1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* memcontrol.c - Memory Controller
4 * Copyright IBM Corporation, 2007
5 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
7 * Copyright 2007 OpenVZ SWsoft Inc
8 * Author: Pavel Emelianov <xemul@openvz.org>
11 * Copyright (C) 2009 Nokia Corporation
12 * Author: Kirill A. Shutemov
14 * Kernel Memory Controller
15 * Copyright (C) 2012 Parallels Inc. and Google Inc.
16 * Authors: Glauber Costa and Suleiman Souhlal
19 * Charge lifetime sanitation
20 * Lockless page tracking & accounting
21 * Unified hierarchy configuration model
22 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
24 * Per memcg lru locking
25 * Copyright (C) 2020 Alibaba, Inc, Alex Shi
28 #include <linux/page_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
31 #include <linux/pagewalk.h>
32 #include <linux/sched/mm.h>
33 #include <linux/shmem_fs.h>
34 #include <linux/hugetlb.h>
35 #include <linux/pagemap.h>
36 #include <linux/vm_event_item.h>
37 #include <linux/smp.h>
38 #include <linux/page-flags.h>
39 #include <linux/backing-dev.h>
40 #include <linux/bit_spinlock.h>
41 #include <linux/rcupdate.h>
42 #include <linux/limits.h>
43 #include <linux/export.h>
44 #include <linux/mutex.h>
45 #include <linux/rbtree.h>
46 #include <linux/slab.h>
47 #include <linux/swap.h>
48 #include <linux/swapops.h>
49 #include <linux/spinlock.h>
50 #include <linux/eventfd.h>
51 #include <linux/poll.h>
52 #include <linux/sort.h>
54 #include <linux/seq_file.h>
55 #include <linux/vmpressure.h>
56 #include <linux/mm_inline.h>
57 #include <linux/swap_cgroup.h>
58 #include <linux/cpu.h>
59 #include <linux/oom.h>
60 #include <linux/lockdep.h>
61 #include <linux/file.h>
62 #include <linux/tracehook.h>
63 #include <linux/psi.h>
64 #include <linux/seq_buf.h>
70 #include <linux/uaccess.h>
72 #include <trace/events/vmscan.h>
74 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
75 EXPORT_SYMBOL(memory_cgrp_subsys);
77 struct mem_cgroup *root_mem_cgroup __read_mostly;
79 /* Active memory cgroup to use from an interrupt context */
80 DEFINE_PER_CPU(struct mem_cgroup *, int_active_memcg);
82 /* Socket memory accounting disabled? */
83 static bool cgroup_memory_nosocket;
85 /* Kernel memory accounting disabled? */
86 static bool cgroup_memory_nokmem;
88 /* Whether the swap controller is active */
89 #ifdef CONFIG_MEMCG_SWAP
90 bool cgroup_memory_noswap __read_mostly;
92 #define cgroup_memory_noswap 1
95 #ifdef CONFIG_CGROUP_WRITEBACK
96 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
99 /* Whether legacy memory+swap accounting is active */
100 static bool do_memsw_account(void)
102 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_noswap;
105 #define THRESHOLDS_EVENTS_TARGET 128
106 #define SOFTLIMIT_EVENTS_TARGET 1024
109 * Cgroups above their limits are maintained in a RB-Tree, independent of
110 * their hierarchy representation
113 struct mem_cgroup_tree_per_node {
114 struct rb_root rb_root;
115 struct rb_node *rb_rightmost;
119 struct mem_cgroup_tree {
120 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
123 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
126 struct mem_cgroup_eventfd_list {
127 struct list_head list;
128 struct eventfd_ctx *eventfd;
132 * cgroup_event represents events which userspace want to receive.
134 struct mem_cgroup_event {
136 * memcg which the event belongs to.
138 struct mem_cgroup *memcg;
140 * eventfd to signal userspace about the event.
142 struct eventfd_ctx *eventfd;
144 * Each of these stored in a list by the cgroup.
146 struct list_head list;
148 * register_event() callback will be used to add new userspace
149 * waiter for changes related to this event. Use eventfd_signal()
150 * on eventfd to send notification to userspace.
152 int (*register_event)(struct mem_cgroup *memcg,
153 struct eventfd_ctx *eventfd, const char *args);
155 * unregister_event() callback will be called when userspace closes
156 * the eventfd or on cgroup removing. This callback must be set,
157 * if you want provide notification functionality.
159 void (*unregister_event)(struct mem_cgroup *memcg,
160 struct eventfd_ctx *eventfd);
162 * All fields below needed to unregister event when
163 * userspace closes eventfd.
166 wait_queue_head_t *wqh;
167 wait_queue_entry_t wait;
168 struct work_struct remove;
171 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
172 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
174 /* Stuffs for move charges at task migration. */
176 * Types of charges to be moved.
178 #define MOVE_ANON 0x1U
179 #define MOVE_FILE 0x2U
180 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
182 /* "mc" and its members are protected by cgroup_mutex */
183 static struct move_charge_struct {
184 spinlock_t lock; /* for from, to */
185 struct mm_struct *mm;
186 struct mem_cgroup *from;
187 struct mem_cgroup *to;
189 unsigned long precharge;
190 unsigned long moved_charge;
191 unsigned long moved_swap;
192 struct task_struct *moving_task; /* a task moving charges */
193 wait_queue_head_t waitq; /* a waitq for other context */
195 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
196 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
200 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
201 * limit reclaim to prevent infinite loops, if they ever occur.
203 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
204 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
206 /* for encoding cft->private value on file */
215 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
216 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
217 #define MEMFILE_ATTR(val) ((val) & 0xffff)
218 /* Used for OOM nofiier */
219 #define OOM_CONTROL (0)
222 * Iteration constructs for visiting all cgroups (under a tree). If
223 * loops are exited prematurely (break), mem_cgroup_iter_break() must
224 * be used for reference counting.
226 #define for_each_mem_cgroup_tree(iter, root) \
227 for (iter = mem_cgroup_iter(root, NULL, NULL); \
229 iter = mem_cgroup_iter(root, iter, NULL))
231 #define for_each_mem_cgroup(iter) \
232 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
234 iter = mem_cgroup_iter(NULL, iter, NULL))
236 static inline bool should_force_charge(void)
238 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
239 (current->flags & PF_EXITING);
242 /* Some nice accessors for the vmpressure. */
243 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
246 memcg = root_mem_cgroup;
247 return &memcg->vmpressure;
250 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
252 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
255 #ifdef CONFIG_MEMCG_KMEM
256 extern spinlock_t css_set_lock;
258 static void obj_cgroup_release(struct percpu_ref *ref)
260 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
261 struct mem_cgroup *memcg;
262 unsigned int nr_bytes;
263 unsigned int nr_pages;
267 * At this point all allocated objects are freed, and
268 * objcg->nr_charged_bytes can't have an arbitrary byte value.
269 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
271 * The following sequence can lead to it:
272 * 1) CPU0: objcg == stock->cached_objcg
273 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
274 * PAGE_SIZE bytes are charged
275 * 3) CPU1: a process from another memcg is allocating something,
276 * the stock if flushed,
277 * objcg->nr_charged_bytes = PAGE_SIZE - 92
278 * 5) CPU0: we do release this object,
279 * 92 bytes are added to stock->nr_bytes
280 * 6) CPU0: stock is flushed,
281 * 92 bytes are added to objcg->nr_charged_bytes
283 * In the result, nr_charged_bytes == PAGE_SIZE.
284 * This page will be uncharged in obj_cgroup_release().
286 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
287 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
288 nr_pages = nr_bytes >> PAGE_SHIFT;
290 spin_lock_irqsave(&css_set_lock, flags);
291 memcg = obj_cgroup_memcg(objcg);
293 __memcg_kmem_uncharge(memcg, nr_pages);
294 list_del(&objcg->list);
295 mem_cgroup_put(memcg);
296 spin_unlock_irqrestore(&css_set_lock, flags);
298 percpu_ref_exit(ref);
299 kfree_rcu(objcg, rcu);
302 static struct obj_cgroup *obj_cgroup_alloc(void)
304 struct obj_cgroup *objcg;
307 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
311 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
317 INIT_LIST_HEAD(&objcg->list);
321 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
322 struct mem_cgroup *parent)
324 struct obj_cgroup *objcg, *iter;
326 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
328 spin_lock_irq(&css_set_lock);
330 /* Move active objcg to the parent's list */
331 xchg(&objcg->memcg, parent);
332 css_get(&parent->css);
333 list_add(&objcg->list, &parent->objcg_list);
335 /* Move already reparented objcgs to the parent's list */
336 list_for_each_entry(iter, &memcg->objcg_list, list) {
337 css_get(&parent->css);
338 xchg(&iter->memcg, parent);
339 css_put(&memcg->css);
341 list_splice(&memcg->objcg_list, &parent->objcg_list);
343 spin_unlock_irq(&css_set_lock);
345 percpu_ref_kill(&objcg->refcnt);
349 * This will be used as a shrinker list's index.
350 * The main reason for not using cgroup id for this:
351 * this works better in sparse environments, where we have a lot of memcgs,
352 * but only a few kmem-limited. Or also, if we have, for instance, 200
353 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
354 * 200 entry array for that.
356 * The current size of the caches array is stored in memcg_nr_cache_ids. It
357 * will double each time we have to increase it.
359 static DEFINE_IDA(memcg_cache_ida);
360 int memcg_nr_cache_ids;
362 /* Protects memcg_nr_cache_ids */
363 static DECLARE_RWSEM(memcg_cache_ids_sem);
365 void memcg_get_cache_ids(void)
367 down_read(&memcg_cache_ids_sem);
370 void memcg_put_cache_ids(void)
372 up_read(&memcg_cache_ids_sem);
376 * MIN_SIZE is different than 1, because we would like to avoid going through
377 * the alloc/free process all the time. In a small machine, 4 kmem-limited
378 * cgroups is a reasonable guess. In the future, it could be a parameter or
379 * tunable, but that is strictly not necessary.
381 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
382 * this constant directly from cgroup, but it is understandable that this is
383 * better kept as an internal representation in cgroup.c. In any case, the
384 * cgrp_id space is not getting any smaller, and we don't have to necessarily
385 * increase ours as well if it increases.
387 #define MEMCG_CACHES_MIN_SIZE 4
388 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
391 * A lot of the calls to the cache allocation functions are expected to be
392 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
393 * conditional to this static branch, we'll have to allow modules that does
394 * kmem_cache_alloc and the such to see this symbol as well
396 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
397 EXPORT_SYMBOL(memcg_kmem_enabled_key);
400 static int memcg_shrinker_map_size;
401 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
403 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
405 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
408 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
409 int size, int old_size)
411 struct memcg_shrinker_map *new, *old;
414 lockdep_assert_held(&memcg_shrinker_map_mutex);
417 old = rcu_dereference_protected(
418 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
419 /* Not yet online memcg */
423 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
427 /* Set all old bits, clear all new bits */
428 memset(new->map, (int)0xff, old_size);
429 memset((void *)new->map + old_size, 0, size - old_size);
431 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
432 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
438 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
440 struct mem_cgroup_per_node *pn;
441 struct memcg_shrinker_map *map;
444 if (mem_cgroup_is_root(memcg))
448 pn = mem_cgroup_nodeinfo(memcg, nid);
449 map = rcu_dereference_protected(pn->shrinker_map, true);
452 rcu_assign_pointer(pn->shrinker_map, NULL);
456 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
458 struct memcg_shrinker_map *map;
459 int nid, size, ret = 0;
461 if (mem_cgroup_is_root(memcg))
464 mutex_lock(&memcg_shrinker_map_mutex);
465 size = memcg_shrinker_map_size;
467 map = kvzalloc_node(sizeof(*map) + size, GFP_KERNEL, nid);
469 memcg_free_shrinker_maps(memcg);
473 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
475 mutex_unlock(&memcg_shrinker_map_mutex);
480 int memcg_expand_shrinker_maps(int new_id)
482 int size, old_size, ret = 0;
483 struct mem_cgroup *memcg;
485 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
486 old_size = memcg_shrinker_map_size;
487 if (size <= old_size)
490 mutex_lock(&memcg_shrinker_map_mutex);
491 if (!root_mem_cgroup)
494 for_each_mem_cgroup(memcg) {
495 if (mem_cgroup_is_root(memcg))
497 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
499 mem_cgroup_iter_break(NULL, memcg);
505 memcg_shrinker_map_size = size;
506 mutex_unlock(&memcg_shrinker_map_mutex);
510 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
512 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
513 struct memcg_shrinker_map *map;
516 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
517 /* Pairs with smp mb in shrink_slab() */
518 smp_mb__before_atomic();
519 set_bit(shrinker_id, map->map);
525 * mem_cgroup_css_from_page - css of the memcg associated with a page
526 * @page: page of interest
528 * If memcg is bound to the default hierarchy, css of the memcg associated
529 * with @page is returned. The returned css remains associated with @page
530 * until it is released.
532 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
535 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
537 struct mem_cgroup *memcg;
539 memcg = page_memcg(page);
541 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
542 memcg = root_mem_cgroup;
548 * page_cgroup_ino - return inode number of the memcg a page is charged to
551 * Look up the closest online ancestor of the memory cgroup @page is charged to
552 * and return its inode number or 0 if @page is not charged to any cgroup. It
553 * is safe to call this function without holding a reference to @page.
555 * Note, this function is inherently racy, because there is nothing to prevent
556 * the cgroup inode from getting torn down and potentially reallocated a moment
557 * after page_cgroup_ino() returns, so it only should be used by callers that
558 * do not care (such as procfs interfaces).
560 ino_t page_cgroup_ino(struct page *page)
562 struct mem_cgroup *memcg;
563 unsigned long ino = 0;
566 memcg = page_memcg_check(page);
568 while (memcg && !(memcg->css.flags & CSS_ONLINE))
569 memcg = parent_mem_cgroup(memcg);
571 ino = cgroup_ino(memcg->css.cgroup);
576 static struct mem_cgroup_per_node *
577 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
579 int nid = page_to_nid(page);
581 return memcg->nodeinfo[nid];
584 static struct mem_cgroup_tree_per_node *
585 soft_limit_tree_node(int nid)
587 return soft_limit_tree.rb_tree_per_node[nid];
590 static struct mem_cgroup_tree_per_node *
591 soft_limit_tree_from_page(struct page *page)
593 int nid = page_to_nid(page);
595 return soft_limit_tree.rb_tree_per_node[nid];
598 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
599 struct mem_cgroup_tree_per_node *mctz,
600 unsigned long new_usage_in_excess)
602 struct rb_node **p = &mctz->rb_root.rb_node;
603 struct rb_node *parent = NULL;
604 struct mem_cgroup_per_node *mz_node;
605 bool rightmost = true;
610 mz->usage_in_excess = new_usage_in_excess;
611 if (!mz->usage_in_excess)
615 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
617 if (mz->usage_in_excess < mz_node->usage_in_excess) {
626 mctz->rb_rightmost = &mz->tree_node;
628 rb_link_node(&mz->tree_node, parent, p);
629 rb_insert_color(&mz->tree_node, &mctz->rb_root);
633 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
634 struct mem_cgroup_tree_per_node *mctz)
639 if (&mz->tree_node == mctz->rb_rightmost)
640 mctz->rb_rightmost = rb_prev(&mz->tree_node);
642 rb_erase(&mz->tree_node, &mctz->rb_root);
646 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
647 struct mem_cgroup_tree_per_node *mctz)
651 spin_lock_irqsave(&mctz->lock, flags);
652 __mem_cgroup_remove_exceeded(mz, mctz);
653 spin_unlock_irqrestore(&mctz->lock, flags);
656 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
658 unsigned long nr_pages = page_counter_read(&memcg->memory);
659 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
660 unsigned long excess = 0;
662 if (nr_pages > soft_limit)
663 excess = nr_pages - soft_limit;
668 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
670 unsigned long excess;
671 struct mem_cgroup_per_node *mz;
672 struct mem_cgroup_tree_per_node *mctz;
674 mctz = soft_limit_tree_from_page(page);
678 * Necessary to update all ancestors when hierarchy is used.
679 * because their event counter is not touched.
681 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
682 mz = mem_cgroup_page_nodeinfo(memcg, page);
683 excess = soft_limit_excess(memcg);
685 * We have to update the tree if mz is on RB-tree or
686 * mem is over its softlimit.
688 if (excess || mz->on_tree) {
691 spin_lock_irqsave(&mctz->lock, flags);
692 /* if on-tree, remove it */
694 __mem_cgroup_remove_exceeded(mz, mctz);
696 * Insert again. mz->usage_in_excess will be updated.
697 * If excess is 0, no tree ops.
699 __mem_cgroup_insert_exceeded(mz, mctz, excess);
700 spin_unlock_irqrestore(&mctz->lock, flags);
705 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
707 struct mem_cgroup_tree_per_node *mctz;
708 struct mem_cgroup_per_node *mz;
712 mz = mem_cgroup_nodeinfo(memcg, nid);
713 mctz = soft_limit_tree_node(nid);
715 mem_cgroup_remove_exceeded(mz, mctz);
719 static struct mem_cgroup_per_node *
720 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
722 struct mem_cgroup_per_node *mz;
726 if (!mctz->rb_rightmost)
727 goto done; /* Nothing to reclaim from */
729 mz = rb_entry(mctz->rb_rightmost,
730 struct mem_cgroup_per_node, tree_node);
732 * Remove the node now but someone else can add it back,
733 * we will to add it back at the end of reclaim to its correct
734 * position in the tree.
736 __mem_cgroup_remove_exceeded(mz, mctz);
737 if (!soft_limit_excess(mz->memcg) ||
738 !css_tryget(&mz->memcg->css))
744 static struct mem_cgroup_per_node *
745 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
747 struct mem_cgroup_per_node *mz;
749 spin_lock_irq(&mctz->lock);
750 mz = __mem_cgroup_largest_soft_limit_node(mctz);
751 spin_unlock_irq(&mctz->lock);
756 * __mod_memcg_state - update cgroup memory statistics
757 * @memcg: the memory cgroup
758 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
759 * @val: delta to add to the counter, can be negative
761 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
763 long x, threshold = MEMCG_CHARGE_BATCH;
765 if (mem_cgroup_disabled())
768 if (memcg_stat_item_in_bytes(idx))
769 threshold <<= PAGE_SHIFT;
771 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
772 if (unlikely(abs(x) > threshold)) {
773 struct mem_cgroup *mi;
776 * Batch local counters to keep them in sync with
777 * the hierarchical ones.
779 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
780 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
781 atomic_long_add(x, &mi->vmstats[idx]);
784 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
787 static struct mem_cgroup_per_node *
788 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
790 struct mem_cgroup *parent;
792 parent = parent_mem_cgroup(pn->memcg);
795 return mem_cgroup_nodeinfo(parent, nid);
798 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
801 struct mem_cgroup_per_node *pn;
802 struct mem_cgroup *memcg;
803 long x, threshold = MEMCG_CHARGE_BATCH;
805 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
809 __mod_memcg_state(memcg, idx, val);
812 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
814 if (vmstat_item_in_bytes(idx))
815 threshold <<= PAGE_SHIFT;
817 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
818 if (unlikely(abs(x) > threshold)) {
819 pg_data_t *pgdat = lruvec_pgdat(lruvec);
820 struct mem_cgroup_per_node *pi;
822 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
823 atomic_long_add(x, &pi->lruvec_stat[idx]);
826 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
830 * __mod_lruvec_state - update lruvec memory statistics
831 * @lruvec: the lruvec
832 * @idx: the stat item
833 * @val: delta to add to the counter, can be negative
835 * The lruvec is the intersection of the NUMA node and a cgroup. This
836 * function updates the all three counters that are affected by a
837 * change of state at this level: per-node, per-cgroup, per-lruvec.
839 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
843 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
845 /* Update memcg and lruvec */
846 if (!mem_cgroup_disabled())
847 __mod_memcg_lruvec_state(lruvec, idx, val);
850 void __mod_lruvec_page_state(struct page *page, enum node_stat_item idx,
853 struct page *head = compound_head(page); /* rmap on tail pages */
854 struct mem_cgroup *memcg = page_memcg(head);
855 pg_data_t *pgdat = page_pgdat(page);
856 struct lruvec *lruvec;
858 /* Untracked pages have no memcg, no lruvec. Update only the node */
860 __mod_node_page_state(pgdat, idx, val);
864 lruvec = mem_cgroup_lruvec(memcg, pgdat);
865 __mod_lruvec_state(lruvec, idx, val);
867 EXPORT_SYMBOL(__mod_lruvec_page_state);
869 void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val)
871 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
872 struct mem_cgroup *memcg;
873 struct lruvec *lruvec;
876 memcg = mem_cgroup_from_obj(p);
879 * Untracked pages have no memcg, no lruvec. Update only the
880 * node. If we reparent the slab objects to the root memcg,
881 * when we free the slab object, we need to update the per-memcg
882 * vmstats to keep it correct for the root memcg.
