pr_info("Tasks in ");
2175 pr_cont_cgroup_path(memcg->css.cgroup);
2176 pr_cont(" are going to be killed due to memory.oom.group set\n");
2180 * lock_page_memcg - lock a page and memcg binding
2183 * This function protects unlocked LRU pages from being moved to
2186 * It ensures lifetime of the returned memcg. Caller is responsible
2187 * for the lifetime of the page; __unlock_page_memcg() is available
2188 * when @page might get freed inside the locked section.
2190 struct mem_cgroup *lock_page_memcg(struct page *page)
2192 struct page *head = compound_head(page); /* rmap on tail pages */
2193 struct mem_cgroup *memcg;
2194 unsigned long flags;
2197 * The RCU lock is held throughout the transaction. The fast
2198 * path can get away without acquiring the memcg->move_lock
2199 * because page moving starts with an RCU grace period.
2201 * The RCU lock also protects the memcg from being freed when
2202 * the page state that is going to change is the only thing
2203 * preventing the page itself from being freed. E.g. writeback
2204 * doesn't hold a page reference and relies on PG_writeback to
2205 * keep off truncation, migration and so forth.
2209 if (mem_cgroup_disabled())
2212 memcg = page_memcg(head);
2213 if (unlikely(!memcg))
2216 #ifdef CONFIG_PROVE_LOCKING
2217 local_irq_save(flags);
2218 might_lock(&memcg->move_lock);
2219 local_irq_restore(flags);
2222 if (atomic_read(&memcg->moving_account) <= 0)
2225 spin_lock_irqsave(&memcg->move_lock, flags);
2226 if (memcg != page_memcg(head)) {
2227 spin_unlock_irqrestore(&memcg->move_lock, flags);
2232 * When charge migration first begins, we can have locked and
2233 * unlocked page stat updates happening concurrently. Track
2234 * the task who has the lock for unlock_page_memcg().
2236 memcg->move_lock_task = current;
2237 memcg->move_lock_flags = flags;
2241 EXPORT_SYMBOL(lock_page_memcg);
2244 * __unlock_page_memcg - unlock and unpin a memcg
2247 * Unlock and unpin a memcg returned by lock_page_memcg().
2249 void __unlock_page_memcg(struct mem_cgroup *memcg)
2251 if (memcg && memcg->move_lock_task == current) {
2252 unsigned long flags = memcg->move_lock_flags;
2254 memcg->move_lock_task = NULL;
2255 memcg->move_lock_flags = 0;
2257 spin_unlock_irqrestore(&memcg->move_lock, flags);
2264 * unlock_page_memcg - unlock a page and memcg binding
2267 void unlock_page_memcg(struct page *page)
2269 struct page *head = compound_head(page);
2271 __unlock_page_memcg(page_memcg(head));
2273 EXPORT_SYMBOL(unlock_page_memcg);
2275 struct memcg_stock_pcp {
2276 struct mem_cgroup *cached; /* this never be root cgroup */
2277 unsigned int nr_pages;
2279 #ifdef CONFIG_MEMCG_KMEM
2280 struct obj_cgroup *cached_objcg;
2281 unsigned int nr_bytes;
2284 struct work_struct work;
2285 unsigned long flags;
2286 #define FLUSHING_CACHED_CHARGE 0
2288 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2289 static DEFINE_MUTEX(percpu_charge_mutex);
2291 #ifdef CONFIG_MEMCG_KMEM
2292 static void drain_obj_stock(struct memcg_stock_pcp *stock);
2293 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2294 struct mem_cgroup *root_memcg);
2297 static inline void drain_obj_stock(struct memcg_stock_pcp *stock)
2300 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2301 struct mem_cgroup *root_memcg)
2308 * consume_stock: Try to consume stocked charge on this cpu.
2309 * @memcg: memcg to consume from.
2310 * @nr_pages: how many pages to charge.
2312 * The charges will only happen if @memcg matches the current cpu's memcg
2313 * stock, and at least @nr_pages are available in that stock. Failure to
2314 * service an allocation will refill the stock.
2316 * returns true if successful, false otherwise.
2318 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2320 struct memcg_stock_pcp *stock;
2321 unsigned long flags;
2324 if (nr_pages > MEMCG_CHARGE_BATCH)
2327 local_irq_save(flags);
2329 stock = this_cpu_ptr(&memcg_stock);
2330 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2331 stock->nr_pages -= nr_pages;
2335 local_irq_restore(flags);
2341 * Returns stocks cached in percpu and reset cached information.
2343 static void drain_stock(struct memcg_stock_pcp *stock)
2345 struct mem_cgroup *old = stock->cached;
2350 if (stock->nr_pages) {
2351 page_counter_uncharge(&old->memory, stock->nr_pages);
2352 if (do_memsw_account())
2353 page_counter_uncharge(&old->memsw, stock->nr_pages);
2354 stock->nr_pages = 0;
2358 stock->cached = NULL;
2361 static void drain_local_stock(struct work_struct *dummy)
2363 struct memcg_stock_pcp *stock;
2364 unsigned long flags;
2367 * The only protection from memory hotplug vs. drain_stock races is
2368 * that we always operate on local CPU stock here with IRQ disabled
2370 local_irq_save(flags);
2372 stock = this_cpu_ptr(&memcg_stock);
2373 drain_obj_stock(stock);
2375 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2377 local_irq_restore(flags);
2381 * Cache charges(val) to local per_cpu area.
2382 * This will be consumed by consume_stock() function, later.
2384 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2386 struct memcg_stock_pcp *stock;
2387 unsigned long flags;
2389 local_irq_save(flags);
2391 stock = this_cpu_ptr(&memcg_stock);
2392 if (stock->cached != memcg) { /* reset if necessary */
2394 css_get(&memcg->css);
2395 stock->cached = memcg;
2397 stock->nr_pages += nr_pages;
2399 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2402 local_irq_restore(flags);
2406 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2407 * of the hierarchy under it.
2409 static void drain_all_stock(struct mem_cgroup *root_memcg)
2413 /* If someone's already draining, avoid adding running more workers. */
2414 if (!mutex_trylock(&percpu_charge_mutex))
2417 * Notify other cpus that system-wide "drain" is running
2418 * We do not care about races with the cpu hotplug because cpu down
2419 * as well as workers from this path always operate on the local
2420 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2423 for_each_online_cpu(cpu) {
2424 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2425 struct mem_cgroup *memcg;
2429 memcg = stock->cached;
2430 if (memcg && stock->nr_pages &&
2431 mem_cgroup_is_descendant(memcg, root_memcg))
2433 if (obj_stock_flush_required(stock, root_memcg))
2438 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2440 drain_local_stock(&stock->work);
2442 schedule_work_on(cpu, &stock->work);
2446 mutex_unlock(&percpu_charge_mutex);
2449 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2451 struct memcg_stock_pcp *stock;
2452 struct mem_cgroup *memcg, *mi;
2454 stock = &per_cpu(memcg_stock, cpu);
2457 for_each_mem_cgroup(memcg) {
2460 for (i = 0; i < MEMCG_NR_STAT; i++) {
2464 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2466 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2467 atomic_long_add(x, &memcg->vmstats[i]);
2469 if (i >= NR_VM_NODE_STAT_ITEMS)
2472 for_each_node(nid) {
2473 struct mem_cgroup_per_node *pn;
2475 pn = mem_cgroup_nodeinfo(memcg, nid);
2476 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2479 atomic_long_add(x, &pn->lruvec_stat[i]);
2480 } while ((pn = parent_nodeinfo(pn, nid)));
2484 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2487 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2489 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2490 atomic_long_add(x, &memcg->vmevents[i]);
2497 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2498 unsigned int nr_pages,
2501 unsigned long nr_reclaimed = 0;
2504 unsigned long pflags;
2506 if (page_counter_read(&memcg->memory) <=
2507 READ_ONCE(memcg->memory.high))
2510 memcg_memory_event(memcg, MEMCG_HIGH);
2512 psi_memstall_enter(&pflags);
2513 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2515 psi_memstall_leave(&pflags);
2516 } while ((memcg = parent_mem_cgroup(memcg)) &&
2517 !mem_cgroup_is_root(memcg));
2519 return nr_reclaimed;
2522 static void high_work_func(struct work_struct *work)
2524 struct mem_cgroup *memcg;
2526 memcg = container_of(work, struct mem_cgroup, high_work);
2527 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2531 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2532 * enough to still cause a significant slowdown in most cases, while still
2533 * allowing diagnostics and tracing to proceed without becoming stuck.
2535 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2538 * When calculating the delay, we use these either side of the exponentiation to
2539 * maintain precision and scale to a reasonable number of jiffies (see the table
2542 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2543 * overage ratio to a delay.
2544 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2545 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2546 * to produce a reasonable delay curve.
2548 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2549 * reasonable delay curve compared to precision-adjusted overage, not
2550 * penalising heavily at first, but still making sure that growth beyond the
2551 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2552 * example, with a high of 100 megabytes:
2554 * +-------+------------------------+
2555 * | usage | time to allocate in ms |
2556 * +-------+------------------------+
2578 * +-------+------------------------+
2580 #define MEMCG_DELAY_PRECISION_SHIFT 20
2581 #define MEMCG_DELAY_SCALING_SHIFT 14
2583 static u64 calculate_overage(unsigned long usage, unsigned long high)
2591 * Prevent division by 0 in overage calculation by acting as if
2592 * it was a threshold of 1 page
2594 high = max(high, 1UL);
2596 overage = usage - high;
2597 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2598 return div64_u64(overage, high);
2601 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2603 u64 overage, max_overage = 0;
2606 overage = calculate_overage(page_counter_read(&memcg->memory),
2607 READ_ONCE(memcg->memory.high));
2608 max_overage = max(overage, max_overage);
2609 } while ((memcg = parent_mem_cgroup(memcg)) &&
2610 !mem_cgroup_is_root(memcg));
2615 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2617 u64 overage, max_overage = 0;
2620 overage = calculate_overage(page_counter_read(&memcg->swap),
2621 READ_ONCE(memcg->swap.high));
2623 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2624 max_overage = max(overage, max_overage);
2625 } while ((memcg = parent_mem_cgroup(memcg)) &&
2626 !mem_cgroup_is_root(memcg));
2632 * Get the number of jiffies that we should penalise a mischievous cgroup which
2633 * is exceeding its memory.high by checking both it and its ancestors.
2635 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2636 unsigned int nr_pages,
2639 unsigned long penalty_jiffies;
2645 * We use overage compared to memory.high to calculate the number of
2646 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2647 * fairly lenient on small overages, and increasingly harsh when the
2648 * memcg in question makes it clear that it has no intention of stopping
2649 * its crazy behaviour, so we exponentially increase the delay based on
2652 penalty_jiffies = max_overage * max_overage * HZ;
2653 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2654 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2657 * Factor in the task's own contribution to the overage, such that four
2658 * N-sized allocations are throttled approximately the same as one
2659 * 4N-sized allocation.
2661 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2662 * larger the current charge patch is than that.
2664 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2668 * Scheduled by try_charge() to be executed from the userland return path
2669 * and reclaims memory over the high limit.
2671 void mem_cgroup_handle_over_high(void)
2673 unsigned long penalty_jiffies;
2674 unsigned long pflags;
2675 unsigned long nr_reclaimed;
2676 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2677 int nr_retries = MAX_RECLAIM_RETRIES;
2678 struct mem_cgroup *memcg;
2679 bool in_retry = false;
2681 if (likely(!nr_pages))
2684 memcg = get_mem_cgroup_from_mm(current->mm);
2685 current->memcg_nr_pages_over_high = 0;
2689 * The allocating task should reclaim at least the batch size, but for
2690 * subsequent retries we only want to do what's necessary to prevent oom
2691 * or breaching resource isolation.
2693 * This is distinct from memory.max or page allocator behaviour because
2694 * memory.high is currently batched, whereas memory.max and the page
2695 * allocator run every time an allocation is made.
2697 nr_reclaimed = reclaim_high(memcg,
2698 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2702 * memory.high is breached and reclaim is unable to keep up. Throttle
2703 * allocators proactively to slow down excessive growth.
2705 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2706 mem_find_max_overage(memcg));
2708 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2709 swap_find_max_overage(memcg));
2712 * Clamp the max delay per usermode return so as to still keep the
2713 * application moving forwards and also permit diagnostics, albeit
2716 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2719 * Don't sleep if the amount of jiffies this memcg owes us is so low
2720 * that it's not even worth doing, in an attempt to be nice to those who
2721 * go only a small amount over their memory.high value and maybe haven't
2722 * been aggressively reclaimed enough yet.
2724 if (penalty_jiffies <= HZ / 100)
2728 * If reclaim is making forward progress but we're still over
2729 * memory.high, we want to encourage that rather than doing allocator
2732 if (nr_reclaimed || nr_retries--) {
2738 * If we exit early, we're guaranteed to die (since
2739 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2740 * need to account for any ill-begotten jiffies to pay them off later.
2742 psi_memstall_enter(&pflags);
2743 schedule_timeout_killable(penalty_jiffies);
2744 psi_memstall_leave(&pflags);
2747 css_put(&memcg->css);
2750 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2751 unsigned int nr_pages)
2753 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2754 int nr_retries = MAX_RECLAIM_RETRIES;
2755 struct mem_cgroup *mem_over_limit;
2756 struct page_counter *counter;
2757 enum oom_status oom_status;
2758 unsigned long nr_reclaimed;
2759 bool may_swap = true;
2760 bool drained = false;
2761 unsigned long pflags;
2763 if (mem_cgroup_is_root(memcg))
2766 if (consume_stock(memcg, nr_pages))
2769 if (!do_memsw_account() ||
2770 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2771 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2773 if (do_memsw_account())
2774 page_counter_uncharge(&memcg->memsw, batch);
2775 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2777 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2781 if (batch > nr_pages) {
2787 * Memcg doesn't have a dedicated reserve for atomic
2788 * allocations. But like the global atomic pool, we need to
2789 * put the burden of reclaim on regular allocation requests
2790 * and let these go through as privileged allocations.
2792 if (gfp_mask & __GFP_ATOMIC)
2796 * Unlike in global OOM situations, memcg is not in a physical
2797 * memory shortage. Allow dying and OOM-killed tasks to
2798 * bypass the last charges so that they can exit quickly and
2799 * free their memory.
2801 if (unlikely(should_force_charge()))
2805 * Prevent unbounded recursion when reclaim operations need to
2806 * allocate memory. This might exceed the limits temporarily,
2807 * but we prefer facilitating memory reclaim and getting back
2808 * under the limit over triggering OOM kills in these cases.
