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_uncharge_pages(struct obj_cgroup *objcg,
259 unsigned int nr_pages);
261 static void obj_cgroup_release(struct percpu_ref *ref)
263 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
264 struct mem_cgroup *memcg;
265 unsigned int nr_bytes;
266 unsigned int nr_pages;
270 * At this point all allocated objects are freed, and
271 * objcg->nr_charged_bytes can't have an arbitrary byte value.
272 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
274 * The following sequence can lead to it:
275 * 1) CPU0: objcg == stock->cached_objcg
276 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
277 * PAGE_SIZE bytes are charged
278 * 3) CPU1: a process from another memcg is allocating something,
279 * the stock if flushed,
280 * objcg->nr_charged_bytes = PAGE_SIZE - 92
281 * 5) CPU0: we do release this object,
282 * 92 bytes are added to stock->nr_bytes
283 * 6) CPU0: stock is flushed,
284 * 92 bytes are added to objcg->nr_charged_bytes
286 * In the result, nr_charged_bytes == PAGE_SIZE.
287 * This page will be uncharged in obj_cgroup_release().
289 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
290 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
291 nr_pages = nr_bytes >> PAGE_SHIFT;
293 spin_lock_irqsave(&css_set_lock, flags);
294 memcg = obj_cgroup_memcg(objcg);
296 obj_cgroup_uncharge_pages(objcg, nr_pages);
297 list_del(&objcg->list);
298 mem_cgroup_put(memcg);
299 spin_unlock_irqrestore(&css_set_lock, flags);
301 percpu_ref_exit(ref);
302 kfree_rcu(objcg, rcu);
305 static struct obj_cgroup *obj_cgroup_alloc(void)
307 struct obj_cgroup *objcg;
310 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
314 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
320 INIT_LIST_HEAD(&objcg->list);
324 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
325 struct mem_cgroup *parent)
327 struct obj_cgroup *objcg, *iter;
329 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
331 spin_lock_irq(&css_set_lock);
333 /* Move active objcg to the parent's list */
334 xchg(&objcg->memcg, parent);
335 css_get(&parent->css);
336 list_add(&objcg->list, &parent->objcg_list);
338 /* Move already reparented objcgs to the parent's list */
339 list_for_each_entry(iter, &memcg->objcg_list, list) {
340 css_get(&parent->css);
341 xchg(&iter->memcg, parent);
342 css_put(&memcg->css);
344 list_splice(&memcg->objcg_list, &parent->objcg_list);
346 spin_unlock_irq(&css_set_lock);
348 percpu_ref_kill(&objcg->refcnt);
352 * This will be used as a shrinker list's index.
353 * The main reason for not using cgroup id for this:
354 * this works better in sparse environments, where we have a lot of memcgs,
355 * but only a few kmem-limited. Or also, if we have, for instance, 200
356 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
357 * 200 entry array for that.
359 * The current size of the caches array is stored in memcg_nr_cache_ids. It
360 * will double each time we have to increase it.
362 static DEFINE_IDA(memcg_cache_ida);
363 int memcg_nr_cache_ids;
365 /* Protects memcg_nr_cache_ids */
366 static DECLARE_RWSEM(memcg_cache_ids_sem);
368 void memcg_get_cache_ids(void)
370 down_read(&memcg_cache_ids_sem);
373 void memcg_put_cache_ids(void)
375 up_read(&memcg_cache_ids_sem);
379 * MIN_SIZE is different than 1, because we would like to avoid going through
380 * the alloc/free process all the time. In a small machine, 4 kmem-limited
381 * cgroups is a reasonable guess. In the future, it could be a parameter or
382 * tunable, but that is strictly not necessary.
384 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
385 * this constant directly from cgroup, but it is understandable that this is
386 * better kept as an internal representation in cgroup.c. In any case, the
387 * cgrp_id space is not getting any smaller, and we don't have to necessarily
388 * increase ours as well if it increases.
390 #define MEMCG_CACHES_MIN_SIZE 4
391 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
394 * A lot of the calls to the cache allocation functions are expected to be
395 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
396 * conditional to this static branch, we'll have to allow modules that does
397 * kmem_cache_alloc and the such to see this symbol as well
399 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
400 EXPORT_SYMBOL(memcg_kmem_enabled_key);
403 static int memcg_shrinker_map_size;
404 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
406 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
408 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
411 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
412 int size, int old_size)
414 struct memcg_shrinker_map *new, *old;
415 struct mem_cgroup_per_node *pn;
418 lockdep_assert_held(&memcg_shrinker_map_mutex);
421 pn = memcg->nodeinfo[nid];
422 old = rcu_dereference_protected(pn->shrinker_map, true);
423 /* Not yet online memcg */
427 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
431 /* Set all old bits, clear all new bits */
432 memset(new->map, (int)0xff, old_size);
433 memset((void *)new->map + old_size, 0, size - old_size);
435 rcu_assign_pointer(pn->shrinker_map, new);
436 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
442 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
444 struct mem_cgroup_per_node *pn;
445 struct memcg_shrinker_map *map;
448 if (mem_cgroup_is_root(memcg))
452 pn = memcg->nodeinfo[nid];
453 map = rcu_dereference_protected(pn->shrinker_map, true);
455 rcu_assign_pointer(pn->shrinker_map, NULL);
459 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
461 struct memcg_shrinker_map *map;
462 int nid, size, ret = 0;
464 if (mem_cgroup_is_root(memcg))
467 mutex_lock(&memcg_shrinker_map_mutex);
468 size = memcg_shrinker_map_size;
470 map = kvzalloc_node(sizeof(*map) + size, GFP_KERNEL, nid);
472 memcg_free_shrinker_maps(memcg);
476 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
478 mutex_unlock(&memcg_shrinker_map_mutex);
483 int memcg_expand_shrinker_maps(int new_id)
485 int size, old_size, ret = 0;
486 struct mem_cgroup *memcg;
488 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
489 old_size = memcg_shrinker_map_size;
490 if (size <= old_size)
493 mutex_lock(&memcg_shrinker_map_mutex);
494 if (!root_mem_cgroup)
497 for_each_mem_cgroup(memcg) {
498 if (mem_cgroup_is_root(memcg))
500 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
502 mem_cgroup_iter_break(NULL, memcg);
508 memcg_shrinker_map_size = size;
509 mutex_unlock(&memcg_shrinker_map_mutex);
513 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
515 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
516 struct memcg_shrinker_map *map;
519 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
520 /* Pairs with smp mb in shrink_slab() */
521 smp_mb__before_atomic();
522 set_bit(shrinker_id, map->map);
528 * mem_cgroup_css_from_page - css of the memcg associated with a page
529 * @page: page of interest
531 * If memcg is bound to the default hierarchy, css of the memcg associated
532 * with @page is returned. The returned css remains associated with @page
533 * until it is released.
535 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
538 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
540 struct mem_cgroup *memcg;
542 memcg = page_memcg(page);
544 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
545 memcg = root_mem_cgroup;
551 * page_cgroup_ino - return inode number of the memcg a page is charged to
554 * Look up the closest online ancestor of the memory cgroup @page is charged to
555 * and return its inode number or 0 if @page is not charged to any cgroup. It
556 * is safe to call this function without holding a reference to @page.
558 * Note, this function is inherently racy, because there is nothing to prevent
559 * the cgroup inode from getting torn down and potentially reallocated a moment
560 * after page_cgroup_ino() returns, so it only should be used by callers that
561 * do not care (such as procfs interfaces).
563 ino_t page_cgroup_ino(struct page *page)
565 struct mem_cgroup *memcg;
566 unsigned long ino = 0;
569 memcg = page_memcg_check(page);
571 while (memcg && !(memcg->css.flags & CSS_ONLINE))
572 memcg = parent_mem_cgroup(memcg);
574 ino = cgroup_ino(memcg->css.cgroup);
579 static struct mem_cgroup_per_node *
580 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
582 int nid = page_to_nid(page);
584 return memcg->nodeinfo[nid];
587 static struct mem_cgroup_tree_per_node *
588 soft_limit_tree_node(int nid)
590 return soft_limit_tree.rb_tree_per_node[nid];
593 static struct mem_cgroup_tree_per_node *
594 soft_limit_tree_from_page(struct page *page)
596 int nid = page_to_nid(page);
598 return soft_limit_tree.rb_tree_per_node[nid];
601 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
602 struct mem_cgroup_tree_per_node *mctz,
603 unsigned long new_usage_in_excess)
605 struct rb_node **p = &mctz->rb_root.rb_node;
606 struct rb_node *parent = NULL;
607 struct mem_cgroup_per_node *mz_node;
608 bool rightmost = true;
613 mz->usage_in_excess = new_usage_in_excess;
614 if (!mz->usage_in_excess)
618 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
620 if (mz->usage_in_excess < mz_node->usage_in_excess) {
629 mctz->rb_rightmost = &mz->tree_node;
631 rb_link_node(&mz->tree_node, parent, p);
632 rb_insert_color(&mz->tree_node, &mctz->rb_root);
636 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
637 struct mem_cgroup_tree_per_node *mctz)
642 if (&mz->tree_node == mctz->rb_rightmost)
643 mctz->rb_rightmost = rb_prev(&mz->tree_node);
645 rb_erase(&mz->tree_node, &mctz->rb_root);
649 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
650 struct mem_cgroup_tree_per_node *mctz)
654 spin_lock_irqsave(&mctz->lock, flags);
655 __mem_cgroup_remove_exceeded(mz, mctz);
656 spin_unlock_irqrestore(&mctz->lock, flags);
659 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
661 unsigned long nr_pages = page_counter_read(&memcg->memory);
662 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
663 unsigned long excess = 0;
665 if (nr_pages > soft_limit)
666 excess = nr_pages - soft_limit;
671 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
673 unsigned long excess;
674 struct mem_cgroup_per_node *mz;
675 struct mem_cgroup_tree_per_node *mctz;
677 mctz = soft_limit_tree_from_page(page);
681 * Necessary to update all ancestors when hierarchy is used.
682 * because their event counter is not touched.
684 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
685 mz = mem_cgroup_page_nodeinfo(memcg, page);
686 excess = soft_limit_excess(memcg);
688 * We have to update the tree if mz is on RB-tree or
689 * mem is over its softlimit.
691 if (excess || mz->on_tree) {
694 spin_lock_irqsave(&mctz->lock, flags);
695 /* if on-tree, remove it */
697 __mem_cgroup_remove_exceeded(mz, mctz);
699 * Insert again. mz->usage_in_excess will be updated.
700 * If excess is 0, no tree ops.
702 __mem_cgroup_insert_exceeded(mz, mctz, excess);
703 spin_unlock_irqrestore(&mctz->lock, flags);
708 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
710 struct mem_cgroup_tree_per_node *mctz;
711 struct mem_cgroup_per_node *mz;
715 mz = memcg->nodeinfo[nid];
716 mctz = soft_limit_tree_node(nid);
718 mem_cgroup_remove_exceeded(mz, mctz);
722 static struct mem_cgroup_per_node *
723 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
725 struct mem_cgroup_per_node *mz;
729 if (!mctz->rb_rightmost)
730 goto done; /* Nothing to reclaim from */
732 mz = rb_entry(mctz->rb_rightmost,
733 struct mem_cgroup_per_node, tree_node);
735 * Remove the node now but someone else can add it back,
736 * we will to add it back at the end of reclaim to its correct
737 * position in the tree.
739 __mem_cgroup_remove_exceeded(mz, mctz);
740 if (!soft_limit_excess(mz->memcg) ||
741 !css_tryget(&mz->memcg->css))
747 static struct mem_cgroup_per_node *
748 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
750 struct mem_cgroup_per_node *mz;
752 spin_lock_irq(&mctz->lock);
753 mz = __mem_cgroup_largest_soft_limit_node(mctz);
754 spin_unlock_irq(&mctz->lock);
759 * __mod_memcg_state - update cgroup memory statistics
760 * @memcg: the memory cgroup
761 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
762 * @val: delta to add to the counter, can be negative
764 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
766 if (mem_cgroup_disabled())
769 __this_cpu_add(memcg->vmstats_percpu->state[idx], val);
770 cgroup_rstat_updated(memcg->css.cgroup, smp_processor_id());
773 /* idx can be of type enum memcg_stat_item or node_stat_item. */
774 static unsigned long memcg_page_state(struct mem_cgroup *memcg, int idx)
776 long x = READ_ONCE(memcg->vmstats.state[idx]);
784 /* idx can be of type enum memcg_stat_item or node_stat_item. */
785 static unsigned long memcg_page_state_local(struct mem_cgroup *memcg, int idx)
790 for_each_possible_cpu(cpu)
791 x += per_cpu(memcg->vmstats_percpu->state[idx], cpu);
799 static struct mem_cgroup_per_node *
800 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
802 struct mem_cgroup *parent;
804 parent = parent_mem_cgroup(pn->memcg);
807 return parent->nodeinfo[nid];
810 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
813 struct mem_cgroup_per_node *pn;
814 struct mem_cgroup *memcg;
815 long x, threshold = MEMCG_CHARGE_BATCH;
817 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
821 __mod_memcg_state(memcg, idx, val);
824 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
826 if (vmstat_item_in_bytes(idx))
827 threshold <<= PAGE_SHIFT;
829 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
830 if (unlikely(abs(x) > threshold)) {
831 pg_data_t *pgdat = lruvec_pgdat(lruvec);
832 struct mem_cgroup_per_node *pi;
834 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
835 atomic_long_add(x, &pi->lruvec_stat[idx]);
838 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
842 * __mod_lruvec_state - update lruvec memory statistics
843 * @lruvec: the lruvec
844 * @idx: the stat item
845 * @val: delta to add to the counter, can be negative
847 * The lruvec is the intersection of the NUMA node and a cgroup. This
848 * function updates the all three counters that are affected by a
849 * change of state at this level: per-node, per-cgroup, per-lruvec.
851 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
855 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
857 /* Update memcg and lruvec */
858 if (!mem_cgroup_disabled())
859 __mod_memcg_lruvec_state(lruvec, idx, val);
862 void __mod_lruvec_page_state(struct page *page, enum node_stat_item idx,
865 struct page *head = compound_head(page); /* rmap on tail pages */
866 struct mem_cgroup *memcg;
867 pg_data_t *pgdat = page_pgdat(page);
868 struct lruvec *lruvec;
871 memcg = page_memcg(head);
872 /* Untracked pages have no memcg, no lruvec. Update only the node */
875 __mod_node_page_state(pgdat, idx, val);
879 lruvec = mem_cgroup_lruvec(memcg, pgdat);
880 __mod_lruvec_state(lruvec, idx, val);
883 EXPORT_SYMBOL(__mod_lruvec_page_state);
885 void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val)
887 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
888 struct mem_cgroup *memcg;
889 struct lruvec *lruvec;
892 memcg = mem_cgroup_from_obj(p);
895 * Untracked pages have no memcg, no lruvec. Update only the
896 * node. If we reparent the slab objects to the root memcg,
897 * when we free the slab object, we need to update the per-memcg
898 * vmstats to keep it correct for the root memcg.
901 __mod_node_page_state(pgdat, idx, val);
903 lruvec = mem_cgroup_lruvec(memcg, pgdat);
904 __mod_lruvec_state(lruvec, idx, val);
910 * __count_memcg_events - account VM events in a cgroup
911 * @memcg: the memory cgroup
912 * @idx: the event item
913 * @count: the number of events that occured
915 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
918 if (mem_cgroup_disabled())
921 __this_cpu_add(memcg->vmstats_percpu->events[idx], count);
922 cgroup_rstat_updated(memcg->css.cgroup, smp_processor_id());
925 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
927 return READ_ONCE(memcg->vmstats.events[event]);
930 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
935 for_each_possible_cpu(cpu)
936 x += per_cpu(memcg->vmstats_percpu->events[event], cpu);
940 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
944 /* pagein of a big page is an event. So, ignore page size */
946 __count_memcg_events(memcg, PGPGIN, 1);
948 __count_memcg_events(memcg, PGPGOUT, 1);
949 nr_pages = -nr_pages; /* for event */
952 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
955 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
956 enum mem_cgroup_events_target target)
958 unsigned long val, next;
960 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
961 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
962 /* from time_after() in jiffies.h */
963 if ((long)(next - val) < 0) {
965 case MEM_CGROUP_TARGET_THRESH:
966 next = val + THRESHOLDS_EVENTS_TARGET;
968 case MEM_CGROUP_TARGET_SOFTLIMIT:
969 next = val + SOFTLIMIT_EVENTS_TARGET;
974 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
981 * Check events in order.
984 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
986 /* threshold event is triggered in finer grain than soft limit */
987 if (unlikely(mem_cgroup_event_ratelimit(memcg,
988 MEM_CGROUP_TARGET_THRESH))) {
991 do_softlimit = mem_cgroup_event_ratelimit(memcg,
992 MEM_CGROUP_TARGET_SOFTLIMIT);
993 mem_cgroup_threshold(memcg);
994 if (unlikely(do_softlimit))
995 mem_cgroup_update_tree(memcg, page);
999 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1002 * mm_update_next_owner() may clear mm->owner to NULL
1003 * if it races with swapoff, page migration, etc.
1004 * So this can be called with p == NULL.
1009 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1011 EXPORT_SYMBOL(mem_cgroup_from_task);
1014 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1015 * @mm: mm from which memcg should be extracted. It can be NULL.
1017 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
1018 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
1021 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1023 struct mem_cgroup *memcg;
1025 if (mem_cgroup_disabled())
1031 * Page cache insertions can happen withou an
1032 * actual mm context, e.g. during disk probing
1033 * on boot, loopback IO, acct() writes etc.
1036 memcg = root_mem_cgroup;
1038 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1039 if (unlikely(!memcg))
1040 memcg = root_mem_cgroup;
1042 } while (!css_tryget(&memcg->css));
1046 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1048 static __always_inline struct mem_cgroup *active_memcg(void)
1051 return this_cpu_read(int_active_memcg);
1053 return current->active_memcg;
1056 static __always_inline bool memcg_kmem_bypass(void)
1058 /* Allow remote memcg charging from any context. */
1059 if (unlikely(active_memcg()))
1062 /* Memcg to charge can't be determined. */
1063 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
1070 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1071 * @root: hierarchy root
1072 * @prev: previously returned memcg, NULL on first invocation
1073 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1075 * Returns references to children of the hierarchy below @root, or
1076 * @root itself, or %NULL after a full round-trip.
1078 * Caller must pass the return value in @prev on subsequent
1079 * invocations for reference counting, or use mem_cgroup_iter_break()
1080 * to cancel a hierarchy walk before the round-trip is complete.
1082 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1083 * in the hierarchy among all concurrent reclaimers operating on the
1086 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1087 struct mem_cgroup *prev,
1088 struct mem_cgroup_reclaim_cookie *reclaim)
1090 struct mem_cgroup_reclaim_iter *iter;
1091 struct cgroup_subsys_state *css = NULL;
1092 struct mem_cgroup *memcg = NULL;
1093 struct mem_cgroup *pos = NULL;
1095 if (mem_cgroup_disabled())
1099 root = root_mem_cgroup;
1101 if (prev && !reclaim)
1107 struct mem_cgroup_per_node *mz;
1109 mz = root->nodeinfo[reclaim->pgdat->node_id];
1112 if (prev && reclaim->generation != iter->generation)
1116 pos = READ_ONCE(iter->position);
1117 if (!pos || css_tryget(&pos->css))
1120 * css reference reached zero, so iter->position will
1121 * be cleared by ->css_released. However, we should not
1122 * rely on this happening soon, because ->css_released
1123 * is called from a work queue, and by busy-waiting we
1124 * might block it. So we clear iter->position right
1127 (void)cmpxchg(&iter->position, pos, NULL);
1135 css = css_next_descendant_pre(css, &root->css);
1138 * Reclaimers share the hierarchy walk, and a
1139 * new one might jump in right at the end of
1140 * the hierarchy - make sure they see at least
1141 * one group and restart from the beginning.
