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 * These functions are safe to use under any of the following conditions:
1352 * - lock_page_memcg()
1353 * - page->_refcount is zero
1355 struct lruvec *lock_page_lruvec(struct page *page)
1357 struct lruvec *lruvec;
1358 struct pglist_data *pgdat = page_pgdat(page);
1361 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1362 spin_lock(&lruvec->lru_lock);
1365 lruvec_memcg_debug(lruvec, page);
1370 struct lruvec *lock_page_lruvec_irq(struct page *page)
1372 struct lruvec *lruvec;
1373 struct pglist_data *pgdat = page_pgdat(page);
1376 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1377 spin_lock_irq(&lruvec->lru_lock);
1380 lruvec_memcg_debug(lruvec, page);
1385 struct lruvec *lock_page_lruvec_irqsave(struct page *page, unsigned long *flags)
1387 struct lruvec *lruvec;
1388 struct pglist_data *pgdat = page_pgdat(page);
1391 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1392 spin_lock_irqsave(&lruvec->lru_lock, *flags);
1395 lruvec_memcg_debug(lruvec, page);
1401 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1402 * @lruvec: mem_cgroup per zone lru vector
1403 * @lru: index of lru list the page is sitting on
1404 * @zid: zone id of the accounted pages
1405 * @nr_pages: positive when adding or negative when removing
1407 * This function must be called under lru_lock, just before a page is added
1408 * to or just after a page is removed from an lru list (that ordering being
1409 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1411 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1412 int zid, int nr_pages)
1414 struct mem_cgroup_per_node *mz;
1415 unsigned long *lru_size;
1418 if (mem_cgroup_disabled())
1421 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1422 lru_size = &mz->lru_zone_size[zid][lru];
1425 *lru_size += nr_pages;
1428 if (WARN_ONCE(size < 0,
1429 "%s(%p, %d, %d): lru_size %ld\n",
1430 __func__, lruvec, lru, nr_pages, size)) {
1436 *lru_size += nr_pages;
1440 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1441 * @memcg: the memory cgroup
1443 * Returns the maximum amount of memory @mem can be charged with, in
1446 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1448 unsigned long margin = 0;
1449 unsigned long count;
1450 unsigned long limit;
1452 count = page_counter_read(&memcg->memory);
1453 limit = READ_ONCE(memcg->memory.max);
1455 margin = limit - count;
1457 if (do_memsw_account()) {
1458 count = page_counter_read(&memcg->memsw);
1459 limit = READ_ONCE(memcg->memsw.max);
1461 margin = min(margin, limit - count);
1470 * A routine for checking "mem" is under move_account() or not.
1472 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1473 * moving cgroups. This is for waiting at high-memory pressure
1476 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1478 struct mem_cgroup *from;
1479 struct mem_cgroup *to;
1482 * Unlike task_move routines, we access mc.to, mc.from not under
1483 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1485 spin_lock(&mc.lock);
1491 ret = mem_cgroup_is_descendant(from, memcg) ||
1492 mem_cgroup_is_descendant(to, memcg);
1494 spin_unlock(&mc.lock);
1498 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1500 if (mc.moving_task && current != mc.moving_task) {
1501 if (mem_cgroup_under_move(memcg)) {
1503 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1504 /* moving charge context might have finished. */
1507 finish_wait(&mc.waitq, &wait);
1514 struct memory_stat {
1519 static const struct memory_stat memory_stats[] = {
1520 { "anon", NR_ANON_MAPPED },
1521 { "file", NR_FILE_PAGES },
1522 { "kernel_stack", NR_KERNEL_STACK_KB },
1523 { "pagetables", NR_PAGETABLE },
1524 { "percpu", MEMCG_PERCPU_B },
1525 { "sock", MEMCG_SOCK },
1526 { "shmem", NR_SHMEM },
1527 { "file_mapped", NR_FILE_MAPPED },
1528 { "file_dirty", NR_FILE_DIRTY },
1529 { "file_writeback", NR_WRITEBACK },
1530 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1531 { "anon_thp", NR_ANON_THPS },
1532 { "file_thp", NR_FILE_THPS },
1533 { "shmem_thp", NR_SHMEM_THPS },
1535 { "inactive_anon", NR_INACTIVE_ANON },
1536 { "active_anon", NR_ACTIVE_ANON },
1537 { "inactive_file", NR_INACTIVE_FILE },
1538 { "active_file", NR_ACTIVE_FILE },
1539 { "unevictable", NR_UNEVICTABLE },
1540 { "slab_reclaimable", NR_SLAB_RECLAIMABLE_B },
1541 { "slab_unreclaimable", NR_SLAB_UNRECLAIMABLE_B },
1543 /* The memory events */
1544 { "workingset_refault_anon", WORKINGSET_REFAULT_ANON },
1545 { "workingset_refault_file", WORKINGSET_REFAULT_FILE },
1546 { "workingset_activate_anon", WORKINGSET_ACTIVATE_ANON },
1547 { "workingset_activate_file", WORKINGSET_ACTIVATE_FILE },
1548 { "workingset_restore_anon", WORKINGSET_RESTORE_ANON },
1549 { "workingset_restore_file", WORKINGSET_RESTORE_FILE },
1550 { "workingset_nodereclaim", WORKINGSET_NODERECLAIM },
1553 /* Translate stat items to the correct unit for memory.stat output */
1554 static int memcg_page_state_unit(int item)
1557 case MEMCG_PERCPU_B:
1558 case NR_SLAB_RECLAIMABLE_B:
1559 case NR_SLAB_UNRECLAIMABLE_B:
1560 case WORKINGSET_REFAULT_ANON:
1561 case WORKINGSET_REFAULT_FILE:
1562 case WORKINGSET_ACTIVATE_ANON:
1563 case WORKINGSET_ACTIVATE_FILE:
1564 case WORKINGSET_RESTORE_ANON:
1565 case WORKINGSET_RESTORE_FILE:
1566 case WORKINGSET_NODERECLAIM:
1568 case NR_KERNEL_STACK_KB:
1575 static inline unsigned long memcg_page_state_output(struct mem_cgroup *memcg,
1578 return memcg_page_state(memcg, item) * memcg_page_state_unit(item);
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_output(memcg, memory_stats[i].idx);
1605 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1607 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1608 size += memcg_page_state_output(memcg,
1609 NR_SLAB_RECLAIMABLE_B);
1610 seq_buf_printf(&s, "slab %llu\n", size);
1614 /* Accumulated memory events */
1616 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1617 memcg_events(memcg, PGFAULT));
1618 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1619 memcg_events(memcg, PGMAJFAULT));
1620 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1621 memcg_events(memcg, PGREFILL));
1622 seq_buf_printf(&s, "pgscan %lu\n",
1623 memcg_events(memcg, PGSCAN_KSWAPD) +
1624 memcg_events(memcg, PGSCAN_DIRECT));
1625 seq_buf_printf(&s, "pgsteal %lu\n",
1626 memcg_events(memcg, PGSTEAL_KSWAPD) +
1627 memcg_events(memcg, PGSTEAL_DIRECT));
1628 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1629 memcg_events(memcg, PGACTIVATE));
1630 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1631 memcg_events(memcg, PGDEACTIVATE));
1632 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1633 memcg_events(memcg, PGLAZYFREE));
1634 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1635 memcg_events(memcg, PGLAZYFREED));
1637 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1638 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1639 memcg_events(memcg, THP_FAULT_ALLOC));
1640 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1641 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1642 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1644 /* The above should easily fit into one page */
1645 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1650 #define K(x) ((x) << (PAGE_SHIFT-10))
1652 * mem_cgroup_print_oom_context: Print OOM information relevant to
1653 * memory controller.
1654 * @memcg: The memory cgroup that went over limit
1655 * @p: Task that is going to be killed
1657 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1660 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1665 pr_cont(",oom_memcg=");
1666 pr_cont_cgroup_path(memcg->css.cgroup);
1668 pr_cont(",global_oom");
1670 pr_cont(",task_memcg=");
1671 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1677 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1678 * memory controller.
1679 * @memcg: The memory cgroup that went over limit
1681 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1685 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1686 K((u64)page_counter_read(&memcg->memory)),
1687 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1688 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1689 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1690 K((u64)page_counter_read(&memcg->swap)),
1691 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1693 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1694 K((u64)page_counter_read(&memcg->memsw)),
1695 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1696 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1697 K((u64)page_counter_read(&memcg->kmem)),
1698 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1701 pr_info("Memory cgroup stats for ");
1702 pr_cont_cgroup_path(memcg->css.cgroup);
1704 buf = memory_stat_format(memcg);
1712 * Return the memory (and swap, if configured) limit for a memcg.
1714 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1716 unsigned long max = READ_ONCE(memcg->memory.max);
1718 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1719 if (mem_cgroup_swappiness(memcg))
1720 max += min(READ_ONCE(memcg->swap.max),
1721 (unsigned long)total_swap_pages);
1723 if (mem_cgroup_swappiness(memcg)) {
1724 /* Calculate swap excess capacity from memsw limit */
1725 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1727 max += min(swap, (unsigned long)total_swap_pages);
1733 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1735 return page_counter_read(&memcg->memory);
1738 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1741 struct oom_control oc = {
1745 .gfp_mask = gfp_mask,
1750 if (mutex_lock_killable(&oom_lock))
1753 if (mem_cgroup_margin(memcg) >= (1 << order))
1757 * A few threads which were not waiting at mutex_lock_killable() can
1758 * fail to bail out. Therefore, check again after holding oom_lock.
1760 ret = should_force_charge() || out_of_memory(&oc);
1763 mutex_unlock(&oom_lock);
1767 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1770 unsigned long *total_scanned)
1772 struct mem_cgroup *victim = NULL;
1775 unsigned long excess;
1776 unsigned long nr_scanned;
1777 struct mem_cgroup_reclaim_cookie reclaim = {
1781 excess = soft_limit_excess(root_memcg);
1784 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1789 * If we have not been able to reclaim
1790 * anything, it might because there are
1791 * no reclaimable pages under this hierarchy
1796 * We want to do more targeted reclaim.
1797 * excess >> 2 is not to excessive so as to
1798 * reclaim too much, nor too less that we keep
1799 * coming back to reclaim from this cgroup
1801 if (total >= (excess >> 2) ||
1802 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1807 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1808 pgdat, &nr_scanned);
1809 *total_scanned += nr_scanned;
1810 if (!soft_limit_excess(root_memcg))
1813 mem_cgroup_iter_break(root_memcg, victim);
1817 #ifdef CONFIG_LOCKDEP
1818 static struct lockdep_map memcg_oom_lock_dep_map = {
1819 .name = "memcg_oom_lock",
1823 static DEFINE_SPINLOCK(memcg_oom_lock);
1826 * Check OOM-Killer is already running under our hierarchy.
1827 * If someone is running, return false.
1829 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1831 struct mem_cgroup *iter, *failed = NULL;
1833 spin_lock(&memcg_oom_lock);
1835 for_each_mem_cgroup_tree(iter, memcg) {
1836 if (iter->oom_lock) {
1838 * this subtree of our hierarchy is already locked
1839 * so we cannot give a lock.
1842 mem_cgroup_iter_break(memcg, iter);
1845 iter->oom_lock = true;
1850 * OK, we failed to lock the whole subtree so we have
1851 * to clean up what we set up to the failing subtree
1853 for_each_mem_cgroup_tree(iter, memcg) {
1854 if (iter == failed) {
1855 mem_cgroup_iter_break(memcg, iter);
1858 iter->oom_lock = false;
1861 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1863 spin_unlock(&memcg_oom_lock);
1868 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1870 struct mem_cgroup *iter;
1872 spin_lock(&memcg_oom_lock);
1873 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1874 for_each_mem_cgroup_tree(iter, memcg)
1875 iter->oom_lock = false;
1876 spin_unlock(&memcg_oom_lock);
1879 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1881 struct mem_cgroup *iter;
1883 spin_lock(&memcg_oom_lock);
1884 for_each_mem_cgroup_tree(iter, memcg)
1886 spin_unlock(&memcg_oom_lock);
1889 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1891 struct mem_cgroup *iter;
1894 * Be careful about under_oom underflows becase a child memcg
1895 * could have been added after mem_cgroup_mark_under_oom.
1897 spin_lock(&memcg_oom_lock);
1898 for_each_mem_cgroup_tree(iter, memcg)
1899 if (iter->under_oom > 0)
1901 spin_unlock(&memcg_oom_lock);
1904 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1906 struct oom_wait_info {
1907 struct mem_cgroup *memcg;
1908 wait_queue_entry_t wait;
1911 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1912 unsigned mode, int sync, void *arg)
1914 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1915 struct mem_cgroup *oom_wait_memcg;
1916 struct oom_wait_info *oom_wait_info;
1918 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1919 oom_wait_memcg = oom_wait_info->memcg;
1921 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1922 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1924 return autoremove_wake_function(wait, mode, sync, arg);
1927 static void memcg_oom_recover(struct mem_cgroup *memcg)
1930 * For the following lockless ->under_oom test, the only required
1931 * guarantee is that it must see the state asserted by an OOM when
1932 * this function is called as a result of userland actions
1933 * triggered by the notification of the OOM. This is trivially
1934 * achieved by invoking mem_cgroup_mark_under_oom() before
1935 * triggering notification.
1937 if (memcg && memcg->under_oom)
1938 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1948 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1950 enum oom_status ret;
1953 if (order > PAGE_ALLOC_COSTLY_ORDER)
1956 memcg_memory_event(memcg, MEMCG_OOM);
1959 * We are in the middle of the charge context here, so we
1960 * don't want to block when potentially sitting on a callstack
1961 * that holds all kinds of filesystem and mm locks.
1963 * cgroup1 allows disabling the OOM killer and waiting for outside
1964 * handling until the charge can succeed; remember the context and put
1965 * the task to sleep at the end of the page fault when all locks are
1968 * On the other hand, in-kernel OOM killer allows for an async victim
1969 * memory reclaim (oom_reaper) and that means that we are not solely
1970 * relying on the oom victim to make a forward progress and we can
1971 * invoke the oom killer here.
1973 * Please note that mem_cgroup_out_of_memory might fail to find a
1974 * victim and then we have to bail out from the charge path.
1976 if (memcg->oom_kill_disable) {
1977 if (!current->in_user_fault)
1979 css_get(&memcg->css);
1980 current->memcg_in_oom = memcg;
1981 current->memcg_oom_gfp_mask = mask;
1982 current->memcg_oom_order = order;
1987 mem_cgroup_mark_under_oom(memcg);
1989 locked = mem_cgroup_oom_trylock(memcg);
1992 mem_cgroup_oom_notify(memcg);
1994 mem_cgroup_unmark_under_oom(memcg);
1995 if (mem_cgroup_out_of_memory(memcg, mask, order))
2001 mem_cgroup_oom_unlock(memcg);
2007 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2008 * @handle: actually kill/wait or just clean up the OOM state
2010 * This has to be called at the end of a page fault if the memcg OOM
2011 * handler was enabled.
2013 * Memcg supports userspace OOM handling where failed allocations must
2014 * sleep on a waitqueue until the userspace task resolves the
2015 * situation. Sleeping directly in the charge context with all kinds
2016 * of locks held is not a good idea, instead we remember an OOM state
2017 * in the task and mem_cgroup_oom_synchronize() has to be called at
2018 * the end of the page fault to complete the OOM handling.
2020 * Returns %true if an ongoing memcg OOM situation was detected and
2021 * completed, %false otherwise.
2023 bool mem_cgroup_oom_synchronize(bool handle)
2025 struct mem_cgroup *memcg = current->memcg_in_oom;
2026 struct oom_wait_info owait;
2029 /* OOM is global, do not handle */
2036 owait.memcg = memcg;
2037 owait.wait.flags = 0;
2038 owait.wait.func = memcg_oom_wake_function;
2039 owait.wait.private = current;
2040 INIT_LIST_HEAD(&owait.wait.entry);
2042 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2043 mem_cgroup_mark_under_oom(memcg);
2045 locked = mem_cgroup_oom_trylock(memcg);
2048 mem_cgroup_oom_notify(memcg);
2050 if (locked && !memcg->oom_kill_disable) {
2051 mem_cgroup_unmark_under_oom(memcg);
2052 finish_wait(&memcg_oom_waitq, &owait.wait);
2053 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
2054 current->memcg_oom_order);
2057 mem_cgroup_unmark_under_oom(memcg);
2058 finish_wait(&memcg_oom_waitq, &owait.wait);
2062 mem_cgroup_oom_unlock(memcg);
2064 * There is no guarantee that an OOM-lock contender
2065 * sees the wakeups triggered by the OOM kill
2066 * uncharges. Wake any sleepers explicitely.