2810 if (unlikely(current->flags & PF_MEMALLOC))
2813 if (unlikely(task_in_memcg_oom(current)))
2816 if (!gfpflags_allow_blocking(gfp_mask))
2819 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2821 psi_memstall_enter(&pflags);
2822 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2823 gfp_mask, may_swap);
2824 psi_memstall_leave(&pflags);
2826 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2830 drain_all_stock(mem_over_limit);
2835 if (gfp_mask & __GFP_NORETRY)
2838 * Even though the limit is exceeded at this point, reclaim
2839 * may have been able to free some pages. Retry the charge
2840 * before killing the task.
2842 * Only for regular pages, though: huge pages are rather
2843 * unlikely to succeed so close to the limit, and we fall back
2844 * to regular pages anyway in case of failure.
2846 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2849 * At task move, charge accounts can be doubly counted. So, it's
2850 * better to wait until the end of task_move if something is going on.
2852 if (mem_cgroup_wait_acct_move(mem_over_limit))
2858 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2861 if (gfp_mask & __GFP_NOFAIL)
2864 if (fatal_signal_pending(current))
2868 * keep retrying as long as the memcg oom killer is able to make
2869 * a forward progress or bypass the charge if the oom killer
2870 * couldn't make any progress.
2872 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2873 get_order(nr_pages * PAGE_SIZE));
2874 switch (oom_status) {
2876 nr_retries = MAX_RECLAIM_RETRIES;
2884 if (!(gfp_mask & __GFP_NOFAIL))
2888 * The allocation either can't fail or will lead to more memory
2889 * being freed very soon. Allow memory usage go over the limit
2890 * temporarily by force charging it.
2892 page_counter_charge(&memcg->memory, nr_pages);
2893 if (do_memsw_account())
2894 page_counter_charge(&memcg->memsw, nr_pages);
2899 if (batch > nr_pages)
2900 refill_stock(memcg, batch - nr_pages);
2903 * If the hierarchy is above the normal consumption range, schedule
2904 * reclaim on returning to userland. We can perform reclaim here
2905 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2906 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2907 * not recorded as it most likely matches current's and won't
2908 * change in the meantime. As high limit is checked again before
2909 * reclaim, the cost of mismatch is negligible.
2912 bool mem_high, swap_high;
2914 mem_high = page_counter_read(&memcg->memory) >
2915 READ_ONCE(memcg->memory.high);
2916 swap_high = page_counter_read(&memcg->swap) >
2917 READ_ONCE(memcg->swap.high);
2919 /* Don't bother a random interrupted task */
2920 if (in_interrupt()) {
2922 schedule_work(&memcg->high_work);
2928 if (mem_high || swap_high) {
2930 * The allocating tasks in this cgroup will need to do
2931 * reclaim or be throttled to prevent further growth
2932 * of the memory or swap footprints.
2934 * Target some best-effort fairness between the tasks,
2935 * and distribute reclaim work and delay penalties
2936 * based on how much each task is actually allocating.
2938 current->memcg_nr_pages_over_high += batch;
2939 set_notify_resume(current);
2942 } while ((memcg = parent_mem_cgroup(memcg)));
2947 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2948 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2950 if (mem_cgroup_is_root(memcg))
2953 page_counter_uncharge(&memcg->memory, nr_pages);
2954 if (do_memsw_account())
2955 page_counter_uncharge(&memcg->memsw, nr_pages);
2959 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2961 VM_BUG_ON_PAGE(page_memcg(page), page);
2963 * Any of the following ensures page's memcg stability:
2967 * - lock_page_memcg()
2968 * - exclusive reference
2970 page->memcg_data = (unsigned long)memcg;
2973 #ifdef CONFIG_MEMCG_KMEM
2974 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2977 unsigned int objects = objs_per_slab_page(s, page);
2980 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2985 if (!set_page_objcgs(page, vec))
2988 kmemleak_not_leak(vec);
2994 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2996 * A passed kernel object can be a slab object or a generic kernel page, so
2997 * different mechanisms for getting the memory cgroup pointer should be used.
2998 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2999 * can not know for sure how the kernel object is implemented.
3000 * mem_cgroup_from_obj() can be safely used in such cases.
3002 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
3003 * cgroup_mutex, etc.
3005 struct mem_cgroup *mem_cgroup_from_obj(void *p)
3009 if (mem_cgroup_disabled())
3012 page = virt_to_head_page(p);
3015 * Slab objects are accounted individually, not per-page.
3016 * Memcg membership data for each individual object is saved in
3017 * the page->obj_cgroups.
3019 if (page_objcgs_check(page)) {
3020 struct obj_cgroup *objcg;
3023 off = obj_to_index(page->slab_cache, page, p);
3024 objcg = page_objcgs(page)[off];
3026 return obj_cgroup_memcg(objcg);
3032 * page_memcg_check() is used here, because page_has_obj_cgroups()
3033 * check above could fail because the object cgroups vector wasn't set
3034 * at that moment, but it can be set concurrently.
3035 * page_memcg_check(page) will guarantee that a proper memory
3036 * cgroup pointer or NULL will be returned.
3038 return page_memcg_check(page);
3041 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
3043 struct obj_cgroup *objcg = NULL;
3044 struct mem_cgroup *memcg;
3046 if (memcg_kmem_bypass())
3050 if (unlikely(active_memcg()))
3051 memcg = active_memcg();
3053 memcg = mem_cgroup_from_task(current);
3055 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
3056 objcg = rcu_dereference(memcg->objcg);
3057 if (objcg && obj_cgroup_tryget(objcg))
3066 static int memcg_alloc_cache_id(void)
3071 id = ida_simple_get(&memcg_cache_ida,
3072 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3076 if (id < memcg_nr_cache_ids)
3080 * There's no space for the new id in memcg_caches arrays,
3081 * so we have to grow them.
3083 down_write(&memcg_cache_ids_sem);
3085 size = 2 * (id + 1);
3086 if (size < MEMCG_CACHES_MIN_SIZE)
3087 size = MEMCG_CACHES_MIN_SIZE;
3088 else if (size > MEMCG_CACHES_MAX_SIZE)
3089 size = MEMCG_CACHES_MAX_SIZE;
3091 err = memcg_update_all_list_lrus(size);
3093 memcg_nr_cache_ids = size;
3095 up_write(&memcg_cache_ids_sem);
3098 ida_simple_remove(&memcg_cache_ida, id);
3104 static void memcg_free_cache_id(int id)
3106 ida_simple_remove(&memcg_cache_ida, id);
3110 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
3111 * @memcg: memory cgroup to charge
3112 * @gfp: reclaim mode
3113 * @nr_pages: number of pages to charge
3115 * Returns 0 on success, an error code on failure.
3117 int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
3118 unsigned int nr_pages)
3120 struct page_counter *counter;
3123 ret = try_charge(memcg, gfp, nr_pages);
3127 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3128 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3131 * Enforce __GFP_NOFAIL allocation because callers are not
3132 * prepared to see failures and likely do not have any failure
3135 if (gfp & __GFP_NOFAIL) {
3136 page_counter_charge(&memcg->kmem, nr_pages);
3139 cancel_charge(memcg, nr_pages);
3146 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
3147 * @memcg: memcg to uncharge
3148 * @nr_pages: number of pages to uncharge
3150 void __memcg_kmem_uncharge(struct mem_cgroup *memcg, unsigned int nr_pages)
3152 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3153 page_counter_uncharge(&memcg->kmem, nr_pages);
3155 page_counter_uncharge(&memcg->memory, nr_pages);
3156 if (do_memsw_account())
3157 page_counter_uncharge(&memcg->memsw, nr_pages);
3161 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3162 * @page: page to charge
3163 * @gfp: reclaim mode
3164 * @order: allocation order
3166 * Returns 0 on success, an error code on failure.
3168 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3170 struct mem_cgroup *memcg;
3173 memcg = get_mem_cgroup_from_current();
3174 if (memcg && !mem_cgroup_is_root(memcg)) {
3175 ret = __memcg_kmem_charge(memcg, gfp, 1 << order);
3177 page->memcg_data = (unsigned long)memcg |
3181 css_put(&memcg->css);
3187 * __memcg_kmem_uncharge_page: uncharge a kmem page
3188 * @page: page to uncharge
3189 * @order: allocation order
3191 void __memcg_kmem_uncharge_page(struct page *page, int order)
3193 struct mem_cgroup *memcg = page_memcg(page);
3194 unsigned int nr_pages = 1 << order;
3199 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3200 __memcg_kmem_uncharge(memcg, nr_pages);
3201 page->memcg_data = 0;
3202 css_put(&memcg->css);
3205 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3207 struct memcg_stock_pcp *stock;
3208 unsigned long flags;
3211 local_irq_save(flags);
3213 stock = this_cpu_ptr(&memcg_stock);
3214 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3215 stock->nr_bytes -= nr_bytes;
3219 local_irq_restore(flags);
3224 static void drain_obj_stock(struct memcg_stock_pcp *stock)
3226 struct obj_cgroup *old = stock->cached_objcg;
3231 if (stock->nr_bytes) {
3232 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3233 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3237 __memcg_kmem_uncharge(obj_cgroup_memcg(old), nr_pages);
3242 * The leftover is flushed to the centralized per-memcg value.
3243 * On the next attempt to refill obj stock it will be moved
3244 * to a per-cpu stock (probably, on an other CPU), see
3245 * refill_obj_stock().
3247 * How often it's flushed is a trade-off between the memory
3248 * limit enforcement accuracy and potential CPU contention,
3249 * so it might be changed in the future.
3251 atomic_add(nr_bytes, &old->nr_charged_bytes);
3252 stock->nr_bytes = 0;
3255 obj_cgroup_put(old);
3256 stock->cached_objcg = NULL;
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)
3275 struct memcg_stock_pcp *stock;
3276 unsigned long flags;
3278 local_irq_save(flags);
3280 stock = this_cpu_ptr(&memcg_stock);
3281 if (stock->cached_objcg != objcg) { /* reset if necessary */
3282 drain_obj_stock(stock);
3283 obj_cgroup_get(objcg);
3284 stock->cached_objcg = objcg;
3285 stock->nr_bytes = atomic_xchg(&objcg->nr_charged_bytes, 0);
3287 stock->nr_bytes += nr_bytes;
3289 if (stock->nr_bytes > PAGE_SIZE)
3290 drain_obj_stock(stock);
3292 local_irq_restore(flags);
3295 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3297 struct mem_cgroup *memcg;
3298 unsigned int nr_pages, nr_bytes;
3301 if (consume_obj_stock(objcg, size))
3305 * In theory, memcg->nr_charged_bytes can have enough
3306 * pre-charged bytes to satisfy the allocation. However,
3307 * flushing memcg->nr_charged_bytes requires two atomic
3308 * operations, and memcg->nr_charged_bytes can't be big,
3309 * so it's better to ignore it and try grab some new pages.
3310 * memcg->nr_charged_bytes will be flushed in
3311 * refill_obj_stock(), called from this function or
3312 * independently later.
3316 memcg = obj_cgroup_memcg(objcg);
3317 if (unlikely(!css_tryget(&memcg->css)))
3321 nr_pages = size >> PAGE_SHIFT;
3322 nr_bytes = size & (PAGE_SIZE - 1);
3327 ret = __memcg_kmem_charge(memcg, gfp, nr_pages);
3328 if (!ret && nr_bytes)
3329 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes);
3331 css_put(&memcg->css);
3335 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3337 refill_obj_stock(objcg, size);
3340 #endif /* CONFIG_MEMCG_KMEM */
3342 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3344 * Because page_memcg(head) is not set on compound tails, set it now.
3346 void mem_cgroup_split_huge_fixup(struct page *head)
3348 struct mem_cgroup *memcg = page_memcg(head);
3351 if (mem_cgroup_disabled())
3354 for (i = 1; i < HPAGE_PMD_NR; i++) {
3355 css_get(&memcg->css);
3356 head[i].memcg_data = (unsigned long)memcg;
3359 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3361 #ifdef CONFIG_MEMCG_SWAP
3363 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3364 * @entry: swap entry to be moved
3365 * @from: mem_cgroup which the entry is moved from
3366 * @to: mem_cgroup which the entry is moved to
3368 * It succeeds only when the swap_cgroup's record for this entry is the same
3369 * as the mem_cgroup's id of @from.
3371 * Returns 0 on success, -EINVAL on failure.
3373 * The caller must have charged to @to, IOW, called page_counter_charge() about
3374 * both res and memsw, and called css_get().
3376 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3377 struct mem_cgroup *from, struct mem_cgroup *to)
3379 unsigned short old_id, new_id;
3381 old_id = mem_cgroup_id(from);
3382 new_id = mem_cgroup_id(to);
3384 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3385 mod_memcg_state(from, MEMCG_SWAP, -1);
3386 mod_memcg_state(to, MEMCG_SWAP, 1);
3392 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3393 struct mem_cgroup *from, struct mem_cgroup *to)
3399 static DEFINE_MUTEX(memcg_max_mutex);
3401 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3402 unsigned long max, bool memsw)
3404 bool enlarge = false;
3405 bool drained = false;
3407 bool limits_invariant;
3408 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3411 if (signal_pending(current)) {
3416 mutex_lock(&memcg_max_mutex);
3418 * Make sure that the new limit (memsw or memory limit) doesn't
3419 * break our basic invariant rule memory.max <= memsw.max.
3421 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3422 max <= memcg->memsw.max;
3423 if (!limits_invariant) {
3424 mutex_unlock(&memcg_max_mutex);
3428 if (max > counter->max)
3430 ret = page_counter_set_max(counter, max);
3431 mutex_unlock(&memcg_max_mutex);
3437 drain_all_stock(memcg);
3442 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3443 GFP_KERNEL, !memsw)) {
3449 if (!ret && enlarge)
3450 memcg_oom_recover(memcg);
3455 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3457 unsigned long *total_scanned)
3459 unsigned long nr_reclaimed = 0;
3460 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3461 unsigned long reclaimed;
3463 struct mem_cgroup_tree_per_node *mctz;
3464 unsigned long excess;
3465 unsigned long nr_scanned;
3470 mctz = soft_limit_tree_node(pgdat->node_id);
3473 * Do not even bother to check the largest node if the root
3474 * is empty. Do it lockless to prevent lock bouncing. Races
3475 * are acceptable as soft limit is best effort anyway.