1149 * Verify the css and acquire a reference. The root
1150 * is provided by the caller, so we know it's alive
1151 * and kicking, and don't take an extra reference.
1153 memcg = mem_cgroup_from_css(css);
1155 if (css == &root->css)
1158 if (css_tryget(css))
1166 * The position could have already been updated by a competing
1167 * thread, so check that the value hasn't changed since we read
1168 * it to avoid reclaiming from the same cgroup twice.
1170 (void)cmpxchg(&iter->position, pos, memcg);
1178 reclaim->generation = iter->generation;
1183 if (prev && prev != root)
1184 css_put(&prev->css);
1190 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1191 * @root: hierarchy root
1192 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1194 void mem_cgroup_iter_break(struct mem_cgroup *root,
1195 struct mem_cgroup *prev)
1198 root = root_mem_cgroup;
1199 if (prev && prev != root)
1200 css_put(&prev->css);
1203 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1204 struct mem_cgroup *dead_memcg)
1206 struct mem_cgroup_reclaim_iter *iter;
1207 struct mem_cgroup_per_node *mz;
1210 for_each_node(nid) {
1211 mz = from->nodeinfo[nid];
1213 cmpxchg(&iter->position, dead_memcg, NULL);
1217 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1219 struct mem_cgroup *memcg = dead_memcg;
1220 struct mem_cgroup *last;
1223 __invalidate_reclaim_iterators(memcg, dead_memcg);
1225 } while ((memcg = parent_mem_cgroup(memcg)));
1228 * When cgruop1 non-hierarchy mode is used,
1229 * parent_mem_cgroup() does not walk all the way up to the
1230 * cgroup root (root_mem_cgroup). So we have to handle
1231 * dead_memcg from cgroup root separately.
1233 if (last != root_mem_cgroup)
1234 __invalidate_reclaim_iterators(root_mem_cgroup,
1239 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1240 * @memcg: hierarchy root
1241 * @fn: function to call for each task
1242 * @arg: argument passed to @fn
1244 * This function iterates over tasks attached to @memcg or to any of its
1245 * descendants and calls @fn for each task. If @fn returns a non-zero
1246 * value, the function breaks the iteration loop and returns the value.
1247 * Otherwise, it will iterate over all tasks and return 0.
1249 * This function must not be called for the root memory cgroup.
1251 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1252 int (*fn)(struct task_struct *, void *), void *arg)
1254 struct mem_cgroup *iter;
1257 BUG_ON(memcg == root_mem_cgroup);
1259 for_each_mem_cgroup_tree(iter, memcg) {
1260 struct css_task_iter it;
1261 struct task_struct *task;
1263 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1264 while (!ret && (task = css_task_iter_next(&it)))
1265 ret = fn(task, arg);
1266 css_task_iter_end(&it);
1268 mem_cgroup_iter_break(memcg, iter);
1275 #ifdef CONFIG_DEBUG_VM
1276 void lruvec_memcg_debug(struct lruvec *lruvec, struct page *page)
1278 struct mem_cgroup *memcg;
1280 if (mem_cgroup_disabled())
1283 memcg = page_memcg(page);
1286 VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != root_mem_cgroup, page);
1288 VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != memcg, page);
1293 * lock_page_lruvec - lock and return lruvec for a given page.
1296 * These functions are safe to use under any of the following conditions:
1299 * - lock_page_memcg()
1300 * - page->_refcount is zero
1302 struct lruvec *lock_page_lruvec(struct page *page)
1304 struct lruvec *lruvec;
1305 struct pglist_data *pgdat = page_pgdat(page);
1307 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1308 spin_lock(&lruvec->lru_lock);
1310 lruvec_memcg_debug(lruvec, page);
1315 struct lruvec *lock_page_lruvec_irq(struct page *page)
1317 struct lruvec *lruvec;
1318 struct pglist_data *pgdat = page_pgdat(page);
1320 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1321 spin_lock_irq(&lruvec->lru_lock);
1323 lruvec_memcg_debug(lruvec, page);
1328 struct lruvec *lock_page_lruvec_irqsave(struct page *page, unsigned long *flags)
1330 struct lruvec *lruvec;
1331 struct pglist_data *pgdat = page_pgdat(page);
1333 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1334 spin_lock_irqsave(&lruvec->lru_lock, *flags);
1336 lruvec_memcg_debug(lruvec, page);
1342 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1343 * @lruvec: mem_cgroup per zone lru vector
1344 * @lru: index of lru list the page is sitting on
1345 * @zid: zone id of the accounted pages
1346 * @nr_pages: positive when adding or negative when removing
1348 * This function must be called under lru_lock, just before a page is added
1349 * to or just after a page is removed from an lru list (that ordering being
1350 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1352 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1353 int zid, int nr_pages)
1355 struct mem_cgroup_per_node *mz;
1356 unsigned long *lru_size;
1359 if (mem_cgroup_disabled())
1362 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1363 lru_size = &mz->lru_zone_size[zid][lru];
1366 *lru_size += nr_pages;
1369 if (WARN_ONCE(size < 0,
1370 "%s(%p, %d, %d): lru_size %ld\n",
1371 __func__, lruvec, lru, nr_pages, size)) {
1377 *lru_size += nr_pages;
1381 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1382 * @memcg: the memory cgroup
1384 * Returns the maximum amount of memory @mem can be charged with, in
1387 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1389 unsigned long margin = 0;
1390 unsigned long count;
1391 unsigned long limit;
1393 count = page_counter_read(&memcg->memory);
1394 limit = READ_ONCE(memcg->memory.max);
1396 margin = limit - count;
1398 if (do_memsw_account()) {
1399 count = page_counter_read(&memcg->memsw);
1400 limit = READ_ONCE(memcg->memsw.max);
1402 margin = min(margin, limit - count);
1411 * A routine for checking "mem" is under move_account() or not.
1413 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1414 * moving cgroups. This is for waiting at high-memory pressure
1417 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1419 struct mem_cgroup *from;
1420 struct mem_cgroup *to;
1423 * Unlike task_move routines, we access mc.to, mc.from not under
1424 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1426 spin_lock(&mc.lock);
1432 ret = mem_cgroup_is_descendant(from, memcg) ||
1433 mem_cgroup_is_descendant(to, memcg);
1435 spin_unlock(&mc.lock);
1439 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1441 if (mc.moving_task && current != mc.moving_task) {
1442 if (mem_cgroup_under_move(memcg)) {
1444 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1445 /* moving charge context might have finished. */
1448 finish_wait(&mc.waitq, &wait);
1455 struct memory_stat {
1460 static const struct memory_stat memory_stats[] = {
1461 { "anon", NR_ANON_MAPPED },
1462 { "file", NR_FILE_PAGES },
1463 { "kernel_stack", NR_KERNEL_STACK_KB },
1464 { "pagetables", NR_PAGETABLE },
1465 { "percpu", MEMCG_PERCPU_B },
1466 { "sock", MEMCG_SOCK },
1467 { "shmem", NR_SHMEM },
1468 { "file_mapped", NR_FILE_MAPPED },
1469 { "file_dirty", NR_FILE_DIRTY },
1470 { "file_writeback", NR_WRITEBACK },
1472 { "swapcached", NR_SWAPCACHE },
1474 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1475 { "anon_thp", NR_ANON_THPS },
1476 { "file_thp", NR_FILE_THPS },
1477 { "shmem_thp", NR_SHMEM_THPS },
1479 { "inactive_anon", NR_INACTIVE_ANON },
1480 { "active_anon", NR_ACTIVE_ANON },
1481 { "inactive_file", NR_INACTIVE_FILE },
1482 { "active_file", NR_ACTIVE_FILE },
1483 { "unevictable", NR_UNEVICTABLE },
1484 { "slab_reclaimable", NR_SLAB_RECLAIMABLE_B },
1485 { "slab_unreclaimable", NR_SLAB_UNRECLAIMABLE_B },
1487 /* The memory events */
1488 { "workingset_refault_anon", WORKINGSET_REFAULT_ANON },
1489 { "workingset_refault_file", WORKINGSET_REFAULT_FILE },
1490 { "workingset_activate_anon", WORKINGSET_ACTIVATE_ANON },
1491 { "workingset_activate_file", WORKINGSET_ACTIVATE_FILE },
1492 { "workingset_restore_anon", WORKINGSET_RESTORE_ANON },
1493 { "workingset_restore_file", WORKINGSET_RESTORE_FILE },
1494 { "workingset_nodereclaim", WORKINGSET_NODERECLAIM },
1497 /* Translate stat items to the correct unit for memory.stat output */
1498 static int memcg_page_state_unit(int item)
1501 case MEMCG_PERCPU_B:
1502 case NR_SLAB_RECLAIMABLE_B:
1503 case NR_SLAB_UNRECLAIMABLE_B:
1504 case WORKINGSET_REFAULT_ANON:
1505 case WORKINGSET_REFAULT_FILE:
1506 case WORKINGSET_ACTIVATE_ANON:
1507 case WORKINGSET_ACTIVATE_FILE:
1508 case WORKINGSET_RESTORE_ANON:
1509 case WORKINGSET_RESTORE_FILE:
1510 case WORKINGSET_NODERECLAIM:
1512 case NR_KERNEL_STACK_KB:
1519 static inline unsigned long memcg_page_state_output(struct mem_cgroup *memcg,
1522 return memcg_page_state(memcg, item) * memcg_page_state_unit(item);
1525 static char *memory_stat_format(struct mem_cgroup *memcg)
1530 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1535 * Provide statistics on the state of the memory subsystem as
1536 * well as cumulative event counters that show past behavior.
1538 * This list is ordered following a combination of these gradients:
1539 * 1) generic big picture -> specifics and details
1540 * 2) reflecting userspace activity -> reflecting kernel heuristics
1542 * Current memory state:
1544 cgroup_rstat_flush(memcg->css.cgroup);
1546 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1549 size = memcg_page_state_output(memcg, memory_stats[i].idx);
1550 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1552 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1553 size += memcg_page_state_output(memcg,
1554 NR_SLAB_RECLAIMABLE_B);
1555 seq_buf_printf(&s, "slab %llu\n", size);
1559 /* Accumulated memory events */
1561 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1562 memcg_events(memcg, PGFAULT));
1563 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1564 memcg_events(memcg, PGMAJFAULT));
1565 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1566 memcg_events(memcg, PGREFILL));
1567 seq_buf_printf(&s, "pgscan %lu\n",
1568 memcg_events(memcg, PGSCAN_KSWAPD) +
1569 memcg_events(memcg, PGSCAN_DIRECT));
1570 seq_buf_printf(&s, "pgsteal %lu\n",
1571 memcg_events(memcg, PGSTEAL_KSWAPD) +
1572 memcg_events(memcg, PGSTEAL_DIRECT));
1573 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1574 memcg_events(memcg, PGACTIVATE));
1575 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1576 memcg_events(memcg, PGDEACTIVATE));
1577 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1578 memcg_events(memcg, PGLAZYFREE));
1579 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1580 memcg_events(memcg, PGLAZYFREED));
1582 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1583 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1584 memcg_events(memcg, THP_FAULT_ALLOC));
1585 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1586 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1587 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1589 /* The above should easily fit into one page */
1590 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1595 #define K(x) ((x) << (PAGE_SHIFT-10))
1597 * mem_cgroup_print_oom_context: Print OOM information relevant to
1598 * memory controller.
1599 * @memcg: The memory cgroup that went over limit
1600 * @p: Task that is going to be killed
1602 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1605 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1610 pr_cont(",oom_memcg=");
1611 pr_cont_cgroup_path(memcg->css.cgroup);
1613 pr_cont(",global_oom");
1615 pr_cont(",task_memcg=");
1616 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1622 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1623 * memory controller.
1624 * @memcg: The memory cgroup that went over limit
1626 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1630 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1631 K((u64)page_counter_read(&memcg->memory)),
1632 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1633 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1634 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1635 K((u64)page_counter_read(&memcg->swap)),
1636 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1638 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1639 K((u64)page_counter_read(&memcg->memsw)),
1640 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1641 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1642 K((u64)page_counter_read(&memcg->kmem)),
1643 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1646 pr_info("Memory cgroup stats for ");
1647 pr_cont_cgroup_path(memcg->css.cgroup);
1649 buf = memory_stat_format(memcg);
1657 * Return the memory (and swap, if configured) limit for a memcg.
1659 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1661 unsigned long max = READ_ONCE(memcg->memory.max);
1663 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1664 if (mem_cgroup_swappiness(memcg))
1665 max += min(READ_ONCE(memcg->swap.max),
1666 (unsigned long)total_swap_pages);
1668 if (mem_cgroup_swappiness(memcg)) {
1669 /* Calculate swap excess capacity from memsw limit */
1670 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1672 max += min(swap, (unsigned long)total_swap_pages);
1678 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1680 return page_counter_read(&memcg->memory);
1683 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1686 struct oom_control oc = {
1690 .gfp_mask = gfp_mask,
1695 if (mutex_lock_killable(&oom_lock))
1698 if (mem_cgroup_margin(memcg) >= (1 << order))
1702 * A few threads which were not waiting at mutex_lock_killable() can
1703 * fail to bail out. Therefore, check again after holding oom_lock.
1705 ret = should_force_charge() || out_of_memory(&oc);
1708 mutex_unlock(&oom_lock);
1712 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1715 unsigned long *total_scanned)
1717 struct mem_cgroup *victim = NULL;
1720 unsigned long excess;
1721 unsigned long nr_scanned;
1722 struct mem_cgroup_reclaim_cookie reclaim = {
1726 excess = soft_limit_excess(root_memcg);
1729 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1734 * If we have not been able to reclaim
1735 * anything, it might because there are
1736 * no reclaimable pages under this hierarchy
1741 * We want to do more targeted reclaim.
1742 * excess >> 2 is not to excessive so as to
1743 * reclaim too much, nor too less that we keep
1744 * coming back to reclaim from this cgroup
1746 if (total >= (excess >> 2) ||
1747 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1752 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1753 pgdat, &nr_scanned);
1754 *total_scanned += nr_scanned;
1755 if (!soft_limit_excess(root_memcg))
1758 mem_cgroup_iter_break(root_memcg, victim);
1762 #ifdef CONFIG_LOCKDEP
1763 static struct lockdep_map memcg_oom_lock_dep_map = {
1764 .name = "memcg_oom_lock",
1768 static DEFINE_SPINLOCK(memcg_oom_lock);
1771 * Check OOM-Killer is already running under our hierarchy.
1772 * If someone is running, return false.
1774 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1776 struct mem_cgroup *iter, *failed = NULL;
1778 spin_lock(&memcg_oom_lock);
1780 for_each_mem_cgroup_tree(iter, memcg) {
1781 if (iter->oom_lock) {
1783 * this subtree of our hierarchy is already locked
1784 * so we cannot give a lock.
1787 mem_cgroup_iter_break(memcg, iter);
1790 iter->oom_lock = true;
1795 * OK, we failed to lock the whole subtree so we have
1796 * to clean up what we set up to the failing subtree
1798 for_each_mem_cgroup_tree(iter, memcg) {
1799 if (iter == failed) {
1800 mem_cgroup_iter_break(memcg, iter);
1803 iter->oom_lock = false;
1806 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1808 spin_unlock(&memcg_oom_lock);
1813 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1815 struct mem_cgroup *iter;
1817 spin_lock(&memcg_oom_lock);
1818 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1819 for_each_mem_cgroup_tree(iter, memcg)
1820 iter->oom_lock = false;
1821 spin_unlock(&memcg_oom_lock);
1824 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1826 struct mem_cgroup *iter;
1828 spin_lock(&memcg_oom_lock);
1829 for_each_mem_cgroup_tree(iter, memcg)
1831 spin_unlock(&memcg_oom_lock);
1834 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1836 struct mem_cgroup *iter;
1839 * Be careful about under_oom underflows becase a child memcg
1840 * could have been added after mem_cgroup_mark_under_oom.
1842 spin_lock(&memcg_oom_lock);
1843 for_each_mem_cgroup_tree(iter, memcg)
1844 if (iter->under_oom > 0)
1846 spin_unlock(&memcg_oom_lock);
1849 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1851 struct oom_wait_info {
1852 struct mem_cgroup *memcg;
1853 wait_queue_entry_t wait;
1856 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1857 unsigned mode, int sync, void *arg)
1859 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1860 struct mem_cgroup *oom_wait_memcg;
1861 struct oom_wait_info *oom_wait_info;
1863 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1864 oom_wait_memcg = oom_wait_info->memcg;
1866 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1867 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1869 return autoremove_wake_function(wait, mode, sync, arg);
1872 static void memcg_oom_recover(struct mem_cgroup *memcg)
1875 * For the following lockless ->under_oom test, the only required
1876 * guarantee is that it must see the state asserted by an OOM when
1877 * this function is called as a result of userland actions
1878 * triggered by the notification of the OOM. This is trivially
1879 * achieved by invoking mem_cgroup_mark_under_oom() before
1880 * triggering notification.
1882 if (memcg && memcg->under_oom)
1883 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1893 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1895 enum oom_status ret;
1898 if (order > PAGE_ALLOC_COSTLY_ORDER)
1901 memcg_memory_event(memcg, MEMCG_OOM);
1904 * We are in the middle of the charge context here, so we
1905 * don't want to block when potentially sitting on a callstack
1906 * that holds all kinds of filesystem and mm locks.
1908 * cgroup1 allows disabling the OOM killer and waiting for outside
1909 * handling until the charge can succeed; remember the context and put
1910 * the task to sleep at the end of the page fault when all locks are
1913 * On the other hand, in-kernel OOM killer allows for an async victim
1914 * memory reclaim (oom_reaper) and that means that we are not solely
1915 * relying on the oom victim to make a forward progress and we can
1916 * invoke the oom killer here.
1918 * Please note that mem_cgroup_out_of_memory might fail to find a
1919 * victim and then we have to bail out from the charge path.
1921 if (memcg->oom_kill_disable) {
1922 if (!current->in_user_fault)
1924 css_get(&memcg->css);
1925 current->memcg_in_oom = memcg;
1926 current->memcg_oom_gfp_mask = mask;
1927 current->memcg_oom_order = order;
1932 mem_cgroup_mark_under_oom(memcg);
1934 locked = mem_cgroup_oom_trylock(memcg);
1937 mem_cgroup_oom_notify(memcg);
1939 mem_cgroup_unmark_under_oom(memcg);
1940 if (mem_cgroup_out_of_memory(memcg, mask, order))
1946 mem_cgroup_oom_unlock(memcg);
1952 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1953 * @handle: actually kill/wait or just clean up the OOM state
1955 * This has to be called at the end of a page fault if the memcg OOM
1956 * handler was enabled.
1958 * Memcg supports userspace OOM handling where failed allocations must
1959 * sleep on a waitqueue until the userspace task resolves the
1960 * situation. Sleeping directly in the charge context with all kinds
1961 * of locks held is not a good idea, instead we remember an OOM state
1962 * in the task and mem_cgroup_oom_synchronize() has to be called at
1963 * the end of the page fault to complete the OOM handling.
1965 * Returns %true if an ongoing memcg OOM situation was detected and
1966 * completed, %false otherwise.
1968 bool mem_cgroup_oom_synchronize(bool handle)
1970 struct mem_cgroup *memcg = current->memcg_in_oom;
1971 struct oom_wait_info owait;
1974 /* OOM is global, do not handle */
1981 owait.memcg = memcg;
1982 owait.wait.flags = 0;
1983 owait.wait.func = memcg_oom_wake_function;
1984 owait.wait.private = current;
1985 INIT_LIST_HEAD(&owait.wait.entry);
1987 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1988 mem_cgroup_mark_under_oom(memcg);
1990 locked = mem_cgroup_oom_trylock(memcg);
1993 mem_cgroup_oom_notify(memcg);
1995 if (locked && !memcg->oom_kill_disable) {
1996 mem_cgroup_unmark_under_oom(memcg);
1997 finish_wait(&memcg_oom_waitq, &owait.wait);
1998 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1999 current->memcg_oom_order);
2002 mem_cgroup_unmark_under_oom(memcg);
2003 finish_wait(&memcg_oom_waitq, &owait.wait);
2007 mem_cgroup_oom_unlock(memcg);
2009 * There is no guarantee that an OOM-lock contender
2010 * sees the wakeups triggered by the OOM kill
2011 * uncharges. Wake any sleepers explicitely.