885 __mod_node_page_state(pgdat, idx, val);
887 lruvec = mem_cgroup_lruvec(memcg, pgdat);
888 __mod_lruvec_state(lruvec, idx, val);
894 * __count_memcg_events - account VM events in a cgroup
895 * @memcg: the memory cgroup
896 * @idx: the event item
897 * @count: the number of events that occured
899 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
904 if (mem_cgroup_disabled())
907 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
908 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
909 struct mem_cgroup *mi;
912 * Batch local counters to keep them in sync with
913 * the hierarchical ones.
915 __this_cpu_add(memcg->vmstats_local->events[idx], x);
916 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
917 atomic_long_add(x, &mi->vmevents[idx]);
920 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
923 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
925 return atomic_long_read(&memcg->vmevents[event]);
928 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
933 for_each_possible_cpu(cpu)
934 x += per_cpu(memcg->vmstats_local->events[event], cpu);
938 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
942 /* pagein of a big page is an event. So, ignore page size */
944 __count_memcg_events(memcg, PGPGIN, 1);
946 __count_memcg_events(memcg, PGPGOUT, 1);
947 nr_pages = -nr_pages; /* for event */
950 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
953 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
954 enum mem_cgroup_events_target target)
956 unsigned long val, next;
958 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
959 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
960 /* from time_after() in jiffies.h */
961 if ((long)(next - val) < 0) {
963 case MEM_CGROUP_TARGET_THRESH:
964 next = val + THRESHOLDS_EVENTS_TARGET;
966 case MEM_CGROUP_TARGET_SOFTLIMIT:
967 next = val + SOFTLIMIT_EVENTS_TARGET;
972 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
979 * Check events in order.
982 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
984 /* threshold event is triggered in finer grain than soft limit */
985 if (unlikely(mem_cgroup_event_ratelimit(memcg,
986 MEM_CGROUP_TARGET_THRESH))) {
989 do_softlimit = mem_cgroup_event_ratelimit(memcg,
990 MEM_CGROUP_TARGET_SOFTLIMIT);
991 mem_cgroup_threshold(memcg);
992 if (unlikely(do_softlimit))
993 mem_cgroup_update_tree(memcg, page);
997 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1000 * mm_update_next_owner() may clear mm->owner to NULL
1001 * if it races with swapoff, page migration, etc.
1002 * So this can be called with p == NULL.
1007 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1009 EXPORT_SYMBOL(mem_cgroup_from_task);
1012 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1013 * @mm: mm from which memcg should be extracted. It can be NULL.
1015 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
1016 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
1019 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1021 struct mem_cgroup *memcg;
1023 if (mem_cgroup_disabled())
1029 * Page cache insertions can happen withou an
1030 * actual mm context, e.g. during disk probing
1031 * on boot, loopback IO, acct() writes etc.
1034 memcg = root_mem_cgroup;
1036 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1037 if (unlikely(!memcg))
1038 memcg = root_mem_cgroup;
1040 } while (!css_tryget(&memcg->css));
1044 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1047 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
1048 * @page: page from which memcg should be extracted.
1050 * Obtain a reference on page->memcg and returns it if successful. Otherwise
1051 * root_mem_cgroup is returned.
1053 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
1055 struct mem_cgroup *memcg = page_memcg(page);
1057 if (mem_cgroup_disabled())
1061 /* Page should not get uncharged and freed memcg under us. */
1062 if (!memcg || WARN_ON_ONCE(!css_tryget(&memcg->css)))
1063 memcg = root_mem_cgroup;
1067 EXPORT_SYMBOL(get_mem_cgroup_from_page);
1069 static __always_inline struct mem_cgroup *active_memcg(void)
1072 return this_cpu_read(int_active_memcg);
1074 return current->active_memcg;
1077 static __always_inline struct mem_cgroup *get_active_memcg(void)
1079 struct mem_cgroup *memcg;
1082 memcg = active_memcg();
1084 /* current->active_memcg must hold a ref. */
1085 if (WARN_ON_ONCE(!css_tryget(&memcg->css)))
1086 memcg = root_mem_cgroup;
1088 memcg = current->active_memcg;
1095 static __always_inline bool memcg_kmem_bypass(void)
1097 /* Allow remote memcg charging from any context. */
1098 if (unlikely(active_memcg()))
1101 /* Memcg to charge can't be determined. */
1102 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
1109 * If active memcg is set, do not fallback to current->mm->memcg.
1111 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
1113 if (memcg_kmem_bypass())
1116 if (unlikely(active_memcg()))
1117 return get_active_memcg();
1119 return get_mem_cgroup_from_mm(current->mm);
1123 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1124 * @root: hierarchy root
1125 * @prev: previously returned memcg, NULL on first invocation
1126 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1128 * Returns references to children of the hierarchy below @root, or
1129 * @root itself, or %NULL after a full round-trip.
1131 * Caller must pass the return value in @prev on subsequent
1132 * invocations for reference counting, or use mem_cgroup_iter_break()
1133 * to cancel a hierarchy walk before the round-trip is complete.
1135 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1136 * in the hierarchy among all concurrent reclaimers operating on the
1139 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1140 struct mem_cgroup *prev,
1141 struct mem_cgroup_reclaim_cookie *reclaim)
1143 struct mem_cgroup_reclaim_iter *iter;
1144 struct cgroup_subsys_state *css = NULL;
1145 struct mem_cgroup *memcg = NULL;
1146 struct mem_cgroup *pos = NULL;
1148 if (mem_cgroup_disabled())
1152 root = root_mem_cgroup;
1154 if (prev && !reclaim)
1160 struct mem_cgroup_per_node *mz;
1162 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1165 if (prev && reclaim->generation != iter->generation)
1169 pos = READ_ONCE(iter->position);
1170 if (!pos || css_tryget(&pos->css))
1173 * css reference reached zero, so iter->position will
1174 * be cleared by ->css_released. However, we should not
1175 * rely on this happening soon, because ->css_released
1176 * is called from a work queue, and by busy-waiting we
1177 * might block it. So we clear iter->position right
1180 (void)cmpxchg(&iter->position, pos, NULL);
1188 css = css_next_descendant_pre(css, &root->css);
1191 * Reclaimers share the hierarchy walk, and a
1192 * new one might jump in right at the end of
1193 * the hierarchy - make sure they see at least
1194 * one group and restart from the beginning.
1202 * Verify the css and acquire a reference. The root
1203 * is provided by the caller, so we know it's alive
1204 * and kicking, and don't take an extra reference.
1206 memcg = mem_cgroup_from_css(css);
1208 if (css == &root->css)
1211 if (css_tryget(css))
1219 * The position could have already been updated by a competing
1220 * thread, so check that the value hasn't changed since we read
1221 * it to avoid reclaiming from the same cgroup twice.
1223 (void)cmpxchg(&iter->position, pos, memcg);
1231 reclaim->generation = iter->generation;
1236 if (prev && prev != root)
1237 css_put(&prev->css);
1243 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1244 * @root: hierarchy root
1245 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1247 void mem_cgroup_iter_break(struct mem_cgroup *root,
1248 struct mem_cgroup *prev)
1251 root = root_mem_cgroup;
1252 if (prev && prev != root)
1253 css_put(&prev->css);
1256 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1257 struct mem_cgroup *dead_memcg)
1259 struct mem_cgroup_reclaim_iter *iter;
1260 struct mem_cgroup_per_node *mz;
1263 for_each_node(nid) {
1264 mz = mem_cgroup_nodeinfo(from, nid);
1266 cmpxchg(&iter->position, dead_memcg, NULL);
1270 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1272 struct mem_cgroup *memcg = dead_memcg;
1273 struct mem_cgroup *last;
1276 __invalidate_reclaim_iterators(memcg, dead_memcg);
1278 } while ((memcg = parent_mem_cgroup(memcg)));
1281 * When cgruop1 non-hierarchy mode is used,
1282 * parent_mem_cgroup() does not walk all the way up to the
1283 * cgroup root (root_mem_cgroup). So we have to handle
1284 * dead_memcg from cgroup root separately.
1286 if (last != root_mem_cgroup)
1287 __invalidate_reclaim_iterators(root_mem_cgroup,
1292 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1293 * @memcg: hierarchy root
1294 * @fn: function to call for each task
1295 * @arg: argument passed to @fn
1297 * This function iterates over tasks attached to @memcg or to any of its
1298 * descendants and calls @fn for each task. If @fn returns a non-zero
1299 * value, the function breaks the iteration loop and returns the value.
1300 * Otherwise, it will iterate over all tasks and return 0.
1302 * This function must not be called for the root memory cgroup.
1304 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1305 int (*fn)(struct task_struct *, void *), void *arg)
1307 struct mem_cgroup *iter;
1310 BUG_ON(memcg == root_mem_cgroup);
1312 for_each_mem_cgroup_tree(iter, memcg) {
1313 struct css_task_iter it;
1314 struct task_struct *task;
1316 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1317 while (!ret && (task = css_task_iter_next(&it)))
1318 ret = fn(task, arg);
1319 css_task_iter_end(&it);
1321 mem_cgroup_iter_break(memcg, iter);
1328 #ifdef CONFIG_DEBUG_VM
1329 void lruvec_memcg_debug(struct lruvec *lruvec, struct page *page)
1331 struct mem_cgroup *memcg;
1333 if (mem_cgroup_disabled())
1336 memcg = page_memcg(page);
1339 VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != root_mem_cgroup, page);
1341 VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != memcg, page);
1346 * lock_page_lruvec - lock and return lruvec for a given page.
1349 * This series functions should be used in either conditions:
1350 * PageLRU is cleared or unset
1351 * or page->_refcount is zero
1352 * or page is locked.
1354 struct lruvec *lock_page_lruvec(struct page *page)
1356 struct lruvec *lruvec;
1357 struct pglist_data *pgdat = page_pgdat(page);
1360 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1361 spin_lock(&lruvec->lru_lock);
1364 lruvec_memcg_debug(lruvec, page);
1369 struct lruvec *lock_page_lruvec_irq(struct page *page)
1371 struct lruvec *lruvec;
1372 struct pglist_data *pgdat = page_pgdat(page);
1375 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1376 spin_lock_irq(&lruvec->lru_lock);
1379 lruvec_memcg_debug(lruvec, page);
1384 struct lruvec *lock_page_lruvec_irqsave(struct page *page, unsigned long *flags)
1386 struct lruvec *lruvec;
1387 struct pglist_data *pgdat = page_pgdat(page);
1390 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1391 spin_lock_irqsave(&lruvec->lru_lock, *flags);
1394 lruvec_memcg_debug(lruvec, page);
1400 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1401 * @lruvec: mem_cgroup per zone lru vector
1402 * @lru: index of lru list the page is sitting on
1403 * @zid: zone id of the accounted pages
1404 * @nr_pages: positive when adding or negative when removing
1406 * This function must be called under lru_lock, just before a page is added
1407 * to or just after a page is removed from an lru list (that ordering being
1408 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1410 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1411 int zid, int nr_pages)
1413 struct mem_cgroup_per_node *mz;
1414 unsigned long *lru_size;
1417 if (mem_cgroup_disabled())
1420 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1421 lru_size = &mz->lru_zone_size[zid][lru];
1424 *lru_size += nr_pages;
1427 if (WARN_ONCE(size < 0,
1428 "%s(%p, %d, %d): lru_size %ld\n",
1429 __func__, lruvec, lru, nr_pages, size)) {
1435 *lru_size += nr_pages;
1439 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1440 * @memcg: the memory cgroup
1442 * Returns the maximum amount of memory @mem can be charged with, in
1445 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1447 unsigned long margin = 0;
1448 unsigned long count;
1449 unsigned long limit;
1451 count = page_counter_read(&memcg->memory);
1452 limit = READ_ONCE(memcg->memory.max);
1454 margin = limit - count;
1456 if (do_memsw_account()) {
1457 count = page_counter_read(&memcg->memsw);
1458 limit = READ_ONCE(memcg->memsw.max);
1460 margin = min(margin, limit - count);
1469 * A routine for checking "mem" is under move_account() or not.
1471 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1472 * moving cgroups. This is for waiting at high-memory pressure
1475 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1477 struct mem_cgroup *from;
1478 struct mem_cgroup *to;
1481 * Unlike task_move routines, we access mc.to, mc.from not under
1482 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1484 spin_lock(&mc.lock);
1490 ret = mem_cgroup_is_descendant(from, memcg) ||
1491 mem_cgroup_is_descendant(to, memcg);
1493 spin_unlock(&mc.lock);
1497 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1499 if (mc.moving_task && current != mc.moving_task) {
1500 if (mem_cgroup_under_move(memcg)) {
1502 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1503 /* moving charge context might have finished. */
1506 finish_wait(&mc.waitq, &wait);
1513 struct memory_stat {
1519 static struct memory_stat memory_stats[] = {
1520 { "anon", PAGE_SIZE, NR_ANON_MAPPED },
1521 { "file", PAGE_SIZE, NR_FILE_PAGES },
1522 { "kernel_stack", 1024, NR_KERNEL_STACK_KB },
1523 { "pagetables", PAGE_SIZE, NR_PAGETABLE },
1524 { "percpu", 1, MEMCG_PERCPU_B },
1525 { "sock", PAGE_SIZE, MEMCG_SOCK },
1526 { "shmem", PAGE_SIZE, NR_SHMEM },
1527 { "file_mapped", PAGE_SIZE, NR_FILE_MAPPED },
1528 { "file_dirty", PAGE_SIZE, NR_FILE_DIRTY },
1529 { "file_writeback", PAGE_SIZE, NR_WRITEBACK },
1530 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1532 * The ratio will be initialized in memory_stats_init(). Because
1533 * on some architectures, the macro of HPAGE_PMD_SIZE is not
1534 * constant(e.g. powerpc).
1536 { "anon_thp", PAGE_SIZE, NR_ANON_THPS },
1537 { "file_thp", 0, NR_FILE_THPS },
1538 { "shmem_thp", 0, NR_SHMEM_THPS },
1540 { "inactive_anon", PAGE_SIZE, NR_INACTIVE_ANON },
1541 { "active_anon", PAGE_SIZE, NR_ACTIVE_ANON },
1542 { "inactive_file", PAGE_SIZE, NR_INACTIVE_FILE },
1543 { "active_file", PAGE_SIZE, NR_ACTIVE_FILE },
1544 { "unevictable", PAGE_SIZE, NR_UNEVICTABLE },
1547 * Note: The slab_reclaimable and slab_unreclaimable must be
1548 * together and slab_reclaimable must be in front.
1550 { "slab_reclaimable", 1, NR_SLAB_RECLAIMABLE_B },
1551 { "slab_unreclaimable", 1, NR_SLAB_UNRECLAIMABLE_B },
1553 /* The memory events */
1554 { "workingset_refault_anon", 1, WORKINGSET_REFAULT_ANON },
1555 { "workingset_refault_file", 1, WORKINGSET_REFAULT_FILE },
1556 { "workingset_activate_anon", 1, WORKINGSET_ACTIVATE_ANON },
1557 { "workingset_activate_file", 1, WORKINGSET_ACTIVATE_FILE },
1558 { "workingset_restore_anon", 1, WORKINGSET_RESTORE_ANON },
1559 { "workingset_restore_file", 1, WORKINGSET_RESTORE_FILE },
1560 { "workingset_nodereclaim", 1, WORKINGSET_NODERECLAIM },
1563 static int __init memory_stats_init(void)
1567 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1568 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1569 if (memory_stats[i].idx == NR_FILE_THPS ||
1570 memory_stats[i].idx == NR_SHMEM_THPS)
1571 memory_stats[i].ratio = HPAGE_PMD_SIZE;
1573 VM_BUG_ON(!memory_stats[i].ratio);
1574 VM_BUG_ON(memory_stats[i].idx >= MEMCG_NR_STAT);
1579 pure_initcall(memory_stats_init);
1581 static char *memory_stat_format(struct mem_cgroup *memcg)
1586 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1591 * Provide statistics on the state of the memory subsystem as
1592 * well as cumulative event counters that show past behavior.
1594 * This list is ordered following a combination of these gradients:
1595 * 1) generic big picture -> specifics and details
1596 * 2) reflecting userspace activity -> reflecting kernel heuristics
1598 * Current memory state:
1601 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1604 size = memcg_page_state(memcg, memory_stats[i].idx);
1605 size *= memory_stats[i].ratio;
1606 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1608 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1609 size = memcg_page_state(memcg, NR_SLAB_RECLAIMABLE_B) +
1610 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE_B);
1611 seq_buf_printf(&s, "slab %llu\n", size);
1615 /* Accumulated memory events */
1617 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1618 memcg_events(memcg, PGFAULT));
1619 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1620 memcg_events(memcg, PGMAJFAULT));
1621 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1622 memcg_events(memcg, PGREFILL));
1623 seq_buf_printf(&s, "pgscan %lu\n",
1624 memcg_events(memcg, PGSCAN_KSWAPD) +
1625 memcg_events(memcg, PGSCAN_DIRECT));
1626 seq_buf_printf(&s, "pgsteal %lu\n",
1627 memcg_events(memcg, PGSTEAL_KSWAPD) +
1628 memcg_events(memcg, PGSTEAL_DIRECT));
1629 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1630 memcg_events(memcg, PGACTIVATE));
1631 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1632 memcg_events(memcg, PGDEACTIVATE));
1633 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1634 memcg_events(memcg, PGLAZYFREE));
1635 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1636 memcg_events(memcg, PGLAZYFREED));
1638 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1639 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1640 memcg_events(memcg, THP_FAULT_ALLOC));
1641 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1642 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1643 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1645 /* The above should easily fit into one page */
1646 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1651 #define K(x) ((x) << (PAGE_SHIFT-10))
1653 * mem_cgroup_print_oom_context: Print OOM information relevant to
1654 * memory controller.
1655 * @memcg: The memory cgroup that went over limit
1656 * @p: Task that is going to be killed
1658 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1661 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1666 pr_cont(",oom_memcg=");
1667 pr_cont_cgroup_path(memcg->css.cgroup);
1669 pr_cont(",global_oom");
1671 pr_cont(",task_memcg=");
1672 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1678 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1679 * memory controller.
1680 * @memcg: The memory cgroup that went over limit
1682 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1686 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1687 K((u64)page_counter_read(&memcg->memory)),
1688 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1689 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1690 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1691 K((u64)page_counter_read(&memcg->swap)),
1692 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1694 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1695 K((u64)page_counter_read(&memcg->memsw)),
1696 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1697 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1698 K((u64)page_counter_read(&memcg->kmem)),
1699 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1702 pr_info("Memory cgroup stats for ");
1703 pr_cont_cgroup_path(memcg->css.cgroup);
1705 buf = memory_stat_format(memcg);
1713 * Return the memory (and swap, if configured) limit for a memcg.
1715 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1717 unsigned long max = READ_ONCE(memcg->memory.max);
1719 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1720 if (mem_cgroup_swappiness(memcg))
1721 max += min(READ_ONCE(memcg->swap.max),
1722 (unsigned long)total_swap_pages);
1724 if (mem_cgroup_swappiness(memcg)) {
1725 /* Calculate swap excess capacity from memsw limit */
1726 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1728 max += min(swap, (unsigned long)total_swap_pages);
1734 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1736 return page_counter_read(&memcg->memory);
1739 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1742 struct oom_control oc = {
1746 .gfp_mask = gfp_mask,
1751 if (mutex_lock_killable(&oom_lock))
1754 if (mem_cgroup_margin(memcg) >= (1 << order))
1758 * A few threads which were not waiting at mutex_lock_killable() can
1759 * fail to bail out. Therefore, check again after holding oom_lock.
1761 ret = should_force_charge() || out_of_memory(&oc);
1764 mutex_unlock(&oom_lock);
1768 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1771 unsigned long *total_scanned)
1773 struct mem_cgroup *victim = NULL;
1776 unsigned long excess;
1777 unsigned long nr_scanned;
1778 struct mem_cgroup_reclaim_cookie reclaim = {
1782 excess = soft_limit_excess(root_memcg);
1785 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1790 * If we have not been able to reclaim
1791 * anything, it might because there are
1792 * no reclaimable pages under this hierarchy
1797 * We want to do more targeted reclaim.