3477 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3481 * This loop can run a while, specially if mem_cgroup's continuously
3482 * keep exceeding their soft limit and putting the system under
3489 mz = mem_cgroup_largest_soft_limit_node(mctz);
3494 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3495 gfp_mask, &nr_scanned);
3496 nr_reclaimed += reclaimed;
3497 *total_scanned += nr_scanned;
3498 spin_lock_irq(&mctz->lock);
3499 __mem_cgroup_remove_exceeded(mz, mctz);
3502 * If we failed to reclaim anything from this memory cgroup
3503 * it is time to move on to the next cgroup
3507 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3509 excess = soft_limit_excess(mz->memcg);
3511 * One school of thought says that we should not add
3512 * back the node to the tree if reclaim returns 0.
3513 * But our reclaim could return 0, simply because due
3514 * to priority we are exposing a smaller subset of
3515 * memory to reclaim from. Consider this as a longer
3518 /* If excess == 0, no tree ops */
3519 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3520 spin_unlock_irq(&mctz->lock);
3521 css_put(&mz->memcg->css);
3524 * Could not reclaim anything and there are no more
3525 * mem cgroups to try or we seem to be looping without
3526 * reclaiming anything.
3528 if (!nr_reclaimed &&
3530 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3532 } while (!nr_reclaimed);
3534 css_put(&next_mz->memcg->css);
3535 return nr_reclaimed;
3539 * Reclaims as many pages from the given memcg as possible.
3541 * Caller is responsible for holding css reference for memcg.
3543 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3545 int nr_retries = MAX_RECLAIM_RETRIES;
3547 /* we call try-to-free pages for make this cgroup empty */
3548 lru_add_drain_all();
3550 drain_all_stock(memcg);
3552 /* try to free all pages in this cgroup */
3553 while (nr_retries && page_counter_read(&memcg->memory)) {
3556 if (signal_pending(current))
3559 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3563 /* maybe some writeback is necessary */
3564 congestion_wait(BLK_RW_ASYNC, HZ/10);
3572 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3573 char *buf, size_t nbytes,
3576 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3578 if (mem_cgroup_is_root(memcg))
3580 return mem_cgroup_force_empty(memcg) ?: nbytes;
3583 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3589 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3590 struct cftype *cft, u64 val)
3595 pr_warn_once("Non-hierarchical mode is deprecated. "
3596 "Please report your usecase to linux-mm@kvack.org if you "
3597 "depend on this functionality.\n");
3602 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3606 if (mem_cgroup_is_root(memcg)) {
3607 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3608 memcg_page_state(memcg, NR_ANON_MAPPED);
3610 val += memcg_page_state(memcg, MEMCG_SWAP);
3613 val = page_counter_read(&memcg->memory);
3615 val = page_counter_read(&memcg->memsw);
3628 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3631 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3632 struct page_counter *counter;
3634 switch (MEMFILE_TYPE(cft->private)) {
3636 counter = &memcg->memory;
3639 counter = &memcg->memsw;
3642 counter = &memcg->kmem;
3645 counter = &memcg->tcpmem;
3651 switch (MEMFILE_ATTR(cft->private)) {
3653 if (counter == &memcg->memory)
3654 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3655 if (counter == &memcg->memsw)
3656 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3657 return (u64)page_counter_read(counter) * PAGE_SIZE;
3659 return (u64)counter->max * PAGE_SIZE;
3661 return (u64)counter->watermark * PAGE_SIZE;
3663 return counter->failcnt;
3664 case RES_SOFT_LIMIT:
3665 return (u64)memcg->soft_limit * PAGE_SIZE;
3671 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3673 unsigned long stat[MEMCG_NR_STAT] = {0};
3674 struct mem_cgroup *mi;
3677 for_each_online_cpu(cpu)
3678 for (i = 0; i < MEMCG_NR_STAT; i++)
3679 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3681 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3682 for (i = 0; i < MEMCG_NR_STAT; i++)
3683 atomic_long_add(stat[i], &mi->vmstats[i]);
3685 for_each_node(node) {
3686 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3687 struct mem_cgroup_per_node *pi;
3689 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3692 for_each_online_cpu(cpu)
3693 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3695 pn->lruvec_stat_cpu->count[i], cpu);
3697 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3698 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3699 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3703 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3705 unsigned long events[NR_VM_EVENT_ITEMS];
3706 struct mem_cgroup *mi;
3709 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3712 for_each_online_cpu(cpu)
3713 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3714 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3717 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3718 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3719 atomic_long_add(events[i], &mi->vmevents[i]);
3722 #ifdef CONFIG_MEMCG_KMEM
3723 static int memcg_online_kmem(struct mem_cgroup *memcg)
3725 struct obj_cgroup *objcg;
3728 if (cgroup_memory_nokmem)
3731 BUG_ON(memcg->kmemcg_id >= 0);
3732 BUG_ON(memcg->kmem_state);
3734 memcg_id = memcg_alloc_cache_id();
3738 objcg = obj_cgroup_alloc();
3740 memcg_free_cache_id(memcg_id);
3743 objcg->memcg = memcg;
3744 rcu_assign_pointer(memcg->objcg, objcg);
3746 static_branch_enable(&memcg_kmem_enabled_key);
3748 memcg->kmemcg_id = memcg_id;
3749 memcg->kmem_state = KMEM_ONLINE;
3754 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3756 struct cgroup_subsys_state *css;
3757 struct mem_cgroup *parent, *child;
3760 if (memcg->kmem_state != KMEM_ONLINE)
3763 memcg->kmem_state = KMEM_ALLOCATED;
3765 parent = parent_mem_cgroup(memcg);
3767 parent = root_mem_cgroup;
3769 memcg_reparent_objcgs(memcg, parent);
3771 kmemcg_id = memcg->kmemcg_id;
3772 BUG_ON(kmemcg_id < 0);
3775 * Change kmemcg_id of this cgroup and all its descendants to the
3776 * parent's id, and then move all entries from this cgroup's list_lrus
3777 * to ones of the parent. After we have finished, all list_lrus
3778 * corresponding to this cgroup are guaranteed to remain empty. The
3779 * ordering is imposed by list_lru_node->lock taken by
3780 * memcg_drain_all_list_lrus().
3782 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3783 css_for_each_descendant_pre(css, &memcg->css) {
3784 child = mem_cgroup_from_css(css);
3785 BUG_ON(child->kmemcg_id != kmemcg_id);
3786 child->kmemcg_id = parent->kmemcg_id;
3790 memcg_drain_all_list_lrus(kmemcg_id, parent);
3792 memcg_free_cache_id(kmemcg_id);
3795 static void memcg_free_kmem(struct mem_cgroup *memcg)
3797 /* css_alloc() failed, offlining didn't happen */
3798 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3799 memcg_offline_kmem(memcg);
3802 static int memcg_online_kmem(struct mem_cgroup *memcg)
3806 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3809 static void memcg_free_kmem(struct mem_cgroup *memcg)
3812 #endif /* CONFIG_MEMCG_KMEM */
3814 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3819 mutex_lock(&memcg_max_mutex);
3820 ret = page_counter_set_max(&memcg->kmem, max);
3821 mutex_unlock(&memcg_max_mutex);
3825 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3829 mutex_lock(&memcg_max_mutex);
3831 ret = page_counter_set_max(&memcg->tcpmem, max);
3835 if (!memcg->tcpmem_active) {
3837 * The active flag needs to be written after the static_key
3838 * update. This is what guarantees that the socket activation
3839 * function is the last one to run. See mem_cgroup_sk_alloc()
3840 * for details, and note that we don't mark any socket as
3841 * belonging to this memcg until that flag is up.
3843 * We need to do this, because static_keys will span multiple
3844 * sites, but we can't control their order. If we mark a socket
3845 * as accounted, but the accounting functions are not patched in
3846 * yet, we'll lose accounting.
3848 * We never race with the readers in mem_cgroup_sk_alloc(),
3849 * because when this value change, the code to process it is not
3852 static_branch_inc(&memcg_sockets_enabled_key);
3853 memcg->tcpmem_active = true;
3856 mutex_unlock(&memcg_max_mutex);
3861 * The user of this function is...
3864 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3865 char *buf, size_t nbytes, loff_t off)
3867 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3868 unsigned long nr_pages;
3871 buf = strstrip(buf);
3872 ret = page_counter_memparse(buf, "-1", &nr_pages);
3876 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3878 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3882 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3884 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3887 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3890 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3891 "Please report your usecase to linux-mm@kvack.org if you "
3892 "depend on this functionality.\n");
3893 ret = memcg_update_kmem_max(memcg, nr_pages);
3896 ret = memcg_update_tcp_max(memcg, nr_pages);
3900 case RES_SOFT_LIMIT:
3901 memcg->soft_limit = nr_pages;
3905 return ret ?: nbytes;
3908 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3909 size_t nbytes, loff_t off)
3911 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3912 struct page_counter *counter;
3914 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3916 counter = &memcg->memory;
3919 counter = &memcg->memsw;
3922 counter = &memcg->kmem;
3925 counter = &memcg->tcpmem;
3931 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3933 page_counter_reset_watermark(counter);
3936 counter->failcnt = 0;
3945 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3948 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3952 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3953 struct cftype *cft, u64 val)
3955 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3957 if (val & ~MOVE_MASK)
3961 * No kind of locking is needed in here, because ->can_attach() will
3962 * check this value once in the beginning of the process, and then carry
3963 * on with stale data. This means that changes to this value will only
3964 * affect task migrations starting after the change.
3966 memcg->move_charge_at_immigrate = val;
3970 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3971 struct cftype *cft, u64 val)
3979 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3980 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3981 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3983 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3984 int nid, unsigned int lru_mask, bool tree)
3986 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3987 unsigned long nr = 0;
3990 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3993 if (!(BIT(lru) & lru_mask))
3996 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3998 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
4003 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
4004 unsigned int lru_mask,
4007 unsigned long nr = 0;
4011 if (!(BIT(lru) & lru_mask))
4014 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
4016 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
4021 static int memcg_numa_stat_show(struct seq_file *m, void *v)
4025 unsigned int lru_mask;
4028 static const struct numa_stat stats[] = {
4029 { "total", LRU_ALL },
4030 { "file", LRU_ALL_FILE },
4031 { "anon", LRU_ALL_ANON },
4032 { "unevictable", BIT(LRU_UNEVICTABLE) },
4034 const struct numa_stat *stat;
4036 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4038 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4039 seq_printf(m, "%s=%lu", stat->name,
4040 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4042 for_each_node_state(nid, N_MEMORY)
4043 seq_printf(m, " N%d=%lu", nid,
4044 mem_cgroup_node_nr_lru_pages(memcg, nid,
4045 stat->lru_mask, false));
4049 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4051 seq_printf(m, "hierarchical_%s=%lu", stat->name,
4052 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4054 for_each_node_state(nid, N_MEMORY)
4055 seq_printf(m, " N%d=%lu", nid,
4056 mem_cgroup_node_nr_lru_pages(memcg, nid,
4057 stat->lru_mask, true));
4063 #endif /* CONFIG_NUMA */
4065 static const unsigned int memcg1_stats[] = {
4068 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4078 static const char *const memcg1_stat_names[] = {
4081 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4091 /* Universal VM events cgroup1 shows, original sort order */
4092 static const unsigned int memcg1_events[] = {
4099 static int memcg_stat_show(struct seq_file *m, void *v)
4101 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4102 unsigned long memory, memsw;
4103 struct mem_cgroup *mi;
4106 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4108 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4111 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4113 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4114 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4115 if (memcg1_stats[i] == NR_ANON_THPS)
4118 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4121 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4122 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4123 memcg_events_local(memcg, memcg1_events[i]));
4125 for (i = 0; i < NR_LRU_LISTS; i++)
4126 seq_printf(m, "%s %lu\n", lru_list_name(i),
4127 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4130 /* Hierarchical information */
4131 memory = memsw = PAGE_COUNTER_MAX;
4132 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4133 memory = min(memory, READ_ONCE(mi->memory.max));
4134 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4136 seq_printf(m, "hierarchical_memory_limit %llu\n",
4137 (u64)memory * PAGE_SIZE);
4138 if (do_memsw_account())
4139 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4140 (u64)memsw * PAGE_SIZE);
4142 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4145 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4147 nr = memcg_page_state(memcg, memcg1_stats[i]);
4148 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4149 if (memcg1_stats[i] == NR_ANON_THPS)
4152 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4153 (u64)nr * PAGE_SIZE);
4156 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4157 seq_printf(m, "total_%s %llu\n",
4158 vm_event_name(memcg1_events[i]),
4159 (u64)memcg_events(memcg, memcg1_events[i]));
4161 for (i = 0; i < NR_LRU_LISTS; i++)
4162 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4163 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4166 #ifdef CONFIG_DEBUG_VM
4169 struct mem_cgroup_per_node *mz;
4170 unsigned long anon_cost = 0;
4171 unsigned long file_cost = 0;
4173 for_each_online_pgdat(pgdat) {
4174 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
4176 anon_cost += mz->lruvec.anon_cost;
4177 file_cost += mz->lruvec.file_cost;
4179 seq_printf(m, "anon_cost %lu\n", anon_cost);
4180 seq_printf(m, "file_cost %lu\n", file_cost);
4187 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4190 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4192 return mem_cgroup_swappiness(memcg);
4195 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4196 struct cftype *cft, u64 val)
4198 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4204 memcg->swappiness = val;
4206 vm_swappiness = val;
4211 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4213 struct mem_cgroup_threshold_ary *t;
4214 unsigned long usage;
4219 t = rcu_dereference(memcg->thresholds.primary);
4221 t = rcu_dereference(memcg->memsw_thresholds.primary);
4226 usage = mem_cgroup_usage(memcg, swap);
4229 * current_threshold points to threshold just below or equal to usage.
4230 * If it's not true, a threshold was crossed after last
4231 * call of __mem_cgroup_threshold().
4233 i = t->current_threshold;
4236 * Iterate backward over array of thresholds starting from
4237 * current_threshold and check if a threshold is crossed.
4238 * If none of thresholds below usage is crossed, we read
4239 * only one element of the array here.
4241 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4242 eventfd_signal(t->entries[i].eventfd, 1);
4244 /* i = current_threshold + 1 */
4248 * Iterate forward over array of thresholds starting from
4249 * current_threshold+1 and check if a threshold is crossed.
4250 * If none of thresholds above usage is crossed, we read
4251 * only one element of the array here.