2068 memcg_oom_recover(memcg);
2071 current->memcg_in_oom = NULL;
2072 css_put(&memcg->css);
2077 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2078 * @victim: task to be killed by the OOM killer
2079 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2081 * Returns a pointer to a memory cgroup, which has to be cleaned up
2082 * by killing all belonging OOM-killable tasks.
2084 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2086 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2087 struct mem_cgroup *oom_domain)
2089 struct mem_cgroup *oom_group = NULL;
2090 struct mem_cgroup *memcg;
2092 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2096 oom_domain = root_mem_cgroup;
2100 memcg = mem_cgroup_from_task(victim);
2101 if (memcg == root_mem_cgroup)
2105 * If the victim task has been asynchronously moved to a different
2106 * memory cgroup, we might end up killing tasks outside oom_domain.
2107 * In this case it's better to ignore memory.group.oom.
2109 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2113 * Traverse the memory cgroup hierarchy from the victim task's
2114 * cgroup up to the OOMing cgroup (or root) to find the
2115 * highest-level memory cgroup with oom.group set.
2117 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2118 if (memcg->oom_group)
2121 if (memcg == oom_domain)
2126 css_get(&oom_group->css);
2133 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2135 pr_info("Tasks in ");
2136 pr_cont_cgroup_path(memcg->css.cgroup);
2137 pr_cont(" are going to be killed due to memory.oom.group set\n");
2141 * lock_page_memcg - lock a page and memcg binding
2144 * This function protects unlocked LRU pages from being moved to
2147 * It ensures lifetime of the returned memcg. Caller is responsible
2148 * for the lifetime of the page; __unlock_page_memcg() is available
2149 * when @page might get freed inside the locked section.
2151 struct mem_cgroup *lock_page_memcg(struct page *page)
2153 struct page *head = compound_head(page); /* rmap on tail pages */
2154 struct mem_cgroup *memcg;
2155 unsigned long flags;
2158 * The RCU lock is held throughout the transaction. The fast
2159 * path can get away without acquiring the memcg->move_lock
2160 * because page moving starts with an RCU grace period.
2162 * The RCU lock also protects the memcg from being freed when
2163 * the page state that is going to change is the only thing
2164 * preventing the page itself from being freed. E.g. writeback
2165 * doesn't hold a page reference and relies on PG_writeback to
2166 * keep off truncation, migration and so forth.
2170 if (mem_cgroup_disabled())
2173 memcg = page_memcg(head);
2174 if (unlikely(!memcg))
2177 #ifdef CONFIG_PROVE_LOCKING
2178 local_irq_save(flags);
2179 might_lock(&memcg->move_lock);
2180 local_irq_restore(flags);
2183 if (atomic_read(&memcg->moving_account) <= 0)
2186 spin_lock_irqsave(&memcg->move_lock, flags);
2187 if (memcg != page_memcg(head)) {
2188 spin_unlock_irqrestore(&memcg->move_lock, flags);
2193 * When charge migration first begins, we can have locked and
2194 * unlocked page stat updates happening concurrently. Track
2195 * the task who has the lock for unlock_page_memcg().
2197 memcg->move_lock_task = current;
2198 memcg->move_lock_flags = flags;
2202 EXPORT_SYMBOL(lock_page_memcg);
2205 * __unlock_page_memcg - unlock and unpin a memcg
2208 * Unlock and unpin a memcg returned by lock_page_memcg().
2210 void __unlock_page_memcg(struct mem_cgroup *memcg)
2212 if (memcg && memcg->move_lock_task == current) {
2213 unsigned long flags = memcg->move_lock_flags;
2215 memcg->move_lock_task = NULL;
2216 memcg->move_lock_flags = 0;
2218 spin_unlock_irqrestore(&memcg->move_lock, flags);
2225 * unlock_page_memcg - unlock a page and memcg binding
2228 void unlock_page_memcg(struct page *page)
2230 struct page *head = compound_head(page);
2232 __unlock_page_memcg(page_memcg(head));
2234 EXPORT_SYMBOL(unlock_page_memcg);
2236 struct memcg_stock_pcp {
2237 struct mem_cgroup *cached; /* this never be root cgroup */
2238 unsigned int nr_pages;
2240 #ifdef CONFIG_MEMCG_KMEM
2241 struct obj_cgroup *cached_objcg;
2242 unsigned int nr_bytes;
2245 struct work_struct work;
2246 unsigned long flags;
2247 #define FLUSHING_CACHED_CHARGE 0
2249 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2250 static DEFINE_MUTEX(percpu_charge_mutex);
2252 #ifdef CONFIG_MEMCG_KMEM
2253 static void drain_obj_stock(struct memcg_stock_pcp *stock);
2254 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2255 struct mem_cgroup *root_memcg);
2258 static inline void drain_obj_stock(struct memcg_stock_pcp *stock)
2261 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2262 struct mem_cgroup *root_memcg)
2269 * consume_stock: Try to consume stocked charge on this cpu.
2270 * @memcg: memcg to consume from.
2271 * @nr_pages: how many pages to charge.
2273 * The charges will only happen if @memcg matches the current cpu's memcg
2274 * stock, and at least @nr_pages are available in that stock. Failure to
2275 * service an allocation will refill the stock.
2277 * returns true if successful, false otherwise.
2279 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2281 struct memcg_stock_pcp *stock;
2282 unsigned long flags;
2285 if (nr_pages > MEMCG_CHARGE_BATCH)
2288 local_irq_save(flags);
2290 stock = this_cpu_ptr(&memcg_stock);
2291 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2292 stock->nr_pages -= nr_pages;
2296 local_irq_restore(flags);
2302 * Returns stocks cached in percpu and reset cached information.
2304 static void drain_stock(struct memcg_stock_pcp *stock)
2306 struct mem_cgroup *old = stock->cached;
2311 if (stock->nr_pages) {
2312 page_counter_uncharge(&old->memory, stock->nr_pages);
2313 if (do_memsw_account())
2314 page_counter_uncharge(&old->memsw, stock->nr_pages);
2315 stock->nr_pages = 0;
2319 stock->cached = NULL;
2322 static void drain_local_stock(struct work_struct *dummy)
2324 struct memcg_stock_pcp *stock;
2325 unsigned long flags;
2328 * The only protection from memory hotplug vs. drain_stock races is
2329 * that we always operate on local CPU stock here with IRQ disabled
2331 local_irq_save(flags);
2333 stock = this_cpu_ptr(&memcg_stock);
2334 drain_obj_stock(stock);
2336 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2338 local_irq_restore(flags);
2342 * Cache charges(val) to local per_cpu area.
2343 * This will be consumed by consume_stock() function, later.
2345 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2347 struct memcg_stock_pcp *stock;
2348 unsigned long flags;
2350 local_irq_save(flags);
2352 stock = this_cpu_ptr(&memcg_stock);
2353 if (stock->cached != memcg) { /* reset if necessary */
2355 css_get(&memcg->css);
2356 stock->cached = memcg;
2358 stock->nr_pages += nr_pages;
2360 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2363 local_irq_restore(flags);
2367 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2368 * of the hierarchy under it.
2370 static void drain_all_stock(struct mem_cgroup *root_memcg)
2374 /* If someone's already draining, avoid adding running more workers. */
2375 if (!mutex_trylock(&percpu_charge_mutex))
2378 * Notify other cpus that system-wide "drain" is running
2379 * We do not care about races with the cpu hotplug because cpu down
2380 * as well as workers from this path always operate on the local
2381 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2384 for_each_online_cpu(cpu) {
2385 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2386 struct mem_cgroup *memcg;
2390 memcg = stock->cached;
2391 if (memcg && stock->nr_pages &&
2392 mem_cgroup_is_descendant(memcg, root_memcg))
2394 if (obj_stock_flush_required(stock, root_memcg))
2399 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2401 drain_local_stock(&stock->work);
2403 schedule_work_on(cpu, &stock->work);
2407 mutex_unlock(&percpu_charge_mutex);
2410 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2412 struct memcg_stock_pcp *stock;
2413 struct mem_cgroup *memcg, *mi;
2415 stock = &per_cpu(memcg_stock, cpu);
2418 for_each_mem_cgroup(memcg) {
2421 for (i = 0; i < MEMCG_NR_STAT; i++) {
2425 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2427 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2428 atomic_long_add(x, &memcg->vmstats[i]);
2430 if (i >= NR_VM_NODE_STAT_ITEMS)
2433 for_each_node(nid) {
2434 struct mem_cgroup_per_node *pn;
2436 pn = mem_cgroup_nodeinfo(memcg, nid);
2437 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2440 atomic_long_add(x, &pn->lruvec_stat[i]);
2441 } while ((pn = parent_nodeinfo(pn, nid)));
2445 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2448 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2450 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2451 atomic_long_add(x, &memcg->vmevents[i]);
2458 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2459 unsigned int nr_pages,
2462 unsigned long nr_reclaimed = 0;
2465 unsigned long pflags;
2467 if (page_counter_read(&memcg->memory) <=
2468 READ_ONCE(memcg->memory.high))
2471 memcg_memory_event(memcg, MEMCG_HIGH);
2473 psi_memstall_enter(&pflags);
2474 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2476 psi_memstall_leave(&pflags);
2477 } while ((memcg = parent_mem_cgroup(memcg)) &&
2478 !mem_cgroup_is_root(memcg));
2480 return nr_reclaimed;
2483 static void high_work_func(struct work_struct *work)
2485 struct mem_cgroup *memcg;
2487 memcg = container_of(work, struct mem_cgroup, high_work);
2488 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2492 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2493 * enough to still cause a significant slowdown in most cases, while still
2494 * allowing diagnostics and tracing to proceed without becoming stuck.
2496 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2499 * When calculating the delay, we use these either side of the exponentiation to
2500 * maintain precision and scale to a reasonable number of jiffies (see the table
2503 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2504 * overage ratio to a delay.
2505 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2506 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2507 * to produce a reasonable delay curve.
2509 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2510 * reasonable delay curve compared to precision-adjusted overage, not
2511 * penalising heavily at first, but still making sure that growth beyond the
2512 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2513 * example, with a high of 100 megabytes:
2515 * +-------+------------------------+
2516 * | usage | time to allocate in ms |
2517 * +-------+------------------------+
2539 * +-------+------------------------+
2541 #define MEMCG_DELAY_PRECISION_SHIFT 20
2542 #define MEMCG_DELAY_SCALING_SHIFT 14
2544 static u64 calculate_overage(unsigned long usage, unsigned long high)
2552 * Prevent division by 0 in overage calculation by acting as if
2553 * it was a threshold of 1 page
2555 high = max(high, 1UL);
2557 overage = usage - high;
2558 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2559 return div64_u64(overage, high);
2562 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2564 u64 overage, max_overage = 0;
2567 overage = calculate_overage(page_counter_read(&memcg->memory),
2568 READ_ONCE(memcg->memory.high));
2569 max_overage = max(overage, max_overage);
2570 } while ((memcg = parent_mem_cgroup(memcg)) &&
2571 !mem_cgroup_is_root(memcg));
2576 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2578 u64 overage, max_overage = 0;
2581 overage = calculate_overage(page_counter_read(&memcg->swap),
2582 READ_ONCE(memcg->swap.high));
2584 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2585 max_overage = max(overage, max_overage);
2586 } while ((memcg = parent_mem_cgroup(memcg)) &&
2587 !mem_cgroup_is_root(memcg));
2593 * Get the number of jiffies that we should penalise a mischievous cgroup which
2594 * is exceeding its memory.high by checking both it and its ancestors.
2596 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2597 unsigned int nr_pages,
2600 unsigned long penalty_jiffies;
2606 * We use overage compared to memory.high to calculate the number of
2607 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2608 * fairly lenient on small overages, and increasingly harsh when the
2609 * memcg in question makes it clear that it has no intention of stopping
2610 * its crazy behaviour, so we exponentially increase the delay based on
2613 penalty_jiffies = max_overage * max_overage * HZ;
2614 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2615 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2618 * Factor in the task's own contribution to the overage, such that four
2619 * N-sized allocations are throttled approximately the same as one
2620 * 4N-sized allocation.
2622 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2623 * larger the current charge patch is than that.
2625 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2629 * Scheduled by try_charge() to be executed from the userland return path
2630 * and reclaims memory over the high limit.
2632 void mem_cgroup_handle_over_high(void)
2634 unsigned long penalty_jiffies;
2635 unsigned long pflags;
2636 unsigned long nr_reclaimed;
2637 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2638 int nr_retries = MAX_RECLAIM_RETRIES;
2639 struct mem_cgroup *memcg;
2640 bool in_retry = false;
2642 if (likely(!nr_pages))
2645 memcg = get_mem_cgroup_from_mm(current->mm);
2646 current->memcg_nr_pages_over_high = 0;
2650 * The allocating task should reclaim at least the batch size, but for
2651 * subsequent retries we only want to do what's necessary to prevent oom
2652 * or breaching resource isolation.
2654 * This is distinct from memory.max or page allocator behaviour because
2655 * memory.high is currently batched, whereas memory.max and the page
2656 * allocator run every time an allocation is made.
2658 nr_reclaimed = reclaim_high(memcg,
2659 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2663 * memory.high is breached and reclaim is unable to keep up. Throttle
2664 * allocators proactively to slow down excessive growth.
2666 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2667 mem_find_max_overage(memcg));
2669 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2670 swap_find_max_overage(memcg));
2673 * Clamp the max delay per usermode return so as to still keep the
2674 * application moving forwards and also permit diagnostics, albeit
2677 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2680 * Don't sleep if the amount of jiffies this memcg owes us is so low
2681 * that it's not even worth doing, in an attempt to be nice to those who
2682 * go only a small amount over their memory.high value and maybe haven't
2683 * been aggressively reclaimed enough yet.
2685 if (penalty_jiffies <= HZ / 100)
2689 * If reclaim is making forward progress but we're still over
2690 * memory.high, we want to encourage that rather than doing allocator
2693 if (nr_reclaimed || nr_retries--) {
2699 * If we exit early, we're guaranteed to die (since
2700 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2701 * need to account for any ill-begotten jiffies to pay them off later.
2703 psi_memstall_enter(&pflags);
2704 schedule_timeout_killable(penalty_jiffies);
2705 psi_memstall_leave(&pflags);
2708 css_put(&memcg->css);
2711 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2712 unsigned int nr_pages)
2714 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2715 int nr_retries = MAX_RECLAIM_RETRIES;
2716 struct mem_cgroup *mem_over_limit;
2717 struct page_counter *counter;
2718 enum oom_status oom_status;
2719 unsigned long nr_reclaimed;
2720 bool may_swap = true;
2721 bool drained = false;
2722 unsigned long pflags;
2724 if (mem_cgroup_is_root(memcg))
2727 if (consume_stock(memcg, nr_pages))
2730 if (!do_memsw_account() ||
2731 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2732 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2734 if (do_memsw_account())
2735 page_counter_uncharge(&memcg->memsw, batch);
2736 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2738 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2742 if (batch > nr_pages) {
2748 * Memcg doesn't have a dedicated reserve for atomic
2749 * allocations. But like the global atomic pool, we need to
2750 * put the burden of reclaim on regular allocation requests
2751 * and let these go through as privileged allocations.
2753 if (gfp_mask & __GFP_ATOMIC)
2757 * Unlike in global OOM situations, memcg is not in a physical
2758 * memory shortage. Allow dying and OOM-killed tasks to
2759 * bypass the last charges so that they can exit quickly and
2760 * free their memory.
2762 if (unlikely(should_force_charge()))
2766 * Prevent unbounded recursion when reclaim operations need to
2767 * allocate memory. This might exceed the limits temporarily,
2768 * but we prefer facilitating memory reclaim and getting back
2769 * under the limit over triggering OOM kills in these cases.
2771 if (unlikely(current->flags & PF_MEMALLOC))
2774 if (unlikely(task_in_memcg_oom(current)))
2777 if (!gfpflags_allow_blocking(gfp_mask))
2780 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2782 psi_memstall_enter(&pflags);
2783 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2784 gfp_mask, may_swap);
2785 psi_memstall_leave(&pflags);
2787 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2791 drain_all_stock(mem_over_limit);
2796 if (gfp_mask & __GFP_NORETRY)
2799 * Even though the limit is exceeded at this point, reclaim
2800 * may have been able to free some pages. Retry the charge
2801 * before killing the task.
2803 * Only for regular pages, though: huge pages are rather
2804 * unlikely to succeed so close to the limit, and we fall back
2805 * to regular pages anyway in case of failure.
2807 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2810 * At task move, charge accounts can be doubly counted. So, it's
2811 * better to wait until the end of task_move if something is going on.
2813 if (mem_cgroup_wait_acct_move(mem_over_limit))
2819 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2822 if (gfp_mask & __GFP_NOFAIL)
2825 if (fatal_signal_pending(current))
2829 * keep retrying as long as the memcg oom killer is able to make
2830 * a forward progress or bypass the charge if the oom killer
2831 * couldn't make any progress.