1798 * excess >> 2 is not to excessive so as to
1799 * reclaim too much, nor too less that we keep
1800 * coming back to reclaim from this cgroup
1802 if (total >= (excess >> 2) ||
1803 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1808 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1809 pgdat, &nr_scanned);
1810 *total_scanned += nr_scanned;
1811 if (!soft_limit_excess(root_memcg))
1814 mem_cgroup_iter_break(root_memcg, victim);
1818 #ifdef CONFIG_LOCKDEP
1819 static struct lockdep_map memcg_oom_lock_dep_map = {
1820 .name = "memcg_oom_lock",
1824 static DEFINE_SPINLOCK(memcg_oom_lock);
1827 * Check OOM-Killer is already running under our hierarchy.
1828 * If someone is running, return false.
1830 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1832 struct mem_cgroup *iter, *failed = NULL;
1834 spin_lock(&memcg_oom_lock);
1836 for_each_mem_cgroup_tree(iter, memcg) {
1837 if (iter->oom_lock) {
1839 * this subtree of our hierarchy is already locked
1840 * so we cannot give a lock.
1843 mem_cgroup_iter_break(memcg, iter);
1846 iter->oom_lock = true;
1851 * OK, we failed to lock the whole subtree so we have
1852 * to clean up what we set up to the failing subtree
1854 for_each_mem_cgroup_tree(iter, memcg) {
1855 if (iter == failed) {
1856 mem_cgroup_iter_break(memcg, iter);
1859 iter->oom_lock = false;
1862 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1864 spin_unlock(&memcg_oom_lock);
1869 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1871 struct mem_cgroup *iter;
1873 spin_lock(&memcg_oom_lock);
1874 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1875 for_each_mem_cgroup_tree(iter, memcg)
1876 iter->oom_lock = false;
1877 spin_unlock(&memcg_oom_lock);
1880 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1882 struct mem_cgroup *iter;
1884 spin_lock(&memcg_oom_lock);
1885 for_each_mem_cgroup_tree(iter, memcg)
1887 spin_unlock(&memcg_oom_lock);
1890 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1892 struct mem_cgroup *iter;
1895 * Be careful about under_oom underflows becase a child memcg
1896 * could have been added after mem_cgroup_mark_under_oom.
1898 spin_lock(&memcg_oom_lock);
1899 for_each_mem_cgroup_tree(iter, memcg)
1900 if (iter->under_oom > 0)
1902 spin_unlock(&memcg_oom_lock);
1905 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1907 struct oom_wait_info {
1908 struct mem_cgroup *memcg;
1909 wait_queue_entry_t wait;
1912 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1913 unsigned mode, int sync, void *arg)
1915 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1916 struct mem_cgroup *oom_wait_memcg;
1917 struct oom_wait_info *oom_wait_info;
1919 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1920 oom_wait_memcg = oom_wait_info->memcg;
1922 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1923 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1925 return autoremove_wake_function(wait, mode, sync, arg);
1928 static void memcg_oom_recover(struct mem_cgroup *memcg)
1931 * For the following lockless ->under_oom test, the only required
1932 * guarantee is that it must see the state asserted by an OOM when
1933 * this function is called as a result of userland actions
1934 * triggered by the notification of the OOM. This is trivially
1935 * achieved by invoking mem_cgroup_mark_under_oom() before
1936 * triggering notification.
1938 if (memcg && memcg->under_oom)
1939 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1949 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1951 enum oom_status ret;
1954 if (order > PAGE_ALLOC_COSTLY_ORDER)
1957 memcg_memory_event(memcg, MEMCG_OOM);
1960 * We are in the middle of the charge context here, so we
1961 * don't want to block when potentially sitting on a callstack
1962 * that holds all kinds of filesystem and mm locks.
1964 * cgroup1 allows disabling the OOM killer and waiting for outside
1965 * handling until the charge can succeed; remember the context and put
1966 * the task to sleep at the end of the page fault when all locks are
1969 * On the other hand, in-kernel OOM killer allows for an async victim
1970 * memory reclaim (oom_reaper) and that means that we are not solely
1971 * relying on the oom victim to make a forward progress and we can
1972 * invoke the oom killer here.
1974 * Please note that mem_cgroup_out_of_memory might fail to find a
1975 * victim and then we have to bail out from the charge path.
1977 if (memcg->oom_kill_disable) {
1978 if (!current->in_user_fault)
1980 css_get(&memcg->css);
1981 current->memcg_in_oom = memcg;
1982 current->memcg_oom_gfp_mask = mask;
1983 current->memcg_oom_order = order;
1988 mem_cgroup_mark_under_oom(memcg);
1990 locked = mem_cgroup_oom_trylock(memcg);
1993 mem_cgroup_oom_notify(memcg);
1995 mem_cgroup_unmark_under_oom(memcg);
1996 if (mem_cgroup_out_of_memory(memcg, mask, order))
2002 mem_cgroup_oom_unlock(memcg);
2008 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2009 * @handle: actually kill/wait or just clean up the OOM state
2011 * This has to be called at the end of a page fault if the memcg OOM
2012 * handler was enabled.
2014 * Memcg supports userspace OOM handling where failed allocations must
2015 * sleep on a waitqueue until the userspace task resolves the
2016 * situation. Sleeping directly in the charge context with all kinds
2017 * of locks held is not a good idea, instead we remember an OOM state
2018 * in the task and mem_cgroup_oom_synchronize() has to be called at
2019 * the end of the page fault to complete the OOM handling.
2021 * Returns %true if an ongoing memcg OOM situation was detected and
2022 * completed, %false otherwise.
2024 bool mem_cgroup_oom_synchronize(bool handle)
2026 struct mem_cgroup *memcg = current->memcg_in_oom;
2027 struct oom_wait_info owait;
2030 /* OOM is global, do not handle */
2037 owait.memcg = memcg;
2038 owait.wait.flags = 0;
2039 owait.wait.func = memcg_oom_wake_function;
2040 owait.wait.private = current;
2041 INIT_LIST_HEAD(&owait.wait.entry);
2043 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2044 mem_cgroup_mark_under_oom(memcg);
2046 locked = mem_cgroup_oom_trylock(memcg);
2049 mem_cgroup_oom_notify(memcg);
2051 if (locked && !memcg->oom_kill_disable) {
2052 mem_cgroup_unmark_under_oom(memcg);
2053 finish_wait(&memcg_oom_waitq, &owait.wait);
2054 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
2055 current->memcg_oom_order);
2058 mem_cgroup_unmark_under_oom(memcg);
2059 finish_wait(&memcg_oom_waitq, &owait.wait);
2063 mem_cgroup_oom_unlock(memcg);
2065 * There is no guarantee that an OOM-lock contender
2066 * sees the wakeups triggered by the OOM kill
2067 * uncharges. Wake any sleepers explicitely.
2069 memcg_oom_recover(memcg);
2072 current->memcg_in_oom = NULL;
2073 css_put(&memcg->css);
2078 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2079 * @victim: task to be killed by the OOM killer
2080 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2082 * Returns a pointer to a memory cgroup, which has to be cleaned up
2083 * by killing all belonging OOM-killable tasks.
2085 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2087 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2088 struct mem_cgroup *oom_domain)
2090 struct mem_cgroup *oom_group = NULL;
2091 struct mem_cgroup *memcg;
2093 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2097 oom_domain = root_mem_cgroup;
2101 memcg = mem_cgroup_from_task(victim);
2102 if (memcg == root_mem_cgroup)
2106 * If the victim task has been asynchronously moved to a different
2107 * memory cgroup, we might end up killing tasks outside oom_domain.
2108 * In this case it's better to ignore memory.group.oom.
2110 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2114 * Traverse the memory cgroup hierarchy from the victim task's
2115 * cgroup up to the OOMing cgroup (or root) to find the
2116 * highest-level memory cgroup with oom.group set.
2118 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2119 if (memcg->oom_group)
2122 if (memcg == oom_domain)
2127 css_get(&oom_group->css);
2134 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2136 pr_info("Tasks in ");
2137 pr_cont_cgroup_path(memcg->css.cgroup);
2138 pr_cont(" are going to be killed due to memory.oom.group set\n");
2142 * lock_page_memcg - lock a page and memcg binding
2145 * This function protects unlocked LRU pages from being moved to
2148 * It ensures lifetime of the returned memcg. Caller is responsible
2149 * for the lifetime of the page; __unlock_page_memcg() is available
2150 * when @page might get freed inside the locked section.
2152 struct mem_cgroup *lock_page_memcg(struct page *page)
2154 struct page *head = compound_head(page); /* rmap on tail pages */
2155 struct mem_cgroup *memcg;
2156 unsigned long flags;
2159 * The RCU lock is held throughout the transaction. The fast
2160 * path can get away without acquiring the memcg->move_lock
2161 * because page moving starts with an RCU grace period.
2163 * The RCU lock also protects the memcg from being freed when
2164 * the page state that is going to change is the only thing
2165 * preventing the page itself from being freed. E.g. writeback
2166 * doesn't hold a page reference and relies on PG_writeback to
2167 * keep off truncation, migration and so forth.
2171 if (mem_cgroup_disabled())
2174 memcg = page_memcg(head);
2175 if (unlikely(!memcg))
2178 #ifdef CONFIG_PROVE_LOCKING
2179 local_irq_save(flags);
2180 might_lock(&memcg->move_lock);
2181 local_irq_restore(flags);
2184 if (atomic_read(&memcg->moving_account) <= 0)
2187 spin_lock_irqsave(&memcg->move_lock, flags);
2188 if (memcg != page_memcg(head)) {
2189 spin_unlock_irqrestore(&memcg->move_lock, flags);
2194 * When charge migration first begins, we can have locked and
2195 * unlocked page stat updates happening concurrently. Track
2196 * the task who has the lock for unlock_page_memcg().
2198 memcg->move_lock_task = current;
2199 memcg->move_lock_flags = flags;
2203 EXPORT_SYMBOL(lock_page_memcg);
2206 * __unlock_page_memcg - unlock and unpin a memcg
2209 * Unlock and unpin a memcg returned by lock_page_memcg().
2211 void __unlock_page_memcg(struct mem_cgroup *memcg)
2213 if (memcg && memcg->move_lock_task == current) {
2214 unsigned long flags = memcg->move_lock_flags;
2216 memcg->move_lock_task = NULL;
2217 memcg->move_lock_flags = 0;
2219 spin_unlock_irqrestore(&memcg->move_lock, flags);
2226 * unlock_page_memcg - unlock a page and memcg binding
2229 void unlock_page_memcg(struct page *page)
2231 struct page *head = compound_head(page);
2233 __unlock_page_memcg(page_memcg(head));
2235 EXPORT_SYMBOL(unlock_page_memcg);
2237 struct memcg_stock_pcp {
2238 struct mem_cgroup *cached; /* this never be root cgroup */
2239 unsigned int nr_pages;
2241 #ifdef CONFIG_MEMCG_KMEM
2242 struct obj_cgroup *cached_objcg;
2243 unsigned int nr_bytes;
2246 struct work_struct work;
2247 unsigned long flags;
2248 #define FLUSHING_CACHED_CHARGE 0
2250 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2251 static DEFINE_MUTEX(percpu_charge_mutex);
2253 #ifdef CONFIG_MEMCG_KMEM
2254 static void drain_obj_stock(struct memcg_stock_pcp *stock);
2255 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2256 struct mem_cgroup *root_memcg);
2259 static inline void drain_obj_stock(struct memcg_stock_pcp *stock)
2262 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2263 struct mem_cgroup *root_memcg)
2270 * consume_stock: Try to consume stocked charge on this cpu.
2271 * @memcg: memcg to consume from.
2272 * @nr_pages: how many pages to charge.
2274 * The charges will only happen if @memcg matches the current cpu's memcg
2275 * stock, and at least @nr_pages are available in that stock. Failure to
2276 * service an allocation will refill the stock.
2278 * returns true if successful, false otherwise.
2280 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2282 struct memcg_stock_pcp *stock;
2283 unsigned long flags;
2286 if (nr_pages > MEMCG_CHARGE_BATCH)
2289 local_irq_save(flags);
2291 stock = this_cpu_ptr(&memcg_stock);
2292 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2293 stock->nr_pages -= nr_pages;
2297 local_irq_restore(flags);
2303 * Returns stocks cached in percpu and reset cached information.
2305 static void drain_stock(struct memcg_stock_pcp *stock)
2307 struct mem_cgroup *old = stock->cached;
2312 if (stock->nr_pages) {
2313 page_counter_uncharge(&old->memory, stock->nr_pages);
2314 if (do_memsw_account())
2315 page_counter_uncharge(&old->memsw, stock->nr_pages);
2316 stock->nr_pages = 0;
2320 stock->cached = NULL;
2323 static void drain_local_stock(struct work_struct *dummy)
2325 struct memcg_stock_pcp *stock;
2326 unsigned long flags;
2329 * The only protection from memory hotplug vs. drain_stock races is
2330 * that we always operate on local CPU stock here with IRQ disabled
2332 local_irq_save(flags);
2334 stock = this_cpu_ptr(&memcg_stock);
2335 drain_obj_stock(stock);
2337 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2339 local_irq_restore(flags);
2343 * Cache charges(val) to local per_cpu area.
2344 * This will be consumed by consume_stock() function, later.
2346 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2348 struct memcg_stock_pcp *stock;
2349 unsigned long flags;
2351 local_irq_save(flags);
2353 stock = this_cpu_ptr(&memcg_stock);
2354 if (stock->cached != memcg) { /* reset if necessary */
2356 css_get(&memcg->css);
2357 stock->cached = memcg;
2359 stock->nr_pages += nr_pages;
2361 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2364 local_irq_restore(flags);
2368 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2369 * of the hierarchy under it.
2371 static void drain_all_stock(struct mem_cgroup *root_memcg)
2375 /* If someone's already draining, avoid adding running more workers. */
2376 if (!mutex_trylock(&percpu_charge_mutex))
2379 * Notify other cpus that system-wide "drain" is running
2380 * We do not care about races with the cpu hotplug because cpu down
2381 * as well as workers from this path always operate on the local
2382 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2385 for_each_online_cpu(cpu) {
2386 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2387 struct mem_cgroup *memcg;
2391 memcg = stock->cached;
2392 if (memcg && stock->nr_pages &&
2393 mem_cgroup_is_descendant(memcg, root_memcg))
2395 if (obj_stock_flush_required(stock, root_memcg))
2400 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2402 drain_local_stock(&stock->work);
2404 schedule_work_on(cpu, &stock->work);
2408 mutex_unlock(&percpu_charge_mutex);
2411 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2413 struct memcg_stock_pcp *stock;
2414 struct mem_cgroup *memcg, *mi;
2416 stock = &per_cpu(memcg_stock, cpu);
2419 for_each_mem_cgroup(memcg) {
2422 for (i = 0; i < MEMCG_NR_STAT; i++) {
2426 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2428 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2429 atomic_long_add(x, &memcg->vmstats[i]);
2431 if (i >= NR_VM_NODE_STAT_ITEMS)
2434 for_each_node(nid) {
2435 struct mem_cgroup_per_node *pn;
2437 pn = mem_cgroup_nodeinfo(memcg, nid);
2438 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2441 atomic_long_add(x, &pn->lruvec_stat[i]);
2442 } while ((pn = parent_nodeinfo(pn, nid)));
2446 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2449 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2451 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2452 atomic_long_add(x, &memcg->vmevents[i]);
2459 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2460 unsigned int nr_pages,
2463 unsigned long nr_reclaimed = 0;
2466 unsigned long pflags;
2468 if (page_counter_read(&memcg->memory) <=
2469 READ_ONCE(memcg->memory.high))
2472 memcg_memory_event(memcg, MEMCG_HIGH);
2474 psi_memstall_enter(&pflags);
2475 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2477 psi_memstall_leave(&pflags);
2478 } while ((memcg = parent_mem_cgroup(memcg)) &&
2479 !mem_cgroup_is_root(memcg));
2481 return nr_reclaimed;
2484 static void high_work_func(struct work_struct *work)
2486 struct mem_cgroup *memcg;
2488 memcg = container_of(work, struct mem_cgroup, high_work);
2489 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2493 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2494 * enough to still cause a significant slowdown in most cases, while still
2495 * allowing diagnostics and tracing to proceed without becoming stuck.
2497 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2500 * When calculating the delay, we use these either side of the exponentiation to
2501 * maintain precision and scale to a reasonable number of jiffies (see the table
2504 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2505 * overage ratio to a delay.
2506 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2507 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2508 * to produce a reasonable delay curve.
2510 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2511 * reasonable delay curve compared to precision-adjusted overage, not
2512 * penalising heavily at first, but still making sure that growth beyond the
2513 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2514 * example, with a high of 100 megabytes:
2516 * +-------+------------------------+
2517 * | usage | time to allocate in ms |
2518 * +-------+------------------------+
2540 * +-------+------------------------+
2542 #define MEMCG_DELAY_PRECISION_SHIFT 20
2543 #define MEMCG_DELAY_SCALING_SHIFT 14
2545 static u64 calculate_overage(unsigned long usage, unsigned long high)
2553 * Prevent division by 0 in overage calculation by acting as if
2554 * it was a threshold of 1 page
2556 high = max(high, 1UL);
2558 overage = usage - high;
2559 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2560 return div64_u64(overage, high);
2563 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2565 u64 overage, max_overage = 0;
2568 overage = calculate_overage(page_counter_read(&memcg->memory),
2569 READ_ONCE(memcg->memory.high));
2570 max_overage = max(overage, max_overage);
2571 } while ((memcg = parent_mem_cgroup(memcg)) &&
2572 !mem_cgroup_is_root(memcg));
2577 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2579 u64 overage, max_overage = 0;
2582 overage = calculate_overage(page_counter_read(&memcg->swap),
2583 READ_ONCE(memcg->swap.high));
2585 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2586 max_overage = max(overage, max_overage);
2587 } while ((memcg = parent_mem_cgroup(memcg)) &&
2588 !mem_cgroup_is_root(memcg));
2594 * Get the number of jiffies that we should penalise a mischievous cgroup which
2595 * is exceeding its memory.high by checking both it and its ancestors.
2597 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2598 unsigned int nr_pages,
2601 unsigned long penalty_jiffies;
2607 * We use overage compared to memory.high to calculate the number of
2608 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2609 * fairly lenient on small overages, and increasingly harsh when the
2610 * memcg in question makes it clear that it has no intention of stopping
2611 * its crazy behaviour, so we exponentially increase the delay based on
2614 penalty_jiffies = max_overage * max_overage * HZ;
2615 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2616 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2619 * Factor in the task's own contribution to the overage, such that four
2620 * N-sized allocations are throttled approximately the same as one
2621 * 4N-sized allocation.
2623 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2624 * larger the current charge patch is than that.
2626 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2630 * Scheduled by try_charge() to be executed from the userland return path
2631 * and reclaims memory over the high limit.
2633 void mem_cgroup_handle_over_high(void)
2635 unsigned long penalty_jiffies;
2636 unsigned long pflags;
2637 unsigned long nr_reclaimed;
2638 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2639 int nr_retries = MAX_RECLAIM_RETRIES;
2640 struct mem_cgroup *memcg;
2641 bool in_retry = false;
2643 if (likely(!nr_pages))
2646 memcg = get_mem_cgroup_from_mm(current->mm);
2647 current->memcg_nr_pages_over_high = 0;
2651 * The allocating task should reclaim at least the batch size, but for
2652 * subsequent retries we only want to do what's necessary to prevent oom
2653 * or breaching resource isolation.
2655 * This is distinct from memory.max or page allocator behaviour because
2656 * memory.high is currently batched, whereas memory.max and the page
2657 * allocator run every time an allocation is made.
2659 nr_reclaimed = reclaim_high(memcg,
2660 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2664 * memory.high is breached and reclaim is unable to keep up. Throttle
2665 * allocators proactively to slow down excessive growth.
2667 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2668 mem_find_max_overage(memcg));
2670 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2671 swap_find_max_overage(memcg));
2674 * Clamp the max delay per usermode return so as to still keep the
2675 * application moving forwards and also permit diagnostics, albeit
2678 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2681 * Don't sleep if the amount of jiffies this memcg owes us is so low
2682 * that it's not even worth doing, in an attempt to be nice to those who
2683 * go only a small amount over their memory.high value and maybe haven't
2684 * been aggressively reclaimed enough yet.
2686 if (penalty_jiffies <= HZ / 100)
2690 * If reclaim is making forward progress but we're still over
2691 * memory.high, we want to encourage that rather than doing allocator
2694 if (nr_reclaimed || nr_retries--) {
2700 * If we exit early, we're guaranteed to die (since
2701 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2702 * need to account for any ill-begotten jiffies to pay them off later.