4253 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4254 eventfd_signal(t->entries[i].eventfd, 1);
4256 /* Update current_threshold */
4257 t->current_threshold = i - 1;
4262 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4265 __mem_cgroup_threshold(memcg, false);
4266 if (do_memsw_account())
4267 __mem_cgroup_threshold(memcg, true);
4269 memcg = parent_mem_cgroup(memcg);
4273 static int compare_thresholds(const void *a, const void *b)
4275 const struct mem_cgroup_threshold *_a = a;
4276 const struct mem_cgroup_threshold *_b = b;
4278 if (_a->threshold > _b->threshold)
4281 if (_a->threshold < _b->threshold)
4287 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4289 struct mem_cgroup_eventfd_list *ev;
4291 spin_lock(&memcg_oom_lock);
4293 list_for_each_entry(ev, &memcg->oom_notify, list)
4294 eventfd_signal(ev->eventfd, 1);
4296 spin_unlock(&memcg_oom_lock);
4300 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4302 struct mem_cgroup *iter;
4304 for_each_mem_cgroup_tree(iter, memcg)
4305 mem_cgroup_oom_notify_cb(iter);
4308 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4309 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4311 struct mem_cgroup_thresholds *thresholds;
4312 struct mem_cgroup_threshold_ary *new;
4313 unsigned long threshold;
4314 unsigned long usage;
4317 ret = page_counter_memparse(args, "-1", &threshold);
4321 mutex_lock(&memcg->thresholds_lock);
4324 thresholds = &memcg->thresholds;
4325 usage = mem_cgroup_usage(memcg, false);
4326 } else if (type == _MEMSWAP) {
4327 thresholds = &memcg->memsw_thresholds;
4328 usage = mem_cgroup_usage(memcg, true);
4332 /* Check if a threshold crossed before adding a new one */
4333 if (thresholds->primary)
4334 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4336 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4338 /* Allocate memory for new array of thresholds */
4339 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4346 /* Copy thresholds (if any) to new array */
4347 if (thresholds->primary)
4348 memcpy(new->entries, thresholds->primary->entries,
4349 flex_array_size(new, entries, size - 1));
4351 /* Add new threshold */
4352 new->entries[size - 1].eventfd = eventfd;
4353 new->entries[size - 1].threshold = threshold;
4355 /* Sort thresholds. Registering of new threshold isn't time-critical */
4356 sort(new->entries, size, sizeof(*new->entries),
4357 compare_thresholds, NULL);
4359 /* Find current threshold */
4360 new->current_threshold = -1;
4361 for (i = 0; i < size; i++) {
4362 if (new->entries[i].threshold <= usage) {
4364 * new->current_threshold will not be used until
4365 * rcu_assign_pointer(), so it's safe to increment
4368 ++new->current_threshold;
4373 /* Free old spare buffer and save old primary buffer as spare */
4374 kfree(thresholds->spare);
4375 thresholds->spare = thresholds->primary;
4377 rcu_assign_pointer(thresholds->primary, new);
4379 /* To be sure that nobody uses thresholds */
4383 mutex_unlock(&memcg->thresholds_lock);
4388 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4389 struct eventfd_ctx *eventfd, const char *args)
4391 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4394 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4395 struct eventfd_ctx *eventfd, const char *args)
4397 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4400 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4401 struct eventfd_ctx *eventfd, enum res_type type)
4403 struct mem_cgroup_thresholds *thresholds;
4404 struct mem_cgroup_threshold_ary *new;
4405 unsigned long usage;
4406 int i, j, size, entries;
4408 mutex_lock(&memcg->thresholds_lock);
4411 thresholds = &memcg->thresholds;
4412 usage = mem_cgroup_usage(memcg, false);
4413 } else if (type == _MEMSWAP) {
4414 thresholds = &memcg->memsw_thresholds;
4415 usage = mem_cgroup_usage(memcg, true);
4419 if (!thresholds->primary)
4422 /* Check if a threshold crossed before removing */
4423 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4425 /* Calculate new number of threshold */
4427 for (i = 0; i < thresholds->primary->size; i++) {
4428 if (thresholds->primary->entries[i].eventfd != eventfd)
4434 new = thresholds->spare;
4436 /* If no items related to eventfd have been cleared, nothing to do */
4440 /* Set thresholds array to NULL if we don't have thresholds */
4449 /* Copy thresholds and find current threshold */
4450 new->current_threshold = -1;
4451 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4452 if (thresholds->primary->entries[i].eventfd == eventfd)
4455 new->entries[j] = thresholds->primary->entries[i];
4456 if (new->entries[j].threshold <= usage) {
4458 * new->current_threshold will not be used
4459 * until rcu_assign_pointer(), so it's safe to increment
4462 ++new->current_threshold;
4468 /* Swap primary and spare array */
4469 thresholds->spare = thresholds->primary;
4471 rcu_assign_pointer(thresholds->primary, new);
4473 /* To be sure that nobody uses thresholds */
4476 /* If all events are unregistered, free the spare array */
4478 kfree(thresholds->spare);
4479 thresholds->spare = NULL;
4482 mutex_unlock(&memcg->thresholds_lock);
4485 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4486 struct eventfd_ctx *eventfd)
4488 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4491 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4492 struct eventfd_ctx *eventfd)
4494 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4497 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4498 struct eventfd_ctx *eventfd, const char *args)
4500 struct mem_cgroup_eventfd_list *event;
4502 event = kmalloc(sizeof(*event), GFP_KERNEL);
4506 spin_lock(&memcg_oom_lock);
4508 event->eventfd = eventfd;
4509 list_add(&event->list, &memcg->oom_notify);
4511 /* already in OOM ? */
4512 if (memcg->under_oom)
4513 eventfd_signal(eventfd, 1);
4514 spin_unlock(&memcg_oom_lock);
4519 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4520 struct eventfd_ctx *eventfd)
4522 struct mem_cgroup_eventfd_list *ev, *tmp;
4524 spin_lock(&memcg_oom_lock);
4526 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4527 if (ev->eventfd == eventfd) {
4528 list_del(&ev->list);
4533 spin_unlock(&memcg_oom_lock);
4536 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4538 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4540 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4541 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4542 seq_printf(sf, "oom_kill %lu\n",
4543 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4547 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4548 struct cftype *cft, u64 val)
4550 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4552 /* cannot set to root cgroup and only 0 and 1 are allowed */
4553 if (!css->parent || !((val == 0) || (val == 1)))
4556 memcg->oom_kill_disable = val;
4558 memcg_oom_recover(memcg);
4563 #ifdef CONFIG_CGROUP_WRITEBACK
4565 #include <trace/events/writeback.h>
4567 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4569 return wb_domain_init(&memcg->cgwb_domain, gfp);
4572 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4574 wb_domain_exit(&memcg->cgwb_domain);
4577 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4579 wb_domain_size_changed(&memcg->cgwb_domain);
4582 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4584 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4586 if (!memcg->css.parent)
4589 return &memcg->cgwb_domain;
4593 * idx can be of type enum memcg_stat_item or node_stat_item.
4594 * Keep in sync with memcg_exact_page().
4596 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4598 long x = atomic_long_read(&memcg->vmstats[idx]);
4601 for_each_online_cpu(cpu)
4602 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4609 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4610 * @wb: bdi_writeback in question
4611 * @pfilepages: out parameter for number of file pages
4612 * @pheadroom: out parameter for number of allocatable pages according to memcg
4613 * @pdirty: out parameter for number of dirty pages
4614 * @pwriteback: out parameter for number of pages under writeback
4616 * Determine the numbers of file, headroom, dirty, and writeback pages in
4617 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4618 * is a bit more involved.
4620 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4621 * headroom is calculated as the lowest headroom of itself and the
4622 * ancestors. Note that this doesn't consider the actual amount of
4623 * available memory in the system. The caller should further cap
4624 * *@pheadroom accordingly.
4626 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4627 unsigned long *pheadroom, unsigned long *pdirty,
4628 unsigned long *pwriteback)
4630 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4631 struct mem_cgroup *parent;
4633 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4635 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4636 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4637 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4638 *pheadroom = PAGE_COUNTER_MAX;
4640 while ((parent = parent_mem_cgroup(memcg))) {
4641 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4642 READ_ONCE(memcg->memory.high));
4643 unsigned long used = page_counter_read(&memcg->memory);
4645 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4651 * Foreign dirty flushing
4653 * There's an inherent mismatch between memcg and writeback. The former
4654 * trackes ownership per-page while the latter per-inode. This was a
4655 * deliberate design decision because honoring per-page ownership in the
4656 * writeback path is complicated, may lead to higher CPU and IO overheads
4657 * and deemed unnecessary given that write-sharing an inode across
4658 * different cgroups isn't a common use-case.
4660 * Combined with inode majority-writer ownership switching, this works well
4661 * enough in most cases but there are some pathological cases. For
4662 * example, let's say there are two cgroups A and B which keep writing to
4663 * different but confined parts of the same inode. B owns the inode and
4664 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4665 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4666 * triggering background writeback. A will be slowed down without a way to
4667 * make writeback of the dirty pages happen.
4669 * Conditions like the above can lead to a cgroup getting repatedly and
4670 * severely throttled after making some progress after each
4671 * dirty_expire_interval while the underyling IO device is almost
4674 * Solving this problem completely requires matching the ownership tracking
4675 * granularities between memcg and writeback in either direction. However,
4676 * the more egregious behaviors can be avoided by simply remembering the
4677 * most recent foreign dirtying events and initiating remote flushes on
4678 * them when local writeback isn't enough to keep the memory clean enough.
4680 * The following two functions implement such mechanism. When a foreign
4681 * page - a page whose memcg and writeback ownerships don't match - is
4682 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4683 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4684 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4685 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4686 * foreign bdi_writebacks which haven't expired. Both the numbers of
4687 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4688 * limited to MEMCG_CGWB_FRN_CNT.
4690 * The mechanism only remembers IDs and doesn't hold any object references.
4691 * As being wrong occasionally doesn't matter, updates and accesses to the
4692 * records are lockless and racy.
4694 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4695 struct bdi_writeback *wb)
4697 struct mem_cgroup *memcg = page_memcg(page);
4698 struct memcg_cgwb_frn *frn;
4699 u64 now = get_jiffies_64();
4700 u64 oldest_at = now;
4704 trace_track_foreign_dirty(page, wb);
4707 * Pick the slot to use. If there is already a slot for @wb, keep
4708 * using it. If not replace the oldest one which isn't being
4711 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4712 frn = &memcg->cgwb_frn[i];
4713 if (frn->bdi_id == wb->bdi->id &&
4714 frn->memcg_id == wb->memcg_css->id)
4716 if (time_before64(frn->at, oldest_at) &&
4717 atomic_read(&frn->done.cnt) == 1) {
4719 oldest_at = frn->at;
4723 if (i < MEMCG_CGWB_FRN_CNT) {
4725 * Re-using an existing one. Update timestamp lazily to
4726 * avoid making the cacheline hot. We want them to be
4727 * reasonably up-to-date and significantly shorter than
4728 * dirty_expire_interval as that's what expires the record.
4729 * Use the shorter of 1s and dirty_expire_interval / 8.
4731 unsigned long update_intv =
4732 min_t(unsigned long, HZ,
4733 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4735 if (time_before64(frn->at, now - update_intv))
4737 } else if (oldest >= 0) {
4738 /* replace the oldest free one */
4739 frn = &memcg->cgwb_frn[oldest];
4740 frn->bdi_id = wb->bdi->id;
4741 frn->memcg_id = wb->memcg_css->id;
4746 /* issue foreign writeback flushes for recorded foreign dirtying events */
4747 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4749 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4750 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4751 u64 now = jiffies_64;
4754 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4755 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4758 * If the record is older than dirty_expire_interval,
4759 * writeback on it has already started. No need to kick it
4760 * off again. Also, don't start a new one if there's
4761 * already one in flight.
4763 if (time_after64(frn->at, now - intv) &&
4764 atomic_read(&frn->done.cnt) == 1) {
4766 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4767 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4768 WB_REASON_FOREIGN_FLUSH,
4774 #else /* CONFIG_CGROUP_WRITEBACK */
4776 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4781 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4785 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4789 #endif /* CONFIG_CGROUP_WRITEBACK */
4792 * DO NOT USE IN NEW FILES.
4794 * "cgroup.event_control" implementation.
4796 * This is way over-engineered. It tries to support fully configurable
4797 * events for each user. Such level of flexibility is completely
4798 * unnecessary especially in the light of the planned unified hierarchy.
4800 * Please deprecate this and replace with something simpler if at all
4805 * Unregister event and free resources.
4807 * Gets called from workqueue.
4809 static void memcg_event_remove(struct work_struct *work)
4811 struct mem_cgroup_event *event =
4812 container_of(work, struct mem_cgroup_event, remove);
4813 struct mem_cgroup *memcg = event->memcg;
4815 remove_wait_queue(event->wqh, &event->wait);
4817 event->unregister_event(memcg, event->eventfd);
4819 /* Notify userspace the event is going away. */
4820 eventfd_signal(event->eventfd, 1);
4822 eventfd_ctx_put(event->eventfd);
4824 css_put(&memcg->css);
4828 * Gets called on EPOLLHUP on eventfd when user closes it.
4830 * Called with wqh->lock held and interrupts disabled.
4832 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4833 int sync, void *key)
4835 struct mem_cgroup_event *event =
4836 container_of(wait, struct mem_cgroup_event, wait);
4837 struct mem_cgroup *memcg = event->memcg;
4838 __poll_t flags = key_to_poll(key);
4840 if (flags & EPOLLHUP) {
4842 * If the event has been detached at cgroup removal, we
4843 * can simply return knowing the other side will cleanup
4846 * We can't race against event freeing since the other
4847 * side will require wqh->lock via remove_wait_queue(),
4850 spin_lock(&memcg->event_list_lock);
4851 if (!list_empty(&event->list)) {
4852 list_del_init(&event->list);
4854 * We are in atomic context, but cgroup_event_remove()
4855 * may sleep, so we have to call it in workqueue.
4857 schedule_work(&event->remove);
4859 spin_unlock(&memcg->event_list_lock);
4865 static void memcg_event_ptable_queue_proc(struct file *file,
4866 wait_queue_head_t *wqh, poll_table *pt)
4868 struct mem_cgroup_event *event =
4869 container_of(pt, struct mem_cgroup_event, pt);
4872 add_wait_queue(wqh, &event->wait);
4876 * DO NOT USE IN NEW FILES.
4878 * Parse input and register new cgroup event handler.
