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
25 #include <linux/page_counter.h>
26 #include <linux/memcontrol.h>
27 #include <linux/cgroup.h>
28 #include <linux/pagewalk.h>
29 #include <linux/sched/mm.h>
30 #include <linux/shmem_fs.h>
31 #include <linux/hugetlb.h>
32 #include <linux/pagemap.h>
33 #include <linux/vm_event_item.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/swap_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
59 #include <linux/tracehook.h>
60 #include <linux/psi.h>
61 #include <linux/seq_buf.h>
66 #include <linux/local_lock.h>
68 #include <linux/uaccess.h>
70 #include <trace/events/vmscan.h>
72 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
73 EXPORT_SYMBOL(memory_cgrp_subsys);
75 struct mem_cgroup *root_mem_cgroup __read_mostly;
77 /* Active memory cgroup to use from an interrupt context */
78 DEFINE_PER_CPU(struct mem_cgroup *, int_active_memcg);
80 /* Socket memory accounting disabled? */
81 static bool cgroup_memory_nosocket;
83 /* Kernel memory accounting disabled? */
84 static bool cgroup_memory_nokmem;
86 /* Whether the swap controller is active */
87 #ifdef CONFIG_MEMCG_SWAP
88 bool cgroup_memory_noswap __read_mostly;
90 #define cgroup_memory_noswap 1
93 #ifdef CONFIG_CGROUP_WRITEBACK
94 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
100 static DEFINE_PER_CPU(struct event_lock, event_lock) = {
101 .l = INIT_LOCAL_LOCK(l),
104 /* Whether legacy memory+swap accounting is active */
105 static bool do_memsw_account(void)
107 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_noswap;
110 #define THRESHOLDS_EVENTS_TARGET 128
111 #define SOFTLIMIT_EVENTS_TARGET 1024
114 * Cgroups above their limits are maintained in a RB-Tree, independent of
115 * their hierarchy representation
118 struct mem_cgroup_tree_per_node {
119 struct rb_root rb_root;
120 struct rb_node *rb_rightmost;
124 struct mem_cgroup_tree {
125 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
128 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
131 struct mem_cgroup_eventfd_list {
132 struct list_head list;
133 struct eventfd_ctx *eventfd;
137 * cgroup_event represents events which userspace want to receive.
139 struct mem_cgroup_event {
141 * memcg which the event belongs to.
143 struct mem_cgroup *memcg;
145 * eventfd to signal userspace about the event.
147 struct eventfd_ctx *eventfd;
149 * Each of these stored in a list by the cgroup.
151 struct list_head list;
153 * register_event() callback will be used to add new userspace
154 * waiter for changes related to this event. Use eventfd_signal()
155 * on eventfd to send notification to userspace.
157 int (*register_event)(struct mem_cgroup *memcg,
158 struct eventfd_ctx *eventfd, const char *args);
160 * unregister_event() callback will be called when userspace closes
161 * the eventfd or on cgroup removing. This callback must be set,
162 * if you want provide notification functionality.
164 void (*unregister_event)(struct mem_cgroup *memcg,
165 struct eventfd_ctx *eventfd);
167 * All fields below needed to unregister event when
168 * userspace closes eventfd.
171 wait_queue_head_t *wqh;
172 wait_queue_entry_t wait;
173 struct work_struct remove;
176 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
177 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
179 /* Stuffs for move charges at task migration. */
181 * Types of charges to be moved.
183 #define MOVE_ANON 0x1U
184 #define MOVE_FILE 0x2U
185 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
187 /* "mc" and its members are protected by cgroup_mutex */
188 static struct move_charge_struct {
189 spinlock_t lock; /* for from, to */
190 struct mm_struct *mm;
191 struct mem_cgroup *from;
192 struct mem_cgroup *to;
194 unsigned long precharge;
195 unsigned long moved_charge;
196 unsigned long moved_swap;
197 struct task_struct *moving_task; /* a task moving charges */
198 wait_queue_head_t waitq; /* a waitq for other context */
200 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
201 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
205 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
206 * limit reclaim to prevent infinite loops, if they ever occur.
208 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
209 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
211 /* for encoding cft->private value on file */
220 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
221 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
222 #define MEMFILE_ATTR(val) ((val) & 0xffff)
223 /* Used for OOM nofiier */
224 #define OOM_CONTROL (0)
227 * Iteration constructs for visiting all cgroups (under a tree). If
228 * loops are exited prematurely (break), mem_cgroup_iter_break() must
229 * be used for reference counting.
231 #define for_each_mem_cgroup_tree(iter, root) \
232 for (iter = mem_cgroup_iter(root, NULL, NULL); \
234 iter = mem_cgroup_iter(root, iter, NULL))
236 #define for_each_mem_cgroup(iter) \
237 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
239 iter = mem_cgroup_iter(NULL, iter, NULL))
241 static inline bool task_is_dying(void)
243 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
244 (current->flags & PF_EXITING);
247 /* Some nice accessors for the vmpressure. */
248 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
251 memcg = root_mem_cgroup;
252 return &memcg->vmpressure;
255 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
257 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
260 #ifdef CONFIG_MEMCG_KMEM
261 extern spinlock_t css_set_lock;
263 static void obj_cgroup_release(struct percpu_ref *ref)
265 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
266 struct mem_cgroup *memcg;
267 unsigned int nr_bytes;
268 unsigned int nr_pages;
272 * At this point all allocated objects are freed, and
273 * objcg->nr_charged_bytes can't have an arbitrary byte value.
274 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
276 * The following sequence can lead to it:
277 * 1) CPU0: objcg == stock->cached_objcg
278 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
279 * PAGE_SIZE bytes are charged
280 * 3) CPU1: a process from another memcg is allocating something,
281 * the stock if flushed,
282 * objcg->nr_charged_bytes = PAGE_SIZE - 92
283 * 5) CPU0: we do release this object,
284 * 92 bytes are added to stock->nr_bytes
285 * 6) CPU0: stock is flushed,
286 * 92 bytes are added to objcg->nr_charged_bytes
288 * In the result, nr_charged_bytes == PAGE_SIZE.
289 * This page will be uncharged in obj_cgroup_release().
291 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
292 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
293 nr_pages = nr_bytes >> PAGE_SHIFT;
295 spin_lock_irqsave(&css_set_lock, flags);
296 memcg = obj_cgroup_memcg(objcg);
298 __memcg_kmem_uncharge(memcg, nr_pages);
299 list_del(&objcg->list);
300 mem_cgroup_put(memcg);
301 spin_unlock_irqrestore(&css_set_lock, flags);
303 percpu_ref_exit(ref);
304 kfree_rcu(objcg, rcu);
307 static struct obj_cgroup *obj_cgroup_alloc(void)
309 struct obj_cgroup *objcg;
312 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
316 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
322 INIT_LIST_HEAD(&objcg->list);
326 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
327 struct mem_cgroup *parent)
329 struct obj_cgroup *objcg, *iter;
331 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
333 spin_lock_irq(&css_set_lock);
335 /* Move active objcg to the parent's list */
336 xchg(&objcg->memcg, parent);
337 css_get(&parent->css);
338 list_add(&objcg->list, &parent->objcg_list);
340 /* Move already reparented objcgs to the parent's list */
341 list_for_each_entry(iter, &memcg->objcg_list, list) {
342 css_get(&parent->css);
343 xchg(&iter->memcg, parent);
344 css_put(&memcg->css);
346 list_splice(&memcg->objcg_list, &parent->objcg_list);
348 spin_unlock_irq(&css_set_lock);
350 percpu_ref_kill(&objcg->refcnt);
354 * This will be used as a shrinker list's index.
355 * The main reason for not using cgroup id for this:
356 * this works better in sparse environments, where we have a lot of memcgs,
357 * but only a few kmem-limited. Or also, if we have, for instance, 200
358 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
359 * 200 entry array for that.
361 * The current size of the caches array is stored in memcg_nr_cache_ids. It
362 * will double each time we have to increase it.