2704 psi_memstall_enter(&pflags);
2705 schedule_timeout_killable(penalty_jiffies);
2706 psi_memstall_leave(&pflags);
2709 css_put(&memcg->css);
2712 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2713 unsigned int nr_pages)
2715 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2716 int nr_retries = MAX_RECLAIM_RETRIES;
2717 struct mem_cgroup *mem_over_limit;
2718 struct page_counter *counter;
2719 enum oom_status oom_status;
2720 unsigned long nr_reclaimed;
2721 bool may_swap = true;
2722 bool drained = false;
2723 unsigned long pflags;
2725 if (mem_cgroup_is_root(memcg))
2728 if (consume_stock(memcg, nr_pages))
2731 if (!do_memsw_account() ||
2732 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2733 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2735 if (do_memsw_account())
2736 page_counter_uncharge(&memcg->memsw, batch);
2737 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2739 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2743 if (batch > nr_pages) {
2749 * Memcg doesn't have a dedicated reserve for atomic
2750 * allocations. But like the global atomic pool, we need to
2751 * put the burden of reclaim on regular allocation requests
2752 * and let these go through as privileged allocations.
2754 if (gfp_mask & __GFP_ATOMIC)
2758 * Unlike in global OOM situations, memcg is not in a physical
2759 * memory shortage. Allow dying and OOM-killed tasks to
2760 * bypass the last charges so that they can exit quickly and
2761 * free their memory.
2763 if (unlikely(should_force_charge()))
2767 * Prevent unbounded recursion when reclaim operations need to
2768 * allocate memory. This might exceed the limits temporarily,
2769 * but we prefer facilitating memory reclaim and getting back
2770 * under the limit over triggering OOM kills in these cases.
2772 if (unlikely(current->flags & PF_MEMALLOC))
2775 if (unlikely(task_in_memcg_oom(current)))
2778 if (!gfpflags_allow_blocking(gfp_mask))
2781 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2783 psi_memstall_enter(&pflags);
2784 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2785 gfp_mask, may_swap);
2786 psi_memstall_leave(&pflags);
2788 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2792 drain_all_stock(mem_over_limit);
2797 if (gfp_mask & __GFP_NORETRY)
2800 * Even though the limit is exceeded at this point, reclaim
2801 * may have been able to free some pages. Retry the charge
2802 * before killing the task.
2804 * Only for regular pages, though: huge pages are rather
2805 * unlikely to succeed so close to the limit, and we fall back
2806 * to regular pages anyway in case of failure.
2808 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2811 * At task move, charge accounts can be doubly counted. So, it's
2812 * better to wait until the end of task_move if something is going on.
2814 if (mem_cgroup_wait_acct_move(mem_over_limit))
2820 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2823 if (gfp_mask & __GFP_NOFAIL)
2826 if (fatal_signal_pending(current))
2830 * keep retrying as long as the memcg oom killer is able to make
2831 * a forward progress or bypass the charge if the oom killer
2832 * couldn't make any progress.
2834 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2835 get_order(nr_pages * PAGE_SIZE));
2836 switch (oom_status) {
2838 nr_retries = MAX_RECLAIM_RETRIES;
2846 if (!(gfp_mask & __GFP_NOFAIL))
2850 * The allocation either can't fail or will lead to more memory
2851 * being freed very soon. Allow memory usage go over the limit
2852 * temporarily by force charging it.
2854 page_counter_charge(&memcg->memory, nr_pages);
2855 if (do_memsw_account())
2856 page_counter_charge(&memcg->memsw, nr_pages);
2861 if (batch > nr_pages)
2862 refill_stock(memcg, batch - nr_pages);
2865 * If the hierarchy is above the normal consumption range, schedule
2866 * reclaim on returning to userland. We can perform reclaim here
2867 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2868 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2869 * not recorded as it most likely matches current's and won't
2870 * change in the meantime. As high limit is checked again before
2871 * reclaim, the cost of mismatch is negligible.
2874 bool mem_high, swap_high;
2876 mem_high = page_counter_read(&memcg->memory) >
2877 READ_ONCE(memcg->memory.high);
2878 swap_high = page_counter_read(&memcg->swap) >
2879 READ_ONCE(memcg->swap.high);
2881 /* Don't bother a random interrupted task */
2882 if (in_interrupt()) {
2884 schedule_work(&memcg->high_work);
2890 if (mem_high || swap_high) {
2892 * The allocating tasks in this cgroup will need to do
2893 * reclaim or be throttled to prevent further growth
2894 * of the memory or swap footprints.
2896 * Target some best-effort fairness between the tasks,
2897 * and distribute reclaim work and delay penalties
2898 * based on how much each task is actually allocating.
2900 current->memcg_nr_pages_over_high += batch;
2901 set_notify_resume(current);
2904 } while ((memcg = parent_mem_cgroup(memcg)));
2909 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2910 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2912 if (mem_cgroup_is_root(memcg))
2915 page_counter_uncharge(&memcg->memory, nr_pages);
2916 if (do_memsw_account())
2917 page_counter_uncharge(&memcg->memsw, nr_pages);
2921 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2923 VM_BUG_ON_PAGE(page_memcg(page), page);
2925 * Any of the following ensures page's memcg stability:
2929 * - lock_page_memcg()
2930 * - exclusive reference
2932 page->memcg_data = (unsigned long)memcg;
2935 #ifdef CONFIG_MEMCG_KMEM
2936 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2937 gfp_t gfp, bool new_page)
2939 unsigned int objects = objs_per_slab_page(s, page);
2940 unsigned long memcg_data;
2943 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2948 memcg_data = (unsigned long) vec | MEMCG_DATA_OBJCGS;
2951 * If the slab page is brand new and nobody can yet access
2952 * it's memcg_data, no synchronization is required and
2953 * memcg_data can be simply assigned.
2955 page->memcg_data = memcg_data;
2956 } else if (cmpxchg(&page->memcg_data, 0, memcg_data)) {
2958 * If the slab page is already in use, somebody can allocate
2959 * and assign obj_cgroups in parallel. In this case the existing
2960 * objcg vector should be reused.
2966 kmemleak_not_leak(vec);
2971 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2973 * A passed kernel object can be a slab object or a generic kernel page, so
2974 * different mechanisms for getting the memory cgroup pointer should be used.
2975 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2976 * can not know for sure how the kernel object is implemented.
2977 * mem_cgroup_from_obj() can be safely used in such cases.
2979 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2980 * cgroup_mutex, etc.
2982 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2986 if (mem_cgroup_disabled())
2989 page = virt_to_head_page(p);
2992 * Slab objects are accounted individually, not per-page.
2993 * Memcg membership data for each individual object is saved in
2994 * the page->obj_cgroups.
2996 if (page_objcgs_check(page)) {
2997 struct obj_cgroup *objcg;
3000 off = obj_to_index(page->slab_cache, page, p);
3001 objcg = page_objcgs(page)[off];
3003 return obj_cgroup_memcg(objcg);
3009 * page_memcg_check() is used here, because page_has_obj_cgroups()
3010 * check above could fail because the object cgroups vector wasn't set
3011 * at that moment, but it can be set concurrently.
3012 * page_memcg_check(page) will guarantee that a proper memory
3013 * cgroup pointer or NULL will be returned.
3015 return page_memcg_check(page);
3018 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
3020 struct obj_cgroup *objcg = NULL;
3021 struct mem_cgroup *memcg;
3023 if (memcg_kmem_bypass())
3027 if (unlikely(active_memcg()))
3028 memcg = active_memcg();
3030 memcg = mem_cgroup_from_task(current);
3032 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
3033 objcg = rcu_dereference(memcg->objcg);
3034 if (objcg && obj_cgroup_tryget(objcg))
3043 static int memcg_alloc_cache_id(void)
3048 id = ida_simple_get(&memcg_cache_ida,
3049 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3053 if (id < memcg_nr_cache_ids)
3057 * There's no space for the new id in memcg_caches arrays,
3058 * so we have to grow them.
3060 down_write(&memcg_cache_ids_sem);
3062 size = 2 * (id + 1);
3063 if (size < MEMCG_CACHES_MIN_SIZE)
3064 size = MEMCG_CACHES_MIN_SIZE;
3065 else if (size > MEMCG_CACHES_MAX_SIZE)
3066 size = MEMCG_CACHES_MAX_SIZE;
3068 err = memcg_update_all_list_lrus(size);
3070 memcg_nr_cache_ids = size;
3072 up_write(&memcg_cache_ids_sem);
3075 ida_simple_remove(&memcg_cache_ida, id);
3081 static void memcg_free_cache_id(int id)
3083 ida_simple_remove(&memcg_cache_ida, id);
3087 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
3088 * @memcg: memory cgroup to charge
3089 * @gfp: reclaim mode
3090 * @nr_pages: number of pages to charge
3092 * Returns 0 on success, an error code on failure.
3094 int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
3095 unsigned int nr_pages)
3097 struct page_counter *counter;
3100 ret = try_charge(memcg, gfp, nr_pages);
3104 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3105 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3108 * Enforce __GFP_NOFAIL allocation because callers are not
3109 * prepared to see failures and likely do not have any failure
3112 if (gfp & __GFP_NOFAIL) {
3113 page_counter_charge(&memcg->kmem, nr_pages);
3116 cancel_charge(memcg, nr_pages);
3123 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
3124 * @memcg: memcg to uncharge
3125 * @nr_pages: number of pages to uncharge
3127 void __memcg_kmem_uncharge(struct mem_cgroup *memcg, unsigned int nr_pages)
3129 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3130 page_counter_uncharge(&memcg->kmem, nr_pages);
3132 refill_stock(memcg, nr_pages);
3136 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3137 * @page: page to charge
3138 * @gfp: reclaim mode
3139 * @order: allocation order
3141 * Returns 0 on success, an error code on failure.
3143 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3145 struct mem_cgroup *memcg;
3148 memcg = get_mem_cgroup_from_current();
3149 if (memcg && !mem_cgroup_is_root(memcg)) {
3150 ret = __memcg_kmem_charge(memcg, gfp, 1 << order);
3152 page->memcg_data = (unsigned long)memcg |
3156 css_put(&memcg->css);
3162 * __memcg_kmem_uncharge_page: uncharge a kmem page
3163 * @page: page to uncharge
3164 * @order: allocation order
3166 void __memcg_kmem_uncharge_page(struct page *page, int order)
3168 struct mem_cgroup *memcg = page_memcg(page);
3169 unsigned int nr_pages = 1 << order;
3174 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3175 __memcg_kmem_uncharge(memcg, nr_pages);
3176 page->memcg_data = 0;
3177 css_put(&memcg->css);
3180 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3182 struct memcg_stock_pcp *stock;
3183 unsigned long flags;
3186 local_irq_save(flags);
3188 stock = this_cpu_ptr(&memcg_stock);
3189 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3190 stock->nr_bytes -= nr_bytes;
3194 local_irq_restore(flags);
3199 static void drain_obj_stock(struct memcg_stock_pcp *stock)
3201 struct obj_cgroup *old = stock->cached_objcg;
3206 if (stock->nr_bytes) {
3207 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3208 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3212 __memcg_kmem_uncharge(obj_cgroup_memcg(old), nr_pages);
3217 * The leftover is flushed to the centralized per-memcg value.
3218 * On the next attempt to refill obj stock it will be moved
3219 * to a per-cpu stock (probably, on an other CPU), see
3220 * refill_obj_stock().
3222 * How often it's flushed is a trade-off between the memory
3223 * limit enforcement accuracy and potential CPU contention,
3224 * so it might be changed in the future.
3226 atomic_add(nr_bytes, &old->nr_charged_bytes);
3227 stock->nr_bytes = 0;
3230 obj_cgroup_put(old);
3231 stock->cached_objcg = NULL;
3234 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3235 struct mem_cgroup *root_memcg)
3237 struct mem_cgroup *memcg;
3239 if (stock->cached_objcg) {
3240 memcg = obj_cgroup_memcg(stock->cached_objcg);
3241 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3248 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3250 struct memcg_stock_pcp *stock;
3251 unsigned long flags;
3253 local_irq_save(flags);
3255 stock = this_cpu_ptr(&memcg_stock);
3256 if (stock->cached_objcg != objcg) { /* reset if necessary */
3257 drain_obj_stock(stock);
3258 obj_cgroup_get(objcg);
3259 stock->cached_objcg = objcg;
3260 stock->nr_bytes = atomic_xchg(&objcg->nr_charged_bytes, 0);
3262 stock->nr_bytes += nr_bytes;
3264 if (stock->nr_bytes > PAGE_SIZE)
3265 drain_obj_stock(stock);
3267 local_irq_restore(flags);
3270 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3272 struct mem_cgroup *memcg;
3273 unsigned int nr_pages, nr_bytes;
3276 if (consume_obj_stock(objcg, size))
3280 * In theory, memcg->nr_charged_bytes can have enough
3281 * pre-charged bytes to satisfy the allocation. However,
3282 * flushing memcg->nr_charged_bytes requires two atomic
3283 * operations, and memcg->nr_charged_bytes can't be big,
3284 * so it's better to ignore it and try grab some new pages.
3285 * memcg->nr_charged_bytes will be flushed in
3286 * refill_obj_stock(), called from this function or
3287 * independently later.
3291 memcg = obj_cgroup_memcg(objcg);
3292 if (unlikely(!css_tryget(&memcg->css)))
3296 nr_pages = size >> PAGE_SHIFT;
3297 nr_bytes = size & (PAGE_SIZE - 1);
3302 ret = __memcg_kmem_charge(memcg, gfp, nr_pages);
3303 if (!ret && nr_bytes)
3304 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes);
3306 css_put(&memcg->css);
3310 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3312 refill_obj_stock(objcg, size);
3315 #endif /* CONFIG_MEMCG_KMEM */
3317 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3319 * Because page_memcg(head) is not set on compound tails, set it now.
3321 void mem_cgroup_split_huge_fixup(struct page *head)
3323 struct mem_cgroup *memcg = page_memcg(head);
3326 if (mem_cgroup_disabled())
3329 for (i = 1; i < HPAGE_PMD_NR; i++) {
3330 css_get(&memcg->css);
3331 head[i].memcg_data = (unsigned long)memcg;
3334 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3336 #ifdef CONFIG_MEMCG_SWAP
3338 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3339 * @entry: swap entry to be moved
3340 * @from: mem_cgroup which the entry is moved from
3341 * @to: mem_cgroup which the entry is moved to
3343 * It succeeds only when the swap_cgroup's record for this entry is the same
3344 * as the mem_cgroup's id of @from.
3346 * Returns 0 on success, -EINVAL on failure.
3348 * The caller must have charged to @to, IOW, called page_counter_charge() about
3349 * both res and memsw, and called css_get().
3351 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3352 struct mem_cgroup *from, struct mem_cgroup *to)
3354 unsigned short old_id, new_id;
3356 old_id = mem_cgroup_id(from);
3357 new_id = mem_cgroup_id(to);
3359 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3360 mod_memcg_state(from, MEMCG_SWAP, -1);
3361 mod_memcg_state(to, MEMCG_SWAP, 1);
3367 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3368 struct mem_cgroup *from, struct mem_cgroup *to)
3374 static DEFINE_MUTEX(memcg_max_mutex);
3376 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3377 unsigned long max, bool memsw)
3379 bool enlarge = false;
3380 bool drained = false;
3382 bool limits_invariant;
3383 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3386 if (signal_pending(current)) {
3391 mutex_lock(&memcg_max_mutex);
3393 * Make sure that the new limit (memsw or memory limit) doesn't
3394 * break our basic invariant rule memory.max <= memsw.max.
3396 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3397 max <= memcg->memsw.max;
3398 if (!limits_invariant) {
3399 mutex_unlock(&memcg_max_mutex);
3403 if (max > counter->max)
3405 ret = page_counter_set_max(counter, max);
3406 mutex_unlock(&memcg_max_mutex);
3412 drain_all_stock(memcg);
3417 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3418 GFP_KERNEL, !memsw)) {
3424 if (!ret && enlarge)
3425 memcg_oom_recover(memcg);
3430 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3432 unsigned long *total_scanned)
3434 unsigned long nr_reclaimed = 0;
3435 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3436 unsigned long reclaimed;
3438 struct mem_cgroup_tree_per_node *mctz;
3439 unsigned long excess;
3440 unsigned long nr_scanned;
3445 mctz = soft_limit_tree_node(pgdat->node_id);
3448 * Do not even bother to check the largest node if the root
3449 * is empty. Do it lockless to prevent lock bouncing. Races
3450 * are acceptable as soft limit is best effort anyway.
3452 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3456 * This loop can run a while, specially if mem_cgroup's continuously
3457 * keep exceeding their soft limit and putting the system under
3464 mz = mem_cgroup_largest_soft_limit_node(mctz);
3469 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3470 gfp_mask, &nr_scanned);
3471 nr_reclaimed += reclaimed;
3472 *total_scanned += nr_scanned;
3473 spin_lock_irq(&mctz->lock);
3474 __mem_cgroup_remove_exceeded(mz, mctz);
3477 * If we failed to reclaim anything from this memory cgroup
3478 * it is time to move on to the next cgroup
3482 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3484 excess = soft_limit_excess(mz->memcg);
3486 * One school of thought says that we should not add
3487 * back the node to the tree if reclaim returns 0.
3488 * But our reclaim could return 0, simply because due
3489 * to priority we are exposing a smaller subset of
3490 * memory to reclaim from. Consider this as a longer
3493 /* If excess == 0, no tree ops */
3494 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3495 spin_unlock_irq(&mctz->lock);
3496 css_put(&mz->memcg->css);
3499 * Could not reclaim anything and there are no more
3500 * mem cgroups to try or we seem to be looping without
3501 * reclaiming anything.
3503 if (!nr_reclaimed &&
3505 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3507 } while (!nr_reclaimed);
3509 css_put(&next_mz->memcg->css);
3510 return nr_reclaimed;
3514 * Reclaims as many pages from the given memcg as possible.
3516 * Caller is responsible for holding css reference for memcg.
3518 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3520 int nr_retries = MAX_RECLAIM_RETRIES;
3522 /* we call try-to-free pages for make this cgroup empty */
3523 lru_add_drain_all();
3525 drain_all_stock(memcg);
3527 /* try to free all pages in this cgroup */
3528 while (nr_retries && page_counter_read(&memcg->memory)) {
3531 if (signal_pending(current))
3534 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3538 /* maybe some writeback is necessary */
3539 congestion_wait(BLK_RW_ASYNC, HZ/10);
3547 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3548 char *buf, size_t nbytes,
3551 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3553 if (mem_cgroup_is_root(memcg))
3555 return mem_cgroup_force_empty(memcg) ?: nbytes;
3558 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3564 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3565 struct cftype *cft, u64 val)
3570 pr_warn_once("Non-hierarchical mode is deprecated. "
3571 "Please report your usecase to linux-mm@kvack.org if you "
3572 "depend on this functionality.\n");
3577 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3581 if (mem_cgroup_is_root(memcg)) {
3582 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3583 memcg_page_state(memcg, NR_ANON_MAPPED);
3585 val += memcg_page_state(memcg, MEMCG_SWAP);
3588 val = page_counter_read(&memcg->memory);
3590 val = page_counter_read(&memcg->memsw);
3603 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3606 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3607 struct page_counter *counter;
3609 switch (MEMFILE_TYPE(cft->private)) {
3611 counter = &memcg->memory;
3614 counter = &memcg->memsw;
3617 counter = &memcg->kmem;
3620 counter = &memcg->tcpmem;
3626 switch (MEMFILE_ATTR(cft->private)) {
3628 if (counter == &memcg->memory)
3629 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3630 if (counter == &memcg->memsw)
3631 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3632 return (u64)page_counter_read(counter) * PAGE_SIZE;
3634 return (u64)counter->max * PAGE_SIZE;
3636 return (u64)counter->watermark * PAGE_SIZE;
3638 return counter->failcnt;
3639 case RES_SOFT_LIMIT:
3640 return (u64)memcg->soft_limit * PAGE_SIZE;
3646 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3648 unsigned long stat[MEMCG_NR_STAT] = {0};
3649 struct mem_cgroup *mi;
3652 for_each_online_cpu(cpu)
3653 for (i = 0; i < MEMCG_NR_STAT; i++)
3654 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3656 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3657 for (i = 0; i < MEMCG_NR_STAT; i++)
3658 atomic_long_add(stat[i], &mi->vmstats[i]);
3660 for_each_node(node) {
3661 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3662 struct mem_cgroup_per_node *pi;
3664 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3667 for_each_online_cpu(cpu)
3668 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3670 pn->lruvec_stat_cpu->count[i], cpu);
3672 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3673 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3674 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3678 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3680 unsigned long events[NR_VM_EVENT_ITEMS];
3681 struct mem_cgroup *mi;
3684 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3687 for_each_online_cpu(cpu)
3688 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3689 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3692 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3693 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3694 atomic_long_add(events[i], &mi->vmevents[i]);
3697 #ifdef CONFIG_MEMCG_KMEM
3698 static int memcg_online_kmem(struct mem_cgroup *memcg)
3700 struct obj_cgroup *objcg;
3703 if (cgroup_memory_nokmem)
3706 BUG_ON(memcg->kmemcg_id >= 0);
3707 BUG_ON(memcg->kmem_state);
3709 memcg_id = memcg_alloc_cache_id();
3713 objcg = obj_cgroup_alloc();
3715 memcg_free_cache_id(memcg_id);
3718 objcg->memcg = memcg;
3719 rcu_assign_pointer(memcg->objcg, objcg);
3721 static_branch_enable(&memcg_kmem_enabled_key);
3723 memcg->kmemcg_id = memcg_id;
3724 memcg->kmem_state = KMEM_ONLINE;
3729 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3731 struct cgroup_subsys_state *css;
3732 struct mem_cgroup *parent, *child;
3735 if (memcg->kmem_state != KMEM_ONLINE)
3738 memcg->kmem_state = KMEM_ALLOCATED;
3740 parent = parent_mem_cgroup(memcg);
3742 parent = root_mem_cgroup;
3744 memcg_reparent_objcgs(memcg, parent);
3746 kmemcg_id = memcg->kmemcg_id;
3747 BUG_ON(kmemcg_id < 0);
3750 * Change kmemcg_id of this cgroup and all its descendants to the
3751 * parent's id, and then move all entries from this cgroup's list_lrus
3752 * to ones of the parent. After we have finished, all list_lrus
3753 * corresponding to this cgroup are guaranteed to remain empty. The
3754 * ordering is imposed by list_lru_node->lock taken by
3755 * memcg_drain_all_list_lrus().