4880 * Input must be in format '<event_fd> <control_fd> <args>'.
4881 * Interpretation of args is defined by control file implementation.
4883 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4884 char *buf, size_t nbytes, loff_t off)
4886 struct cgroup_subsys_state *css = of_css(of);
4887 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4888 struct mem_cgroup_event *event;
4889 struct cgroup_subsys_state *cfile_css;
4890 unsigned int efd, cfd;
4897 buf = strstrip(buf);
4899 efd = simple_strtoul(buf, &endp, 10);
4904 cfd = simple_strtoul(buf, &endp, 10);
4905 if ((*endp != ' ') && (*endp != '\0'))
4909 event = kzalloc(sizeof(*event), GFP_KERNEL);
4913 event->memcg = memcg;
4914 INIT_LIST_HEAD(&event->list);
4915 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4916 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4917 INIT_WORK(&event->remove, memcg_event_remove);
4925 event->eventfd = eventfd_ctx_fileget(efile.file);
4926 if (IS_ERR(event->eventfd)) {
4927 ret = PTR_ERR(event->eventfd);
4934 goto out_put_eventfd;
4937 /* the process need read permission on control file */
4938 /* AV: shouldn't we check that it's been opened for read instead? */
4939 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4944 * Determine the event callbacks and set them in @event. This used
4945 * to be done via struct cftype but cgroup core no longer knows
4946 * about these events. The following is crude but the whole thing
4947 * is for compatibility anyway.
4949 * DO NOT ADD NEW FILES.
4951 name = cfile.file->f_path.dentry->d_name.name;
4953 if (!strcmp(name, "memory.usage_in_bytes")) {
4954 event->register_event = mem_cgroup_usage_register_event;
4955 event->unregister_event = mem_cgroup_usage_unregister_event;
4956 } else if (!strcmp(name, "memory.oom_control")) {
4957 event->register_event = mem_cgroup_oom_register_event;
4958 event->unregister_event = mem_cgroup_oom_unregister_event;
4959 } else if (!strcmp(name, "memory.pressure_level")) {
4960 event->register_event = vmpressure_register_event;
4961 event->unregister_event = vmpressure_unregister_event;
4962 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4963 event->register_event = memsw_cgroup_usage_register_event;
4964 event->unregister_event = memsw_cgroup_usage_unregister_event;
4971 * Verify @cfile should belong to @css. Also, remaining events are
4972 * automatically removed on cgroup destruction but the removal is
4973 * asynchronous, so take an extra ref on @css.
4975 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4976 &memory_cgrp_subsys);
4978 if (IS_ERR(cfile_css))
4980 if (cfile_css != css) {
4985 ret = event->register_event(memcg, event->eventfd, buf);
4989 vfs_poll(efile.file, &event->pt);
4991 spin_lock(&memcg->event_list_lock);
4992 list_add(&event->list, &memcg->event_list);
4993 spin_unlock(&memcg->event_list_lock);
5005 eventfd_ctx_put(event->eventfd);
5014 static struct cftype mem_cgroup_legacy_files[] = {
5016 .name = "usage_in_bytes",
5017 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5018 .read_u64 = mem_cgroup_read_u64,
5021 .name = "max_usage_in_bytes",
5022 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5023 .write = mem_cgroup_reset,
5024 .read_u64 = mem_cgroup_read_u64,
5027 .name = "limit_in_bytes",
5028 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5029 .write = mem_cgroup_write,
5030 .read_u64 = mem_cgroup_read_u64,
5033 .name = "soft_limit_in_bytes",
5034 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5035 .write = mem_cgroup_write,
5036 .read_u64 = mem_cgroup_read_u64,
5040 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5041 .write = mem_cgroup_reset,
5042 .read_u64 = mem_cgroup_read_u64,
5046 .seq_show = memcg_stat_show,
5049 .name = "force_empty",
5050 .write = mem_cgroup_force_empty_write,
5053 .name = "use_hierarchy",
5054 .write_u64 = mem_cgroup_hierarchy_write,
5055 .read_u64 = mem_cgroup_hierarchy_read,
5058 .name = "cgroup.event_control", /* XXX: for compat */
5059 .write = memcg_write_event_control,
5060 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
5063 .name = "swappiness",
5064 .read_u64 = mem_cgroup_swappiness_read,
5065 .write_u64 = mem_cgroup_swappiness_write,
5068 .name = "move_charge_at_immigrate",
5069 .read_u64 = mem_cgroup_move_charge_read,
5070 .write_u64 = mem_cgroup_move_charge_write,
5073 .name = "oom_control",
5074 .seq_show = mem_cgroup_oom_control_read,
5075 .write_u64 = mem_cgroup_oom_control_write,
5076 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5079 .name = "pressure_level",
5083 .name = "numa_stat",
5084 .seq_show = memcg_numa_stat_show,
5088 .name = "kmem.limit_in_bytes",
5089 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5090 .write = mem_cgroup_write,
5091 .read_u64 = mem_cgroup_read_u64,
5094 .name = "kmem.usage_in_bytes",
5095 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5096 .read_u64 = mem_cgroup_read_u64,
5099 .name = "kmem.failcnt",
5100 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5101 .write = mem_cgroup_reset,
5102 .read_u64 = mem_cgroup_read_u64,
5105 .name = "kmem.max_usage_in_bytes",
5106 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5107 .write = mem_cgroup_reset,
5108 .read_u64 = mem_cgroup_read_u64,
5110 #if defined(CONFIG_MEMCG_KMEM) && \
5111 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5113 .name = "kmem.slabinfo",
5114 .seq_show = memcg_slab_show,
5118 .name = "kmem.tcp.limit_in_bytes",
5119 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5120 .write = mem_cgroup_write,
5121 .read_u64 = mem_cgroup_read_u64,
5124 .name = "kmem.tcp.usage_in_bytes",
5125 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5126 .read_u64 = mem_cgroup_read_u64,
5129 .name = "kmem.tcp.failcnt",
5130 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5131 .write = mem_cgroup_reset,
5132 .read_u64 = mem_cgroup_read_u64,
5135 .name = "kmem.tcp.max_usage_in_bytes",
5136 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5137 .write = mem_cgroup_reset,
5138 .read_u64 = mem_cgroup_read_u64,
5140 { }, /* terminate */
5144 * Private memory cgroup IDR
5146 * Swap-out records and page cache shadow entries need to store memcg
5147 * references in constrained space, so we maintain an ID space that is
5148 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5149 * memory-controlled cgroups to 64k.
5151 * However, there usually are many references to the offline CSS after
5152 * the cgroup has been destroyed, such as page cache or reclaimable
5153 * slab objects, that don't need to hang on to the ID. We want to keep
5154 * those dead CSS from occupying IDs, or we might quickly exhaust the
5155 * relatively small ID space and prevent the creation of new cgroups
5156 * even when there are much fewer than 64k cgroups - possibly none.
5158 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5159 * be freed and recycled when it's no longer needed, which is usually
5160 * when the CSS is offlined.
5162 * The only exception to that are records of swapped out tmpfs/shmem
5163 * pages that need to be attributed to live ancestors on swapin. But
5164 * those references are manageable from userspace.
5167 static DEFINE_IDR(mem_cgroup_idr);
5169 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5171 if (memcg->id.id > 0) {
5172 idr_remove(&mem_cgroup_idr, memcg->id.id);
5177 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5180 refcount_add(n, &memcg->id.ref);
5183 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5185 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5186 mem_cgroup_id_remove(memcg);
5188 /* Memcg ID pins CSS */
5189 css_put(&memcg->css);
5193 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5195 mem_cgroup_id_put_many(memcg, 1);
5199 * mem_cgroup_from_id - look up a memcg from a memcg id
5200 * @id: the memcg id to look up
5202 * Caller must hold rcu_read_lock().
5204 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5206 WARN_ON_ONCE(!rcu_read_lock_held());
5207 return idr_find(&mem_cgroup_idr, id);
5210 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5212 struct mem_cgroup_per_node *pn;
5215 * This routine is called against possible nodes.
5216 * But it's BUG to call kmalloc() against offline node.
5218 * TODO: this routine can waste much memory for nodes which will
5219 * never be onlined. It's better to use memory hotplug callback
5222 if (!node_state(node, N_NORMAL_MEMORY))
5224 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5228 pn->lruvec_stat_local = alloc_percpu_gfp(struct lruvec_stat,
5229 GFP_KERNEL_ACCOUNT);
5230 if (!pn->lruvec_stat_local) {
5235 pn->lruvec_stat_cpu = alloc_percpu_gfp(struct lruvec_stat,
5236 GFP_KERNEL_ACCOUNT);
5237 if (!pn->lruvec_stat_cpu) {
5238 free_percpu(pn->lruvec_stat_local);
5243 lruvec_init(&pn->lruvec);
5244 pn->usage_in_excess = 0;
5245 pn->on_tree = false;
5248 memcg->nodeinfo[node] = pn;
5252 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5254 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5259 free_percpu(pn->lruvec_stat_cpu);
5260 free_percpu(pn->lruvec_stat_local);
5264 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5269 free_mem_cgroup_per_node_info(memcg, node);
5270 free_percpu(memcg->vmstats_percpu);
5271 free_percpu(memcg->vmstats_local);
5275 static void mem_cgroup_free(struct mem_cgroup *memcg)
5277 memcg_wb_domain_exit(memcg);
5279 * Flush percpu vmstats and vmevents to guarantee the value correctness
5280 * on parent's and all ancestor levels.
5282 memcg_flush_percpu_vmstats(memcg);
5283 memcg_flush_percpu_vmevents(memcg);
5284 __mem_cgroup_free(memcg);
5287 static struct mem_cgroup *mem_cgroup_alloc(void)
5289 struct mem_cgroup *memcg;
5292 int __maybe_unused i;
5293 long error = -ENOMEM;
5295 size = sizeof(struct mem_cgroup);
5296 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5298 memcg = kzalloc(size, GFP_KERNEL);
5300 return ERR_PTR(error);
5302 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5303 1, MEM_CGROUP_ID_MAX,
5305 if (memcg->id.id < 0) {
5306 error = memcg->id.id;
5310 memcg->vmstats_local = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5311 GFP_KERNEL_ACCOUNT);
5312 if (!memcg->vmstats_local)
5315 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5316 GFP_KERNEL_ACCOUNT);
5317 if (!memcg->vmstats_percpu)
5321 if (alloc_mem_cgroup_per_node_info(memcg, node))
5324 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5327 INIT_WORK(&memcg->high_work, high_work_func);
5328 INIT_LIST_HEAD(&memcg->oom_notify);
5329 mutex_init(&memcg->thresholds_lock);
5330 spin_lock_init(&memcg->move_lock);
5331 vmpressure_init(&memcg->vmpressure);
5332 INIT_LIST_HEAD(&memcg->event_list);
5333 spin_lock_init(&memcg->event_list_lock);
5334 memcg->socket_pressure = jiffies;
5335 #ifdef CONFIG_MEMCG_KMEM
5336 memcg->kmemcg_id = -1;
5337 INIT_LIST_HEAD(&memcg->objcg_list);
5339 #ifdef CONFIG_CGROUP_WRITEBACK
5340 INIT_LIST_HEAD(&memcg->cgwb_list);
5341 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5342 memcg->cgwb_frn[i].done =
5343 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5345 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5346 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5347 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5348 memcg->deferred_split_queue.split_queue_len = 0;
5350 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5353 mem_cgroup_id_remove(memcg);
5354 __mem_cgroup_free(memcg);
5355 return ERR_PTR(error);
5358 static struct cgroup_subsys_state * __ref
5359 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5361 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5362 struct mem_cgroup *memcg, *old_memcg;
5363 long error = -ENOMEM;
5365 old_memcg = set_active_memcg(parent);
5366 memcg = mem_cgroup_alloc();
5367 set_active_memcg(old_memcg);
5369 return ERR_CAST(memcg);
5371 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5372 memcg->soft_limit = PAGE_COUNTER_MAX;
5373 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5375 memcg->swappiness = mem_cgroup_swappiness(parent);
5376 memcg->oom_kill_disable = parent->oom_kill_disable;
5378 page_counter_init(&memcg->memory, &parent->memory);
5379 page_counter_init(&memcg->swap, &parent->swap);
5380 page_counter_init(&memcg->kmem, &parent->kmem);
5381 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5383 page_counter_init(&memcg->memory, NULL);
5384 page_counter_init(&memcg->swap, NULL);
5385 page_counter_init(&memcg->kmem, NULL);
5386 page_counter_init(&memcg->tcpmem, NULL);
5388 root_mem_cgroup = memcg;
5392 /* The following stuff does not apply to the root */
5393 error = memcg_online_kmem(memcg);
5397 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5398 static_branch_inc(&memcg_sockets_enabled_key);
5402 mem_cgroup_id_remove(memcg);
5403 mem_cgroup_free(memcg);
5404 return ERR_PTR(error);
5407 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5409 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5412 * A memcg must be visible for memcg_expand_shrinker_maps()
5413 * by the time the maps are allocated. So, we allocate maps
5414 * here, when for_each_mem_cgroup() can't skip it.
5416 if (memcg_alloc_shrinker_maps(memcg)) {
5417 mem_cgroup_id_remove(memcg);
5421 /* Online state pins memcg ID, memcg ID pins CSS */
5422 refcount_set(&memcg->id.ref, 1);
5427 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5429 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5430 struct mem_cgroup_event *event, *tmp;
5433 * Unregister events and notify userspace.
5434 * Notify userspace about cgroup removing only after rmdir of cgroup
5435 * directory to avoid race between userspace and kernelspace.
5437 spin_lock(&memcg->event_list_lock);
5438 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5439 list_del_init(&event->list);
5440 schedule_work(&event->remove);
5442 spin_unlock(&memcg->event_list_lock);
5444 page_counter_set_min(&memcg->memory, 0);
5445 page_counter_set_low(&memcg->memory, 0);
5447 memcg_offline_kmem(memcg);
5448 wb_memcg_offline(memcg);
5450 drain_all_stock(memcg);
5452 mem_cgroup_id_put(memcg);
5455 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5457 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5459 invalidate_reclaim_iterators(memcg);
5462 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5464 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5465 int __maybe_unused i;
5467 #ifdef CONFIG_CGROUP_WRITEBACK
5468 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5469 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5471 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5472 static_branch_dec(&memcg_sockets_enabled_key);
5474 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5475 static_branch_dec(&memcg_sockets_enabled_key);
5477 vmpressure_cleanup(&memcg->vmpressure);
5478 cancel_work_sync(&memcg->high_work);
5479 mem_cgroup_remove_from_trees(memcg);
5480 memcg_free_shrinker_maps(memcg);
5481 memcg_free_kmem(memcg);
5482 mem_cgroup_free(memcg);
5486 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5487 * @css: the target css
5489 * Reset the states of the mem_cgroup associated with @css. This is
5490 * invoked when the userland requests disabling on the default hierarchy
5491 * but the memcg is pinned through dependency. The memcg should stop
5492 * applying policies and should revert to the vanilla state as it may be
5493 * made visible again.
