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->mem_cgroup;
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->mem_cgroup;
569 * The lowest bit set means that memcg isn't a valid
570 * memcg pointer, but a obj_cgroups pointer.
571 * In this case the page is shared and doesn't belong
572 * to any specific memory cgroup.
574 if ((unsigned long) memcg & 0x1UL)
577 while (memcg && !(memcg->css.flags & CSS_ONLINE))
578 memcg = parent_mem_cgroup(memcg);
580 ino = cgroup_ino(memcg->css.cgroup);
585 static struct mem_cgroup_per_node *
586 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
588 int nid = page_to_nid(page);
590 return memcg->nodeinfo[nid];
593 static struct mem_cgroup_tree_per_node *
594 soft_limit_tree_node(int nid)
596 return soft_limit_tree.rb_tree_per_node[nid];
599 static struct mem_cgroup_tree_per_node *
600 soft_limit_tree_from_page(struct page *page)
602 int nid = page_to_nid(page);
604 return soft_limit_tree.rb_tree_per_node[nid];
607 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
608 struct mem_cgroup_tree_per_node *mctz,
609 unsigned long new_usage_in_excess)
611 struct rb_node **p = &mctz->rb_root.rb_node;
612 struct rb_node *parent = NULL;
613 struct mem_cgroup_per_node *mz_node;
614 bool rightmost = true;
619 mz->usage_in_excess = new_usage_in_excess;
620 if (!mz->usage_in_excess)
624 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
626 if (mz->usage_in_excess < mz_node->usage_in_excess) {
635 mctz->rb_rightmost = &mz->tree_node;
637 rb_link_node(&mz->tree_node, parent, p);
638 rb_insert_color(&mz->tree_node, &mctz->rb_root);
642 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
643 struct mem_cgroup_tree_per_node *mctz)
648 if (&mz->tree_node == mctz->rb_rightmost)
649 mctz->rb_rightmost = rb_prev(&mz->tree_node);
651 rb_erase(&mz->tree_node, &mctz->rb_root);
655 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
656 struct mem_cgroup_tree_per_node *mctz)
660 spin_lock_irqsave(&mctz->lock, flags);
661 __mem_cgroup_remove_exceeded(mz, mctz);
662 spin_unlock_irqrestore(&mctz->lock, flags);
665 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
667 unsigned long nr_pages = page_counter_read(&memcg->memory);
668 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
669 unsigned long excess = 0;
671 if (nr_pages > soft_limit)
672 excess = nr_pages - soft_limit;
677 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
679 unsigned long excess;
680 struct mem_cgroup_per_node *mz;
681 struct mem_cgroup_tree_per_node *mctz;
683 mctz = soft_limit_tree_from_page(page);
687 * Necessary to update all ancestors when hierarchy is used.
688 * because their event counter is not touched.
690 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
691 mz = mem_cgroup_page_nodeinfo(memcg, page);
692 excess = soft_limit_excess(memcg);
694 * We have to update the tree if mz is on RB-tree or
695 * mem is over its softlimit.
697 if (excess || mz->on_tree) {
700 spin_lock_irqsave(&mctz->lock, flags);
701 /* if on-tree, remove it */
703 __mem_cgroup_remove_exceeded(mz, mctz);
705 * Insert again. mz->usage_in_excess will be updated.
706 * If excess is 0, no tree ops.
708 __mem_cgroup_insert_exceeded(mz, mctz, excess);
709 spin_unlock_irqrestore(&mctz->lock, flags);
714 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
716 struct mem_cgroup_tree_per_node *mctz;
717 struct mem_cgroup_per_node *mz;
721 mz = mem_cgroup_nodeinfo(memcg, nid);
722 mctz = soft_limit_tree_node(nid);
724 mem_cgroup_remove_exceeded(mz, mctz);
728 static struct mem_cgroup_per_node *
729 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
731 struct mem_cgroup_per_node *mz;
735 if (!mctz->rb_rightmost)
736 goto done; /* Nothing to reclaim from */
738 mz = rb_entry(mctz->rb_rightmost,
739 struct mem_cgroup_per_node, tree_node);
741 * Remove the node now but someone else can add it back,
742 * we will to add it back at the end of reclaim to its correct
743 * position in the tree.
745 __mem_cgroup_remove_exceeded(mz, mctz);
746 if (!soft_limit_excess(mz->memcg) ||
747 !css_tryget(&mz->memcg->css))
753 static struct mem_cgroup_per_node *
754 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
756 struct mem_cgroup_per_node *mz;
758 spin_lock_irq(&mctz->lock);
759 mz = __mem_cgroup_largest_soft_limit_node(mctz);
760 spin_unlock_irq(&mctz->lock);
765 * __mod_memcg_state - update cgroup memory statistics
766 * @memcg: the memory cgroup
767 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
768 * @val: delta to add to the counter, can be negative
770 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
772 long x, threshold = MEMCG_CHARGE_BATCH;
774 if (mem_cgroup_disabled())
777 if (memcg_stat_item_in_bytes(idx))
778 threshold <<= PAGE_SHIFT;
780 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
781 if (unlikely(abs(x) > threshold)) {
782 struct mem_cgroup *mi;
785 * Batch local counters to keep them in sync with
786 * the hierarchical ones.
788 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
789 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
790 atomic_long_add(x, &mi->vmstats[idx]);
793 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
796 static struct mem_cgroup_per_node *
797 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
799 struct mem_cgroup *parent;
801 parent = parent_mem_cgroup(pn->memcg);
804 return mem_cgroup_nodeinfo(parent, nid);
807 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
810 struct mem_cgroup_per_node *pn;
811 struct mem_cgroup *memcg;
812 long x, threshold = MEMCG_CHARGE_BATCH;
814 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
818 __mod_memcg_state(memcg, idx, val);
821 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
823 if (vmstat_item_in_bytes(idx))
824 threshold <<= PAGE_SHIFT;
826 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
827 if (unlikely(abs(x) > threshold)) {
828 pg_data_t *pgdat = lruvec_pgdat(lruvec);
829 struct mem_cgroup_per_node *pi;
831 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
832 atomic_long_add(x, &pi->lruvec_stat[idx]);
835 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
839 * __mod_lruvec_state - update lruvec memory statistics
840 * @lruvec: the lruvec
841 * @idx: the stat item
842 * @val: delta to add to the counter, can be negative
844 * The lruvec is the intersection of the NUMA node and a cgroup. This
845 * function updates the all three counters that are affected by a
846 * change of state at this level: per-node, per-cgroup, per-lruvec.
848 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
852 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
854 /* Update memcg and lruvec */
855 if (!mem_cgroup_disabled())
856 __mod_memcg_lruvec_state(lruvec, idx, val);
859 void __mod_lruvec_page_state(struct page *page, enum node_stat_item idx,
862 struct page *head = compound_head(page); /* rmap on tail pages */
863 pg_data_t *pgdat = page_pgdat(page);
864 struct lruvec *lruvec;
866 /* Untracked pages have no memcg, no lruvec. Update only the node */
867 if (!head->mem_cgroup) {
868 __mod_node_page_state(pgdat, idx, val);
872 lruvec = mem_cgroup_lruvec(head->mem_cgroup, pgdat);
873 __mod_lruvec_state(lruvec, idx, val);
875 EXPORT_SYMBOL(__mod_lruvec_page_state);
877 void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val)
879 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
880 struct mem_cgroup *memcg;
881 struct lruvec *lruvec;
884 memcg = mem_cgroup_from_obj(p);
887 * Untracked pages have no memcg, no lruvec. Update only the
888 * node. If we reparent the slab objects to the root memcg,
889 * when we free the slab object, we need to update the per-memcg
890 * vmstats to keep it correct for the root memcg.
893 __mod_node_page_state(pgdat, idx, val);
895 lruvec = mem_cgroup_lruvec(memcg, pgdat);
896 __mod_lruvec_state(lruvec, idx, val);
902 * __count_memcg_events - account VM events in a cgroup
903 * @memcg: the memory cgroup
904 * @idx: the event item
905 * @count: the number of events that occured
907 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
912 if (mem_cgroup_disabled())
915 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
916 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
917 struct mem_cgroup *mi;
920 * Batch local counters to keep them in sync with
921 * the hierarchical ones.
923 __this_cpu_add(memcg->vmstats_local->events[idx], x);
924 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
925 atomic_long_add(x, &mi->vmevents[idx]);
928 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
931 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
933 return atomic_long_read(&memcg->vmevents[event]);
936 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
941 for_each_possible_cpu(cpu)
942 x += per_cpu(memcg->vmstats_local->events[event], cpu);
946 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
950 /* pagein of a big page is an event. So, ignore page size */
952 __count_memcg_events(memcg, PGPGIN, 1);
954 __count_memcg_events(memcg, PGPGOUT, 1);
955 nr_pages = -nr_pages; /* for event */
958 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
961 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
962 enum mem_cgroup_events_target target)
964 unsigned long val, next;
966 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
967 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
968 /* from time_after() in jiffies.h */
969 if ((long)(next - val) < 0) {
971 case MEM_CGROUP_TARGET_THRESH:
972 next = val + THRESHOLDS_EVENTS_TARGET;
974 case MEM_CGROUP_TARGET_SOFTLIMIT:
975 next = val + SOFTLIMIT_EVENTS_TARGET;
980 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
987 * Check events in order.
990 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
992 /* threshold event is triggered in finer grain than soft limit */
993 if (unlikely(mem_cgroup_event_ratelimit(memcg,
994 MEM_CGROUP_TARGET_THRESH))) {
997 do_softlimit = mem_cgroup_event_ratelimit(memcg,
998 MEM_CGROUP_TARGET_SOFTLIMIT);
999 mem_cgroup_threshold(memcg);
1000 if (unlikely(do_softlimit))
1001 mem_cgroup_update_tree(memcg, page);
1005 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1008 * mm_update_next_owner() may clear mm->owner to NULL
1009 * if it races with swapoff, page migration, etc.
1010 * So this can be called with p == NULL.
1015 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1017 EXPORT_SYMBOL(mem_cgroup_from_task);
1020 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1021 * @mm: mm from which memcg should be extracted. It can be NULL.
1023 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
1024 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
1027 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1029 struct mem_cgroup *memcg;
1031 if (mem_cgroup_disabled())
1037 * Page cache insertions can happen withou an
1038 * actual mm context, e.g. during disk probing
1039 * on boot, loopback IO, acct() writes etc.
1042 memcg = root_mem_cgroup;
1044 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1045 if (unlikely(!memcg))
1046 memcg = root_mem_cgroup;
1048 } while (!css_tryget(&memcg->css));
1052 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1055 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
1056 * @page: page from which memcg should be extracted.
1058 * Obtain a reference on page->memcg and returns it if successful. Otherwise
1059 * root_mem_cgroup is returned.
1061 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
1063 struct mem_cgroup *memcg = page->mem_cgroup;
1065 if (mem_cgroup_disabled())
1069 /* Page should not get uncharged and freed memcg under us. */
1070 if (!memcg || WARN_ON_ONCE(!css_tryget(&memcg->css)))
1071 memcg = root_mem_cgroup;
1075 EXPORT_SYMBOL(get_mem_cgroup_from_page);
1077 static __always_inline struct mem_cgroup *active_memcg(void)
1080 return this_cpu_read(int_active_memcg);
1082 return current->active_memcg;
1085 static __always_inline struct mem_cgroup *get_active_memcg(void)
1087 struct mem_cgroup *memcg;
1090 memcg = active_memcg();
1092 /* current->active_memcg must hold a ref. */
1093 if (WARN_ON_ONCE(!css_tryget(&memcg->css)))
1094 memcg = root_mem_cgroup;
1096 memcg = current->active_memcg;
1103 static __always_inline bool memcg_kmem_bypass(void)
1105 /* Allow remote memcg charging from any context. */
1106 if (unlikely(active_memcg()))
1109 /* Memcg to charge can't be determined. */
1110 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
1117 * If active memcg is set, do not fallback to current->mm->memcg.
1119 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
1121 if (memcg_kmem_bypass())
1124 if (unlikely(active_memcg()))
1125 return get_active_memcg();
1127 return get_mem_cgroup_from_mm(current->mm);
1131 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1132 * @root: hierarchy root
1133 * @prev: previously returned memcg, NULL on first invocation
1134 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1136 * Returns references to children of the hierarchy below @root, or
1137 * @root itself, or %NULL after a full round-trip.
1139 * Caller must pass the return value in @prev on subsequent
1140 * invocations for reference counting, or use mem_cgroup_iter_break()
1141 * to cancel a hierarchy walk before the round-trip is complete.
1143 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1144 * in the hierarchy among all concurrent reclaimers operating on the
1147 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1148 struct mem_cgroup *prev,
1149 struct mem_cgroup_reclaim_cookie *reclaim)
1151 struct mem_cgroup_reclaim_iter *iter;
1152 struct cgroup_subsys_state *css = NULL;
1153 struct mem_cgroup *memcg = NULL;
1154 struct mem_cgroup *pos = NULL;
1156 if (mem_cgroup_disabled())
1160 root = root_mem_cgroup;
1162 if (prev && !reclaim)
1168 struct mem_cgroup_per_node *mz;
1170 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1173 if (prev && reclaim->generation != iter->generation)
1177 pos = READ_ONCE(iter->position);
1178 if (!pos || css_tryget(&pos->css))
1181 * css reference reached zero, so iter->position will
1182 * be cleared by ->css_released. However, we should not
1183 * rely on this happening soon, because ->css_released
1184 * is called from a work queue, and by busy-waiting we
1185 * might block it. So we clear iter->position right
1188 (void)cmpxchg(&iter->position, pos, NULL);
1196 css = css_next_descendant_pre(css, &root->css);
1199 * Reclaimers share the hierarchy walk, and a
1200 * new one might jump in right at the end of
1201 * the hierarchy - make sure they see at least
1202 * one group and restart from the beginning.
1210 * Verify the css and acquire a reference. The root
1211 * is provided by the caller, so we know it's alive
1212 * and kicking, and don't take an extra reference.
1214 memcg = mem_cgroup_from_css(css);
1216 if (css == &root->css)
1219 if (css_tryget(css))
1227 * The position could have already been updated by a competing
1228 * thread, so check that the value hasn't changed since we read
1229 * it to avoid reclaiming from the same cgroup twice.
1231 (void)cmpxchg(&iter->position, pos, memcg);
1239 reclaim->generation = iter->generation;
1244 if (prev && prev != root)
1245 css_put(&prev->css);
1251 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1252 * @root: hierarchy root
1253 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1255 void mem_cgroup_iter_break(struct mem_cgroup *root,
1256 struct mem_cgroup *prev)
1259 root = root_mem_cgroup;
1260 if (prev && prev != root)
1261 css_put(&prev->css);
1264 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1265 struct mem_cgroup *dead_memcg)
1267 struct mem_cgroup_reclaim_iter *iter;
1268 struct mem_cgroup_per_node *mz;
1271 for_each_node(nid) {
1272 mz = mem_cgroup_nodeinfo(from, nid);
1274 cmpxchg(&iter->position, dead_memcg, NULL);
1278 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1280 struct mem_cgroup *memcg = dead_memcg;
1281 struct mem_cgroup *last;
1284 __invalidate_reclaim_iterators(memcg, dead_memcg);
1286 } while ((memcg = parent_mem_cgroup(memcg)));
1289 * When cgruop1 non-hierarchy mode is used,
1290 * parent_mem_cgroup() does not walk all the way up to the
1291 * cgroup root (root_mem_cgroup). So we have to handle
1292 * dead_memcg from cgroup root separately.
1294 if (last != root_mem_cgroup)
1295 __invalidate_reclaim_iterators(root_mem_cgroup,
1300 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1301 * @memcg: hierarchy root
1302 * @fn: function to call for each task
1303 * @arg: argument passed to @fn
1305 * This function iterates over tasks attached to @memcg or to any of its
1306 * descendants and calls @fn for each task. If @fn returns a non-zero
1307 * value, the function breaks the iteration loop and returns the value.
1308 * Otherwise, it will iterate over all tasks and return 0.
1310 * This function must not be called for the root memory cgroup.
1312 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1313 int (*fn)(struct task_struct *, void *), void *arg)
1315 struct mem_cgroup *iter;
1318 BUG_ON(memcg == root_mem_cgroup);
1320 for_each_mem_cgroup_tree(iter, memcg) {
1321 struct css_task_iter it;
1322 struct task_struct *task;
1324 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1325 while (!ret && (task = css_task_iter_next(&it)))
1326 ret = fn(task, arg);
1327 css_task_iter_end(&it);
1329 mem_cgroup_iter_break(memcg, iter);
1336 #ifdef CONFIG_DEBUG_VM
1337 void lruvec_memcg_debug(struct lruvec *lruvec, struct page *page)
1339 struct mem_cgroup *memcg;
1341 if (mem_cgroup_disabled())
1344 memcg = page_memcg(page);
1347 VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != root_mem_cgroup, page);
1349 VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != memcg, page);
1354 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1356 * @pgdat: pgdat of the page
1358 * This function relies on page's memcg being stable - see the
1359 * access rules in commit_charge().
1361 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1363 struct mem_cgroup_per_node *mz;
1364 struct mem_cgroup *memcg;
1365 struct lruvec *lruvec;
1367 if (mem_cgroup_disabled()) {
1368 lruvec = &pgdat->__lruvec;
1372 memcg = page->mem_cgroup;
1374 * Swapcache readahead pages are added to the LRU - and
1375 * possibly migrated - before they are charged.
1378 memcg = root_mem_cgroup;
1380 mz = mem_cgroup_page_nodeinfo(memcg, page);
1381 lruvec = &mz->lruvec;
1384 * Since a node can be onlined after the mem_cgroup was created,
1385 * we have to be prepared to initialize lruvec->zone here;
1386 * and if offlined then reonlined, we need to reinitialize it.
1388 if (unlikely(lruvec->pgdat != pgdat))
1389 lruvec->pgdat = pgdat;
1394 * lock_page_lruvec - lock and return lruvec for a given page.
1397 * This series functions should be used in either conditions:
1398 * PageLRU is cleared or unset
1399 * or page->_refcount is zero
1400 * or page is locked.