3757 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3758 css_for_each_descendant_pre(css, &memcg->css) {
3759 child = mem_cgroup_from_css(css);
3760 BUG_ON(child->kmemcg_id != kmemcg_id);
3761 child->kmemcg_id = parent->kmemcg_id;
3765 memcg_drain_all_list_lrus(kmemcg_id, parent);
3767 memcg_free_cache_id(kmemcg_id);
3770 static void memcg_free_kmem(struct mem_cgroup *memcg)
3772 /* css_alloc() failed, offlining didn't happen */
3773 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3774 memcg_offline_kmem(memcg);
3777 static int memcg_online_kmem(struct mem_cgroup *memcg)
3781 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3784 static void memcg_free_kmem(struct mem_cgroup *memcg)
3787 #endif /* CONFIG_MEMCG_KMEM */
3789 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3794 mutex_lock(&memcg_max_mutex);
3795 ret = page_counter_set_max(&memcg->kmem, max);
3796 mutex_unlock(&memcg_max_mutex);
3800 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3804 mutex_lock(&memcg_max_mutex);
3806 ret = page_counter_set_max(&memcg->tcpmem, max);
3810 if (!memcg->tcpmem_active) {
3812 * The active flag needs to be written after the static_key
3813 * update. This is what guarantees that the socket activation
3814 * function is the last one to run. See mem_cgroup_sk_alloc()
3815 * for details, and note that we don't mark any socket as
3816 * belonging to this memcg until that flag is up.
3818 * We need to do this, because static_keys will span multiple
3819 * sites, but we can't control their order. If we mark a socket
3820 * as accounted, but the accounting functions are not patched in
3821 * yet, we'll lose accounting.
3823 * We never race with the readers in mem_cgroup_sk_alloc(),
3824 * because when this value change, the code to process it is not
3827 static_branch_inc(&memcg_sockets_enabled_key);
3828 memcg->tcpmem_active = true;
3831 mutex_unlock(&memcg_max_mutex);
3836 * The user of this function is...
3839 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3840 char *buf, size_t nbytes, loff_t off)
3842 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3843 unsigned long nr_pages;
3846 buf = strstrip(buf);
3847 ret = page_counter_memparse(buf, "-1", &nr_pages);
3851 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3853 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3857 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3859 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3862 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3865 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3866 "Please report your usecase to linux-mm@kvack.org if you "
3867 "depend on this functionality.\n");
3868 ret = memcg_update_kmem_max(memcg, nr_pages);
3871 ret = memcg_update_tcp_max(memcg, nr_pages);
3875 case RES_SOFT_LIMIT:
3876 memcg->soft_limit = nr_pages;
3880 return ret ?: nbytes;
3883 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3884 size_t nbytes, loff_t off)
3886 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3887 struct page_counter *counter;
3889 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3891 counter = &memcg->memory;
3894 counter = &memcg->memsw;
3897 counter = &memcg->kmem;
3900 counter = &memcg->tcpmem;
3906 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3908 page_counter_reset_watermark(counter);
3911 counter->failcnt = 0;
3920 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3923 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3927 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3928 struct cftype *cft, u64 val)
3930 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3932 if (val & ~MOVE_MASK)
3936 * No kind of locking is needed in here, because ->can_attach() will
3937 * check this value once in the beginning of the process, and then carry
3938 * on with stale data. This means that changes to this value will only
3939 * affect task migrations starting after the change.
3941 memcg->move_charge_at_immigrate = val;
3945 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3946 struct cftype *cft, u64 val)
3954 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3955 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3956 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3958 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3959 int nid, unsigned int lru_mask, bool tree)
3961 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3962 unsigned long nr = 0;
3965 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3968 if (!(BIT(lru) & lru_mask))
3971 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3973 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3978 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3979 unsigned int lru_mask,
3982 unsigned long nr = 0;
3986 if (!(BIT(lru) & lru_mask))
3989 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3991 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3996 static int memcg_numa_stat_show(struct seq_file *m, void *v)
4000 unsigned int lru_mask;
4003 static const struct numa_stat stats[] = {
4004 { "total", LRU_ALL },
4005 { "file", LRU_ALL_FILE },
4006 { "anon", LRU_ALL_ANON },
4007 { "unevictable", BIT(LRU_UNEVICTABLE) },
4009 const struct numa_stat *stat;
4011 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4013 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4014 seq_printf(m, "%s=%lu", stat->name,
4015 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4017 for_each_node_state(nid, N_MEMORY)
4018 seq_printf(m, " N%d=%lu", nid,
4019 mem_cgroup_node_nr_lru_pages(memcg, nid,
4020 stat->lru_mask, false));
4024 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4026 seq_printf(m, "hierarchical_%s=%lu", stat->name,
4027 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4029 for_each_node_state(nid, N_MEMORY)
4030 seq_printf(m, " N%d=%lu", nid,
4031 mem_cgroup_node_nr_lru_pages(memcg, nid,
4032 stat->lru_mask, true));
4038 #endif /* CONFIG_NUMA */
4040 static const unsigned int memcg1_stats[] = {
4043 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4053 static const char *const memcg1_stat_names[] = {
4056 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4066 /* Universal VM events cgroup1 shows, original sort order */
4067 static const unsigned int memcg1_events[] = {
4074 static int memcg_stat_show(struct seq_file *m, void *v)
4076 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4077 unsigned long memory, memsw;
4078 struct mem_cgroup *mi;
4081 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4083 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4086 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4088 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4089 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4092 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4093 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4094 memcg_events_local(memcg, memcg1_events[i]));
4096 for (i = 0; i < NR_LRU_LISTS; i++)
4097 seq_printf(m, "%s %lu\n", lru_list_name(i),
4098 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4101 /* Hierarchical information */
4102 memory = memsw = PAGE_COUNTER_MAX;
4103 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4104 memory = min(memory, READ_ONCE(mi->memory.max));
4105 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4107 seq_printf(m, "hierarchical_memory_limit %llu\n",
4108 (u64)memory * PAGE_SIZE);
4109 if (do_memsw_account())
4110 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4111 (u64)memsw * PAGE_SIZE);
4113 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4116 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4118 nr = memcg_page_state(memcg, memcg1_stats[i]);
4119 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4120 (u64)nr * PAGE_SIZE);
4123 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4124 seq_printf(m, "total_%s %llu\n",
4125 vm_event_name(memcg1_events[i]),
4126 (u64)memcg_events(memcg, memcg1_events[i]));
4128 for (i = 0; i < NR_LRU_LISTS; i++)
4129 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4130 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4133 #ifdef CONFIG_DEBUG_VM
4136 struct mem_cgroup_per_node *mz;
4137 unsigned long anon_cost = 0;
4138 unsigned long file_cost = 0;
4140 for_each_online_pgdat(pgdat) {
4141 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
4143 anon_cost += mz->lruvec.anon_cost;
4144 file_cost += mz->lruvec.file_cost;
4146 seq_printf(m, "anon_cost %lu\n", anon_cost);
4147 seq_printf(m, "file_cost %lu\n", file_cost);
4154 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4157 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4159 return mem_cgroup_swappiness(memcg);
4162 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4163 struct cftype *cft, u64 val)
4165 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4171 memcg->swappiness = val;
4173 vm_swappiness = val;
4178 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4180 struct mem_cgroup_threshold_ary *t;
4181 unsigned long usage;
4186 t = rcu_dereference(memcg->thresholds.primary);
4188 t = rcu_dereference(memcg->memsw_thresholds.primary);
4193 usage = mem_cgroup_usage(memcg, swap);
4196 * current_threshold points to threshold just below or equal to usage.
4197 * If it's not true, a threshold was crossed after last
4198 * call of __mem_cgroup_threshold().
4200 i = t->current_threshold;
4203 * Iterate backward over array of thresholds starting from
4204 * current_threshold and check if a threshold is crossed.
4205 * If none of thresholds below usage is crossed, we read
4206 * only one element of the array here.
4208 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4209 eventfd_signal(t->entries[i].eventfd, 1);
4211 /* i = current_threshold + 1 */
4215 * Iterate forward over array of thresholds starting from
4216 * current_threshold+1 and check if a threshold is crossed.
4217 * If none of thresholds above usage is crossed, we read
4218 * only one element of the array here.
4220 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4221 eventfd_signal(t->entries[i].eventfd, 1);
4223 /* Update current_threshold */
4224 t->current_threshold = i - 1;
4229 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4232 __mem_cgroup_threshold(memcg, false);
4233 if (do_memsw_account())
4234 __mem_cgroup_threshold(memcg, true);
4236 memcg = parent_mem_cgroup(memcg);
4240 static int compare_thresholds(const void *a, const void *b)
4242 const struct mem_cgroup_threshold *_a = a;
4243 const struct mem_cgroup_threshold *_b = b;
4245 if (_a->threshold > _b->threshold)
4248 if (_a->threshold < _b->threshold)
4254 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4256 struct mem_cgroup_eventfd_list *ev;
4258 spin_lock(&memcg_oom_lock);
4260 list_for_each_entry(ev, &memcg->oom_notify, list)
4261 eventfd_signal(ev->eventfd, 1);
4263 spin_unlock(&memcg_oom_lock);
4267 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4269 struct mem_cgroup *iter;
4271 for_each_mem_cgroup_tree(iter, memcg)
4272 mem_cgroup_oom_notify_cb(iter);
4275 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4276 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4278 struct mem_cgroup_thresholds *thresholds;
4279 struct mem_cgroup_threshold_ary *new;
4280 unsigned long threshold;
4281 unsigned long usage;
4284 ret = page_counter_memparse(args, "-1", &threshold);
4288 mutex_lock(&memcg->thresholds_lock);
4291 thresholds = &memcg->thresholds;
4292 usage = mem_cgroup_usage(memcg, false);
4293 } else if (type == _MEMSWAP) {
4294 thresholds = &memcg->memsw_thresholds;
4295 usage = mem_cgroup_usage(memcg, true);
4299 /* Check if a threshold crossed before adding a new one */
4300 if (thresholds->primary)
4301 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4303 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4305 /* Allocate memory for new array of thresholds */
4306 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4313 /* Copy thresholds (if any) to new array */
4314 if (thresholds->primary)
4315 memcpy(new->entries, thresholds->primary->entries,
4316 flex_array_size(new, entries, size - 1));
4318 /* Add new threshold */
4319 new->entries[size - 1].eventfd = eventfd;
4320 new->entries[size - 1].threshold = threshold;
4322 /* Sort thresholds. Registering of new threshold isn't time-critical */
4323 sort(new->entries, size, sizeof(*new->entries),
4324 compare_thresholds, NULL);
4326 /* Find current threshold */
4327 new->current_threshold = -1;
4328 for (i = 0; i < size; i++) {
4329 if (new->entries[i].threshold <= usage) {
4331 * new->current_threshold will not be used until
4332 * rcu_assign_pointer(), so it's safe to increment
4335 ++new->current_threshold;
4340 /* Free old spare buffer and save old primary buffer as spare */
4341 kfree(thresholds->spare);
4342 thresholds->spare = thresholds->primary;
4344 rcu_assign_pointer(thresholds->primary, new);
4346 /* To be sure that nobody uses thresholds */
4350 mutex_unlock(&memcg->thresholds_lock);
4355 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4356 struct eventfd_ctx *eventfd, const char *args)
4358 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4361 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4362 struct eventfd_ctx *eventfd, const char *args)
4364 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4367 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4368 struct eventfd_ctx *eventfd, enum res_type type)
4370 struct mem_cgroup_thresholds *thresholds;
4371 struct mem_cgroup_threshold_ary *new;
4372 unsigned long usage;
4373 int i, j, size, entries;
4375 mutex_lock(&memcg->thresholds_lock);
4378 thresholds = &memcg->thresholds;
4379 usage = mem_cgroup_usage(memcg, false);
4380 } else if (type == _MEMSWAP) {
4381 thresholds = &memcg->memsw_thresholds;
4382 usage = mem_cgroup_usage(memcg, true);
4386 if (!thresholds->primary)
4389 /* Check if a threshold crossed before removing */
4390 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4392 /* Calculate new number of threshold */
4394 for (i = 0; i < thresholds->primary->size; i++) {
4395 if (thresholds->primary->entries[i].eventfd != eventfd)
4401 new = thresholds->spare;
4403 /* If no items related to eventfd have been cleared, nothing to do */
4407 /* Set thresholds array to NULL if we don't have thresholds */
4416 /* Copy thresholds and find current threshold */
4417 new->current_threshold = -1;
4418 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4419 if (thresholds->primary->entries[i].eventfd == eventfd)
4422 new->entries[j] = thresholds->primary->entries[i];
4423 if (new->entries[j].threshold <= usage) {
4425 * new->current_threshold will not be used
4426 * until rcu_assign_pointer(), so it's safe to increment
4429 ++new->current_threshold;
4435 /* Swap primary and spare array */
4436 thresholds->spare = thresholds->primary;
4438 rcu_assign_pointer(thresholds->primary, new);
4440 /* To be sure that nobody uses thresholds */
4443 /* If all events are unregistered, free the spare array */
4445 kfree(thresholds->spare);
4446 thresholds->spare = NULL;
4449 mutex_unlock(&memcg->thresholds_lock);
4452 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4453 struct eventfd_ctx *eventfd)
4455 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4458 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4459 struct eventfd_ctx *eventfd)
4461 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4464 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4465 struct eventfd_ctx *eventfd, const char *args)
4467 struct mem_cgroup_eventfd_list *event;
4469 event = kmalloc(sizeof(*event), GFP_KERNEL);
4473 spin_lock(&memcg_oom_lock);
4475 event->eventfd = eventfd;
4476 list_add(&event->list, &memcg->oom_notify);
4478 /* already in OOM ? */
4479 if (memcg->under_oom)
4480 eventfd_signal(eventfd, 1);
4481 spin_unlock(&memcg_oom_lock);
4486 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4487 struct eventfd_ctx *eventfd)
4489 struct mem_cgroup_eventfd_list *ev, *tmp;
4491 spin_lock(&memcg_oom_lock);
4493 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4494 if (ev->eventfd == eventfd) {
4495 list_del(&ev->list);
4500 spin_unlock(&memcg_oom_lock);
4503 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4505 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4507 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4508 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4509 seq_printf(sf, "oom_kill %lu\n",
4510 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4514 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4515 struct cftype *cft, u64 val)
4517 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4519 /* cannot set to root cgroup and only 0 and 1 are allowed */
4520 if (!css->parent || !((val == 0) || (val == 1)))
4523 memcg->oom_kill_disable = val;
4525 memcg_oom_recover(memcg);
4530 #ifdef CONFIG_CGROUP_WRITEBACK
4532 #include <trace/events/writeback.h>
4534 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4536 return wb_domain_init(&memcg->cgwb_domain, gfp);
4539 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4541 wb_domain_exit(&memcg->cgwb_domain);
4544 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4546 wb_domain_size_changed(&memcg->cgwb_domain);
4549 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4551 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4553 if (!memcg->css.parent)
4556 return &memcg->cgwb_domain;
4560 * idx can be of type enum memcg_stat_item or node_stat_item.
4561 * Keep in sync with memcg_exact_page().
4563 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4565 long x = atomic_long_read(&memcg->vmstats[idx]);
4568 for_each_online_cpu(cpu)
4569 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4576 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4577 * @wb: bdi_writeback in question
4578 * @pfilepages: out parameter for number of file pages
4579 * @pheadroom: out parameter for number of allocatable pages according to memcg
4580 * @pdirty: out parameter for number of dirty pages
4581 * @pwriteback: out parameter for number of pages under writeback
4583 * Determine the numbers of file, headroom, dirty, and writeback pages in
4584 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4585 * is a bit more involved.
4587 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4588 * headroom is calculated as the lowest headroom of itself and the
4589 * ancestors. Note that this doesn't consider the actual amount of
4590 * available memory in the system. The caller should further cap
4591 * *@pheadroom accordingly.
4593 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4594 unsigned long *pheadroom, unsigned long *pdirty,
4595 unsigned long *pwriteback)
4597 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4598 struct mem_cgroup *parent;
4600 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4602 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4603 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4604 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4605 *pheadroom = PAGE_COUNTER_MAX;
4607 while ((parent = parent_mem_cgroup(memcg))) {
4608 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4609 READ_ONCE(memcg->memory.high));
4610 unsigned long used = page_counter_read(&memcg->memory);
4612 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4618 * Foreign dirty flushing
4620 * There's an inherent mismatch between memcg and writeback. The former
4621 * trackes ownership per-page while the latter per-inode. This was a
4622 * deliberate design decision because honoring per-page ownership in the
4623 * writeback path is complicated, may lead to higher CPU and IO overheads
4624 * and deemed unnecessary given that write-sharing an inode across
4625 * different cgroups isn't a common use-case.
4627 * Combined with inode majority-writer ownership switching, this works well
4628 * enough in most cases but there are some pathological cases. For
4629 * example, let's say there are two cgroups A and B which keep writing to
4630 * different but confined parts of the same inode. B owns the inode and
4631 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4632 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4633 * triggering background writeback. A will be slowed down without a way to
4634 * make writeback of the dirty pages happen.
4636 * Conditions like the above can lead to a cgroup getting repatedly and
4637 * severely throttled after making some progress after each
4638 * dirty_expire_interval while the underyling IO device is almost
4641 * Solving this problem completely requires matching the ownership tracking
4642 * granularities between memcg and writeback in either direction. However,
4643 * the more egregious behaviors can be avoided by simply remembering the
4644 * most recent foreign dirtying events and initiating remote flushes on
4645 * them when local writeback isn't enough to keep the memory clean enough.
4647 * The following two functions implement such mechanism. When a foreign
4648 * page - a page whose memcg and writeback ownerships don't match - is
4649 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4650 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4651 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4652 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4653 * foreign bdi_writebacks which haven't expired. Both the numbers of
4654 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4655 * limited to MEMCG_CGWB_FRN_CNT.
4657 * The mechanism only remembers IDs and doesn't hold any object references.
4658 * As being wrong occasionally doesn't matter, updates and accesses to the
4659 * records are lockless and racy.
4661 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4662 struct bdi_writeback *wb)
4664 struct mem_cgroup *memcg = page_memcg(page);
4665 struct memcg_cgwb_frn *frn;
4666 u64 now = get_jiffies_64();
4667 u64 oldest_at = now;
4671 trace_track_foreign_dirty(page, wb);
4674 * Pick the slot to use. If there is already a slot for @wb, keep
4675 * using it. If not replace the oldest one which isn't being
4678 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4679 frn = &memcg->cgwb_frn[i];
4680 if (frn->bdi_id == wb->bdi->id &&
4681 frn->memcg_id == wb->memcg_css->id)
4683 if (time_before64(frn->at, oldest_at) &&
4684 atomic_read(&frn->done.cnt) == 1) {
4686 oldest_at = frn->at;
4690 if (i < MEMCG_CGWB_FRN_CNT) {
4692 * Re-using an existing one. Update timestamp lazily to
4693 * avoid making the cacheline hot. We want them to be
4694 * reasonably up-to-date and significantly shorter than
4695 * dirty_expire_interval as that's what expires the record.