5495 * The current implementation only resets the essential configurations.
5496 * This needs to be expanded to cover all the visible parts.
5498 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5500 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5502 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5503 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5504 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5505 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5506 page_counter_set_min(&memcg->memory, 0);
5507 page_counter_set_low(&memcg->memory, 0);
5508 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5509 memcg->soft_limit = PAGE_COUNTER_MAX;
5510 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5511 memcg_wb_domain_size_changed(memcg);
5515 /* Handlers for move charge at task migration. */
5516 static int mem_cgroup_do_precharge(unsigned long count)
5520 /* Try a single bulk charge without reclaim first, kswapd may wake */
5521 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5523 mc.precharge += count;
5527 /* Try charges one by one with reclaim, but do not retry */
5529 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5543 enum mc_target_type {
5550 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5551 unsigned long addr, pte_t ptent)
5553 struct page *page = vm_normal_page(vma, addr, ptent);
5555 if (!page || !page_mapped(page))
5557 if (PageAnon(page)) {
5558 if (!(mc.flags & MOVE_ANON))
5561 if (!(mc.flags & MOVE_FILE))
5564 if (!get_page_unless_zero(page))
5570 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5571 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5572 pte_t ptent, swp_entry_t *entry)
5574 struct page *page = NULL;
5575 swp_entry_t ent = pte_to_swp_entry(ptent);
5577 if (!(mc.flags & MOVE_ANON))
5581 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5582 * a device and because they are not accessible by CPU they are store
5583 * as special swap entry in the CPU page table.
5585 if (is_device_private_entry(ent)) {
5586 page = device_private_entry_to_page(ent);
5588 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5589 * a refcount of 1 when free (unlike normal page)
5591 if (!page_ref_add_unless(page, 1, 1))
5596 if (non_swap_entry(ent))
5600 * Because lookup_swap_cache() updates some statistics counter,
5601 * we call find_get_page() with swapper_space directly.
5603 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5604 entry->val = ent.val;
5609 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5610 pte_t ptent, swp_entry_t *entry)
5616 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5617 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5619 if (!vma->vm_file) /* anonymous vma */
5621 if (!(mc.flags & MOVE_FILE))
5624 /* page is moved even if it's not RSS of this task(page-faulted). */
5625 /* shmem/tmpfs may report page out on swap: account for that too. */
5626 return find_get_incore_page(vma->vm_file->f_mapping,
5627 linear_page_index(vma, addr));
5631 * mem_cgroup_move_account - move account of the page
5633 * @compound: charge the page as compound or small page
5634 * @from: mem_cgroup which the page is moved from.
5635 * @to: mem_cgroup which the page is moved to. @from != @to.
5637 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5639 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5642 static int mem_cgroup_move_account(struct page *page,
5644 struct mem_cgroup *from,
5645 struct mem_cgroup *to)
5647 struct lruvec *from_vec, *to_vec;
5648 struct pglist_data *pgdat;
5649 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5652 VM_BUG_ON(from == to);
5653 VM_BUG_ON_PAGE(PageLRU(page), page);
5654 VM_BUG_ON(compound && !PageTransHuge(page));
5657 * Prevent mem_cgroup_migrate() from looking at
5658 * page's memory cgroup of its source page while we change it.
5661 if (!trylock_page(page))
5665 if (page_memcg(page) != from)
5668 pgdat = page_pgdat(page);
5669 from_vec = mem_cgroup_lruvec(from, pgdat);
5670 to_vec = mem_cgroup_lruvec(to, pgdat);
5672 lock_page_memcg(page);
5674 if (PageAnon(page)) {
5675 if (page_mapped(page)) {
5676 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5677 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5678 if (PageTransHuge(page)) {
5679 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5681 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5687 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5688 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5690 if (PageSwapBacked(page)) {
5691 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5692 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5695 if (page_mapped(page)) {
5696 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5697 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5700 if (PageDirty(page)) {
5701 struct address_space *mapping = page_mapping(page);
5703 if (mapping_can_writeback(mapping)) {
5704 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5706 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5712 if (PageWriteback(page)) {
5713 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5714 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5718 * All state has been migrated, let's switch to the new memcg.
5720 * It is safe to change page's memcg here because the page
5721 * is referenced, charged, isolated, and locked: we can't race
5722 * with (un)charging, migration, LRU putback, or anything else
5723 * that would rely on a stable page's memory cgroup.
5725 * Note that lock_page_memcg is a memcg lock, not a page lock,
5726 * to save space. As soon as we switch page's memory cgroup to a
5727 * new memcg that isn't locked, the above state can change
5728 * concurrently again. Make sure we're truly done with it.
5733 css_put(&from->css);
5735 page->memcg_data = (unsigned long)to;
5737 __unlock_page_memcg(from);
5741 local_irq_disable();
5742 mem_cgroup_charge_statistics(to, page, nr_pages);
5743 memcg_check_events(to, page);
5744 mem_cgroup_charge_statistics(from, page, -nr_pages);
5745 memcg_check_events(from, page);
5754 * get_mctgt_type - get target type of moving charge
5755 * @vma: the vma the pte to be checked belongs
5756 * @addr: the address corresponding to the pte to be checked
5757 * @ptent: the pte to be checked
5758 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5761 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5762 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5763 * move charge. if @target is not NULL, the page is stored in target->page
5764 * with extra refcnt got(Callers should handle it).
5765 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5766 * target for charge migration. if @target is not NULL, the entry is stored
5768 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5769 * (so ZONE_DEVICE page and thus not on the lru).
5770 * For now we such page is charge like a regular page would be as for all
5771 * intent and purposes it is just special memory taking the place of a
5774 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5776 * Called with pte lock held.
5779 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5780 unsigned long addr, pte_t ptent, union mc_target *target)
5782 struct page *page = NULL;
5783 enum mc_target_type ret = MC_TARGET_NONE;
5784 swp_entry_t ent = { .val = 0 };
5786 if (pte_present(ptent))
5787 page = mc_handle_present_pte(vma, addr, ptent);
5788 else if (is_swap_pte(ptent))
5789 page = mc_handle_swap_pte(vma, ptent, &ent);
5790 else if (pte_none(ptent))
5791 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5793 if (!page && !ent.val)
5797 * Do only loose check w/o serialization.
5798 * mem_cgroup_move_account() checks the page is valid or
5799 * not under LRU exclusion.
5801 if (page_memcg(page) == mc.from) {
5802 ret = MC_TARGET_PAGE;
5803 if (is_device_private_page(page))
5804 ret = MC_TARGET_DEVICE;
5806 target->page = page;
5808 if (!ret || !target)
5812 * There is a swap entry and a page doesn't exist or isn't charged.
5813 * But we cannot move a tail-page in a THP.
5815 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5816 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5817 ret = MC_TARGET_SWAP;
5824 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5826 * We don't consider PMD mapped swapping or file mapped pages because THP does
5827 * not support them for now.
5828 * Caller should make sure that pmd_trans_huge(pmd) is true.
5830 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5831 unsigned long addr, pmd_t pmd, union mc_target *target)
5833 struct page *page = NULL;
5834 enum mc_target_type ret = MC_TARGET_NONE;
5836 if (unlikely(is_swap_pmd(pmd))) {
5837 VM_BUG_ON(thp_migration_supported() &&
5838 !is_pmd_migration_entry(pmd));
5841 page = pmd_page(pmd);
5842 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5843 if (!(mc.flags & MOVE_ANON))
5845 if (page_memcg(page) == mc.from) {
5846 ret = MC_TARGET_PAGE;
5849 target->page = page;
5855 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5856 unsigned long addr, pmd_t pmd, union mc_target *target)
5858 return MC_TARGET_NONE;
5862 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5863 unsigned long addr, unsigned long end,
5864 struct mm_walk *walk)
5866 struct vm_area_struct *vma = walk->vma;
5870 ptl = pmd_trans_huge_lock(pmd, vma);
5873 * Note their can not be MC_TARGET_DEVICE for now as we do not
5874 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5875 * this might change.
5877 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5878 mc.precharge += HPAGE_PMD_NR;
5883 if (pmd_trans_unstable(pmd))
5885 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5886 for (; addr != end; pte++, addr += PAGE_SIZE)
5887 if (get_mctgt_type(vma, addr, *pte, NULL))
5888 mc.precharge++; /* increment precharge temporarily */
5889 pte_unmap_unlock(pte - 1, ptl);
5895 static const struct mm_walk_ops precharge_walk_ops = {
5896 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5899 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5901 unsigned long precharge;
5904 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5905 mmap_read_unlock(mm);
5907 precharge = mc.precharge;
5913 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5915 unsigned long precharge = mem_cgroup_count_precharge(mm);
5917 VM_BUG_ON(mc.moving_task);
5918 mc.moving_task = current;
5919 return mem_cgroup_do_precharge(precharge);
5922 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5923 static void __mem_cgroup_clear_mc(void)
5925 struct mem_cgroup *from = mc.from;
5926 struct mem_cgroup *to = mc.to;
5928 /* we must uncharge all the leftover precharges from mc.to */
5930 cancel_charge(mc.to, mc.precharge);
5934 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5935 * we must uncharge here.
5937 if (mc.moved_charge) {
5938 cancel_charge(mc.from, mc.moved_charge);
5939 mc.moved_charge = 0;
5941 /* we must fixup refcnts and charges */
5942 if (mc.moved_swap) {
5943 /* uncharge swap account from the old cgroup */
5944 if (!mem_cgroup_is_root(mc.from))
5945 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5947 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5950 * we charged both to->memory and to->memsw, so we
5951 * should uncharge to->memory.
5953 if (!mem_cgroup_is_root(mc.to))
5954 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5958 memcg_oom_recover(from);
5959 memcg_oom_recover(to);
5960 wake_up_all(&mc.waitq);
5963 static void mem_cgroup_clear_mc(void)
5965 struct mm_struct *mm = mc.mm;
5968 * we must clear moving_task before waking up waiters at the end of
5971 mc.moving_task = NULL;
5972 __mem_cgroup_clear_mc();
5973 spin_lock(&mc.lock);
5977 spin_unlock(&mc.lock);
5982 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5984 struct cgroup_subsys_state *css;
5985 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5986 struct mem_cgroup *from;
5987 struct task_struct *leader, *p;
5988 struct mm_struct *mm;
5989 unsigned long move_flags;
5992 /* charge immigration isn't supported on the default hierarchy */
5993 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5997 * Multi-process migrations only happen on the default hierarchy
5998 * where charge immigration is not used. Perform charge
5999 * immigration if @tset contains a leader and whine if there are
6003 cgroup_taskset_for_each_leader(leader, css, tset) {
6006 memcg = mem_cgroup_from_css(css);
6012 * We are now commited to this value whatever it is. Changes in this
6013 * tunable will only affect upcoming migrations, not the current one.
6014 * So we need to save it, and keep it going.
6016 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
6020 from = mem_cgroup_from_task(p);
6022 VM_BUG_ON(from == memcg);
6024 mm = get_task_mm(p);
6027 /* We move charges only when we move a owner of the mm */
6028 if (mm->owner == p) {
6031 VM_BUG_ON(mc.precharge);
6032 VM_BUG_ON(mc.moved_charge);
6033 VM_BUG_ON(mc.moved_swap);
6035 spin_lock(&mc.lock);
6039 mc.flags = move_flags;
6040 spin_unlock(&mc.lock);
6041 /* We set mc.moving_task later */
6043 ret = mem_cgroup_precharge_mc(mm);
6045 mem_cgroup_clear_mc();
6052 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6055 mem_cgroup_clear_mc();
6058 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6059 unsigned long addr, unsigned long end,
6060 struct mm_walk *walk)
6063 struct vm_area_struct *vma = walk->vma;
6066 enum mc_target_type target_type;
6067 union mc_target target;
6070 ptl = pmd_trans_huge_lock(pmd, vma);
6072 if (mc.precharge < HPAGE_PMD_NR) {
6076 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6077 if (target_type == MC_TARGET_PAGE) {
6079 if (!isolate_lru_page(page)) {
6080 if (!mem_cgroup_move_account(page, true,
6082 mc.precharge -= HPAGE_PMD_NR;
6083 mc.moved_charge += HPAGE_PMD_NR;
6085 putback_lru_page(page);
6088 } else if (target_type == MC_TARGET_DEVICE) {
6090 if (!mem_cgroup_move_account(page, true,
6092 mc.precharge -= HPAGE_PMD_NR;
6093 mc.moved_charge += HPAGE_PMD_NR;
6101 if (pmd_trans_unstable(pmd))
6104 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6105 for (; addr != end; addr += PAGE_SIZE) {
6106 pte_t ptent = *(pte++);
6107 bool device = false;
6113 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6114 case MC_TARGET_DEVICE:
6117 case MC_TARGET_PAGE:
6120 * We can have a part of the split pmd here. Moving it
6121 * can be done but it would be too convoluted so simply
6122 * ignore such a partial THP and keep it in original
6123 * memcg. There should be somebody mapping the head.
6125 if (PageTransCompound(page))
6127 if (!device && isolate_lru_page(page))
6129 if (!mem_cgroup_move_account(page, false,
6132 /* we uncharge from mc.from later. */
6136 putback_lru_page(page);
6137 put: /* get_mctgt_type() gets the page */
6140 case MC_TARGET_SWAP:
6142 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6144 mem_cgroup_id_get_many(mc.to, 1);
6145 /* we fixup other refcnts and charges later. */
6153 pte_unmap_unlock(pte - 1, ptl);
6158 * We have consumed all precharges we got in can_attach().
6159 * We try charge one by one, but don't do any additional
6160 * charges to mc.to if we have failed in charge once in attach()
6163 ret = mem_cgroup_do_precharge(1);
6171 static const struct mm_walk_ops charge_walk_ops = {
6172 .pmd_entry = mem_cgroup_move_charge_pte_range,
6175 static void mem_cgroup_move_charge(void)
6177 lru_add_drain_all();
6179 * Signal lock_page_memcg() to take the memcg's move_lock
6180 * while we're moving its pages to another memcg. Then wait
6181 * for already started RCU-only updates to finish.