2013 memcg_oom_recover(memcg);
2016 current->memcg_in_oom = NULL;
2017 css_put(&memcg->css);
2022 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2023 * @victim: task to be killed by the OOM killer
2024 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2026 * Returns a pointer to a memory cgroup, which has to be cleaned up
2027 * by killing all belonging OOM-killable tasks.
2029 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2031 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2032 struct mem_cgroup *oom_domain)
2034 struct mem_cgroup *oom_group = NULL;
2035 struct mem_cgroup *memcg;
2037 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2041 oom_domain = root_mem_cgroup;
2045 memcg = mem_cgroup_from_task(victim);
2046 if (memcg == root_mem_cgroup)
2050 * If the victim task has been asynchronously moved to a different
2051 * memory cgroup, we might end up killing tasks outside oom_domain.
2052 * In this case it's better to ignore memory.group.oom.
2054 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2058 * Traverse the memory cgroup hierarchy from the victim task's
2059 * cgroup up to the OOMing cgroup (or root) to find the
2060 * highest-level memory cgroup with oom.group set.
2062 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2063 if (memcg->oom_group)
2066 if (memcg == oom_domain)
2071 css_get(&oom_group->css);
2078 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2080 pr_info("Tasks in ");
2081 pr_cont_cgroup_path(memcg->css.cgroup);
2082 pr_cont(" are going to be killed due to memory.oom.group set\n");
2086 * lock_page_memcg - lock a page and memcg binding
2089 * This function protects unlocked LRU pages from being moved to
2092 * It ensures lifetime of the locked memcg. Caller is responsible
2093 * for the lifetime of the page.
2095 void lock_page_memcg(struct page *page)
2097 struct page *head = compound_head(page); /* rmap on tail pages */
2098 struct mem_cgroup *memcg;
2099 unsigned long flags;
2102 * The RCU lock is held throughout the transaction. The fast
2103 * path can get away without acquiring the memcg->move_lock
2104 * because page moving starts with an RCU grace period.
2108 if (mem_cgroup_disabled())
2111 memcg = page_memcg(head);
2112 if (unlikely(!memcg))
2115 #ifdef CONFIG_PROVE_LOCKING
2116 local_irq_save(flags);
2117 might_lock(&memcg->move_lock);
2118 local_irq_restore(flags);
2121 if (atomic_read(&memcg->moving_account) <= 0)
2124 spin_lock_irqsave(&memcg->move_lock, flags);
2125 if (memcg != page_memcg(head)) {
2126 spin_unlock_irqrestore(&memcg->move_lock, flags);
2131 * When charge migration first begins, we can have multiple
2132 * critical sections holding the fast-path RCU lock and one
2133 * holding the slowpath move_lock. Track the task who has the
2134 * move_lock for unlock_page_memcg().
2136 memcg->move_lock_task = current;
2137 memcg->move_lock_flags = flags;
2139 EXPORT_SYMBOL(lock_page_memcg);
2141 static void __unlock_page_memcg(struct mem_cgroup *memcg)
2143 if (memcg && memcg->move_lock_task == current) {
2144 unsigned long flags = memcg->move_lock_flags;
2146 memcg->move_lock_task = NULL;
2147 memcg->move_lock_flags = 0;
2149 spin_unlock_irqrestore(&memcg->move_lock, flags);
2156 * unlock_page_memcg - unlock a page and memcg binding
2159 void unlock_page_memcg(struct page *page)
2161 struct page *head = compound_head(page);
2163 __unlock_page_memcg(page_memcg(head));
2165 EXPORT_SYMBOL(unlock_page_memcg);
2167 struct memcg_stock_pcp {
2168 struct mem_cgroup *cached; /* this never be root cgroup */
2169 unsigned int nr_pages;
2171 #ifdef CONFIG_MEMCG_KMEM
2172 struct obj_cgroup *cached_objcg;
2173 unsigned int nr_bytes;
2176 struct work_struct work;
2177 unsigned long flags;
2178 #define FLUSHING_CACHED_CHARGE 0
2180 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2181 static DEFINE_MUTEX(percpu_charge_mutex);
2183 #ifdef CONFIG_MEMCG_KMEM
2184 static void drain_obj_stock(struct memcg_stock_pcp *stock);
2185 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2186 struct mem_cgroup *root_memcg);
2189 static inline void drain_obj_stock(struct memcg_stock_pcp *stock)
2192 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2193 struct mem_cgroup *root_memcg)
2200 * consume_stock: Try to consume stocked charge on this cpu.
2201 * @memcg: memcg to consume from.
2202 * @nr_pages: how many pages to charge.
2204 * The charges will only happen if @memcg matches the current cpu's memcg
2205 * stock, and at least @nr_pages are available in that stock. Failure to
2206 * service an allocation will refill the stock.
2208 * returns true if successful, false otherwise.
2210 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2212 struct memcg_stock_pcp *stock;
2213 unsigned long flags;
2216 if (nr_pages > MEMCG_CHARGE_BATCH)
2219 local_irq_save(flags);
2221 stock = this_cpu_ptr(&memcg_stock);
2222 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2223 stock->nr_pages -= nr_pages;
2227 local_irq_restore(flags);
2233 * Returns stocks cached in percpu and reset cached information.
2235 static void drain_stock(struct memcg_stock_pcp *stock)
2237 struct mem_cgroup *old = stock->cached;
2242 if (stock->nr_pages) {
2243 page_counter_uncharge(&old->memory, stock->nr_pages);
2244 if (do_memsw_account())
2245 page_counter_uncharge(&old->memsw, stock->nr_pages);
2246 stock->nr_pages = 0;
2250 stock->cached = NULL;
2253 static void drain_local_stock(struct work_struct *dummy)
2255 struct memcg_stock_pcp *stock;
2256 unsigned long flags;
2259 * The only protection from memory hotplug vs. drain_stock races is
2260 * that we always operate on local CPU stock here with IRQ disabled
2262 local_irq_save(flags);
2264 stock = this_cpu_ptr(&memcg_stock);
2265 drain_obj_stock(stock);
2267 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2269 local_irq_restore(flags);
2273 * Cache charges(val) to local per_cpu area.
2274 * This will be consumed by consume_stock() function, later.
2276 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2278 struct memcg_stock_pcp *stock;
2279 unsigned long flags;
2281 local_irq_save(flags);
2283 stock = this_cpu_ptr(&memcg_stock);
2284 if (stock->cached != memcg) { /* reset if necessary */
2286 css_get(&memcg->css);
2287 stock->cached = memcg;
2289 stock->nr_pages += nr_pages;
2291 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2294 local_irq_restore(flags);
2298 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2299 * of the hierarchy under it.
2301 static void drain_all_stock(struct mem_cgroup *root_memcg)
2305 /* If someone's already draining, avoid adding running more workers. */
2306 if (!mutex_trylock(&percpu_charge_mutex))
2309 * Notify other cpus that system-wide "drain" is running
2310 * We do not care about races with the cpu hotplug because cpu down
2311 * as well as workers from this path always operate on the local
2312 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2315 for_each_online_cpu(cpu) {
2316 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2317 struct mem_cgroup *memcg;
2321 memcg = stock->cached;
2322 if (memcg && stock->nr_pages &&
2323 mem_cgroup_is_descendant(memcg, root_memcg))
2325 if (obj_stock_flush_required(stock, root_memcg))
2330 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2332 drain_local_stock(&stock->work);
2334 schedule_work_on(cpu, &stock->work);
2338 mutex_unlock(&percpu_charge_mutex);
2341 static void memcg_flush_lruvec_page_state(struct mem_cgroup *memcg, int cpu)
2345 for_each_node(nid) {
2346 struct mem_cgroup_per_node *pn = memcg->nodeinfo[nid];
2347 unsigned long stat[NR_VM_NODE_STAT_ITEMS];
2348 struct batched_lruvec_stat *lstatc;
2351 lstatc = per_cpu_ptr(pn->lruvec_stat_cpu, cpu);
2352 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
2353 stat[i] = lstatc->count[i];
2354 lstatc->count[i] = 0;
2358 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
2359 atomic_long_add(stat[i], &pn->lruvec_stat[i]);
2360 } while ((pn = parent_nodeinfo(pn, nid)));
2364 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2366 struct memcg_stock_pcp *stock;
2367 struct mem_cgroup *memcg;
2369 stock = &per_cpu(memcg_stock, cpu);
2372 for_each_mem_cgroup(memcg)
2373 memcg_flush_lruvec_page_state(memcg, cpu);
2378 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2379 unsigned int nr_pages,
2382 unsigned long nr_reclaimed = 0;
2385 unsigned long pflags;
2387 if (page_counter_read(&memcg->memory) <=
2388 READ_ONCE(memcg->memory.high))
2391 memcg_memory_event(memcg, MEMCG_HIGH);
2393 psi_memstall_enter(&pflags);
2394 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2396 psi_memstall_leave(&pflags);
2397 } while ((memcg = parent_mem_cgroup(memcg)) &&
2398 !mem_cgroup_is_root(memcg));
2400 return nr_reclaimed;
2403 static void high_work_func(struct work_struct *work)
2405 struct mem_cgroup *memcg;
2407 memcg = container_of(work, struct mem_cgroup, high_work);
2408 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2412 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2413 * enough to still cause a significant slowdown in most cases, while still
2414 * allowing diagnostics and tracing to proceed without becoming stuck.
2416 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2419 * When calculating the delay, we use these either side of the exponentiation to
2420 * maintain precision and scale to a reasonable number of jiffies (see the table
2423 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2424 * overage ratio to a delay.
2425 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2426 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2427 * to produce a reasonable delay curve.
2429 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2430 * reasonable delay curve compared to precision-adjusted overage, not
2431 * penalising heavily at first, but still making sure that growth beyond the
2432 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2433 * example, with a high of 100 megabytes:
2435 * +-------+------------------------+
2436 * | usage | time to allocate in ms |
2437 * +-------+------------------------+
2459 * +-------+------------------------+
2461 #define MEMCG_DELAY_PRECISION_SHIFT 20
2462 #define MEMCG_DELAY_SCALING_SHIFT 14
2464 static u64 calculate_overage(unsigned long usage, unsigned long high)
2472 * Prevent division by 0 in overage calculation by acting as if
2473 * it was a threshold of 1 page
2475 high = max(high, 1UL);
2477 overage = usage - high;
2478 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2479 return div64_u64(overage, high);
2482 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2484 u64 overage, max_overage = 0;
2487 overage = calculate_overage(page_counter_read(&memcg->memory),
2488 READ_ONCE(memcg->memory.high));
2489 max_overage = max(overage, max_overage);
2490 } while ((memcg = parent_mem_cgroup(memcg)) &&
2491 !mem_cgroup_is_root(memcg));
2496 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2498 u64 overage, max_overage = 0;
2501 overage = calculate_overage(page_counter_read(&memcg->swap),
2502 READ_ONCE(memcg->swap.high));
2504 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2505 max_overage = max(overage, max_overage);
2506 } while ((memcg = parent_mem_cgroup(memcg)) &&
2507 !mem_cgroup_is_root(memcg));
2513 * Get the number of jiffies that we should penalise a mischievous cgroup which
2514 * is exceeding its memory.high by checking both it and its ancestors.
2516 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2517 unsigned int nr_pages,
2520 unsigned long penalty_jiffies;
2526 * We use overage compared to memory.high to calculate the number of
2527 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2528 * fairly lenient on small overages, and increasingly harsh when the
2529 * memcg in question makes it clear that it has no intention of stopping
2530 * its crazy behaviour, so we exponentially increase the delay based on
2533 penalty_jiffies = max_overage * max_overage * HZ;
2534 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2535 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2538 * Factor in the task's own contribution to the overage, such that four
2539 * N-sized allocations are throttled approximately the same as one
2540 * 4N-sized allocation.
2542 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2543 * larger the current charge patch is than that.
2545 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2549 * Scheduled by try_charge() to be executed from the userland return path
2550 * and reclaims memory over the high limit.
2552 void mem_cgroup_handle_over_high(void)
2554 unsigned long penalty_jiffies;
2555 unsigned long pflags;
2556 unsigned long nr_reclaimed;
2557 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2558 int nr_retries = MAX_RECLAIM_RETRIES;
2559 struct mem_cgroup *memcg;
2560 bool in_retry = false;
2562 if (likely(!nr_pages))
2565 memcg = get_mem_cgroup_from_mm(current->mm);
2566 current->memcg_nr_pages_over_high = 0;
2570 * The allocating task should reclaim at least the batch size, but for
2571 * subsequent retries we only want to do what's necessary to prevent oom
2572 * or breaching resource isolation.
2574 * This is distinct from memory.max or page allocator behaviour because
2575 * memory.high is currently batched, whereas memory.max and the page
2576 * allocator run every time an allocation is made.
2578 nr_reclaimed = reclaim_high(memcg,
2579 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2583 * memory.high is breached and reclaim is unable to keep up. Throttle
2584 * allocators proactively to slow down excessive growth.
2586 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2587 mem_find_max_overage(memcg));
2589 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2590 swap_find_max_overage(memcg));
2593 * Clamp the max delay per usermode return so as to still keep the
2594 * application moving forwards and also permit diagnostics, albeit
2597 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2600 * Don't sleep if the amount of jiffies this memcg owes us is so low
2601 * that it's not even worth doing, in an attempt to be nice to those who
2602 * go only a small amount over their memory.high value and maybe haven't
2603 * been aggressively reclaimed enough yet.
2605 if (penalty_jiffies <= HZ / 100)
2609 * If reclaim is making forward progress but we're still over
2610 * memory.high, we want to encourage that rather than doing allocator
2613 if (nr_reclaimed || nr_retries--) {
2619 * If we exit early, we're guaranteed to die (since
2620 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2621 * need to account for any ill-begotten jiffies to pay them off later.
2623 psi_memstall_enter(&pflags);
2624 schedule_timeout_killable(penalty_jiffies);
2625 psi_memstall_leave(&pflags);
2628 css_put(&memcg->css);
2631 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2632 unsigned int nr_pages)
2634 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2635 int nr_retries = MAX_RECLAIM_RETRIES;
2636 struct mem_cgroup *mem_over_limit;
2637 struct page_counter *counter;
2638 enum oom_status oom_status;
2639 unsigned long nr_reclaimed;
2640 bool may_swap = true;
2641 bool drained = false;
2642 unsigned long pflags;
2644 if (mem_cgroup_is_root(memcg))
2647 if (consume_stock(memcg, nr_pages))
2650 if (!do_memsw_account() ||
2651 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2652 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2654 if (do_memsw_account())
2655 page_counter_uncharge(&memcg->memsw, batch);
2656 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2658 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2662 if (batch > nr_pages) {
2668 * Memcg doesn't have a dedicated reserve for atomic
2669 * allocations. But like the global atomic pool, we need to
2670 * put the burden of reclaim on regular allocation requests
2671 * and let these go through as privileged allocations.
2673 if (gfp_mask & __GFP_ATOMIC)
2677 * Unlike in global OOM situations, memcg is not in a physical
2678 * memory shortage. Allow dying and OOM-killed tasks to
2679 * bypass the last charges so that they can exit quickly and
2680 * free their memory.
2682 if (unlikely(should_force_charge()))
2686 * Prevent unbounded recursion when reclaim operations need to
2687 * allocate memory. This might exceed the limits temporarily,
2688 * but we prefer facilitating memory reclaim and getting back
2689 * under the limit over triggering OOM kills in these cases.
2691 if (unlikely(current->flags & PF_MEMALLOC))
2694 if (unlikely(task_in_memcg_oom(current)))
2697 if (!gfpflags_allow_blocking(gfp_mask))
2700 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2702 psi_memstall_enter(&pflags);
2703 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2704 gfp_mask, may_swap);
2705 psi_memstall_leave(&pflags);
2707 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2711 drain_all_stock(mem_over_limit);
2716 if (gfp_mask & __GFP_NORETRY)
2719 * Even though the limit is exceeded at this point, reclaim
2720 * may have been able to free some pages. Retry the charge
2721 * before killing the task.
2723 * Only for regular pages, though: huge pages are rather
2724 * unlikely to succeed so close to the limit, and we fall back
2725 * to regular pages anyway in case of failure.
2727 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2730 * At task move, charge accounts can be doubly counted. So, it's
2731 * better to wait until the end of task_move if something is going on.
2733 if (mem_cgroup_wait_acct_move(mem_over_limit))
2739 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2742 if (fatal_signal_pending(current))
2746 * keep retrying as long as the memcg oom killer is able to make
2747 * a forward progress or bypass the charge if the oom killer
2748 * couldn't make any progress.
2750 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2751 get_order(nr_pages * PAGE_SIZE));
2752 switch (oom_status) {
2754 nr_retries = MAX_RECLAIM_RETRIES;
2762 if (!(gfp_mask & __GFP_NOFAIL))
2766 * The allocation either can't fail or will lead to more memory
2767 * being freed very soon. Allow memory usage go over the limit
2768 * temporarily by force charging it.
2770 page_counter_charge(&memcg->memory, nr_pages);
2771 if (do_memsw_account())
2772 page_counter_charge(&memcg->memsw, nr_pages);
2777 if (batch > nr_pages)
2778 refill_stock(memcg, batch - nr_pages);
2781 * If the hierarchy is above the normal consumption range, schedule
2782 * reclaim on returning to userland. We can perform reclaim here
2783 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2784 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2785 * not recorded as it most likely matches current's and won't
2786 * change in the meantime. As high limit is checked again before
2787 * reclaim, the cost of mismatch is negligible.
2790 bool mem_high, swap_high;
2792 mem_high = page_counter_read(&memcg->memory) >
2793 READ_ONCE(memcg->memory.high);
2794 swap_high = page_counter_read(&memcg->swap) >
2795 READ_ONCE(memcg->swap.high);
2797 /* Don't bother a random interrupted task */
2798 if (in_interrupt()) {
2800 schedule_work(&memcg->high_work);
2806 if (mem_high || swap_high) {
2808 * The allocating tasks in this cgroup will need to do
2809 * reclaim or be throttled to prevent further growth
2810 * of the memory or swap footprints.
2812 * Target some best-effort fairness between the tasks,
2813 * and distribute reclaim work and delay penalties
2814 * based on how much each task is actually allocating.
2816 current->memcg_nr_pages_over_high += batch;
2817 set_notify_resume(current);
2820 } while ((memcg = parent_mem_cgroup(memcg)));
2825 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2826 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2828 if (mem_cgroup_is_root(memcg))
2831 page_counter_uncharge(&memcg->memory, nr_pages);
2832 if (do_memsw_account())
2833 page_counter_uncharge(&memcg->memsw, nr_pages);
2837 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2839 VM_BUG_ON_PAGE(page_memcg(page), page);
2841 * Any of the following ensures page's memcg stability:
2845 * - lock_page_memcg()
2846 * - exclusive reference
2848 page->memcg_data = (unsigned long)memcg;
2851 static struct mem_cgroup *get_mem_cgroup_from_objcg(struct obj_cgroup *objcg)
2853 struct mem_cgroup *memcg;
2857 memcg = obj_cgroup_memcg(objcg);
2858 if (unlikely(!css_tryget(&memcg->css)))
2865 #ifdef CONFIG_MEMCG_KMEM
2866 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2867 gfp_t gfp, bool new_page)
2869 unsigned int objects = objs_per_slab_page(s, page);
2870 unsigned long memcg_data;
2873 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2878 memcg_data = (unsigned long) vec | MEMCG_DATA_OBJCGS;
2881 * If the slab page is brand new and nobody can yet access
2882 * it's memcg_data, no synchronization is required and
2883 * memcg_data can be simply assigned.