4696 * Use the shorter of 1s and dirty_expire_interval / 8.
4698 unsigned long update_intv =
4699 min_t(unsigned long, HZ,
4700 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4702 if (time_before64(frn->at, now - update_intv))
4704 } else if (oldest >= 0) {
4705 /* replace the oldest free one */
4706 frn = &memcg->cgwb_frn[oldest];
4707 frn->bdi_id = wb->bdi->id;
4708 frn->memcg_id = wb->memcg_css->id;
4713 /* issue foreign writeback flushes for recorded foreign dirtying events */
4714 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4716 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4717 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4718 u64 now = jiffies_64;
4721 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4722 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4725 * If the record is older than dirty_expire_interval,
4726 * writeback on it has already started. No need to kick it
4727 * off again. Also, don't start a new one if there's
4728 * already one in flight.
4730 if (time_after64(frn->at, now - intv) &&
4731 atomic_read(&frn->done.cnt) == 1) {
4733 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4734 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4735 WB_REASON_FOREIGN_FLUSH,
4741 #else /* CONFIG_CGROUP_WRITEBACK */
4743 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4748 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4752 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4756 #endif /* CONFIG_CGROUP_WRITEBACK */
4759 * DO NOT USE IN NEW FILES.
4761 * "cgroup.event_control" implementation.
4763 * This is way over-engineered. It tries to support fully configurable
4764 * events for each user. Such level of flexibility is completely
4765 * unnecessary especially in the light of the planned unified hierarchy.
4767 * Please deprecate this and replace with something simpler if at all
4772 * Unregister event and free resources.
4774 * Gets called from workqueue.
4776 static void memcg_event_remove(struct work_struct *work)
4778 struct mem_cgroup_event *event =
4779 container_of(work, struct mem_cgroup_event, remove);
4780 struct mem_cgroup *memcg = event->memcg;
4782 remove_wait_queue(event->wqh, &event->wait);
4784 event->unregister_event(memcg, event->eventfd);
4786 /* Notify userspace the event is going away. */
4787 eventfd_signal(event->eventfd, 1);
4789 eventfd_ctx_put(event->eventfd);
4791 css_put(&memcg->css);
4795 * Gets called on EPOLLHUP on eventfd when user closes it.
4797 * Called with wqh->lock held and interrupts disabled.
4799 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4800 int sync, void *key)
4802 struct mem_cgroup_event *event =
4803 container_of(wait, struct mem_cgroup_event, wait);
4804 struct mem_cgroup *memcg = event->memcg;
4805 __poll_t flags = key_to_poll(key);
4807 if (flags & EPOLLHUP) {
4809 * If the event has been detached at cgroup removal, we
4810 * can simply return knowing the other side will cleanup
4813 * We can't race against event freeing since the other
4814 * side will require wqh->lock via remove_wait_queue(),
4817 spin_lock(&memcg->event_list_lock);
4818 if (!list_empty(&event->list)) {
4819 list_del_init(&event->list);
4821 * We are in atomic context, but cgroup_event_remove()
4822 * may sleep, so we have to call it in workqueue.
4824 schedule_work(&event->remove);
4826 spin_unlock(&memcg->event_list_lock);
4832 static void memcg_event_ptable_queue_proc(struct file *file,
4833 wait_queue_head_t *wqh, poll_table *pt)
4835 struct mem_cgroup_event *event =
4836 container_of(pt, struct mem_cgroup_event, pt);
4839 add_wait_queue(wqh, &event->wait);
4843 * DO NOT USE IN NEW FILES.
4845 * Parse input and register new cgroup event handler.
4847 * Input must be in format '<event_fd> <control_fd> <args>'.
4848 * Interpretation of args is defined by control file implementation.
4850 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4851 char *buf, size_t nbytes, loff_t off)
4853 struct cgroup_subsys_state *css = of_css(of);
4854 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4855 struct mem_cgroup_event *event;
4856 struct cgroup_subsys_state *cfile_css;
4857 unsigned int efd, cfd;
4864 buf = strstrip(buf);
4866 efd = simple_strtoul(buf, &endp, 10);
4871 cfd = simple_strtoul(buf, &endp, 10);
4872 if ((*endp != ' ') && (*endp != '\0'))
4876 event = kzalloc(sizeof(*event), GFP_KERNEL);
4880 event->memcg = memcg;
4881 INIT_LIST_HEAD(&event->list);
4882 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4883 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4884 INIT_WORK(&event->remove, memcg_event_remove);
4892 event->eventfd = eventfd_ctx_fileget(efile.file);
4893 if (IS_ERR(event->eventfd)) {
4894 ret = PTR_ERR(event->eventfd);
4901 goto out_put_eventfd;
4904 /* the process need read permission on control file */
4905 /* AV: shouldn't we check that it's been opened for read instead? */
4906 ret = file_permission(cfile.file, MAY_READ);
4911 * Determine the event callbacks and set them in @event. This used
4912 * to be done via struct cftype but cgroup core no longer knows
4913 * about these events. The following is crude but the whole thing
4914 * is for compatibility anyway.
4916 * DO NOT ADD NEW FILES.
4918 name = cfile.file->f_path.dentry->d_name.name;
4920 if (!strcmp(name, "memory.usage_in_bytes")) {
4921 event->register_event = mem_cgroup_usage_register_event;
4922 event->unregister_event = mem_cgroup_usage_unregister_event;
4923 } else if (!strcmp(name, "memory.oom_control")) {
4924 event->register_event = mem_cgroup_oom_register_event;
4925 event->unregister_event = mem_cgroup_oom_unregister_event;
4926 } else if (!strcmp(name, "memory.pressure_level")) {
4927 event->register_event = vmpressure_register_event;
4928 event->unregister_event = vmpressure_unregister_event;
4929 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4930 event->register_event = memsw_cgroup_usage_register_event;
4931 event->unregister_event = memsw_cgroup_usage_unregister_event;
4938 * Verify @cfile should belong to @css. Also, remaining events are
4939 * automatically removed on cgroup destruction but the removal is
4940 * asynchronous, so take an extra ref on @css.
4942 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4943 &memory_cgrp_subsys);
4945 if (IS_ERR(cfile_css))
4947 if (cfile_css != css) {
4952 ret = event->register_event(memcg, event->eventfd, buf);
4956 vfs_poll(efile.file, &event->pt);
4958 spin_lock(&memcg->event_list_lock);
4959 list_add(&event->list, &memcg->event_list);
4960 spin_unlock(&memcg->event_list_lock);
4972 eventfd_ctx_put(event->eventfd);
4981 static struct cftype mem_cgroup_legacy_files[] = {
4983 .name = "usage_in_bytes",
4984 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4985 .read_u64 = mem_cgroup_read_u64,
4988 .name = "max_usage_in_bytes",
4989 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4990 .write = mem_cgroup_reset,
4991 .read_u64 = mem_cgroup_read_u64,
4994 .name = "limit_in_bytes",
4995 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4996 .write = mem_cgroup_write,
4997 .read_u64 = mem_cgroup_read_u64,
5000 .name = "soft_limit_in_bytes",
5001 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5002 .write = mem_cgroup_write,
5003 .read_u64 = mem_cgroup_read_u64,
5007 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5008 .write = mem_cgroup_reset,
5009 .read_u64 = mem_cgroup_read_u64,
5013 .seq_show = memcg_stat_show,
5016 .name = "force_empty",
5017 .write = mem_cgroup_force_empty_write,
5020 .name = "use_hierarchy",
5021 .write_u64 = mem_cgroup_hierarchy_write,
5022 .read_u64 = mem_cgroup_hierarchy_read,
5025 .name = "cgroup.event_control", /* XXX: for compat */
5026 .write = memcg_write_event_control,
5027 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
5030 .name = "swappiness",
5031 .read_u64 = mem_cgroup_swappiness_read,
5032 .write_u64 = mem_cgroup_swappiness_write,
5035 .name = "move_charge_at_immigrate",
5036 .read_u64 = mem_cgroup_move_charge_read,
5037 .write_u64 = mem_cgroup_move_charge_write,
5040 .name = "oom_control",
5041 .seq_show = mem_cgroup_oom_control_read,
5042 .write_u64 = mem_cgroup_oom_control_write,
5043 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5046 .name = "pressure_level",
5050 .name = "numa_stat",
5051 .seq_show = memcg_numa_stat_show,
5055 .name = "kmem.limit_in_bytes",
5056 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5057 .write = mem_cgroup_write,
5058 .read_u64 = mem_cgroup_read_u64,
5061 .name = "kmem.usage_in_bytes",
5062 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5063 .read_u64 = mem_cgroup_read_u64,
5066 .name = "kmem.failcnt",
5067 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5068 .write = mem_cgroup_reset,
5069 .read_u64 = mem_cgroup_read_u64,
5072 .name = "kmem.max_usage_in_bytes",
5073 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5074 .write = mem_cgroup_reset,
5075 .read_u64 = mem_cgroup_read_u64,
5077 #if defined(CONFIG_MEMCG_KMEM) && \
5078 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5080 .name = "kmem.slabinfo",
5081 .seq_show = memcg_slab_show,
5085 .name = "kmem.tcp.limit_in_bytes",
5086 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5087 .write = mem_cgroup_write,
5088 .read_u64 = mem_cgroup_read_u64,
5091 .name = "kmem.tcp.usage_in_bytes",
5092 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5093 .read_u64 = mem_cgroup_read_u64,
5096 .name = "kmem.tcp.failcnt",
5097 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5098 .write = mem_cgroup_reset,
5099 .read_u64 = mem_cgroup_read_u64,
5102 .name = "kmem.tcp.max_usage_in_bytes",
5103 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5104 .write = mem_cgroup_reset,
5105 .read_u64 = mem_cgroup_read_u64,
5107 { }, /* terminate */
5111 * Private memory cgroup IDR
5113 * Swap-out records and page cache shadow entries need to store memcg
5114 * references in constrained space, so we maintain an ID space that is
5115 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5116 * memory-controlled cgroups to 64k.
5118 * However, there usually are many references to the offline CSS after
5119 * the cgroup has been destroyed, such as page cache or reclaimable
5120 * slab objects, that don't need to hang on to the ID. We want to keep
5121 * those dead CSS from occupying IDs, or we might quickly exhaust the
5122 * relatively small ID space and prevent the creation of new cgroups
5123 * even when there are much fewer than 64k cgroups - possibly none.
5125 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5126 * be freed and recycled when it's no longer needed, which is usually
5127 * when the CSS is offlined.
5129 * The only exception to that are records of swapped out tmpfs/shmem
5130 * pages that need to be attributed to live ancestors on swapin. But
5131 * those references are manageable from userspace.
5134 static DEFINE_IDR(mem_cgroup_idr);
5136 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5138 if (memcg->id.id > 0) {
5139 idr_remove(&mem_cgroup_idr, memcg->id.id);
5144 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5147 refcount_add(n, &memcg->id.ref);
5150 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5152 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5153 mem_cgroup_id_remove(memcg);
5155 /* Memcg ID pins CSS */
5156 css_put(&memcg->css);
5160 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5162 mem_cgroup_id_put_many(memcg, 1);
5166 * mem_cgroup_from_id - look up a memcg from a memcg id
5167 * @id: the memcg id to look up
5169 * Caller must hold rcu_read_lock().
5171 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5173 WARN_ON_ONCE(!rcu_read_lock_held());
5174 return idr_find(&mem_cgroup_idr, id);
5177 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5179 struct mem_cgroup_per_node *pn;
5182 * This routine is called against possible nodes.
5183 * But it's BUG to call kmalloc() against offline node.
5185 * TODO: this routine can waste much memory for nodes which will
5186 * never be onlined. It's better to use memory hotplug callback
5189 if (!node_state(node, N_NORMAL_MEMORY))
5191 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5195 pn->lruvec_stat_local = alloc_percpu_gfp(struct lruvec_stat,
5196 GFP_KERNEL_ACCOUNT);
5197 if (!pn->lruvec_stat_local) {
5202 pn->lruvec_stat_cpu = alloc_percpu_gfp(struct batched_lruvec_stat,
5203 GFP_KERNEL_ACCOUNT);
5204 if (!pn->lruvec_stat_cpu) {
5205 free_percpu(pn->lruvec_stat_local);
5210 lruvec_init(&pn->lruvec);
5211 pn->usage_in_excess = 0;
5212 pn->on_tree = false;
5215 memcg->nodeinfo[node] = pn;
5219 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5221 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5226 free_percpu(pn->lruvec_stat_cpu);
5227 free_percpu(pn->lruvec_stat_local);
5231 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5236 free_mem_cgroup_per_node_info(memcg, node);
5237 free_percpu(memcg->vmstats_percpu);
5238 free_percpu(memcg->vmstats_local);
5242 static void mem_cgroup_free(struct mem_cgroup *memcg)
5244 memcg_wb_domain_exit(memcg);
5246 * Flush percpu vmstats and vmevents to guarantee the value correctness
5247 * on parent's and all ancestor levels.
5249 memcg_flush_percpu_vmstats(memcg);
5250 memcg_flush_percpu_vmevents(memcg);
5251 __mem_cgroup_free(memcg);
5254 static struct mem_cgroup *mem_cgroup_alloc(void)
5256 struct mem_cgroup *memcg;
5259 int __maybe_unused i;
5260 long error = -ENOMEM;
5262 size = sizeof(struct mem_cgroup);
5263 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5265 memcg = kzalloc(size, GFP_KERNEL);
5267 return ERR_PTR(error);
5269 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5270 1, MEM_CGROUP_ID_MAX,
5272 if (memcg->id.id < 0) {
5273 error = memcg->id.id;
5277 memcg->vmstats_local = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5278 GFP_KERNEL_ACCOUNT);
5279 if (!memcg->vmstats_local)
5282 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5283 GFP_KERNEL_ACCOUNT);
5284 if (!memcg->vmstats_percpu)
5288 if (alloc_mem_cgroup_per_node_info(memcg, node))
5291 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5294 INIT_WORK(&memcg->high_work, high_work_func);
5295 INIT_LIST_HEAD(&memcg->oom_notify);
5296 mutex_init(&memcg->thresholds_lock);
5297 spin_lock_init(&memcg->move_lock);
5298 vmpressure_init(&memcg->vmpressure);
5299 INIT_LIST_HEAD(&memcg->event_list);
5300 spin_lock_init(&memcg->event_list_lock);
5301 memcg->socket_pressure = jiffies;
5302 #ifdef CONFIG_MEMCG_KMEM
5303 memcg->kmemcg_id = -1;
5304 INIT_LIST_HEAD(&memcg->objcg_list);
5306 #ifdef CONFIG_CGROUP_WRITEBACK
5307 INIT_LIST_HEAD(&memcg->cgwb_list);
5308 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5309 memcg->cgwb_frn[i].done =
5310 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5312 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5313 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5314 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5315 memcg->deferred_split_queue.split_queue_len = 0;
5317 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5320 mem_cgroup_id_remove(memcg);
5321 __mem_cgroup_free(memcg);
5322 return ERR_PTR(error);
5325 static struct cgroup_subsys_state * __ref
5326 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5328 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5329 struct mem_cgroup *memcg, *old_memcg;
5330 long error = -ENOMEM;
5332 old_memcg = set_active_memcg(parent);
5333 memcg = mem_cgroup_alloc();
5334 set_active_memcg(old_memcg);
5336 return ERR_CAST(memcg);
5338 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5339 memcg->soft_limit = PAGE_COUNTER_MAX;
5340 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5342 memcg->swappiness = mem_cgroup_swappiness(parent);
5343 memcg->oom_kill_disable = parent->oom_kill_disable;
5345 page_counter_init(&memcg->memory, &parent->memory);
5346 page_counter_init(&memcg->swap, &parent->swap);
5347 page_counter_init(&memcg->kmem, &parent->kmem);
5348 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5350 page_counter_init(&memcg->memory, NULL);
5351 page_counter_init(&memcg->swap, NULL);
5352 page_counter_init(&memcg->kmem, NULL);
5353 page_counter_init(&memcg->tcpmem, NULL);
5355 root_mem_cgroup = memcg;
5359 /* The following stuff does not apply to the root */
5360 error = memcg_online_kmem(memcg);
5364 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5365 static_branch_inc(&memcg_sockets_enabled_key);
5369 mem_cgroup_id_remove(memcg);
5370 mem_cgroup_free(memcg);
5371 return ERR_PTR(error);
5374 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5376 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5379 * A memcg must be visible for memcg_expand_shrinker_maps()
5380 * by the time the maps are allocated. So, we allocate maps
5381 * here, when for_each_mem_cgroup() can't skip it.
5383 if (memcg_alloc_shrinker_maps(memcg)) {
5384 mem_cgroup_id_remove(memcg);
5388 /* Online state pins memcg ID, memcg ID pins CSS */
5389 refcount_set(&memcg->id.ref, 1);
5394 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5396 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5397 struct mem_cgroup_event *event, *tmp;
5400 * Unregister events and notify userspace.
5401 * Notify userspace about cgroup removing only after rmdir of cgroup
5402 * directory to avoid race between userspace and kernelspace.
5404 spin_lock(&memcg->event_list_lock);
5405 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5406 list_del_init(&event->list);
5407 schedule_work(&event->remove);
5409 spin_unlock(&memcg->event_list_lock);
5411 page_counter_set_min(&memcg->memory, 0);
5412 page_counter_set_low(&memcg->memory, 0);
5414 memcg_offline_kmem(memcg);
5415 wb_memcg_offline(memcg);
5417 drain_all_stock(memcg);
5419 mem_cgroup_id_put(memcg);
5422 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5424 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5426 invalidate_reclaim_iterators(memcg);
5429 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5431 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5432 int __maybe_unused i;
5434 #ifdef CONFIG_CGROUP_WRITEBACK
5435 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5436 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5438 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5439 static_branch_dec(&memcg_sockets_enabled_key);
5441 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5442 static_branch_dec(&memcg_sockets_enabled_key);
5444 vmpressure_cleanup(&memcg->vmpressure);
5445 cancel_work_sync(&memcg->high_work);
5446 mem_cgroup_remove_from_trees(memcg);
5447 memcg_free_shrinker_maps(memcg);
5448 memcg_free_kmem(memcg);
5449 mem_cgroup_free(memcg);
5453 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5454 * @css: the target css
5456 * Reset the states of the mem_cgroup associated with @css. This is
5457 * invoked when the userland requests disabling on the default hierarchy
5458 * but the memcg is pinned through dependency. The memcg should stop
5459 * applying policies and should revert to the vanilla state as it may be
5460 * made visible again.
5462 * The current implementation only resets the essential configurations.
5463 * This needs to be expanded to cover all the visible parts.
5465 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5467 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5469 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5470 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5471 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5472 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5473 page_counter_set_min(&memcg->memory, 0);
5474 page_counter_set_low(&memcg->memory, 0);
5475 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5476 memcg->soft_limit = PAGE_COUNTER_MAX;
5477 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5478 memcg_wb_domain_size_changed(memcg);
5482 /* Handlers for move charge at task migration. */
5483 static int mem_cgroup_do_precharge(unsigned long count)
5487 /* Try a single bulk charge without reclaim first, kswapd may wake */
5488 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5490 mc.precharge += count;
5494 /* Try charges one by one with reclaim, but do not retry */
5496 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5510 enum mc_target_type {
5517 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5518 unsigned long addr, pte_t ptent)
5520 struct page *page = vm_normal_page(vma, addr, ptent);
5522 if (!page || !page_mapped(page))
5524 if (PageAnon(page)) {
5525 if (!(mc.flags & MOVE_ANON))
5528 if (!(mc.flags & MOVE_FILE))
5531 if (!get_page_unless_zero(page))
5537 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5538 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5539 pte_t ptent, swp_entry_t *entry)
5541 struct page *page = NULL;
5542 swp_entry_t ent = pte_to_swp_entry(ptent);
5544 if (!(mc.flags & MOVE_ANON))
5548 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5549 * a device and because they are not accessible by CPU they are store
5550 * as special swap entry in the CPU page table.