364 static DEFINE_IDA(memcg_cache_ida);
365 int memcg_nr_cache_ids;
367 /* Protects memcg_nr_cache_ids */
368 static DECLARE_RWSEM(memcg_cache_ids_sem);
370 void memcg_get_cache_ids(void)
372 down_read(&memcg_cache_ids_sem);
375 void memcg_put_cache_ids(void)
377 up_read(&memcg_cache_ids_sem);
381 * MIN_SIZE is different than 1, because we would like to avoid going through
382 * the alloc/free process all the time. In a small machine, 4 kmem-limited
383 * cgroups is a reasonable guess. In the future, it could be a parameter or
384 * tunable, but that is strictly not necessary.
386 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
387 * this constant directly from cgroup, but it is understandable that this is
388 * better kept as an internal representation in cgroup.c. In any case, the
389 * cgrp_id space is not getting any smaller, and we don't have to necessarily
390 * increase ours as well if it increases.
392 #define MEMCG_CACHES_MIN_SIZE 4
393 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
396 * A lot of the calls to the cache allocation functions are expected to be
397 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
398 * conditional to this static branch, we'll have to allow modules that does
399 * kmem_cache_alloc and the such to see this symbol as well
401 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
402 EXPORT_SYMBOL(memcg_kmem_enabled_key);
405 static int memcg_shrinker_map_size;
406 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
408 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
410 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
413 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
414 int size, int old_size)
416 struct memcg_shrinker_map *new, *old;
419 lockdep_assert_held(&memcg_shrinker_map_mutex);
422 old = rcu_dereference_protected(
423 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
424 /* Not yet online memcg */
428 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
432 /* Set all old bits, clear all new bits */
433 memset(new->map, (int)0xff, old_size);
434 memset((void *)new->map + old_size, 0, size - old_size);
436 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
437 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
443 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
445 struct mem_cgroup_per_node *pn;
446 struct memcg_shrinker_map *map;
449 if (mem_cgroup_is_root(memcg))
453 pn = mem_cgroup_nodeinfo(memcg, nid);
454 map = rcu_dereference_protected(pn->shrinker_map, true);
457 rcu_assign_pointer(pn->shrinker_map, NULL);
461 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
463 struct memcg_shrinker_map *map;
464 int nid, size, ret = 0;
466 if (mem_cgroup_is_root(memcg))
469 mutex_lock(&memcg_shrinker_map_mutex);
470 size = memcg_shrinker_map_size;
472 map = kvzalloc_node(sizeof(*map) + size, GFP_KERNEL, nid);
474 memcg_free_shrinker_maps(memcg);
478 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
480 mutex_unlock(&memcg_shrinker_map_mutex);
485 int memcg_expand_shrinker_maps(int new_id)
487 int size, old_size, ret = 0;
488 struct mem_cgroup *memcg;
490 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
491 old_size = memcg_shrinker_map_size;
492 if (size <= old_size)
495 mutex_lock(&memcg_shrinker_map_mutex);
496 if (!root_mem_cgroup)
499 for_each_mem_cgroup(memcg) {
500 if (mem_cgroup_is_root(memcg))
502 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
504 mem_cgroup_iter_break(NULL, memcg);
510 memcg_shrinker_map_size = size;
511 mutex_unlock(&memcg_shrinker_map_mutex);
515 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
517 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
518 struct memcg_shrinker_map *map;
521 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
522 /* Pairs with smp mb in shrink_slab() */
523 smp_mb__before_atomic();
524 set_bit(shrinker_id, map->map);
530 * mem_cgroup_css_from_page - css of the memcg associated with a page
531 * @page: page of interest
533 * If memcg is bound to the default hierarchy, css of the memcg associated
534 * with @page is returned. The returned css remains associated with @page
535 * until it is released.
537 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
540 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
542 struct mem_cgroup *memcg;
544 memcg = page->mem_cgroup;
546 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
547 memcg = root_mem_cgroup;
553 * page_cgroup_ino - return inode number of the memcg a page is charged to
556 * Look up the closest online ancestor of the memory cgroup @page is charged to
557 * and return its inode number or 0 if @page is not charged to any cgroup. It
558 * is safe to call this function without holding a reference to @page.
560 * Note, this function is inherently racy, because there is nothing to prevent
561 * the cgroup inode from getting torn down and potentially reallocated a moment
562 * after page_cgroup_ino() returns, so it only should be used by callers that
563 * do not care (such as procfs interfaces).
565 ino_t page_cgroup_ino(struct page *page)
567 struct mem_cgroup *memcg;
568 unsigned long ino = 0;
571 memcg = page->mem_cgroup;
574 * The lowest bit set means that memcg isn't a valid
575 * memcg pointer, but a obj_cgroups pointer.
576 * In this case the page is shared and doesn't belong
577 * to any specific memory cgroup.
579 if ((unsigned long) memcg & 0x1UL)
582 while (memcg && !(memcg->css.flags & CSS_ONLINE))
583 memcg = parent_mem_cgroup(memcg);
585 ino = cgroup_ino(memcg->css.cgroup);
590 static struct mem_cgroup_per_node *
591 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
593 int nid = page_to_nid(page);
595 return memcg->nodeinfo[nid];
598 static struct mem_cgroup_tree_per_node *
599 soft_limit_tree_node(int nid)
601 return soft_limit_tree.rb_tree_per_node[nid];
604 static struct mem_cgroup_tree_per_node *
605 soft_limit_tree_from_page(struct page *page)
607 int nid = page_to_nid(page);
609 return soft_limit_tree.rb_tree_per_node[nid];
612 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
613 struct mem_cgroup_tree_per_node *mctz,
614 unsigned long new_usage_in_excess)
616 struct rb_node **p = &mctz->rb_root.rb_node;
617 struct rb_node *parent = NULL;
618 struct mem_cgroup_per_node *mz_node;
619 bool rightmost = true;
624 mz->usage_in_excess = new_usage_in_excess;
625 if (!mz->usage_in_excess)
629 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
631 if (mz->usage_in_excess < mz_node->usage_in_excess) {
637 * We can't avoid mem cgroups that are over their soft
638 * limit by the same amount
640 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
645 mctz->rb_rightmost = &mz->tree_node;
647 rb_link_node(&mz->tree_node, parent, p);
648 rb_insert_color(&mz->tree_node, &mctz->rb_root);
652 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
653 struct mem_cgroup_tree_per_node *mctz)
658 if (&mz->tree_node == mctz->rb_rightmost)
659 mctz->rb_rightmost = rb_prev(&mz->tree_node);
661 rb_erase(&mz->tree_node, &mctz->rb_root);
665 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
666 struct mem_cgroup_tree_per_node *mctz)
670 spin_lock_irqsave(&mctz->lock, flags);
671 __mem_cgroup_remove_exceeded(mz, mctz);
672 spin_unlock_irqrestore(&mctz->lock, flags);
675 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
677 unsigned long nr_pages = page_counter_read(&memcg->memory);
678 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
679 unsigned long excess = 0;
681 if (nr_pages > soft_limit)
682 excess = nr_pages - soft_limit;
687 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
689 unsigned long excess;
690 struct mem_cgroup_per_node *mz;
691 struct mem_cgroup_tree_per_node *mctz;
693 mctz = soft_limit_tree_from_page(page);
697 * Necessary to update all ancestors when hierarchy is used.
698 * because their event counter is not touched.
700 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
701 mz = mem_cgroup_page_nodeinfo(memcg, page);
702 excess = soft_limit_excess(memcg);
704 * We have to update the tree if mz is on RB-tree or
705 * mem is over its softlimit.
707 if (excess || mz->on_tree) {
710 spin_lock_irqsave(&mctz->lock, flags);
711 /* if on-tree, remove it */
713 __mem_cgroup_remove_exceeded(mz, mctz);
715 * Insert again. mz->usage_in_excess will be updated.
716 * If excess is 0, no tree ops.