1402 struct lruvec *lock_page_lruvec(struct page *page)
1404 struct lruvec *lruvec;
1405 struct pglist_data *pgdat = page_pgdat(page);
1408 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1409 spin_lock(&lruvec->lru_lock);
1412 lruvec_memcg_debug(lruvec, page);
1417 struct lruvec *lock_page_lruvec_irq(struct page *page)
1419 struct lruvec *lruvec;
1420 struct pglist_data *pgdat = page_pgdat(page);
1423 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1424 spin_lock_irq(&lruvec->lru_lock);
1427 lruvec_memcg_debug(lruvec, page);
1432 struct lruvec *lock_page_lruvec_irqsave(struct page *page, unsigned long *flags)
1434 struct lruvec *lruvec;
1435 struct pglist_data *pgdat = page_pgdat(page);
1438 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1439 spin_lock_irqsave(&lruvec->lru_lock, *flags);
1442 lruvec_memcg_debug(lruvec, page);
1448 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1449 * @lruvec: mem_cgroup per zone lru vector
1450 * @lru: index of lru list the page is sitting on
1451 * @zid: zone id of the accounted pages
1452 * @nr_pages: positive when adding or negative when removing
1454 * This function must be called under lru_lock, just before a page is added
1455 * to or just after a page is removed from an lru list (that ordering being
1456 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1458 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1459 int zid, int nr_pages)
1461 struct mem_cgroup_per_node *mz;
1462 unsigned long *lru_size;
1465 if (mem_cgroup_disabled())
1468 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1469 lru_size = &mz->lru_zone_size[zid][lru];
1472 *lru_size += nr_pages;
1475 if (WARN_ONCE(size < 0,
1476 "%s(%p, %d, %d): lru_size %ld\n",
1477 __func__, lruvec, lru, nr_pages, size)) {
1483 *lru_size += nr_pages;
1487 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1488 * @memcg: the memory cgroup
1490 * Returns the maximum amount of memory @mem can be charged with, in
1493 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1495 unsigned long margin = 0;
1496 unsigned long count;
1497 unsigned long limit;
1499 count = page_counter_read(&memcg->memory);
1500 limit = READ_ONCE(memcg->memory.max);
1502 margin = limit - count;
1504 if (do_memsw_account()) {
1505 count = page_counter_read(&memcg->memsw);
1506 limit = READ_ONCE(memcg->memsw.max);
1508 margin = min(margin, limit - count);
1517 * A routine for checking "mem" is under move_account() or not.
1519 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1520 * moving cgroups. This is for waiting at high-memory pressure
1523 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1525 struct mem_cgroup *from;
1526 struct mem_cgroup *to;
1529 * Unlike task_move routines, we access mc.to, mc.from not under
1530 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1532 spin_lock(&mc.lock);
1538 ret = mem_cgroup_is_descendant(from, memcg) ||
1539 mem_cgroup_is_descendant(to, memcg);
1541 spin_unlock(&mc.lock);
1545 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1547 if (mc.moving_task && current != mc.moving_task) {
1548 if (mem_cgroup_under_move(memcg)) {
1550 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1551 /* moving charge context might have finished. */
1554 finish_wait(&mc.waitq, &wait);
1561 struct memory_stat {
1567 static struct memory_stat memory_stats[] = {
1568 { "anon", PAGE_SIZE, NR_ANON_MAPPED },
1569 { "file", PAGE_SIZE, NR_FILE_PAGES },
1570 { "kernel_stack", 1024, NR_KERNEL_STACK_KB },
1571 { "pagetables", PAGE_SIZE, NR_PAGETABLE },
1572 { "percpu", 1, MEMCG_PERCPU_B },
1573 { "sock", PAGE_SIZE, MEMCG_SOCK },
1574 { "shmem", PAGE_SIZE, NR_SHMEM },
1575 { "file_mapped", PAGE_SIZE, NR_FILE_MAPPED },
1576 { "file_dirty", PAGE_SIZE, NR_FILE_DIRTY },
1577 { "file_writeback", PAGE_SIZE, NR_WRITEBACK },
1578 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1580 * The ratio will be initialized in memory_stats_init(). Because
1581 * on some architectures, the macro of HPAGE_PMD_SIZE is not
1582 * constant(e.g. powerpc).
1584 { "anon_thp", 0, NR_ANON_THPS },
1585 { "file_thp", 0, NR_FILE_THPS },
1586 { "shmem_thp", 0, NR_SHMEM_THPS },
1588 { "inactive_anon", PAGE_SIZE, NR_INACTIVE_ANON },
1589 { "active_anon", PAGE_SIZE, NR_ACTIVE_ANON },
1590 { "inactive_file", PAGE_SIZE, NR_INACTIVE_FILE },
1591 { "active_file", PAGE_SIZE, NR_ACTIVE_FILE },
1592 { "unevictable", PAGE_SIZE, NR_UNEVICTABLE },
1595 * Note: The slab_reclaimable and slab_unreclaimable must be
1596 * together and slab_reclaimable must be in front.
1598 { "slab_reclaimable", 1, NR_SLAB_RECLAIMABLE_B },
1599 { "slab_unreclaimable", 1, NR_SLAB_UNRECLAIMABLE_B },
1601 /* The memory events */
1602 { "workingset_refault_anon", 1, WORKINGSET_REFAULT_ANON },
1603 { "workingset_refault_file", 1, WORKINGSET_REFAULT_FILE },
1604 { "workingset_activate_anon", 1, WORKINGSET_ACTIVATE_ANON },
1605 { "workingset_activate_file", 1, WORKINGSET_ACTIVATE_FILE },
1606 { "workingset_restore_anon", 1, WORKINGSET_RESTORE_ANON },
1607 { "workingset_restore_file", 1, WORKINGSET_RESTORE_FILE },
1608 { "workingset_nodereclaim", 1, WORKINGSET_NODERECLAIM },
1611 static int __init memory_stats_init(void)
1615 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1616 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1617 if (memory_stats[i].idx == NR_ANON_THPS ||
1618 memory_stats[i].idx == NR_FILE_THPS ||
1619 memory_stats[i].idx == NR_SHMEM_THPS)
1620 memory_stats[i].ratio = HPAGE_PMD_SIZE;
1622 VM_BUG_ON(!memory_stats[i].ratio);
1623 VM_BUG_ON(memory_stats[i].idx >= MEMCG_NR_STAT);
1628 pure_initcall(memory_stats_init);
1630 static char *memory_stat_format(struct mem_cgroup *memcg)
1635 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1640 * Provide statistics on the state of the memory subsystem as
1641 * well as cumulative event counters that show past behavior.
1643 * This list is ordered following a combination of these gradients:
1644 * 1) generic big picture -> specifics and details
1645 * 2) reflecting userspace activity -> reflecting kernel heuristics
1647 * Current memory state:
1650 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1653 size = memcg_page_state(memcg, memory_stats[i].idx);
1654 size *= memory_stats[i].ratio;
1655 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1657 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1658 size = memcg_page_state(memcg, NR_SLAB_RECLAIMABLE_B) +
1659 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE_B);
1660 seq_buf_printf(&s, "slab %llu\n", size);
1664 /* Accumulated memory events */
1666 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1667 memcg_events(memcg, PGFAULT));
1668 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1669 memcg_events(memcg, PGMAJFAULT));
1670 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1671 memcg_events(memcg, PGREFILL));
1672 seq_buf_printf(&s, "pgscan %lu\n",
1673 memcg_events(memcg, PGSCAN_KSWAPD) +
1674 memcg_events(memcg, PGSCAN_DIRECT));
1675 seq_buf_printf(&s, "pgsteal %lu\n",
1676 memcg_events(memcg, PGSTEAL_KSWAPD) +
1677 memcg_events(memcg, PGSTEAL_DIRECT));
1678 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1679 memcg_events(memcg, PGACTIVATE));
1680 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1681 memcg_events(memcg, PGDEACTIVATE));
1682 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1683 memcg_events(memcg, PGLAZYFREE));
1684 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1685 memcg_events(memcg, PGLAZYFREED));
1687 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1688 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1689 memcg_events(memcg, THP_FAULT_ALLOC));
1690 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1691 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1692 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1694 /* The above should easily fit into one page */
1695 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1700 #define K(x) ((x) << (PAGE_SHIFT-10))
1702 * mem_cgroup_print_oom_context: Print OOM information relevant to
1703 * memory controller.
1704 * @memcg: The memory cgroup that went over limit
1705 * @p: Task that is going to be killed
1707 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1710 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1715 pr_cont(",oom_memcg=");
1716 pr_cont_cgroup_path(memcg->css.cgroup);
1718 pr_cont(",global_oom");
1720 pr_cont(",task_memcg=");
1721 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1727 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1728 * memory controller.
1729 * @memcg: The memory cgroup that went over limit
1731 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1735 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1736 K((u64)page_counter_read(&memcg->memory)),
1737 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1738 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1739 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1740 K((u64)page_counter_read(&memcg->swap)),
1741 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1743 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1744 K((u64)page_counter_read(&memcg->memsw)),
1745 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1746 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1747 K((u64)page_counter_read(&memcg->kmem)),
1748 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1751 pr_info("Memory cgroup stats for ");
1752 pr_cont_cgroup_path(memcg->css.cgroup);
1754 buf = memory_stat_format(memcg);
1762 * Return the memory (and swap, if configured) limit for a memcg.
1764 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1766 unsigned long max = READ_ONCE(memcg->memory.max);
1768 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1769 if (mem_cgroup_swappiness(memcg))
1770 max += min(READ_ONCE(memcg->swap.max),
1771 (unsigned long)total_swap_pages);
1773 if (mem_cgroup_swappiness(memcg)) {
1774 /* Calculate swap excess capacity from memsw limit */
1775 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1777 max += min(swap, (unsigned long)total_swap_pages);
1783 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1785 return page_counter_read(&memcg->memory);
1788 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1791 struct oom_control oc = {
1795 .gfp_mask = gfp_mask,
1800 if (mutex_lock_killable(&oom_lock))
1803 if (mem_cgroup_margin(memcg) >= (1 << order))
1807 * A few threads which were not waiting at mutex_lock_killable() can
1808 * fail to bail out. Therefore, check again after holding oom_lock.
1810 ret = should_force_charge() || out_of_memory(&oc);
1813 mutex_unlock(&oom_lock);
1817 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1820 unsigned long *total_scanned)
1822 struct mem_cgroup *victim = NULL;
1825 unsigned long excess;
1826 unsigned long nr_scanned;
1827 struct mem_cgroup_reclaim_cookie reclaim = {
1831 excess = soft_limit_excess(root_memcg);
1834 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1839 * If we have not been able to reclaim
1840 * anything, it might because there are
1841 * no reclaimable pages under this hierarchy
1846 * We want to do more targeted reclaim.
1847 * excess >> 2 is not to excessive so as to
1848 * reclaim too much, nor too less that we keep
1849 * coming back to reclaim from this cgroup
1851 if (total >= (excess >> 2) ||
1852 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1857 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1858 pgdat, &nr_scanned);
1859 *total_scanned += nr_scanned;
1860 if (!soft_limit_excess(root_memcg))
1863 mem_cgroup_iter_break(root_memcg, victim);
1867 #ifdef CONFIG_LOCKDEP
1868 static struct lockdep_map memcg_oom_lock_dep_map = {
1869 .name = "memcg_oom_lock",
1873 static DEFINE_SPINLOCK(memcg_oom_lock);
1876 * Check OOM-Killer is already running under our hierarchy.
1877 * If someone is running, return false.
1879 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1881 struct mem_cgroup *iter, *failed = NULL;
1883 spin_lock(&memcg_oom_lock);
1885 for_each_mem_cgroup_tree(iter, memcg) {
1886 if (iter->oom_lock) {
1888 * this subtree of our hierarchy is already locked
1889 * so we cannot give a lock.
1892 mem_cgroup_iter_break(memcg, iter);
1895 iter->oom_lock = true;
1900 * OK, we failed to lock the whole subtree so we have
1901 * to clean up what we set up to the failing subtree
1903 for_each_mem_cgroup_tree(iter, memcg) {
1904 if (iter == failed) {
1905 mem_cgroup_iter_break(memcg, iter);
1908 iter->oom_lock = false;
1911 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1913 spin_unlock(&memcg_oom_lock);
1918 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1920 struct mem_cgroup *iter;
1922 spin_lock(&memcg_oom_lock);
1923 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1924 for_each_mem_cgroup_tree(iter, memcg)
1925 iter->oom_lock = false;
1926 spin_unlock(&memcg_oom_lock);
1929 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1931 struct mem_cgroup *iter;
1933 spin_lock(&memcg_oom_lock);
1934 for_each_mem_cgroup_tree(iter, memcg)
1936 spin_unlock(&memcg_oom_lock);
1939 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1941 struct mem_cgroup *iter;
1944 * Be careful about under_oom underflows becase a child memcg
1945 * could have been added after mem_cgroup_mark_under_oom.
1947 spin_lock(&memcg_oom_lock);
1948 for_each_mem_cgroup_tree(iter, memcg)
1949 if (iter->under_oom > 0)
1951 spin_unlock(&memcg_oom_lock);
1954 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1956 struct oom_wait_info {
1957 struct mem_cgroup *memcg;
1958 wait_queue_entry_t wait;
1961 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1962 unsigned mode, int sync, void *arg)
1964 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1965 struct mem_cgroup *oom_wait_memcg;
1966 struct oom_wait_info *oom_wait_info;
1968 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1969 oom_wait_memcg = oom_wait_info->memcg;
1971 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1972 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1974 return autoremove_wake_function(wait, mode, sync, arg);
1977 static void memcg_oom_recover(struct mem_cgroup *memcg)
1980 * For the following lockless ->under_oom test, the only required
1981 * guarantee is that it must see the state asserted by an OOM when
1982 * this function is called as a result of userland actions
1983 * triggered by the notification of the OOM. This is trivially
1984 * achieved by invoking mem_cgroup_mark_under_oom() before
1985 * triggering notification.
1987 if (memcg && memcg->under_oom)
1988 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1998 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2000 enum oom_status ret;
2003 if (order > PAGE_ALLOC_COSTLY_ORDER)
2006 memcg_memory_event(memcg, MEMCG_OOM);
2009 * We are in the middle of the charge context here, so we
2010 * don't want to block when potentially sitting on a callstack
2011 * that holds all kinds of filesystem and mm locks.
2013 * cgroup1 allows disabling the OOM killer and waiting for outside
2014 * handling until the charge can succeed; remember the context and put
2015 * the task to sleep at the end of the page fault when all locks are
2018 * On the other hand, in-kernel OOM killer allows for an async victim
2019 * memory reclaim (oom_reaper) and that means that we are not solely
2020 * relying on the oom victim to make a forward progress and we can
2021 * invoke the oom killer here.
2023 * Please note that mem_cgroup_out_of_memory might fail to find a
2024 * victim and then we have to bail out from the charge path.
2026 if (memcg->oom_kill_disable) {
2027 if (!current->in_user_fault)
2029 css_get(&memcg->css);
2030 current->memcg_in_oom = memcg;
2031 current->memcg_oom_gfp_mask = mask;
2032 current->memcg_oom_order = order;
2037 mem_cgroup_mark_under_oom(memcg);
2039 locked = mem_cgroup_oom_trylock(memcg);
2042 mem_cgroup_oom_notify(memcg);
2044 mem_cgroup_unmark_under_oom(memcg);
2045 if (mem_cgroup_out_of_memory(memcg, mask, order))
2051 mem_cgroup_oom_unlock(memcg);
2057 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2058 * @handle: actually kill/wait or just clean up the OOM state
2060 * This has to be called at the end of a page fault if the memcg OOM
2061 * handler was enabled.
2063 * Memcg supports userspace OOM handling where failed allocations must
2064 * sleep on a waitqueue until the userspace task resolves the
2065 * situation. Sleeping directly in the charge context with all kinds
2066 * of locks held is not a good idea, instead we remember an OOM state
2067 * in the task and mem_cgroup_oom_synchronize() has to be called at
2068 * the end of the page fault to complete the OOM handling.
2070 * Returns %true if an ongoing memcg OOM situation was detected and
2071 * completed, %false otherwise.
2073 bool mem_cgroup_oom_synchronize(bool handle)
2075 struct mem_cgroup *memcg = current->memcg_in_oom;
2076 struct oom_wait_info owait;
2079 /* OOM is global, do not handle */
2086 owait.memcg = memcg;
2087 owait.wait.flags = 0;
2088 owait.wait.func = memcg_oom_wake_function;
2089 owait.wait.private = current;
2090 INIT_LIST_HEAD(&owait.wait.entry);
2092 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2093 mem_cgroup_mark_under_oom(memcg);
2095 locked = mem_cgroup_oom_trylock(memcg);
2098 mem_cgroup_oom_notify(memcg);
2100 if (locked && !memcg->oom_kill_disable) {
2101 mem_cgroup_unmark_under_oom(memcg);
2102 finish_wait(&memcg_oom_waitq, &owait.wait);
2103 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
2104 current->memcg_oom_order);
2107 mem_cgroup_unmark_under_oom(memcg);
2108 finish_wait(&memcg_oom_waitq, &owait.wait);
2112 mem_cgroup_oom_unlock(memcg);
2114 * There is no guarantee that an OOM-lock contender
2115 * sees the wakeups triggered by the OOM kill
2116 * uncharges. Wake any sleepers explicitely.
2118 memcg_oom_recover(memcg);
2121 current->memcg_in_oom = NULL;
2122 css_put(&memcg->css);
2127 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2128 * @victim: task to be killed by the OOM killer
2129 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2131 * Returns a pointer to a memory cgroup, which has to be cleaned up
2132 * by killing all belonging OOM-killable tasks.
2134 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2136 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2137 struct mem_cgroup *oom_domain)
2139 struct mem_cgroup *oom_group = NULL;
2140 struct mem_cgroup *memcg;
2142 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2146 oom_domain = root_mem_cgroup;
2150 memcg = mem_cgroup_from_task(victim);
2151 if (memcg == root_mem_cgroup)
2155 * If the victim task has been asynchronously moved to a different
2156 * memory cgroup, we might end up killing tasks outside oom_domain.
2157 * In this case it's better to ignore memory.group.oom.
2159 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2163 * Traverse the memory cgroup hierarchy from the victim task's
2164 * cgroup up to the OOMing cgroup (or root) to find the
2165 * highest-level memory cgroup with oom.group set.
2167 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2168 if (memcg->oom_group)
2171 if (memcg == oom_domain)
2176 css_get(&oom_group->css);
2183 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2185 pr_info("Tasks in ");
2186 pr_cont_cgroup_path(memcg->css.cgroup);
2187 pr_cont(" are going to be killed due to memory.oom.group set\n");
2191 * lock_page_memcg - lock a page->mem_cgroup binding
2194 * This function protects unlocked LRU pages from being moved to
2197 * It ensures lifetime of the returned memcg. Caller is responsible
2198 * for the lifetime of the page; __unlock_page_memcg() is available
2199 * when @page might get freed inside the locked section.
2201 struct mem_cgroup *lock_page_memcg(struct page *page)
2203 struct page *head = compound_head(page); /* rmap on tail pages */
2204 struct mem_cgroup *memcg;
2205 unsigned long flags;
2208 * The RCU lock is held throughout the transaction. The fast
2209 * path can get away without acquiring the memcg->move_lock
2210 * because page moving starts with an RCU grace period.
2212 * The RCU lock also protects the memcg from being freed when
2213 * the page state that is going to change is the only thing
2214 * preventing the page itself from being freed. E.g. writeback
2215 * doesn't hold a page reference and relies on PG_writeback to
2216 * keep off truncation, migration and so forth.