2833 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2834 get_order(nr_pages * PAGE_SIZE));
2835 switch (oom_status) {
2837 nr_retries = MAX_RECLAIM_RETRIES;
2845 if (!(gfp_mask & __GFP_NOFAIL))
2849 * The allocation either can't fail or will lead to more memory
2850 * being freed very soon. Allow memory usage go over the limit
2851 * temporarily by force charging it.
2853 page_counter_charge(&memcg->memory, nr_pages);
2854 if (do_memsw_account())
2855 page_counter_charge(&memcg->memsw, nr_pages);
2860 if (batch > nr_pages)
2861 refill_stock(memcg, batch - nr_pages);
2864 * If the hierarchy is above the normal consumption range, schedule
2865 * reclaim on returning to userland. We can perform reclaim here
2866 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2867 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2868 * not recorded as it most likely matches current's and won't
2869 * change in the meantime. As high limit is checked again before
2870 * reclaim, the cost of mismatch is negligible.
2873 bool mem_high, swap_high;
2875 mem_high = page_counter_read(&memcg->memory) >
2876 READ_ONCE(memcg->memory.high);
2877 swap_high = page_counter_read(&memcg->swap) >
2878 READ_ONCE(memcg->swap.high);
2880 /* Don't bother a random interrupted task */
2881 if (in_interrupt()) {
2883 schedule_work(&memcg->high_work);
2889 if (mem_high || swap_high) {
2891 * The allocating tasks in this cgroup will need to do
2892 * reclaim or be throttled to prevent further growth
2893 * of the memory or swap footprints.
2895 * Target some best-effort fairness between the tasks,
2896 * and distribute reclaim work and delay penalties
2897 * based on how much each task is actually allocating.
2899 current->memcg_nr_pages_over_high += batch;
2900 set_notify_resume(current);
2903 } while ((memcg = parent_mem_cgroup(memcg)));
2908 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2909 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2911 if (mem_cgroup_is_root(memcg))
2914 page_counter_uncharge(&memcg->memory, nr_pages);
2915 if (do_memsw_account())
2916 page_counter_uncharge(&memcg->memsw, nr_pages);
2920 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2922 VM_BUG_ON_PAGE(page_memcg(page), page);
2924 * Any of the following ensures page's memcg stability:
2928 * - lock_page_memcg()
2929 * - exclusive reference
2931 page->memcg_data = (unsigned long)memcg;
2934 #ifdef CONFIG_MEMCG_KMEM
2935 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2936 gfp_t gfp, bool new_page)
2938 unsigned int objects = objs_per_slab_page(s, page);
2939 unsigned long memcg_data;
2942 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2947 memcg_data = (unsigned long) vec | MEMCG_DATA_OBJCGS;
2950 * If the slab page is brand new and nobody can yet access
2951 * it's memcg_data, no synchronization is required and
2952 * memcg_data can be simply assigned.
2954 page->memcg_data = memcg_data;
2955 } else if (cmpxchg(&page->memcg_data, 0, memcg_data)) {
2957 * If the slab page is already in use, somebody can allocate
2958 * and assign obj_cgroups in parallel. In this case the existing
2959 * objcg vector should be reused.
2965 kmemleak_not_leak(vec);
2970 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2972 * A passed kernel object can be a slab object or a generic kernel page, so
2973 * different mechanisms for getting the memory cgroup pointer should be used.
2974 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2975 * can not know for sure how the kernel object is implemented.
2976 * mem_cgroup_from_obj() can be safely used in such cases.
2978 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2979 * cgroup_mutex, etc.
2981 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2985 if (mem_cgroup_disabled())
2988 page = virt_to_head_page(p);
2991 * Slab objects are accounted individually, not per-page.
2992 * Memcg membership data for each individual object is saved in
2993 * the page->obj_cgroups.
2995 if (page_objcgs_check(page)) {
2996 struct obj_cgroup *objcg;
2999 off = obj_to_index(page->slab_cache, page, p);
3000 objcg = page_objcgs(page)[off];
3002 return obj_cgroup_memcg(objcg);
3008 * page_memcg_check() is used here, because page_has_obj_cgroups()
3009 * check above could fail because the object cgroups vector wasn't set
3010 * at that moment, but it can be set concurrently.
3011 * page_memcg_check(page) will guarantee that a proper memory
3012 * cgroup pointer or NULL will be returned.
3014 return page_memcg_check(page);
3017 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
3019 struct obj_cgroup *objcg = NULL;
3020 struct mem_cgroup *memcg;
3022 if (memcg_kmem_bypass())
3026 if (unlikely(active_memcg()))
3027 memcg = active_memcg();
3029 memcg = mem_cgroup_from_task(current);
3031 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
3032 objcg = rcu_dereference(memcg->objcg);
3033 if (objcg && obj_cgroup_tryget(objcg))
3042 static int memcg_alloc_cache_id(void)
3047 id = ida_simple_get(&memcg_cache_ida,
3048 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3052 if (id < memcg_nr_cache_ids)
3056 * There's no space for the new id in memcg_caches arrays,
3057 * so we have to grow them.
3059 down_write(&memcg_cache_ids_sem);
3061 size = 2 * (id + 1);
3062 if (size < MEMCG_CACHES_MIN_SIZE)
3063 size = MEMCG_CACHES_MIN_SIZE;
3064 else if (size > MEMCG_CACHES_MAX_SIZE)
3065 size = MEMCG_CACHES_MAX_SIZE;
3067 err = memcg_update_all_list_lrus(size);
3069 memcg_nr_cache_ids = size;
3071 up_write(&memcg_cache_ids_sem);
3074 ida_simple_remove(&memcg_cache_ida, id);
3080 static void memcg_free_cache_id(int id)
3082 ida_simple_remove(&memcg_cache_ida, id);
3086 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
3087 * @memcg: memory cgroup to charge
3088 * @gfp: reclaim mode
3089 * @nr_pages: number of pages to charge
3091 * Returns 0 on success, an error code on failure.
3093 int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
3094 unsigned int nr_pages)
3096 struct page_counter *counter;
3099 ret = try_charge(memcg, gfp, nr_pages);
3103 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3104 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3107 * Enforce __GFP_NOFAIL allocation because callers are not
3108 * prepared to see failures and likely do not have any failure
3111 if (gfp & __GFP_NOFAIL) {
3112 page_counter_charge(&memcg->kmem, nr_pages);
3115 cancel_charge(memcg, nr_pages);
3122 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
3123 * @memcg: memcg to uncharge
3124 * @nr_pages: number of pages to uncharge
3126 void __memcg_kmem_uncharge(struct mem_cgroup *memcg, unsigned int nr_pages)
3128 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3129 page_counter_uncharge(&memcg->kmem, nr_pages);
3131 refill_stock(memcg, nr_pages);
3135 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3136 * @page: page to charge
3137 * @gfp: reclaim mode
3138 * @order: allocation order
3140 * Returns 0 on success, an error code on failure.
3142 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3144 struct mem_cgroup *memcg;
3147 memcg = get_mem_cgroup_from_current();
3148 if (memcg && !mem_cgroup_is_root(memcg)) {
3149 ret = __memcg_kmem_charge(memcg, gfp, 1 << order);
3151 page->memcg_data = (unsigned long)memcg |
3155 css_put(&memcg->css);
3161 * __memcg_kmem_uncharge_page: uncharge a kmem page
3162 * @page: page to uncharge
3163 * @order: allocation order
3165 void __memcg_kmem_uncharge_page(struct page *page, int order)
3167 struct mem_cgroup *memcg = page_memcg(page);
3168 unsigned int nr_pages = 1 << order;
3173 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3174 __memcg_kmem_uncharge(memcg, nr_pages);
3175 page->memcg_data = 0;
3176 css_put(&memcg->css);
3179 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3181 struct memcg_stock_pcp *stock;
3182 unsigned long flags;
3185 local_irq_save(flags);
3187 stock = this_cpu_ptr(&memcg_stock);
3188 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3189 stock->nr_bytes -= nr_bytes;
3193 local_irq_restore(flags);
3198 static void drain_obj_stock(struct memcg_stock_pcp *stock)
3200 struct obj_cgroup *old = stock->cached_objcg;
3205 if (stock->nr_bytes) {
3206 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3207 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3211 __memcg_kmem_uncharge(obj_cgroup_memcg(old), nr_pages);
3216 * The leftover is flushed to the centralized per-memcg value.
3217 * On the next attempt to refill obj stock it will be moved
3218 * to a per-cpu stock (probably, on an other CPU), see
3219 * refill_obj_stock().
3221 * How often it's flushed is a trade-off between the memory
3222 * limit enforcement accuracy and potential CPU contention,
3223 * so it might be changed in the future.
3225 atomic_add(nr_bytes, &old->nr_charged_bytes);
3226 stock->nr_bytes = 0;
3229 obj_cgroup_put(old);
3230 stock->cached_objcg = NULL;
3233 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3234 struct mem_cgroup *root_memcg)
3236 struct mem_cgroup *memcg;
3238 if (stock->cached_objcg) {
3239 memcg = obj_cgroup_memcg(stock->cached_objcg);
3240 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3247 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3249 struct memcg_stock_pcp *stock;
3250 unsigned long flags;
3252 local_irq_save(flags);
3254 stock = this_cpu_ptr(&memcg_stock);
3255 if (stock->cached_objcg != objcg) { /* reset if necessary */
3256 drain_obj_stock(stock);
3257 obj_cgroup_get(objcg);
3258 stock->cached_objcg = objcg;
3259 stock->nr_bytes = atomic_xchg(&objcg->nr_charged_bytes, 0);
3261 stock->nr_bytes += nr_bytes;
3263 if (stock->nr_bytes > PAGE_SIZE)
3264 drain_obj_stock(stock);
3266 local_irq_restore(flags);
3269 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3271 struct mem_cgroup *memcg;
3272 unsigned int nr_pages, nr_bytes;
3275 if (consume_obj_stock(objcg, size))
3279 * In theory, memcg->nr_charged_bytes can have enough
3280 * pre-charged bytes to satisfy the allocation. However,
3281 * flushing memcg->nr_charged_bytes requires two atomic
3282 * operations, and memcg->nr_charged_bytes can't be big,
3283 * so it's better to ignore it and try grab some new pages.
3284 * memcg->nr_charged_bytes will be flushed in
3285 * refill_obj_stock(), called from this function or
3286 * independently later.
3290 memcg = obj_cgroup_memcg(objcg);
3291 if (unlikely(!css_tryget(&memcg->css)))
3295 nr_pages = size >> PAGE_SHIFT;
3296 nr_bytes = size & (PAGE_SIZE - 1);
3301 ret = __memcg_kmem_charge(memcg, gfp, nr_pages);
3302 if (!ret && nr_bytes)
3303 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes);
3305 css_put(&memcg->css);
3309 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3311 refill_obj_stock(objcg, size);
3314 #endif /* CONFIG_MEMCG_KMEM */
3316 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3318 * Because page_memcg(head) is not set on compound tails, set it now.
3320 void mem_cgroup_split_huge_fixup(struct page *head)
3322 struct mem_cgroup *memcg = page_memcg(head);
3325 if (mem_cgroup_disabled())
3328 for (i = 1; i < HPAGE_PMD_NR; i++) {
3329 css_get(&memcg->css);
3330 head[i].memcg_data = (unsigned long)memcg;
3333 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3335 #ifdef CONFIG_MEMCG_SWAP
3337 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3338 * @entry: swap entry to be moved
3339 * @from: mem_cgroup which the entry is moved from
3340 * @to: mem_cgroup which the entry is moved to
3342 * It succeeds only when the swap_cgroup's record for this entry is the same
3343 * as the mem_cgroup's id of @from.
3345 * Returns 0 on success, -EINVAL on failure.
3347 * The caller must have charged to @to, IOW, called page_counter_charge() about
3348 * both res and memsw, and called css_get().
3350 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3351 struct mem_cgroup *from, struct mem_cgroup *to)
3353 unsigned short old_id, new_id;
3355 old_id = mem_cgroup_id(from);
3356 new_id = mem_cgroup_id(to);
3358 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3359 mod_memcg_state(from, MEMCG_SWAP, -1);
3360 mod_memcg_state(to, MEMCG_SWAP, 1);
3366 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3367 struct mem_cgroup *from, struct mem_cgroup *to)
3373 static DEFINE_MUTEX(memcg_max_mutex);
3375 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3376 unsigned long max, bool memsw)
3378 bool enlarge = false;
3379 bool drained = false;
3381 bool limits_invariant;
3382 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3385 if (signal_pending(current)) {
3390 mutex_lock(&memcg_max_mutex);
3392 * Make sure that the new limit (memsw or memory limit) doesn't
3393 * break our basic invariant rule memory.max <= memsw.max.
3395 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3396 max <= memcg->memsw.max;
3397 if (!limits_invariant) {
3398 mutex_unlock(&memcg_max_mutex);
3402 if (max > counter->max)
3404 ret = page_counter_set_max(counter, max);
3405 mutex_unlock(&memcg_max_mutex);
3411 drain_all_stock(memcg);
3416 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3417 GFP_KERNEL, !memsw)) {
3423 if (!ret && enlarge)
3424 memcg_oom_recover(memcg);
3429 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3431 unsigned long *total_scanned)
3433 unsigned long nr_reclaimed = 0;
3434 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3435 unsigned long reclaimed;
3437 struct mem_cgroup_tree_per_node *mctz;
3438 unsigned long excess;
3439 unsigned long nr_scanned;
3444 mctz = soft_limit_tree_node(pgdat->node_id);
3447 * Do not even bother to check the largest node if the root
3448 * is empty. Do it lockless to prevent lock bouncing. Races
3449 * are acceptable as soft limit is best effort anyway.
3451 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3455 * This loop can run a while, specially if mem_cgroup's continuously
3456 * keep exceeding their soft limit and putting the system under
3463 mz = mem_cgroup_largest_soft_limit_node(mctz);
3468 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3469 gfp_mask, &nr_scanned);
3470 nr_reclaimed += reclaimed;
3471 *total_scanned += nr_scanned;
3472 spin_lock_irq(&mctz->lock);
3473 __mem_cgroup_remove_exceeded(mz, mctz);
3476 * If we failed to reclaim anything from this memory cgroup
3477 * it is time to move on to the next cgroup
3481 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3483 excess = soft_limit_excess(mz->memcg);
3485 * One school of thought says that we should not add
3486 * back the node to the tree if reclaim returns 0.
3487 * But our reclaim could return 0, simply because due
3488 * to priority we are exposing a smaller subset of
3489 * memory to reclaim from. Consider this as a longer
3492 /* If excess == 0, no tree ops */
3493 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3494 spin_unlock_irq(&mctz->lock);
3495 css_put(&mz->memcg->css);
3498 * Could not reclaim anything and there are no more
3499 * mem cgroups to try or we seem to be looping without
3500 * reclaiming anything.
3502 if (!nr_reclaimed &&
3504 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3506 } while (!nr_reclaimed);
3508 css_put(&next_mz->memcg->css);
3509 return nr_reclaimed;
3513 * Reclaims as many pages from the given memcg as possible.
3515 * Caller is responsible for holding css reference for memcg.