2885 page->memcg_data = memcg_data;
2886 } else if (cmpxchg(&page->memcg_data, 0, memcg_data)) {
2888 * If the slab page is already in use, somebody can allocate
2889 * and assign obj_cgroups in parallel. In this case the existing
2890 * objcg vector should be reused.
2896 kmemleak_not_leak(vec);
2901 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2903 * A passed kernel object can be a slab object or a generic kernel page, so
2904 * different mechanisms for getting the memory cgroup pointer should be used.
2905 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2906 * can not know for sure how the kernel object is implemented.
2907 * mem_cgroup_from_obj() can be safely used in such cases.
2909 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2910 * cgroup_mutex, etc.
2912 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2916 if (mem_cgroup_disabled())
2919 page = virt_to_head_page(p);
2922 * Slab objects are accounted individually, not per-page.
2923 * Memcg membership data for each individual object is saved in
2924 * the page->obj_cgroups.
2926 if (page_objcgs_check(page)) {
2927 struct obj_cgroup *objcg;
2930 off = obj_to_index(page->slab_cache, page, p);
2931 objcg = page_objcgs(page)[off];
2933 return obj_cgroup_memcg(objcg);
2939 * page_memcg_check() is used here, because page_has_obj_cgroups()
2940 * check above could fail because the object cgroups vector wasn't set
2941 * at that moment, but it can be set concurrently.
2942 * page_memcg_check(page) will guarantee that a proper memory
2943 * cgroup pointer or NULL will be returned.
2945 return page_memcg_check(page);
2948 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2950 struct obj_cgroup *objcg = NULL;
2951 struct mem_cgroup *memcg;
2953 if (memcg_kmem_bypass())
2957 if (unlikely(active_memcg()))
2958 memcg = active_memcg();
2960 memcg = mem_cgroup_from_task(current);
2962 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2963 objcg = rcu_dereference(memcg->objcg);
2964 if (objcg && obj_cgroup_tryget(objcg))
2973 static int memcg_alloc_cache_id(void)
2978 id = ida_simple_get(&memcg_cache_ida,
2979 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2983 if (id < memcg_nr_cache_ids)
2987 * There's no space for the new id in memcg_caches arrays,
2988 * so we have to grow them.
2990 down_write(&memcg_cache_ids_sem);
2992 size = 2 * (id + 1);
2993 if (size < MEMCG_CACHES_MIN_SIZE)
2994 size = MEMCG_CACHES_MIN_SIZE;
2995 else if (size > MEMCG_CACHES_MAX_SIZE)
2996 size = MEMCG_CACHES_MAX_SIZE;
2998 err = memcg_update_all_list_lrus(size);
3000 memcg_nr_cache_ids = size;
3002 up_write(&memcg_cache_ids_sem);
3005 ida_simple_remove(&memcg_cache_ida, id);
3011 static void memcg_free_cache_id(int id)
3013 ida_simple_remove(&memcg_cache_ida, id);
3017 * obj_cgroup_uncharge_pages: uncharge a number of kernel pages from a objcg
3018 * @objcg: object cgroup to uncharge
3019 * @nr_pages: number of pages to uncharge
3021 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
3022 unsigned int nr_pages)
3024 struct mem_cgroup *memcg;
3026 memcg = get_mem_cgroup_from_objcg(objcg);
3028 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3029 page_counter_uncharge(&memcg->kmem, nr_pages);
3030 refill_stock(memcg, nr_pages);
3032 css_put(&memcg->css);
3036 * obj_cgroup_charge_pages: charge a number of kernel pages to a objcg
3037 * @objcg: object cgroup to charge
3038 * @gfp: reclaim mode
3039 * @nr_pages: number of pages to charge
3041 * Returns 0 on success, an error code on failure.
3043 static int obj_cgroup_charge_pages(struct obj_cgroup *objcg, gfp_t gfp,
3044 unsigned int nr_pages)
3046 struct page_counter *counter;
3047 struct mem_cgroup *memcg;
3050 memcg = get_mem_cgroup_from_objcg(objcg);
3052 ret = try_charge(memcg, gfp, nr_pages);
3056 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3057 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3060 * Enforce __GFP_NOFAIL allocation because callers are not
3061 * prepared to see failures and likely do not have any failure
3064 if (gfp & __GFP_NOFAIL) {
3065 page_counter_charge(&memcg->kmem, nr_pages);
3068 cancel_charge(memcg, nr_pages);
3072 css_put(&memcg->css);
3078 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3079 * @page: page to charge
3080 * @gfp: reclaim mode
3081 * @order: allocation order
3083 * Returns 0 on success, an error code on failure.
3085 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3087 struct obj_cgroup *objcg;
3090 objcg = get_obj_cgroup_from_current();
3092 ret = obj_cgroup_charge_pages(objcg, gfp, 1 << order);
3094 page->memcg_data = (unsigned long)objcg |
3098 obj_cgroup_put(objcg);
3104 * __memcg_kmem_uncharge_page: uncharge a kmem page
3105 * @page: page to uncharge
3106 * @order: allocation order
3108 void __memcg_kmem_uncharge_page(struct page *page, int order)
3110 struct obj_cgroup *objcg;
3111 unsigned int nr_pages = 1 << order;
3113 if (!PageMemcgKmem(page))
3116 objcg = __page_objcg(page);
3117 obj_cgroup_uncharge_pages(objcg, nr_pages);
3118 page->memcg_data = 0;
3119 obj_cgroup_put(objcg);
3122 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3124 struct memcg_stock_pcp *stock;
3125 unsigned long flags;
3128 local_irq_save(flags);
3130 stock = this_cpu_ptr(&memcg_stock);
3131 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3132 stock->nr_bytes -= nr_bytes;
3136 local_irq_restore(flags);
3141 static void drain_obj_stock(struct memcg_stock_pcp *stock)
3143 struct obj_cgroup *old = stock->cached_objcg;
3148 if (stock->nr_bytes) {
3149 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3150 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3153 obj_cgroup_uncharge_pages(old, nr_pages);
3156 * The leftover is flushed to the centralized per-memcg value.
3157 * On the next attempt to refill obj stock it will be moved
3158 * to a per-cpu stock (probably, on an other CPU), see
3159 * refill_obj_stock().
3161 * How often it's flushed is a trade-off between the memory
3162 * limit enforcement accuracy and potential CPU contention,
3163 * so it might be changed in the future.
3165 atomic_add(nr_bytes, &old->nr_charged_bytes);
3166 stock->nr_bytes = 0;
3169 obj_cgroup_put(old);
3170 stock->cached_objcg = NULL;
3173 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3174 struct mem_cgroup *root_memcg)
3176 struct mem_cgroup *memcg;
3178 if (stock->cached_objcg) {
3179 memcg = obj_cgroup_memcg(stock->cached_objcg);
3180 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3187 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3189 struct memcg_stock_pcp *stock;
3190 unsigned long flags;
3192 local_irq_save(flags);
3194 stock = this_cpu_ptr(&memcg_stock);
3195 if (stock->cached_objcg != objcg) { /* reset if necessary */
3196 drain_obj_stock(stock);
3197 obj_cgroup_get(objcg);
3198 stock->cached_objcg = objcg;
3199 stock->nr_bytes = atomic_xchg(&objcg->nr_charged_bytes, 0);
3201 stock->nr_bytes += nr_bytes;
3203 if (stock->nr_bytes > PAGE_SIZE)
3204 drain_obj_stock(stock);
3206 local_irq_restore(flags);
3209 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3211 unsigned int nr_pages, nr_bytes;
3214 if (consume_obj_stock(objcg, size))
3218 * In theory, memcg->nr_charged_bytes can have enough
3219 * pre-charged bytes to satisfy the allocation. However,
3220 * flushing memcg->nr_charged_bytes requires two atomic
3221 * operations, and memcg->nr_charged_bytes can't be big,
3222 * so it's better to ignore it and try grab some new pages.
3223 * memcg->nr_charged_bytes will be flushed in
3224 * refill_obj_stock(), called from this function or
3225 * independently later.
3227 nr_pages = size >> PAGE_SHIFT;
3228 nr_bytes = size & (PAGE_SIZE - 1);
3233 ret = obj_cgroup_charge_pages(objcg, gfp, nr_pages);
3234 if (!ret && nr_bytes)
3235 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes);
3240 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3242 refill_obj_stock(objcg, size);
3245 #endif /* CONFIG_MEMCG_KMEM */
3248 * Because page_memcg(head) is not set on tails, set it now.
3250 void split_page_memcg(struct page *head, unsigned int nr)
3252 struct mem_cgroup *memcg = page_memcg(head);
3255 if (mem_cgroup_disabled() || !memcg)
3258 for (i = 1; i < nr; i++)
3259 head[i].memcg_data = head->memcg_data;
3261 if (PageMemcgKmem(head))
3262 obj_cgroup_get_many(__page_objcg(head), nr - 1);
3264 css_get_many(&memcg->css, nr - 1);
3267 #ifdef CONFIG_MEMCG_SWAP
3269 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3270 * @entry: swap entry to be moved
3271 * @from: mem_cgroup which the entry is moved from
3272 * @to: mem_cgroup which the entry is moved to
3274 * It succeeds only when the swap_cgroup's record for this entry is the same
3275 * as the mem_cgroup's id of @from.
3277 * Returns 0 on success, -EINVAL on failure.
3279 * The caller must have charged to @to, IOW, called page_counter_charge() about
3280 * both res and memsw, and called css_get().
3282 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3283 struct mem_cgroup *from, struct mem_cgroup *to)
3285 unsigned short old_id, new_id;
3287 old_id = mem_cgroup_id(from);
3288 new_id = mem_cgroup_id(to);
3290 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3291 mod_memcg_state(from, MEMCG_SWAP, -1);
3292 mod_memcg_state(to, MEMCG_SWAP, 1);
3298 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3299 struct mem_cgroup *from, struct mem_cgroup *to)
3305 static DEFINE_MUTEX(memcg_max_mutex);
3307 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3308 unsigned long max, bool memsw)
3310 bool enlarge = false;
3311 bool drained = false;
3313 bool limits_invariant;
3314 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3317 if (signal_pending(current)) {
3322 mutex_lock(&memcg_max_mutex);
3324 * Make sure that the new limit (memsw or memory limit) doesn't
3325 * break our basic invariant rule memory.max <= memsw.max.
3327 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3328 max <= memcg->memsw.max;
3329 if (!limits_invariant) {
3330 mutex_unlock(&memcg_max_mutex);
3334 if (max > counter->max)
3336 ret = page_counter_set_max(counter, max);
3337 mutex_unlock(&memcg_max_mutex);
3343 drain_all_stock(memcg);
3348 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3349 GFP_KERNEL, !memsw)) {
3355 if (!ret && enlarge)
3356 memcg_oom_recover(memcg);
3361 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3363 unsigned long *total_scanned)
3365 unsigned long nr_reclaimed = 0;
3366 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3367 unsigned long reclaimed;
3369 struct mem_cgroup_tree_per_node *mctz;
3370 unsigned long excess;
3371 unsigned long nr_scanned;
3376 mctz = soft_limit_tree_node(pgdat->node_id);
3379 * Do not even bother to check the largest node if the root
3380 * is empty. Do it lockless to prevent lock bouncing. Races
3381 * are acceptable as soft limit is best effort anyway.
3383 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3387 * This loop can run a while, specially if mem_cgroup's continuously
3388 * keep exceeding their soft limit and putting the system under
3395 mz = mem_cgroup_largest_soft_limit_node(mctz);
3400 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3401 gfp_mask, &nr_scanned);
3402 nr_reclaimed += reclaimed;
3403 *total_scanned += nr_scanned;
3404 spin_lock_irq(&mctz->lock);
3405 __mem_cgroup_remove_exceeded(mz, mctz);
3408 * If we failed to reclaim anything from this memory cgroup
3409 * it is time to move on to the next cgroup
3413 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3415 excess = soft_limit_excess(mz->memcg);
3417 * One school of thought says that we should not add
3418 * back the node to the tree if reclaim returns 0.
3419 * But our reclaim could return 0, simply because due
3420 * to priority we are exposing a smaller subset of
3421 * memory to reclaim from. Consider this as a longer
3424 /* If excess == 0, no tree ops */
3425 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3426 spin_unlock_irq(&mctz->lock);
3427 css_put(&mz->memcg->css);
3430 * Could not reclaim anything and there are no more
3431 * mem cgroups to try or we seem to be looping without
3432 * reclaiming anything.
3434 if (!nr_reclaimed &&
3436 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3438 } while (!nr_reclaimed);
3440 css_put(&next_mz->memcg->css);
3441 return nr_reclaimed;
3445 * Reclaims as many pages from the given memcg as possible.
3447 * Caller is responsible for holding css reference for memcg.
3449 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3451 int nr_retries = MAX_RECLAIM_RETRIES;
3453 /* we call try-to-free pages for make this cgroup empty */
3454 lru_add_drain_all();
3456 drain_all_stock(memcg);
3458 /* try to free all pages in this cgroup */
3459 while (nr_retries && page_counter_read(&memcg->memory)) {
3462 if (signal_pending(current))
3465 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3469 /* maybe some writeback is necessary */
3470 congestion_wait(BLK_RW_ASYNC, HZ/10);
3478 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3479 char *buf, size_t nbytes,
3482 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3484 if (mem_cgroup_is_root(memcg))
3486 return mem_cgroup_force_empty(memcg) ?: nbytes;
3489 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3495 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3496 struct cftype *cft, u64 val)
3501 pr_warn_once("Non-hierarchical mode is deprecated. "
3502 "Please report your usecase to linux-mm@kvack.org if you "
3503 "depend on this functionality.\n");
3508 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3512 if (mem_cgroup_is_root(memcg)) {
3513 cgroup_rstat_flush(memcg->css.cgroup);
3514 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3515 memcg_page_state(memcg, NR_ANON_MAPPED);
3517 val += memcg_page_state(memcg, MEMCG_SWAP);
3520 val = page_counter_read(&memcg->memory);
3522 val = page_counter_read(&memcg->memsw);
3535 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3538 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3539 struct page_counter *counter;
3541 switch (MEMFILE_TYPE(cft->private)) {
3543 counter = &memcg->memory;
3546 counter = &memcg->memsw;
3549 counter = &memcg->kmem;
3552 counter = &memcg->tcpmem;
3558 switch (MEMFILE_ATTR(cft->private)) {
3560 if (counter == &memcg->memory)
3561 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3562 if (counter == &memcg->memsw)
3563 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3564 return (u64)page_counter_read(counter) * PAGE_SIZE;
3566 return (u64)counter->max * PAGE_SIZE;
3568 return (u64)counter->watermark * PAGE_SIZE;
3570 return counter->failcnt;
3571 case RES_SOFT_LIMIT:
3572 return (u64)memcg->soft_limit * PAGE_SIZE;
3578 #ifdef CONFIG_MEMCG_KMEM
3579 static int memcg_online_kmem(struct mem_cgroup *memcg)
3581 struct obj_cgroup *objcg;
3584 if (cgroup_memory_nokmem)
3587 BUG_ON(memcg->kmemcg_id >= 0);
3588 BUG_ON(memcg->kmem_state);
3590 memcg_id = memcg_alloc_cache_id();
3594 objcg = obj_cgroup_alloc();
3596 memcg_free_cache_id(memcg_id);
3599 objcg->memcg = memcg;
3600 rcu_assign_pointer(memcg->objcg, objcg);
3602 static_branch_enable(&memcg_kmem_enabled_key);
3604 memcg->kmemcg_id = memcg_id;
3605 memcg->kmem_state = KMEM_ONLINE;
3610 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3612 struct cgroup_subsys_state *css;
3613 struct mem_cgroup *parent, *child;
3616 if (memcg->kmem_state != KMEM_ONLINE)
3619 memcg->kmem_state = KMEM_ALLOCATED;
3621 parent = parent_mem_cgroup(memcg);
3623 parent = root_mem_cgroup;
3625 memcg_reparent_objcgs(memcg, parent);
3627 kmemcg_id = memcg->kmemcg_id;
3628 BUG_ON(kmemcg_id < 0);
3631 * Change kmemcg_id of this cgroup and all its descendants to the
3632 * parent's id, and then move all entries from this cgroup's list_lrus
3633 * to ones of the parent. After we have finished, all list_lrus
3634 * corresponding to this cgroup are guaranteed to remain empty. The
3635 * ordering is imposed by list_lru_node->lock taken by
3636 * memcg_drain_all_list_lrus().
3638 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3639 css_for_each_descendant_pre(css, &memcg->css) {
3640 child = mem_cgroup_from_css(css);
3641 BUG_ON(child->kmemcg_id != kmemcg_id);
3642 child->kmemcg_id = parent->kmemcg_id;
3646 memcg_drain_all_list_lrus(kmemcg_id, parent);
3648 memcg_free_cache_id(kmemcg_id);
3651 static void memcg_free_kmem(struct mem_cgroup *memcg)
3653 /* css_alloc() failed, offlining didn't happen */
3654 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3655 memcg_offline_kmem(memcg);
3658 static int memcg_online_kmem(struct mem_cgroup *memcg)
3662 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3665 static void memcg_free_kmem(struct mem_cgroup *memcg)
3668 #endif /* CONFIG_MEMCG_KMEM */
3670 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3675 mutex_lock(&memcg_max_mutex);
3676 ret = page_counter_set_max(&memcg->kmem, max);
3677 mutex_unlock(&memcg_max_mutex);
3681 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3685 mutex_lock(&memcg_max_mutex);
3687 ret = page_counter_set_max(&memcg->tcpmem, max);
3691 if (!memcg->tcpmem_active) {
3693 * The active flag needs to be written after the static_key
3694 * update. This is what guarantees that the socket activation
3695 * function is the last one to run. See mem_cgroup_sk_alloc()
3696 * for details, and note that we don't mark any socket as
3697 * belonging to this memcg until that flag is up.
3699 * We need to do this, because static_keys will span multiple
3700 * sites, but we can't control their order. If we mark a socket
3701 * as accounted, but the accounting functions are not patched in
3702 * yet, we'll lose accounting.
3704 * We never race with the readers in mem_cgroup_sk_alloc(),
3705 * because when this value change, the code to process it is not
3708 static_branch_inc(&memcg_sockets_enabled_key);
3709 memcg->tcpmem_active = true;
3712 mutex_unlock(&memcg_max_mutex);
3717 * The user of this function is...
3720 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3721 char *buf, size_t nbytes, loff_t off)
3723 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3724 unsigned long nr_pages;
3727 buf = strstrip(buf);
3728 ret = page_counter_memparse(buf, "-1", &nr_pages);
3732 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3734 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3738 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3740 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3743 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3746 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3747 "Please report your usecase to linux-mm@kvack.org if you "
3748 "depend on this functionality.\n");
3749 ret = memcg_update_kmem_max(memcg, nr_pages);
3752 ret = memcg_update_tcp_max(memcg, nr_pages);
3756 case RES_SOFT_LIMIT:
3757 memcg->soft_limit = nr_pages;
3761 return ret ?: nbytes;
3764 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3765 size_t nbytes, loff_t off)
3767 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3768 struct page_counter *counter;
3770 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3772 counter = &memcg->memory;
3775 counter = &memcg->memsw;
3778 counter = &memcg->kmem;
3781 counter = &memcg->tcpmem;
3787 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3789 page_counter_reset_watermark(counter);
3792 counter->failcnt = 0;
3801 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3804 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3808 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3809 struct cftype *cft, u64 val)
3811 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3813 if (val & ~MOVE_MASK)
3817 * No kind of locking is needed in here, because ->can_attach() will
3818 * check this value once in the beginning of the process, and then carry
3819 * on with stale data. This means that changes to this value will only
3820 * affect task migrations starting after the change.