5552 if (is_device_private_entry(ent)) {
5553 page = device_private_entry_to_page(ent);
5555 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5556 * a refcount of 1 when free (unlike normal page)
5558 if (!page_ref_add_unless(page, 1, 1))
5563 if (non_swap_entry(ent))
5567 * Because lookup_swap_cache() updates some statistics counter,
5568 * we call find_get_page() with swapper_space directly.
5570 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5571 entry->val = ent.val;
5576 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5577 pte_t ptent, swp_entry_t *entry)
5583 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5584 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5586 if (!vma->vm_file) /* anonymous vma */
5588 if (!(mc.flags & MOVE_FILE))
5591 /* page is moved even if it's not RSS of this task(page-faulted). */
5592 /* shmem/tmpfs may report page out on swap: account for that too. */
5593 return find_get_incore_page(vma->vm_file->f_mapping,
5594 linear_page_index(vma, addr));
5598 * mem_cgroup_move_account - move account of the page
5600 * @compound: charge the page as compound or small page
5601 * @from: mem_cgroup which the page is moved from.
5602 * @to: mem_cgroup which the page is moved to. @from != @to.
5604 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5606 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5609 static int mem_cgroup_move_account(struct page *page,
5611 struct mem_cgroup *from,
5612 struct mem_cgroup *to)
5614 struct lruvec *from_vec, *to_vec;
5615 struct pglist_data *pgdat;
5616 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5619 VM_BUG_ON(from == to);
5620 VM_BUG_ON_PAGE(PageLRU(page), page);
5621 VM_BUG_ON(compound && !PageTransHuge(page));
5624 * Prevent mem_cgroup_migrate() from looking at
5625 * page's memory cgroup of its source page while we change it.
5628 if (!trylock_page(page))
5632 if (page_memcg(page) != from)
5635 pgdat = page_pgdat(page);
5636 from_vec = mem_cgroup_lruvec(from, pgdat);
5637 to_vec = mem_cgroup_lruvec(to, pgdat);
5639 lock_page_memcg(page);
5641 if (PageAnon(page)) {
5642 if (page_mapped(page)) {
5643 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5644 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5645 if (PageTransHuge(page)) {
5646 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5648 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5653 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5654 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5656 if (PageSwapBacked(page)) {
5657 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5658 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5661 if (page_mapped(page)) {
5662 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5663 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5666 if (PageDirty(page)) {
5667 struct address_space *mapping = page_mapping(page);
5669 if (mapping_can_writeback(mapping)) {
5670 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5672 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5678 if (PageWriteback(page)) {
5679 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5680 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5684 * All state has been migrated, let's switch to the new memcg.
5686 * It is safe to change page's memcg here because the page
5687 * is referenced, charged, isolated, and locked: we can't race
5688 * with (un)charging, migration, LRU putback, or anything else
5689 * that would rely on a stable page's memory cgroup.
5691 * Note that lock_page_memcg is a memcg lock, not a page lock,
5692 * to save space. As soon as we switch page's memory cgroup to a
5693 * new memcg that isn't locked, the above state can change
5694 * concurrently again. Make sure we're truly done with it.
5699 css_put(&from->css);
5701 page->memcg_data = (unsigned long)to;
5703 __unlock_page_memcg(from);
5707 local_irq_disable();
5708 mem_cgroup_charge_statistics(to, page, nr_pages);
5709 memcg_check_events(to, page);
5710 mem_cgroup_charge_statistics(from, page, -nr_pages);
5711 memcg_check_events(from, page);
5720 * get_mctgt_type - get target type of moving charge
5721 * @vma: the vma the pte to be checked belongs
5722 * @addr: the address corresponding to the pte to be checked
5723 * @ptent: the pte to be checked
5724 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5727 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5728 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5729 * move charge. if @target is not NULL, the page is stored in target->page
5730 * with extra refcnt got(Callers should handle it).
5731 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5732 * target for charge migration. if @target is not NULL, the entry is stored
5734 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5735 * (so ZONE_DEVICE page and thus not on the lru).
5736 * For now we such page is charge like a regular page would be as for all
5737 * intent and purposes it is just special memory taking the place of a
5740 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5742 * Called with pte lock held.
5745 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5746 unsigned long addr, pte_t ptent, union mc_target *target)
5748 struct page *page = NULL;
5749 enum mc_target_type ret = MC_TARGET_NONE;
5750 swp_entry_t ent = { .val = 0 };
5752 if (pte_present(ptent))
5753 page = mc_handle_present_pte(vma, addr, ptent);
5754 else if (is_swap_pte(ptent))
5755 page = mc_handle_swap_pte(vma, ptent, &ent);
5756 else if (pte_none(ptent))
5757 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5759 if (!page && !ent.val)
5763 * Do only loose check w/o serialization.
5764 * mem_cgroup_move_account() checks the page is valid or
5765 * not under LRU exclusion.
5767 if (page_memcg(page) == mc.from) {
5768 ret = MC_TARGET_PAGE;
5769 if (is_device_private_page(page))
5770 ret = MC_TARGET_DEVICE;
5772 target->page = page;
5774 if (!ret || !target)
5778 * There is a swap entry and a page doesn't exist or isn't charged.
5779 * But we cannot move a tail-page in a THP.
5781 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5782 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5783 ret = MC_TARGET_SWAP;
5790 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5792 * We don't consider PMD mapped swapping or file mapped pages because THP does
5793 * not support them for now.
5794 * Caller should make sure that pmd_trans_huge(pmd) is true.
5796 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5797 unsigned long addr, pmd_t pmd, union mc_target *target)
5799 struct page *page = NULL;
5800 enum mc_target_type ret = MC_TARGET_NONE;
5802 if (unlikely(is_swap_pmd(pmd))) {
5803 VM_BUG_ON(thp_migration_supported() &&
5804 !is_pmd_migration_entry(pmd));
5807 page = pmd_page(pmd);
5808 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5809 if (!(mc.flags & MOVE_ANON))
5811 if (page_memcg(page) == mc.from) {
5812 ret = MC_TARGET_PAGE;
5815 target->page = page;
5821 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5822 unsigned long addr, pmd_t pmd, union mc_target *target)
5824 return MC_TARGET_NONE;
5828 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5829 unsigned long addr, unsigned long end,
5830 struct mm_walk *walk)
5832 struct vm_area_struct *vma = walk->vma;
5836 ptl = pmd_trans_huge_lock(pmd, vma);
5839 * Note their can not be MC_TARGET_DEVICE for now as we do not
5840 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5841 * this might change.
5843 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5844 mc.precharge += HPAGE_PMD_NR;
5849 if (pmd_trans_unstable(pmd))
5851 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5852 for (; addr != end; pte++, addr += PAGE_SIZE)
5853 if (get_mctgt_type(vma, addr, *pte, NULL))
5854 mc.precharge++; /* increment precharge temporarily */
5855 pte_unmap_unlock(pte - 1, ptl);
5861 static const struct mm_walk_ops precharge_walk_ops = {
5862 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5865 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5867 unsigned long precharge;
5870 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5871 mmap_read_unlock(mm);
5873 precharge = mc.precharge;
5879 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5881 unsigned long precharge = mem_cgroup_count_precharge(mm);
5883 VM_BUG_ON(mc.moving_task);
5884 mc.moving_task = current;
5885 return mem_cgroup_do_precharge(precharge);
5888 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5889 static void __mem_cgroup_clear_mc(void)
5891 struct mem_cgroup *from = mc.from;
5892 struct mem_cgroup *to = mc.to;
5894 /* we must uncharge all the leftover precharges from mc.to */
5896 cancel_charge(mc.to, mc.precharge);
5900 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5901 * we must uncharge here.
5903 if (mc.moved_charge) {
5904 cancel_charge(mc.from, mc.moved_charge);
5905 mc.moved_charge = 0;
5907 /* we must fixup refcnts and charges */
5908 if (mc.moved_swap) {
5909 /* uncharge swap account from the old cgroup */
5910 if (!mem_cgroup_is_root(mc.from))
5911 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5913 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5916 * we charged both to->memory and to->memsw, so we
5917 * should uncharge to->memory.
5919 if (!mem_cgroup_is_root(mc.to))
5920 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5924 memcg_oom_recover(from);
5925 memcg_oom_recover(to);
5926 wake_up_all(&mc.waitq);
5929 static void mem_cgroup_clear_mc(void)
5931 struct mm_struct *mm = mc.mm;
5934 * we must clear moving_task before waking up waiters at the end of
5937 mc.moving_task = NULL;
5938 __mem_cgroup_clear_mc();
5939 spin_lock(&mc.lock);
5943 spin_unlock(&mc.lock);
5948 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5950 struct cgroup_subsys_state *css;
5951 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5952 struct mem_cgroup *from;
5953 struct task_struct *leader, *p;
5954 struct mm_struct *mm;
5955 unsigned long move_flags;
5958 /* charge immigration isn't supported on the default hierarchy */
5959 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5963 * Multi-process migrations only happen on the default hierarchy
5964 * where charge immigration is not used. Perform charge
5965 * immigration if @tset contains a leader and whine if there are
5969 cgroup_taskset_for_each_leader(leader, css, tset) {
5972 memcg = mem_cgroup_from_css(css);
5978 * We are now commited to this value whatever it is. Changes in this
5979 * tunable will only affect upcoming migrations, not the current one.
5980 * So we need to save it, and keep it going.
5982 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5986 from = mem_cgroup_from_task(p);
5988 VM_BUG_ON(from == memcg);
5990 mm = get_task_mm(p);
5993 /* We move charges only when we move a owner of the mm */
5994 if (mm->owner == p) {
5997 VM_BUG_ON(mc.precharge);
5998 VM_BUG_ON(mc.moved_charge);
5999 VM_BUG_ON(mc.moved_swap);
6001 spin_lock(&mc.lock);
6005 mc.flags = move_flags;
6006 spin_unlock(&mc.lock);
6007 /* We set mc.moving_task later */
6009 ret = mem_cgroup_precharge_mc(mm);
6011 mem_cgroup_clear_mc();
6018 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6021 mem_cgroup_clear_mc();
6024 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6025 unsigned long addr, unsigned long end,
6026 struct mm_walk *walk)
6029 struct vm_area_struct *vma = walk->vma;
6032 enum mc_target_type target_type;
6033 union mc_target target;
6036 ptl = pmd_trans_huge_lock(pmd, vma);
6038 if (mc.precharge < HPAGE_PMD_NR) {
6042 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6043 if (target_type == MC_TARGET_PAGE) {
6045 if (!isolate_lru_page(page)) {
6046 if (!mem_cgroup_move_account(page, true,
6048 mc.precharge -= HPAGE_PMD_NR;
6049 mc.moved_charge += HPAGE_PMD_NR;
6051 putback_lru_page(page);
6054 } else if (target_type == MC_TARGET_DEVICE) {
6056 if (!mem_cgroup_move_account(page, true,
6058 mc.precharge -= HPAGE_PMD_NR;
6059 mc.moved_charge += HPAGE_PMD_NR;
6067 if (pmd_trans_unstable(pmd))
6070 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6071 for (; addr != end; addr += PAGE_SIZE) {
6072 pte_t ptent = *(pte++);
6073 bool device = false;
6079 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6080 case MC_TARGET_DEVICE:
6083 case MC_TARGET_PAGE:
6086 * We can have a part of the split pmd here. Moving it
6087 * can be done but it would be too convoluted so simply
6088 * ignore such a partial THP and keep it in original
6089 * memcg. There should be somebody mapping the head.
6091 if (PageTransCompound(page))
6093 if (!device && isolate_lru_page(page))
6095 if (!mem_cgroup_move_account(page, false,
6098 /* we uncharge from mc.from later. */
6102 putback_lru_page(page);
6103 put: /* get_mctgt_type() gets the page */
6106 case MC_TARGET_SWAP:
6108 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6110 mem_cgroup_id_get_many(mc.to, 1);
6111 /* we fixup other refcnts and charges later. */
6119 pte_unmap_unlock(pte - 1, ptl);
6124 * We have consumed all precharges we got in can_attach().
6125 * We try charge one by one, but don't do any additional
6126 * charges to mc.to if we have failed in charge once in attach()
6129 ret = mem_cgroup_do_precharge(1);
6137 static const struct mm_walk_ops charge_walk_ops = {
6138 .pmd_entry = mem_cgroup_move_charge_pte_range,
6141 static void mem_cgroup_move_charge(void)
6143 lru_add_drain_all();
6145 * Signal lock_page_memcg() to take the memcg's move_lock
6146 * while we're moving its pages to another memcg. Then wait
6147 * for already started RCU-only updates to finish.
6149 atomic_inc(&mc.from->moving_account);
6152 if (unlikely(!mmap_read_trylock(mc.mm))) {
6154 * Someone who are holding the mmap_lock might be waiting in
6155 * waitq. So we cancel all extra charges, wake up all waiters,
6156 * and retry. Because we cancel precharges, we might not be able
6157 * to move enough charges, but moving charge is a best-effort
6158 * feature anyway, so it wouldn't be a big problem.
6160 __mem_cgroup_clear_mc();
6165 * When we have consumed all precharges and failed in doing
6166 * additional charge, the page walk just aborts.
6168 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6171 mmap_read_unlock(mc.mm);
6172 atomic_dec(&mc.from->moving_account);
6175 static void mem_cgroup_move_task(void)
6178 mem_cgroup_move_charge();
6179 mem_cgroup_clear_mc();
6182 #else /* !CONFIG_MMU */
6183 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6187 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6190 static void mem_cgroup_move_task(void)
6195 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6197 if (value == PAGE_COUNTER_MAX)
6198 seq_puts(m, "max\n");
6200 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6205 static u64 memory_current_read(struct cgroup_subsys_state *css,
6208 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6210 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6213 static int memory_min_show(struct seq_file *m, void *v)
6215 return seq_puts_memcg_tunable(m,
6216 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6219 static ssize_t memory_min_write(struct kernfs_open_file *of,
6220 char *buf, size_t nbytes, loff_t off)
6222 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6226 buf = strstrip(buf);
6227 err = page_counter_memparse(buf, "max", &min);
6231 page_counter_set_min(&memcg->memory, min);
6236 static int memory_low_show(struct seq_file *m, void *v)
6238 return seq_puts_memcg_tunable(m,
6239 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6242 static ssize_t memory_low_write(struct kernfs_open_file *of,
6243 char *buf, size_t nbytes, loff_t off)
6245 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6249 buf = strstrip(buf);
6250 err = page_counter_memparse(buf, "max", &low);
6254 page_counter_set_low(&memcg->memory, low);
6259 static int memory_high_show(struct seq_file *m, void *v)
6261 return seq_puts_memcg_tunable(m,
6262 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6265 static ssize_t memory_high_write(struct kernfs_open_file *of,
6266 char *buf, size_t nbytes, loff_t off)
6268 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6269 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6270 bool drained = false;
6274 buf = strstrip(buf);
6275 err = page_counter_memparse(buf, "max", &high);
6279 page_counter_set_high(&memcg->memory, high);
6282 unsigned long nr_pages = page_counter_read(&memcg->memory);
6283 unsigned long reclaimed;
6285 if (nr_pages <= high)
6288 if (signal_pending(current))
6292 drain_all_stock(memcg);
6297 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6300 if (!reclaimed && !nr_retries--)
6304 memcg_wb_domain_size_changed(memcg);
6308 static int memory_max_show(struct seq_file *m, void *v)
6310 return seq_puts_memcg_tunable(m,
6311 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6314 static ssize_t memory_max_write(struct kernfs_open_file *of,
6315 char *buf, size_t nbytes, loff_t off)
6317 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6318 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6319 bool drained = false;
6323 buf = strstrip(buf);
6324 err = page_counter_memparse(buf, "max", &max);
6328 xchg(&memcg->memory.max, max);
6331 unsigned long nr_pages = page_counter_read(&memcg->memory);
6333 if (nr_pages <= max)
6336 if (signal_pending(current))
6340 drain_all_stock(memcg);
6346 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6352 memcg_memory_event(memcg, MEMCG_OOM);
6353 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6357 memcg_wb_domain_size_changed(memcg);
6361 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6363 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6364 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6365 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6366 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6367 seq_printf(m, "oom_kill %lu\n",
6368 atomic_long_read(&events[MEMCG_OOM_KILL]));
6371 static int memory_events_show(struct seq_file *m, void *v)
6373 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6375 __memory_events_show(m, memcg->memory_events);
6379 static int memory_events_local_show(struct seq_file *m, void *v)
6381 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6383 __memory_events_show(m, memcg->memory_events_local);
6387 static int memory_stat_show(struct seq_file *m, void *v)
6389 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6392 buf = memory_stat_format(memcg);
6401 static int memory_numa_stat_show(struct seq_file *m, void *v)
6404 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6406 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6409 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6412 seq_printf(m, "%s", memory_stats[i].name);
6413 for_each_node_state(nid, N_MEMORY) {
6415 struct lruvec *lruvec;
6417 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6418 size = lruvec_page_state(lruvec, memory_stats[i].idx);
6419 size *= memory_stats[i].ratio;
6420 seq_printf(m, " N%d=%llu", nid, size);
6429 static int memory_oom_group_show(struct seq_file *m, void *v)
6431 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6433 seq_printf(m, "%d\n", memcg->oom_group);
6438 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6439 char *buf, size_t nbytes, loff_t off)
6441 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6444 buf = strstrip(buf);
6448 ret = kstrtoint(buf, 0, &oom_group);
6452 if (oom_group != 0 && oom_group != 1)
6455 memcg->oom_group = oom_group;
6460 static struct cftype memory_files[] = {
6463 .flags = CFTYPE_NOT_ON_ROOT,
6464 .read_u64 = memory_current_read,
6468 .flags = CFTYPE_NOT_ON_ROOT,
6469 .seq_show = memory_min_show,
6470 .write = memory_min_write,
6474 .flags = CFTYPE_NOT_ON_ROOT,
6475 .seq_show = memory_low_show,
6476 .write = memory_low_write,
6480 .flags = CFTYPE_NOT_ON_ROOT,
6481 .seq_show = memory_high_show,
6482 .write = memory_high_write,
6486 .flags = CFTYPE_NOT_ON_ROOT,
6487 .seq_show = memory_max_show,
6488 .write = memory_max_write,
6492 .flags = CFTYPE_NOT_ON_ROOT,
6493 .file_offset = offsetof(struct mem_cgroup, events_file),
6494 .seq_show = memory_events_show,
6497 .name = "events.local",
6498 .flags = CFTYPE_NOT_ON_ROOT,
6499 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6500 .seq_show = memory_events_local_show,
6504 .seq_show = memory_stat_show,
6508 .name = "numa_stat",
6509 .seq_show = memory_numa_stat_show,
6513 .name = "oom.group",
6514 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6515 .seq_show = memory_oom_group_show,
6516 .write = memory_oom_group_write,
6521 struct cgroup_subsys memory_cgrp_subsys = {
6522 .css_alloc = mem_cgroup_css_alloc,
6523 .css_online = mem_cgroup_css_online,
6524 .css_offline = mem_cgroup_css_offline,
6525 .css_released = mem_cgroup_css_released,
6526 .css_free = mem_cgroup_css_free,
6527 .css_reset = mem_cgroup_css_reset,
6528 .can_attach = mem_cgroup_can_attach,
6529 .cancel_attach = mem_cgroup_cancel_attach,
6530 .post_attach = mem_cgroup_move_task,
6531 .dfl_cftypes = memory_files,
6532 .legacy_cftypes = mem_cgroup_legacy_files,
6537 * This function calculates an individual cgroup's effective
6538 * protection which is derived from its own memory.min/low, its
6539 * parent's and siblings' settings, as well as the actual memory
6540 * distribution in the tree.
6542 * The following rules apply to the effective protection values:
6544 * 1. At the first level of reclaim, effective protection is equal to
6545 * the declared protection in memory.min and memory.low.
6547 * 2. To enable safe delegation of the protection configuration, at
6548 * subsequent levels the effective protection is capped to the
6549 * parent's effective protection.
6551 * 3. To make complex and dynamic subtrees easier to configure, the
6552 * user is allowed to overcommit the declared protection at a given
6553 * level. If that is the case, the parent's effective protection is
6554 * distributed to the children in proportion to how much protection
6555 * they have declared and how much of it they are utilizing.
6557 * This makes distribution proportional, but also work-conserving:
6558 * if one cgroup claims much more protection than it uses memory,
6559 * the unused remainder is available to its siblings.
6561 * 4. Conversely, when the declared protection is undercommitted at a
6562 * given level, the distribution of the larger parental protection
6563 * budget is NOT proportional. A cgroup's protection from a sibling
6564 * is capped to its own memory.min/low setting.