3517 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3519 int nr_retries = MAX_RECLAIM_RETRIES;
3521 /* we call try-to-free pages for make this cgroup empty */
3522 lru_add_drain_all();
3524 drain_all_stock(memcg);
3526 /* try to free all pages in this cgroup */
3527 while (nr_retries && page_counter_read(&memcg->memory)) {
3530 if (signal_pending(current))
3533 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3537 /* maybe some writeback is necessary */
3538 congestion_wait(BLK_RW_ASYNC, HZ/10);
3546 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3547 char *buf, size_t nbytes,
3550 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3552 if (mem_cgroup_is_root(memcg))
3554 return mem_cgroup_force_empty(memcg) ?: nbytes;
3557 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3563 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3564 struct cftype *cft, u64 val)
3569 pr_warn_once("Non-hierarchical mode is deprecated. "
3570 "Please report your usecase to linux-mm@kvack.org if you "
3571 "depend on this functionality.\n");
3576 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3580 if (mem_cgroup_is_root(memcg)) {
3581 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3582 memcg_page_state(memcg, NR_ANON_MAPPED);
3584 val += memcg_page_state(memcg, MEMCG_SWAP);
3587 val = page_counter_read(&memcg->memory);
3589 val = page_counter_read(&memcg->memsw);
3602 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3605 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3606 struct page_counter *counter;
3608 switch (MEMFILE_TYPE(cft->private)) {
3610 counter = &memcg->memory;
3613 counter = &memcg->memsw;
3616 counter = &memcg->kmem;
3619 counter = &memcg->tcpmem;
3625 switch (MEMFILE_ATTR(cft->private)) {
3627 if (counter == &memcg->memory)
3628 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3629 if (counter == &memcg->memsw)
3630 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3631 return (u64)page_counter_read(counter) * PAGE_SIZE;
3633 return (u64)counter->max * PAGE_SIZE;
3635 return (u64)counter->watermark * PAGE_SIZE;
3637 return counter->failcnt;
3638 case RES_SOFT_LIMIT:
3639 return (u64)memcg->soft_limit * PAGE_SIZE;
3645 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3647 unsigned long stat[MEMCG_NR_STAT] = {0};
3648 struct mem_cgroup *mi;
3651 for_each_online_cpu(cpu)
3652 for (i = 0; i < MEMCG_NR_STAT; i++)
3653 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3655 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3656 for (i = 0; i < MEMCG_NR_STAT; i++)
3657 atomic_long_add(stat[i], &mi->vmstats[i]);
3659 for_each_node(node) {
3660 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3661 struct mem_cgroup_per_node *pi;
3663 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3666 for_each_online_cpu(cpu)
3667 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3669 pn->lruvec_stat_cpu->count[i], cpu);
3671 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3672 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3673 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3677 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3679 unsigned long events[NR_VM_EVENT_ITEMS];
3680 struct mem_cgroup *mi;
3683 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3686 for_each_online_cpu(cpu)
3687 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3688 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3691 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3692 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3693 atomic_long_add(events[i], &mi->vmevents[i]);
3696 #ifdef CONFIG_MEMCG_KMEM
3697 static int memcg_online_kmem(struct mem_cgroup *memcg)
3699 struct obj_cgroup *objcg;
3702 if (cgroup_memory_nokmem)
3705 BUG_ON(memcg->kmemcg_id >= 0);
3706 BUG_ON(memcg->kmem_state);
3708 memcg_id = memcg_alloc_cache_id();
3712 objcg = obj_cgroup_alloc();
3714 memcg_free_cache_id(memcg_id);
3717 objcg->memcg = memcg;
3718 rcu_assign_pointer(memcg->objcg, objcg);
3720 static_branch_enable(&memcg_kmem_enabled_key);
3722 memcg->kmemcg_id = memcg_id;
3723 memcg->kmem_state = KMEM_ONLINE;
3728 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3730 struct cgroup_subsys_state *css;
3731 struct mem_cgroup *parent, *child;
3734 if (memcg->kmem_state != KMEM_ONLINE)
3737 memcg->kmem_state = KMEM_ALLOCATED;
3739 parent = parent_mem_cgroup(memcg);
3741 parent = root_mem_cgroup;
3743 memcg_reparent_objcgs(memcg, parent);
3745 kmemcg_id = memcg->kmemcg_id;
3746 BUG_ON(kmemcg_id < 0);
3749 * Change kmemcg_id of this cgroup and all its descendants to the
3750 * parent's id, and then move all entries from this cgroup's list_lrus
3751 * to ones of the parent. After we have finished, all list_lrus
3752 * corresponding to this cgroup are guaranteed to remain empty. The
3753 * ordering is imposed by list_lru_node->lock taken by
3754 * memcg_drain_all_list_lrus().
3756 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3757 css_for_each_descendant_pre(css, &memcg->css) {
3758 child = mem_cgroup_from_css(css);
3759 BUG_ON(child->kmemcg_id != kmemcg_id);
3760 child->kmemcg_id = parent->kmemcg_id;
3764 memcg_drain_all_list_lrus(kmemcg_id, parent);
3766 memcg_free_cache_id(kmemcg_id);
3769 static void memcg_free_kmem(struct mem_cgroup *memcg)
3771 /* css_alloc() failed, offlining didn't happen */
3772 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3773 memcg_offline_kmem(memcg);
3776 static int memcg_online_kmem(struct mem_cgroup *memcg)
3780 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3783 static void memcg_free_kmem(struct mem_cgroup *memcg)
3786 #endif /* CONFIG_MEMCG_KMEM */
3788 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3793 mutex_lock(&memcg_max_mutex);
3794 ret = page_counter_set_max(&memcg->kmem, max);
3795 mutex_unlock(&memcg_max_mutex);
3799 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3803 mutex_lock(&memcg_max_mutex);
3805 ret = page_counter_set_max(&memcg->tcpmem, max);
3809 if (!memcg->tcpmem_active) {
3811 * The active flag needs to be written after the static_key
3812 * update. This is what guarantees that the socket activation
3813 * function is the last one to run. See mem_cgroup_sk_alloc()
3814 * for details, and note that we don't mark any socket as
3815 * belonging to this memcg until that flag is up.
3817 * We need to do this, because static_keys will span multiple
3818 * sites, but we can't control their order. If we mark a socket
3819 * as accounted, but the accounting functions are not patched in
3820 * yet, we'll lose accounting.
3822 * We never race with the readers in mem_cgroup_sk_alloc(),
3823 * because when this value change, the code to process it is not
3826 static_branch_inc(&memcg_sockets_enabled_key);
3827 memcg->tcpmem_active = true;
3830 mutex_unlock(&memcg_max_mutex);
3835 * The user of this function is...
3838 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3839 char *buf, size_t nbytes, loff_t off)
3841 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3842 unsigned long nr_pages;
3845 buf = strstrip(buf);
3846 ret = page_counter_memparse(buf, "-1", &nr_pages);
3850 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3852 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3856 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3858 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3861 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3864 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3865 "Please report your usecase to linux-mm@kvack.org if you "
3866 "depend on this functionality.\n");
3867 ret = memcg_update_kmem_max(memcg, nr_pages);
3870 ret = memcg_update_tcp_max(memcg, nr_pages);
3874 case RES_SOFT_LIMIT:
3875 memcg->soft_limit = nr_pages;
3879 return ret ?: nbytes;
3882 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3883 size_t nbytes, loff_t off)
3885 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3886 struct page_counter *counter;
3888 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3890 counter = &memcg->memory;
3893 counter = &memcg->memsw;
3896 counter = &memcg->kmem;
3899 counter = &memcg->tcpmem;
3905 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3907 page_counter_reset_watermark(counter);
3910 counter->failcnt = 0;
3919 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3922 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3926 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3927 struct cftype *cft, u64 val)
3929 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3931 if (val & ~MOVE_MASK)
3935 * No kind of locking is needed in here, because ->can_attach() will
3936 * check this value once in the beginning of the process, and then carry
3937 * on with stale data. This means that changes to this value will only
3938 * affect task migrations starting after the change.
3940 memcg->move_charge_at_immigrate = val;
3944 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3945 struct cftype *cft, u64 val)
3953 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3954 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3955 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3957 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3958 int nid, unsigned int lru_mask, bool tree)
3960 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3961 unsigned long nr = 0;
3964 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3967 if (!(BIT(lru) & lru_mask))
3970 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3972 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3977 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3978 unsigned int lru_mask,
3981 unsigned long nr = 0;
3985 if (!(BIT(lru) & lru_mask))
3988 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3990 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3995 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3999 unsigned int lru_mask;
4002 static const struct numa_stat stats[] = {
4003 { "total", LRU_ALL },
4004 { "file", LRU_ALL_FILE },
4005 { "anon", LRU_ALL_ANON },
4006 { "unevictable", BIT(LRU_UNEVICTABLE) },
4008 const struct numa_stat *stat;
4010 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4012 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4013 seq_printf(m, "%s=%lu", stat->name,
4014 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4016 for_each_node_state(nid, N_MEMORY)
4017 seq_printf(m, " N%d=%lu", nid,
4018 mem_cgroup_node_nr_lru_pages(memcg, nid,
4019 stat->lru_mask, false));
4023 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4025 seq_printf(m, "hierarchical_%s=%lu", stat->name,
4026 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4028 for_each_node_state(nid, N_MEMORY)
4029 seq_printf(m, " N%d=%lu", nid,
4030 mem_cgroup_node_nr_lru_pages(memcg, nid,
4031 stat->lru_mask, true));
4037 #endif /* CONFIG_NUMA */
4039 static const unsigned int memcg1_stats[] = {
4042 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4052 static const char *const memcg1_stat_names[] = {
4055 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4065 /* Universal VM events cgroup1 shows, original sort order */
4066 static const unsigned int memcg1_events[] = {
4073 static int memcg_stat_show(struct seq_file *m, void *v)
4075 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4076 unsigned long memory, memsw;
4077 struct mem_cgroup *mi;
4080 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4082 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4085 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4087 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4088 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4091 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4092 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4093 memcg_events_local(memcg, memcg1_events[i]));
4095 for (i = 0; i < NR_LRU_LISTS; i++)
4096 seq_printf(m, "%s %lu\n", lru_list_name(i),
4097 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4100 /* Hierarchical information */
4101 memory = memsw = PAGE_COUNTER_MAX;
4102 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4103 memory = min(memory, READ_ONCE(mi->memory.max));
4104 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4106 seq_printf(m, "hierarchical_memory_limit %llu\n",
4107 (u64)memory * PAGE_SIZE);
4108 if (do_memsw_account())
4109 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4110 (u64)memsw * PAGE_SIZE);
4112 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4115 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4117 nr = memcg_page_state(memcg, memcg1_stats[i]);
4118 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4119 (u64)nr * PAGE_SIZE);
4122 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4123 seq_printf(m, "total_%s %llu\n",
4124 vm_event_name(memcg1_events[i]),
4125 (u64)memcg_events(memcg, memcg1_events[i]));
4127 for (i = 0; i < NR_LRU_LISTS; i++)
4128 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4129 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4132 #ifdef CONFIG_DEBUG_VM
4135 struct mem_cgroup_per_node *mz;
4136 unsigned long anon_cost = 0;
4137 unsigned long file_cost = 0;
4139 for_each_online_pgdat(pgdat) {
4140 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
4142 anon_cost += mz->lruvec.anon_cost;
4143 file_cost += mz->lruvec.file_cost;
4145 seq_printf(m, "anon_cost %lu\n", anon_cost);
4146 seq_printf(m, "file_cost %lu\n", file_cost);
4153 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4156 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4158 return mem_cgroup_swappiness(memcg);
4161 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4162 struct cftype *cft, u64 val)
4164 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4170 memcg->swappiness = val;
4172 vm_swappiness = val;
4177 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4179 struct mem_cgroup_threshold_ary *t;
4180 unsigned long usage;
4185 t = rcu_dereference(memcg->thresholds.primary);
4187 t = rcu_dereference(memcg->memsw_thresholds.primary);
4192 usage = mem_cgroup_usage(memcg, swap);
4195 * current_threshold points to threshold just below or equal to usage.
4196 * If it's not true, a threshold was crossed after last
4197 * call of __mem_cgroup_threshold().
4199 i = t->current_threshold;
4202 * Iterate backward over array of thresholds starting from
4203 * current_threshold and check if a threshold is crossed.
4204 * If none of thresholds below usage is crossed, we read
4205 * only one element of the array here.
4207 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4208 eventfd_signal(t->entries[i].eventfd, 1);
4210 /* i = current_threshold + 1 */
4214 * Iterate forward over array of thresholds starting from
4215 * current_threshold+1 and check if a threshold is crossed.
4216 * If none of thresholds above usage is crossed, we read
4217 * only one element of the array here.
4219 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4220 eventfd_signal(t->entries[i].eventfd, 1);
4222 /* Update current_threshold */
4223 t->current_threshold = i - 1;
4228 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4231 __mem_cgroup_threshold(memcg, false);
4232 if (do_memsw_account())
4233 __mem_cgroup_threshold(memcg, true);
4235 memcg = parent_mem_cgroup(memcg);
4239 static int compare_thresholds(const void *a, const void *b)
4241 const struct mem_cgroup_threshold *_a = a;
4242 const struct mem_cgroup_threshold *_b = b;
4244 if (_a->threshold > _b->threshold)
4247 if (_a->threshold < _b->threshold)
4253 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4255 struct mem_cgroup_eventfd_list *ev;
4257 spin_lock(&memcg_oom_lock);
4259 list_for_each_entry(ev, &memcg->oom_notify, list)
4260 eventfd_signal(ev->eventfd, 1);
4262 spin_unlock(&memcg_oom_lock);
4266 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4268 struct mem_cgroup *iter;
4270 for_each_mem_cgroup_tree(iter, memcg)
4271 mem_cgroup_oom_notify_cb(iter);
4274 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4275 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4277 struct mem_cgroup_thresholds *thresholds;
4278 struct mem_cgroup_threshold_ary *new;
4279 unsigned long threshold;
4280 unsigned long usage;
4283 ret = page_counter_memparse(args, "-1", &threshold);
4287 mutex_lock(&memcg->thresholds_lock);
4290 thresholds = &memcg->thresholds;
4291 usage = mem_cgroup_usage(memcg, false);
4292 } else if (type == _MEMSWAP) {
4293 thresholds = &memcg->memsw_thresholds;
4294 usage = mem_cgroup_usage(memcg, true);
4298 /* Check if a threshold crossed before adding a new one */
4299 if (thresholds->primary)
4300 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4302 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4304 /* Allocate memory for new array of thresholds */
4305 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4312 /* Copy thresholds (if any) to new array */
4313 if (thresholds->primary)
4314 memcpy(new->entries, thresholds->primary->entries,
4315 flex_array_size(new, entries, size - 1));
4317 /* Add new threshold */
4318 new->entries[size - 1].eventfd = eventfd;
4319 new->entries[size - 1].threshold = threshold;
4321 /* Sort thresholds. Registering of new threshold isn't time-critical */
4322 sort(new->entries, size, sizeof(*new->entries),
4323 compare_thresholds, NULL);
4325 /* Find current threshold */
4326 new->current_threshold = -1;
4327 for (i = 0; i < size; i++) {
4328 if (new->entries[i].threshold <= usage) {
4330 * new->current_threshold will not be used until
4331 * rcu_assign_pointer(), so it's safe to increment
4334 ++new->current_threshold;
4339 /* Free old spare buffer and save old primary buffer as spare */
4340 kfree(thresholds->spare);
4341 thresholds->spare = thresholds->primary;
4343 rcu_assign_pointer(thresholds->primary, new);
4345 /* To be sure that nobody uses thresholds */
4349 mutex_unlock(&memcg->thresholds_lock);
4354 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4355 struct eventfd_ctx *eventfd, const char *args)
4357 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4360 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4361 struct eventfd_ctx *eventfd, const char *args)
4363 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4366 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4367 struct eventfd_ctx *eventfd, enum res_type type)
4369 struct mem_cgroup_thresholds *thresholds;
4370 struct mem_cgroup_threshold_ary *new;
4371 unsigned long usage;
4372 int i, j, size, entries;
4374 mutex_lock(&memcg->thresholds_lock);
4377 thresholds = &memcg->thresholds;
4378 usage = mem_cgroup_usage(memcg, false);
4379 } else if (type == _MEMSWAP) {
4380 thresholds = &memcg->memsw_thresholds;
4381 usage = mem_cgroup_usage(memcg, true);
4385 if (!thresholds->primary)
4388 /* Check if a threshold crossed before removing */
4389 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4391 /* Calculate new number of threshold */
4393 for (i = 0; i < thresholds->primary->size; i++) {
4394 if (thresholds->primary->entries[i].eventfd != eventfd)
4400 new = thresholds->spare;
4402 /* If no items related to eventfd have been cleared, nothing to do */
4406 /* Set thresholds array to NULL if we don't have thresholds */
4415 /* Copy thresholds and find current threshold */
4416 new->current_threshold = -1;
4417 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4418 if (thresholds->primary->entries[i].eventfd == eventfd)
4421 new->entries[j] = thresholds->primary->entries[i];
4422 if (new->entries[j].threshold <= usage) {
4424 * new->current_threshold will not be used
4425 * until rcu_assign_pointer(), so it's safe to increment
4428 ++new->current_threshold;
4434 /* Swap primary and spare array */
4435 thresholds->spare = thresholds->primary;
4437 rcu_assign_pointer(thresholds->primary, new);
4439 /* To be sure that nobody uses thresholds */
4442 /* If all events are unregistered, free the spare array */
4444 kfree(thresholds->spare);
4445 thresholds->spare = NULL;
4448 mutex_unlock(&memcg->thresholds_lock);
4451 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4452 struct eventfd_ctx *eventfd)
4454 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4457 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4458 struct eventfd_ctx *eventfd)
4460 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4463 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4464 struct eventfd_ctx *eventfd, const char *args)
4466 struct mem_cgroup_eventfd_list *event;
4468 event = kmalloc(sizeof(*event), GFP_KERNEL);
4472 spin_lock(&memcg_oom_lock);
4474 event->eventfd = eventfd;
4475 list_add(&event->list, &memcg->oom_notify);
4477 /* already in OOM ? */
4478 if (memcg->under_oom)
4479 eventfd_signal(eventfd, 1);
4480 spin_unlock(&memcg_oom_lock);
4485 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4486 struct eventfd_ctx *eventfd)
4488 struct mem_cgroup_eventfd_list *ev, *tmp;
4490 spin_lock(&memcg_oom_lock);
4492 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4493 if (ev->eventfd == eventfd) {
4494 list_del(&ev->list);
4499 spin_unlock(&memcg_oom_lock);
4502 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4504 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4506 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4507 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4508 seq_printf(sf, "oom_kill %lu\n",
4509 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4513 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4514 struct cftype *cft, u64 val)
4516 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4518 /* cannot set to root cgroup and only 0 and 1 are allowed */
4519 if (!css->parent || !((val == 0) || (val == 1)))
4522 memcg->oom_kill_disable = val;
4524 memcg_oom_recover(memcg);
4529 #ifdef CONFIG_CGROUP_WRITEBACK
4531 #include <trace/events/writeback.h>
4533 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4535 return wb_domain_init(&memcg->cgwb_domain, gfp);
4538 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4540 wb_domain_exit(&memcg->cgwb_domain);
4543 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4545 wb_domain_size_changed(&memcg->cgwb_domain);
4548 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4550 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4552 if (!memcg->css.parent)
4555 return &memcg->cgwb_domain;
4559 * idx can be of type enum memcg_stat_item or node_stat_item.