718 __mem_cgroup_insert_exceeded(mz, mctz, excess);
719 spin_unlock_irqrestore(&mctz->lock, flags);
724 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
726 struct mem_cgroup_tree_per_node *mctz;
727 struct mem_cgroup_per_node *mz;
731 mz = mem_cgroup_nodeinfo(memcg, nid);
732 mctz = soft_limit_tree_node(nid);
734 mem_cgroup_remove_exceeded(mz, mctz);
738 static struct mem_cgroup_per_node *
739 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
741 struct mem_cgroup_per_node *mz;
745 if (!mctz->rb_rightmost)
746 goto done; /* Nothing to reclaim from */
748 mz = rb_entry(mctz->rb_rightmost,
749 struct mem_cgroup_per_node, tree_node);
751 * Remove the node now but someone else can add it back,
752 * we will to add it back at the end of reclaim to its correct
753 * position in the tree.
755 __mem_cgroup_remove_exceeded(mz, mctz);
756 if (!soft_limit_excess(mz->memcg) ||
757 !css_tryget(&mz->memcg->css))
763 static struct mem_cgroup_per_node *
764 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
766 struct mem_cgroup_per_node *mz;
768 spin_lock_irq(&mctz->lock);
769 mz = __mem_cgroup_largest_soft_limit_node(mctz);
770 spin_unlock_irq(&mctz->lock);
775 * __mod_memcg_state - update cgroup memory statistics
776 * @memcg: the memory cgroup
777 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
778 * @val: delta to add to the counter, can be negative
780 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
782 long x, threshold = MEMCG_CHARGE_BATCH;
784 if (mem_cgroup_disabled())
787 if (memcg_stat_item_in_bytes(idx))
788 threshold <<= PAGE_SHIFT;
790 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
791 if (unlikely(abs(x) > threshold)) {
792 struct mem_cgroup *mi;
795 * Batch local counters to keep them in sync with
796 * the hierarchical ones.
798 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
799 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
800 atomic_long_add(x, &mi->vmstats[idx]);
803 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
806 static struct mem_cgroup_per_node *
807 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
809 struct mem_cgroup *parent;
811 parent = parent_mem_cgroup(pn->memcg);
814 return mem_cgroup_nodeinfo(parent, nid);
817 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
820 struct mem_cgroup_per_node *pn;
821 struct mem_cgroup *memcg;
822 long x, threshold = MEMCG_CHARGE_BATCH;
824 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
827 preempt_disable_rt();
829 __mod_memcg_state(memcg, idx, val);
832 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
834 if (vmstat_item_in_bytes(idx))
835 threshold <<= PAGE_SHIFT;
837 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
838 if (unlikely(abs(x) > threshold)) {
839 pg_data_t *pgdat = lruvec_pgdat(lruvec);
840 struct mem_cgroup_per_node *pi;
842 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
843 atomic_long_add(x, &pi->lruvec_stat[idx]);
846 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
851 * __mod_lruvec_state - update lruvec memory statistics
852 * @lruvec: the lruvec
853 * @idx: the stat item
854 * @val: delta to add to the counter, can be negative
856 * The lruvec is the intersection of the NUMA node and a cgroup. This
857 * function updates the all three counters that are affected by a
858 * change of state at this level: per-node, per-cgroup, per-lruvec.
860 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
864 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
866 /* Update memcg and lruvec */
867 if (!mem_cgroup_disabled())
868 __mod_memcg_lruvec_state(lruvec, idx, val);
871 void __mod_lruvec_slab_state(void *p, enum node_stat_item idx, int val)
873 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
874 struct mem_cgroup *memcg;
875 struct lruvec *lruvec;
878 memcg = mem_cgroup_from_obj(p);
881 * Untracked pages have no memcg, no lruvec. Update only the
882 * node. If we reparent the slab objects to the root memcg,
883 * when we free the slab object, we need to update the per-memcg
884 * vmstats to keep it correct for the root memcg.
887 __mod_node_page_state(pgdat, idx, val);
889 lruvec = mem_cgroup_lruvec(memcg, pgdat);
890 __mod_lruvec_state(lruvec, idx, val);
895 void mod_memcg_obj_state(void *p, int idx, int val)
897 struct mem_cgroup *memcg;
900 memcg = mem_cgroup_from_obj(p);
902 mod_memcg_state(memcg, idx, val);
907 * __count_memcg_events - account VM events in a cgroup
908 * @memcg: the memory cgroup
909 * @idx: the event item
910 * @count: the number of events that occured
912 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
917 if (mem_cgroup_disabled())
920 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
921 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
922 struct mem_cgroup *mi;
925 * Batch local counters to keep them in sync with
926 * the hierarchical ones.
928 __this_cpu_add(memcg->vmstats_local->events[idx], x);
929 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
930 atomic_long_add(x, &mi->vmevents[idx]);
933 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
936 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
938 return atomic_long_read(&memcg->vmevents[event]);
941 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
946 for_each_possible_cpu(cpu)
947 x += per_cpu(memcg->vmstats_local->events[event], cpu);
951 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
955 /* pagein of a big page is an event. So, ignore page size */
957 __count_memcg_events(memcg, PGPGIN, 1);
959 __count_memcg_events(memcg, PGPGOUT, 1);
960 nr_pages = -nr_pages; /* for event */
963 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
966 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
967 enum mem_cgroup_events_target target)
969 unsigned long val, next;
971 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
972 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
973 /* from time_after() in jiffies.h */
974 if ((long)(next - val) < 0) {
976 case MEM_CGROUP_TARGET_THRESH:
977 next = val + THRESHOLDS_EVENTS_TARGET;
979 case MEM_CGROUP_TARGET_SOFTLIMIT:
980 next = val + SOFTLIMIT_EVENTS_TARGET;
985 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
992 * Check events in order.
995 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
997 /* threshold event is triggered in finer grain than soft limit */
998 if (unlikely(mem_cgroup_event_ratelimit(memcg,
999 MEM_CGROUP_TARGET_THRESH))) {
1002 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1003 MEM_CGROUP_TARGET_SOFTLIMIT);
1004 mem_cgroup_threshold(memcg);
1005 if (unlikely(do_softlimit))
1006 mem_cgroup_update_tree(memcg, page);
1010 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1013 * mm_update_next_owner() may clear mm->owner to NULL
1014 * if it races with swapoff, page migration, etc.
1015 * So this can be called with p == NULL.
1020 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1022 EXPORT_SYMBOL(mem_cgroup_from_task);
1025 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1026 * @mm: mm from which memcg should be extracted. It can be NULL.
1028 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
1029 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
1032 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1034 struct mem_cgroup *memcg;
1036 if (mem_cgroup_disabled())
1042 * Page cache insertions can happen withou an
1043 * actual mm context, e.g. during disk probing
1044 * on boot, loopback IO, acct() writes etc.
1047 memcg = root_mem_cgroup;
1049 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1050 if (unlikely(!memcg))
1051 memcg = root_mem_cgroup;
1053 } while (!css_tryget(&memcg->css));
1057 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1060 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
1061 * @page: page from which memcg should be extracted.
1063 * Obtain a reference on page->memcg and returns it if successful. Otherwise
1064 * root_mem_cgroup is returned.
1066 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
1068 struct mem_cgroup *memcg = page->mem_cgroup;
1070 if (mem_cgroup_disabled())
1074 /* Page should not get uncharged and freed memcg under us. */
1075 if (!memcg || WARN_ON_ONCE(!css_tryget(&memcg->css)))
1076 memcg = root_mem_cgroup;
1080 EXPORT_SYMBOL(get_mem_cgroup_from_page);
1082 static __always_inline struct mem_cgroup *active_memcg(void)
1085 return this_cpu_read(int_active_memcg);
1087 return current->active_memcg;
1090 static __always_inline struct mem_cgroup *get_active_memcg(void)
1092 struct mem_cgroup *memcg;
1095 memcg = active_memcg();
1096 /* remote memcg must hold a ref. */
1097 if (memcg && WARN_ON_ONCE(!css_tryget(&memcg->css)))
1098 memcg = root_mem_cgroup;
1104 static __always_inline bool memcg_kmem_bypass(void)
1106 /* Allow remote memcg charging from any context. */
1107 if (unlikely(active_memcg()))
1110 /* Memcg to charge can't be determined. */
1111 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
1118 * If active memcg is set, do not fallback to current->mm->memcg.