2220 if (mem_cgroup_disabled())
2223 memcg = head->mem_cgroup;
2224 if (unlikely(!memcg))
2227 #ifdef CONFIG_PROVE_LOCKING
2228 local_irq_save(flags);
2229 might_lock(&memcg->move_lock);
2230 local_irq_restore(flags);
2233 if (atomic_read(&memcg->moving_account) <= 0)
2236 spin_lock_irqsave(&memcg->move_lock, flags);
2237 if (memcg != head->mem_cgroup) {
2238 spin_unlock_irqrestore(&memcg->move_lock, flags);
2243 * When charge migration first begins, we can have locked and
2244 * unlocked page stat updates happening concurrently. Track
2245 * the task who has the lock for unlock_page_memcg().
2247 memcg->move_lock_task = current;
2248 memcg->move_lock_flags = flags;
2252 EXPORT_SYMBOL(lock_page_memcg);
2255 * __unlock_page_memcg - unlock and unpin a memcg
2258 * Unlock and unpin a memcg returned by lock_page_memcg().
2260 void __unlock_page_memcg(struct mem_cgroup *memcg)
2262 if (memcg && memcg->move_lock_task == current) {
2263 unsigned long flags = memcg->move_lock_flags;
2265 memcg->move_lock_task = NULL;
2266 memcg->move_lock_flags = 0;
2268 spin_unlock_irqrestore(&memcg->move_lock, flags);
2275 * unlock_page_memcg - unlock a page->mem_cgroup binding
2278 void unlock_page_memcg(struct page *page)
2280 struct page *head = compound_head(page);
2282 __unlock_page_memcg(head->mem_cgroup);
2284 EXPORT_SYMBOL(unlock_page_memcg);
2286 struct memcg_stock_pcp {
2287 struct mem_cgroup *cached; /* this never be root cgroup */
2288 unsigned int nr_pages;
2290 #ifdef CONFIG_MEMCG_KMEM
2291 struct obj_cgroup *cached_objcg;
2292 unsigned int nr_bytes;
2295 struct work_struct work;
2296 unsigned long flags;
2297 #define FLUSHING_CACHED_CHARGE 0
2299 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2300 static DEFINE_MUTEX(percpu_charge_mutex);
2302 #ifdef CONFIG_MEMCG_KMEM
2303 static void drain_obj_stock(struct memcg_stock_pcp *stock);
2304 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2305 struct mem_cgroup *root_memcg);
2308 static inline void drain_obj_stock(struct memcg_stock_pcp *stock)
2311 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2312 struct mem_cgroup *root_memcg)
2319 * consume_stock: Try to consume stocked charge on this cpu.
2320 * @memcg: memcg to consume from.
2321 * @nr_pages: how many pages to charge.
2323 * The charges will only happen if @memcg matches the current cpu's memcg
2324 * stock, and at least @nr_pages are available in that stock. Failure to
2325 * service an allocation will refill the stock.
2327 * returns true if successful, false otherwise.
2329 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2331 struct memcg_stock_pcp *stock;
2332 unsigned long flags;
2335 if (nr_pages > MEMCG_CHARGE_BATCH)
2338 local_irq_save(flags);
2340 stock = this_cpu_ptr(&memcg_stock);
2341 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2342 stock->nr_pages -= nr_pages;
2346 local_irq_restore(flags);
2352 * Returns stocks cached in percpu and reset cached information.
2354 static void drain_stock(struct memcg_stock_pcp *stock)
2356 struct mem_cgroup *old = stock->cached;
2361 if (stock->nr_pages) {
2362 page_counter_uncharge(&old->memory, stock->nr_pages);
2363 if (do_memsw_account())
2364 page_counter_uncharge(&old->memsw, stock->nr_pages);
2365 stock->nr_pages = 0;
2369 stock->cached = NULL;
2372 static void drain_local_stock(struct work_struct *dummy)
2374 struct memcg_stock_pcp *stock;
2375 unsigned long flags;
2378 * The only protection from memory hotplug vs. drain_stock races is
2379 * that we always operate on local CPU stock here with IRQ disabled
2381 local_irq_save(flags);
2383 stock = this_cpu_ptr(&memcg_stock);
2384 drain_obj_stock(stock);
2386 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2388 local_irq_restore(flags);
2392 * Cache charges(val) to local per_cpu area.
2393 * This will be consumed by consume_stock() function, later.
2395 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2397 struct memcg_stock_pcp *stock;
2398 unsigned long flags;
2400 local_irq_save(flags);
2402 stock = this_cpu_ptr(&memcg_stock);
2403 if (stock->cached != memcg) { /* reset if necessary */
2405 css_get(&memcg->css);
2406 stock->cached = memcg;
2408 stock->nr_pages += nr_pages;
2410 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2413 local_irq_restore(flags);
2417 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2418 * of the hierarchy under it.
2420 static void drain_all_stock(struct mem_cgroup *root_memcg)
2424 /* If someone's already draining, avoid adding running more workers. */
2425 if (!mutex_trylock(&percpu_charge_mutex))
2428 * Notify other cpus that system-wide "drain" is running
2429 * We do not care about races with the cpu hotplug because cpu down
2430 * as well as workers from this path always operate on the local
2431 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2434 for_each_online_cpu(cpu) {
2435 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2436 struct mem_cgroup *memcg;
2440 memcg = stock->cached;
2441 if (memcg && stock->nr_pages &&
2442 mem_cgroup_is_descendant(memcg, root_memcg))
2444 if (obj_stock_flush_required(stock, root_memcg))
2449 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2451 drain_local_stock(&stock->work);
2453 schedule_work_on(cpu, &stock->work);
2457 mutex_unlock(&percpu_charge_mutex);
2460 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2462 struct memcg_stock_pcp *stock;
2463 struct mem_cgroup *memcg, *mi;
2465 stock = &per_cpu(memcg_stock, cpu);
2468 for_each_mem_cgroup(memcg) {
2471 for (i = 0; i < MEMCG_NR_STAT; i++) {
2475 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2477 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2478 atomic_long_add(x, &memcg->vmstats[i]);
2480 if (i >= NR_VM_NODE_STAT_ITEMS)
2483 for_each_node(nid) {
2484 struct mem_cgroup_per_node *pn;
2486 pn = mem_cgroup_nodeinfo(memcg, nid);
2487 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2490 atomic_long_add(x, &pn->lruvec_stat[i]);
2491 } while ((pn = parent_nodeinfo(pn, nid)));
2495 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2498 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2500 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2501 atomic_long_add(x, &memcg->vmevents[i]);
2508 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2509 unsigned int nr_pages,
2512 unsigned long nr_reclaimed = 0;
2515 unsigned long pflags;
2517 if (page_counter_read(&memcg->memory) <=
2518 READ_ONCE(memcg->memory.high))
2521 memcg_memory_event(memcg, MEMCG_HIGH);
2523 psi_memstall_enter(&pflags);
2524 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2526 psi_memstall_leave(&pflags);
2527 } while ((memcg = parent_mem_cgroup(memcg)) &&
2528 !mem_cgroup_is_root(memcg));
2530 return nr_reclaimed;
2533 static void high_work_func(struct work_struct *work)
2535 struct mem_cgroup *memcg;
2537 memcg = container_of(work, struct mem_cgroup, high_work);
2538 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2542 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2543 * enough to still cause a significant slowdown in most cases, while still
2544 * allowing diagnostics and tracing to proceed without becoming stuck.
2546 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2549 * When calculating the delay, we use these either side of the exponentiation to
2550 * maintain precision and scale to a reasonable number of jiffies (see the table
2553 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2554 * overage ratio to a delay.
2555 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2556 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2557 * to produce a reasonable delay curve.
2559 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2560 * reasonable delay curve compared to precision-adjusted overage, not
2561 * penalising heavily at first, but still making sure that growth beyond the
2562 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2563 * example, with a high of 100 megabytes:
2565 * +-------+------------------------+
2566 * | usage | time to allocate in ms |
2567 * +-------+------------------------+
2589 * +-------+------------------------+
2591 #define MEMCG_DELAY_PRECISION_SHIFT 20
2592 #define MEMCG_DELAY_SCALING_SHIFT 14
2594 static u64 calculate_overage(unsigned long usage, unsigned long high)
2602 * Prevent division by 0 in overage calculation by acting as if
2603 * it was a threshold of 1 page
2605 high = max(high, 1UL);
2607 overage = usage - high;
2608 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2609 return div64_u64(overage, high);
2612 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2614 u64 overage, max_overage = 0;
2617 overage = calculate_overage(page_counter_read(&memcg->memory),
2618 READ_ONCE(memcg->memory.high));
2619 max_overage = max(overage, max_overage);
2620 } while ((memcg = parent_mem_cgroup(memcg)) &&
2621 !mem_cgroup_is_root(memcg));
2626 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2628 u64 overage, max_overage = 0;
2631 overage = calculate_overage(page_counter_read(&memcg->swap),
2632 READ_ONCE(memcg->swap.high));
2634 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2635 max_overage = max(overage, max_overage);
2636 } while ((memcg = parent_mem_cgroup(memcg)) &&
2637 !mem_cgroup_is_root(memcg));
2643 * Get the number of jiffies that we should penalise a mischievous cgroup which
2644 * is exceeding its memory.high by checking both it and its ancestors.
2646 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2647 unsigned int nr_pages,
2650 unsigned long penalty_jiffies;
2656 * We use overage compared to memory.high to calculate the number of
2657 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2658 * fairly lenient on small overages, and increasingly harsh when the
2659 * memcg in question makes it clear that it has no intention of stopping
2660 * its crazy behaviour, so we exponentially increase the delay based on
2663 penalty_jiffies = max_overage * max_overage * HZ;
2664 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2665 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2668 * Factor in the task's own contribution to the overage, such that four
2669 * N-sized allocations are throttled approximately the same as one
2670 * 4N-sized allocation.
2672 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2673 * larger the current charge patch is than that.
2675 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2679 * Scheduled by try_charge() to be executed from the userland return path
2680 * and reclaims memory over the high limit.
2682 void mem_cgroup_handle_over_high(void)
2684 unsigned long penalty_jiffies;
2685 unsigned long pflags;
2686 unsigned long nr_reclaimed;
2687 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2688 int nr_retries = MAX_RECLAIM_RETRIES;
2689 struct mem_cgroup *memcg;
2690 bool in_retry = false;
2692 if (likely(!nr_pages))
2695 memcg = get_mem_cgroup_from_mm(current->mm);
2696 current->memcg_nr_pages_over_high = 0;
2700 * The allocating task should reclaim at least the batch size, but for
2701 * subsequent retries we only want to do what's necessary to prevent oom
2702 * or breaching resource isolation.
2704 * This is distinct from memory.max or page allocator behaviour because
2705 * memory.high is currently batched, whereas memory.max and the page
2706 * allocator run every time an allocation is made.
2708 nr_reclaimed = reclaim_high(memcg,
2709 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2713 * memory.high is breached and reclaim is unable to keep up. Throttle
2714 * allocators proactively to slow down excessive growth.
2716 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2717 mem_find_max_overage(memcg));
2719 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2720 swap_find_max_overage(memcg));
2723 * Clamp the max delay per usermode return so as to still keep the
2724 * application moving forwards and also permit diagnostics, albeit
2727 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2730 * Don't sleep if the amount of jiffies this memcg owes us is so low
2731 * that it's not even worth doing, in an attempt to be nice to those who
2732 * go only a small amount over their memory.high value and maybe haven't
2733 * been aggressively reclaimed enough yet.
2735 if (penalty_jiffies <= HZ / 100)
2739 * If reclaim is making forward progress but we're still over
2740 * memory.high, we want to encourage that rather than doing allocator
2743 if (nr_reclaimed || nr_retries--) {
2749 * If we exit early, we're guaranteed to die (since
2750 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2751 * need to account for any ill-begotten jiffies to pay them off later.
2753 psi_memstall_enter(&pflags);
2754 schedule_timeout_killable(penalty_jiffies);
2755 psi_memstall_leave(&pflags);
2758 css_put(&memcg->css);
2761 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2762 unsigned int nr_pages)
2764 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2765 int nr_retries = MAX_RECLAIM_RETRIES;
2766 struct mem_cgroup *mem_over_limit;
2767 struct page_counter *counter;
2768 enum oom_status oom_status;
2769 unsigned long nr_reclaimed;
2770 bool may_swap = true;
2771 bool drained = false;
2772 unsigned long pflags;
2774 if (mem_cgroup_is_root(memcg))
2777 if (consume_stock(memcg, nr_pages))
2780 if (!do_memsw_account() ||
2781 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2782 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2784 if (do_memsw_account())
2785 page_counter_uncharge(&memcg->memsw, batch);
2786 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2788 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2792 if (batch > nr_pages) {
2798 * Memcg doesn't have a dedicated reserve for atomic
2799 * allocations. But like the global atomic pool, we need to
2800 * put the burden of reclaim on regular allocation requests
2801 * and let these go through as privileged allocations.
2803 if (gfp_mask & __GFP_ATOMIC)
2807 * Unlike in global OOM situations, memcg is not in a physical
2808 * memory shortage. Allow dying and OOM-killed tasks to
2809 * bypass the last charges so that they can exit quickly and
2810 * free their memory.
2812 if (unlikely(should_force_charge()))
2816 * Prevent unbounded recursion when reclaim operations need to
2817 * allocate memory. This might exceed the limits temporarily,
2818 * but we prefer facilitating memory reclaim and getting back
2819 * under the limit over triggering OOM kills in these cases.
2821 if (unlikely(current->flags & PF_MEMALLOC))
2824 if (unlikely(task_in_memcg_oom(current)))
2827 if (!gfpflags_allow_blocking(gfp_mask))
2830 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2832 psi_memstall_enter(&pflags);
2833 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2834 gfp_mask, may_swap);
2835 psi_memstall_leave(&pflags);
2837 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2841 drain_all_stock(mem_over_limit);
2846 if (gfp_mask & __GFP_NORETRY)
2849 * Even though the limit is exceeded at this point, reclaim
2850 * may have been able to free some pages. Retry the charge
2851 * before killing the task.
2853 * Only for regular pages, though: huge pages are rather
2854 * unlikely to succeed so close to the limit, and we fall back
2855 * to regular pages anyway in case of failure.
2857 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2860 * At task move, charge accounts can be doubly counted. So, it's
2861 * better to wait until the end of task_move if something is going on.
2863 if (mem_cgroup_wait_acct_move(mem_over_limit))
2869 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2872 if (gfp_mask & __GFP_NOFAIL)
2875 if (fatal_signal_pending(current))
2879 * keep retrying as long as the memcg oom killer is able to make
2880 * a forward progress or bypass the charge if the oom killer
2881 * couldn't make any progress.
2883 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2884 get_order(nr_pages * PAGE_SIZE));
2885 switch (oom_status) {
2887 nr_retries = MAX_RECLAIM_RETRIES;
2895 if (!(gfp_mask & __GFP_NOFAIL))
2899 * The allocation either can't fail or will lead to more memory
2900 * being freed very soon. Allow memory usage go over the limit
2901 * temporarily by force charging it.
2903 page_counter_charge(&memcg->memory, nr_pages);
2904 if (do_memsw_account())
2905 page_counter_charge(&memcg->memsw, nr_pages);
2910 if (batch > nr_pages)
2911 refill_stock(memcg, batch - nr_pages);
2914 * If the hierarchy is above the normal consumption range, schedule
2915 * reclaim on returning to userland. We can perform reclaim here
2916 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2917 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2918 * not recorded as it most likely matches current's and won't
2919 * change in the meantime. As high limit is checked again before
2920 * reclaim, the cost of mismatch is negligible.
2923 bool mem_high, swap_high;
2925 mem_high = page_counter_read(&memcg->memory) >
2926 READ_ONCE(memcg->memory.high);
2927 swap_high = page_counter_read(&memcg->swap) >
2928 READ_ONCE(memcg->swap.high);
2930 /* Don't bother a random interrupted task */
2931 if (in_interrupt()) {
2933 schedule_work(&memcg->high_work);
2939 if (mem_high || swap_high) {
2941 * The allocating tasks in this cgroup will need to do
2942 * reclaim or be throttled to prevent further growth
2943 * of the memory or swap footprints.
2945 * Target some best-effort fairness between the tasks,
2946 * and distribute reclaim work and delay penalties
2947 * based on how much each task is actually allocating.
2949 current->memcg_nr_pages_over_high += batch;
2950 set_notify_resume(current);
2953 } while ((memcg = parent_mem_cgroup(memcg)));
2958 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2959 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2961 if (mem_cgroup_is_root(memcg))
2964 page_counter_uncharge(&memcg->memory, nr_pages);
2965 if (do_memsw_account())
2966 page_counter_uncharge(&memcg->memsw, nr_pages);
2970 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2972 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2974 * Any of the following ensures page's memcg stability:
2978 * - lock_page_memcg()
2979 * - exclusive reference
2981 page->mem_cgroup = memcg;
2984 #ifdef CONFIG_MEMCG_KMEM
2985 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2988 unsigned int objects = objs_per_slab_page(s, page);
2991 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2996 if (cmpxchg(&page->obj_cgroups, NULL,
2997 (struct obj_cgroup **) ((unsigned long)vec | 0x1UL)))
3000 kmemleak_not_leak(vec);
3006 * Returns a pointer to the memory cgroup to which the kernel object is charged.
3008 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
3009 * cgroup_mutex, etc.
3011 struct mem_cgroup *mem_cgroup_from_obj(void *p)
3015 if (mem_cgroup_disabled())
3018 page = virt_to_head_page(p);
3021 * If page->mem_cgroup is set, it's either a simple mem_cgroup pointer
3022 * or a pointer to obj_cgroup vector. In the latter case the lowest
3023 * bit of the pointer is set.
3024 * The page->mem_cgroup pointer can be asynchronously changed
3025 * from NULL to (obj_cgroup_vec | 0x1UL), but can't be changed
3026 * from a valid memcg pointer to objcg vector or back.
3028 if (!page->mem_cgroup)
3032 * Slab objects are accounted individually, not per-page.
3033 * Memcg membership data for each individual object is saved in
3034 * the page->obj_cgroups.
3036 if (page_has_obj_cgroups(page)) {
3037 struct obj_cgroup *objcg;
3040 off = obj_to_index(page->slab_cache, page, p);
3041 objcg = page_obj_cgroups(page)[off];
3043 return obj_cgroup_memcg(objcg);
3048 /* All other pages use page->mem_cgroup */
3049 return page->mem_cgroup;
3052 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
3054 struct obj_cgroup *objcg = NULL;
3055 struct mem_cgroup *memcg;
3057 if (memcg_kmem_bypass())
3061 if (unlikely(active_memcg()))
3062 memcg = active_memcg();
3064 memcg = mem_cgroup_from_task(current);
3066 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
3067 objcg = rcu_dereference(memcg->objcg);
3068 if (objcg && obj_cgroup_tryget(objcg))
3077 static int memcg_alloc_cache_id(void)
3082 id = ida_simple_get(&memcg_cache_ida,
3083 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3087 if (id < memcg_nr_cache_ids)
3091 * There's no space for the new id in memcg_caches arrays,
3092 * so we have to grow them.