4560 * Keep in sync with memcg_exact_page().
4562 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4564 long x = atomic_long_read(&memcg->vmstats[idx]);
4567 for_each_online_cpu(cpu)
4568 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4575 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4576 * @wb: bdi_writeback in question
4577 * @pfilepages: out parameter for number of file pages
4578 * @pheadroom: out parameter for number of allocatable pages according to memcg
4579 * @pdirty: out parameter for number of dirty pages
4580 * @pwriteback: out parameter for number of pages under writeback
4582 * Determine the numbers of file, headroom, dirty, and writeback pages in
4583 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4584 * is a bit more involved.
4586 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4587 * headroom is calculated as the lowest headroom of itself and the
4588 * ancestors. Note that this doesn't consider the actual amount of
4589 * available memory in the system. The caller should further cap
4590 * *@pheadroom accordingly.
4592 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4593 unsigned long *pheadroom, unsigned long *pdirty,
4594 unsigned long *pwriteback)
4596 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4597 struct mem_cgroup *parent;
4599 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4601 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4602 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4603 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4604 *pheadroom = PAGE_COUNTER_MAX;
4606 while ((parent = parent_mem_cgroup(memcg))) {
4607 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4608 READ_ONCE(memcg->memory.high));
4609 unsigned long used = page_counter_read(&memcg->memory);
4611 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4617 * Foreign dirty flushing
4619 * There's an inherent mismatch between memcg and writeback. The former
4620 * trackes ownership per-page while the latter per-inode. This was a
4621 * deliberate design decision because honoring per-page ownership in the
4622 * writeback path is complicated, may lead to higher CPU and IO overheads
4623 * and deemed unnecessary given that write-sharing an inode across
4624 * different cgroups isn't a common use-case.
4626 * Combined with inode majority-writer ownership switching, this works well
4627 * enough in most cases but there are some pathological cases. For
4628 * example, let's say there are two cgroups A and B which keep writing to
4629 * different but confined parts of the same inode. B owns the inode and
4630 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4631 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4632 * triggering background writeback. A will be slowed down without a way to
4633 * make writeback of the dirty pages happen.
4635 * Conditions like the above can lead to a cgroup getting repatedly and
4636 * severely throttled after making some progress after each
4637 * dirty_expire_interval while the underyling IO device is almost
4640 * Solving this problem completely requires matching the ownership tracking
4641 * granularities between memcg and writeback in either direction. However,
4642 * the more egregious behaviors can be avoided by simply remembering the
4643 * most recent foreign dirtying events and initiating remote flushes on
4644 * them when local writeback isn't enough to keep the memory clean enough.
4646 * The following two functions implement such mechanism. When a foreign
4647 * page - a page whose memcg and writeback ownerships don't match - is
4648 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4649 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4650 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4651 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4652 * foreign bdi_writebacks which haven't expired. Both the numbers of
4653 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4654 * limited to MEMCG_CGWB_FRN_CNT.
4656 * The mechanism only remembers IDs and doesn't hold any object references.
4657 * As being wrong occasionally doesn't matter, updates and accesses to the
4658 * records are lockless and racy.
4660 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4661 struct bdi_writeback *wb)
4663 struct mem_cgroup *memcg = page_memcg(page);
4664 struct memcg_cgwb_frn *frn;
4665 u64 now = get_jiffies_64();
4666 u64 oldest_at = now;
4670 trace_track_foreign_dirty(page, wb);
4673 * Pick the slot to use. If there is already a slot for @wb, keep
4674 * using it. If not replace the oldest one which isn't being
4677 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4678 frn = &memcg->cgwb_frn[i];
4679 if (frn->bdi_id == wb->bdi->id &&
4680 frn->memcg_id == wb->memcg_css->id)
4682 if (time_before64(frn->at, oldest_at) &&
4683 atomic_read(&frn->done.cnt) == 1) {
4685 oldest_at = frn->at;
4689 if (i < MEMCG_CGWB_FRN_CNT) {
4691 * Re-using an existing one. Update timestamp lazily to
4692 * avoid making the cacheline hot. We want them to be
4693 * reasonably up-to-date and significantly shorter than
4694 * dirty_expire_interval as that's what expires the record.
4695 * Use the shorter of 1s and dirty_expire_interval / 8.
4697 unsigned long update_intv =
4698 min_t(unsigned long, HZ,
4699 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4701 if (time_before64(frn->at, now - update_intv))
4703 } else if (oldest >= 0) {
4704 /* replace the oldest free one */
4705 frn = &memcg->cgwb_frn[oldest];
4706 frn->bdi_id = wb->bdi->id;
4707 frn->memcg_id = wb->memcg_css->id;
4712 /* issue foreign writeback flushes for recorded foreign dirtying events */
4713 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4715 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4716 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4717 u64 now = jiffies_64;
4720 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4721 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4724 * If the record is older than dirty_expire_interval,
4725 * writeback on it has already started. No need to kick it
4726 * off again. Also, don't start a new one if there's
4727 * already one in flight.
4729 if (time_after64(frn->at, now - intv) &&
4730 atomic_read(&frn->done.cnt) == 1) {
4732 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4733 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4734 WB_REASON_FOREIGN_FLUSH,
4740 #else /* CONFIG_CGROUP_WRITEBACK */
4742 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4747 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4751 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4755 #endif /* CONFIG_CGROUP_WRITEBACK */
4758 * DO NOT USE IN NEW FILES.
4760 * "cgroup.event_control" implementation.
4762 * This is way over-engineered. It tries to support fully configurable
4763 * events for each user. Such level of flexibility is completely
4764 * unnecessary especially in the light of the planned unified hierarchy.
4766 * Please deprecate this and replace with something simpler if at all
4771 * Unregister event and free resources.
4773 * Gets called from workqueue.
4775 static void memcg_event_remove(struct work_struct *work)
4777 struct mem_cgroup_event *event =
4778 container_of(work, struct mem_cgroup_event, remove);
4779 struct mem_cgroup *memcg = event->memcg;
4781 remove_wait_queue(event->wqh, &event->wait);
4783 event->unregister_event(memcg, event->eventfd);
4785 /* Notify userspace the event is going away. */
4786 eventfd_signal(event->eventfd, 1);
4788 eventfd_ctx_put(event->eventfd);
4790 css_put(&memcg->css);
4794 * Gets called on EPOLLHUP on eventfd when user closes it.
4796 * Called with wqh->lock held and interrupts disabled.
4798 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4799 int sync, void *key)
4801 struct mem_cgroup_event *event =
4802 container_of(wait, struct mem_cgroup_event, wait);
4803 struct mem_cgroup *memcg = event->memcg;
4804 __poll_t flags = key_to_poll(key);
4806 if (flags & EPOLLHUP) {
4808 * If the event has been detached at cgroup removal, we
4809 * can simply return knowing the other side will cleanup
4812 * We can't race against event freeing since the other
4813 * side will require wqh->lock via remove_wait_queue(),
4816 spin_lock(&memcg->event_list_lock);
4817 if (!list_empty(&event->list)) {
4818 list_del_init(&event->list);
4820 * We are in atomic context, but cgroup_event_remove()
4821 * may sleep, so we have to call it in workqueue.
4823 schedule_work(&event->remove);
4825 spin_unlock(&memcg->event_list_lock);
4831 static void memcg_event_ptable_queue_proc(struct file *file,
4832 wait_queue_head_t *wqh, poll_table *pt)
4834 struct mem_cgroup_event *event =
4835 container_of(pt, struct mem_cgroup_event, pt);
4838 add_wait_queue(wqh, &event->wait);
4842 * DO NOT USE IN NEW FILES.
4844 * Parse input and register new cgroup event handler.
4846 * Input must be in format '<event_fd> <control_fd> <args>'.
4847 * Interpretation of args is defined by control file implementation.
4849 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4850 char *buf, size_t nbytes, loff_t off)
4852 struct cgroup_subsys_state *css = of_css(of);
4853 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4854 struct mem_cgroup_event *event;
4855 struct cgroup_subsys_state *cfile_css;
4856 unsigned int efd, cfd;
4863 buf = strstrip(buf);
4865 efd = simple_strtoul(buf, &endp, 10);
4870 cfd = simple_strtoul(buf, &endp, 10);
4871 if ((*endp != ' ') && (*endp != '\0'))
4875 event = kzalloc(sizeof(*event), GFP_KERNEL);
4879 event->memcg = memcg;
4880 INIT_LIST_HEAD(&event->list);
4881 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4882 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4883 INIT_WORK(&event->remove, memcg_event_remove);
4891 event->eventfd = eventfd_ctx_fileget(efile.file);
4892 if (IS_ERR(event->eventfd)) {
4893 ret = PTR_ERR(event->eventfd);
4900 goto out_put_eventfd;
4903 /* the process need read permission on control file */
4904 /* AV: shouldn't we check that it's been opened for read instead? */
4905 ret = file_permission(cfile.file, MAY_READ);
4910 * Determine the event callbacks and set them in @event. This used
4911 * to be done via struct cftype but cgroup core no longer knows
4912 * about these events. The following is crude but the whole thing
4913 * is for compatibility anyway.
4915 * DO NOT ADD NEW FILES.
4917 name = cfile.file->f_path.dentry->d_name.name;
4919 if (!strcmp(name, "memory.usage_in_bytes")) {
4920 event->register_event = mem_cgroup_usage_register_event;
4921 event->unregister_event = mem_cgroup_usage_unregister_event;
4922 } else if (!strcmp(name, "memory.oom_control")) {
4923 event->register_event = mem_cgroup_oom_register_event;
4924 event->unregister_event = mem_cgroup_oom_unregister_event;
4925 } else if (!strcmp(name, "memory.pressure_level")) {
4926 event->register_event = vmpressure_register_event;
4927 event->unregister_event = vmpressure_unregister_event;
4928 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4929 event->register_event = memsw_cgroup_usage_register_event;
4930 event->unregister_event = memsw_cgroup_usage_unregister_event;
4937 * Verify @cfile should belong to @css. Also, remaining events are
4938 * automatically removed on cgroup destruction but the removal is
4939 * asynchronous, so take an extra ref on @css.
4941 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4942 &memory_cgrp_subsys);
4944 if (IS_ERR(cfile_css))
4946 if (cfile_css != css) {
4951 ret = event->register_event(memcg, event->eventfd, buf);
4955 vfs_poll(efile.file, &event->pt);
4957 spin_lock(&memcg->event_list_lock);
4958 list_add(&event->list, &memcg->event_list);
4959 spin_unlock(&memcg->event_list_lock);
4971 eventfd_ctx_put(event->eventfd);
4980 static struct cftype mem_cgroup_legacy_files[] = {
4982 .name = "usage_in_bytes",
4983 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4984 .read_u64 = mem_cgroup_read_u64,
4987 .name = "max_usage_in_bytes",
4988 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4989 .write = mem_cgroup_reset,
4990 .read_u64 = mem_cgroup_read_u64,
4993 .name = "limit_in_bytes",
4994 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4995 .write = mem_cgroup_write,
4996 .read_u64 = mem_cgroup_read_u64,
4999 .name = "soft_limit_in_bytes",
5000 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5001 .write = mem_cgroup_write,
5002 .read_u64 = mem_cgroup_read_u64,
5006 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5007 .write = mem_cgroup_reset,
5008 .read_u64 = mem_cgroup_read_u64,
5012 .seq_show = memcg_stat_show,
5015 .name = "force_empty",
5016 .write = mem_cgroup_force_empty_write,
5019 .name = "use_hierarchy",
5020 .write_u64 = mem_cgroup_hierarchy_write,
5021 .read_u64 = mem_cgroup_hierarchy_read,
5024 .name = "cgroup.event_control", /* XXX: for compat */
5025 .write = memcg_write_event_control,
5026 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
5029 .name = "swappiness",
5030 .read_u64 = mem_cgroup_swappiness_read,
5031 .write_u64 = mem_cgroup_swappiness_write,
5034 .name = "move_charge_at_immigrate",
5035 .read_u64 = mem_cgroup_move_charge_read,
5036 .write_u64 = mem_cgroup_move_charge_write,
5039 .name = "oom_control",
5040 .seq_show = mem_cgroup_oom_control_read,
5041 .write_u64 = mem_cgroup_oom_control_write,
5042 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5045 .name = "pressure_level",
5049 .name = "numa_stat",
5050 .seq_show = memcg_numa_stat_show,
5054 .name = "kmem.limit_in_bytes",
5055 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5056 .write = mem_cgroup_write,
5057 .read_u64 = mem_cgroup_read_u64,
5060 .name = "kmem.usage_in_bytes",
5061 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5062 .read_u64 = mem_cgroup_read_u64,
5065 .name = "kmem.failcnt",
5066 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5067 .write = mem_cgroup_reset,
5068 .read_u64 = mem_cgroup_read_u64,
5071 .name = "kmem.max_usage_in_bytes",
5072 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5073 .write = mem_cgroup_reset,
5074 .read_u64 = mem_cgroup_read_u64,
5076 #if defined(CONFIG_MEMCG_KMEM) && \
5077 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5079 .name = "kmem.slabinfo",
5080 .seq_show = memcg_slab_show,
5084 .name = "kmem.tcp.limit_in_bytes",
5085 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5086 .write = mem_cgroup_write,
5087 .read_u64 = mem_cgroup_read_u64,
5090 .name = "kmem.tcp.usage_in_bytes",
5091 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5092 .read_u64 = mem_cgroup_read_u64,
5095 .name = "kmem.tcp.failcnt",
5096 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5097 .write = mem_cgroup_reset,
5098 .read_u64 = mem_cgroup_read_u64,
5101 .name = "kmem.tcp.max_usage_in_bytes",
5102 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5103 .write = mem_cgroup_reset,
5104 .read_u64 = mem_cgroup_read_u64,
5106 { }, /* terminate */
5110 * Private memory cgroup IDR
5112 * Swap-out records and page cache shadow entries need to store memcg
5113 * references in constrained space, so we maintain an ID space that is
5114 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5115 * memory-controlled cgroups to 64k.
5117 * However, there usually are many references to the offline CSS after
5118 * the cgroup has been destroyed, such as page cache or reclaimable
5119 * slab objects, that don't need to hang on to the ID. We want to keep
5120 * those dead CSS from occupying IDs, or we might quickly exhaust the
5121 * relatively small ID space and prevent the creation of new cgroups
5122 * even when there are much fewer than 64k cgroups - possibly none.
5124 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5125 * be freed and recycled when it's no longer needed, which is usually
5126 * when the CSS is offlined.
5128 * The only exception to that are records of swapped out tmpfs/shmem
5129 * pages that need to be attributed to live ancestors on swapin. But
5130 * those references are manageable from userspace.
5133 static DEFINE_IDR(mem_cgroup_idr);
5135 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5137 if (memcg->id.id > 0) {
5138 idr_remove(&mem_cgroup_idr, memcg->id.id);
5143 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5146 refcount_add(n, &memcg->id.ref);
5149 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5151 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5152 mem_cgroup_id_remove(memcg);
5154 /* Memcg ID pins CSS */
5155 css_put(&memcg->css);
5159 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5161 mem_cgroup_id_put_many(memcg, 1);
5165 * mem_cgroup_from_id - look up a memcg from a memcg id
5166 * @id: the memcg id to look up
5168 * Caller must hold rcu_read_lock().
5170 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5172 WARN_ON_ONCE(!rcu_read_lock_held());
5173 return idr_find(&mem_cgroup_idr, id);
5176 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5178 struct mem_cgroup_per_node *pn;
5181 * This routine is called against possible nodes.