1120 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
1122 if (memcg_kmem_bypass())
1125 if (unlikely(active_memcg()))
1126 return get_active_memcg();
1128 return get_mem_cgroup_from_mm(current->mm);
1132 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1133 * @root: hierarchy root
1134 * @prev: previously returned memcg, NULL on first invocation
1135 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1137 * Returns references to children of the hierarchy below @root, or
1138 * @root itself, or %NULL after a full round-trip.
1140 * Caller must pass the return value in @prev on subsequent
1141 * invocations for reference counting, or use mem_cgroup_iter_break()
1142 * to cancel a hierarchy walk before the round-trip is complete.
1144 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1145 * in the hierarchy among all concurrent reclaimers operating on the
1148 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1149 struct mem_cgroup *prev,
1150 struct mem_cgroup_reclaim_cookie *reclaim)
1152 struct mem_cgroup_reclaim_iter *iter;
1153 struct cgroup_subsys_state *css = NULL;
1154 struct mem_cgroup *memcg = NULL;
1155 struct mem_cgroup *pos = NULL;
1157 if (mem_cgroup_disabled())
1161 root = root_mem_cgroup;
1163 if (prev && !reclaim)
1166 if (!root->use_hierarchy && root != root_mem_cgroup) {
1175 struct mem_cgroup_per_node *mz;
1177 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1180 if (prev && reclaim->generation != iter->generation)
1184 pos = READ_ONCE(iter->position);
1185 if (!pos || css_tryget(&pos->css))
1188 * css reference reached zero, so iter->position will
1189 * be cleared by ->css_released. However, we should not
1190 * rely on this happening soon, because ->css_released
1191 * is called from a work queue, and by busy-waiting we
1192 * might block it. So we clear iter->position right
1195 (void)cmpxchg(&iter->position, pos, NULL);
1203 css = css_next_descendant_pre(css, &root->css);
1206 * Reclaimers share the hierarchy walk, and a
1207 * new one might jump in right at the end of
1208 * the hierarchy - make sure they see at least
1209 * one group and restart from the beginning.
1217 * Verify the css and acquire a reference. The root
1218 * is provided by the caller, so we know it's alive
1219 * and kicking, and don't take an extra reference.
1221 memcg = mem_cgroup_from_css(css);
1223 if (css == &root->css)
1226 if (css_tryget(css))
1234 * The position could have already been updated by a competing
1235 * thread, so check that the value hasn't changed since we read
1236 * it to avoid reclaiming from the same cgroup twice.
1238 (void)cmpxchg(&iter->position, pos, memcg);
1246 reclaim->generation = iter->generation;
1252 if (prev && prev != root)
1253 css_put(&prev->css);
1259 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1260 * @root: hierarchy root
1261 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1263 void mem_cgroup_iter_break(struct mem_cgroup *root,
1264 struct mem_cgroup *prev)
1267 root = root_mem_cgroup;
1268 if (prev && prev != root)
1269 css_put(&prev->css);
1272 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1273 struct mem_cgroup *dead_memcg)
1275 struct mem_cgroup_reclaim_iter *iter;
1276 struct mem_cgroup_per_node *mz;
1279 for_each_node(nid) {
1280 mz = mem_cgroup_nodeinfo(from, nid);
1282 cmpxchg(&iter->position, dead_memcg, NULL);
1286 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1288 struct mem_cgroup *memcg = dead_memcg;
1289 struct mem_cgroup *last;
1292 __invalidate_reclaim_iterators(memcg, dead_memcg);
1294 } while ((memcg = parent_mem_cgroup(memcg)));
1297 * When cgruop1 non-hierarchy mode is used,
1298 * parent_mem_cgroup() does not walk all the way up to the
1299 * cgroup root (root_mem_cgroup). So we have to handle
1300 * dead_memcg from cgroup root separately.
1302 if (last != root_mem_cgroup)
1303 __invalidate_reclaim_iterators(root_mem_cgroup,
1308 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1309 * @memcg: hierarchy root
1310 * @fn: function to call for each task
1311 * @arg: argument passed to @fn
1313 * This function iterates over tasks attached to @memcg or to any of its
1314 * descendants and calls @fn for each task. If @fn returns a non-zero
1315 * value, the function breaks the iteration loop and returns the value.
1316 * Otherwise, it will iterate over all tasks and return 0.
1318 * This function must not be called for the root memory cgroup.
1320 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1321 int (*fn)(struct task_struct *, void *), void *arg)
1323 struct mem_cgroup *iter;
1326 BUG_ON(memcg == root_mem_cgroup);
1328 for_each_mem_cgroup_tree(iter, memcg) {
1329 struct css_task_iter it;
1330 struct task_struct *task;
1332 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1333 while (!ret && (task = css_task_iter_next(&it)))
1334 ret = fn(task, arg);
1335 css_task_iter_end(&it);
1337 mem_cgroup_iter_break(memcg, iter);
1345 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1347 * @pgdat: pgdat of the page
1349 * This function relies on page->mem_cgroup being stable - see the
1350 * access rules in commit_charge().
1352 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1354 struct mem_cgroup_per_node *mz;
1355 struct mem_cgroup *memcg;
1356 struct lruvec *lruvec;
1358 if (mem_cgroup_disabled()) {
1359 lruvec = &pgdat->__lruvec;
1363 memcg = page->mem_cgroup;
1365 * Swapcache readahead pages are added to the LRU - and
1366 * possibly migrated - before they are charged.
1369 memcg = root_mem_cgroup;
1371 mz = mem_cgroup_page_nodeinfo(memcg, page);
1372 lruvec = &mz->lruvec;
1375 * Since a node can be onlined after the mem_cgroup was created,
1376 * we have to be prepared to initialize lruvec->zone here;
1377 * and if offlined then reonlined, we need to reinitialize it.
1379 if (unlikely(lruvec->pgdat != pgdat))
1380 lruvec->pgdat = pgdat;
1385 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1386 * @lruvec: mem_cgroup per zone lru vector
1387 * @lru: index of lru list the page is sitting on
1388 * @zid: zone id of the accounted pages
1389 * @nr_pages: positive when adding or negative when removing
1391 * This function must be called under lru_lock, just before a page is added
1392 * to or just after a page is removed from an lru list (that ordering being
1393 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1395 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1396 int zid, int nr_pages)
1398 struct mem_cgroup_per_node *mz;
1399 unsigned long *lru_size;
1402 if (mem_cgroup_disabled())
1405 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1406 lru_size = &mz->lru_zone_size[zid][lru];
1409 *lru_size += nr_pages;
1412 if (WARN_ONCE(size < 0,
1413 "%s(%p, %d, %d): lru_size %ld\n",
1414 __func__, lruvec, lru, nr_pages, size)) {
1420 *lru_size += nr_pages;
1424 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1425 * @memcg: the memory cgroup
1427 * Returns the maximum amount of memory @mem can be charged with, in
1430 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1432 unsigned long margin = 0;
1433 unsigned long count;
1434 unsigned long limit;
1436 count = page_counter_read(&memcg->memory);
1437 limit = READ_ONCE(memcg->memory.max);
1439 margin = limit - count;
1441 if (do_memsw_account()) {
1442 count = page_counter_read(&memcg->memsw);
1443 limit = READ_ONCE(memcg->memsw.max);
1445 margin = min(margin, limit - count);
1454 * A routine for checking "mem" is under move_account() or not.