3822 memcg->move_charge_at_immigrate = val;
3826 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3827 struct cftype *cft, u64 val)
3835 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3836 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3837 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3839 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3840 int nid, unsigned int lru_mask, bool tree)
3842 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3843 unsigned long nr = 0;
3846 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3849 if (!(BIT(lru) & lru_mask))
3852 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3854 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3859 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3860 unsigned int lru_mask,
3863 unsigned long nr = 0;
3867 if (!(BIT(lru) & lru_mask))
3870 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3872 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3877 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3881 unsigned int lru_mask;
3884 static const struct numa_stat stats[] = {
3885 { "total", LRU_ALL },
3886 { "file", LRU_ALL_FILE },
3887 { "anon", LRU_ALL_ANON },
3888 { "unevictable", BIT(LRU_UNEVICTABLE) },
3890 const struct numa_stat *stat;
3892 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3894 cgroup_rstat_flush(memcg->css.cgroup);
3896 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3897 seq_printf(m, "%s=%lu", stat->name,
3898 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3900 for_each_node_state(nid, N_MEMORY)
3901 seq_printf(m, " N%d=%lu", nid,
3902 mem_cgroup_node_nr_lru_pages(memcg, nid,
3903 stat->lru_mask, false));
3907 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3909 seq_printf(m, "hierarchical_%s=%lu", stat->name,
3910 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3912 for_each_node_state(nid, N_MEMORY)
3913 seq_printf(m, " N%d=%lu", nid,
3914 mem_cgroup_node_nr_lru_pages(memcg, nid,
3915 stat->lru_mask, true));
3921 #endif /* CONFIG_NUMA */
3923 static const unsigned int memcg1_stats[] = {
3926 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3936 static const char *const memcg1_stat_names[] = {
3939 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3949 /* Universal VM events cgroup1 shows, original sort order */
3950 static const unsigned int memcg1_events[] = {
3957 static int memcg_stat_show(struct seq_file *m, void *v)
3959 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3960 unsigned long memory, memsw;
3961 struct mem_cgroup *mi;
3964 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3966 cgroup_rstat_flush(memcg->css.cgroup);
3968 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3971 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3973 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
3974 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
3977 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3978 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
3979 memcg_events_local(memcg, memcg1_events[i]));
3981 for (i = 0; i < NR_LRU_LISTS; i++)
3982 seq_printf(m, "%s %lu\n", lru_list_name(i),
3983 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3986 /* Hierarchical information */
3987 memory = memsw = PAGE_COUNTER_MAX;
3988 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3989 memory = min(memory, READ_ONCE(mi->memory.max));
3990 memsw = min(memsw, READ_ONCE(mi->memsw.max));
3992 seq_printf(m, "hierarchical_memory_limit %llu\n",
3993 (u64)memory * PAGE_SIZE);
3994 if (do_memsw_account())
3995 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3996 (u64)memsw * PAGE_SIZE);
3998 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4001 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4003 nr = memcg_page_state(memcg, memcg1_stats[i]);
4004 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4005 (u64)nr * PAGE_SIZE);
4008 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4009 seq_printf(m, "total_%s %llu\n",
4010 vm_event_name(memcg1_events[i]),
4011 (u64)memcg_events(memcg, memcg1_events[i]));
4013 for (i = 0; i < NR_LRU_LISTS; i++)
4014 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4015 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4018 #ifdef CONFIG_DEBUG_VM
4021 struct mem_cgroup_per_node *mz;
4022 unsigned long anon_cost = 0;
4023 unsigned long file_cost = 0;
4025 for_each_online_pgdat(pgdat) {
4026 mz = memcg->nodeinfo[pgdat->node_id];
4028 anon_cost += mz->lruvec.anon_cost;
4029 file_cost += mz->lruvec.file_cost;
4031 seq_printf(m, "anon_cost %lu\n", anon_cost);
4032 seq_printf(m, "file_cost %lu\n", file_cost);
4039 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4042 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4044 return mem_cgroup_swappiness(memcg);
4047 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4048 struct cftype *cft, u64 val)
4050 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4055 if (!mem_cgroup_is_root(memcg))
4056 memcg->swappiness = val;
4058 vm_swappiness = val;
4063 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4065 struct mem_cgroup_threshold_ary *t;
4066 unsigned long usage;
4071 t = rcu_dereference(memcg->thresholds.primary);
4073 t = rcu_dereference(memcg->memsw_thresholds.primary);
4078 usage = mem_cgroup_usage(memcg, swap);
4081 * current_threshold points to threshold just below or equal to usage.
4082 * If it's not true, a threshold was crossed after last
4083 * call of __mem_cgroup_threshold().
4085 i = t->current_threshold;
4088 * Iterate backward over array of thresholds starting from
4089 * current_threshold and check if a threshold is crossed.
4090 * If none of thresholds below usage is crossed, we read
4091 * only one element of the array here.
4093 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4094 eventfd_signal(t->entries[i].eventfd, 1);
4096 /* i = current_threshold + 1 */
4100 * Iterate forward over array of thresholds starting from
4101 * current_threshold+1 and check if a threshold is crossed.
4102 * If none of thresholds above usage is crossed, we read
4103 * only one element of the array here.
4105 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4106 eventfd_signal(t->entries[i].eventfd, 1);
4108 /* Update current_threshold */
4109 t->current_threshold = i - 1;
4114 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4117 __mem_cgroup_threshold(memcg, false);
4118 if (do_memsw_account())
4119 __mem_cgroup_threshold(memcg, true);
4121 memcg = parent_mem_cgroup(memcg);
4125 static int compare_thresholds(const void *a, const void *b)
4127 const struct mem_cgroup_threshold *_a = a;
4128 const struct mem_cgroup_threshold *_b = b;
4130 if (_a->threshold > _b->threshold)
4133 if (_a->threshold < _b->threshold)
4139 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4141 struct mem_cgroup_eventfd_list *ev;
4143 spin_lock(&memcg_oom_lock);
4145 list_for_each_entry(ev, &memcg->oom_notify, list)
4146 eventfd_signal(ev->eventfd, 1);
4148 spin_unlock(&memcg_oom_lock);
4152 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4154 struct mem_cgroup *iter;
4156 for_each_mem_cgroup_tree(iter, memcg)
4157 mem_cgroup_oom_notify_cb(iter);
4160 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4161 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4163 struct mem_cgroup_thresholds *thresholds;
4164 struct mem_cgroup_threshold_ary *new;
4165 unsigned long threshold;
4166 unsigned long usage;
4169 ret = page_counter_memparse(args, "-1", &threshold);
4173 mutex_lock(&memcg->thresholds_lock);
4176 thresholds = &memcg->thresholds;
4177 usage = mem_cgroup_usage(memcg, false);
4178 } else if (type == _MEMSWAP) {
4179 thresholds = &memcg->memsw_thresholds;
4180 usage = mem_cgroup_usage(memcg, true);
4184 /* Check if a threshold crossed before adding a new one */
4185 if (thresholds->primary)
4186 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4188 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4190 /* Allocate memory for new array of thresholds */
4191 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4198 /* Copy thresholds (if any) to new array */
4199 if (thresholds->primary)
4200 memcpy(new->entries, thresholds->primary->entries,
4201 flex_array_size(new, entries, size - 1));
4203 /* Add new threshold */
4204 new->entries[size - 1].eventfd = eventfd;
4205 new->entries[size - 1].threshold = threshold;
4207 /* Sort thresholds. Registering of new threshold isn't time-critical */
4208 sort(new->entries, size, sizeof(*new->entries),
4209 compare_thresholds, NULL);
4211 /* Find current threshold */
4212 new->current_threshold = -1;
4213 for (i = 0; i < size; i++) {
4214 if (new->entries[i].threshold <= usage) {
4216 * new->current_threshold will not be used until
4217 * rcu_assign_pointer(), so it's safe to increment
4220 ++new->current_threshold;
4225 /* Free old spare buffer and save old primary buffer as spare */
4226 kfree(thresholds->spare);
4227 thresholds->spare = thresholds->primary;
4229 rcu_assign_pointer(thresholds->primary, new);
4231 /* To be sure that nobody uses thresholds */
4235 mutex_unlock(&memcg->thresholds_lock);
4240 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4241 struct eventfd_ctx *eventfd, const char *args)
4243 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4246 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4247 struct eventfd_ctx *eventfd, const char *args)
4249 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4252 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4253 struct eventfd_ctx *eventfd, enum res_type type)
4255 struct mem_cgroup_thresholds *thresholds;
4256 struct mem_cgroup_threshold_ary *new;
4257 unsigned long usage;
4258 int i, j, size, entries;
4260 mutex_lock(&memcg->thresholds_lock);
4263 thresholds = &memcg->thresholds;
4264 usage = mem_cgroup_usage(memcg, false);
4265 } else if (type == _MEMSWAP) {
4266 thresholds = &memcg->memsw_thresholds;
4267 usage = mem_cgroup_usage(memcg, true);
4271 if (!thresholds->primary)
4274 /* Check if a threshold crossed before removing */
4275 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4277 /* Calculate new number of threshold */
4279 for (i = 0; i < thresholds->primary->size; i++) {
4280 if (thresholds->primary->entries[i].eventfd != eventfd)
4286 new = thresholds->spare;
4288 /* If no items related to eventfd have been cleared, nothing to do */
4292 /* Set thresholds array to NULL if we don't have thresholds */
4301 /* Copy thresholds and find current threshold */
4302 new->current_threshold = -1;
4303 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4304 if (thresholds->primary->entries[i].eventfd == eventfd)
4307 new->entries[j] = thresholds->primary->entries[i];
4308 if (new->entries[j].threshold <= usage) {
4310 * new->current_threshold will not be used
4311 * until rcu_assign_pointer(), so it's safe to increment
4314 ++new->current_threshold;
4320 /* Swap primary and spare array */
4321 thresholds->spare = thresholds->primary;
4323 rcu_assign_pointer(thresholds->primary, new);
4325 /* To be sure that nobody uses thresholds */
4328 /* If all events are unregistered, free the spare array */
4330 kfree(thresholds->spare);
4331 thresholds->spare = NULL;
4334 mutex_unlock(&memcg->thresholds_lock);
4337 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4338 struct eventfd_ctx *eventfd)
4340 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4343 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4344 struct eventfd_ctx *eventfd)
4346 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4349 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4350 struct eventfd_ctx *eventfd, const char *args)
4352 struct mem_cgroup_eventfd_list *event;
4354 event = kmalloc(sizeof(*event), GFP_KERNEL);
4358 spin_lock(&memcg_oom_lock);
4360 event->eventfd = eventfd;
4361 list_add(&event->list, &memcg->oom_notify);
4363 /* already in OOM ? */
4364 if (memcg->under_oom)
4365 eventfd_signal(eventfd, 1);
4366 spin_unlock(&memcg_oom_lock);
4371 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4372 struct eventfd_ctx *eventfd)
4374 struct mem_cgroup_eventfd_list *ev, *tmp;
4376 spin_lock(&memcg_oom_lock);
4378 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4379 if (ev->eventfd == eventfd) {
4380 list_del(&ev->list);
4385 spin_unlock(&memcg_oom_lock);
4388 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4390 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4392 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4393 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4394 seq_printf(sf, "oom_kill %lu\n",
4395 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4399 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4400 struct cftype *cft, u64 val)
4402 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4404 /* cannot set to root cgroup and only 0 and 1 are allowed */
4405 if (mem_cgroup_is_root(memcg) || !((val == 0) || (val == 1)))
4408 memcg->oom_kill_disable = val;
4410 memcg_oom_recover(memcg);
4415 #ifdef CONFIG_CGROUP_WRITEBACK
4417 #include <trace/events/writeback.h>
4419 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4421 return wb_domain_init(&memcg->cgwb_domain, gfp);
4424 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4426 wb_domain_exit(&memcg->cgwb_domain);
4429 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4431 wb_domain_size_changed(&memcg->cgwb_domain);
4434 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4436 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4438 if (!memcg->css.parent)
4441 return &memcg->cgwb_domain;
4445 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4446 * @wb: bdi_writeback in question
4447 * @pfilepages: out parameter for number of file pages
4448 * @pheadroom: out parameter for number of allocatable pages according to memcg
4449 * @pdirty: out parameter for number of dirty pages
4450 * @pwriteback: out parameter for number of pages under writeback
4452 * Determine the numbers of file, headroom, dirty, and writeback pages in
4453 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4454 * is a bit more involved.
4456 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4457 * headroom is calculated as the lowest headroom of itself and the
4458 * ancestors. Note that this doesn't consider the actual amount of
4459 * available memory in the system. The caller should further cap
4460 * *@pheadroom accordingly.
4462 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4463 unsigned long *pheadroom, unsigned long *pdirty,
4464 unsigned long *pwriteback)
4466 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4467 struct mem_cgroup *parent;
4469 cgroup_rstat_flush_irqsafe(memcg->css.cgroup);
4471 *pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
4472 *pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
4473 *pfilepages = memcg_page_state(memcg, NR_INACTIVE_FILE) +
4474 memcg_page_state(memcg, NR_ACTIVE_FILE);
4476 *pheadroom = PAGE_COUNTER_MAX;
4477 while ((parent = parent_mem_cgroup(memcg))) {
4478 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4479 READ_ONCE(memcg->memory.high));
4480 unsigned long used = page_counter_read(&memcg->memory);
4482 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4488 * Foreign dirty flushing
4490 * There's an inherent mismatch between memcg and writeback. The former
4491 * trackes ownership per-page while the latter per-inode. This was a
4492 * deliberate design decision because honoring per-page ownership in the
4493 * writeback path is complicated, may lead to higher CPU and IO overheads
4494 * and deemed unnecessary given that write-sharing an inode across
4495 * different cgroups isn't a common use-case.
4497 * Combined with inode majority-writer ownership switching, this works well
4498 * enough in most cases but there are some pathological cases. For
4499 * example, let's say there are two cgroups A and B which keep writing to
4500 * different but confined parts of the same inode. B owns the inode and
4501 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4502 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4503 * triggering background writeback. A will be slowed down without a way to
4504 * make writeback of the dirty pages happen.
4506 * Conditions like the above can lead to a cgroup getting repatedly and
4507 * severely throttled after making some progress after each
4508 * dirty_expire_interval while the underyling IO device is almost
4511 * Solving this problem completely requires matching the ownership tracking
4512 * granularities between memcg and writeback in either direction. However,
4513 * the more egregious behaviors can be avoided by simply remembering the
4514 * most recent foreign dirtying events and initiating remote flushes on
4515 * them when local writeback isn't enough to keep the memory clean enough.
4517 * The following two functions implement such mechanism. When a foreign
4518 * page - a page whose memcg and writeback ownerships don't match - is
4519 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4520 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4521 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4522 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4523 * foreign bdi_writebacks which haven't expired. Both the numbers of
4524 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4525 * limited to MEMCG_CGWB_FRN_CNT.
4527 * The mechanism only remembers IDs and doesn't hold any object references.
4528 * As being wrong occasionally doesn't matter, updates and accesses to the
4529 * records are lockless and racy.
4531 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4532 struct bdi_writeback *wb)
4534 struct mem_cgroup *memcg = page_memcg(page);
4535 struct memcg_cgwb_frn *frn;
4536 u64 now = get_jiffies_64();
4537 u64 oldest_at = now;
4541 trace_track_foreign_dirty(page, wb);
4544 * Pick the slot to use. If there is already a slot for @wb, keep
4545 * using it. If not replace the oldest one which isn't being
4548 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4549 frn = &memcg->cgwb_frn[i];
4550 if (frn->bdi_id == wb->bdi->id &&
4551 frn->memcg_id == wb->memcg_css->id)
4553 if (time_before64(frn->at, oldest_at) &&
4554 atomic_read(&frn->done.cnt) == 1) {
4556 oldest_at = frn->at;
4560 if (i < MEMCG_CGWB_FRN_CNT) {
4562 * Re-using an existing one. Update timestamp lazily to
4563 * avoid making the cacheline hot. We want them to be
4564 * reasonably up-to-date and significantly shorter than
4565 * dirty_expire_interval as that's what expires the record.
4566 * Use the shorter of 1s and dirty_expire_interval / 8.
4568 unsigned long update_intv =
4569 min_t(unsigned long, HZ,
4570 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4572 if (time_before64(frn->at, now - update_intv))
4574 } else if (oldest >= 0) {
4575 /* replace the oldest free one */
4576 frn = &memcg->cgwb_frn[oldest];
4577 frn->bdi_id = wb->bdi->id;
4578 frn->memcg_id = wb->memcg_css->id;
4583 /* issue foreign writeback flushes for recorded foreign dirtying events */
4584 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4586 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4587 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4588 u64 now = jiffies_64;
4591 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4592 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4595 * If the record is older than dirty_expire_interval,
4596 * writeback on it has already started. No need to kick it
4597 * off again. Also, don't start a new one if there's
4598 * already one in flight.
4600 if (time_after64(frn->at, now - intv) &&
4601 atomic_read(&frn->done.cnt) == 1) {
4603 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4604 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4605 WB_REASON_FOREIGN_FLUSH,
4611 #else /* CONFIG_CGROUP_WRITEBACK */
4613 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4618 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4622 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4626 #endif /* CONFIG_CGROUP_WRITEBACK */
4629 * DO NOT USE IN NEW FILES.
4631 * "cgroup.event_control" implementation.
4633 * This is way over-engineered. It tries to support fully configurable
4634 * events for each user. Such level of flexibility is completely
4635 * unnecessary especially in the light of the planned unified hierarchy.
4637 * Please deprecate this and replace with something simpler if at all
4642 * Unregister event and free resources.
4644 * Gets called from workqueue.
4646 static void memcg_event_remove(struct work_struct *work)
4648 struct mem_cgroup_event *event =
4649 container_of(work, struct mem_cgroup_event, remove);
4650 struct mem_cgroup *memcg = event->memcg;
4652 remove_wait_queue(event->wqh, &event->wait);
4654 event->unregister_event(memcg, event->eventfd);
4656 /* Notify userspace the event is going away. */
4657 eventfd_signal(event->eventfd, 1);
4659 eventfd_ctx_put(event->eventfd);
4661 css_put(&memcg->css);
4665 * Gets called on EPOLLHUP on eventfd when user closes it.
4667 * Called with wqh->lock held and interrupts disabled.
4669 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4670 int sync, void *key)
4672 struct mem_cgroup_event *event =
4673 container_of(wait, struct mem_cgroup_event, wait);
4674 struct mem_cgroup *memcg = event->memcg;
4675 __poll_t flags = key_to_poll(key);
4677 if (flags & EPOLLHUP) {
4679 * If the event has been detached at cgroup removal, we
4680 * can simply return knowing the other side will cleanup
4683 * We can't race against event freeing since the other
4684 * side will require wqh->lock via remove_wait_queue(),
4687 spin_lock(&memcg->event_list_lock);
4688 if (!list_empty(&event->list)) {
4689 list_del_init(&event->list);
4691 * We are in atomic context, but cgroup_event_remove()
4692 * may sleep, so we have to call it in workqueue.
4694 schedule_work(&event->remove);
4696 spin_unlock(&memcg->event_list_lock);
4702 static void memcg_event_ptable_queue_proc(struct file *file,
4703 wait_queue_head_t *wqh, poll_table *pt)
4705 struct mem_cgroup_event *event =
4706 container_of(pt, struct mem_cgroup_event, pt);
4709 add_wait_queue(wqh, &event->wait);
4713 * DO NOT USE IN NEW FILES.
4715 * Parse input and register new cgroup event handler.
4717 * Input must be in format '<event_fd> <control_fd> <args>'.
4718 * Interpretation of args is defined by control file implementation.
4720 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4721 char *buf, size_t nbytes, loff_t off)
4723 struct cgroup_subsys_state *css = of_css(of);
4724 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4725 struct mem_cgroup_event *event;
4726 struct cgroup_subsys_state *cfile_css;
4727 unsigned int efd, cfd;
4734 buf = strstrip(buf);
4736 efd = simple_strtoul(buf, &endp, 10);
4741 cfd = simple_strtoul(buf, &endp, 10);
4742 if ((*endp != ' ') && (*endp != '\0'))
4746 event = kzalloc(sizeof(*event), GFP_KERNEL);
4750 event->memcg = memcg;
4751 INIT_LIST_HEAD(&event->list);
4752 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4753 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4754 INIT_WORK(&event->remove, memcg_event_remove);
4762 event->eventfd = eventfd_ctx_fileget(efile.file);
4763 if (IS_ERR(event->eventfd)) {
4764 ret = PTR_ERR(event->eventfd);
4771 goto out_put_eventfd;
4774 /* the process need read permission on control file */
4775 /* AV: shouldn't we check that it's been opened for read instead? */
4776 ret = file_permission(cfile.file, MAY_READ);
4781 * Determine the event callbacks and set them in @event. This used
4782 * to be done via struct cftype but cgroup core no longer knows
4783 * about these events. The following is crude but the whole thing
4784 * is for compatibility anyway.