5182 * But it's BUG to call kmalloc() against offline node.
5184 * TODO: this routine can waste much memory for nodes which will
5185 * never be onlined. It's better to use memory hotplug callback
5188 if (!node_state(node, N_NORMAL_MEMORY))
5190 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5194 pn->lruvec_stat_local = alloc_percpu_gfp(struct lruvec_stat,
5195 GFP_KERNEL_ACCOUNT);
5196 if (!pn->lruvec_stat_local) {
5201 pn->lruvec_stat_cpu = alloc_percpu_gfp(struct batched_lruvec_stat,
5202 GFP_KERNEL_ACCOUNT);
5203 if (!pn->lruvec_stat_cpu) {
5204 free_percpu(pn->lruvec_stat_local);
5209 lruvec_init(&pn->lruvec);
5210 pn->usage_in_excess = 0;
5211 pn->on_tree = false;
5214 memcg->nodeinfo[node] = pn;
5218 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5220 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5225 free_percpu(pn->lruvec_stat_cpu);
5226 free_percpu(pn->lruvec_stat_local);
5230 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5235 free_mem_cgroup_per_node_info(memcg, node);
5236 free_percpu(memcg->vmstats_percpu);
5237 free_percpu(memcg->vmstats_local);
5241 static void mem_cgroup_free(struct mem_cgroup *memcg)
5243 memcg_wb_domain_exit(memcg);
5245 * Flush percpu vmstats and vmevents to guarantee the value correctness
5246 * on parent's and all ancestor levels.
5248 memcg_flush_percpu_vmstats(memcg);
5249 memcg_flush_percpu_vmevents(memcg);
5250 __mem_cgroup_free(memcg);
5253 static struct mem_cgroup *mem_cgroup_alloc(void)
5255 struct mem_cgroup *memcg;
5258 int __maybe_unused i;
5259 long error = -ENOMEM;
5261 size = sizeof(struct mem_cgroup);
5262 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5264 memcg = kzalloc(size, GFP_KERNEL);
5266 return ERR_PTR(error);
5268 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5269 1, MEM_CGROUP_ID_MAX,
5271 if (memcg->id.id < 0) {
5272 error = memcg->id.id;
5276 memcg->vmstats_local = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5277 GFP_KERNEL_ACCOUNT);
5278 if (!memcg->vmstats_local)
5281 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5282 GFP_KERNEL_ACCOUNT);
5283 if (!memcg->vmstats_percpu)
5287 if (alloc_mem_cgroup_per_node_info(memcg, node))
5290 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5293 INIT_WORK(&memcg->high_work, high_work_func);
5294 INIT_LIST_HEAD(&memcg->oom_notify);
5295 mutex_init(&memcg->thresholds_lock);
5296 spin_lock_init(&memcg->move_lock);
5297 vmpressure_init(&memcg->vmpressure);
5298 INIT_LIST_HEAD(&memcg->event_list);
5299 spin_lock_init(&memcg->event_list_lock);
5300 memcg->socket_pressure = jiffies;
5301 #ifdef CONFIG_MEMCG_KMEM
5302 memcg->kmemcg_id = -1;
5303 INIT_LIST_HEAD(&memcg->objcg_list);
5305 #ifdef CONFIG_CGROUP_WRITEBACK
5306 INIT_LIST_HEAD(&memcg->cgwb_list);
5307 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5308 memcg->cgwb_frn[i].done =
5309 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5311 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5312 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5313 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5314 memcg->deferred_split_queue.split_queue_len = 0;
5316 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5319 mem_cgroup_id_remove(memcg);
5320 __mem_cgroup_free(memcg);
5321 return ERR_PTR(error);
5324 static struct cgroup_subsys_state * __ref
5325 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5327 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5328 struct mem_cgroup *memcg, *old_memcg;
5329 long error = -ENOMEM;
5331 old_memcg = set_active_memcg(parent);
5332 memcg = mem_cgroup_alloc();
5333 set_active_memcg(old_memcg);
5335 return ERR_CAST(memcg);
5337 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5338 memcg->soft_limit = PAGE_COUNTER_MAX;
5339 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5341 memcg->swappiness = mem_cgroup_swappiness(parent);
5342 memcg->oom_kill_disable = parent->oom_kill_disable;
5344 page_counter_init(&memcg->memory, &parent->memory);
5345 page_counter_init(&memcg->swap, &parent->swap);
5346 page_counter_init(&memcg->kmem, &parent->kmem);
5347 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5349 page_counter_init(&memcg->memory, NULL);
5350 page_counter_init(&memcg->swap, NULL);
5351 page_counter_init(&memcg->kmem, NULL);
5352 page_counter_init(&memcg->tcpmem, NULL);
5354 root_mem_cgroup = memcg;
5358 /* The following stuff does not apply to the root */
5359 error = memcg_online_kmem(memcg);
5363 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5364 static_branch_inc(&memcg_sockets_enabled_key);
5368 mem_cgroup_id_remove(memcg);
5369 mem_cgroup_free(memcg);
5370 return ERR_PTR(error);
5373 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5375 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5378 * A memcg must be visible for memcg_expand_shrinker_maps()
5379 * by the time the maps are allocated. So, we allocate maps
5380 * here, when for_each_mem_cgroup() can't skip it.
5382 if (memcg_alloc_shrinker_maps(memcg)) {
5383 mem_cgroup_id_remove(memcg);
5387 /* Online state pins memcg ID, memcg ID pins CSS */
5388 refcount_set(&memcg->id.ref, 1);
5393 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5395 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5396 struct mem_cgroup_event *event, *tmp;
5399 * Unregister events and notify userspace.
5400 * Notify userspace about cgroup removing only after rmdir of cgroup
5401 * directory to avoid race between userspace and kernelspace.
5403 spin_lock(&memcg->event_list_lock);
5404 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5405 list_del_init(&event->list);
5406 schedule_work(&event->remove);
5408 spin_unlock(&memcg->event_list_lock);
5410 page_counter_set_min(&memcg->memory, 0);
5411 page_counter_set_low(&memcg->memory, 0);
5413 memcg_offline_kmem(memcg);
5414 wb_memcg_offline(memcg);
5416 drain_all_stock(memcg);
5418 mem_cgroup_id_put(memcg);
5421 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5423 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5425 invalidate_reclaim_iterators(memcg);
5428 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5430 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5431 int __maybe_unused i;
5433 #ifdef CONFIG_CGROUP_WRITEBACK
5434 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5435 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5437 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5438 static_branch_dec(&memcg_sockets_enabled_key);
5440 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5441 static_branch_dec(&memcg_sockets_enabled_key);
5443 vmpressure_cleanup(&memcg->vmpressure);
5444 cancel_work_sync(&memcg->high_work);
5445 mem_cgroup_remove_from_trees(memcg);
5446 memcg_free_shrinker_maps(memcg);
5447 memcg_free_kmem(memcg);
5448 mem_cgroup_free(memcg);
5452 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5453 * @css: the target css
5455 * Reset the states of the mem_cgroup associated with @css. This is
5456 * invoked when the userland requests disabling on the default hierarchy
5457 * but the memcg is pinned through dependency. The memcg should stop
5458 * applying policies and should revert to the vanilla state as it may be
5459 * made visible again.
5461 * The current implementation only resets the essential configurations.
5462 * This needs to be expanded to cover all the visible parts.
5464 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5466 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5468 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5469 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5470 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5471 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5472 page_counter_set_min(&memcg->memory, 0);
5473 page_counter_set_low(&memcg->memory, 0);
5474 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5475 memcg->soft_limit = PAGE_COUNTER_MAX;
5476 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5477 memcg_wb_domain_size_changed(memcg);
5481 /* Handlers for move charge at task migration. */
5482 static int mem_cgroup_do_precharge(unsigned long count)
5486 /* Try a single bulk charge without reclaim first, kswapd may wake */
5487 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5489 mc.precharge += count;
5493 /* Try charges one by one with reclaim, but do not retry */
5495 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5509 enum mc_target_type {
5516 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5517 unsigned long addr, pte_t ptent)
5519 struct page *page = vm_normal_page(vma, addr, ptent);
5521 if (!page || !page_mapped(page))
5523 if (PageAnon(page)) {
5524 if (!(mc.flags & MOVE_ANON))
5527 if (!(mc.flags & MOVE_FILE))
5530 if (!get_page_unless_zero(page))
5536 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5537 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5538 pte_t ptent, swp_entry_t *entry)
5540 struct page *page = NULL;
5541 swp_entry_t ent = pte_to_swp_entry(ptent);
5543 if (!(mc.flags & MOVE_ANON))
5547 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5548 * a device and because they are not accessible by CPU they are store
5549 * as special swap entry in the CPU page table.
5551 if (is_device_private_entry(ent)) {
5552 page = device_private_entry_to_page(ent);
5554 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5555 * a refcount of 1 when free (unlike normal page)
5557 if (!page_ref_add_unless(page, 1, 1))
5562 if (non_swap_entry(ent))
5566 * Because lookup_swap_cache() updates some statistics counter,
5567 * we call find_get_page() with swapper_space directly.
5569 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5570 entry->val = ent.val;
5575 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5576 pte_t ptent, swp_entry_t *entry)
5582 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5583 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5585 if (!vma->vm_file) /* anonymous vma */
5587 if (!(mc.flags & MOVE_FILE))
5590 /* page is moved even if it's not RSS of this task(page-faulted). */
5591 /* shmem/tmpfs may report page out on swap: account for that too. */
5592 return find_get_incore_page(vma->vm_file->f_mapping,
5593 linear_page_index(vma, addr));
5597 * mem_cgroup_move_account - move account of the page
5599 * @compound: charge the page as compound or small page
5600 * @from: mem_cgroup which the page is moved from.
5601 * @to: mem_cgroup which the page is moved to. @from != @to.
5603 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5605 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5608 static int mem_cgroup_move_account(struct page *page,
5610 struct mem_cgroup *from,
5611 struct mem_cgroup *to)
5613 struct lruvec *from_vec, *to_vec;
5614 struct pglist_data *pgdat;
5615 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5618 VM_BUG_ON(from == to);
5619 VM_BUG_ON_PAGE(PageLRU(page), page);
5620 VM_BUG_ON(compound && !PageTransHuge(page));
5623 * Prevent mem_cgroup_migrate() from looking at
5624 * page's memory cgroup of its source page while we change it.
5627 if (!trylock_page(page))
5631 if (page_memcg(page) != from)
5634 pgdat = page_pgdat(page);
5635 from_vec = mem_cgroup_lruvec(from, pgdat);
5636 to_vec = mem_cgroup_lruvec(to, pgdat);
5638 lock_page_memcg(page);
5640 if (PageAnon(page)) {
5641 if (page_mapped(page)) {
5642 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5643 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5644 if (PageTransHuge(page)) {
5645 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5647 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5652 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5653 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5655 if (PageSwapBacked(page)) {
5656 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5657 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5660 if (page_mapped(page)) {
5661 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5662 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5665 if (PageDirty(page)) {
5666 struct address_space *mapping = page_mapping(page);
5668 if (mapping_can_writeback(mapping)) {
5669 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5671 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5677 if (PageWriteback(page)) {
5678 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5679 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5683 * All state has been migrated, let's switch to the new memcg.
5685 * It is safe to change page's memcg here because the page
5686 * is referenced, charged, isolated, and locked: we can't race
5687 * with (un)charging, migration, LRU putback, or anything else
5688 * that would rely on a stable page's memory cgroup.
5690 * Note that lock_page_memcg is a memcg lock, not a page lock,
5691 * to save space. As soon as we switch page's memory cgroup to a
5692 * new memcg that isn't locked, the above state can change
5693 * concurrently again. Make sure we're truly done with it.
5698 css_put(&from->css);
5700 page->memcg_data = (unsigned long)to;
5702 __unlock_page_memcg(from);
5706 local_irq_disable();
5707 mem_cgroup_charge_statistics(to, page, nr_pages);
5708 memcg_check_events(to, page);
5709 mem_cgroup_charge_statistics(from, page, -nr_pages);
5710 memcg_check_events(from, page);
5719 * get_mctgt_type - get target type of moving charge
5720 * @vma: the vma the pte to be checked belongs
5721 * @addr: the address corresponding to the pte to be checked
5722 * @ptent: the pte to be checked
5723 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5726 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5727 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5728 * move charge. if @target is not NULL, the page is stored in target->page
5729 * with extra refcnt got(Callers should handle it).
5730 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5731 * target for charge migration. if @target is not NULL, the entry is stored
5733 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5734 * (so ZONE_DEVICE page and thus not on the lru).
5735 * For now we such page is charge like a regular page would be as for all
5736 * intent and purposes it is just special memory taking the place of a
5739 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5741 * Called with pte lock held.
5744 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5745 unsigned long addr, pte_t ptent, union mc_target *target)
5747 struct page *page = NULL;
5748 enum mc_target_type ret = MC_TARGET_NONE;
5749 swp_entry_t ent = { .val = 0 };
5751 if (pte_present(ptent))
5752 page = mc_handle_present_pte(vma, addr, ptent);
5753 else if (is_swap_pte(ptent))
5754 page = mc_handle_swap_pte(vma, ptent, &ent);
5755 else if (pte_none(ptent))
5756 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5758 if (!page && !ent.val)
5762 * Do only loose check w/o serialization.
5763 * mem_cgroup_move_account() checks the page is valid or
5764 * not under LRU exclusion.
5766 if (page_memcg(page) == mc.from) {
5767 ret = MC_TARGET_PAGE;
5768 if (is_device_private_page(page))
5769 ret = MC_TARGET_DEVICE;
5771 target->page = page;
5773 if (!ret || !target)
5777 * There is a swap entry and a page doesn't exist or isn't charged.
5778 * But we cannot move a tail-page in a THP.
5780 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5781 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5782 ret = MC_TARGET_SWAP;
5789 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5791 * We don't consider PMD mapped swapping or file mapped pages because THP does
5792 * not support them for now.
5793 * Caller should make sure that pmd_trans_huge(pmd) is true.
5795 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5796 unsigned long addr, pmd_t pmd, union mc_target *target)
5798 struct page *page = NULL;
5799 enum mc_target_type ret = MC_TARGET_NONE;
5801 if (unlikely(is_swap_pmd(pmd))) {
5802 VM_BUG_ON(thp_migration_supported() &&
5803 !is_pmd_migration_entry(pmd));
5806 page = pmd_page(pmd);
5807 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5808 if (!(mc.flags & MOVE_ANON))
5810 if (page_memcg(page) == mc.from) {
5811 ret = MC_TARGET_PAGE;
5814 target->page = page;
5820 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5821 unsigned long addr, pmd_t pmd, union mc_target *target)
5823 return MC_TARGET_NONE;
5827 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5828 unsigned long addr, unsigned long end,
5829 struct mm_walk *walk)
5831 struct vm_area_struct *vma = walk->vma;
5835 ptl = pmd_trans_huge_lock(pmd, vma);
5838 * Note their can not be MC_TARGET_DEVICE for now as we do not
5839 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5840 * this might change.
5842 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5843 mc.precharge += HPAGE_PMD_NR;
5848 if (pmd_trans_unstable(pmd))
5850 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5851 for (; addr != end; pte++, addr += PAGE_SIZE)
5852 if (get_mctgt_type(vma, addr, *pte, NULL))
5853 mc.precharge++; /* increment precharge temporarily */
5854 pte_unmap_unlock(pte - 1, ptl);
5860 static const struct mm_walk_ops precharge_walk_ops = {
5861 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5864 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5866 unsigned long precharge;
5869 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5870 mmap_read_unlock(mm);
5872 precharge = mc.precharge;
5878 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5880 unsigned long precharge = mem_cgroup_count_precharge(mm);
5882 VM_BUG_ON(mc.moving_task);
5883 mc.moving_task = current;
5884 return mem_cgroup_do_precharge(precharge);
5887 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5888 static void __mem_cgroup_clear_mc(void)
5890 struct mem_cgroup *from = mc.from;
5891 struct mem_cgroup *to = mc.to;
5893 /* we must uncharge all the leftover precharges from mc.to */
5895 cancel_charge(mc.to, mc.precharge);
5899 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5900 * we must uncharge here.
5902 if (mc.moved_charge) {
5903 cancel_charge(mc.from, mc.moved_charge);
5904 mc.moved_charge = 0;
5906 /* we must fixup refcnts and charges */
5907 if (mc.moved_swap) {
5908 /* uncharge swap account from the old cgroup */
5909 if (!mem_cgroup_is_root(mc.from))
5910 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5912 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5915 * we charged both to->memory and to->memsw, so we
5916 * should uncharge to->memory.