1456 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1457 * moving cgroups. This is for waiting at high-memory pressure
1460 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1462 struct mem_cgroup *from;
1463 struct mem_cgroup *to;
1466 * Unlike task_move routines, we access mc.to, mc.from not under
1467 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1469 spin_lock(&mc.lock);
1475 ret = mem_cgroup_is_descendant(from, memcg) ||
1476 mem_cgroup_is_descendant(to, memcg);
1478 spin_unlock(&mc.lock);
1482 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1484 if (mc.moving_task && current != mc.moving_task) {
1485 if (mem_cgroup_under_move(memcg)) {
1487 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1488 /* moving charge context might have finished. */
1491 finish_wait(&mc.waitq, &wait);
1498 struct memory_stat {
1504 static struct memory_stat memory_stats[] = {
1505 { "anon", PAGE_SIZE, NR_ANON_MAPPED },
1506 { "file", PAGE_SIZE, NR_FILE_PAGES },
1507 { "kernel_stack", 1024, NR_KERNEL_STACK_KB },
1508 { "percpu", 1, MEMCG_PERCPU_B },
1509 { "sock", PAGE_SIZE, MEMCG_SOCK },
1510 { "shmem", PAGE_SIZE, NR_SHMEM },
1511 { "file_mapped", PAGE_SIZE, NR_FILE_MAPPED },
1512 { "file_dirty", PAGE_SIZE, NR_FILE_DIRTY },
1513 { "file_writeback", PAGE_SIZE, NR_WRITEBACK },
1514 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1516 * The ratio will be initialized in memory_stats_init(). Because
1517 * on some architectures, the macro of HPAGE_PMD_SIZE is not
1518 * constant(e.g. powerpc).
1520 { "anon_thp", 0, NR_ANON_THPS },
1522 { "inactive_anon", PAGE_SIZE, NR_INACTIVE_ANON },
1523 { "active_anon", PAGE_SIZE, NR_ACTIVE_ANON },
1524 { "inactive_file", PAGE_SIZE, NR_INACTIVE_FILE },
1525 { "active_file", PAGE_SIZE, NR_ACTIVE_FILE },
1526 { "unevictable", PAGE_SIZE, NR_UNEVICTABLE },
1529 * Note: The slab_reclaimable and slab_unreclaimable must be
1530 * together and slab_reclaimable must be in front.
1532 { "slab_reclaimable", 1, NR_SLAB_RECLAIMABLE_B },
1533 { "slab_unreclaimable", 1, NR_SLAB_UNRECLAIMABLE_B },
1535 /* The memory events */
1536 { "workingset_refault_anon", 1, WORKINGSET_REFAULT_ANON },
1537 { "workingset_refault_file", 1, WORKINGSET_REFAULT_FILE },
1538 { "workingset_activate_anon", 1, WORKINGSET_ACTIVATE_ANON },
1539 { "workingset_activate_file", 1, WORKINGSET_ACTIVATE_FILE },
1540 { "workingset_restore_anon", 1, WORKINGSET_RESTORE_ANON },
1541 { "workingset_restore_file", 1, WORKINGSET_RESTORE_FILE },
1542 { "workingset_nodereclaim", 1, WORKINGSET_NODERECLAIM },
1545 static int __init memory_stats_init(void)
1549 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1550 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1551 if (memory_stats[i].idx == NR_ANON_THPS)
1552 memory_stats[i].ratio = HPAGE_PMD_SIZE;
1554 VM_BUG_ON(!memory_stats[i].ratio);
1555 VM_BUG_ON(memory_stats[i].idx >= MEMCG_NR_STAT);
1560 pure_initcall(memory_stats_init);
1562 static char *memory_stat_format(struct mem_cgroup *memcg)
1567 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1572 * Provide statistics on the state of the memory subsystem as
1573 * well as cumulative event counters that show past behavior.
1575 * This list is ordered following a combination of these gradients:
1576 * 1) generic big picture -> specifics and details
1577 * 2) reflecting userspace activity -> reflecting kernel heuristics
1579 * Current memory state:
1582 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1585 size = memcg_page_state(memcg, memory_stats[i].idx);
1586 size *= memory_stats[i].ratio;
1587 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1589 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1590 size = memcg_page_state(memcg, NR_SLAB_RECLAIMABLE_B) +
1591 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE_B);
1592 seq_buf_printf(&s, "slab %llu\n", size);
1596 /* Accumulated memory events */
1598 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1599 memcg_events(memcg, PGFAULT));
1600 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1601 memcg_events(memcg, PGMAJFAULT));
1602 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1603 memcg_events(memcg, PGREFILL));
1604 seq_buf_printf(&s, "pgscan %lu\n",
1605 memcg_events(memcg, PGSCAN_KSWAPD) +
1606 memcg_events(memcg, PGSCAN_DIRECT));
1607 seq_buf_printf(&s, "pgsteal %lu\n",
1608 memcg_events(memcg, PGSTEAL_KSWAPD) +
1609 memcg_events(memcg, PGSTEAL_DIRECT));
1610 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1611 memcg_events(memcg, PGACTIVATE));
1612 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1613 memcg_events(memcg, PGDEACTIVATE));
1614 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1615 memcg_events(memcg, PGLAZYFREE));
1616 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1617 memcg_events(memcg, PGLAZYFREED));
1619 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1620 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1621 memcg_events(memcg, THP_FAULT_ALLOC));
1622 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1623 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1624 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1626 /* The above should easily fit into one page */
1627 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1632 #define K(x) ((x) << (PAGE_SHIFT-10))
1634 * mem_cgroup_print_oom_context: Print OOM information relevant to
1635 * memory controller.
1636 * @memcg: The memory cgroup that went over limit
1637 * @p: Task that is going to be killed
1639 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1642 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1647 pr_cont(",oom_memcg=");
1648 pr_cont_cgroup_path(memcg->css.cgroup);
1650 pr_cont(",global_oom");
1652 pr_cont(",task_memcg=");
1653 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1659 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1660 * memory controller.
1661 * @memcg: The memory cgroup that went over limit
1663 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1667 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1668 K((u64)page_counter_read(&memcg->memory)),
1669 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1670 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1671 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1672 K((u64)page_counter_read(&memcg->swap)),
1673 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1675 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1676 K((u64)page_counter_read(&memcg->memsw)),
1677 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1678 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1679 K((u64)page_counter_read(&memcg->kmem)),
1680 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1683 pr_info("Memory cgroup stats for ");
1684 pr_cont_cgroup_path(memcg->css.cgroup);
1686 buf = memory_stat_format(memcg);
1694 * Return the memory (and swap, if configured) limit for a memcg.
1696 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1698 unsigned long max = READ_ONCE(memcg->memory.max);
1700 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1701 if (mem_cgroup_swappiness(memcg))
1702 max += min(READ_ONCE(memcg->swap.max),
1703 (unsigned long)total_swap_pages);
1705 if (mem_cgroup_swappiness(memcg)) {
1706 /* Calculate swap excess capacity from memsw limit */
1707 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1709 max += min(swap, (unsigned long)total_swap_pages);
1715 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1717 return page_counter_read(&memcg->memory);
1720 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1723 struct oom_control oc = {
1727 .gfp_mask = gfp_mask,
1732 if (mutex_lock_killable(&oom_lock))
1735 if (mem_cgroup_margin(memcg) >= (1 << order))
1739 * A few threads which were not waiting at mutex_lock_killable() can
1740 * fail to bail out. Therefore, check again after holding oom_lock.
1742 ret = task_is_dying() || out_of_memory(&oc);
1745 mutex_unlock(&oom_lock);
1749 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1752 unsigned long *total_scanned)
1754 struct mem_cgroup *victim = NULL;
1757 unsigned long excess;
1758 unsigned long nr_scanned;
1759 struct mem_cgroup_reclaim_cookie reclaim = {
1763 excess = soft_limit_excess(root_memcg);
1766 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1771 * If we have not been able to reclaim
1772 * anything, it might because there are
1773 * no reclaimable pages under this hierarchy
1778 * We want to do more targeted reclaim.
1779 * excess >> 2 is not to excessive so as to
1780 * reclaim too much, nor too less that we keep
1781 * coming back to reclaim from this cgroup
1783 if (total >= (excess >> 2) ||
1784 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1789 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1790 pgdat, &nr_scanned);
1791 *total_scanned += nr_scanned;
1792 if (!soft_limit_excess(root_memcg))
1795 mem_cgroup_iter_break(root_memcg, victim);
1799 #ifdef CONFIG_LOCKDEP
1800 static struct lockdep_map memcg_oom_lock_dep_map = {
1801 .name = "memcg_oom_lock",
1805 static DEFINE_SPINLOCK(memcg_oom_lock);
1808 * Check OOM-Killer is already running under our hierarchy.