3094 down_write(&memcg_cache_ids_sem);
3096 size = 2 * (id + 1);
3097 if (size < MEMCG_CACHES_MIN_SIZE)
3098 size = MEMCG_CACHES_MIN_SIZE;
3099 else if (size > MEMCG_CACHES_MAX_SIZE)
3100 size = MEMCG_CACHES_MAX_SIZE;
3102 err = memcg_update_all_list_lrus(size);
3104 memcg_nr_cache_ids = size;
3106 up_write(&memcg_cache_ids_sem);
3109 ida_simple_remove(&memcg_cache_ida, id);
3115 static void memcg_free_cache_id(int id)
3117 ida_simple_remove(&memcg_cache_ida, id);
3121 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
3122 * @memcg: memory cgroup to charge
3123 * @gfp: reclaim mode
3124 * @nr_pages: number of pages to charge
3126 * Returns 0 on success, an error code on failure.
3128 int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
3129 unsigned int nr_pages)
3131 struct page_counter *counter;
3134 ret = try_charge(memcg, gfp, nr_pages);
3138 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3139 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3142 * Enforce __GFP_NOFAIL allocation because callers are not
3143 * prepared to see failures and likely do not have any failure
3146 if (gfp & __GFP_NOFAIL) {
3147 page_counter_charge(&memcg->kmem, nr_pages);
3150 cancel_charge(memcg, nr_pages);
3157 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
3158 * @memcg: memcg to uncharge
3159 * @nr_pages: number of pages to uncharge
3161 void __memcg_kmem_uncharge(struct mem_cgroup *memcg, unsigned int nr_pages)
3163 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3164 page_counter_uncharge(&memcg->kmem, nr_pages);
3166 page_counter_uncharge(&memcg->memory, nr_pages);
3167 if (do_memsw_account())
3168 page_counter_uncharge(&memcg->memsw, nr_pages);
3172 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3173 * @page: page to charge
3174 * @gfp: reclaim mode
3175 * @order: allocation order
3177 * Returns 0 on success, an error code on failure.
3179 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3181 struct mem_cgroup *memcg;
3184 memcg = get_mem_cgroup_from_current();
3185 if (memcg && !mem_cgroup_is_root(memcg)) {
3186 ret = __memcg_kmem_charge(memcg, gfp, 1 << order);
3188 page->mem_cgroup = memcg;
3189 __SetPageKmemcg(page);
3192 css_put(&memcg->css);
3198 * __memcg_kmem_uncharge_page: uncharge a kmem page
3199 * @page: page to uncharge
3200 * @order: allocation order
3202 void __memcg_kmem_uncharge_page(struct page *page, int order)
3204 struct mem_cgroup *memcg = page->mem_cgroup;
3205 unsigned int nr_pages = 1 << order;
3210 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3211 __memcg_kmem_uncharge(memcg, nr_pages);
3212 page->mem_cgroup = NULL;
3213 css_put(&memcg->css);
3215 /* slab pages do not have PageKmemcg flag set */
3216 if (PageKmemcg(page))
3217 __ClearPageKmemcg(page);
3220 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3222 struct memcg_stock_pcp *stock;
3223 unsigned long flags;
3226 local_irq_save(flags);
3228 stock = this_cpu_ptr(&memcg_stock);
3229 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3230 stock->nr_bytes -= nr_bytes;
3234 local_irq_restore(flags);
3239 static void drain_obj_stock(struct memcg_stock_pcp *stock)
3241 struct obj_cgroup *old = stock->cached_objcg;
3246 if (stock->nr_bytes) {
3247 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3248 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3252 __memcg_kmem_uncharge(obj_cgroup_memcg(old), nr_pages);
3257 * The leftover is flushed to the centralized per-memcg value.
3258 * On the next attempt to refill obj stock it will be moved
3259 * to a per-cpu stock (probably, on an other CPU), see
3260 * refill_obj_stock().
3262 * How often it's flushed is a trade-off between the memory
3263 * limit enforcement accuracy and potential CPU contention,
3264 * so it might be changed in the future.
3266 atomic_add(nr_bytes, &old->nr_charged_bytes);
3267 stock->nr_bytes = 0;
3270 obj_cgroup_put(old);
3271 stock->cached_objcg = NULL;
3274 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3275 struct mem_cgroup *root_memcg)
3277 struct mem_cgroup *memcg;
3279 if (stock->cached_objcg) {
3280 memcg = obj_cgroup_memcg(stock->cached_objcg);
3281 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3288 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3290 struct memcg_stock_pcp *stock;
3291 unsigned long flags;
3293 local_irq_save(flags);
3295 stock = this_cpu_ptr(&memcg_stock);
3296 if (stock->cached_objcg != objcg) { /* reset if necessary */
3297 drain_obj_stock(stock);
3298 obj_cgroup_get(objcg);
3299 stock->cached_objcg = objcg;
3300 stock->nr_bytes = atomic_xchg(&objcg->nr_charged_bytes, 0);
3302 stock->nr_bytes += nr_bytes;
3304 if (stock->nr_bytes > PAGE_SIZE)
3305 drain_obj_stock(stock);
3307 local_irq_restore(flags);
3310 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3312 struct mem_cgroup *memcg;
3313 unsigned int nr_pages, nr_bytes;
3316 if (consume_obj_stock(objcg, size))
3320 * In theory, memcg->nr_charged_bytes can have enough
3321 * pre-charged bytes to satisfy the allocation. However,
3322 * flushing memcg->nr_charged_bytes requires two atomic
3323 * operations, and memcg->nr_charged_bytes can't be big,
3324 * so it's better to ignore it and try grab some new pages.
3325 * memcg->nr_charged_bytes will be flushed in
3326 * refill_obj_stock(), called from this function or
3327 * independently later.
3331 memcg = obj_cgroup_memcg(objcg);
3332 if (unlikely(!css_tryget(&memcg->css)))
3336 nr_pages = size >> PAGE_SHIFT;
3337 nr_bytes = size & (PAGE_SIZE - 1);
3342 ret = __memcg_kmem_charge(memcg, gfp, nr_pages);
3343 if (!ret && nr_bytes)
3344 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes);
3346 css_put(&memcg->css);
3350 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3352 refill_obj_stock(objcg, size);
3355 #endif /* CONFIG_MEMCG_KMEM */
3357 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3359 * Because page_memcg(head) is not set on compound tails, set it now.
3361 void mem_cgroup_split_huge_fixup(struct page *head)
3363 struct mem_cgroup *memcg = head->mem_cgroup;
3366 if (mem_cgroup_disabled())
3369 for (i = 1; i < HPAGE_PMD_NR; i++) {
3370 css_get(&memcg->css);
3371 head[i].mem_cgroup = memcg;
3374 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3376 #ifdef CONFIG_MEMCG_SWAP
3378 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3379 * @entry: swap entry to be moved
3380 * @from: mem_cgroup which the entry is moved from
3381 * @to: mem_cgroup which the entry is moved to
3383 * It succeeds only when the swap_cgroup's record for this entry is the same
3384 * as the mem_cgroup's id of @from.
3386 * Returns 0 on success, -EINVAL on failure.
3388 * The caller must have charged to @to, IOW, called page_counter_charge() about
3389 * both res and memsw, and called css_get().
3391 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3392 struct mem_cgroup *from, struct mem_cgroup *to)
3394 unsigned short old_id, new_id;
3396 old_id = mem_cgroup_id(from);
3397 new_id = mem_cgroup_id(to);
3399 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3400 mod_memcg_state(from, MEMCG_SWAP, -1);
3401 mod_memcg_state(to, MEMCG_SWAP, 1);
3407 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3408 struct mem_cgroup *from, struct mem_cgroup *to)
3414 static DEFINE_MUTEX(memcg_max_mutex);
3416 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3417 unsigned long max, bool memsw)
3419 bool enlarge = false;
3420 bool drained = false;
3422 bool limits_invariant;
3423 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3426 if (signal_pending(current)) {
3431 mutex_lock(&memcg_max_mutex);
3433 * Make sure that the new limit (memsw or memory limit) doesn't
3434 * break our basic invariant rule memory.max <= memsw.max.
3436 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3437 max <= memcg->memsw.max;
3438 if (!limits_invariant) {
3439 mutex_unlock(&memcg_max_mutex);
3443 if (max > counter->max)
3445 ret = page_counter_set_max(counter, max);
3446 mutex_unlock(&memcg_max_mutex);
3452 drain_all_stock(memcg);
3457 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3458 GFP_KERNEL, !memsw)) {
3464 if (!ret && enlarge)
3465 memcg_oom_recover(memcg);
3470 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3472 unsigned long *total_scanned)
3474 unsigned long nr_reclaimed = 0;
3475 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3476 unsigned long reclaimed;
3478 struct mem_cgroup_tree_per_node *mctz;
3479 unsigned long excess;
3480 unsigned long nr_scanned;
3485 mctz = soft_limit_tree_node(pgdat->node_id);
3488 * Do not even bother to check the largest node if the root
3489 * is empty. Do it lockless to prevent lock bouncing. Races
3490 * are acceptable as soft limit is best effort anyway.
3492 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3496 * This loop can run a while, specially if mem_cgroup's continuously
3497 * keep exceeding their soft limit and putting the system under
3504 mz = mem_cgroup_largest_soft_limit_node(mctz);
3509 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3510 gfp_mask, &nr_scanned);
3511 nr_reclaimed += reclaimed;
3512 *total_scanned += nr_scanned;
3513 spin_lock_irq(&mctz->lock);
3514 __mem_cgroup_remove_exceeded(mz, mctz);
3517 * If we failed to reclaim anything from this memory cgroup
3518 * it is time to move on to the next cgroup
3522 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3524 excess = soft_limit_excess(mz->memcg);
3526 * One school of thought says that we should not add
3527 * back the node to the tree if reclaim returns 0.
3528 * But our reclaim could return 0, simply because due
3529 * to priority we are exposing a smaller subset of
3530 * memory to reclaim from. Consider this as a longer
3533 /* If excess == 0, no tree ops */
3534 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3535 spin_unlock_irq(&mctz->lock);
3536 css_put(&mz->memcg->css);
3539 * Could not reclaim anything and there are no more
3540 * mem cgroups to try or we seem to be looping without
3541 * reclaiming anything.
3543 if (!nr_reclaimed &&
3545 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3547 } while (!nr_reclaimed);
3549 css_put(&next_mz->memcg->css);
3550 return nr_reclaimed;
3554 * Reclaims as many pages from the given memcg as possible.
3556 * Caller is responsible for holding css reference for memcg.
3558 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3560 int nr_retries = MAX_RECLAIM_RETRIES;
3562 /* we call try-to-free pages for make this cgroup empty */
3563 lru_add_drain_all();
3565 drain_all_stock(memcg);
3567 /* try to free all pages in this cgroup */
3568 while (nr_retries && page_counter_read(&memcg->memory)) {
3571 if (signal_pending(current))
3574 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3578 /* maybe some writeback is necessary */
3579 congestion_wait(BLK_RW_ASYNC, HZ/10);
3587 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3588 char *buf, size_t nbytes,
3591 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3593 if (mem_cgroup_is_root(memcg))
3595 return mem_cgroup_force_empty(memcg) ?: nbytes;
3598 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3604 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3605 struct cftype *cft, u64 val)
3610 pr_warn_once("Non-hierarchical mode is deprecated. "
3611 "Please report your usecase to linux-mm@kvack.org if you "
3612 "depend on this functionality.\n");
3617 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3621 if (mem_cgroup_is_root(memcg)) {
3622 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3623 memcg_page_state(memcg, NR_ANON_MAPPED);
3625 val += memcg_page_state(memcg, MEMCG_SWAP);
3628 val = page_counter_read(&memcg->memory);
3630 val = page_counter_read(&memcg->memsw);
3643 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3646 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3647 struct page_counter *counter;
3649 switch (MEMFILE_TYPE(cft->private)) {
3651 counter = &memcg->memory;
3654 counter = &memcg->memsw;
3657 counter = &memcg->kmem;
3660 counter = &memcg->tcpmem;
3666 switch (MEMFILE_ATTR(cft->private)) {
3668 if (counter == &memcg->memory)
3669 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3670 if (counter == &memcg->memsw)
3671 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3672 return (u64)page_counter_read(counter) * PAGE_SIZE;
3674 return (u64)counter->max * PAGE_SIZE;
3676 return (u64)counter->watermark * PAGE_SIZE;
3678 return counter->failcnt;
3679 case RES_SOFT_LIMIT:
3680 return (u64)memcg->soft_limit * PAGE_SIZE;
3686 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3688 unsigned long stat[MEMCG_NR_STAT] = {0};
3689 struct mem_cgroup *mi;
3692 for_each_online_cpu(cpu)
3693 for (i = 0; i < MEMCG_NR_STAT; i++)
3694 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3696 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3697 for (i = 0; i < MEMCG_NR_STAT; i++)
3698 atomic_long_add(stat[i], &mi->vmstats[i]);
3700 for_each_node(node) {
3701 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3702 struct mem_cgroup_per_node *pi;
3704 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3707 for_each_online_cpu(cpu)
3708 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3710 pn->lruvec_stat_cpu->count[i], cpu);
3712 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3713 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3714 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3718 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3720 unsigned long events[NR_VM_EVENT_ITEMS];
3721 struct mem_cgroup *mi;
3724 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3727 for_each_online_cpu(cpu)
3728 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3729 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3732 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3733 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3734 atomic_long_add(events[i], &mi->vmevents[i]);
3737 #ifdef CONFIG_MEMCG_KMEM
3738 static int memcg_online_kmem(struct mem_cgroup *memcg)
3740 struct obj_cgroup *objcg;
3743 if (cgroup_memory_nokmem)
3746 BUG_ON(memcg->kmemcg_id >= 0);
3747 BUG_ON(memcg->kmem_state);
3749 memcg_id = memcg_alloc_cache_id();
3753 objcg = obj_cgroup_alloc();
3755 memcg_free_cache_id(memcg_id);
3758 objcg->memcg = memcg;
3759 rcu_assign_pointer(memcg->objcg, objcg);
3761 static_branch_enable(&memcg_kmem_enabled_key);
3763 memcg->kmemcg_id = memcg_id;
3764 memcg->kmem_state = KMEM_ONLINE;
3769 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3771 struct cgroup_subsys_state *css;
3772 struct mem_cgroup *parent, *child;
3775 if (memcg->kmem_state != KMEM_ONLINE)
3778 memcg->kmem_state = KMEM_ALLOCATED;
3780 parent = parent_mem_cgroup(memcg);
3782 parent = root_mem_cgroup;
3784 memcg_reparent_objcgs(memcg, parent);
3786 kmemcg_id = memcg->kmemcg_id;
3787 BUG_ON(kmemcg_id < 0);
3790 * Change kmemcg_id of this cgroup and all its descendants to the
3791 * parent's id, and then move all entries from this cgroup's list_lrus
3792 * to ones of the parent. After we have finished, all list_lrus
3793 * corresponding to this cgroup are guaranteed to remain empty. The
3794 * ordering is imposed by list_lru_node->lock taken by
3795 * memcg_drain_all_list_lrus().
3797 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3798 css_for_each_descendant_pre(css, &memcg->css) {
3799 child = mem_cgroup_from_css(css);
3800 BUG_ON(child->kmemcg_id != kmemcg_id);
3801 child->kmemcg_id = parent->kmemcg_id;
3805 memcg_drain_all_list_lrus(kmemcg_id, parent);
3807 memcg_free_cache_id(kmemcg_id);
3810 static void memcg_free_kmem(struct mem_cgroup *memcg)
3812 /* css_alloc() failed, offlining didn't happen */
3813 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3814 memcg_offline_kmem(memcg);
3817 static int memcg_online_kmem(struct mem_cgroup *memcg)
3821 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3824 static void memcg_free_kmem(struct mem_cgroup *memcg)
3827 #endif /* CONFIG_MEMCG_KMEM */
3829 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3834 mutex_lock(&memcg_max_mutex);
3835 ret = page_counter_set_max(&memcg->kmem, max);
3836 mutex_unlock(&memcg_max_mutex);
3840 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3844 mutex_lock(&memcg_max_mutex);
3846 ret = page_counter_set_max(&memcg->tcpmem, max);
3850 if (!memcg->tcpmem_active) {
3852 * The active flag needs to be written after the static_key
3853 * update. This is what guarantees that the socket activation
3854 * function is the last one to run. See mem_cgroup_sk_alloc()
3855 * for details, and note that we don't mark any socket as
3856 * belonging to this memcg until that flag is up.
3858 * We need to do this, because static_keys will span multiple
3859 * sites, but we can't control their order. If we mark a socket
3860 * as accounted, but the accounting functions are not patched in
3861 * yet, we'll lose accounting.
3863 * We never race with the readers in mem_cgroup_sk_alloc(),
3864 * because when this value change, the code to process it is not
3867 static_branch_inc(&memcg_sockets_enabled_key);
3868 memcg->tcpmem_active = true;
3871 mutex_unlock(&memcg_max_mutex);
3876 * The user of this function is...
3879 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3880 char *buf, size_t nbytes, loff_t off)
3882 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3883 unsigned long nr_pages;
3886 buf = strstrip(buf);
3887 ret = page_counter_memparse(buf, "-1", &nr_pages);
3891 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3893 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3897 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3899 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3902 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3905 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3906 "Please report your usecase to linux-mm@kvack.org if you "
3907 "depend on this functionality.\n");
3908 ret = memcg_update_kmem_max(memcg, nr_pages);
3911 ret = memcg_update_tcp_max(memcg, nr_pages);
3915 case RES_SOFT_LIMIT:
3916 memcg->soft_limit = nr_pages;
3920 return ret ?: nbytes;
3923 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3924 size_t nbytes, loff_t off)
3926 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3927 struct page_counter *counter;
3929 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3931 counter = &memcg->memory;
3934 counter = &memcg->memsw;
3937 counter = &memcg->kmem;
3940 counter = &memcg->tcpmem;
3946 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3948 page_counter_reset_watermark(counter);
3951 counter->failcnt = 0;
3960 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3963 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3967 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3968 struct cftype *cft, u64 val)
3970 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3972 if (val & ~MOVE_MASK)
3976 * No kind of locking is needed in here, because ->can_attach() will
3977 * check this value once in the beginning of the process, and then carry
3978 * on with stale data. This means that changes to this value will only
3979 * affect task migrations starting after the change.