6566 * 5. However, to allow protecting recursive subtrees from each other
6567 * without having to declare each individual cgroup's fixed share
6568 * of the ancestor's claim to protection, any unutilized -
6569 * "floating" - protection from up the tree is distributed in
6570 * proportion to each cgroup's *usage*. This makes the protection
6571 * neutral wrt sibling cgroups and lets them compete freely over
6572 * the shared parental protection budget, but it protects the
6573 * subtree as a whole from neighboring subtrees.
6575 * Note that 4. and 5. are not in conflict: 4. is about protecting
6576 * against immediate siblings whereas 5. is about protecting against
6577 * neighboring subtrees.
6579 static unsigned long effective_protection(unsigned long usage,
6580 unsigned long parent_usage,
6581 unsigned long setting,
6582 unsigned long parent_effective,
6583 unsigned long siblings_protected)
6585 unsigned long protected;
6588 protected = min(usage, setting);
6590 * If all cgroups at this level combined claim and use more
6591 * protection then what the parent affords them, distribute
6592 * shares in proportion to utilization.
6594 * We are using actual utilization rather than the statically
6595 * claimed protection in order to be work-conserving: claimed
6596 * but unused protection is available to siblings that would
6597 * otherwise get a smaller chunk than what they claimed.
6599 if (siblings_protected > parent_effective)
6600 return protected * parent_effective / siblings_protected;
6603 * Ok, utilized protection of all children is within what the
6604 * parent affords them, so we know whatever this child claims
6605 * and utilizes is effectively protected.
6607 * If there is unprotected usage beyond this value, reclaim
6608 * will apply pressure in proportion to that amount.
6610 * If there is unutilized protection, the cgroup will be fully
6611 * shielded from reclaim, but we do return a smaller value for
6612 * protection than what the group could enjoy in theory. This
6613 * is okay. With the overcommit distribution above, effective
6614 * protection is always dependent on how memory is actually
6615 * consumed among the siblings anyway.
6620 * If the children aren't claiming (all of) the protection
6621 * afforded to them by the parent, distribute the remainder in
6622 * proportion to the (unprotected) memory of each cgroup. That
6623 * way, cgroups that aren't explicitly prioritized wrt each
6624 * other compete freely over the allowance, but they are
6625 * collectively protected from neighboring trees.
6627 * We're using unprotected memory for the weight so that if
6628 * some cgroups DO claim explicit protection, we don't protect
6629 * the same bytes twice.
6631 * Check both usage and parent_usage against the respective
6632 * protected values. One should imply the other, but they
6633 * aren't read atomically - make sure the division is sane.
6635 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6637 if (parent_effective > siblings_protected &&
6638 parent_usage > siblings_protected &&
6639 usage > protected) {
6640 unsigned long unclaimed;
6642 unclaimed = parent_effective - siblings_protected;
6643 unclaimed *= usage - protected;
6644 unclaimed /= parent_usage - siblings_protected;
6653 * mem_cgroup_protected - check if memory consumption is in the normal range
6654 * @root: the top ancestor of the sub-tree being checked
6655 * @memcg: the memory cgroup to check
6657 * WARNING: This function is not stateless! It can only be used as part
6658 * of a top-down tree iteration, not for isolated queries.
6660 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6661 struct mem_cgroup *memcg)
6663 unsigned long usage, parent_usage;
6664 struct mem_cgroup *parent;
6666 if (mem_cgroup_disabled())
6670 root = root_mem_cgroup;
6673 * Effective values of the reclaim targets are ignored so they
6674 * can be stale. Have a look at mem_cgroup_protection for more
6676 * TODO: calculation should be more robust so that we do not need
6677 * that special casing.
6682 usage = page_counter_read(&memcg->memory);
6686 parent = parent_mem_cgroup(memcg);
6687 /* No parent means a non-hierarchical mode on v1 memcg */
6691 if (parent == root) {
6692 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6693 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6697 parent_usage = page_counter_read(&parent->memory);
6699 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6700 READ_ONCE(memcg->memory.min),
6701 READ_ONCE(parent->memory.emin),
6702 atomic_long_read(&parent->memory.children_min_usage)));
6704 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6705 READ_ONCE(memcg->memory.low),
6706 READ_ONCE(parent->memory.elow),
6707 atomic_long_read(&parent->memory.children_low_usage)));
6711 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6712 * @page: page to charge
6713 * @mm: mm context of the victim
6714 * @gfp_mask: reclaim mode
6716 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6717 * pages according to @gfp_mask if necessary.
6719 * Returns 0 on success. Otherwise, an error code is returned.
6721 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6723 unsigned int nr_pages = thp_nr_pages(page);
6724 struct mem_cgroup *memcg = NULL;
6727 if (mem_cgroup_disabled())
6730 if (PageSwapCache(page)) {
6731 swp_entry_t ent = { .val = page_private(page), };
6735 * Every swap fault against a single page tries to charge the
6736 * page, bail as early as possible. shmem_unuse() encounters
6737 * already charged pages, too. page and memcg binding is
6738 * protected by the page lock, which serializes swap cache
6739 * removal, which in turn serializes uncharging.
6741 VM_BUG_ON_PAGE(!PageLocked(page), page);
6742 if (page_memcg(compound_head(page)))
6745 id = lookup_swap_cgroup_id(ent);
6747 memcg = mem_cgroup_from_id(id);
6748 if (memcg && !css_tryget_online(&memcg->css))
6754 memcg = get_mem_cgroup_from_mm(mm);
6756 ret = try_charge(memcg, gfp_mask, nr_pages);
6760 css_get(&memcg->css);
6761 commit_charge(page, memcg);
6763 local_irq_disable();
6764 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6765 memcg_check_events(memcg, page);
6768 if (PageSwapCache(page)) {
6769 swp_entry_t entry = { .val = page_private(page) };
6771 * The swap entry might not get freed for a long time,
6772 * let's not wait for it. The page already received a
6773 * memory+swap charge, drop the swap entry duplicate.
6775 mem_cgroup_uncharge_swap(entry, nr_pages);
6779 css_put(&memcg->css);
6784 struct uncharge_gather {
6785 struct mem_cgroup *memcg;
6786 unsigned long nr_pages;
6787 unsigned long pgpgout;
6788 unsigned long nr_kmem;
6789 struct page *dummy_page;
6792 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6794 memset(ug, 0, sizeof(*ug));
6797 static void uncharge_batch(const struct uncharge_gather *ug)
6799 unsigned long flags;
6801 if (!mem_cgroup_is_root(ug->memcg)) {
6802 page_counter_uncharge(&ug->memcg->memory, ug->nr_pages);
6803 if (do_memsw_account())
6804 page_counter_uncharge(&ug->memcg->memsw, ug->nr_pages);
6805 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6806 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6807 memcg_oom_recover(ug->memcg);
6810 local_irq_save(flags);
6811 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6812 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_pages);
6813 memcg_check_events(ug->memcg, ug->dummy_page);
6814 local_irq_restore(flags);
6816 /* drop reference from uncharge_page */
6817 css_put(&ug->memcg->css);
6820 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6822 unsigned long nr_pages;
6824 VM_BUG_ON_PAGE(PageLRU(page), page);
6826 if (!page_memcg(page))
6830 * Nobody should be changing or seriously looking at
6831 * page_memcg(page) at this point, we have fully
6832 * exclusive access to the page.
6835 if (ug->memcg != page_memcg(page)) {
6838 uncharge_gather_clear(ug);
6840 ug->memcg = page_memcg(page);
6842 /* pairs with css_put in uncharge_batch */
6843 css_get(&ug->memcg->css);
6846 nr_pages = compound_nr(page);
6847 ug->nr_pages += nr_pages;
6849 if (PageMemcgKmem(page))
6850 ug->nr_kmem += nr_pages;
6854 ug->dummy_page = page;
6855 page->memcg_data = 0;
6856 css_put(&ug->memcg->css);
6859 static void uncharge_list(struct list_head *page_list)
6861 struct uncharge_gather ug;
6862 struct list_head *next;
6864 uncharge_gather_clear(&ug);
6867 * Note that the list can be a single page->lru; hence the
6868 * do-while loop instead of a simple list_for_each_entry().
6870 next = page_list->next;
6874 page = list_entry(next, struct page, lru);
6875 next = page->lru.next;
6877 uncharge_page(page, &ug);
6878 } while (next != page_list);
6881 uncharge_batch(&ug);
6885 * mem_cgroup_uncharge - uncharge a page
6886 * @page: page to uncharge
6888 * Uncharge a page previously charged with mem_cgroup_charge().
6890 void mem_cgroup_uncharge(struct page *page)
6892 struct uncharge_gather ug;
6894 if (mem_cgroup_disabled())
6897 /* Don't touch page->lru of any random page, pre-check: */
6898 if (!page_memcg(page))
6901 uncharge_gather_clear(&ug);
6902 uncharge_page(page, &ug);
6903 uncharge_batch(&ug);
6907 * mem_cgroup_uncharge_list - uncharge a list of page
6908 * @page_list: list of pages to uncharge
6910 * Uncharge a list of pages previously charged with
6911 * mem_cgroup_charge().
6913 void mem_cgroup_uncharge_list(struct list_head *page_list)
6915 if (mem_cgroup_disabled())
6918 if (!list_empty(page_list))
6919 uncharge_list(page_list);
6923 * mem_cgroup_migrate - charge a page's replacement
6924 * @oldpage: currently circulating page
6925 * @newpage: replacement page
6927 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6928 * be uncharged upon free.
6930 * Both pages must be locked, @newpage->mapping must be set up.
6932 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6934 struct mem_cgroup *memcg;
6935 unsigned int nr_pages;
6936 unsigned long flags;
6938 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6939 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6940 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6941 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6944 if (mem_cgroup_disabled())
6947 /* Page cache replacement: new page already charged? */
6948 if (page_memcg(newpage))
6951 memcg = page_memcg(oldpage);
6952 VM_WARN_ON_ONCE_PAGE(!memcg, oldpage);
6956 /* Force-charge the new page. The old one will be freed soon */
6957 nr_pages = thp_nr_pages(newpage);
6959 page_counter_charge(&memcg->memory, nr_pages);
6960 if (do_memsw_account())
6961 page_counter_charge(&memcg->memsw, nr_pages);
6963 css_get(&memcg->css);
6964 commit_charge(newpage, memcg);
6966 local_irq_save(flags);
6967 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
6968 memcg_check_events(memcg, newpage);
6969 local_irq_restore(flags);
6972 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6973 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6975 void mem_cgroup_sk_alloc(struct sock *sk)
6977 struct mem_cgroup *memcg;
6979 if (!mem_cgroup_sockets_enabled)
6982 /* Do not associate the sock with unrelated interrupted task's memcg. */
6987 memcg = mem_cgroup_from_task(current);
6988 if (memcg == root_mem_cgroup)
6990 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6992 if (css_tryget(&memcg->css))
6993 sk->sk_memcg = memcg;
6998 void mem_cgroup_sk_free(struct sock *sk)
7001 css_put(&sk->sk_memcg->css);
7005 * mem_cgroup_charge_skmem - charge socket memory
7006 * @memcg: memcg to charge
7007 * @nr_pages: number of pages to charge
7009 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7010 * @memcg's configured limit, %false if the charge had to be forced.
7012 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7014 gfp_t gfp_mask = GFP_KERNEL;
7016 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7017 struct page_counter *fail;
7019 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7020 memcg->tcpmem_pressure = 0;
7023 page_counter_charge(&memcg->tcpmem, nr_pages);
7024 memcg->tcpmem_pressure = 1;
7028 /* Don't block in the packet receive path */
7030 gfp_mask = GFP_NOWAIT;
7032 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7034 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
7037 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
7042 * mem_cgroup_uncharge_skmem - uncharge socket memory
7043 * @memcg: memcg to uncharge
7044 * @nr_pages: number of pages to uncharge
7046 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7048 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7049 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7053 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7055 refill_stock(memcg, nr_pages);
7058 static int __init cgroup_memory(char *s)
7062 while ((token = strsep(&s, ",")) != NULL) {
7065 if (!strcmp(token, "nosocket"))
7066 cgroup_memory_nosocket = true;
7067 if (!strcmp(token, "nokmem"))
7068 cgroup_memory_nokmem = true;
7072 __setup("cgroup.memory=", cgroup_memory);
7075 * subsys_initcall() for memory controller.
7077 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7078 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7079 * basically everything that doesn't depend on a specific mem_cgroup structure
7080 * should be initialized from here.
7082 static int __init mem_cgroup_init(void)
7087 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7088 * used for per-memcg-per-cpu caching of per-node statistics. In order
7089 * to work fine, we should make sure that the overfill threshold can't
7090 * exceed S32_MAX / PAGE_SIZE.
7092 BUILD_BUG_ON(MEMCG_CHARGE_BATCH > S32_MAX / PAGE_SIZE);
7094 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7095 memcg_hotplug_cpu_dead);
7097 for_each_possible_cpu(cpu)
7098 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7101 for_each_node(node) {
7102 struct mem_cgroup_tree_per_node *rtpn;
7104 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7105 node_online(node) ? node : NUMA_NO_NODE);
7107 rtpn->rb_root = RB_ROOT;
7108 rtpn->rb_rightmost = NULL;
7109 spin_lock_init(&rtpn->lock);
7110 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7115 subsys_initcall(mem_cgroup_init);
7117 #ifdef CONFIG_MEMCG_SWAP
7118 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7120 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7122 * The root cgroup cannot be destroyed, so it's refcount must
7125 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7129 memcg = parent_mem_cgroup(memcg);
7131 memcg = root_mem_cgroup;
7137 * mem_cgroup_swapout - transfer a memsw charge to swap
7138 * @page: page whose memsw charge to transfer
7139 * @entry: swap entry to move the charge to
7141 * Transfer the memsw charge of @page to @entry.
7143 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7145 struct mem_cgroup *memcg, *swap_memcg;
7146 unsigned int nr_entries;
7147 unsigned short oldid;
7149 VM_BUG_ON_PAGE(PageLRU(page), page);
7150 VM_BUG_ON_PAGE(page_count(page), page);
7152 if (mem_cgroup_disabled())
7155 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7158 memcg = page_memcg(page);
7160 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7165 * In case the memcg owning these pages has been offlined and doesn't
7166 * have an ID allocated to it anymore, charge the closest online
7167 * ancestor for the swap instead and transfer the memory+swap charge.
7169 swap_memcg = mem_cgroup_id_get_online(memcg);
7170 nr_entries = thp_nr_pages(page);
7171 /* Get references for the tail pages, too */
7173 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7174 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7176 VM_BUG_ON_PAGE(oldid, page);
7177 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7179 page->memcg_data = 0;
7181 if (!mem_cgroup_is_root(memcg))
7182 page_counter_uncharge(&memcg->memory, nr_entries);
7184 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7185 if (!mem_cgroup_is_root(swap_memcg))
7186 page_counter_charge(&swap_memcg->memsw, nr_entries);
7187 page_counter_uncharge(&memcg->memsw, nr_entries);
7191 * Interrupts should be disabled here because the caller holds the
7192 * i_pages lock which is taken with interrupts-off. It is
7193 * important here to have the interrupts disabled because it is the
7194 * only synchronisation we have for updating the per-CPU variables.
7196 VM_BUG_ON(!irqs_disabled());
7197 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7198 memcg_check_events(memcg, page);
7200 css_put(&memcg->css);
7204 * mem_cgroup_try_charge_swap - try charging swap space for a page
7205 * @page: page being added to swap
7206 * @entry: swap entry to charge
7208 * Try to charge @page's memcg for the swap space at @entry.
7210 * Returns 0 on success, -ENOMEM on failure.
7212 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7214 unsigned int nr_pages = thp_nr_pages(page);
7215 struct page_counter *counter;
7216 struct mem_cgroup *memcg;
7217 unsigned short oldid;
7219 if (mem_cgroup_disabled())
7222 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7225 memcg = page_memcg(page);
7227 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7232 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7236 memcg = mem_cgroup_id_get_online(memcg);
7238 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7239 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7240 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7241 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7242 mem_cgroup_id_put(memcg);
7246 /* Get references for the tail pages, too */
7248 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7249 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7250 VM_BUG_ON_PAGE(oldid, page);
7251 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7257 * mem_cgroup_uncharge_swap - uncharge swap space
7258 * @entry: swap entry to uncharge
7259 * @nr_pages: the amount of swap space to uncharge
7261 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7263 struct mem_cgroup *memcg;
7266 id = swap_cgroup_record(entry, 0, nr_pages);
7268 memcg = mem_cgroup_from_id(id);
7270 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7271 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7272 page_counter_uncharge(&memcg->swap, nr_pages);
7274 page_counter_uncharge(&memcg->memsw, nr_pages);
7276 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7277 mem_cgroup_id_put_many(memcg, nr_pages);
7282 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7284 long nr_swap_pages = get_nr_swap_pages();
7286 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7287 return nr_swap_pages;
7288 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7289 nr_swap_pages = min_t(long, nr_swap_pages,
7290 READ_ONCE(memcg->swap.max) -
7291 page_counter_read(&memcg->swap));
7292 return nr_swap_pages;
7295 bool mem_cgroup_swap_full(struct page *page)
7297 struct mem_cgroup *memcg;
7299 VM_BUG_ON_PAGE(!PageLocked(page), page);
7303 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7306 memcg = page_memcg(page);
7310 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7311 unsigned long usage = page_counter_read(&memcg->swap);
7313 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7314 usage * 2 >= READ_ONCE(memcg->swap.max))
7321 static int __init setup_swap_account(char *s)
7323 if (!strcmp(s, "1"))
7324 cgroup_memory_noswap = false;
7325 else if (!strcmp(s, "0"))
7326 cgroup_memory_noswap = true;
7329 __setup("swapaccount=", setup_swap_account);
7331 static u64 swap_current_read(struct cgroup_subsys_state *css,
7334 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7336 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7339 static int swap_high_show(struct seq_file *m, void *v)
7341 return seq_puts_memcg_tunable(m,
7342 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7345 static ssize_t swap_high_write(struct kernfs_open_file *of,
7346 char *buf, size_t nbytes, loff_t off)
7348 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7352 buf = strstrip(buf);
7353 err = page_counter_memparse(buf, "max", &high);
7357 page_counter_set_high(&memcg->swap, high);
7362 static int swap_max_show(struct seq_file *m, void *v)
7364 return seq_puts_memcg_tunable(m,
7365 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7368 static ssize_t swap_max_write(struct kernfs_open_file *of,
7369 char *buf, size_t nbytes, loff_t off)
7371 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7375 buf = strstrip(buf);
7376 err = page_counter_memparse(buf, "max", &max);
7380 xchg(&memcg->swap.max, max);
7385 static int swap_events_show(struct seq_file *m, void *v)
7387 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7389 seq_printf(m, "high %lu\n",
7390 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7391 seq_printf(m, "max %lu\n",
7392 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7393 seq_printf(m, "fail %lu\n",
7394 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7399 static struct cftype swap_files[] = {
7401 .name = "swap.current",
7402 .flags = CFTYPE_NOT_ON_ROOT,
7403 .read_u64 = swap_current_read,
7406 .name = "swap.high",
7407 .flags = CFTYPE_NOT_ON_ROOT,
7408 .seq_show = swap_high_show,
7409 .write = swap_high_write,
7413 .flags = CFTYPE_NOT_ON_ROOT,
7414 .seq_show = swap_max_show,
7415 .write = swap_max_write,
7418 .name = "swap.events",
7419 .flags = CFTYPE_NOT_ON_ROOT,
7420 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7421 .seq_show = swap_events_show,
7426 static struct cftype memsw_files[] = {
7428 .name = "memsw.usage_in_bytes",
7429 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7430 .read_u64 = mem_cgroup_read_u64,
7433 .name = "memsw.max_usage_in_bytes",
7434 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7435 .write = mem_cgroup_reset,
7436 .read_u64 = mem_cgroup_read_u64,
7439 .name = "memsw.limit_in_bytes",
7440 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7441 .write = mem_cgroup_write,
7442 .read_u64 = mem_cgroup_read_u64,
7445 .name = "memsw.failcnt",
7446 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7447 .write = mem_cgroup_reset,
7448 .read_u64 = mem_cgroup_read_u64,
7450 { }, /* terminate */
7454 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7455 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7456 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7457 * boot parameter. This may result in premature OOPS inside
7458 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7460 static int __init mem_cgroup_swap_init(void)
7462 /* No memory control -> no swap control */
7463 if (mem_cgroup_disabled())
7464 cgroup_memory_noswap = true;
7466 if (cgroup_memory_noswap)
7469 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7470 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7474 core_initcall(mem_cgroup_swap_init);
7476 #endif /* CONFIG_MEMCG_SWAP */