5918 if (!mem_cgroup_is_root(mc.to))
5919 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5923 memcg_oom_recover(from);
5924 memcg_oom_recover(to);
5925 wake_up_all(&mc.waitq);
5928 static void mem_cgroup_clear_mc(void)
5930 struct mm_struct *mm = mc.mm;
5933 * we must clear moving_task before waking up waiters at the end of
5936 mc.moving_task = NULL;
5937 __mem_cgroup_clear_mc();
5938 spin_lock(&mc.lock);
5942 spin_unlock(&mc.lock);
5947 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5949 struct cgroup_subsys_state *css;
5950 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5951 struct mem_cgroup *from;
5952 struct task_struct *leader, *p;
5953 struct mm_struct *mm;
5954 unsigned long move_flags;
5957 /* charge immigration isn't supported on the default hierarchy */
5958 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5962 * Multi-process migrations only happen on the default hierarchy
5963 * where charge immigration is not used. Perform charge
5964 * immigration if @tset contains a leader and whine if there are
5968 cgroup_taskset_for_each_leader(leader, css, tset) {
5971 memcg = mem_cgroup_from_css(css);
5977 * We are now commited to this value whatever it is. Changes in this
5978 * tunable will only affect upcoming migrations, not the current one.
5979 * So we need to save it, and keep it going.
5981 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5985 from = mem_cgroup_from_task(p);
5987 VM_BUG_ON(from == memcg);
5989 mm = get_task_mm(p);
5992 /* We move charges only when we move a owner of the mm */
5993 if (mm->owner == p) {
5996 VM_BUG_ON(mc.precharge);
5997 VM_BUG_ON(mc.moved_charge);
5998 VM_BUG_ON(mc.moved_swap);
6000 spin_lock(&mc.lock);
6004 mc.flags = move_flags;
6005 spin_unlock(&mc.lock);
6006 /* We set mc.moving_task later */
6008 ret = mem_cgroup_precharge_mc(mm);
6010 mem_cgroup_clear_mc();
6017 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6020 mem_cgroup_clear_mc();
6023 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6024 unsigned long addr, unsigned long end,
6025 struct mm_walk *walk)
6028 struct vm_area_struct *vma = walk->vma;
6031 enum mc_target_type target_type;
6032 union mc_target target;
6035 ptl = pmd_trans_huge_lock(pmd, vma);
6037 if (mc.precharge < HPAGE_PMD_NR) {
6041 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6042 if (target_type == MC_TARGET_PAGE) {
6044 if (!isolate_lru_page(page)) {
6045 if (!mem_cgroup_move_account(page, true,
6047 mc.precharge -= HPAGE_PMD_NR;
6048 mc.moved_charge += HPAGE_PMD_NR;
6050 putback_lru_page(page);
6053 } else if (target_type == MC_TARGET_DEVICE) {
6055 if (!mem_cgroup_move_account(page, true,
6057 mc.precharge -= HPAGE_PMD_NR;
6058 mc.moved_charge += HPAGE_PMD_NR;
6066 if (pmd_trans_unstable(pmd))
6069 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6070 for (; addr != end; addr += PAGE_SIZE) {
6071 pte_t ptent = *(pte++);
6072 bool device = false;
6078 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6079 case MC_TARGET_DEVICE:
6082 case MC_TARGET_PAGE:
6085 * We can have a part of the split pmd here. Moving it
6086 * can be done but it would be too convoluted so simply
6087 * ignore such a partial THP and keep it in original
6088 * memcg. There should be somebody mapping the head.
6090 if (PageTransCompound(page))
6092 if (!device && isolate_lru_page(page))
6094 if (!mem_cgroup_move_account(page, false,
6097 /* we uncharge from mc.from later. */
6101 putback_lru_page(page);
6102 put: /* get_mctgt_type() gets the page */
6105 case MC_TARGET_SWAP:
6107 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6109 mem_cgroup_id_get_many(mc.to, 1);
6110 /* we fixup other refcnts and charges later. */
6118 pte_unmap_unlock(pte - 1, ptl);
6123 * We have consumed all precharges we got in can_attach().
6124 * We try charge one by one, but don't do any additional
6125 * charges to mc.to if we have failed in charge once in attach()
6128 ret = mem_cgroup_do_precharge(1);
6136 static const struct mm_walk_ops charge_walk_ops = {
6137 .pmd_entry = mem_cgroup_move_charge_pte_range,
6140 static void mem_cgroup_move_charge(void)
6142 lru_add_drain_all();
6144 * Signal lock_page_memcg() to take the memcg's move_lock
6145 * while we're moving its pages to another memcg. Then wait
6146 * for already started RCU-only updates to finish.
6148 atomic_inc(&mc.from->moving_account);
6151 if (unlikely(!mmap_read_trylock(mc.mm))) {
6153 * Someone who are holding the mmap_lock might be waiting in
6154 * waitq. So we cancel all extra charges, wake up all waiters,
6155 * and retry. Because we cancel precharges, we might not be able
6156 * to move enough charges, but moving charge is a best-effort
6157 * feature anyway, so it wouldn't be a big problem.
6159 __mem_cgroup_clear_mc();
6164 * When we have consumed all precharges and failed in doing
6165 * additional charge, the page walk just aborts.
6167 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6170 mmap_read_unlock(mc.mm);
6171 atomic_dec(&mc.from->moving_account);
6174 static void mem_cgroup_move_task(void)
6177 mem_cgroup_move_charge();
6178 mem_cgroup_clear_mc();
6181 #else /* !CONFIG_MMU */
6182 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6186 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6189 static void mem_cgroup_move_task(void)
6194 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6196 if (value == PAGE_COUNTER_MAX)
6197 seq_puts(m, "max\n");
6199 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6204 static u64 memory_current_read(struct cgroup_subsys_state *css,
6207 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6209 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6212 static int memory_min_show(struct seq_file *m, void *v)
6214 return seq_puts_memcg_tunable(m,
6215 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6218 static ssize_t memory_min_write(struct kernfs_open_file *of,
6219 char *buf, size_t nbytes, loff_t off)
6221 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6225 buf = strstrip(buf);
6226 err = page_counter_memparse(buf, "max", &min);
6230 page_counter_set_min(&memcg->memory, min);
6235 static int memory_low_show(struct seq_file *m, void *v)
6237 return seq_puts_memcg_tunable(m,
6238 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6241 static ssize_t memory_low_write(struct kernfs_open_file *of,
6242 char *buf, size_t nbytes, loff_t off)
6244 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6248 buf = strstrip(buf);
6249 err = page_counter_memparse(buf, "max", &low);
6253 page_counter_set_low(&memcg->memory, low);
6258 static int memory_high_show(struct seq_file *m, void *v)
6260 return seq_puts_memcg_tunable(m,
6261 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6264 static ssize_t memory_high_write(struct kernfs_open_file *of,
6265 char *buf, size_t nbytes, loff_t off)
6267 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6268 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6269 bool drained = false;
6273 buf = strstrip(buf);
6274 err = page_counter_memparse(buf, "max", &high);
6278 page_counter_set_high(&memcg->memory, high);
6281 unsigned long nr_pages = page_counter_read(&memcg->memory);
6282 unsigned long reclaimed;
6284 if (nr_pages <= high)
6287 if (signal_pending(current))
6291 drain_all_stock(memcg);
6296 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6299 if (!reclaimed && !nr_retries--)
6303 memcg_wb_domain_size_changed(memcg);
6307 static int memory_max_show(struct seq_file *m, void *v)
6309 return seq_puts_memcg_tunable(m,
6310 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6313 static ssize_t memory_max_write(struct kernfs_open_file *of,
6314 char *buf, size_t nbytes, loff_t off)
6316 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6317 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6318 bool drained = false;
6322 buf = strstrip(buf);
6323 err = page_counter_memparse(buf, "max", &max);
6327 xchg(&memcg->memory.max, max);
6330 unsigned long nr_pages = page_counter_read(&memcg->memory);
6332 if (nr_pages <= max)
6335 if (signal_pending(current))
6339 drain_all_stock(memcg);
6345 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6351 memcg_memory_event(memcg, MEMCG_OOM);
6352 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6356 memcg_wb_domain_size_changed(memcg);
6360 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6362 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6363 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6364 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6365 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6366 seq_printf(m, "oom_kill %lu\n",
6367 atomic_long_read(&events[MEMCG_OOM_KILL]));
6370 static int memory_events_show(struct seq_file *m, void *v)
6372 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6374 __memory_events_show(m, memcg->memory_events);
6378 static int memory_events_local_show(struct seq_file *m, void *v)
6380 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6382 __memory_events_show(m, memcg->memory_events_local);
6386 static int memory_stat_show(struct seq_file *m, void *v)
6388 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6391 buf = memory_stat_format(memcg);
6400 static inline unsigned long lruvec_page_state_output(struct lruvec *lruvec,
6403 return lruvec_page_state(lruvec, item) * memcg_page_state_unit(item);
6406 static int memory_numa_stat_show(struct seq_file *m, void *v)
6409 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6411 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6414 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6417 seq_printf(m, "%s", memory_stats[i].name);
6418 for_each_node_state(nid, N_MEMORY) {
6420 struct lruvec *lruvec;
6422 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6423 size = lruvec_page_state_output(lruvec,
6424 memory_stats[i].idx);
6425 seq_printf(m, " N%d=%llu", nid, size);
6434 static int memory_oom_group_show(struct seq_file *m, void *v)
6436 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6438 seq_printf(m, "%d\n", memcg->oom_group);
6443 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6444 char *buf, size_t nbytes, loff_t off)
6446 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6449 buf = strstrip(buf);
6453 ret = kstrtoint(buf, 0, &oom_group);
6457 if (oom_group != 0 && oom_group != 1)
6460 memcg->oom_group = oom_group;
6465 static struct cftype memory_files[] = {
6468 .flags = CFTYPE_NOT_ON_ROOT,
6469 .read_u64 = memory_current_read,
6473 .flags = CFTYPE_NOT_ON_ROOT,
6474 .seq_show = memory_min_show,
6475 .write = memory_min_write,
6479 .flags = CFTYPE_NOT_ON_ROOT,
6480 .seq_show = memory_low_show,
6481 .write = memory_low_write,
6485 .flags = CFTYPE_NOT_ON_ROOT,
6486 .seq_show = memory_high_show,
6487 .write = memory_high_write,
6491 .flags = CFTYPE_NOT_ON_ROOT,
6492 .seq_show = memory_max_show,
6493 .write = memory_max_write,
6497 .flags = CFTYPE_NOT_ON_ROOT,
6498 .file_offset = offsetof(struct mem_cgroup, events_file),
6499 .seq_show = memory_events_show,
6502 .name = "events.local",
6503 .flags = CFTYPE_NOT_ON_ROOT,
6504 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6505 .seq_show = memory_events_local_show,
6509 .seq_show = memory_stat_show,
6513 .name = "numa_stat",
6514 .seq_show = memory_numa_stat_show,
6518 .name = "oom.group",
6519 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6520 .seq_show = memory_oom_group_show,
6521 .write = memory_oom_group_write,
6526 struct cgroup_subsys memory_cgrp_subsys = {
6527 .css_alloc = mem_cgroup_css_alloc,
6528 .css_online = mem_cgroup_css_online,
6529 .css_offline = mem_cgroup_css_offline,
6530 .css_released = mem_cgroup_css_released,
6531 .css_free = mem_cgroup_css_free,
6532 .css_reset = mem_cgroup_css_reset,
6533 .can_attach = mem_cgroup_can_attach,
6534 .cancel_attach = mem_cgroup_cancel_attach,
6535 .post_attach = mem_cgroup_move_task,
6536 .dfl_cftypes = memory_files,
6537 .legacy_cftypes = mem_cgroup_legacy_files,
6542 * This function calculates an individual cgroup's effective
6543 * protection which is derived from its own memory.min/low, its
6544 * parent's and siblings' settings, as well as the actual memory
6545 * distribution in the tree.
6547 * The following rules apply to the effective protection values:
6549 * 1. At the first level of reclaim, effective protection is equal to
6550 * the declared protection in memory.min and memory.low.
6552 * 2. To enable safe delegation of the protection configuration, at
6553 * subsequent levels the effective protection is capped to the
6554 * parent's effective protection.
6556 * 3. To make complex and dynamic subtrees easier to configure, the
6557 * user is allowed to overcommit the declared protection at a given
6558 * level. If that is the case, the parent's effective protection is
6559 * distributed to the children in proportion to how much protection
6560 * they have declared and how much of it they are utilizing.
6562 * This makes distribution proportional, but also work-conserving:
6563 * if one cgroup claims much more protection than it uses memory,
6564 * the unused remainder is available to its siblings.
6566 * 4. Conversely, when the declared protection is undercommitted at a
6567 * given level, the distribution of the larger parental protection
6568 * budget is NOT proportional. A cgroup's protection from a sibling
6569 * is capped to its own memory.min/low setting.
6571 * 5. However, to allow protecting recursive subtrees from each other
6572 * without having to declare each individual cgroup's fixed share
6573 * of the ancestor's claim to protection, any unutilized -
6574 * "floating" - protection from up the tree is distributed in
6575 * proportion to each cgroup's *usage*. This makes the protection
6576 * neutral wrt sibling cgroups and lets them compete freely over
6577 * the shared parental protection budget, but it protects the
6578 * subtree as a whole from neighboring subtrees.
6580 * Note that 4. and 5. are not in conflict: 4. is about protecting
6581 * against immediate siblings whereas 5. is about protecting against
6582 * neighboring subtrees.
6584 static unsigned long effective_protection(unsigned long usage,
6585 unsigned long parent_usage,
6586 unsigned long setting,
6587 unsigned long parent_effective,
6588 unsigned long siblings_protected)
6590 unsigned long protected;
6593 protected = min(usage, setting);
6595 * If all cgroups at this level combined claim and use more
6596 * protection then what the parent affords them, distribute
6597 * shares in proportion to utilization.
6599 * We are using actual utilization rather than the statically
6600 * claimed protection in order to be work-conserving: claimed
6601 * but unused protection is available to siblings that would
6602 * otherwise get a smaller chunk than what they claimed.
6604 if (siblings_protected > parent_effective)
6605 return protected * parent_effective / siblings_protected;
6608 * Ok, utilized protection of all children is within what the
6609 * parent affords them, so we know whatever this child claims
6610 * and utilizes is effectively protected.
6612 * If there is unprotected usage beyond this value, reclaim
6613 * will apply pressure in proportion to that amount.
6615 * If there is unutilized protection, the cgroup will be fully
6616 * shielded from reclaim, but we do return a smaller value for
6617 * protection than what the group could enjoy in theory. This
6618 * is okay. With the overcommit distribution above, effective
6619 * protection is always dependent on how memory is actually
6620 * consumed among the siblings anyway.
6625 * If the children aren't claiming (all of) the protection
6626 * afforded to them by the parent, distribute the remainder in
6627 * proportion to the (unprotected) memory of each cgroup. That
6628 * way, cgroups that aren't explicitly prioritized wrt each
6629 * other compete freely over the allowance, but they are
6630 * collectively protected from neighboring trees.
6632 * We're using unprotected memory for the weight so that if
6633 * some cgroups DO claim explicit protection, we don't protect
6634 * the same bytes twice.
6636 * Check both usage and parent_usage against the respective
6637 * protected values. One should imply the other, but they
6638 * aren't read atomically - make sure the division is sane.
6640 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6642 if (parent_effective > siblings_protected &&
6643 parent_usage > siblings_protected &&
6644 usage > protected) {
6645 unsigned long unclaimed;
6647 unclaimed = parent_effective - siblings_protected;
6648 unclaimed *= usage - protected;
6649 unclaimed /= parent_usage - siblings_protected;
6658 * mem_cgroup_protected - check if memory consumption is in the normal range
6659 * @root: the top ancestor of the sub-tree being checked
6660 * @memcg: the memory cgroup to check
6662 * WARNING: This function is not stateless! It can only be used as part
6663 * of a top-down tree iteration, not for isolated queries.
6665 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6666 struct mem_cgroup *memcg)
6668 unsigned long usage, parent_usage;
6669 struct mem_cgroup *parent;
6671 if (mem_cgroup_disabled())
6675 root = root_mem_cgroup;
6678 * Effective values of the reclaim targets are ignored so they
6679 * can be stale. Have a look at mem_cgroup_protection for more
6681 * TODO: calculation should be more robust so that we do not need
6682 * that special casing.
6687 usage = page_counter_read(&memcg->memory);
6691 parent = parent_mem_cgroup(memcg);
6692 /* No parent means a non-hierarchical mode on v1 memcg */
6696 if (parent == root) {
6697 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6698 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6702 parent_usage = page_counter_read(&parent->memory);
6704 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6705 READ_ONCE(memcg->memory.min),
6706 READ_ONCE(parent->memory.emin),
6707 atomic_long_read(&parent->memory.children_min_usage)));
6709 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6710 READ_ONCE(memcg->memory.low),
6711 READ_ONCE(parent->memory.elow),
6712 atomic_long_read(&parent->memory.children_low_usage)));
6716 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6717 * @page: page to charge
6718 * @mm: mm context of the victim
6719 * @gfp_mask: reclaim mode
6721 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6722 * pages according to @gfp_mask if necessary.