3981 memcg->move_charge_at_immigrate = val;
3985 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3986 struct cftype *cft, u64 val)
3994 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3995 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3996 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3998 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3999 int nid, unsigned int lru_mask, bool tree)
4001 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
4002 unsigned long nr = 0;
4005 VM_BUG_ON((unsigned)nid >= nr_node_ids);
4008 if (!(BIT(lru) & lru_mask))
4011 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
4013 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
4018 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
4019 unsigned int lru_mask,
4022 unsigned long nr = 0;
4026 if (!(BIT(lru) & lru_mask))
4029 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
4031 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
4036 static int memcg_numa_stat_show(struct seq_file *m, void *v)
4040 unsigned int lru_mask;
4043 static const struct numa_stat stats[] = {
4044 { "total", LRU_ALL },
4045 { "file", LRU_ALL_FILE },
4046 { "anon", LRU_ALL_ANON },
4047 { "unevictable", BIT(LRU_UNEVICTABLE) },
4049 const struct numa_stat *stat;
4051 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4053 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4054 seq_printf(m, "%s=%lu", stat->name,
4055 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4057 for_each_node_state(nid, N_MEMORY)
4058 seq_printf(m, " N%d=%lu", nid,
4059 mem_cgroup_node_nr_lru_pages(memcg, nid,
4060 stat->lru_mask, false));
4064 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4066 seq_printf(m, "hierarchical_%s=%lu", stat->name,
4067 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4069 for_each_node_state(nid, N_MEMORY)
4070 seq_printf(m, " N%d=%lu", nid,
4071 mem_cgroup_node_nr_lru_pages(memcg, nid,
4072 stat->lru_mask, true));
4078 #endif /* CONFIG_NUMA */
4080 static const unsigned int memcg1_stats[] = {
4083 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4093 static const char *const memcg1_stat_names[] = {
4096 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4106 /* Universal VM events cgroup1 shows, original sort order */
4107 static const unsigned int memcg1_events[] = {
4114 static int memcg_stat_show(struct seq_file *m, void *v)
4116 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4117 unsigned long memory, memsw;
4118 struct mem_cgroup *mi;
4121 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4123 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4126 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4128 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4129 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4130 if (memcg1_stats[i] == NR_ANON_THPS)
4133 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4136 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4137 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4138 memcg_events_local(memcg, memcg1_events[i]));
4140 for (i = 0; i < NR_LRU_LISTS; i++)
4141 seq_printf(m, "%s %lu\n", lru_list_name(i),
4142 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4145 /* Hierarchical information */
4146 memory = memsw = PAGE_COUNTER_MAX;
4147 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4148 memory = min(memory, READ_ONCE(mi->memory.max));
4149 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4151 seq_printf(m, "hierarchical_memory_limit %llu\n",
4152 (u64)memory * PAGE_SIZE);
4153 if (do_memsw_account())
4154 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4155 (u64)memsw * PAGE_SIZE);
4157 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4160 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4162 nr = memcg_page_state(memcg, memcg1_stats[i]);
4163 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4164 if (memcg1_stats[i] == NR_ANON_THPS)
4167 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4168 (u64)nr * PAGE_SIZE);
4171 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4172 seq_printf(m, "total_%s %llu\n",
4173 vm_event_name(memcg1_events[i]),
4174 (u64)memcg_events(memcg, memcg1_events[i]));
4176 for (i = 0; i < NR_LRU_LISTS; i++)
4177 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4178 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4181 #ifdef CONFIG_DEBUG_VM
4184 struct mem_cgroup_per_node *mz;
4185 unsigned long anon_cost = 0;
4186 unsigned long file_cost = 0;
4188 for_each_online_pgdat(pgdat) {
4189 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
4191 anon_cost += mz->lruvec.anon_cost;
4192 file_cost += mz->lruvec.file_cost;
4194 seq_printf(m, "anon_cost %lu\n", anon_cost);
4195 seq_printf(m, "file_cost %lu\n", file_cost);
4202 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4205 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4207 return mem_cgroup_swappiness(memcg);
4210 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4211 struct cftype *cft, u64 val)
4213 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4219 memcg->swappiness = val;
4221 vm_swappiness = val;
4226 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4228 struct mem_cgroup_threshold_ary *t;
4229 unsigned long usage;
4234 t = rcu_dereference(memcg->thresholds.primary);
4236 t = rcu_dereference(memcg->memsw_thresholds.primary);
4241 usage = mem_cgroup_usage(memcg, swap);
4244 * current_threshold points to threshold just below or equal to usage.
4245 * If it's not true, a threshold was crossed after last
4246 * call of __mem_cgroup_threshold().
4248 i = t->current_threshold;
4251 * Iterate backward over array of thresholds starting from
4252 * current_threshold and check if a threshold is crossed.
4253 * If none of thresholds below usage is crossed, we read
4254 * only one element of the array here.
4256 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4257 eventfd_signal(t->entries[i].eventfd, 1);
4259 /* i = current_threshold + 1 */
4263 * Iterate forward over array of thresholds starting from
4264 * current_threshold+1 and check if a threshold is crossed.
4265 * If none of thresholds above usage is crossed, we read
4266 * only one element of the array here.
4268 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4269 eventfd_signal(t->entries[i].eventfd, 1);
4271 /* Update current_threshold */
4272 t->current_threshold = i - 1;
4277 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4280 __mem_cgroup_threshold(memcg, false);
4281 if (do_memsw_account())
4282 __mem_cgroup_threshold(memcg, true);
4284 memcg = parent_mem_cgroup(memcg);
4288 static int compare_thresholds(const void *a, const void *b)
4290 const struct mem_cgroup_threshold *_a = a;
4291 const struct mem_cgroup_threshold *_b = b;
4293 if (_a->threshold > _b->threshold)
4296 if (_a->threshold < _b->threshold)
4302 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4304 struct mem_cgroup_eventfd_list *ev;
4306 spin_lock(&memcg_oom_lock);
4308 list_for_each_entry(ev, &memcg->oom_notify, list)
4309 eventfd_signal(ev->eventfd, 1);
4311 spin_unlock(&memcg_oom_lock);
4315 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4317 struct mem_cgroup *iter;
4319 for_each_mem_cgroup_tree(iter, memcg)
4320 mem_cgroup_oom_notify_cb(iter);
4323 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4324 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4326 struct mem_cgroup_thresholds *thresholds;
4327 struct mem_cgroup_threshold_ary *new;
4328 unsigned long threshold;
4329 unsigned long usage;
4332 ret = page_counter_memparse(args, "-1", &threshold);
4336 mutex_lock(&memcg->thresholds_lock);
4339 thresholds = &memcg->thresholds;
4340 usage = mem_cgroup_usage(memcg, false);
4341 } else if (type == _MEMSWAP) {
4342 thresholds = &memcg->memsw_thresholds;
4343 usage = mem_cgroup_usage(memcg, true);
4347 /* Check if a threshold crossed before adding a new one */
4348 if (thresholds->primary)
4349 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4351 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4353 /* Allocate memory for new array of thresholds */
4354 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4361 /* Copy thresholds (if any) to new array */
4362 if (thresholds->primary)
4363 memcpy(new->entries, thresholds->primary->entries,
4364 flex_array_size(new, entries, size - 1));
4366 /* Add new threshold */
4367 new->entries[size - 1].eventfd = eventfd;
4368 new->entries[size - 1].threshold = threshold;
4370 /* Sort thresholds. Registering of new threshold isn't time-critical */
4371 sort(new->entries, size, sizeof(*new->entries),
4372 compare_thresholds, NULL);
4374 /* Find current threshold */
4375 new->current_threshold = -1;
4376 for (i = 0; i < size; i++) {
4377 if (new->entries[i].threshold <= usage) {
4379 * new->current_threshold will not be used until
4380 * rcu_assign_pointer(), so it's safe to increment
4383 ++new->current_threshold;
4388 /* Free old spare buffer and save old primary buffer as spare */
4389 kfree(thresholds->spare);
4390 thresholds->spare = thresholds->primary;
4392 rcu_assign_pointer(thresholds->primary, new);
4394 /* To be sure that nobody uses thresholds */
4398 mutex_unlock(&memcg->thresholds_lock);
4403 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4404 struct eventfd_ctx *eventfd, const char *args)
4406 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4409 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4410 struct eventfd_ctx *eventfd, const char *args)
4412 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4415 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4416 struct eventfd_ctx *eventfd, enum res_type type)
4418 struct mem_cgroup_thresholds *thresholds;
4419 struct mem_cgroup_threshold_ary *new;
4420 unsigned long usage;
4421 int i, j, size, entries;
4423 mutex_lock(&memcg->thresholds_lock);
4426 thresholds = &memcg->thresholds;
4427 usage = mem_cgroup_usage(memcg, false);
4428 } else if (type == _MEMSWAP) {
4429 thresholds = &memcg->memsw_thresholds;
4430 usage = mem_cgroup_usage(memcg, true);
4434 if (!thresholds->primary)
4437 /* Check if a threshold crossed before removing */
4438 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4440 /* Calculate new number of threshold */
4442 for (i = 0; i < thresholds->primary->size; i++) {
4443 if (thresholds->primary->entries[i].eventfd != eventfd)
4449 new = thresholds->spare;
4451 /* If no items related to eventfd have been cleared, nothing to do */
4455 /* Set thresholds array to NULL if we don't have thresholds */
4464 /* Copy thresholds and find current threshold */
4465 new->current_threshold = -1;
4466 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4467 if (thresholds->primary->entries[i].eventfd == eventfd)
4470 new->entries[j] = thresholds->primary->entries[i];
4471 if (new->entries[j].threshold <= usage) {
4473 * new->current_threshold will not be used
4474 * until rcu_assign_pointer(), so it's safe to increment
4477 ++new->current_threshold;
4483 /* Swap primary and spare array */
4484 thresholds->spare = thresholds->primary;
4486 rcu_assign_pointer(thresholds->primary, new);
4488 /* To be sure that nobody uses thresholds */
4491 /* If all events are unregistered, free the spare array */
4493 kfree(thresholds->spare);
4494 thresholds->spare = NULL;
4497 mutex_unlock(&memcg->thresholds_lock);
4500 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4501 struct eventfd_ctx *eventfd)
4503 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4506 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4507 struct eventfd_ctx *eventfd)
4509 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4512 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4513 struct eventfd_ctx *eventfd, const char *args)
4515 struct mem_cgroup_eventfd_list *event;
4517 event = kmalloc(sizeof(*event), GFP_KERNEL);
4521 spin_lock(&memcg_oom_lock);
4523 event->eventfd = eventfd;
4524 list_add(&event->list, &memcg->oom_notify);
4526 /* already in OOM ? */
4527 if (memcg->under_oom)
4528 eventfd_signal(eventfd, 1);
4529 spin_unlock(&memcg_oom_lock);
4534 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4535 struct eventfd_ctx *eventfd)
4537 struct mem_cgroup_eventfd_list *ev, *tmp;
4539 spin_lock(&memcg_oom_lock);
4541 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4542 if (ev->eventfd == eventfd) {
4543 list_del(&ev->list);
4548 spin_unlock(&memcg_oom_lock);
4551 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4553 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4555 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4556 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4557 seq_printf(sf, "oom_kill %lu\n",
4558 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4562 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4563 struct cftype *cft, u64 val)
4565 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4567 /* cannot set to root cgroup and only 0 and 1 are allowed */
4568 if (!css->parent || !((val == 0) || (val == 1)))
4571 memcg->oom_kill_disable = val;
4573 memcg_oom_recover(memcg);
4578 #ifdef CONFIG_CGROUP_WRITEBACK
4580 #include <trace/events/writeback.h>
4582 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4584 return wb_domain_init(&memcg->cgwb_domain, gfp);
4587 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4589 wb_domain_exit(&memcg->cgwb_domain);
4592 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4594 wb_domain_size_changed(&memcg->cgwb_domain);
4597 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4599 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4601 if (!memcg->css.parent)
4604 return &memcg->cgwb_domain;
4608 * idx can be of type enum memcg_stat_item or node_stat_item.
4609 * Keep in sync with memcg_exact_page().
4611 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4613 long x = atomic_long_read(&memcg->vmstats[idx]);
4616 for_each_online_cpu(cpu)
4617 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4624 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4625 * @wb: bdi_writeback in question
4626 * @pfilepages: out parameter for number of file pages
4627 * @pheadroom: out parameter for number of allocatable pages according to memcg
4628 * @pdirty: out parameter for number of dirty pages
4629 * @pwriteback: out parameter for number of pages under writeback
4631 * Determine the numbers of file, headroom, dirty, and writeback pages in
4632 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4633 * is a bit more involved.
4635 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4636 * headroom is calculated as the lowest headroom of itself and the
4637 * ancestors. Note that this doesn't consider the actual amount of
4638 * available memory in the system. The caller should further cap
4639 * *@pheadroom accordingly.
4641 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4642 unsigned long *pheadroom, unsigned long *pdirty,
4643 unsigned long *pwriteback)
4645 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4646 struct mem_cgroup *parent;
4648 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4650 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4651 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4652 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4653 *pheadroom = PAGE_COUNTER_MAX;
4655 while ((parent = parent_mem_cgroup(memcg))) {
4656 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4657 READ_ONCE(memcg->memory.high));
4658 unsigned long used = page_counter_read(&memcg->memory);
4660 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4666 * Foreign dirty flushing
4668 * There's an inherent mismatch between memcg and writeback. The former
4669 * trackes ownership per-page while the latter per-inode. This was a
4670 * deliberate design decision because honoring per-page ownership in the
4671 * writeback path is complicated, may lead to higher CPU and IO overheads
4672 * and deemed unnecessary given that write-sharing an inode across
4673 * different cgroups isn't a common use-case.
4675 * Combined with inode majority-writer ownership switching, this works well
4676 * enough in most cases but there are some pathological cases. For
4677 * example, let's say there are two cgroups A and B which keep writing to
4678 * different but confined parts of the same inode. B owns the inode and
4679 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4680 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4681 * triggering background writeback. A will be slowed down without a way to
4682 * make writeback of the dirty pages happen.
4684 * Conditions like the above can lead to a cgroup getting repatedly and
4685 * severely throttled after making some progress after each
4686 * dirty_expire_interval while the underyling IO device is almost
4689 * Solving this problem completely requires matching the ownership tracking
4690 * granularities between memcg and writeback in either direction. However,
4691 * the more egregious behaviors can be avoided by simply remembering the
4692 * most recent foreign dirtying events and initiating remote flushes on
4693 * them when local writeback isn't enough to keep the memory clean enough.
4695 * The following two functions implement such mechanism. When a foreign
4696 * page - a page whose memcg and writeback ownerships don't match - is
4697 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4698 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4699 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4700 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4701 * foreign bdi_writebacks which haven't expired. Both the numbers of
4702 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4703 * limited to MEMCG_CGWB_FRN_CNT.
4705 * The mechanism only remembers IDs and doesn't hold any object references.
4706 * As being wrong occasionally doesn't matter, updates and accesses to the
4707 * records are lockless and racy.
4709 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4710 struct bdi_writeback *wb)
4712 struct mem_cgroup *memcg = page->mem_cgroup;
4713 struct memcg_cgwb_frn *frn;
4714 u64 now = get_jiffies_64();
4715 u64 oldest_at = now;
4719 trace_track_foreign_dirty(page, wb);
4722 * Pick the slot to use. If there is already a slot for @wb, keep
4723 * using it. If not replace the oldest one which isn't being
4726 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4727 frn = &memcg->cgwb_frn[i];
4728 if (frn->bdi_id == wb->bdi->id &&
4729 frn->memcg_id == wb->memcg_css->id)
4731 if (time_before64(frn->at, oldest_at) &&
4732 atomic_read(&frn->done.cnt) == 1) {
4734 oldest_at = frn->at;
4738 if (i < MEMCG_CGWB_FRN_CNT) {
4740 * Re-using an existing one. Update timestamp lazily to
4741 * avoid making the cacheline hot. We want them to be
4742 * reasonably up-to-date and significantly shorter than
4743 * dirty_expire_interval as that's what expires the record.
4744 * Use the shorter of 1s and dirty_expire_interval / 8.
4746 unsigned long update_intv =
4747 min_t(unsigned long, HZ,
4748 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4750 if (time_before64(frn->at, now - update_intv))
4752 } else if (oldest >= 0) {
4753 /* replace the oldest free one */
4754 frn = &memcg->cgwb_frn[oldest];
4755 frn->bdi_id = wb->bdi->id;
4756 frn->memcg_id = wb->memcg_css->id;
4761 /* issue foreign writeback flushes for recorded foreign dirtying events */
4762 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4764 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4765 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4766 u64 now = jiffies_64;
4769 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4770 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4773 * If the record is older than dirty_expire_interval,
4774 * writeback on it has already started. No need to kick it
4775 * off again. Also, don't start a new one if there's
4776 * already one in flight.
4778 if (time_after64(frn->at, now - intv) &&
4779 atomic_read(&frn->done.cnt) == 1) {
4781 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4782 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4783 WB_REASON_FOREIGN_FLUSH,
4789 #else /* CONFIG_CGROUP_WRITEBACK */
4791 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4796 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4800 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4804 #endif /* CONFIG_CGROUP_WRITEBACK */
4807 * DO NOT USE IN NEW FILES.
4809 * "cgroup.event_control" implementation.
4811 * This is way over-engineered. It tries to support fully configurable
4812 * events for each user. Such level of flexibility is completely
4813 * unnecessary especially in the light of the planned unified hierarchy.
4815 * Please deprecate this and replace with something simpler if at all
4820 * Unregister event and free resources.
4822 * Gets called from workqueue.
4824 static void memcg_event_remove(struct work_struct *work)
4826 struct mem_cgroup_event *event =
4827 container_of(work, struct mem_cgroup_event, remove);
4828 struct mem_cgroup *memcg = event->memcg;
4830 remove_wait_queue(event->wqh, &event->wait);
4832 event->unregister_event(memcg, event->eventfd);
4834 /* Notify userspace the event is going away. */
4835 eventfd_signal(event->eventfd, 1);
4837 eventfd_ctx_put(event->eventfd);
4839 css_put(&memcg->css);
4843 * Gets called on EPOLLHUP on eventfd when user closes it.
4845 * Called with wqh->lock held and interrupts disabled.
4847 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4848 int sync, void *key)
4850 struct mem_cgroup_event *event =
4851 container_of(wait, struct mem_cgroup_event, wait);
4852 struct mem_cgroup *memcg = event->memcg;
4853 __poll_t flags = key_to_poll(key);
4855 if (flags & EPOLLHUP) {
4857 * If the event has been detached at cgroup removal, we
4858 * can simply return knowing the other side will cleanup
4861 * We can't race against event freeing since the other
4862 * side will require wqh->lock via remove_wait_queue(),
4865 spin_lock(&memcg->event_list_lock);
4866 if (!list_empty(&event->list)) {
4867 list_del_init(&event->list);
4869 * We are in atomic context, but cgroup_event_remove()
4870 * may sleep, so we have to call it in workqueue.
4872 schedule_work(&event->remove);
4874 spin_unlock(&memcg->event_list_lock);
4880 static void memcg_event_ptable_queue_proc(struct file *file,
4881 wait_queue_head_t *wqh, poll_table *pt)
4883 struct mem_cgroup_event *event =
4884 container_of(pt, struct mem_cgroup_event, pt);
4887 add_wait_queue(wqh, &event->wait);
4891 * DO NOT USE IN NEW FILES.
4893 * Parse input and register new cgroup event handler.
4895 * Input must be in format '<event_fd> <control_fd> <args>'.
4896 * Interpretation of args is defined by control file implementation.