1809 * If someone is running, return false.
1811 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1813 struct mem_cgroup *iter, *failed = NULL;
1815 spin_lock(&memcg_oom_lock);
1817 for_each_mem_cgroup_tree(iter, memcg) {
1818 if (iter->oom_lock) {
1820 * this subtree of our hierarchy is already locked
1821 * so we cannot give a lock.
1824 mem_cgroup_iter_break(memcg, iter);
1827 iter->oom_lock = true;
1832 * OK, we failed to lock the whole subtree so we have
1833 * to clean up what we set up to the failing subtree
1835 for_each_mem_cgroup_tree(iter, memcg) {
1836 if (iter == failed) {
1837 mem_cgroup_iter_break(memcg, iter);
1840 iter->oom_lock = false;
1843 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1845 spin_unlock(&memcg_oom_lock);
1850 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1852 struct mem_cgroup *iter;
1854 spin_lock(&memcg_oom_lock);
1855 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1856 for_each_mem_cgroup_tree(iter, memcg)
1857 iter->oom_lock = false;
1858 spin_unlock(&memcg_oom_lock);
1861 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1863 struct mem_cgroup *iter;
1865 spin_lock(&memcg_oom_lock);
1866 for_each_mem_cgroup_tree(iter, memcg)
1868 spin_unlock(&memcg_oom_lock);
1871 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1873 struct mem_cgroup *iter;
1876 * Be careful about under_oom underflows becase a child memcg
1877 * could have been added after mem_cgroup_mark_under_oom.
1879 spin_lock(&memcg_oom_lock);
1880 for_each_mem_cgroup_tree(iter, memcg)
1881 if (iter->under_oom > 0)
1883 spin_unlock(&memcg_oom_lock);
1886 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1888 struct oom_wait_info {
1889 struct mem_cgroup *memcg;
1890 wait_queue_entry_t wait;
1893 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1894 unsigned mode, int sync, void *arg)
1896 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1897 struct mem_cgroup *oom_wait_memcg;
1898 struct oom_wait_info *oom_wait_info;
1900 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1901 oom_wait_memcg = oom_wait_info->memcg;
1903 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1904 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1906 return autoremove_wake_function(wait, mode, sync, arg);
1909 static void memcg_oom_recover(struct mem_cgroup *memcg)
1912 * For the following lockless ->under_oom test, the only required
1913 * guarantee is that it must see the state asserted by an OOM when
1914 * this function is called as a result of userland actions
1915 * triggered by the notification of the OOM. This is trivially
1916 * achieved by invoking mem_cgroup_mark_under_oom() before
1917 * triggering notification.
1919 if (memcg && memcg->under_oom)
1920 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1930 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1932 enum oom_status ret;
1935 if (order > PAGE_ALLOC_COSTLY_ORDER)
1938 memcg_memory_event(memcg, MEMCG_OOM);
1941 * We are in the middle of the charge context here, so we
1942 * don't want to block when potentially sitting on a callstack
1943 * that holds all kinds of filesystem and mm locks.
1945 * cgroup1 allows disabling the OOM killer and waiting for outside
1946 * handling until the charge can succeed; remember the context and put
1947 * the task to sleep at the end of the page fault when all locks are
1950 * On the other hand, in-kernel OOM killer allows for an async victim
1951 * memory reclaim (oom_reaper) and that means that we are not solely
1952 * relying on the oom victim to make a forward progress and we can
1953 * invoke the oom killer here.
1955 * Please note that mem_cgroup_out_of_memory might fail to find a
1956 * victim and then we have to bail out from the charge path.
1958 if (memcg->oom_kill_disable) {
1959 if (!current->in_user_fault)
1961 css_get(&memcg->css);
1962 current->memcg_in_oom = memcg;
1963 current->memcg_oom_gfp_mask = mask;
1964 current->memcg_oom_order = order;
1969 mem_cgroup_mark_under_oom(memcg);
1971 locked = mem_cgroup_oom_trylock(memcg);
1974 mem_cgroup_oom_notify(memcg);
1976 mem_cgroup_unmark_under_oom(memcg);
1977 if (mem_cgroup_out_of_memory(memcg, mask, order))
1983 mem_cgroup_oom_unlock(memcg);
1989 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1990 * @handle: actually kill/wait or just clean up the OOM state
1992 * This has to be called at the end of a page fault if the memcg OOM
1993 * handler was enabled.
1995 * Memcg supports userspace OOM handling where failed allocations must
1996 * sleep on a waitqueue until the userspace task resolves the
1997 * situation. Sleeping directly in the charge context with all kinds
1998 * of locks held is not a good idea, instead we remember an OOM state
1999 * in the task and mem_cgroup_oom_synchronize() has to be called at
2000 * the end of the page fault to complete the OOM handling.
2002 * Returns %true if an ongoing memcg OOM situation was detected and
2003 * completed, %false otherwise.
2005 bool mem_cgroup_oom_synchronize(bool handle)
2007 struct mem_cgroup *memcg = current->memcg_in_oom;
2008 struct oom_wait_info owait;
2011 /* OOM is global, do not handle */
2018 owait.memcg = memcg;
2019 owait.wait.flags = 0;
2020 owait.wait.func = memcg_oom_wake_function;
2021 owait.wait.private = current;
2022 INIT_LIST_HEAD(&owait.wait.entry);
2024 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2025 mem_cgroup_mark_under_oom(memcg);
2027 locked = mem_cgroup_oom_trylock(memcg);
2030 mem_cgroup_oom_notify(memcg);
2032 if (locked && !memcg->oom_kill_disable) {
2033 mem_cgroup_unmark_under_oom(memcg);
2034 finish_wait(&memcg_oom_waitq, &owait.wait);
2035 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
2036 current->memcg_oom_order);
2039 mem_cgroup_unmark_under_oom(memcg);
2040 finish_wait(&memcg_oom_waitq, &owait.wait);
2044 mem_cgroup_oom_unlock(memcg);
2046 * There is no guarantee that an OOM-lock contender
2047 * sees the wakeups triggered by the OOM kill
2048 * uncharges. Wake any sleepers explicitely.
2050 memcg_oom_recover(memcg);
2053 current->memcg_in_oom = NULL;
2054 css_put(&memcg->css);
2059 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2060 * @victim: task to be killed by the OOM killer
2061 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2063 * Returns a pointer to a memory cgroup, which has to be cleaned up
2064 * by killing all belonging OOM-killable tasks.
2066 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2068 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2069 struct mem_cgroup *oom_domain)
2071 struct mem_cgroup *oom_group = NULL;
2072 struct mem_cgroup *memcg;
2074 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2078 oom_domain = root_mem_cgroup;
2082 memcg = mem_cgroup_from_task(victim);
2083 if (memcg == root_mem_cgroup)
2087 * If the victim task has been asynchronously moved to a different
2088 * memory cgroup, we might end up killing tasks outside oom_domain.
2089 * In this case it's better to ignore memory.group.oom.
2091 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2095 * Traverse the memory cgroup hierarchy from the victim task's
2096 * cgroup up to the OOMing cgroup (or root) to find the
2097 * highest-level memory cgroup with oom.group set.
2099 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2100 if (memcg->oom_group)
2103 if (memcg == oom_domain)
2108 css_get(&oom_group->css);
2115 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2117 pr_info("Tasks in ");
2118 pr_cont_cgroup_path(memcg->css.cgroup);
2119 pr_cont(" are going to be killed due to memory.oom.group set\n");
2123 * lock_page_memcg - lock a page->mem_cgroup binding
2126 * This function protects unlocked LRU pages from being moved to
2129 * It ensures lifetime of the returned memcg. Caller is responsible
2130 * for the lifetime of the page; __unlock_page_memcg() is available
2131 * when @page might get freed inside the locked section.