6183 atomic_inc(&mc.from->moving_account);
6186 if (unlikely(!mmap_read_trylock(mc.mm))) {
6188 * Someone who are holding the mmap_lock might be waiting in
6189 * waitq. So we cancel all extra charges, wake up all waiters,
6190 * and retry. Because we cancel precharges, we might not be able
6191 * to move enough charges, but moving charge is a best-effort
6192 * feature anyway, so it wouldn't be a big problem.
6194 __mem_cgroup_clear_mc();
6199 * When we have consumed all precharges and failed in doing
6200 * additional charge, the page walk just aborts.
6202 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6205 mmap_read_unlock(mc.mm);
6206 atomic_dec(&mc.from->moving_account);
6209 static void mem_cgroup_move_task(void)
6212 mem_cgroup_move_charge();
6213 mem_cgroup_clear_mc();
6216 #else /* !CONFIG_MMU */
6217 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6221 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6224 static void mem_cgroup_move_task(void)
6229 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6231 if (value == PAGE_COUNTER_MAX)
6232 seq_puts(m, "max\n");
6234 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6239 static u64 memory_current_read(struct cgroup_subsys_state *css,
6242 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6244 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6247 static int memory_min_show(struct seq_file *m, void *v)
6249 return seq_puts_memcg_tunable(m,
6250 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6253 static ssize_t memory_min_write(struct kernfs_open_file *of,
6254 char *buf, size_t nbytes, loff_t off)
6256 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6260 buf = strstrip(buf);
6261 err = page_counter_memparse(buf, "max", &min);
6265 page_counter_set_min(&memcg->memory, min);
6270 static int memory_low_show(struct seq_file *m, void *v)
6272 return seq_puts_memcg_tunable(m,
6273 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6276 static ssize_t memory_low_write(struct kernfs_open_file *of,
6277 char *buf, size_t nbytes, loff_t off)
6279 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6283 buf = strstrip(buf);
6284 err = page_counter_memparse(buf, "max", &low);
6288 page_counter_set_low(&memcg->memory, low);
6293 static int memory_high_show(struct seq_file *m, void *v)
6295 return seq_puts_memcg_tunable(m,
6296 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6299 static ssize_t memory_high_write(struct kernfs_open_file *of,
6300 char *buf, size_t nbytes, loff_t off)
6302 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6303 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6304 bool drained = false;
6308 buf = strstrip(buf);
6309 err = page_counter_memparse(buf, "max", &high);
6314 unsigned long nr_pages = page_counter_read(&memcg->memory);
6315 unsigned long reclaimed;
6317 if (nr_pages <= high)
6320 if (signal_pending(current))
6324 drain_all_stock(memcg);
6329 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6332 if (!reclaimed && !nr_retries--)
6336 page_counter_set_high(&memcg->memory, high);
6338 memcg_wb_domain_size_changed(memcg);
6343 static int memory_max_show(struct seq_file *m, void *v)
6345 return seq_puts_memcg_tunable(m,
6346 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6349 static ssize_t memory_max_write(struct kernfs_open_file *of,
6350 char *buf, size_t nbytes, loff_t off)
6352 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6353 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6354 bool drained = false;
6358 buf = strstrip(buf);
6359 err = page_counter_memparse(buf, "max", &max);
6363 xchg(&memcg->memory.max, max);
6366 unsigned long nr_pages = page_counter_read(&memcg->memory);
6368 if (nr_pages <= max)
6371 if (signal_pending(current))
6375 drain_all_stock(memcg);
6381 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6387 memcg_memory_event(memcg, MEMCG_OOM);
6388 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6392 memcg_wb_domain_size_changed(memcg);
6396 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6398 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6399 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6400 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6401 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6402 seq_printf(m, "oom_kill %lu\n",
6403 atomic_long_read(&events[MEMCG_OOM_KILL]));
6406 static int memory_events_show(struct seq_file *m, void *v)
6408 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6410 __memory_events_show(m, memcg->memory_events);
6414 static int memory_events_local_show(struct seq_file *m, void *v)
6416 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6418 __memory_events_show(m, memcg->memory_events_local);
6422 static int memory_stat_show(struct seq_file *m, void *v)
6424 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6427 buf = memory_stat_format(memcg);
6436 static int memory_numa_stat_show(struct seq_file *m, void *v)
6439 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6441 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6444 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6447 seq_printf(m, "%s", memory_stats[i].name);
6448 for_each_node_state(nid, N_MEMORY) {
6450 struct lruvec *lruvec;
6452 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6453 size = lruvec_page_state(lruvec, memory_stats[i].idx);
6454 size *= memory_stats[i].ratio;
6455 seq_printf(m, " N%d=%llu", nid, size);
6464 static int memory_oom_group_show(struct seq_file *m, void *v)
6466 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6468 seq_printf(m, "%d\n", memcg->oom_group);
6473 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6474 char *buf, size_t nbytes, loff_t off)
6476 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6479 buf = strstrip(buf);
6483 ret = kstrtoint(buf, 0, &oom_group);
6487 if (oom_group != 0 && oom_group != 1)
6490 memcg->oom_group = oom_group;
6495 static struct cftype memory_files[] = {
6498 .flags = CFTYPE_NOT_ON_ROOT,
6499 .read_u64 = memory_current_read,
6503 .flags = CFTYPE_NOT_ON_ROOT,
6504 .seq_show = memory_min_show,
6505 .write = memory_min_write,
6509 .flags = CFTYPE_NOT_ON_ROOT,
6510 .seq_show = memory_low_show,
6511 .write = memory_low_write,
6515 .flags = CFTYPE_NOT_ON_ROOT,
6516 .seq_show = memory_high_show,
6517 .write = memory_high_write,
6521 .flags = CFTYPE_NOT_ON_ROOT,
6522 .seq_show = memory_max_show,
6523 .write = memory_max_write,
6527 .flags = CFTYPE_NOT_ON_ROOT,
6528 .file_offset = offsetof(struct mem_cgroup, events_file),
6529 .seq_show = memory_events_show,
6532 .name = "events.local",
6533 .flags = CFTYPE_NOT_ON_ROOT,
6534 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6535 .seq_show = memory_events_local_show,
6539 .seq_show = memory_stat_show,
6543 .name = "numa_stat",
6544 .seq_show = memory_numa_stat_show,
6548 .name = "oom.group",
6549 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6550 .seq_show = memory_oom_group_show,
6551 .write = memory_oom_group_write,
6556 struct cgroup_subsys memory_cgrp_subsys = {
6557 .css_alloc = mem_cgroup_css_alloc,
6558 .css_online = mem_cgroup_css_online,
6559 .css_offline = mem_cgroup_css_offline,
6560 .css_released = mem_cgroup_css_released,
6561 .css_free = mem_cgroup_css_free,
6562 .css_reset = mem_cgroup_css_reset,
6563 .can_attach = mem_cgroup_can_attach,
6564 .cancel_attach = mem_cgroup_cancel_attach,
6565 .post_attach = mem_cgroup_move_task,
6566 .dfl_cftypes = memory_files,
6567 .legacy_cftypes = mem_cgroup_legacy_files,
6572 * This function calculates an individual cgroup's effective
6573 * protection which is derived from its own memory.min/low, its
6574 * parent's and siblings' settings, as well as the actual memory
6575 * distribution in the tree.
6577 * The following rules apply to the effective protection values:
6579 * 1. At the first level of reclaim, effective protection is equal to
6580 * the declared protection in memory.min and memory.low.
6582 * 2. To enable safe delegation of the protection configuration, at
6583 * subsequent levels the effective protection is capped to the
6584 * parent's effective protection.
6586 * 3. To make complex and dynamic subtrees easier to configure, the
6587 * user is allowed to overcommit the declared protection at a given
6588 * level. If that is the case, the parent's effective protection is
6589 * distributed to the children in proportion to how much protection
6590 * they have declared and how much of it they are utilizing.
6592 * This makes distribution proportional, but also work-conserving:
6593 * if one cgroup claims much more protection than it uses memory,
6594 * the unused remainder is available to its siblings.
6596 * 4. Conversely, when the declared protection is undercommitted at a
6597 * given level, the distribution of the larger parental protection
6598 * budget is NOT proportional. A cgroup's protection from a sibling
6599 * is capped to its own memory.min/low setting.
6601 * 5. However, to allow protecting recursive subtrees from each other
6602 * without having to declare each individual cgroup's fixed share
6603 * of the ancestor's claim to protection, any unutilized -
6604 * "floating" - protection from up the tree is distributed in
6605 * proportion to each cgroup's *usage*. This makes the protection
6606 * neutral wrt sibling cgroups and lets them compete freely over
6607 * the shared parental protection budget, but it protects the
6608 * subtree as a whole from neighboring subtrees.
6610 * Note that 4. and 5. are not in conflict: 4. is about protecting
6611 * against immediate siblings whereas 5. is about protecting against
6612 * neighboring subtrees.
6614 static unsigned long effective_protection(unsigned long usage,
6615 unsigned long parent_usage,
6616 unsigned long setting,
6617 unsigned long parent_effective,
6618 unsigned long siblings_protected)
6620 unsigned long protected;
6623 protected = min(usage, setting);
6625 * If all cgroups at this level combined claim and use more
6626 * protection then what the parent affords them, distribute
6627 * shares in proportion to utilization.
6629 * We are using actual utilization rather than the statically
6630 * claimed protection in order to be work-conserving: claimed
6631 * but unused protection is available to siblings that would
6632 * otherwise get a smaller chunk than what they claimed.
6634 if (siblings_protected > parent_effective)
6635 return protected * parent_effective / siblings_protected;
6638 * Ok, utilized protection of all children is within what the
6639 * parent affords them, so we know whatever this child claims
6640 * and utilizes is effectively protected.
6642 * If there is unprotected usage beyond this value, reclaim
6643 * will apply pressure in proportion to that amount.
6645 * If there is unutilized protection, the cgroup will be fully
6646 * shielded from reclaim, but we do return a smaller value for
6647 * protection than what the group could enjoy in theory. This
6648 * is okay. With the overcommit distribution above, effective
6649 * protection is always dependent on how memory is actually
6650 * consumed among the siblings anyway.
6655 * If the children aren't claiming (all of) the protection
6656 * afforded to them by the parent, distribute the remainder in
6657 * proportion to the (unprotected) memory of each cgroup. That
6658 * way, cgroups that aren't explicitly prioritized wrt each
6659 * other compete freely over the allowance, but they are
6660 * collectively protected from neighboring trees.
6662 * We're using unprotected memory for the weight so that if
6663 * some cgroups DO claim explicit protection, we don't protect
6664 * the same bytes twice.
6666 * Check both usage and parent_usage against the respective
6667 * protected values. One should imply the other, but they
6668 * aren't read atomically - make sure the division is sane.
6670 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6672 if (parent_effective > siblings_protected &&
6673 parent_usage > siblings_protected &&
6674 usage > protected) {
6675 unsigned long unclaimed;
6677 unclaimed = parent_effective - siblings_protected;
6678 unclaimed *= usage - protected;
6679 unclaimed /= parent_usage - siblings_protected;
6688 * mem_cgroup_protected - check if memory consumption is in the normal range
6689 * @root: the top ancestor of the sub-tree being checked
6690 * @memcg: the memory cgroup to check
6692 * WARNING: This function is not stateless! It can only be used as part
6693 * of a top-down tree iteration, not for isolated queries.
6695 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6696 struct mem_cgroup *memcg)
6698 unsigned long usage, parent_usage;
6699 struct mem_cgroup *parent;
6701 if (mem_cgroup_disabled())
6705 root = root_mem_cgroup;
6708 * Effective values of the reclaim targets are ignored so they
6709 * can be stale. Have a look at mem_cgroup_protection for more
6711 * TODO: calculation should be more robust so that we do not need
6712 * that special casing.
6717 usage = page_counter_read(&memcg->memory);
6721 parent = parent_mem_cgroup(memcg);
6722 /* No parent means a non-hierarchical mode on v1 memcg */
6726 if (parent == root) {
6727 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6728 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6732 parent_usage = page_counter_read(&parent->memory);
6734 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6735 READ_ONCE(memcg->memory.min),
6736 READ_ONCE(parent->memory.emin),
6737 atomic_long_read(&parent->memory.children_min_usage)));
6739 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6740 READ_ONCE(memcg->memory.low),
6741 READ_ONCE(parent->memory.elow),
6742 atomic_long_read(&parent->memory.children_low_usage)));
6746 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6747 * @page: page to charge
6748 * @mm: mm context of the victim
6749 * @gfp_mask: reclaim mode
6751 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6752 * pages according to @gfp_mask if necessary.
6754 * Returns 0 on success. Otherwise, an error code is returned.
6756 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6758 unsigned int nr_pages = thp_nr_pages(page);
6759 struct mem_cgroup *memcg = NULL;
6762 if (mem_cgroup_disabled())
6765 if (PageSwapCache(page)) {
6766 swp_entry_t ent = { .val = page_private(page), };
6770 * Every swap fault against a single page tries to charge the
6771 * page, bail as early as possible. shmem_unuse() encounters
6772 * already charged pages, too. page and memcg binding is
6773 * protected by the page lock, which serializes swap cache
6774 * removal, which in turn serializes uncharging.
6776 VM_BUG_ON_PAGE(!PageLocked(page), page);
6777 if (page_memcg(compound_head(page)))
6780 id = lookup_swap_cgroup_id(ent);
6782 memcg = mem_cgroup_from_id(id);
6783 if (memcg && !css_tryget_online(&memcg->css))
6789 memcg = get_mem_cgroup_from_mm(mm);
6791 ret = try_charge(memcg, gfp_mask, nr_pages);
6795 css_get(&memcg->css);
6796 commit_charge(page, memcg);
6798 local_irq_disable();
6799 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6800 memcg_check_events(memcg, page);
6803 if (PageSwapCache(page)) {
6804 swp_entry_t entry = { .val = page_private(page) };
6806 * The swap entry might not get freed for a long time,
6807 * let's not wait for it. The page already received a
6808 * memory+swap charge, drop the swap entry duplicate.