4898 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4899 char *buf, size_t nbytes, loff_t off)
4901 struct cgroup_subsys_state *css = of_css(of);
4902 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4903 struct mem_cgroup_event *event;
4904 struct cgroup_subsys_state *cfile_css;
4905 unsigned int efd, cfd;
4912 buf = strstrip(buf);
4914 efd = simple_strtoul(buf, &endp, 10);
4919 cfd = simple_strtoul(buf, &endp, 10);
4920 if ((*endp != ' ') && (*endp != '\0'))
4924 event = kzalloc(sizeof(*event), GFP_KERNEL);
4928 event->memcg = memcg;
4929 INIT_LIST_HEAD(&event->list);
4930 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4931 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4932 INIT_WORK(&event->remove, memcg_event_remove);
4940 event->eventfd = eventfd_ctx_fileget(efile.file);
4941 if (IS_ERR(event->eventfd)) {
4942 ret = PTR_ERR(event->eventfd);
4949 goto out_put_eventfd;
4952 /* the process need read permission on control file */
4953 /* AV: shouldn't we check that it's been opened for read instead? */
4954 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4959 * Determine the event callbacks and set them in @event. This used
4960 * to be done via struct cftype but cgroup core no longer knows
4961 * about these events. The following is crude but the whole thing
4962 * is for compatibility anyway.
4964 * DO NOT ADD NEW FILES.
4966 name = cfile.file->f_path.dentry->d_name.name;
4968 if (!strcmp(name, "memory.usage_in_bytes")) {
4969 event->register_event = mem_cgroup_usage_register_event;
4970 event->unregister_event = mem_cgroup_usage_unregister_event;
4971 } else if (!strcmp(name, "memory.oom_control")) {
4972 event->register_event = mem_cgroup_oom_register_event;
4973 event->unregister_event = mem_cgroup_oom_unregister_event;
4974 } else if (!strcmp(name, "memory.pressure_level")) {
4975 event->register_event = vmpressure_register_event;
4976 event->unregister_event = vmpressure_unregister_event;
4977 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4978 event->register_event = memsw_cgroup_usage_register_event;
4979 event->unregister_event = memsw_cgroup_usage_unregister_event;
4986 * Verify @cfile should belong to @css. Also, remaining events are
4987 * automatically removed on cgroup destruction but the removal is
4988 * asynchronous, so take an extra ref on @css.
4990 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4991 &memory_cgrp_subsys);
4993 if (IS_ERR(cfile_css))
4995 if (cfile_css != css) {
5000 ret = event->register_event(memcg, event->eventfd, buf);
5004 vfs_poll(efile.file, &event->pt);
5006 spin_lock(&memcg->event_list_lock);
5007 list_add(&event->list, &memcg->event_list);
5008 spin_unlock(&memcg->event_list_lock);
5020 eventfd_ctx_put(event->eventfd);
5029 static struct cftype mem_cgroup_legacy_files[] = {
5031 .name = "usage_in_bytes",
5032 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5033 .read_u64 = mem_cgroup_read_u64,
5036 .name = "max_usage_in_bytes",
5037 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5038 .write = mem_cgroup_reset,
5039 .read_u64 = mem_cgroup_read_u64,
5042 .name = "limit_in_bytes",
5043 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5044 .write = mem_cgroup_write,
5045 .read_u64 = mem_cgroup_read_u64,
5048 .name = "soft_limit_in_bytes",
5049 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5050 .write = mem_cgroup_write,
5051 .read_u64 = mem_cgroup_read_u64,
5055 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5056 .write = mem_cgroup_reset,
5057 .read_u64 = mem_cgroup_read_u64,
5061 .seq_show = memcg_stat_show,
5064 .name = "force_empty",
5065 .write = mem_cgroup_force_empty_write,
5068 .name = "use_hierarchy",
5069 .write_u64 = mem_cgroup_hierarchy_write,
5070 .read_u64 = mem_cgroup_hierarchy_read,
5073 .name = "cgroup.event_control", /* XXX: for compat */
5074 .write = memcg_write_event_control,
5075 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
5078 .name = "swappiness",
5079 .read_u64 = mem_cgroup_swappiness_read,
5080 .write_u64 = mem_cgroup_swappiness_write,
5083 .name = "move_charge_at_immigrate",
5084 .read_u64 = mem_cgroup_move_charge_read,
5085 .write_u64 = mem_cgroup_move_charge_write,
5088 .name = "oom_control",
5089 .seq_show = mem_cgroup_oom_control_read,
5090 .write_u64 = mem_cgroup_oom_control_write,
5091 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5094 .name = "pressure_level",
5098 .name = "numa_stat",
5099 .seq_show = memcg_numa_stat_show,
5103 .name = "kmem.limit_in_bytes",
5104 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5105 .write = mem_cgroup_write,
5106 .read_u64 = mem_cgroup_read_u64,
5109 .name = "kmem.usage_in_bytes",
5110 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5111 .read_u64 = mem_cgroup_read_u64,
5114 .name = "kmem.failcnt",
5115 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5116 .write = mem_cgroup_reset,
5117 .read_u64 = mem_cgroup_read_u64,
5120 .name = "kmem.max_usage_in_bytes",
5121 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5122 .write = mem_cgroup_reset,
5123 .read_u64 = mem_cgroup_read_u64,
5125 #if defined(CONFIG_MEMCG_KMEM) && \
5126 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5128 .name = "kmem.slabinfo",
5129 .seq_show = memcg_slab_show,
5133 .name = "kmem.tcp.limit_in_bytes",
5134 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5135 .write = mem_cgroup_write,
5136 .read_u64 = mem_cgroup_read_u64,
5139 .name = "kmem.tcp.usage_in_bytes",
5140 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5141 .read_u64 = mem_cgroup_read_u64,
5144 .name = "kmem.tcp.failcnt",
5145 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5146 .write = mem_cgroup_reset,
5147 .read_u64 = mem_cgroup_read_u64,
5150 .name = "kmem.tcp.max_usage_in_bytes",
5151 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5152 .write = mem_cgroup_reset,
5153 .read_u64 = mem_cgroup_read_u64,
5155 { }, /* terminate */
5159 * Private memory cgroup IDR
5161 * Swap-out records and page cache shadow entries need to store memcg
5162 * references in constrained space, so we maintain an ID space that is
5163 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5164 * memory-controlled cgroups to 64k.
5166 * However, there usually are many references to the offline CSS after
5167 * the cgroup has been destroyed, such as page cache or reclaimable
5168 * slab objects, that don't need to hang on to the ID. We want to keep
5169 * those dead CSS from occupying IDs, or we might quickly exhaust the
5170 * relatively small ID space and prevent the creation of new cgroups
5171 * even when there are much fewer than 64k cgroups - possibly none.
5173 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5174 * be freed and recycled when it's no longer needed, which is usually
5175 * when the CSS is offlined.
5177 * The only exception to that are records of swapped out tmpfs/shmem
5178 * pages that need to be attributed to live ancestors on swapin. But
5179 * those references are manageable from userspace.
5182 static DEFINE_IDR(mem_cgroup_idr);
5184 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5186 if (memcg->id.id > 0) {
5187 idr_remove(&mem_cgroup_idr, memcg->id.id);
5192 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5195 refcount_add(n, &memcg->id.ref);
5198 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5200 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5201 mem_cgroup_id_remove(memcg);
5203 /* Memcg ID pins CSS */
5204 css_put(&memcg->css);
5208 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5210 mem_cgroup_id_put_many(memcg, 1);
5214 * mem_cgroup_from_id - look up a memcg from a memcg id
5215 * @id: the memcg id to look up
5217 * Caller must hold rcu_read_lock().
5219 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5221 WARN_ON_ONCE(!rcu_read_lock_held());
5222 return idr_find(&mem_cgroup_idr, id);
5225 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5227 struct mem_cgroup_per_node *pn;
5230 * This routine is called against possible nodes.
5231 * But it's BUG to call kmalloc() against offline node.
5233 * TODO: this routine can waste much memory for nodes which will
5234 * never be onlined. It's better to use memory hotplug callback
5237 if (!node_state(node, N_NORMAL_MEMORY))
5239 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5243 pn->lruvec_stat_local = alloc_percpu_gfp(struct lruvec_stat,
5244 GFP_KERNEL_ACCOUNT);
5245 if (!pn->lruvec_stat_local) {
5250 pn->lruvec_stat_cpu = alloc_percpu_gfp(struct lruvec_stat,
5251 GFP_KERNEL_ACCOUNT);
5252 if (!pn->lruvec_stat_cpu) {
5253 free_percpu(pn->lruvec_stat_local);
5258 lruvec_init(&pn->lruvec);
5259 pn->usage_in_excess = 0;
5260 pn->on_tree = false;
5263 memcg->nodeinfo[node] = pn;
5267 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5269 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5274 free_percpu(pn->lruvec_stat_cpu);
5275 free_percpu(pn->lruvec_stat_local);
5279 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5284 free_mem_cgroup_per_node_info(memcg, node);
5285 free_percpu(memcg->vmstats_percpu);
5286 free_percpu(memcg->vmstats_local);
5290 static void mem_cgroup_free(struct mem_cgroup *memcg)
5292 memcg_wb_domain_exit(memcg);
5294 * Flush percpu vmstats and vmevents to guarantee the value correctness
5295 * on parent's and all ancestor levels.
5297 memcg_flush_percpu_vmstats(memcg);
5298 memcg_flush_percpu_vmevents(memcg);
5299 __mem_cgroup_free(memcg);
5302 static struct mem_cgroup *mem_cgroup_alloc(void)
5304 struct mem_cgroup *memcg;
5307 int __maybe_unused i;
5308 long error = -ENOMEM;
5310 size = sizeof(struct mem_cgroup);
5311 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5313 memcg = kzalloc(size, GFP_KERNEL);
5315 return ERR_PTR(error);
5317 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5318 1, MEM_CGROUP_ID_MAX,
5320 if (memcg->id.id < 0) {
5321 error = memcg->id.id;
5325 memcg->vmstats_local = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5326 GFP_KERNEL_ACCOUNT);
5327 if (!memcg->vmstats_local)
5330 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5331 GFP_KERNEL_ACCOUNT);
5332 if (!memcg->vmstats_percpu)
5336 if (alloc_mem_cgroup_per_node_info(memcg, node))
5339 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5342 INIT_WORK(&memcg->high_work, high_work_func);
5343 INIT_LIST_HEAD(&memcg->oom_notify);
5344 mutex_init(&memcg->thresholds_lock);
5345 spin_lock_init(&memcg->move_lock);
5346 vmpressure_init(&memcg->vmpressure);
5347 INIT_LIST_HEAD(&memcg->event_list);
5348 spin_lock_init(&memcg->event_list_lock);
5349 memcg->socket_pressure = jiffies;
5350 #ifdef CONFIG_MEMCG_KMEM
5351 memcg->kmemcg_id = -1;
5352 INIT_LIST_HEAD(&memcg->objcg_list);
5354 #ifdef CONFIG_CGROUP_WRITEBACK
5355 INIT_LIST_HEAD(&memcg->cgwb_list);
5356 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5357 memcg->cgwb_frn[i].done =
5358 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5360 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5361 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5362 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5363 memcg->deferred_split_queue.split_queue_len = 0;
5365 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5368 mem_cgroup_id_remove(memcg);
5369 __mem_cgroup_free(memcg);
5370 return ERR_PTR(error);
5373 static struct cgroup_subsys_state * __ref
5374 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5376 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5377 struct mem_cgroup *memcg, *old_memcg;
5378 long error = -ENOMEM;
5380 old_memcg = set_active_memcg(parent);
5381 memcg = mem_cgroup_alloc();
5382 set_active_memcg(old_memcg);
5384 return ERR_CAST(memcg);
5386 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5387 memcg->soft_limit = PAGE_COUNTER_MAX;
5388 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5390 memcg->swappiness = mem_cgroup_swappiness(parent);
5391 memcg->oom_kill_disable = parent->oom_kill_disable;
5393 page_counter_init(&memcg->memory, &parent->memory);
5394 page_counter_init(&memcg->swap, &parent->swap);
5395 page_counter_init(&memcg->kmem, &parent->kmem);
5396 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5398 page_counter_init(&memcg->memory, NULL);
5399 page_counter_init(&memcg->swap, NULL);
5400 page_counter_init(&memcg->kmem, NULL);
5401 page_counter_init(&memcg->tcpmem, NULL);
5403 root_mem_cgroup = memcg;
5407 /* The following stuff does not apply to the root */
5408 error = memcg_online_kmem(memcg);
5412 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5413 static_branch_inc(&memcg_sockets_enabled_key);
5417 mem_cgroup_id_remove(memcg);
5418 mem_cgroup_free(memcg);
5419 return ERR_PTR(error);
5422 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5424 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5427 * A memcg must be visible for memcg_expand_shrinker_maps()
5428 * by the time the maps are allocated. So, we allocate maps
5429 * here, when for_each_mem_cgroup() can't skip it.
5431 if (memcg_alloc_shrinker_maps(memcg)) {
5432 mem_cgroup_id_remove(memcg);
5436 /* Online state pins memcg ID, memcg ID pins CSS */
5437 refcount_set(&memcg->id.ref, 1);
5442 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5444 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5445 struct mem_cgroup_event *event, *tmp;
5448 * Unregister events and notify userspace.
5449 * Notify userspace about cgroup removing only after rmdir of cgroup
5450 * directory to avoid race between userspace and kernelspace.
5452 spin_lock(&memcg->event_list_lock);
5453 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5454 list_del_init(&event->list);
5455 schedule_work(&event->remove);
5457 spin_unlock(&memcg->event_list_lock);
5459 page_counter_set_min(&memcg->memory, 0);
5460 page_counter_set_low(&memcg->memory, 0);
5462 memcg_offline_kmem(memcg);
5463 wb_memcg_offline(memcg);
5465 drain_all_stock(memcg);
5467 mem_cgroup_id_put(memcg);
5470 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5472 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5474 invalidate_reclaim_iterators(memcg);
5477 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5479 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5480 int __maybe_unused i;
5482 #ifdef CONFIG_CGROUP_WRITEBACK
5483 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5484 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5486 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5487 static_branch_dec(&memcg_sockets_enabled_key);
5489 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5490 static_branch_dec(&memcg_sockets_enabled_key);
5492 vmpressure_cleanup(&memcg->vmpressure);
5493 cancel_work_sync(&memcg->high_work);
5494 mem_cgroup_remove_from_trees(memcg);
5495 memcg_free_shrinker_maps(memcg);
5496 memcg_free_kmem(memcg);
5497 mem_cgroup_free(memcg);
5501 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5502 * @css: the target css
5504 * Reset the states of the mem_cgroup associated with @css. This is
5505 * invoked when the userland requests disabling on the default hierarchy
5506 * but the memcg is pinned through dependency. The memcg should stop
5507 * applying policies and should revert to the vanilla state as it may be
5508 * made visible again.
5510 * The current implementation only resets the essential configurations.
5511 * This needs to be expanded to cover all the visible parts.
5513 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5515 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5517 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5518 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5519 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5520 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5521 page_counter_set_min(&memcg->memory, 0);
5522 page_counter_set_low(&memcg->memory, 0);
5523 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5524 memcg->soft_limit = PAGE_COUNTER_MAX;
5525 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5526 memcg_wb_domain_size_changed(memcg);
5530 /* Handlers for move charge at task migration. */
5531 static int mem_cgroup_do_precharge(unsigned long count)
5535 /* Try a single bulk charge without reclaim first, kswapd may wake */
5536 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5538 mc.precharge += count;
5542 /* Try charges one by one with reclaim, but do not retry */
5544 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5558 enum mc_target_type {
5565 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5566 unsigned long addr, pte_t ptent)
5568 struct page *page = vm_normal_page(vma, addr, ptent);
5570 if (!page || !page_mapped(page))
5572 if (PageAnon(page)) {
5573 if (!(mc.flags & MOVE_ANON))
5576 if (!(mc.flags & MOVE_FILE))
5579 if (!get_page_unless_zero(page))
5585 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5586 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5587 pte_t ptent, swp_entry_t *entry)
5589 struct page *page = NULL;
5590 swp_entry_t ent = pte_to_swp_entry(ptent);
5592 if (!(mc.flags & MOVE_ANON))
5596 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5597 * a device and because they are not accessible by CPU they are store
5598 * as special swap entry in the CPU page table.
5600 if (is_device_private_entry(ent)) {
5601 page = device_private_entry_to_page(ent);
5603 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5604 * a refcount of 1 when free (unlike normal page)
5606 if (!page_ref_add_unless(page, 1, 1))
5611 if (non_swap_entry(ent))
5615 * Because lookup_swap_cache() updates some statistics counter,
5616 * we call find_get_page() with swapper_space directly.
5618 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5619 entry->val = ent.val;
5624 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5625 pte_t ptent, swp_entry_t *entry)
5631 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5632 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5634 if (!vma->vm_file) /* anonymous vma */
5636 if (!(mc.flags & MOVE_FILE))
5639 /* page is moved even if it's not RSS of this task(page-faulted). */
5640 /* shmem/tmpfs may report page out on swap: account for that too. */
5641 return find_get_incore_page(vma->vm_file->f_mapping,
5642 linear_page_index(vma, addr));
5646 * mem_cgroup_move_account - move account of the page
5648 * @compound: charge the page as compound or small page
5649 * @from: mem_cgroup which the page is moved from.
5650 * @to: mem_cgroup which the page is moved to. @from != @to.
5652 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5654 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5657 static int mem_cgroup_move_account(struct page *page,
5659 struct mem_cgroup *from,
5660 struct mem_cgroup *to)
5662 struct lruvec *from_vec, *to_vec;
5663 struct pglist_data *pgdat;
5664 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5667 VM_BUG_ON(from == to);
5668 VM_BUG_ON_PAGE(PageLRU(page), page);
5669 VM_BUG_ON(compound && !PageTransHuge(page));
5672 * Prevent mem_cgroup_migrate() from looking at
5673 * page->mem_cgroup of its source page while we change it.
5676 if (!trylock_page(page))
5680 if (page->mem_cgroup != from)
5683 pgdat = page_pgdat(page);
5684 from_vec = mem_cgroup_lruvec(from, pgdat);
5685 to_vec = mem_cgroup_lruvec(to, pgdat);
5687 lock_page_memcg(page);
5689 if (PageAnon(page)) {
5690 if (page_mapped(page)) {
5691 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5692 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5693 if (PageTransHuge(page)) {
5694 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5696 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5702 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5703 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5705 if (PageSwapBacked(page)) {
5706 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5707 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5710 if (page_mapped(page)) {
5711 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5712 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5715 if (PageDirty(page)) {
5716 struct address_space *mapping = page_mapping(page);
5718 if (mapping_can_writeback(mapping)) {
5719 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5721 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5727 if (PageWriteback(page)) {
5728 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5729 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5733 * All state has been migrated, let's switch to the new memcg.
5735 * It is safe to change page->mem_cgroup here because the page
5736 * is referenced, charged, isolated, and locked: we can't race
5737 * with (un)charging, migration, LRU putback, or anything else
5738 * that would rely on a stable page->mem_cgroup.
5740 * Note that lock_page_memcg is a memcg lock, not a page lock,
5741 * to save space. As soon as we switch page->mem_cgroup to a
5742 * new memcg that isn't locked, the above state can change
5743 * concurrently again. Make sure we're truly done with it.