2133 struct mem_cgroup *lock_page_memcg(struct page *page)
2135 struct page *head = compound_head(page); /* rmap on tail pages */
2136 struct mem_cgroup *memcg;
2137 unsigned long flags;
2140 * The RCU lock is held throughout the transaction. The fast
2141 * path can get away without acquiring the memcg->move_lock
2142 * because page moving starts with an RCU grace period.
2144 * The RCU lock also protects the memcg from being freed when
2145 * the page state that is going to change is the only thing
2146 * preventing the page itself from being freed. E.g. writeback
2147 * doesn't hold a page reference and relies on PG_writeback to
2148 * keep off truncation, migration and so forth.
2152 if (mem_cgroup_disabled())
2155 memcg = head->mem_cgroup;
2156 if (unlikely(!memcg))
2159 if (atomic_read(&memcg->moving_account) <= 0)
2162 spin_lock_irqsave(&memcg->move_lock, flags);
2163 if (memcg != head->mem_cgroup) {
2164 spin_unlock_irqrestore(&memcg->move_lock, flags);
2169 * When charge migration first begins, we can have locked and
2170 * unlocked page stat updates happening concurrently. Track
2171 * the task who has the lock for unlock_page_memcg().
2173 memcg->move_lock_task = current;
2174 memcg->move_lock_flags = flags;
2178 EXPORT_SYMBOL(lock_page_memcg);
2181 * __unlock_page_memcg - unlock and unpin a memcg
2184 * Unlock and unpin a memcg returned by lock_page_memcg().
2186 void __unlock_page_memcg(struct mem_cgroup *memcg)
2188 if (memcg && memcg->move_lock_task == current) {
2189 unsigned long flags = memcg->move_lock_flags;
2191 memcg->move_lock_task = NULL;
2192 memcg->move_lock_flags = 0;
2194 spin_unlock_irqrestore(&memcg->move_lock, flags);
2201 * unlock_page_memcg - unlock a page->mem_cgroup binding
2204 void unlock_page_memcg(struct page *page)
2206 struct page *head = compound_head(page);
2208 __unlock_page_memcg(head->mem_cgroup);
2210 EXPORT_SYMBOL(unlock_page_memcg);
2212 struct memcg_stock_pcp {
2214 struct mem_cgroup *cached; /* this never be root cgroup */
2215 unsigned int nr_pages;
2217 #ifdef CONFIG_MEMCG_KMEM
2218 struct obj_cgroup *cached_objcg;
2219 unsigned int nr_bytes;
2222 struct work_struct work;
2223 unsigned long flags;
2224 #define FLUSHING_CACHED_CHARGE 0
2226 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2227 static DEFINE_MUTEX(percpu_charge_mutex);
2229 #ifdef CONFIG_MEMCG_KMEM
2230 static void drain_obj_stock(struct memcg_stock_pcp *stock);
2231 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2232 struct mem_cgroup *root_memcg);
2235 static inline void drain_obj_stock(struct memcg_stock_pcp *stock)
2238 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2239 struct mem_cgroup *root_memcg)
2246 * consume_stock: Try to consume stocked charge on this cpu.
2247 * @memcg: memcg to consume from.
2248 * @nr_pages: how many pages to charge.
2250 * The charges will only happen if @memcg matches the current cpu's memcg
2251 * stock, and at least @nr_pages are available in that stock. Failure to
2252 * service an allocation will refill the stock.
2254 * returns true if successful, false otherwise.
2256 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2258 struct memcg_stock_pcp *stock;
2259 unsigned long flags;
2262 if (nr_pages > MEMCG_CHARGE_BATCH)
2265 local_lock_irqsave(&memcg_stock.lock, flags);
2267 stock = this_cpu_ptr(&memcg_stock);
2268 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2269 stock->nr_pages -= nr_pages;
2273 local_unlock_irqrestore(&memcg_stock.lock, flags);
2279 * Returns stocks cached in percpu and reset cached information.
2281 static void drain_stock(struct memcg_stock_pcp *stock)
2283 struct mem_cgroup *old = stock->cached;
2288 if (stock->nr_pages) {
2289 page_counter_uncharge(&old->memory, stock->nr_pages);
2290 if (do_memsw_account())
2291 page_counter_uncharge(&old->memsw, stock->nr_pages);
2292 stock->nr_pages = 0;
2296 stock->cached = NULL;
2299 static void drain_local_stock(struct work_struct *dummy)
2301 struct memcg_stock_pcp *stock;
2302 unsigned long flags;
2305 * The only protection from memory hotplug vs. drain_stock races is
2306 * that we always operate on local CPU stock here with IRQ disabled
2308 local_lock_irqsave(&memcg_stock.lock, flags);
2310 stock = this_cpu_ptr(&memcg_stock);
2311 drain_obj_stock(stock);
2313 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2315 local_unlock_irqrestore(&memcg_stock.lock, flags);
2319 * Cache charges(val) to local per_cpu area.
2320 * This will be consumed by consume_stock() function, later.
2322 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2324 struct memcg_stock_pcp *stock;
2325 unsigned long flags;
2327 local_lock_irqsave(&memcg_stock.lock, flags);
2329 stock = this_cpu_ptr(&memcg_stock);
2330 if (stock->cached != memcg) { /* reset if necessary */
2332 css_get(&memcg->css);
2333 stock->cached = memcg;
2335 stock->nr_pages += nr_pages;
2337 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2340 local_unlock_irqrestore(&memcg_stock.lock, flags);
2344 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2345 * of the hierarchy under it.
2347 static void drain_all_stock(struct mem_cgroup *root_memcg)
2351 /* If someone's already draining, avoid adding running more workers. */
2352 if (!mutex_trylock(&percpu_charge_mutex))
2355 * Notify other cpus that system-wide "drain" is running
2356 * We do not care about races with the cpu hotplug because cpu down
2357 * as well as workers from this path always operate on the local
2358 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2360 curcpu = get_cpu_light();
2361 for_each_online_cpu(cpu) {
2362 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2363 struct mem_cgroup *memcg;
2367 memcg = stock->cached;
2368 if (memcg && stock->nr_pages &&
2369 mem_cgroup_is_descendant(memcg, root_memcg))
2371 if (obj_stock_flush_required(stock, root_memcg))
2376 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2378 drain_local_stock(&stock->work);
2380 schedule_work_on(cpu, &stock->work);
2384 mutex_unlock(&percpu_charge_mutex);
2387 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2389 struct memcg_stock_pcp *stock;
2390 struct mem_cgroup *memcg, *mi;
2392 stock = &per_cpu(memcg_stock, cpu);
2395 for_each_mem_cgroup(memcg) {
2398 for (i = 0; i < MEMCG_NR_STAT; i++) {
2402 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2404 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2405 atomic_long_add(x, &memcg->vmstats[i]);
2407 if (i >= NR_VM_NODE_STAT_ITEMS)
2410 for_each_node(nid) {
2411 struct mem_cgroup_per_node *pn;
2413 pn = mem_cgroup_nodeinfo(memcg, nid);
2414 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2417 atomic_long_add(x, &pn->lruvec_stat[i]);
2418 } while ((pn = parent_nodeinfo(pn, nid)));
2422 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2425 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2427 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2428 atomic_long_add(x, &memcg->vmevents[i]);
2435 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2436 unsigned int nr_pages,
2439 unsigned long nr_reclaimed = 0;
2442 unsigned long pflags;
2444 if (page_counter_read(&memcg->memory) <=
2445 READ_ONCE(memcg->memory.high))
2448 memcg_memory_event(memcg, MEMCG_HIGH);
2450 psi_memstall_enter(&pflags);
2451 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2453 psi_memstall_leave(&pflags);
2454 } while ((memcg = parent_mem_cgroup(memcg)) &&
2455 !mem_cgroup_is_root(memcg));
2457 return nr_reclaimed;
2460 static void high_work_func(struct work_struct *work)
2462 struct mem_cgroup *memcg;
2464 memcg = container_of(work, struct mem_cgroup, high_work);
2465 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2469 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2470 * enough to still cause a significant slowdown in most cases, while still
2471 * allowing diagnostics and tracing to proceed without becoming stuck.