4786 * DO NOT ADD NEW FILES.
4788 name = cfile.file->f_path.dentry->d_name.name;
4790 if (!strcmp(name, "memory.usage_in_bytes")) {
4791 event->register_event = mem_cgroup_usage_register_event;
4792 event->unregister_event = mem_cgroup_usage_unregister_event;
4793 } else if (!strcmp(name, "memory.oom_control")) {
4794 event->register_event = mem_cgroup_oom_register_event;
4795 event->unregister_event = mem_cgroup_oom_unregister_event;
4796 } else if (!strcmp(name, "memory.pressure_level")) {
4797 event->register_event = vmpressure_register_event;
4798 event->unregister_event = vmpressure_unregister_event;
4799 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4800 event->register_event = memsw_cgroup_usage_register_event;
4801 event->unregister_event = memsw_cgroup_usage_unregister_event;
4808 * Verify @cfile should belong to @css. Also, remaining events are
4809 * automatically removed on cgroup destruction but the removal is
4810 * asynchronous, so take an extra ref on @css.
4812 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4813 &memory_cgrp_subsys);
4815 if (IS_ERR(cfile_css))
4817 if (cfile_css != css) {
4822 ret = event->register_event(memcg, event->eventfd, buf);
4826 vfs_poll(efile.file, &event->pt);
4828 spin_lock(&memcg->event_list_lock);
4829 list_add(&event->list, &memcg->event_list);
4830 spin_unlock(&memcg->event_list_lock);
4842 eventfd_ctx_put(event->eventfd);
4851 static struct cftype mem_cgroup_legacy_files[] = {
4853 .name = "usage_in_bytes",
4854 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4855 .read_u64 = mem_cgroup_read_u64,
4858 .name = "max_usage_in_bytes",
4859 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4860 .write = mem_cgroup_reset,
4861 .read_u64 = mem_cgroup_read_u64,
4864 .name = "limit_in_bytes",
4865 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4866 .write = mem_cgroup_write,
4867 .read_u64 = mem_cgroup_read_u64,
4870 .name = "soft_limit_in_bytes",
4871 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4872 .write = mem_cgroup_write,
4873 .read_u64 = mem_cgroup_read_u64,
4877 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4878 .write = mem_cgroup_reset,
4879 .read_u64 = mem_cgroup_read_u64,
4883 .seq_show = memcg_stat_show,
4886 .name = "force_empty",
4887 .write = mem_cgroup_force_empty_write,
4890 .name = "use_hierarchy",
4891 .write_u64 = mem_cgroup_hierarchy_write,
4892 .read_u64 = mem_cgroup_hierarchy_read,
4895 .name = "cgroup.event_control", /* XXX: for compat */
4896 .write = memcg_write_event_control,
4897 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4900 .name = "swappiness",
4901 .read_u64 = mem_cgroup_swappiness_read,
4902 .write_u64 = mem_cgroup_swappiness_write,
4905 .name = "move_charge_at_immigrate",
4906 .read_u64 = mem_cgroup_move_charge_read,
4907 .write_u64 = mem_cgroup_move_charge_write,
4910 .name = "oom_control",
4911 .seq_show = mem_cgroup_oom_control_read,
4912 .write_u64 = mem_cgroup_oom_control_write,
4913 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4916 .name = "pressure_level",
4920 .name = "numa_stat",
4921 .seq_show = memcg_numa_stat_show,
4925 .name = "kmem.limit_in_bytes",
4926 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4927 .write = mem_cgroup_write,
4928 .read_u64 = mem_cgroup_read_u64,
4931 .name = "kmem.usage_in_bytes",
4932 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4933 .read_u64 = mem_cgroup_read_u64,
4936 .name = "kmem.failcnt",
4937 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4938 .write = mem_cgroup_reset,
4939 .read_u64 = mem_cgroup_read_u64,
4942 .name = "kmem.max_usage_in_bytes",
4943 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4944 .write = mem_cgroup_reset,
4945 .read_u64 = mem_cgroup_read_u64,
4947 #if defined(CONFIG_MEMCG_KMEM) && \
4948 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
4950 .name = "kmem.slabinfo",
4951 .seq_show = memcg_slab_show,
4955 .name = "kmem.tcp.limit_in_bytes",
4956 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4957 .write = mem_cgroup_write,
4958 .read_u64 = mem_cgroup_read_u64,
4961 .name = "kmem.tcp.usage_in_bytes",
4962 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4963 .read_u64 = mem_cgroup_read_u64,
4966 .name = "kmem.tcp.failcnt",
4967 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4968 .write = mem_cgroup_reset,
4969 .read_u64 = mem_cgroup_read_u64,
4972 .name = "kmem.tcp.max_usage_in_bytes",
4973 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4974 .write = mem_cgroup_reset,
4975 .read_u64 = mem_cgroup_read_u64,
4977 { }, /* terminate */
4981 * Private memory cgroup IDR
4983 * Swap-out records and page cache shadow entries need to store memcg
4984 * references in constrained space, so we maintain an ID space that is
4985 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4986 * memory-controlled cgroups to 64k.
4988 * However, there usually are many references to the offline CSS after
4989 * the cgroup has been destroyed, such as page cache or reclaimable
4990 * slab objects, that don't need to hang on to the ID. We want to keep
4991 * those dead CSS from occupying IDs, or we might quickly exhaust the
4992 * relatively small ID space and prevent the creation of new cgroups
4993 * even when there are much fewer than 64k cgroups - possibly none.
4995 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4996 * be freed and recycled when it's no longer needed, which is usually
4997 * when the CSS is offlined.
4999 * The only exception to that are records of swapped out tmpfs/shmem
5000 * pages that need to be attributed to live ancestors on swapin. But
5001 * those references are manageable from userspace.
5004 static DEFINE_IDR(mem_cgroup_idr);
5006 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5008 if (memcg->id.id > 0) {
5009 idr_remove(&mem_cgroup_idr, memcg->id.id);
5014 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5017 refcount_add(n, &memcg->id.ref);
5020 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5022 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5023 mem_cgroup_id_remove(memcg);
5025 /* Memcg ID pins CSS */
5026 css_put(&memcg->css);
5030 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5032 mem_cgroup_id_put_many(memcg, 1);
5036 * mem_cgroup_from_id - look up a memcg from a memcg id
5037 * @id: the memcg id to look up
5039 * Caller must hold rcu_read_lock().
5041 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5043 WARN_ON_ONCE(!rcu_read_lock_held());
5044 return idr_find(&mem_cgroup_idr, id);
5047 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5049 struct mem_cgroup_per_node *pn;
5052 * This routine is called against possible nodes.
5053 * But it's BUG to call kmalloc() against offline node.
5055 * TODO: this routine can waste much memory for nodes which will
5056 * never be onlined. It's better to use memory hotplug callback
5059 if (!node_state(node, N_NORMAL_MEMORY))
5061 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5065 pn->lruvec_stat_local = alloc_percpu_gfp(struct lruvec_stat,
5066 GFP_KERNEL_ACCOUNT);
5067 if (!pn->lruvec_stat_local) {
5072 pn->lruvec_stat_cpu = alloc_percpu_gfp(struct batched_lruvec_stat,
5073 GFP_KERNEL_ACCOUNT);
5074 if (!pn->lruvec_stat_cpu) {
5075 free_percpu(pn->lruvec_stat_local);
5080 lruvec_init(&pn->lruvec);
5081 pn->usage_in_excess = 0;
5082 pn->on_tree = false;
5085 memcg->nodeinfo[node] = pn;
5089 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5091 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5096 free_percpu(pn->lruvec_stat_cpu);
5097 free_percpu(pn->lruvec_stat_local);
5101 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5106 free_mem_cgroup_per_node_info(memcg, node);
5107 free_percpu(memcg->vmstats_percpu);
5111 static void mem_cgroup_free(struct mem_cgroup *memcg)
5115 memcg_wb_domain_exit(memcg);
5117 * Flush percpu lruvec stats to guarantee the value
5118 * correctness on parent's and all ancestor levels.
5120 for_each_online_cpu(cpu)
5121 memcg_flush_lruvec_page_state(memcg, cpu);
5122 __mem_cgroup_free(memcg);
5125 static struct mem_cgroup *mem_cgroup_alloc(void)
5127 struct mem_cgroup *memcg;
5130 int __maybe_unused i;
5131 long error = -ENOMEM;
5133 size = sizeof(struct mem_cgroup);
5134 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5136 memcg = kzalloc(size, GFP_KERNEL);
5138 return ERR_PTR(error);
5140 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5141 1, MEM_CGROUP_ID_MAX,
5143 if (memcg->id.id < 0) {
5144 error = memcg->id.id;
5148 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5149 GFP_KERNEL_ACCOUNT);
5150 if (!memcg->vmstats_percpu)
5154 if (alloc_mem_cgroup_per_node_info(memcg, node))
5157 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5160 INIT_WORK(&memcg->high_work, high_work_func);
5161 INIT_LIST_HEAD(&memcg->oom_notify);
5162 mutex_init(&memcg->thresholds_lock);
5163 spin_lock_init(&memcg->move_lock);
5164 vmpressure_init(&memcg->vmpressure);
5165 INIT_LIST_HEAD(&memcg->event_list);
5166 spin_lock_init(&memcg->event_list_lock);
5167 memcg->socket_pressure = jiffies;
5168 #ifdef CONFIG_MEMCG_KMEM
5169 memcg->kmemcg_id = -1;
5170 INIT_LIST_HEAD(&memcg->objcg_list);
5172 #ifdef CONFIG_CGROUP_WRITEBACK
5173 INIT_LIST_HEAD(&memcg->cgwb_list);
5174 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5175 memcg->cgwb_frn[i].done =
5176 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5178 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5179 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5180 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5181 memcg->deferred_split_queue.split_queue_len = 0;
5183 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5186 mem_cgroup_id_remove(memcg);
5187 __mem_cgroup_free(memcg);
5188 return ERR_PTR(error);
5191 static struct cgroup_subsys_state * __ref
5192 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5194 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5195 struct mem_cgroup *memcg, *old_memcg;
5196 long error = -ENOMEM;
5198 old_memcg = set_active_memcg(parent);
5199 memcg = mem_cgroup_alloc();
5200 set_active_memcg(old_memcg);
5202 return ERR_CAST(memcg);
5204 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5205 memcg->soft_limit = PAGE_COUNTER_MAX;
5206 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5208 memcg->swappiness = mem_cgroup_swappiness(parent);
5209 memcg->oom_kill_disable = parent->oom_kill_disable;
5211 page_counter_init(&memcg->memory, &parent->memory);
5212 page_counter_init(&memcg->swap, &parent->swap);
5213 page_counter_init(&memcg->kmem, &parent->kmem);
5214 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5216 page_counter_init(&memcg->memory, NULL);
5217 page_counter_init(&memcg->swap, NULL);
5218 page_counter_init(&memcg->kmem, NULL);
5219 page_counter_init(&memcg->tcpmem, NULL);
5221 root_mem_cgroup = memcg;
5225 /* The following stuff does not apply to the root */
5226 error = memcg_online_kmem(memcg);
5230 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5231 static_branch_inc(&memcg_sockets_enabled_key);
5235 mem_cgroup_id_remove(memcg);
5236 mem_cgroup_free(memcg);
5237 return ERR_PTR(error);
5240 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5242 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5245 * A memcg must be visible for memcg_expand_shrinker_maps()
5246 * by the time the maps are allocated. So, we allocate maps
5247 * here, when for_each_mem_cgroup() can't skip it.
5249 if (memcg_alloc_shrinker_maps(memcg)) {
5250 mem_cgroup_id_remove(memcg);
5254 /* Online state pins memcg ID, memcg ID pins CSS */
5255 refcount_set(&memcg->id.ref, 1);
5260 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5262 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5263 struct mem_cgroup_event *event, *tmp;
5266 * Unregister events and notify userspace.
5267 * Notify userspace about cgroup removing only after rmdir of cgroup
5268 * directory to avoid race between userspace and kernelspace.
5270 spin_lock(&memcg->event_list_lock);
5271 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5272 list_del_init(&event->list);
5273 schedule_work(&event->remove);
5275 spin_unlock(&memcg->event_list_lock);
5277 page_counter_set_min(&memcg->memory, 0);
5278 page_counter_set_low(&memcg->memory, 0);
5280 memcg_offline_kmem(memcg);
5281 wb_memcg_offline(memcg);
5283 drain_all_stock(memcg);
5285 mem_cgroup_id_put(memcg);
5288 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5290 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5292 invalidate_reclaim_iterators(memcg);
5295 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5297 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5298 int __maybe_unused i;
5300 #ifdef CONFIG_CGROUP_WRITEBACK
5301 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5302 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5304 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5305 static_branch_dec(&memcg_sockets_enabled_key);
5307 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5308 static_branch_dec(&memcg_sockets_enabled_key);
5310 vmpressure_cleanup(&memcg->vmpressure);
5311 cancel_work_sync(&memcg->high_work);
5312 mem_cgroup_remove_from_trees(memcg);
5313 memcg_free_shrinker_maps(memcg);
5314 memcg_free_kmem(memcg);
5315 mem_cgroup_free(memcg);
5319 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5320 * @css: the target css
5322 * Reset the states of the mem_cgroup associated with @css. This is
5323 * invoked when the userland requests disabling on the default hierarchy
5324 * but the memcg is pinned through dependency. The memcg should stop
5325 * applying policies and should revert to the vanilla state as it may be
5326 * made visible again.
5328 * The current implementation only resets the essential configurations.
5329 * This needs to be expanded to cover all the visible parts.
5331 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5333 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5335 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5336 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5337 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5338 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5339 page_counter_set_min(&memcg->memory, 0);
5340 page_counter_set_low(&memcg->memory, 0);
5341 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5342 memcg->soft_limit = PAGE_COUNTER_MAX;
5343 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5344 memcg_wb_domain_size_changed(memcg);
5347 static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state *css, int cpu)
5349 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5350 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5351 struct memcg_vmstats_percpu *statc;
5355 statc = per_cpu_ptr(memcg->vmstats_percpu, cpu);
5357 for (i = 0; i < MEMCG_NR_STAT; i++) {
5359 * Collect the aggregated propagation counts of groups
5360 * below us. We're in a per-cpu loop here and this is
5361 * a global counter, so the first cycle will get them.
5363 delta = memcg->vmstats.state_pending[i];
5365 memcg->vmstats.state_pending[i] = 0;
5367 /* Add CPU changes on this level since the last flush */
5368 v = READ_ONCE(statc->state[i]);
5369 if (v != statc->state_prev[i]) {
5370 delta += v - statc->state_prev[i];
5371 statc->state_prev[i] = v;
5377 /* Aggregate counts on this level and propagate upwards */
5378 memcg->vmstats.state[i] += delta;
5380 parent->vmstats.state_pending[i] += delta;
5383 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
5384 delta = memcg->vmstats.events_pending[i];
5386 memcg->vmstats.events_pending[i] = 0;
5388 v = READ_ONCE(statc->events[i]);
5389 if (v != statc->events_prev[i]) {
5390 delta += v - statc->events_prev[i];
5391 statc->events_prev[i] = v;
5397 memcg->vmstats.events[i] += delta;
5399 parent->vmstats.events_pending[i] += delta;
5404 /* Handlers for move charge at task migration. */
5405 static int mem_cgroup_do_precharge(unsigned long count)
5409 /* Try a single bulk charge without reclaim first, kswapd may wake */
5410 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5412 mc.precharge += count;
5416 /* Try charges one by one with reclaim, but do not retry */
5418 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5432 enum mc_target_type {
5439 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5440 unsigned long addr, pte_t ptent)
5442 struct page *page = vm_normal_page(vma, addr, ptent);
5444 if (!page || !page_mapped(page))
5446 if (PageAnon(page)) {
5447 if (!(mc.flags & MOVE_ANON))
5450 if (!(mc.flags & MOVE_FILE))
5453 if (!get_page_unless_zero(page))
5459 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5460 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5461 pte_t ptent, swp_entry_t *entry)
5463 struct page *page = NULL;
5464 swp_entry_t ent = pte_to_swp_entry(ptent);
5466 if (!(mc.flags & MOVE_ANON))
5470 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5471 * a device and because they are not accessible by CPU they are store
5472 * as special swap entry in the CPU page table.
5474 if (is_device_private_entry(ent)) {
5475 page = device_private_entry_to_page(ent);
5477 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5478 * a refcount of 1 when free (unlike normal page)
5480 if (!page_ref_add_unless(page, 1, 1))
5485 if (non_swap_entry(ent))
5489 * Because lookup_swap_cache() updates some statistics counter,
5490 * we call find_get_page() with swapper_space directly.
5492 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5493 entry->val = ent.val;
5498 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5499 pte_t ptent, swp_entry_t *entry)
5505 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5506 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5508 if (!vma->vm_file) /* anonymous vma */
5510 if (!(mc.flags & MOVE_FILE))
5513 /* page is moved even if it's not RSS of this task(page-faulted). */
5514 /* shmem/tmpfs may report page out on swap: account for that too. */
5515 return find_get_incore_page(vma->vm_file->f_mapping,
5516 linear_page_index(vma, addr));
5520 * mem_cgroup_move_account - move account of the page
5522 * @compound: charge the page as compound or small page
5523 * @from: mem_cgroup which the page is moved from.
5524 * @to: mem_cgroup which the page is moved to. @from != @to.
5526 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5528 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5531 static int mem_cgroup_move_account(struct page *page,
5533 struct mem_cgroup *from,
5534 struct mem_cgroup *to)
5536 struct lruvec *from_vec, *to_vec;
5537 struct pglist_data *pgdat;
5538 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5541 VM_BUG_ON(from == to);
5542 VM_BUG_ON_PAGE(PageLRU(page), page);
5543 VM_BUG_ON(compound && !PageTransHuge(page));
5546 * Prevent mem_cgroup_migrate() from looking at
5547 * page's memory cgroup of its source page while we change it.
5550 if (!trylock_page(page))
5554 if (page_memcg(page) != from)
5557 pgdat = page_pgdat(page);
5558 from_vec = mem_cgroup_lruvec(from, pgdat);
5559 to_vec = mem_cgroup_lruvec(to, pgdat);
5561 lock_page_memcg(page);
5563 if (PageAnon(page)) {
5564 if (page_mapped(page)) {
5565 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5566 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5567 if (PageTransHuge(page)) {
5568 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5570 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5575 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5576 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5578 if (PageSwapBacked(page)) {
5579 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5580 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5583 if (page_mapped(page)) {
5584 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5585 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5588 if (PageDirty(page)) {
5589 struct address_space *mapping = page_mapping(page);
5591 if (mapping_can_writeback(mapping)) {
5592 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5594 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5600 if (PageWriteback(page)) {
5601 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5602 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5606 * All state has been migrated, let's switch to the new memcg.
5608 * It is safe to change page's memcg here because the page
5609 * is referenced, charged, isolated, and locked: we can't race
5610 * with (un)charging, migration, LRU putback, or anything else
5611 * that would rely on a stable page's memory cgroup.
5613 * Note that lock_page_memcg is a memcg lock, not a page lock,
5614 * to save space. As soon as we switch page's memory cgroup to a
5615 * new memcg that isn't locked, the above state can change
5616 * concurrently again. Make sure we're truly done with it.