5748 css_put(&from->css);
5750 page->mem_cgroup = to;
5752 __unlock_page_memcg(from);
5756 local_irq_disable();
5757 mem_cgroup_charge_statistics(to, page, nr_pages);
5758 memcg_check_events(to, page);
5759 mem_cgroup_charge_statistics(from, page, -nr_pages);
5760 memcg_check_events(from, page);
5769 * get_mctgt_type - get target type of moving charge
5770 * @vma: the vma the pte to be checked belongs
5771 * @addr: the address corresponding to the pte to be checked
5772 * @ptent: the pte to be checked
5773 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5776 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5777 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5778 * move charge. if @target is not NULL, the page is stored in target->page
5779 * with extra refcnt got(Callers should handle it).
5780 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5781 * target for charge migration. if @target is not NULL, the entry is stored
5783 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5784 * (so ZONE_DEVICE page and thus not on the lru).
5785 * For now we such page is charge like a regular page would be as for all
5786 * intent and purposes it is just special memory taking the place of a
5789 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5791 * Called with pte lock held.
5794 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5795 unsigned long addr, pte_t ptent, union mc_target *target)
5797 struct page *page = NULL;
5798 enum mc_target_type ret = MC_TARGET_NONE;
5799 swp_entry_t ent = { .val = 0 };
5801 if (pte_present(ptent))
5802 page = mc_handle_present_pte(vma, addr, ptent);
5803 else if (is_swap_pte(ptent))
5804 page = mc_handle_swap_pte(vma, ptent, &ent);
5805 else if (pte_none(ptent))
5806 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5808 if (!page && !ent.val)
5812 * Do only loose check w/o serialization.
5813 * mem_cgroup_move_account() checks the page is valid or
5814 * not under LRU exclusion.
5816 if (page->mem_cgroup == mc.from) {
5817 ret = MC_TARGET_PAGE;
5818 if (is_device_private_page(page))
5819 ret = MC_TARGET_DEVICE;
5821 target->page = page;
5823 if (!ret || !target)
5827 * There is a swap entry and a page doesn't exist or isn't charged.
5828 * But we cannot move a tail-page in a THP.
5830 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5831 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5832 ret = MC_TARGET_SWAP;
5839 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5841 * We don't consider PMD mapped swapping or file mapped pages because THP does
5842 * not support them for now.
5843 * Caller should make sure that pmd_trans_huge(pmd) is true.
5845 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5846 unsigned long addr, pmd_t pmd, union mc_target *target)
5848 struct page *page = NULL;
5849 enum mc_target_type ret = MC_TARGET_NONE;
5851 if (unlikely(is_swap_pmd(pmd))) {
5852 VM_BUG_ON(thp_migration_supported() &&
5853 !is_pmd_migration_entry(pmd));
5856 page = pmd_page(pmd);
5857 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5858 if (!(mc.flags & MOVE_ANON))
5860 if (page->mem_cgroup == mc.from) {
5861 ret = MC_TARGET_PAGE;
5864 target->page = page;
5870 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5871 unsigned long addr, pmd_t pmd, union mc_target *target)
5873 return MC_TARGET_NONE;
5877 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5878 unsigned long addr, unsigned long end,
5879 struct mm_walk *walk)
5881 struct vm_area_struct *vma = walk->vma;
5885 ptl = pmd_trans_huge_lock(pmd, vma);
5888 * Note their can not be MC_TARGET_DEVICE for now as we do not
5889 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5890 * this might change.
5892 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5893 mc.precharge += HPAGE_PMD_NR;
5898 if (pmd_trans_unstable(pmd))
5900 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5901 for (; addr != end; pte++, addr += PAGE_SIZE)
5902 if (get_mctgt_type(vma, addr, *pte, NULL))
5903 mc.precharge++; /* increment precharge temporarily */
5904 pte_unmap_unlock(pte - 1, ptl);
5910 static const struct mm_walk_ops precharge_walk_ops = {
5911 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5914 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5916 unsigned long precharge;
5919 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5920 mmap_read_unlock(mm);
5922 precharge = mc.precharge;
5928 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5930 unsigned long precharge = mem_cgroup_count_precharge(mm);
5932 VM_BUG_ON(mc.moving_task);
5933 mc.moving_task = current;
5934 return mem_cgroup_do_precharge(precharge);
5937 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5938 static void __mem_cgroup_clear_mc(void)
5940 struct mem_cgroup *from = mc.from;
5941 struct mem_cgroup *to = mc.to;
5943 /* we must uncharge all the leftover precharges from mc.to */
5945 cancel_charge(mc.to, mc.precharge);
5949 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5950 * we must uncharge here.
5952 if (mc.moved_charge) {
5953 cancel_charge(mc.from, mc.moved_charge);
5954 mc.moved_charge = 0;
5956 /* we must fixup refcnts and charges */
5957 if (mc.moved_swap) {
5958 /* uncharge swap account from the old cgroup */
5959 if (!mem_cgroup_is_root(mc.from))
5960 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5962 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5965 * we charged both to->memory and to->memsw, so we
5966 * should uncharge to->memory.
5968 if (!mem_cgroup_is_root(mc.to))
5969 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5973 memcg_oom_recover(from);
5974 memcg_oom_recover(to);
5975 wake_up_all(&mc.waitq);
5978 static void mem_cgroup_clear_mc(void)
5980 struct mm_struct *mm = mc.mm;
5983 * we must clear moving_task before waking up waiters at the end of
5986 mc.moving_task = NULL;
5987 __mem_cgroup_clear_mc();
5988 spin_lock(&mc.lock);
5992 spin_unlock(&mc.lock);
5997 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5999 struct cgroup_subsys_state *css;
6000 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
6001 struct mem_cgroup *from;
6002 struct task_struct *leader, *p;
6003 struct mm_struct *mm;
6004 unsigned long move_flags;
6007 /* charge immigration isn't supported on the default hierarchy */
6008 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6012 * Multi-process migrations only happen on the default hierarchy
6013 * where charge immigration is not used. Perform charge
6014 * immigration if @tset contains a leader and whine if there are
6018 cgroup_taskset_for_each_leader(leader, css, tset) {
6021 memcg = mem_cgroup_from_css(css);
6027 * We are now commited to this value whatever it is. Changes in this
6028 * tunable will only affect upcoming migrations, not the current one.
6029 * So we need to save it, and keep it going.
6031 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
6035 from = mem_cgroup_from_task(p);
6037 VM_BUG_ON(from == memcg);
6039 mm = get_task_mm(p);
6042 /* We move charges only when we move a owner of the mm */
6043 if (mm->owner == p) {
6046 VM_BUG_ON(mc.precharge);
6047 VM_BUG_ON(mc.moved_charge);
6048 VM_BUG_ON(mc.moved_swap);
6050 spin_lock(&mc.lock);
6054 mc.flags = move_flags;
6055 spin_unlock(&mc.lock);
6056 /* We set mc.moving_task later */
6058 ret = mem_cgroup_precharge_mc(mm);
6060 mem_cgroup_clear_mc();
6067 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6070 mem_cgroup_clear_mc();
6073 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6074 unsigned long addr, unsigned long end,
6075 struct mm_walk *walk)
6078 struct vm_area_struct *vma = walk->vma;
6081 enum mc_target_type target_type;
6082 union mc_target target;
6085 ptl = pmd_trans_huge_lock(pmd, vma);
6087 if (mc.precharge < HPAGE_PMD_NR) {
6091 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6092 if (target_type == MC_TARGET_PAGE) {
6094 if (!isolate_lru_page(page)) {
6095 if (!mem_cgroup_move_account(page, true,
6097 mc.precharge -= HPAGE_PMD_NR;
6098 mc.moved_charge += HPAGE_PMD_NR;
6100 putback_lru_page(page);
6103 } else if (target_type == MC_TARGET_DEVICE) {
6105 if (!mem_cgroup_move_account(page, true,
6107 mc.precharge -= HPAGE_PMD_NR;
6108 mc.moved_charge += HPAGE_PMD_NR;
6116 if (pmd_trans_unstable(pmd))
6119 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6120 for (; addr != end; addr += PAGE_SIZE) {
6121 pte_t ptent = *(pte++);
6122 bool device = false;
6128 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6129 case MC_TARGET_DEVICE:
6132 case MC_TARGET_PAGE:
6135 * We can have a part of the split pmd here. Moving it
6136 * can be done but it would be too convoluted so simply
6137 * ignore such a partial THP and keep it in original
6138 * memcg. There should be somebody mapping the head.
6140 if (PageTransCompound(page))
6142 if (!device && isolate_lru_page(page))
6144 if (!mem_cgroup_move_account(page, false,
6147 /* we uncharge from mc.from later. */
6151 putback_lru_page(page);
6152 put: /* get_mctgt_type() gets the page */
6155 case MC_TARGET_SWAP:
6157 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6159 mem_cgroup_id_get_many(mc.to, 1);
6160 /* we fixup other refcnts and charges later. */
6168 pte_unmap_unlock(pte - 1, ptl);
6173 * We have consumed all precharges we got in can_attach().
6174 * We try charge one by one, but don't do any additional
6175 * charges to mc.to if we have failed in charge once in attach()
6178 ret = mem_cgroup_do_precharge(1);
6186 static const struct mm_walk_ops charge_walk_ops = {
6187 .pmd_entry = mem_cgroup_move_charge_pte_range,
6190 static void mem_cgroup_move_charge(void)
6192 lru_add_drain_all();
6194 * Signal lock_page_memcg() to take the memcg's move_lock
6195 * while we're moving its pages to another memcg. Then wait
6196 * for already started RCU-only updates to finish.
6198 atomic_inc(&mc.from->moving_account);
6201 if (unlikely(!mmap_read_trylock(mc.mm))) {
6203 * Someone who are holding the mmap_lock might be waiting in
6204 * waitq. So we cancel all extra charges, wake up all waiters,
6205 * and retry. Because we cancel precharges, we might not be able
6206 * to move enough charges, but moving charge is a best-effort
6207 * feature anyway, so it wouldn't be a big problem.
6209 __mem_cgroup_clear_mc();
6214 * When we have consumed all precharges and failed in doing
6215 * additional charge, the page walk just aborts.
6217 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6220 mmap_read_unlock(mc.mm);
6221 atomic_dec(&mc.from->moving_account);
6224 static void mem_cgroup_move_task(void)
6227 mem_cgroup_move_charge();
6228 mem_cgroup_clear_mc();
6231 #else /* !CONFIG_MMU */
6232 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6236 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6239 static void mem_cgroup_move_task(void)
6244 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6246 if (value == PAGE_COUNTER_MAX)
6247 seq_puts(m, "max\n");
6249 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6254 static u64 memory_current_read(struct cgroup_subsys_state *css,
6257 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6259 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6262 static int memory_min_show(struct seq_file *m, void *v)
6264 return seq_puts_memcg_tunable(m,
6265 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6268 static ssize_t memory_min_write(struct kernfs_open_file *of,
6269 char *buf, size_t nbytes, loff_t off)
6271 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6275 buf = strstrip(buf);
6276 err = page_counter_memparse(buf, "max", &min);
6280 page_counter_set_min(&memcg->memory, min);
6285 static int memory_low_show(struct seq_file *m, void *v)
6287 return seq_puts_memcg_tunable(m,
6288 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6291 static ssize_t memory_low_write(struct kernfs_open_file *of,
6292 char *buf, size_t nbytes, loff_t off)
6294 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6298 buf = strstrip(buf);
6299 err = page_counter_memparse(buf, "max", &low);
6303 page_counter_set_low(&memcg->memory, low);
6308 static int memory_high_show(struct seq_file *m, void *v)
6310 return seq_puts_memcg_tunable(m,
6311 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6314 static ssize_t memory_high_write(struct kernfs_open_file *of,
6315 char *buf, size_t nbytes, loff_t off)
6317 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6318 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6319 bool drained = false;
6323 buf = strstrip(buf);
6324 err = page_counter_memparse(buf, "max", &high);
6329 unsigned long nr_pages = page_counter_read(&memcg->memory);
6330 unsigned long reclaimed;
6332 if (nr_pages <= high)
6335 if (signal_pending(current))
6339 drain_all_stock(memcg);
6344 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6347 if (!reclaimed && !nr_retries--)
6351 page_counter_set_high(&memcg->memory, high);
6353 memcg_wb_domain_size_changed(memcg);
6358 static int memory_max_show(struct seq_file *m, void *v)
6360 return seq_puts_memcg_tunable(m,
6361 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6364 static ssize_t memory_max_write(struct kernfs_open_file *of,
6365 char *buf, size_t nbytes, loff_t off)
6367 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6368 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6369 bool drained = false;
6373 buf = strstrip(buf);
6374 err = page_counter_memparse(buf, "max", &max);
6378 xchg(&memcg->memory.max, max);
6381 unsigned long nr_pages = page_counter_read(&memcg->memory);
6383 if (nr_pages <= max)
6386 if (signal_pending(current))
6390 drain_all_stock(memcg);
6396 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6402 memcg_memory_event(memcg, MEMCG_OOM);
6403 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6407 memcg_wb_domain_size_changed(memcg);
6411 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6413 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6414 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6415 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6416 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6417 seq_printf(m, "oom_kill %lu\n",
6418 atomic_long_read(&events[MEMCG_OOM_KILL]));
6421 static int memory_events_show(struct seq_file *m, void *v)
6423 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6425 __memory_events_show(m, memcg->memory_events);
6429 static int memory_events_local_show(struct seq_file *m, void *v)
6431 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6433 __memory_events_show(m, memcg->memory_events_local);
6437 static int memory_stat_show(struct seq_file *m, void *v)
6439 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6442 buf = memory_stat_format(memcg);
6451 static int memory_numa_stat_show(struct seq_file *m, void *v)
6454 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6456 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6459 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6462 seq_printf(m, "%s", memory_stats[i].name);
6463 for_each_node_state(nid, N_MEMORY) {
6465 struct lruvec *lruvec;
6467 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6468 size = lruvec_page_state(lruvec, memory_stats[i].idx);
6469 size *= memory_stats[i].ratio;
6470 seq_printf(m, " N%d=%llu", nid, size);
6479 static int memory_oom_group_show(struct seq_file *m, void *v)
6481 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6483 seq_printf(m, "%d\n", memcg->oom_group);
6488 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6489 char *buf, size_t nbytes, loff_t off)
6491 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6494 buf = strstrip(buf);
6498 ret = kstrtoint(buf, 0, &oom_group);
6502 if (oom_group != 0 && oom_group != 1)
6505 memcg->oom_group = oom_group;
6510 static struct cftype memory_files[] = {
6513 .flags = CFTYPE_NOT_ON_ROOT,
6514 .read_u64 = memory_current_read,
6518 .flags = CFTYPE_NOT_ON_ROOT,
6519 .seq_show = memory_min_show,
6520 .write = memory_min_write,
6524 .flags = CFTYPE_NOT_ON_ROOT,
6525 .seq_show = memory_low_show,
6526 .write = memory_low_write,
6530 .flags = CFTYPE_NOT_ON_ROOT,
6531 .seq_show = memory_high_show,
6532 .write = memory_high_write,
6536 .flags = CFTYPE_NOT_ON_ROOT,
6537 .seq_show = memory_max_show,
6538 .write = memory_max_write,
6542 .flags = CFTYPE_NOT_ON_ROOT,
6543 .file_offset = offsetof(struct mem_cgroup, events_file),
6544 .seq_show = memory_events_show,
6547 .name = "events.local",
6548 .flags = CFTYPE_NOT_ON_ROOT,
6549 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6550 .seq_show = memory_events_local_show,
6554 .seq_show = memory_stat_show,
6558 .name = "numa_stat",
6559 .seq_show = memory_numa_stat_show,
6563 .name = "oom.group",
6564 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6565 .seq_show = memory_oom_group_show,
6566 .write = memory_oom_group_write,
6571 struct cgroup_subsys memory_cgrp_subsys = {
6572 .css_alloc = mem_cgroup_css_alloc,
6573 .css_online = mem_cgroup_css_online,
6574 .css_offline = mem_cgroup_css_offline,
6575 .css_released = mem_cgroup_css_released,
6576 .css_free = mem_cgroup_css_free,
6577 .css_reset = mem_cgroup_css_reset,
6578 .can_attach = mem_cgroup_can_attach,
6579 .cancel_attach = mem_cgroup_cancel_attach,
6580 .post_attach = mem_cgroup_move_task,
6581 .dfl_cftypes = memory_files,
6582 .legacy_cftypes = mem_cgroup_legacy_files,
6587 * This function calculates an individual cgroup's effective
6588 * protection which is derived from its own memory.min/low, its
6589 * parent's and siblings' settings, as well as the actual memory
6590 * distribution in the tree.
6592 * The following rules apply to the effective protection values:
6594 * 1. At the first level of reclaim, effective protection is equal to
6595 * the declared protection in memory.min and memory.low.
6597 * 2. To enable safe delegation of the protection configuration, at
6598 * subsequent levels the effective protection is capped to the
6599 * parent's effective protection.
6601 * 3. To make complex and dynamic subtrees easier to configure, the
6602 * user is allowed to overcommit the declared protection at a given
6603 * level. If that is the case, the parent's effective protection is
6604 * distributed to the children in proportion to how much protection
6605 * they have declared and how much of it they are utilizing.
6607 * This makes distribution proportional, but also work-conserving:
6608 * if one cgroup claims much more protection than it uses memory,
6609 * the unused remainder is available to its siblings.
6611 * 4. Conversely, when the declared protection is undercommitted at a
6612 * given level, the distribution of the larger parental protection
6613 * budget is NOT proportional. A cgroup's protection from a sibling
6614 * is capped to its own memory.min/low setting.
6616 * 5. However, to allow protecting recursive subtrees from each other
6617 * without having to declare each individual cgroup's fixed share
6618 * of the ancestor's claim to protection, any unutilized -
6619 * "floating" - protection from up the tree is distributed in
6620 * proportion to each cgroup's *usage*. This makes the protection
6621 * neutral wrt sibling cgroups and lets them compete freely over
6622 * the shared parental protection budget, but it protects the
6623 * subtree as a whole from neighboring subtrees.
6625 * Note that 4. and 5. are not in conflict: 4. is about protecting
6626 * against immediate siblings whereas 5. is about protecting against
6627 * neighboring subtrees.
6629 static unsigned long effective_protection(unsigned long usage,
6630 unsigned long parent_usage,
6631 unsigned long setting,
6632 unsigned long parent_effective,
6633 unsigned long siblings_protected)
6635 unsigned long protected;
6638 protected = min(usage, setting);
6640 * If all cgroups at this level combined claim and use more
6641 * protection then what the parent affords them, distribute
6642 * shares in proportion to utilization.
6644 * We are using actual utilization rather than the statically
6645 * claimed protection in order to be work-conserving: claimed
6646 * but unused protection is available to siblings that would
6647 * otherwise get a smaller chunk than what they claimed.
6649 if (siblings_protected > parent_effective)
6650 return protected * parent_effective / siblings_protected;
6653 * Ok, utilized protection of all children is within what the
6654 * parent affords them, so we know whatever this child claims
6655 * and utilizes is effectively protected.