6724 * Returns 0 on success. Otherwise, an error code is returned.
6726 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6728 unsigned int nr_pages = thp_nr_pages(page);
6729 struct mem_cgroup *memcg = NULL;
6732 if (mem_cgroup_disabled())
6735 if (PageSwapCache(page)) {
6736 swp_entry_t ent = { .val = page_private(page), };
6740 * Every swap fault against a single page tries to charge the
6741 * page, bail as early as possible. shmem_unuse() encounters
6742 * already charged pages, too. page and memcg binding is
6743 * protected by the page lock, which serializes swap cache
6744 * removal, which in turn serializes uncharging.
6746 VM_BUG_ON_PAGE(!PageLocked(page), page);
6747 if (page_memcg(compound_head(page)))
6750 id = lookup_swap_cgroup_id(ent);
6752 memcg = mem_cgroup_from_id(id);
6753 if (memcg && !css_tryget_online(&memcg->css))
6759 memcg = get_mem_cgroup_from_mm(mm);
6761 ret = try_charge(memcg, gfp_mask, nr_pages);
6765 css_get(&memcg->css);
6766 commit_charge(page, memcg);
6768 local_irq_disable();
6769 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6770 memcg_check_events(memcg, page);
6773 if (PageSwapCache(page)) {
6774 swp_entry_t entry = { .val = page_private(page) };
6776 * The swap entry might not get freed for a long time,
6777 * let's not wait for it. The page already received a
6778 * memory+swap charge, drop the swap entry duplicate.
6780 mem_cgroup_uncharge_swap(entry, nr_pages);
6784 css_put(&memcg->css);
6789 struct uncharge_gather {
6790 struct mem_cgroup *memcg;
6791 unsigned long nr_pages;
6792 unsigned long pgpgout;
6793 unsigned long nr_kmem;
6794 struct page *dummy_page;
6797 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6799 memset(ug, 0, sizeof(*ug));
6802 static void uncharge_batch(const struct uncharge_gather *ug)
6804 unsigned long flags;
6806 if (!mem_cgroup_is_root(ug->memcg)) {
6807 page_counter_uncharge(&ug->memcg->memory, ug->nr_pages);
6808 if (do_memsw_account())
6809 page_counter_uncharge(&ug->memcg->memsw, ug->nr_pages);
6810 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6811 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6812 memcg_oom_recover(ug->memcg);
6815 local_irq_save(flags);
6816 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6817 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_pages);
6818 memcg_check_events(ug->memcg, ug->dummy_page);
6819 local_irq_restore(flags);
6821 /* drop reference from uncharge_page */
6822 css_put(&ug->memcg->css);
6825 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6827 unsigned long nr_pages;
6829 VM_BUG_ON_PAGE(PageLRU(page), page);
6831 if (!page_memcg(page))
6835 * Nobody should be changing or seriously looking at
6836 * page_memcg(page) at this point, we have fully
6837 * exclusive access to the page.
6840 if (ug->memcg != page_memcg(page)) {
6843 uncharge_gather_clear(ug);
6845 ug->memcg = page_memcg(page);
6847 /* pairs with css_put in uncharge_batch */
6848 css_get(&ug->memcg->css);
6851 nr_pages = compound_nr(page);
6852 ug->nr_pages += nr_pages;
6854 if (PageMemcgKmem(page))
6855 ug->nr_kmem += nr_pages;
6859 ug->dummy_page = page;
6860 page->memcg_data = 0;
6861 css_put(&ug->memcg->css);
6864 static void uncharge_list(struct list_head *page_list)
6866 struct uncharge_gather ug;
6867 struct list_head *next;
6869 uncharge_gather_clear(&ug);
6872 * Note that the list can be a single page->lru; hence the
6873 * do-while loop instead of a simple list_for_each_entry().
6875 next = page_list->next;
6879 page = list_entry(next, struct page, lru);
6880 next = page->lru.next;
6882 uncharge_page(page, &ug);
6883 } while (next != page_list);
6886 uncharge_batch(&ug);
6890 * mem_cgroup_uncharge - uncharge a page
6891 * @page: page to uncharge
6893 * Uncharge a page previously charged with mem_cgroup_charge().
6895 void mem_cgroup_uncharge(struct page *page)
6897 struct uncharge_gather ug;
6899 if (mem_cgroup_disabled())
6902 /* Don't touch page->lru of any random page, pre-check: */
6903 if (!page_memcg(page))
6906 uncharge_gather_clear(&ug);
6907 uncharge_page(page, &ug);
6908 uncharge_batch(&ug);
6912 * mem_cgroup_uncharge_list - uncharge a list of page
6913 * @page_list: list of pages to uncharge
6915 * Uncharge a list of pages previously charged with
6916 * mem_cgroup_charge().
6918 void mem_cgroup_uncharge_list(struct list_head *page_list)
6920 if (mem_cgroup_disabled())
6923 if (!list_empty(page_list))
6924 uncharge_list(page_list);
6928 * mem_cgroup_migrate - charge a page's replacement
6929 * @oldpage: currently circulating page
6930 * @newpage: replacement page
6932 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6933 * be uncharged upon free.
6935 * Both pages must be locked, @newpage->mapping must be set up.
6937 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6939 struct mem_cgroup *memcg;
6940 unsigned int nr_pages;
6941 unsigned long flags;
6943 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6944 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6945 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6946 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6949 if (mem_cgroup_disabled())
6952 /* Page cache replacement: new page already charged? */
6953 if (page_memcg(newpage))
6956 memcg = page_memcg(oldpage);
6957 VM_WARN_ON_ONCE_PAGE(!memcg, oldpage);
6961 /* Force-charge the new page. The old one will be freed soon */
6962 nr_pages = thp_nr_pages(newpage);
6964 page_counter_charge(&memcg->memory, nr_pages);
6965 if (do_memsw_account())
6966 page_counter_charge(&memcg->memsw, nr_pages);
6968 css_get(&memcg->css);
6969 commit_charge(newpage, memcg);
6971 local_irq_save(flags);
6972 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
6973 memcg_check_events(memcg, newpage);
6974 local_irq_restore(flags);
6977 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6978 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6980 void mem_cgroup_sk_alloc(struct sock *sk)
6982 struct mem_cgroup *memcg;
6984 if (!mem_cgroup_sockets_enabled)
6987 /* Do not associate the sock with unrelated interrupted task's memcg. */
6992 memcg = mem_cgroup_from_task(current);
6993 if (memcg == root_mem_cgroup)
6995 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6997 if (css_tryget(&memcg->css))
6998 sk->sk_memcg = memcg;
7003 void mem_cgroup_sk_free(struct sock *sk)
7006 css_put(&sk->sk_memcg->css);
7010 * mem_cgroup_charge_skmem - charge socket memory
7011 * @memcg: memcg to charge
7012 * @nr_pages: number of pages to charge
7014 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7015 * @memcg's configured limit, %false if the charge had to be forced.
7017 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7019 gfp_t gfp_mask = GFP_KERNEL;
7021 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7022 struct page_counter *fail;
7024 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7025 memcg->tcpmem_pressure = 0;
7028 page_counter_charge(&memcg->tcpmem, nr_pages);
7029 memcg->tcpmem_pressure = 1;
7033 /* Don't block in the packet receive path */
7035 gfp_mask = GFP_NOWAIT;
7037 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7039 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
7042 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
7047 * mem_cgroup_uncharge_skmem - uncharge socket memory
7048 * @memcg: memcg to uncharge
7049 * @nr_pages: number of pages to uncharge
7051 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7053 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7054 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7058 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7060 refill_stock(memcg, nr_pages);
7063 static int __init cgroup_memory(char *s)
7067 while ((token = strsep(&s, ",")) != NULL) {
7070 if (!strcmp(token, "nosocket"))
7071 cgroup_memory_nosocket = true;
7072 if (!strcmp(token, "nokmem"))
7073 cgroup_memory_nokmem = true;
7077 __setup("cgroup.memory=", cgroup_memory);
7080 * subsys_initcall() for memory controller.
7082 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7083 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7084 * basically everything that doesn't depend on a specific mem_cgroup structure
7085 * should be initialized from here.
7087 static int __init mem_cgroup_init(void)
7092 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7093 * used for per-memcg-per-cpu caching of per-node statistics. In order
7094 * to work fine, we should make sure that the overfill threshold can't
7095 * exceed S32_MAX / PAGE_SIZE.
7097 BUILD_BUG_ON(MEMCG_CHARGE_BATCH > S32_MAX / PAGE_SIZE);
7099 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7100 memcg_hotplug_cpu_dead);
7102 for_each_possible_cpu(cpu)
7103 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7106 for_each_node(node) {
7107 struct mem_cgroup_tree_per_node *rtpn;
7109 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7110 node_online(node) ? node : NUMA_NO_NODE);
7112 rtpn->rb_root = RB_ROOT;
7113 rtpn->rb_rightmost = NULL;
7114 spin_lock_init(&rtpn->lock);
7115 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7120 subsys_initcall(mem_cgroup_init);
7122 #ifdef CONFIG_MEMCG_SWAP
7123 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7125 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7127 * The root cgroup cannot be destroyed, so it's refcount must
7130 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7134 memcg = parent_mem_cgroup(memcg);
7136 memcg = root_mem_cgroup;
7142 * mem_cgroup_swapout - transfer a memsw charge to swap
7143 * @page: page whose memsw charge to transfer
7144 * @entry: swap entry to move the charge to
7146 * Transfer the memsw charge of @page to @entry.
7148 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7150 struct mem_cgroup *memcg, *swap_memcg;
7151 unsigned int nr_entries;
7152 unsigned short oldid;
7154 VM_BUG_ON_PAGE(PageLRU(page), page);
7155 VM_BUG_ON_PAGE(page_count(page), page);
7157 if (mem_cgroup_disabled())
7160 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7163 memcg = page_memcg(page);
7165 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7170 * In case the memcg owning these pages has been offlined and doesn't
7171 * have an ID allocated to it anymore, charge the closest online
7172 * ancestor for the swap instead and transfer the memory+swap charge.
7174 swap_memcg = mem_cgroup_id_get_online(memcg);
7175 nr_entries = thp_nr_pages(page);
7176 /* Get references for the tail pages, too */
7178 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7179 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7181 VM_BUG_ON_PAGE(oldid, page);
7182 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7184 page->memcg_data = 0;
7186 if (!mem_cgroup_is_root(memcg))
7187 page_counter_uncharge(&memcg->memory, nr_entries);
7189 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7190 if (!mem_cgroup_is_root(swap_memcg))
7191 page_counter_charge(&swap_memcg->memsw, nr_entries);
7192 page_counter_uncharge(&memcg->memsw, nr_entries);
7196 * Interrupts should be disabled here because the caller holds the
7197 * i_pages lock which is taken with interrupts-off. It is
7198 * important here to have the interrupts disabled because it is the
7199 * only synchronisation we have for updating the per-CPU variables.
7201 VM_BUG_ON(!irqs_disabled());
7202 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7203 memcg_check_events(memcg, page);
7205 css_put(&memcg->css);
7209 * mem_cgroup_try_charge_swap - try charging swap space for a page
7210 * @page: page being added to swap
7211 * @entry: swap entry to charge
7213 * Try to charge @page's memcg for the swap space at @entry.
7215 * Returns 0 on success, -ENOMEM on failure.
7217 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7219 unsigned int nr_pages = thp_nr_pages(page);
7220 struct page_counter *counter;
7221 struct mem_cgroup *memcg;
7222 unsigned short oldid;
7224 if (mem_cgroup_disabled())
7227 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7230 memcg = page_memcg(page);
7232 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7237 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7241 memcg = mem_cgroup_id_get_online(memcg);
7243 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7244 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7245 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7246 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7247 mem_cgroup_id_put(memcg);
7251 /* Get references for the tail pages, too */
7253 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7254 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7255 VM_BUG_ON_PAGE(oldid, page);
7256 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7262 * mem_cgroup_uncharge_swap - uncharge swap space
7263 * @entry: swap entry to uncharge
7264 * @nr_pages: the amount of swap space to uncharge
7266 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7268 struct mem_cgroup *memcg;
7271 id = swap_cgroup_record(entry, 0, nr_pages);
7273 memcg = mem_cgroup_from_id(id);
7275 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7276 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7277 page_counter_uncharge(&memcg->swap, nr_pages);
7279 page_counter_uncharge(&memcg->memsw, nr_pages);
7281 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7282 mem_cgroup_id_put_many(memcg, nr_pages);
7287 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7289 long nr_swap_pages = get_nr_swap_pages();
7291 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7292 return nr_swap_pages;
7293 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7294 nr_swap_pages = min_t(long, nr_swap_pages,
7295 READ_ONCE(memcg->swap.max) -
7296 page_counter_read(&memcg->swap));
7297 return nr_swap_pages;
7300 bool mem_cgroup_swap_full(struct page *page)
7302 struct mem_cgroup *memcg;
7304 VM_BUG_ON_PAGE(!PageLocked(page), page);
7308 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7311 memcg = page_memcg(page);
7315 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7316 unsigned long usage = page_counter_read(&memcg->swap);
7318 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7319 usage * 2 >= READ_ONCE(memcg->swap.max))
7326 static int __init setup_swap_account(char *s)
7328 if (!strcmp(s, "1"))
7329 cgroup_memory_noswap = false;
7330 else if (!strcmp(s, "0"))
7331 cgroup_memory_noswap = true;
7334 __setup("swapaccount=", setup_swap_account);
7336 static u64 swap_current_read(struct cgroup_subsys_state *css,
7339 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7341 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7344 static int swap_high_show(struct seq_file *m, void *v)
7346 return seq_puts_memcg_tunable(m,
7347 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7350 static ssize_t swap_high_write(struct kernfs_open_file *of,
7351 char *buf, size_t nbytes, loff_t off)
7353 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7357 buf = strstrip(buf);
7358 err = page_counter_memparse(buf, "max", &high);
7362 page_counter_set_high(&memcg->swap, high);
7367 static int swap_max_show(struct seq_file *m, void *v)
7369 return seq_puts_memcg_tunable(m,
7370 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7373 static ssize_t swap_max_write(struct kernfs_open_file *of,
7374 char *buf, size_t nbytes, loff_t off)
7376 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7380 buf = strstrip(buf);
7381 err = page_counter_memparse(buf, "max", &max);
7385 xchg(&memcg->swap.max, max);
7390 static int swap_events_show(struct seq_file *m, void *v)
7392 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7394 seq_printf(m, "high %lu\n",
7395 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7396 seq_printf(m, "max %lu\n",
7397 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7398 seq_printf(m, "fail %lu\n",
7399 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7404 static struct cftype swap_files[] = {
7406 .name = "swap.current",
7407 .flags = CFTYPE_NOT_ON_ROOT,
7408 .read_u64 = swap_current_read,
7411 .name = "swap.high",
7412 .flags = CFTYPE_NOT_ON_ROOT,
7413 .seq_show = swap_high_show,
7414 .write = swap_high_write,
7418 .flags = CFTYPE_NOT_ON_ROOT,
7419 .seq_show = swap_max_show,
7420 .write = swap_max_write,
7423 .name = "swap.events",
7424 .flags = CFTYPE_NOT_ON_ROOT,
7425 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7426 .seq_show = swap_events_show,
7431 static struct cftype memsw_files[] = {
7433 .name = "memsw.usage_in_bytes",
7434 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7435 .read_u64 = mem_cgroup_read_u64,
7438 .name = "memsw.max_usage_in_bytes",
7439 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7440 .write = mem_cgroup_reset,
7441 .read_u64 = mem_cgroup_read_u64,
7444 .name = "memsw.limit_in_bytes",
7445 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7446 .write = mem_cgroup_write,
7447 .read_u64 = mem_cgroup_read_u64,
7450 .name = "memsw.failcnt",
7451 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7452 .write = mem_cgroup_reset,
7453 .read_u64 = mem_cgroup_read_u64,
7455 { }, /* terminate */
7459 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7460 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7461 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7462 * boot parameter. This may result in premature OOPS inside
7463 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7465 static int __init mem_cgroup_swap_init(void)
7467 /* No memory control -> no swap control */
7468 if (mem_cgroup_disabled())
7469 cgroup_memory_noswap = true;
7471 if (cgroup_memory_noswap)
7474 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7475 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7479 core_initcall(mem_cgroup_swap_init);
7481 #endif /* CONFIG_MEMCG_SWAP */