5621 css_put(&from->css);
5623 page->memcg_data = (unsigned long)to;
5625 __unlock_page_memcg(from);
5629 local_irq_disable();
5630 mem_cgroup_charge_statistics(to, page, nr_pages);
5631 memcg_check_events(to, page);
5632 mem_cgroup_charge_statistics(from, page, -nr_pages);
5633 memcg_check_events(from, page);
5642 * get_mctgt_type - get target type of moving charge
5643 * @vma: the vma the pte to be checked belongs
5644 * @addr: the address corresponding to the pte to be checked
5645 * @ptent: the pte to be checked
5646 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5649 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5650 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5651 * move charge. if @target is not NULL, the page is stored in target->page
5652 * with extra refcnt got(Callers should handle it).
5653 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5654 * target for charge migration. if @target is not NULL, the entry is stored
5656 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5657 * (so ZONE_DEVICE page and thus not on the lru).
5658 * For now we such page is charge like a regular page would be as for all
5659 * intent and purposes it is just special memory taking the place of a
5662 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5664 * Called with pte lock held.
5667 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5668 unsigned long addr, pte_t ptent, union mc_target *target)
5670 struct page *page = NULL;
5671 enum mc_target_type ret = MC_TARGET_NONE;
5672 swp_entry_t ent = { .val = 0 };
5674 if (pte_present(ptent))
5675 page = mc_handle_present_pte(vma, addr, ptent);
5676 else if (is_swap_pte(ptent))
5677 page = mc_handle_swap_pte(vma, ptent, &ent);
5678 else if (pte_none(ptent))
5679 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5681 if (!page && !ent.val)
5685 * Do only loose check w/o serialization.
5686 * mem_cgroup_move_account() checks the page is valid or
5687 * not under LRU exclusion.
5689 if (page_memcg(page) == mc.from) {
5690 ret = MC_TARGET_PAGE;
5691 if (is_device_private_page(page))
5692 ret = MC_TARGET_DEVICE;
5694 target->page = page;
5696 if (!ret || !target)
5700 * There is a swap entry and a page doesn't exist or isn't charged.
5701 * But we cannot move a tail-page in a THP.
5703 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5704 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5705 ret = MC_TARGET_SWAP;
5712 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5714 * We don't consider PMD mapped swapping or file mapped pages because THP does
5715 * not support them for now.
5716 * Caller should make sure that pmd_trans_huge(pmd) is true.
5718 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5719 unsigned long addr, pmd_t pmd, union mc_target *target)
5721 struct page *page = NULL;
5722 enum mc_target_type ret = MC_TARGET_NONE;
5724 if (unlikely(is_swap_pmd(pmd))) {
5725 VM_BUG_ON(thp_migration_supported() &&
5726 !is_pmd_migration_entry(pmd));
5729 page = pmd_page(pmd);
5730 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5731 if (!(mc.flags & MOVE_ANON))
5733 if (page_memcg(page) == mc.from) {
5734 ret = MC_TARGET_PAGE;
5737 target->page = page;
5743 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5744 unsigned long addr, pmd_t pmd, union mc_target *target)
5746 return MC_TARGET_NONE;
5750 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5751 unsigned long addr, unsigned long end,
5752 struct mm_walk *walk)
5754 struct vm_area_struct *vma = walk->vma;
5758 ptl = pmd_trans_huge_lock(pmd, vma);
5761 * Note their can not be MC_TARGET_DEVICE for now as we do not
5762 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5763 * this might change.
5765 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5766 mc.precharge += HPAGE_PMD_NR;
5771 if (pmd_trans_unstable(pmd))
5773 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5774 for (; addr != end; pte++, addr += PAGE_SIZE)
5775 if (get_mctgt_type(vma, addr, *pte, NULL))
5776 mc.precharge++; /* increment precharge temporarily */
5777 pte_unmap_unlock(pte - 1, ptl);
5783 static const struct mm_walk_ops precharge_walk_ops = {
5784 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5787 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5789 unsigned long precharge;
5792 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5793 mmap_read_unlock(mm);
5795 precharge = mc.precharge;
5801 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5803 unsigned long precharge = mem_cgroup_count_precharge(mm);
5805 VM_BUG_ON(mc.moving_task);
5806 mc.moving_task = current;
5807 return mem_cgroup_do_precharge(precharge);
5810 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5811 static void __mem_cgroup_clear_mc(void)
5813 struct mem_cgroup *from = mc.from;
5814 struct mem_cgroup *to = mc.to;
5816 /* we must uncharge all the leftover precharges from mc.to */
5818 cancel_charge(mc.to, mc.precharge);
5822 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5823 * we must uncharge here.
5825 if (mc.moved_charge) {
5826 cancel_charge(mc.from, mc.moved_charge);
5827 mc.moved_charge = 0;
5829 /* we must fixup refcnts and charges */
5830 if (mc.moved_swap) {
5831 /* uncharge swap account from the old cgroup */
5832 if (!mem_cgroup_is_root(mc.from))
5833 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5835 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5838 * we charged both to->memory and to->memsw, so we
5839 * should uncharge to->memory.
5841 if (!mem_cgroup_is_root(mc.to))
5842 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5846 memcg_oom_recover(from);
5847 memcg_oom_recover(to);
5848 wake_up_all(&mc.waitq);
5851 static void mem_cgroup_clear_mc(void)
5853 struct mm_struct *mm = mc.mm;
5856 * we must clear moving_task before waking up waiters at the end of
5859 mc.moving_task = NULL;
5860 __mem_cgroup_clear_mc();
5861 spin_lock(&mc.lock);
5865 spin_unlock(&mc.lock);
5870 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5872 struct cgroup_subsys_state *css;
5873 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5874 struct mem_cgroup *from;
5875 struct task_struct *leader, *p;
5876 struct mm_struct *mm;
5877 unsigned long move_flags;
5880 /* charge immigration isn't supported on the default hierarchy */
5881 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5885 * Multi-process migrations only happen on the default hierarchy
5886 * where charge immigration is not used. Perform charge
5887 * immigration if @tset contains a leader and whine if there are
5891 cgroup_taskset_for_each_leader(leader, css, tset) {
5894 memcg = mem_cgroup_from_css(css);
5900 * We are now commited to this value whatever it is. Changes in this
5901 * tunable will only affect upcoming migrations, not the current one.
5902 * So we need to save it, and keep it going.
5904 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5908 from = mem_cgroup_from_task(p);
5910 VM_BUG_ON(from == memcg);
5912 mm = get_task_mm(p);
5915 /* We move charges only when we move a owner of the mm */
5916 if (mm->owner == p) {
5919 VM_BUG_ON(mc.precharge);
5920 VM_BUG_ON(mc.moved_charge);
5921 VM_BUG_ON(mc.moved_swap);
5923 spin_lock(&mc.lock);
5927 mc.flags = move_flags;
5928 spin_unlock(&mc.lock);
5929 /* We set mc.moving_task later */
5931 ret = mem_cgroup_precharge_mc(mm);
5933 mem_cgroup_clear_mc();
5940 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5943 mem_cgroup_clear_mc();
5946 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5947 unsigned long addr, unsigned long end,
5948 struct mm_walk *walk)
5951 struct vm_area_struct *vma = walk->vma;
5954 enum mc_target_type target_type;
5955 union mc_target target;
5958 ptl = pmd_trans_huge_lock(pmd, vma);
5960 if (mc.precharge < HPAGE_PMD_NR) {
5964 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5965 if (target_type == MC_TARGET_PAGE) {
5967 if (!isolate_lru_page(page)) {
5968 if (!mem_cgroup_move_account(page, true,
5970 mc.precharge -= HPAGE_PMD_NR;
5971 mc.moved_charge += HPAGE_PMD_NR;
5973 putback_lru_page(page);
5976 } else if (target_type == MC_TARGET_DEVICE) {
5978 if (!mem_cgroup_move_account(page, true,
5980 mc.precharge -= HPAGE_PMD_NR;
5981 mc.moved_charge += HPAGE_PMD_NR;
5989 if (pmd_trans_unstable(pmd))
5992 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5993 for (; addr != end; addr += PAGE_SIZE) {
5994 pte_t ptent = *(pte++);
5995 bool device = false;
6001 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6002 case MC_TARGET_DEVICE:
6005 case MC_TARGET_PAGE:
6008 * We can have a part of the split pmd here. Moving it
6009 * can be done but it would be too convoluted so simply
6010 * ignore such a partial THP and keep it in original
6011 * memcg. There should be somebody mapping the head.
6013 if (PageTransCompound(page))
6015 if (!device && isolate_lru_page(page))
6017 if (!mem_cgroup_move_account(page, false,
6020 /* we uncharge from mc.from later. */
6024 putback_lru_page(page);
6025 put: /* get_mctgt_type() gets the page */
6028 case MC_TARGET_SWAP:
6030 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6032 mem_cgroup_id_get_many(mc.to, 1);
6033 /* we fixup other refcnts and charges later. */
6041 pte_unmap_unlock(pte - 1, ptl);
6046 * We have consumed all precharges we got in can_attach().
6047 * We try charge one by one, but don't do any additional
6048 * charges to mc.to if we have failed in charge once in attach()
6051 ret = mem_cgroup_do_precharge(1);
6059 static const struct mm_walk_ops charge_walk_ops = {
6060 .pmd_entry = mem_cgroup_move_charge_pte_range,
6063 static void mem_cgroup_move_charge(void)
6065 lru_add_drain_all();
6067 * Signal lock_page_memcg() to take the memcg's move_lock
6068 * while we're moving its pages to another memcg. Then wait
6069 * for already started RCU-only updates to finish.
6071 atomic_inc(&mc.from->moving_account);
6074 if (unlikely(!mmap_read_trylock(mc.mm))) {
6076 * Someone who are holding the mmap_lock might be waiting in
6077 * waitq. So we cancel all extra charges, wake up all waiters,
6078 * and retry. Because we cancel precharges, we might not be able
6079 * to move enough charges, but moving charge is a best-effort
6080 * feature anyway, so it wouldn't be a big problem.
6082 __mem_cgroup_clear_mc();
6087 * When we have consumed all precharges and failed in doing
6088 * additional charge, the page walk just aborts.
6090 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6093 mmap_read_unlock(mc.mm);
6094 atomic_dec(&mc.from->moving_account);
6097 static void mem_cgroup_move_task(void)
6100 mem_cgroup_move_charge();
6101 mem_cgroup_clear_mc();
6104 #else /* !CONFIG_MMU */
6105 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6109 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6112 static void mem_cgroup_move_task(void)
6117 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6119 if (value == PAGE_COUNTER_MAX)
6120 seq_puts(m, "max\n");
6122 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6127 static u64 memory_current_read(struct cgroup_subsys_state *css,
6130 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6132 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6135 static int memory_min_show(struct seq_file *m, void *v)
6137 return seq_puts_memcg_tunable(m,
6138 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6141 static ssize_t memory_min_write(struct kernfs_open_file *of,
6142 char *buf, size_t nbytes, loff_t off)
6144 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6148 buf = strstrip(buf);
6149 err = page_counter_memparse(buf, "max", &min);
6153 page_counter_set_min(&memcg->memory, min);
6158 static int memory_low_show(struct seq_file *m, void *v)
6160 return seq_puts_memcg_tunable(m,
6161 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6164 static ssize_t memory_low_write(struct kernfs_open_file *of,
6165 char *buf, size_t nbytes, loff_t off)
6167 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6171 buf = strstrip(buf);
6172 err = page_counter_memparse(buf, "max", &low);
6176 page_counter_set_low(&memcg->memory, low);
6181 static int memory_high_show(struct seq_file *m, void *v)
6183 return seq_puts_memcg_tunable(m,
6184 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6187 static ssize_t memory_high_write(struct kernfs_open_file *of,
6188 char *buf, size_t nbytes, loff_t off)
6190 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6191 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6192 bool drained = false;
6196 buf = strstrip(buf);
6197 err = page_counter_memparse(buf, "max", &high);
6201 page_counter_set_high(&memcg->memory, high);
6204 unsigned long nr_pages = page_counter_read(&memcg->memory);
6205 unsigned long reclaimed;
6207 if (nr_pages <= high)
6210 if (signal_pending(current))
6214 drain_all_stock(memcg);
6219 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6222 if (!reclaimed && !nr_retries--)
6226 memcg_wb_domain_size_changed(memcg);
6230 static int memory_max_show(struct seq_file *m, void *v)
6232 return seq_puts_memcg_tunable(m,
6233 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6236 static ssize_t memory_max_write(struct kernfs_open_file *of,
6237 char *buf, size_t nbytes, loff_t off)
6239 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6240 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6241 bool drained = false;
6245 buf = strstrip(buf);
6246 err = page_counter_memparse(buf, "max", &max);
6250 xchg(&memcg->memory.max, max);
6253 unsigned long nr_pages = page_counter_read(&memcg->memory);
6255 if (nr_pages <= max)
6258 if (signal_pending(current))
6262 drain_all_stock(memcg);
6268 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6274 memcg_memory_event(memcg, MEMCG_OOM);
6275 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6279 memcg_wb_domain_size_changed(memcg);
6283 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6285 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6286 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6287 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6288 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6289 seq_printf(m, "oom_kill %lu\n",
6290 atomic_long_read(&events[MEMCG_OOM_KILL]));
6293 static int memory_events_show(struct seq_file *m, void *v)
6295 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6297 __memory_events_show(m, memcg->memory_events);
6301 static int memory_events_local_show(struct seq_file *m, void *v)
6303 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6305 __memory_events_show(m, memcg->memory_events_local);
6309 static int memory_stat_show(struct seq_file *m, void *v)
6311 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6314 buf = memory_stat_format(memcg);
6323 static inline unsigned long lruvec_page_state_output(struct lruvec *lruvec,
6326 return lruvec_page_state(lruvec, item) * memcg_page_state_unit(item);
6329 static int memory_numa_stat_show(struct seq_file *m, void *v)
6332 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6334 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6337 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6340 seq_printf(m, "%s", memory_stats[i].name);
6341 for_each_node_state(nid, N_MEMORY) {
6343 struct lruvec *lruvec;
6345 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6346 size = lruvec_page_state_output(lruvec,
6347 memory_stats[i].idx);
6348 seq_printf(m, " N%d=%llu", nid, size);
6357 static int memory_oom_group_show(struct seq_file *m, void *v)
6359 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6361 seq_printf(m, "%d\n", memcg->oom_group);
6366 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6367 char *buf, size_t nbytes, loff_t off)
6369 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6372 buf = strstrip(buf);
6376 ret = kstrtoint(buf, 0, &oom_group);
6380 if (oom_group != 0 && oom_group != 1)
6383 memcg->oom_group = oom_group;
6388 static struct cftype memory_files[] = {
6391 .flags = CFTYPE_NOT_ON_ROOT,
6392 .read_u64 = memory_current_read,
6396 .flags = CFTYPE_NOT_ON_ROOT,
6397 .seq_show = memory_min_show,
6398 .write = memory_min_write,
6402 .flags = CFTYPE_NOT_ON_ROOT,
6403 .seq_show = memory_low_show,
6404 .write = memory_low_write,
6408 .flags = CFTYPE_NOT_ON_ROOT,
6409 .seq_show = memory_high_show,
6410 .write = memory_high_write,
6414 .flags = CFTYPE_NOT_ON_ROOT,
6415 .seq_show = memory_max_show,
6416 .write = memory_max_write,
6420 .flags = CFTYPE_NOT_ON_ROOT,
6421 .file_offset = offsetof(struct mem_cgroup, events_file),
6422 .seq_show = memory_events_show,
6425 .name = "events.local",
6426 .flags = CFTYPE_NOT_ON_ROOT,
6427 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6428 .seq_show = memory_events_local_show,
6432 .seq_show = memory_stat_show,
6436 .name = "numa_stat",
6437 .seq_show = memory_numa_stat_show,
6441 .name = "oom.group",
6442 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6443 .seq_show = memory_oom_group_show,
6444 .write = memory_oom_group_write,
6449 struct cgroup_subsys memory_cgrp_subsys = {
6450 .css_alloc = mem_cgroup_css_alloc,
6451 .css_online = mem_cgroup_css_online,
6452 .css_offline = mem_cgroup_css_offline,
6453 .css_released = mem_cgroup_css_released,
6454 .css_free = mem_cgroup_css_free,
6455 .css_reset = mem_cgroup_css_reset,
6456 .css_rstat_flush = mem_cgroup_css_rstat_flush,
6457 .can_attach = mem_cgroup_can_attach,
6458 .cancel_attach = mem_cgroup_cancel_attach,
6459 .post_attach = mem_cgroup_move_task,
6460 .dfl_cftypes = memory_files,
6461 .legacy_cftypes = mem_cgroup_legacy_files,
6466 * This function calculates an individual cgroup's effective
6467 * protection which is derived from its own memory.min/low, its
6468 * parent's and siblings' settings, as well as the actual memory
6469 * distribution in the tree.
6471 * The following rules apply to the effective protection values:
6473 * 1. At the first level of reclaim, effective protection is equal to
6474 * the declared protection in memory.min and memory.low.
6476 * 2. To enable safe delegation of the protection configuration, at
6477 * subsequent levels the effective protection is capped to the
6478 * parent's effective protection.
6480 * 3. To make complex and dynamic subtrees easier to configure, the
6481 * user is allowed to overcommit the declared protection at a given
6482 * level. If that is the case, the parent's effective protection is
6483 * distributed to the children in proportion to how much protection
6484 * they have declared and how much of it they are utilizing.
6486 * This makes distribution proportional, but also work-conserving:
6487 * if one cgroup claims much more protection than it uses memory,
6488 * the unused remainder is available to its siblings.
6490 * 4. Conversely, when the declared protection is undercommitted at a
6491 * given level, the distribution of the larger parental protection
6492 * budget is NOT proportional. A cgroup's protection from a sibling
6493 * is capped to its own memory.min/low setting.
6495 * 5. However, to allow protecting recursive subtrees from each other
6496 * without having to declare each individual cgroup's fixed share
6497 * of the ancestor's claim to protection, any unutilized -
6498 * "floating" - protection from up the tree is distributed in
6499 * proportion to each cgroup's *usage*. This makes the protection
6500 * neutral wrt sibling cgroups and lets them compete freely over
6501 * the shared parental protection budget, but it protects the
6502 * subtree as a whole from neighboring subtrees.
6504 * Note that 4. and 5. are not in conflict: 4. is about protecting
6505 * against immediate siblings whereas 5. is about protecting against
6506 * neighboring subtrees.
6508 static unsigned long effective_protection(unsigned long usage,
6509 unsigned long parent_usage,
6510 unsigned long setting,
6511 unsigned long parent_effective,
6512 unsigned long siblings_protected)
6514 unsigned long protected;
6517 protected = min(usage, setting);
6519 * If all cgroups at this level combined claim and use more
6520 * protection then what the parent affords them, distribute
6521 * shares in proportion to utilization.
6523 * We are using actual utilization rather than the statically
6524 * claimed protection in order to be work-conserving: claimed
6525 * but unused protection is available to siblings that would
6526 * otherwise get a smaller chunk than what they claimed.
6528 if (siblings_protected > parent_effective)
6529 return protected * parent_effective / siblings_protected;
6532 * Ok, utilized protection of all children is within what the
6533 * parent affords them, so we know whatever this child claims
6534 * and utilizes is effectively protected.
6536 * If there is unprotected usage beyond this value, reclaim
6537 * will apply pressure in proportion to that amount.
6539 * If there is unutilized protection, the cgroup will be fully
6540 * shielded from reclaim, but we do return a smaller value for
6541 * protection than what the group could enjoy in theory. This
6542 * is okay. With the overcommit distribution above, effective
6543 * protection is always dependent on how memory is actually
6544 * consumed among the siblings anyway.
6549 * If the children aren't claiming (all of) the protection
6550 * afforded to them by the parent, distribute the remainder in
6551 * proportion to the (unprotected) memory of each cgroup. That
6552 * way, cgroups that aren't explicitly prioritized wrt each
6553 * other compete freely over the allowance, but they are
6554 * collectively protected from neighboring trees.