6810 mem_cgroup_uncharge_swap(entry, nr_pages);
6814 css_put(&memcg->css);
6819 struct uncharge_gather {
6820 struct mem_cgroup *memcg;
6821 unsigned long nr_pages;
6822 unsigned long pgpgout;
6823 unsigned long nr_kmem;
6824 struct page *dummy_page;
6827 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6829 memset(ug, 0, sizeof(*ug));
6832 static void uncharge_batch(const struct uncharge_gather *ug)
6834 unsigned long flags;
6836 if (!mem_cgroup_is_root(ug->memcg)) {
6837 page_counter_uncharge(&ug->memcg->memory, ug->nr_pages);
6838 if (do_memsw_account())
6839 page_counter_uncharge(&ug->memcg->memsw, ug->nr_pages);
6840 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6841 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6842 memcg_oom_recover(ug->memcg);
6845 local_irq_save(flags);
6846 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6847 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_pages);
6848 memcg_check_events(ug->memcg, ug->dummy_page);
6849 local_irq_restore(flags);
6851 /* drop reference from uncharge_page */
6852 css_put(&ug->memcg->css);
6855 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6857 unsigned long nr_pages;
6859 VM_BUG_ON_PAGE(PageLRU(page), page);
6861 if (!page_memcg(page))
6865 * Nobody should be changing or seriously looking at
6866 * page_memcg(page) at this point, we have fully
6867 * exclusive access to the page.
6870 if (ug->memcg != page_memcg(page)) {
6873 uncharge_gather_clear(ug);
6875 ug->memcg = page_memcg(page);
6877 /* pairs with css_put in uncharge_batch */
6878 css_get(&ug->memcg->css);
6881 nr_pages = compound_nr(page);
6882 ug->nr_pages += nr_pages;
6884 if (PageMemcgKmem(page))
6885 ug->nr_kmem += nr_pages;
6889 ug->dummy_page = page;
6890 page->memcg_data = 0;
6891 css_put(&ug->memcg->css);
6894 static void uncharge_list(struct list_head *page_list)
6896 struct uncharge_gather ug;
6897 struct list_head *next;
6899 uncharge_gather_clear(&ug);
6902 * Note that the list can be a single page->lru; hence the
6903 * do-while loop instead of a simple list_for_each_entry().
6905 next = page_list->next;
6909 page = list_entry(next, struct page, lru);
6910 next = page->lru.next;
6912 uncharge_page(page, &ug);
6913 } while (next != page_list);
6916 uncharge_batch(&ug);
6920 * mem_cgroup_uncharge - uncharge a page
6921 * @page: page to uncharge
6923 * Uncharge a page previously charged with mem_cgroup_charge().
6925 void mem_cgroup_uncharge(struct page *page)
6927 struct uncharge_gather ug;
6929 if (mem_cgroup_disabled())
6932 /* Don't touch page->lru of any random page, pre-check: */
6933 if (!page_memcg(page))
6936 uncharge_gather_clear(&ug);
6937 uncharge_page(page, &ug);
6938 uncharge_batch(&ug);
6942 * mem_cgroup_uncharge_list - uncharge a list of page
6943 * @page_list: list of pages to uncharge
6945 * Uncharge a list of pages previously charged with
6946 * mem_cgroup_charge().
6948 void mem_cgroup_uncharge_list(struct list_head *page_list)
6950 if (mem_cgroup_disabled())
6953 if (!list_empty(page_list))
6954 uncharge_list(page_list);
6958 * mem_cgroup_migrate - charge a page's replacement
6959 * @oldpage: currently circulating page
6960 * @newpage: replacement page
6962 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6963 * be uncharged upon free.
6965 * Both pages must be locked, @newpage->mapping must be set up.
6967 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6969 struct mem_cgroup *memcg;
6970 unsigned int nr_pages;
6971 unsigned long flags;
6973 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6974 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6975 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6976 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6979 if (mem_cgroup_disabled())
6982 /* Page cache replacement: new page already charged? */
6983 if (page_memcg(newpage))
6986 memcg = page_memcg(oldpage);
6987 VM_WARN_ON_ONCE_PAGE(!memcg, oldpage);
6991 /* Force-charge the new page. The old one will be freed soon */
6992 nr_pages = thp_nr_pages(newpage);
6994 page_counter_charge(&memcg->memory, nr_pages);
6995 if (do_memsw_account())
6996 page_counter_charge(&memcg->memsw, nr_pages);
6998 css_get(&memcg->css);
6999 commit_charge(newpage, memcg);
7001 local_irq_save(flags);
7002 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
7003 memcg_check_events(memcg, newpage);
7004 local_irq_restore(flags);
7007 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
7008 EXPORT_SYMBOL(memcg_sockets_enabled_key);
7010 void mem_cgroup_sk_alloc(struct sock *sk)
7012 struct mem_cgroup *memcg;
7014 if (!mem_cgroup_sockets_enabled)
7017 /* Do not associate the sock with unrelated interrupted task's memcg. */
7022 memcg = mem_cgroup_from_task(current);
7023 if (memcg == root_mem_cgroup)
7025 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
7027 if (css_tryget(&memcg->css))
7028 sk->sk_memcg = memcg;
7033 void mem_cgroup_sk_free(struct sock *sk)
7036 css_put(&sk->sk_memcg->css);
7040 * mem_cgroup_charge_skmem - charge socket memory
7041 * @memcg: memcg to charge
7042 * @nr_pages: number of pages to charge
7044 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7045 * @memcg's configured limit, %false if the charge had to be forced.
7047 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7049 gfp_t gfp_mask = GFP_KERNEL;
7051 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7052 struct page_counter *fail;
7054 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7055 memcg->tcpmem_pressure = 0;
7058 page_counter_charge(&memcg->tcpmem, nr_pages);
7059 memcg->tcpmem_pressure = 1;
7063 /* Don't block in the packet receive path */
7065 gfp_mask = GFP_NOWAIT;
7067 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7069 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
7072 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
7077 * mem_cgroup_uncharge_skmem - uncharge socket memory
7078 * @memcg: memcg to uncharge
7079 * @nr_pages: number of pages to uncharge
7081 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7083 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7084 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7088 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7090 refill_stock(memcg, nr_pages);
7093 static int __init cgroup_memory(char *s)
7097 while ((token = strsep(&s, ",")) != NULL) {
7100 if (!strcmp(token, "nosocket"))
7101 cgroup_memory_nosocket = true;
7102 if (!strcmp(token, "nokmem"))
7103 cgroup_memory_nokmem = true;
7107 __setup("cgroup.memory=", cgroup_memory);
7110 * subsys_initcall() for memory controller.
7112 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7113 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7114 * basically everything that doesn't depend on a specific mem_cgroup structure
7115 * should be initialized from here.
7117 static int __init mem_cgroup_init(void)
7121 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7122 memcg_hotplug_cpu_dead);
7124 for_each_possible_cpu(cpu)
7125 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7128 for_each_node(node) {
7129 struct mem_cgroup_tree_per_node *rtpn;
7131 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7132 node_online(node) ? node : NUMA_NO_NODE);
7134 rtpn->rb_root = RB_ROOT;
7135 rtpn->rb_rightmost = NULL;
7136 spin_lock_init(&rtpn->lock);
7137 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7142 subsys_initcall(mem_cgroup_init);
7144 #ifdef CONFIG_MEMCG_SWAP
7145 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7147 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7149 * The root cgroup cannot be destroyed, so it's refcount must
7152 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7156 memcg = parent_mem_cgroup(memcg);
7158 memcg = root_mem_cgroup;
7164 * mem_cgroup_swapout - transfer a memsw charge to swap
7165 * @page: page whose memsw charge to transfer
7166 * @entry: swap entry to move the charge to
7168 * Transfer the memsw charge of @page to @entry.
7170 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7172 struct mem_cgroup *memcg, *swap_memcg;
7173 unsigned int nr_entries;
7174 unsigned short oldid;
7176 VM_BUG_ON_PAGE(PageLRU(page), page);
7177 VM_BUG_ON_PAGE(page_count(page), page);
7179 if (mem_cgroup_disabled())
7182 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7185 memcg = page_memcg(page);
7187 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7192 * In case the memcg owning these pages has been offlined and doesn't
7193 * have an ID allocated to it anymore, charge the closest online
7194 * ancestor for the swap instead and transfer the memory+swap charge.
7196 swap_memcg = mem_cgroup_id_get_online(memcg);
7197 nr_entries = thp_nr_pages(page);
7198 /* Get references for the tail pages, too */
7200 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7201 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7203 VM_BUG_ON_PAGE(oldid, page);
7204 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7206 page->memcg_data = 0;
7208 if (!mem_cgroup_is_root(memcg))
7209 page_counter_uncharge(&memcg->memory, nr_entries);
7211 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7212 if (!mem_cgroup_is_root(swap_memcg))
7213 page_counter_charge(&swap_memcg->memsw, nr_entries);
7214 page_counter_uncharge(&memcg->memsw, nr_entries);
7218 * Interrupts should be disabled here because the caller holds the
7219 * i_pages lock which is taken with interrupts-off. It is
7220 * important here to have the interrupts disabled because it is the
7221 * only synchronisation we have for updating the per-CPU variables.
7223 VM_BUG_ON(!irqs_disabled());
7224 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7225 memcg_check_events(memcg, page);
7227 css_put(&memcg->css);
7231 * mem_cgroup_try_charge_swap - try charging swap space for a page
7232 * @page: page being added to swap
7233 * @entry: swap entry to charge
7235 * Try to charge @page's memcg for the swap space at @entry.
7237 * Returns 0 on success, -ENOMEM on failure.
7239 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7241 unsigned int nr_pages = thp_nr_pages(page);
7242 struct page_counter *counter;
7243 struct mem_cgroup *memcg;
7244 unsigned short oldid;
7246 if (mem_cgroup_disabled())
7249 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7252 memcg = page_memcg(page);
7254 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7259 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7263 memcg = mem_cgroup_id_get_online(memcg);
7265 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7266 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7267 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7268 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7269 mem_cgroup_id_put(memcg);
7273 /* Get references for the tail pages, too */
7275 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7276 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7277 VM_BUG_ON_PAGE(oldid, page);
7278 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7284 * mem_cgroup_uncharge_swap - uncharge swap space
7285 * @entry: swap entry to uncharge
7286 * @nr_pages: the amount of swap space to uncharge
7288 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7290 struct mem_cgroup *memcg;
7293 id = swap_cgroup_record(entry, 0, nr_pages);
7295 memcg = mem_cgroup_from_id(id);
7297 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7298 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7299 page_counter_uncharge(&memcg->swap, nr_pages);
7301 page_counter_uncharge(&memcg->memsw, nr_pages);
7303 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7304 mem_cgroup_id_put_many(memcg, nr_pages);
7309 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7311 long nr_swap_pages = get_nr_swap_pages();
7313 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7314 return nr_swap_pages;
7315 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7316 nr_swap_pages = min_t(long, nr_swap_pages,
7317 READ_ONCE(memcg->swap.max) -
7318 page_counter_read(&memcg->swap));
7319 return nr_swap_pages;
7322 bool mem_cgroup_swap_full(struct page *page)
7324 struct mem_cgroup *memcg;
7326 VM_BUG_ON_PAGE(!PageLocked(page), page);
7330 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7333 memcg = page_memcg(page);
7337 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7338 unsigned long usage = page_counter_read(&memcg->swap);
7340 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7341 usage * 2 >= READ_ONCE(memcg->swap.max))
7348 static int __init setup_swap_account(char *s)
7350 if (!strcmp(s, "1"))
7351 cgroup_memory_noswap = false;
7352 else if (!strcmp(s, "0"))
7353 cgroup_memory_noswap = true;
7356 __setup("swapaccount=", setup_swap_account);
7358 static u64 swap_current_read(struct cgroup_subsys_state *css,
7361 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7363 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7366 static int swap_high_show(struct seq_file *m, void *v)
7368 return seq_puts_memcg_tunable(m,
7369 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7372 static ssize_t swap_high_write(struct kernfs_open_file *of,
7373 char *buf, size_t nbytes, loff_t off)
7375 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7379 buf = strstrip(buf);
7380 err = page_counter_memparse(buf, "max", &high);
7384 page_counter_set_high(&memcg->swap, high);
7389 static int swap_max_show(struct seq_file *m, void *v)
7391 return seq_puts_memcg_tunable(m,
7392 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7395 static ssize_t swap_max_write(struct kernfs_open_file *of,
7396 char *buf, size_t nbytes, loff_t off)
7398 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7402 buf = strstrip(buf);
7403 err = page_counter_memparse(buf, "max", &max);
7407 xchg(&memcg->swap.max, max);
7412 static int swap_events_show(struct seq_file *m, void *v)
7414 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7416 seq_printf(m, "high %lu\n",
7417 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7418 seq_printf(m, "max %lu\n",
7419 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7420 seq_printf(m, "fail %lu\n",
7421 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7426 static struct cftype swap_files[] = {
7428 .name = "swap.current",
7429 .flags = CFTYPE_NOT_ON_ROOT,
7430 .read_u64 = swap_current_read,
7433 .name = "swap.high",
7434 .flags = CFTYPE_NOT_ON_ROOT,
7435 .seq_show = swap_high_show,
7436 .write = swap_high_write,
7440 .flags = CFTYPE_NOT_ON_ROOT,
7441 .seq_show = swap_max_show,
7442 .write = swap_max_write,
7445 .name = "swap.events",
7446 .flags = CFTYPE_NOT_ON_ROOT,
7447 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7448 .seq_show = swap_events_show,
7453 static struct cftype memsw_files[] = {
7455 .name = "memsw.usage_in_bytes",
7456 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7457 .read_u64 = mem_cgroup_read_u64,
7460 .name = "memsw.max_usage_in_bytes",
7461 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7462 .write = mem_cgroup_reset,
7463 .read_u64 = mem_cgroup_read_u64,
7466 .name = "memsw.limit_in_bytes",
7467 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7468 .write = mem_cgroup_write,
7469 .read_u64 = mem_cgroup_read_u64,
7472 .name = "memsw.failcnt",
7473 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7474 .write = mem_cgroup_reset,
7475 .read_u64 = mem_cgroup_read_u64,
7477 { }, /* terminate */
7481 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7482 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7483 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7484 * boot parameter. This may result in premature OOPS inside
7485 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7487 static int __init mem_cgroup_swap_init(void)
7489 /* No memory control -> no swap control */
7490 if (mem_cgroup_disabled())
7491 cgroup_memory_noswap = true;
7493 if (cgroup_memory_noswap)
7496 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7497 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7501 core_initcall(mem_cgroup_swap_init);
7503 #endif /* CONFIG_MEMCG_SWAP */