6657 * If there is unprotected usage beyond this value, reclaim
6658 * will apply pressure in proportion to that amount.
6660 * If there is unutilized protection, the cgroup will be fully
6661 * shielded from reclaim, but we do return a smaller value for
6662 * protection than what the group could enjoy in theory. This
6663 * is okay. With the overcommit distribution above, effective
6664 * protection is always dependent on how memory is actually
6665 * consumed among the siblings anyway.
6670 * If the children aren't claiming (all of) the protection
6671 * afforded to them by the parent, distribute the remainder in
6672 * proportion to the (unprotected) memory of each cgroup. That
6673 * way, cgroups that aren't explicitly prioritized wrt each
6674 * other compete freely over the allowance, but they are
6675 * collectively protected from neighboring trees.
6677 * We're using unprotected memory for the weight so that if
6678 * some cgroups DO claim explicit protection, we don't protect
6679 * the same bytes twice.
6681 * Check both usage and parent_usage against the respective
6682 * protected values. One should imply the other, but they
6683 * aren't read atomically - make sure the division is sane.
6685 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6687 if (parent_effective > siblings_protected &&
6688 parent_usage > siblings_protected &&
6689 usage > protected) {
6690 unsigned long unclaimed;
6692 unclaimed = parent_effective - siblings_protected;
6693 unclaimed *= usage - protected;
6694 unclaimed /= parent_usage - siblings_protected;
6703 * mem_cgroup_protected - check if memory consumption is in the normal range
6704 * @root: the top ancestor of the sub-tree being checked
6705 * @memcg: the memory cgroup to check
6707 * WARNING: This function is not stateless! It can only be used as part
6708 * of a top-down tree iteration, not for isolated queries.
6710 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6711 struct mem_cgroup *memcg)
6713 unsigned long usage, parent_usage;
6714 struct mem_cgroup *parent;
6716 if (mem_cgroup_disabled())
6720 root = root_mem_cgroup;
6723 * Effective values of the reclaim targets are ignored so they
6724 * can be stale. Have a look at mem_cgroup_protection for more
6726 * TODO: calculation should be more robust so that we do not need
6727 * that special casing.
6732 usage = page_counter_read(&memcg->memory);
6736 parent = parent_mem_cgroup(memcg);
6737 /* No parent means a non-hierarchical mode on v1 memcg */
6741 if (parent == root) {
6742 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6743 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6747 parent_usage = page_counter_read(&parent->memory);
6749 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6750 READ_ONCE(memcg->memory.min),
6751 READ_ONCE(parent->memory.emin),
6752 atomic_long_read(&parent->memory.children_min_usage)));
6754 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6755 READ_ONCE(memcg->memory.low),
6756 READ_ONCE(parent->memory.elow),
6757 atomic_long_read(&parent->memory.children_low_usage)));
6761 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6762 * @page: page to charge
6763 * @mm: mm context of the victim
6764 * @gfp_mask: reclaim mode
6766 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6767 * pages according to @gfp_mask if necessary.
6769 * Returns 0 on success. Otherwise, an error code is returned.
6771 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6773 unsigned int nr_pages = thp_nr_pages(page);
6774 struct mem_cgroup *memcg = NULL;
6777 if (mem_cgroup_disabled())
6780 if (PageSwapCache(page)) {
6781 swp_entry_t ent = { .val = page_private(page), };
6785 * Every swap fault against a single page tries to charge the
6786 * page, bail as early as possible. shmem_unuse() encounters
6787 * already charged pages, too. page->mem_cgroup is protected
6788 * by the page lock, which serializes swap cache removal, which
6789 * in turn serializes uncharging.
6791 VM_BUG_ON_PAGE(!PageLocked(page), page);
6792 if (compound_head(page)->mem_cgroup)
6795 id = lookup_swap_cgroup_id(ent);
6797 memcg = mem_cgroup_from_id(id);
6798 if (memcg && !css_tryget_online(&memcg->css))
6804 memcg = get_mem_cgroup_from_mm(mm);
6806 ret = try_charge(memcg, gfp_mask, nr_pages);
6810 css_get(&memcg->css);
6811 commit_charge(page, memcg);
6813 local_irq_disable();
6814 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6815 memcg_check_events(memcg, page);
6818 if (PageSwapCache(page)) {
6819 swp_entry_t entry = { .val = page_private(page) };
6821 * The swap entry might not get freed for a long time,
6822 * let's not wait for it. The page already received a
6823 * memory+swap charge, drop the swap entry duplicate.
6825 mem_cgroup_uncharge_swap(entry, nr_pages);
6829 css_put(&memcg->css);
6834 struct uncharge_gather {
6835 struct mem_cgroup *memcg;
6836 unsigned long nr_pages;
6837 unsigned long pgpgout;
6838 unsigned long nr_kmem;
6839 struct page *dummy_page;
6842 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6844 memset(ug, 0, sizeof(*ug));
6847 static void uncharge_batch(const struct uncharge_gather *ug)
6849 unsigned long flags;
6851 if (!mem_cgroup_is_root(ug->memcg)) {
6852 page_counter_uncharge(&ug->memcg->memory, ug->nr_pages);
6853 if (do_memsw_account())
6854 page_counter_uncharge(&ug->memcg->memsw, ug->nr_pages);
6855 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6856 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6857 memcg_oom_recover(ug->memcg);
6860 local_irq_save(flags);
6861 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6862 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_pages);
6863 memcg_check_events(ug->memcg, ug->dummy_page);
6864 local_irq_restore(flags);
6866 /* drop reference from uncharge_page */
6867 css_put(&ug->memcg->css);
6870 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6872 unsigned long nr_pages;
6874 VM_BUG_ON_PAGE(PageLRU(page), page);
6876 if (!page->mem_cgroup)
6880 * Nobody should be changing or seriously looking at
6881 * page->mem_cgroup at this point, we have fully
6882 * exclusive access to the page.
6885 if (ug->memcg != page->mem_cgroup) {
6888 uncharge_gather_clear(ug);
6890 ug->memcg = page->mem_cgroup;
6892 /* pairs with css_put in uncharge_batch */
6893 css_get(&ug->memcg->css);
6896 nr_pages = compound_nr(page);
6897 ug->nr_pages += nr_pages;
6899 if (!PageKmemcg(page)) {
6902 ug->nr_kmem += nr_pages;
6903 __ClearPageKmemcg(page);
6906 ug->dummy_page = page;
6907 page->mem_cgroup = NULL;
6908 css_put(&ug->memcg->css);
6911 static void uncharge_list(struct list_head *page_list)
6913 struct uncharge_gather ug;
6914 struct list_head *next;
6916 uncharge_gather_clear(&ug);
6919 * Note that the list can be a single page->lru; hence the
6920 * do-while loop instead of a simple list_for_each_entry().
6922 next = page_list->next;
6926 page = list_entry(next, struct page, lru);
6927 next = page->lru.next;
6929 uncharge_page(page, &ug);
6930 } while (next != page_list);
6933 uncharge_batch(&ug);
6937 * mem_cgroup_uncharge - uncharge a page
6938 * @page: page to uncharge
6940 * Uncharge a page previously charged with mem_cgroup_charge().
6942 void mem_cgroup_uncharge(struct page *page)
6944 struct uncharge_gather ug;
6946 if (mem_cgroup_disabled())
6949 /* Don't touch page->lru of any random page, pre-check: */
6950 if (!page->mem_cgroup)
6953 uncharge_gather_clear(&ug);
6954 uncharge_page(page, &ug);
6955 uncharge_batch(&ug);
6959 * mem_cgroup_uncharge_list - uncharge a list of page
6960 * @page_list: list of pages to uncharge
6962 * Uncharge a list of pages previously charged with
6963 * mem_cgroup_charge().
6965 void mem_cgroup_uncharge_list(struct list_head *page_list)
6967 if (mem_cgroup_disabled())
6970 if (!list_empty(page_list))
6971 uncharge_list(page_list);
6975 * mem_cgroup_migrate - charge a page's replacement
6976 * @oldpage: currently circulating page
6977 * @newpage: replacement page
6979 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6980 * be uncharged upon free.
6982 * Both pages must be locked, @newpage->mapping must be set up.
6984 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6986 struct mem_cgroup *memcg;
6987 unsigned int nr_pages;
6988 unsigned long flags;
6990 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6991 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6992 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6993 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6996 if (mem_cgroup_disabled())
6999 /* Page cache replacement: new page already charged? */
7000 if (newpage->mem_cgroup)
7003 memcg = oldpage->mem_cgroup;
7007 /* Force-charge the new page. The old one will be freed soon */
7008 nr_pages = thp_nr_pages(newpage);
7010 page_counter_charge(&memcg->memory, nr_pages);
7011 if (do_memsw_account())
7012 page_counter_charge(&memcg->memsw, nr_pages);
7014 css_get(&memcg->css);
7015 commit_charge(newpage, memcg);
7017 local_irq_save(flags);
7018 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
7019 memcg_check_events(memcg, newpage);
7020 local_irq_restore(flags);
7023 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
7024 EXPORT_SYMBOL(memcg_sockets_enabled_key);
7026 void mem_cgroup_sk_alloc(struct sock *sk)
7028 struct mem_cgroup *memcg;
7030 if (!mem_cgroup_sockets_enabled)
7033 /* Do not associate the sock with unrelated interrupted task's memcg. */
7038 memcg = mem_cgroup_from_task(current);
7039 if (memcg == root_mem_cgroup)
7041 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
7043 if (css_tryget(&memcg->css))
7044 sk->sk_memcg = memcg;
7049 void mem_cgroup_sk_free(struct sock *sk)
7052 css_put(&sk->sk_memcg->css);
7056 * mem_cgroup_charge_skmem - charge socket memory
7057 * @memcg: memcg to charge
7058 * @nr_pages: number of pages to charge
7060 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7061 * @memcg's configured limit, %false if the charge had to be forced.
7063 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7065 gfp_t gfp_mask = GFP_KERNEL;
7067 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7068 struct page_counter *fail;
7070 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7071 memcg->tcpmem_pressure = 0;
7074 page_counter_charge(&memcg->tcpmem, nr_pages);
7075 memcg->tcpmem_pressure = 1;
7079 /* Don't block in the packet receive path */
7081 gfp_mask = GFP_NOWAIT;
7083 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7085 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
7088 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
7093 * mem_cgroup_uncharge_skmem - uncharge socket memory
7094 * @memcg: memcg to uncharge
7095 * @nr_pages: number of pages to uncharge
7097 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7099 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7100 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7104 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7106 refill_stock(memcg, nr_pages);
7109 static int __init cgroup_memory(char *s)
7113 while ((token = strsep(&s, ",")) != NULL) {
7116 if (!strcmp(token, "nosocket"))
7117 cgroup_memory_nosocket = true;
7118 if (!strcmp(token, "nokmem"))
7119 cgroup_memory_nokmem = true;
7123 __setup("cgroup.memory=", cgroup_memory);
7126 * subsys_initcall() for memory controller.
7128 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7129 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7130 * basically everything that doesn't depend on a specific mem_cgroup structure
7131 * should be initialized from here.
7133 static int __init mem_cgroup_init(void)
7137 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7138 memcg_hotplug_cpu_dead);
7140 for_each_possible_cpu(cpu)
7141 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7144 for_each_node(node) {
7145 struct mem_cgroup_tree_per_node *rtpn;
7147 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7148 node_online(node) ? node : NUMA_NO_NODE);
7150 rtpn->rb_root = RB_ROOT;
7151 rtpn->rb_rightmost = NULL;
7152 spin_lock_init(&rtpn->lock);
7153 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7158 subsys_initcall(mem_cgroup_init);
7160 #ifdef CONFIG_MEMCG_SWAP
7161 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7163 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7165 * The root cgroup cannot be destroyed, so it's refcount must
7168 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7172 memcg = parent_mem_cgroup(memcg);
7174 memcg = root_mem_cgroup;
7180 * mem_cgroup_swapout - transfer a memsw charge to swap
7181 * @page: page whose memsw charge to transfer
7182 * @entry: swap entry to move the charge to
7184 * Transfer the memsw charge of @page to @entry.
7186 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7188 struct mem_cgroup *memcg, *swap_memcg;
7189 unsigned int nr_entries;
7190 unsigned short oldid;
7192 VM_BUG_ON_PAGE(PageLRU(page), page);
7193 VM_BUG_ON_PAGE(page_count(page), page);
7195 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7198 memcg = page->mem_cgroup;
7200 /* Readahead page, never charged */
7205 * In case the memcg owning these pages has been offlined and doesn't
7206 * have an ID allocated to it anymore, charge the closest online
7207 * ancestor for the swap instead and transfer the memory+swap charge.
7209 swap_memcg = mem_cgroup_id_get_online(memcg);
7210 nr_entries = thp_nr_pages(page);
7211 /* Get references for the tail pages, too */
7213 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7214 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7216 VM_BUG_ON_PAGE(oldid, page);
7217 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7219 page->mem_cgroup = NULL;
7221 if (!mem_cgroup_is_root(memcg))
7222 page_counter_uncharge(&memcg->memory, nr_entries);
7224 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7225 if (!mem_cgroup_is_root(swap_memcg))
7226 page_counter_charge(&swap_memcg->memsw, nr_entries);
7227 page_counter_uncharge(&memcg->memsw, nr_entries);
7231 * Interrupts should be disabled here because the caller holds the
7232 * i_pages lock which is taken with interrupts-off. It is
7233 * important here to have the interrupts disabled because it is the
7234 * only synchronisation we have for updating the per-CPU variables.
7236 VM_BUG_ON(!irqs_disabled());
7237 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7238 memcg_check_events(memcg, page);
7240 css_put(&memcg->css);
7244 * mem_cgroup_try_charge_swap - try charging swap space for a page
7245 * @page: page being added to swap
7246 * @entry: swap entry to charge
7248 * Try to charge @page's memcg for the swap space at @entry.
7250 * Returns 0 on success, -ENOMEM on failure.
7252 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7254 unsigned int nr_pages = thp_nr_pages(page);
7255 struct page_counter *counter;
7256 struct mem_cgroup *memcg;
7257 unsigned short oldid;
7259 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7262 memcg = page->mem_cgroup;
7264 /* Readahead page, never charged */
7269 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7273 memcg = mem_cgroup_id_get_online(memcg);
7275 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7276 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7277 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7278 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7279 mem_cgroup_id_put(memcg);
7283 /* Get references for the tail pages, too */
7285 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7286 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7287 VM_BUG_ON_PAGE(oldid, page);
7288 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7294 * mem_cgroup_uncharge_swap - uncharge swap space
7295 * @entry: swap entry to uncharge
7296 * @nr_pages: the amount of swap space to uncharge
7298 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7300 struct mem_cgroup *memcg;
7303 id = swap_cgroup_record(entry, 0, nr_pages);
7305 memcg = mem_cgroup_from_id(id);
7307 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7308 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7309 page_counter_uncharge(&memcg->swap, nr_pages);
7311 page_counter_uncharge(&memcg->memsw, nr_pages);
7313 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7314 mem_cgroup_id_put_many(memcg, nr_pages);
7319 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7321 long nr_swap_pages = get_nr_swap_pages();
7323 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7324 return nr_swap_pages;
7325 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7326 nr_swap_pages = min_t(long, nr_swap_pages,
7327 READ_ONCE(memcg->swap.max) -
7328 page_counter_read(&memcg->swap));
7329 return nr_swap_pages;
7332 bool mem_cgroup_swap_full(struct page *page)
7334 struct mem_cgroup *memcg;
7336 VM_BUG_ON_PAGE(!PageLocked(page), page);
7340 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7343 memcg = page->mem_cgroup;
7347 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7348 unsigned long usage = page_counter_read(&memcg->swap);
7350 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7351 usage * 2 >= READ_ONCE(memcg->swap.max))
7358 static int __init setup_swap_account(char *s)
7360 if (!strcmp(s, "1"))
7361 cgroup_memory_noswap = false;
7362 else if (!strcmp(s, "0"))
7363 cgroup_memory_noswap = true;
7366 __setup("swapaccount=", setup_swap_account);
7368 static u64 swap_current_read(struct cgroup_subsys_state *css,
7371 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7373 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7376 static int swap_high_show(struct seq_file *m, void *v)
7378 return seq_puts_memcg_tunable(m,
7379 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7382 static ssize_t swap_high_write(struct kernfs_open_file *of,
7383 char *buf, size_t nbytes, loff_t off)
7385 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7389 buf = strstrip(buf);
7390 err = page_counter_memparse(buf, "max", &high);
7394 page_counter_set_high(&memcg->swap, high);
7399 static int swap_max_show(struct seq_file *m, void *v)
7401 return seq_puts_memcg_tunable(m,
7402 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7405 static ssize_t swap_max_write(struct kernfs_open_file *of,
7406 char *buf, size_t nbytes, loff_t off)
7408 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7412 buf = strstrip(buf);
7413 err = page_counter_memparse(buf, "max", &max);
7417 xchg(&memcg->swap.max, max);
7422 static int swap_events_show(struct seq_file *m, void *v)
7424 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7426 seq_printf(m, "high %lu\n",
7427 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7428 seq_printf(m, "max %lu\n",
7429 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7430 seq_printf(m, "fail %lu\n",
7431 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7436 static struct cftype swap_files[] = {
7438 .name = "swap.current",
7439 .flags = CFTYPE_NOT_ON_ROOT,
7440 .read_u64 = swap_current_read,
7443 .name = "swap.high",
7444 .flags = CFTYPE_NOT_ON_ROOT,
7445 .seq_show = swap_high_show,
7446 .write = swap_high_write,
7450 .flags = CFTYPE_NOT_ON_ROOT,
7451 .seq_show = swap_max_show,
7452 .write = swap_max_write,
7455 .name = "swap.events",
7456 .flags = CFTYPE_NOT_ON_ROOT,
7457 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7458 .seq_show = swap_events_show,
7463 static struct cftype memsw_files[] = {
7465 .name = "memsw.usage_in_bytes",
7466 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7467 .read_u64 = mem_cgroup_read_u64,
7470 .name = "memsw.max_usage_in_bytes",
7471 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7472 .write = mem_cgroup_reset,
7473 .read_u64 = mem_cgroup_read_u64,
7476 .name = "memsw.limit_in_bytes",
7477 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7478 .write = mem_cgroup_write,
7479 .read_u64 = mem_cgroup_read_u64,
7482 .name = "memsw.failcnt",
7483 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7484 .write = mem_cgroup_reset,
7485 .read_u64 = mem_cgroup_read_u64,
7487 { }, /* terminate */
7491 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7492 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7493 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7494 * boot parameter. This may result in premature OOPS inside
7495 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7497 static int __init mem_cgroup_swap_init(void)
7499 /* No memory control -> no swap control */
7500 if (mem_cgroup_disabled())
7501 cgroup_memory_noswap = true;
7503 if (cgroup_memory_noswap)
7506 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7507 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7511 core_initcall(mem_cgroup_swap_init);
7513 #endif /* CONFIG_MEMCG_SWAP */