6556 * We're using unprotected memory for the weight so that if
6557 * some cgroups DO claim explicit protection, we don't protect
6558 * the same bytes twice.
6560 * Check both usage and parent_usage against the respective
6561 * protected values. One should imply the other, but they
6562 * aren't read atomically - make sure the division is sane.
6564 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6566 if (parent_effective > siblings_protected &&
6567 parent_usage > siblings_protected &&
6568 usage > protected) {
6569 unsigned long unclaimed;
6571 unclaimed = parent_effective - siblings_protected;
6572 unclaimed *= usage - protected;
6573 unclaimed /= parent_usage - siblings_protected;
6582 * mem_cgroup_protected - check if memory consumption is in the normal range
6583 * @root: the top ancestor of the sub-tree being checked
6584 * @memcg: the memory cgroup to check
6586 * WARNING: This function is not stateless! It can only be used as part
6587 * of a top-down tree iteration, not for isolated queries.
6589 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6590 struct mem_cgroup *memcg)
6592 unsigned long usage, parent_usage;
6593 struct mem_cgroup *parent;
6595 if (mem_cgroup_disabled())
6599 root = root_mem_cgroup;
6602 * Effective values of the reclaim targets are ignored so they
6603 * can be stale. Have a look at mem_cgroup_protection for more
6605 * TODO: calculation should be more robust so that we do not need
6606 * that special casing.
6611 usage = page_counter_read(&memcg->memory);
6615 parent = parent_mem_cgroup(memcg);
6616 /* No parent means a non-hierarchical mode on v1 memcg */
6620 if (parent == root) {
6621 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6622 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6626 parent_usage = page_counter_read(&parent->memory);
6628 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6629 READ_ONCE(memcg->memory.min),
6630 READ_ONCE(parent->memory.emin),
6631 atomic_long_read(&parent->memory.children_min_usage)));
6633 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6634 READ_ONCE(memcg->memory.low),
6635 READ_ONCE(parent->memory.elow),
6636 atomic_long_read(&parent->memory.children_low_usage)));
6639 static int __mem_cgroup_charge(struct page *page, struct mem_cgroup *memcg,
6642 unsigned int nr_pages = thp_nr_pages(page);
6645 ret = try_charge(memcg, gfp, nr_pages);
6649 css_get(&memcg->css);
6650 commit_charge(page, memcg);
6652 local_irq_disable();
6653 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6654 memcg_check_events(memcg, page);
6661 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6662 * @page: page to charge
6663 * @mm: mm context of the victim
6664 * @gfp_mask: reclaim mode
6666 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6667 * pages according to @gfp_mask if necessary.
6669 * Do not use this for pages allocated for swapin.
6671 * Returns 0 on success. Otherwise, an error code is returned.
6673 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6675 struct mem_cgroup *memcg;
6678 if (mem_cgroup_disabled())
6681 memcg = get_mem_cgroup_from_mm(mm);
6682 ret = __mem_cgroup_charge(page, memcg, gfp_mask);
6683 css_put(&memcg->css);
6689 * mem_cgroup_swapin_charge_page - charge a newly allocated page for swapin
6690 * @page: page to charge
6691 * @mm: mm context of the victim
6692 * @gfp: reclaim mode
6693 * @entry: swap entry for which the page is allocated
6695 * This function charges a page allocated for swapin. Please call this before
6696 * adding the page to the swapcache.
6698 * Returns 0 on success. Otherwise, an error code is returned.
6700 int mem_cgroup_swapin_charge_page(struct page *page, struct mm_struct *mm,
6701 gfp_t gfp, swp_entry_t entry)
6703 struct mem_cgroup *memcg;
6707 if (mem_cgroup_disabled())
6710 id = lookup_swap_cgroup_id(entry);
6712 memcg = mem_cgroup_from_id(id);
6713 if (!memcg || !css_tryget_online(&memcg->css))
6714 memcg = get_mem_cgroup_from_mm(mm);
6717 ret = __mem_cgroup_charge(page, memcg, gfp);
6719 css_put(&memcg->css);
6724 * mem_cgroup_swapin_uncharge_swap - uncharge swap slot
6725 * @entry: swap entry for which the page is charged
6727 * Call this function after successfully adding the charged page to swapcache.
6729 * Note: This function assumes the page for which swap slot is being uncharged
6732 void mem_cgroup_swapin_uncharge_swap(swp_entry_t entry)
6735 * Cgroup1's unified memory+swap counter has been charged with the
6736 * new swapcache page, finish the transfer by uncharging the swap
6737 * slot. The swap slot would also get uncharged when it dies, but
6738 * it can stick around indefinitely and we'd count the page twice
6741 * Cgroup2 has separate resource counters for memory and swap,
6742 * so this is a non-issue here. Memory and swap charge lifetimes
6743 * correspond 1:1 to page and swap slot lifetimes: we charge the
6744 * page to memory here, and uncharge swap when the slot is freed.
6746 if (!mem_cgroup_disabled() && do_memsw_account()) {
6748 * The swap entry might not get freed for a long time,
6749 * let's not wait for it. The page already received a
6750 * memory+swap charge, drop the swap entry duplicate.
6752 mem_cgroup_uncharge_swap(entry, 1);
6756 struct uncharge_gather {
6757 struct mem_cgroup *memcg;
6758 unsigned long nr_memory;
6759 unsigned long pgpgout;
6760 unsigned long nr_kmem;
6761 struct page *dummy_page;
6764 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6766 memset(ug, 0, sizeof(*ug));
6769 static void uncharge_batch(const struct uncharge_gather *ug)
6771 unsigned long flags;
6773 if (ug->nr_memory) {
6774 page_counter_uncharge(&ug->memcg->memory, ug->nr_memory);
6775 if (do_memsw_account())
6776 page_counter_uncharge(&ug->memcg->memsw, ug->nr_memory);
6777 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6778 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6779 memcg_oom_recover(ug->memcg);
6782 local_irq_save(flags);
6783 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6784 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_memory);
6785 memcg_check_events(ug->memcg, ug->dummy_page);
6786 local_irq_restore(flags);
6788 /* drop reference from uncharge_page */
6789 css_put(&ug->memcg->css);
6792 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6794 unsigned long nr_pages;
6795 struct mem_cgroup *memcg;
6796 struct obj_cgroup *objcg;
6798 VM_BUG_ON_PAGE(PageLRU(page), page);
6801 * Nobody should be changing or seriously looking at
6802 * page memcg or objcg at this point, we have fully
6803 * exclusive access to the page.
6805 if (PageMemcgKmem(page)) {
6806 objcg = __page_objcg(page);
6808 * This get matches the put at the end of the function and
6809 * kmem pages do not hold memcg references anymore.
6811 memcg = get_mem_cgroup_from_objcg(objcg);
6813 memcg = __page_memcg(page);
6819 if (ug->memcg != memcg) {
6822 uncharge_gather_clear(ug);
6825 ug->dummy_page = page;
6827 /* pairs with css_put in uncharge_batch */
6828 css_get(&memcg->css);
6831 nr_pages = compound_nr(page);
6833 if (PageMemcgKmem(page)) {
6834 ug->nr_memory += nr_pages;
6835 ug->nr_kmem += nr_pages;
6837 page->memcg_data = 0;
6838 obj_cgroup_put(objcg);
6840 /* LRU pages aren't accounted at the root level */
6841 if (!mem_cgroup_is_root(memcg))
6842 ug->nr_memory += nr_pages;
6845 page->memcg_data = 0;
6848 css_put(&memcg->css);
6852 * mem_cgroup_uncharge - uncharge a page
6853 * @page: page to uncharge
6855 * Uncharge a page previously charged with mem_cgroup_charge().
6857 void mem_cgroup_uncharge(struct page *page)
6859 struct uncharge_gather ug;
6861 if (mem_cgroup_disabled())
6864 /* Don't touch page->lru of any random page, pre-check: */
6865 if (!page_memcg(page))
6868 uncharge_gather_clear(&ug);
6869 uncharge_page(page, &ug);
6870 uncharge_batch(&ug);
6874 * mem_cgroup_uncharge_list - uncharge a list of page
6875 * @page_list: list of pages to uncharge
6877 * Uncharge a list of pages previously charged with
6878 * mem_cgroup_charge().
6880 void mem_cgroup_uncharge_list(struct list_head *page_list)
6882 struct uncharge_gather ug;
6885 if (mem_cgroup_disabled())
6888 uncharge_gather_clear(&ug);
6889 list_for_each_entry(page, page_list, lru)
6890 uncharge_page(page, &ug);
6892 uncharge_batch(&ug);
6896 * mem_cgroup_migrate - charge a page's replacement
6897 * @oldpage: currently circulating page
6898 * @newpage: replacement page
6900 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6901 * be uncharged upon free.
6903 * Both pages must be locked, @newpage->mapping must be set up.
6905 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6907 struct mem_cgroup *memcg;
6908 unsigned int nr_pages;
6909 unsigned long flags;
6911 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6912 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6913 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6914 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6917 if (mem_cgroup_disabled())
6920 /* Page cache replacement: new page already charged? */
6921 if (page_memcg(newpage))
6924 memcg = page_memcg(oldpage);
6925 VM_WARN_ON_ONCE_PAGE(!memcg, oldpage);
6929 /* Force-charge the new page. The old one will be freed soon */
6930 nr_pages = thp_nr_pages(newpage);
6932 page_counter_charge(&memcg->memory, nr_pages);
6933 if (do_memsw_account())
6934 page_counter_charge(&memcg->memsw, nr_pages);
6936 css_get(&memcg->css);
6937 commit_charge(newpage, memcg);
6939 local_irq_save(flags);
6940 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
6941 memcg_check_events(memcg, newpage);
6942 local_irq_restore(flags);
6945 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6946 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6948 void mem_cgroup_sk_alloc(struct sock *sk)
6950 struct mem_cgroup *memcg;
6952 if (!mem_cgroup_sockets_enabled)
6955 /* Do not associate the sock with unrelated interrupted task's memcg. */
6960 memcg = mem_cgroup_from_task(current);
6961 if (memcg == root_mem_cgroup)
6963 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6965 if (css_tryget(&memcg->css))
6966 sk->sk_memcg = memcg;
6971 void mem_cgroup_sk_free(struct sock *sk)
6974 css_put(&sk->sk_memcg->css);
6978 * mem_cgroup_charge_skmem - charge socket memory
6979 * @memcg: memcg to charge
6980 * @nr_pages: number of pages to charge
6982 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6983 * @memcg's configured limit, %false if the charge had to be forced.
6985 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6987 gfp_t gfp_mask = GFP_KERNEL;
6989 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6990 struct page_counter *fail;
6992 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6993 memcg->tcpmem_pressure = 0;
6996 page_counter_charge(&memcg->tcpmem, nr_pages);
6997 memcg->tcpmem_pressure = 1;
7001 /* Don't block in the packet receive path */
7003 gfp_mask = GFP_NOWAIT;
7005 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7007 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
7010 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
7015 * mem_cgroup_uncharge_skmem - uncharge socket memory
7016 * @memcg: memcg to uncharge
7017 * @nr_pages: number of pages to uncharge
7019 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7021 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7022 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7026 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7028 refill_stock(memcg, nr_pages);
7031 static int __init cgroup_memory(char *s)
7035 while ((token = strsep(&s, ",")) != NULL) {
7038 if (!strcmp(token, "nosocket"))
7039 cgroup_memory_nosocket = true;
7040 if (!strcmp(token, "nokmem"))
7041 cgroup_memory_nokmem = true;
7045 __setup("cgroup.memory=", cgroup_memory);
7048 * subsys_initcall() for memory controller.
7050 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7051 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7052 * basically everything that doesn't depend on a specific mem_cgroup structure
7053 * should be initialized from here.
7055 static int __init mem_cgroup_init(void)
7060 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7061 * used for per-memcg-per-cpu caching of per-node statistics. In order
7062 * to work fine, we should make sure that the overfill threshold can't
7063 * exceed S32_MAX / PAGE_SIZE.
7065 BUILD_BUG_ON(MEMCG_CHARGE_BATCH > S32_MAX / PAGE_SIZE);
7067 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7068 memcg_hotplug_cpu_dead);
7070 for_each_possible_cpu(cpu)
7071 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7074 for_each_node(node) {
7075 struct mem_cgroup_tree_per_node *rtpn;
7077 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7078 node_online(node) ? node : NUMA_NO_NODE);
7080 rtpn->rb_root = RB_ROOT;
7081 rtpn->rb_rightmost = NULL;
7082 spin_lock_init(&rtpn->lock);
7083 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7088 subsys_initcall(mem_cgroup_init);
7090 #ifdef CONFIG_MEMCG_SWAP
7091 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7093 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7095 * The root cgroup cannot be destroyed, so it's refcount must
7098 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7102 memcg = parent_mem_cgroup(memcg);
7104 memcg = root_mem_cgroup;
7110 * mem_cgroup_swapout - transfer a memsw charge to swap
7111 * @page: page whose memsw charge to transfer
7112 * @entry: swap entry to move the charge to
7114 * Transfer the memsw charge of @page to @entry.
7116 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7118 struct mem_cgroup *memcg, *swap_memcg;
7119 unsigned int nr_entries;
7120 unsigned short oldid;
7122 VM_BUG_ON_PAGE(PageLRU(page), page);
7123 VM_BUG_ON_PAGE(page_count(page), page);
7125 if (mem_cgroup_disabled())
7128 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7131 memcg = page_memcg(page);
7133 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7138 * In case the memcg owning these pages has been offlined and doesn't
7139 * have an ID allocated to it anymore, charge the closest online
7140 * ancestor for the swap instead and transfer the memory+swap charge.
7142 swap_memcg = mem_cgroup_id_get_online(memcg);
7143 nr_entries = thp_nr_pages(page);
7144 /* Get references for the tail pages, too */
7146 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7147 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7149 VM_BUG_ON_PAGE(oldid, page);
7150 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7152 page->memcg_data = 0;
7154 if (!mem_cgroup_is_root(memcg))
7155 page_counter_uncharge(&memcg->memory, nr_entries);
7157 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7158 if (!mem_cgroup_is_root(swap_memcg))
7159 page_counter_charge(&swap_memcg->memsw, nr_entries);
7160 page_counter_uncharge(&memcg->memsw, nr_entries);
7164 * Interrupts should be disabled here because the caller holds the
7165 * i_pages lock which is taken with interrupts-off. It is
7166 * important here to have the interrupts disabled because it is the
7167 * only synchronisation we have for updating the per-CPU variables.
7169 VM_BUG_ON(!irqs_disabled());
7170 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7171 memcg_check_events(memcg, page);
7173 css_put(&memcg->css);
7177 * mem_cgroup_try_charge_swap - try charging swap space for a page
7178 * @page: page being added to swap
7179 * @entry: swap entry to charge
7181 * Try to charge @page's memcg for the swap space at @entry.
7183 * Returns 0 on success, -ENOMEM on failure.
7185 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7187 unsigned int nr_pages = thp_nr_pages(page);
7188 struct page_counter *counter;
7189 struct mem_cgroup *memcg;
7190 unsigned short oldid;
7192 if (mem_cgroup_disabled())
7195 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7198 memcg = page_memcg(page);
7200 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7205 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7209 memcg = mem_cgroup_id_get_online(memcg);
7211 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7212 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7213 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7214 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7215 mem_cgroup_id_put(memcg);
7219 /* Get references for the tail pages, too */
7221 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7222 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7223 VM_BUG_ON_PAGE(oldid, page);
7224 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7230 * mem_cgroup_uncharge_swap - uncharge swap space
7231 * @entry: swap entry to uncharge
7232 * @nr_pages: the amount of swap space to uncharge
7234 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7236 struct mem_cgroup *memcg;
7239 id = swap_cgroup_record(entry, 0, nr_pages);
7241 memcg = mem_cgroup_from_id(id);
7243 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7244 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7245 page_counter_uncharge(&memcg->swap, nr_pages);
7247 page_counter_uncharge(&memcg->memsw, nr_pages);
7249 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7250 mem_cgroup_id_put_many(memcg, nr_pages);
7255 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7257 long nr_swap_pages = get_nr_swap_pages();
7259 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7260 return nr_swap_pages;
7261 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7262 nr_swap_pages = min_t(long, nr_swap_pages,
7263 READ_ONCE(memcg->swap.max) -
7264 page_counter_read(&memcg->swap));
7265 return nr_swap_pages;
7268 bool mem_cgroup_swap_full(struct page *page)
7270 struct mem_cgroup *memcg;
7272 VM_BUG_ON_PAGE(!PageLocked(page), page);
7276 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7279 memcg = page_memcg(page);
7283 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7284 unsigned long usage = page_counter_read(&memcg->swap);
7286 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7287 usage * 2 >= READ_ONCE(memcg->swap.max))
7294 static int __init setup_swap_account(char *s)
7296 if (!strcmp(s, "1"))
7297 cgroup_memory_noswap = false;
7298 else if (!strcmp(s, "0"))
7299 cgroup_memory_noswap = true;
7302 __setup("swapaccount=", setup_swap_account);
7304 static u64 swap_current_read(struct cgroup_subsys_state *css,
7307 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7309 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7312 static int swap_high_show(struct seq_file *m, void *v)
7314 return seq_puts_memcg_tunable(m,
7315 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7318 static ssize_t swap_high_write(struct kernfs_open_file *of,
7319 char *buf, size_t nbytes, loff_t off)
7321 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7325 buf = strstrip(buf);
7326 err = page_counter_memparse(buf, "max", &high);
7330 page_counter_set_high(&memcg->swap, high);
7335 static int swap_max_show(struct seq_file *m, void *v)
7337 return seq_puts_memcg_tunable(m,
7338 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7341 static ssize_t swap_max_write(struct kernfs_open_file *of,
7342 char *buf, size_t nbytes, loff_t off)
7344 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7348 buf = strstrip(buf);
7349 err = page_counter_memparse(buf, "max", &max);
7353 xchg(&memcg->swap.max, max);
7358 static int swap_events_show(struct seq_file *m, void *v)
7360 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7362 seq_printf(m, "high %lu\n",
7363 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7364 seq_printf(m, "max %lu\n",
7365 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7366 seq_printf(m, "fail %lu\n",
7367 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7372 static struct cftype swap_files[] = {
7374 .name = "swap.current",
7375 .flags = CFTYPE_NOT_ON_ROOT,
7376 .read_u64 = swap_current_read,
7379 .name = "swap.high",
7380 .flags = CFTYPE_NOT_ON_ROOT,
7381 .seq_show = swap_high_show,
7382 .write = swap_high_write,
7386 .flags = CFTYPE_NOT_ON_ROOT,
7387 .seq_show = swap_max_show,
7388 .write = swap_max_write,
7391 .name = "swap.events",
7392 .flags = CFTYPE_NOT_ON_ROOT,
7393 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7394 .seq_show = swap_events_show,
7399 static struct cftype memsw_files[] = {
7401 .name = "memsw.usage_in_bytes",
7402 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7403 .read_u64 = mem_cgroup_read_u64,
7406 .name = "memsw.max_usage_in_bytes",
7407 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7408 .write = mem_cgroup_reset,
7409 .read_u64 = mem_cgroup_read_u64,
7412 .name = "memsw.limit_in_bytes",
7413 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7414 .write = mem_cgroup_write,
7415 .read_u64 = mem_cgroup_read_u64,
7418 .name = "memsw.failcnt",
7419 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7420 .write = mem_cgroup_reset,
7421 .read_u64 = mem_cgroup_read_u64,
7423 { }, /* terminate */
7427 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7428 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7429 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7430 * boot parameter. This may result in premature OOPS inside
7431 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7433 static int __init mem_cgroup_swap_init(void)
7435 /* No memory control -> no swap control */
7436 if (mem_cgroup_disabled())
7437 cgroup_memory_noswap = true;
7439 if (cgroup_memory_noswap)
7442 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7443 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7447 core_initcall(mem_cgroup_swap_init);
7449 #endif /* CONFIG_MEMCG_SWAP */