2473 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2476 * When calculating the delay, we use these either side of the exponentiation to
2477 * maintain precision and scale to a reasonable number of jiffies (see the table
2480 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2481 * overage ratio to a delay.
2482 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2483 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2484 * to produce a reasonable delay curve.
2486 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2487 * reasonable delay curve compared to precision-adjusted overage, not
2488 * penalising heavily at first, but still making sure that growth beyond the
2489 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2490 * example, with a high of 100 megabytes:
2492 * +-------+------------------------+
2493 * | usage | time to allocate in ms |
2494 * +-------+------------------------+
2516 * +-------+------------------------+
2518 #define MEMCG_DELAY_PRECISION_SHIFT 20
2519 #define MEMCG_DELAY_SCALING_SHIFT 14
2521 static u64 calculate_overage(unsigned long usage, unsigned long high)
2529 * Prevent division by 0 in overage calculation by acting as if
2530 * it was a threshold of 1 page
2532 high = max(high, 1UL);
2534 overage = usage - high;
2535 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2536 return div64_u64(overage, high);
2539 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2541 u64 overage, max_overage = 0;
2544 overage = calculate_overage(page_counter_read(&memcg->memory),
2545 READ_ONCE(memcg->memory.high));
2546 max_overage = max(overage, max_overage);
2547 } while ((memcg = parent_mem_cgroup(memcg)) &&
2548 !mem_cgroup_is_root(memcg));
2553 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2555 u64 overage, max_overage = 0;
2558 overage = calculate_overage(page_counter_read(&memcg->swap),
2559 READ_ONCE(memcg->swap.high));
2561 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2562 max_overage = max(overage, max_overage);
2563 } while ((memcg = parent_mem_cgroup(memcg)) &&
2564 !mem_cgroup_is_root(memcg));
2570 * Get the number of jiffies that we should penalise a mischievous cgroup which
2571 * is exceeding its memory.high by checking both it and its ancestors.
2573 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2574 unsigned int nr_pages,
2577 unsigned long penalty_jiffies;
2583 * We use overage compared to memory.high to calculate the number of
2584 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2585 * fairly lenient on small overages, and increasingly harsh when the
2586 * memcg in question makes it clear that it has no intention of stopping
2587 * its crazy behaviour, so we exponentially increase the delay based on
2590 penalty_jiffies = max_overage * max_overage * HZ;
2591 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2592 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2595 * Factor in the task's own contribution to the overage, such that four
2596 * N-sized allocations are throttled approximately the same as one
2597 * 4N-sized allocation.
2599 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2600 * larger the current charge patch is than that.
2602 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2606 * Scheduled by try_charge() to be executed from the userland return path
2607 * and reclaims memory over the high limit.
2609 void mem_cgroup_handle_over_high(void)
2611 unsigned long penalty_jiffies;
2612 unsigned long pflags;
2613 unsigned long nr_reclaimed;
2614 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2615 int nr_retries = MAX_RECLAIM_RETRIES;
2616 struct mem_cgroup *memcg;
2617 bool in_retry = false;
2619 if (likely(!nr_pages))
2622 memcg = get_mem_cgroup_from_mm(current->mm);
2623 current->memcg_nr_pages_over_high = 0;
2627 * The allocating task should reclaim at least the batch size, but for
2628 * subsequent retries we only want to do what's necessary to prevent oom
2629 * or breaching resource isolation.
2631 * This is distinct from memory.max or page allocator behaviour because
2632 * memory.high is currently batched, whereas memory.max and the page
2633 * allocator run every time an allocation is made.
2635 nr_reclaimed = reclaim_high(memcg,
2636 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2640 * memory.high is breached and reclaim is unable to keep up. Throttle
2641 * allocators proactively to slow down excessive growth.
2643 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2644 mem_find_max_overage(memcg));
2646 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2647 swap_find_max_overage(memcg));
2650 * Clamp the max delay per usermode return so as to still keep the
2651 * application moving forwards and also permit diagnostics, albeit
2654 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2657 * Don't sleep if the amount of jiffies this memcg owes us is so low
2658 * that it's not even worth doing, in an attempt to be nice to those who
2659 * go only a small amount over their memory.high value and maybe haven't
2660 * been aggressively reclaimed enough yet.
2662 if (penalty_jiffies <= HZ / 100)
2666 * If reclaim is making forward progress but we're still over
2667 * memory.high, we want to encourage that rather than doing allocator
2670 if (nr_reclaimed || nr_retries--) {
2676 * If we exit early, we're guaranteed to die (since
2677 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2678 * need to account for any ill-begotten jiffies to pay them off later.
2680 psi_memstall_enter(&pflags);
2681 schedule_timeout_killable(penalty_jiffies);
2682 psi_memstall_leave(&pflags);
2685 css_put(&memcg->css);
2688 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2689 unsigned int nr_pages)
2691 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2692 int nr_retries = MAX_RECLAIM_RETRIES;
2693 struct mem_cgroup *mem_over_limit;
2694 struct page_counter *counter;
2695 enum oom_status oom_status;
2696 unsigned long nr_reclaimed;
2697 bool passed_oom = false;
2698 bool may_swap = true;
2699 bool drained = false;
2700 unsigned long pflags;
2702 if (mem_cgroup_is_root(memcg))
2705 if (consume_stock(memcg, nr_pages))
2708 if (!do_memsw_account() ||
2709 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2710 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2712 if (do_memsw_account())
2713 page_counter_uncharge(&memcg->memsw, batch);
2714 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2716 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2720 if (batch > nr_pages) {
2726 * Memcg doesn't have a dedicated reserve for atomic
2727 * allocations. But like the global atomic pool, we need to
2728 * put the burden of reclaim on regular allocation requests
2729 * and let these go through as privileged allocations.
2731 if (gfp_mask & __GFP_ATOMIC)
2735 * Prevent unbounded recursion when reclaim operations need to
2736 * allocate memory. This might exceed the limits temporarily,
2737 * but we prefer facilitating memory reclaim and getting back
2738 * under the limit over triggering OOM kills in these cases.
2740 if (unlikely(current->flags & PF_MEMALLOC))
2743 if (unlikely(task_in_memcg_oom(current)))
2746 if (!gfpflags_allow_blocking(gfp_mask))
2749 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2751 psi_memstall_enter(&pflags);
2752 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2753 gfp_mask, may_swap);
2754 psi_memstall_leave(&pflags);
2756 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2760 drain_all_stock(mem_over_limit);
2765 if (gfp_mask & __GFP_NORETRY)
2768 * Even though the limit is exceeded at this point, reclaim
2769 * may have been able to free some pages. Retry the charge
2770 * before killing the task.
2772 * Only for regular pages, though: huge pages are rather
2773 * unlikely to succeed so close to the limit, and we fall back
2774 * to regular pages anyway in case of failure.
2776 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2779 * At task move, charge accounts can be doubly counted. So, it's
2780 * better to wait until the end of task_move if something is going on.
2782 if (mem_cgroup_wait_acct_move(mem_over_limit))
2788 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2791 if (gfp_mask & __GFP_NOFAIL)
2794 /* Avoid endless loop for tasks bypassed by the oom killer */
2795 if (passed_oom && task_is_dying())
2799 * keep retrying as long as the memcg oom killer is able to make
2800 * a forward progress or bypass the charge if the oom killer
2801 * couldn't make any progress.
2803 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2804 get_order(nr_pages * PAGE_SIZE));
2805 if (oom_status == OOM_SUCCESS) {
2807 nr_retries = MAX_RECLAIM_RETRIES;
2811 if (!(gfp_mask & __GFP_NOFAIL))
2815 * The allocation either can't fail or will lead to more memory
2816 * being freed very soon. Allow memory usage go over the limit
2817 * temporarily by force charging it.
2819 page_counter_charge(&memcg->memory, nr_pages);
2820 if (do_memsw_account())
2821 page_counter_charge(&memcg->memsw, nr_pages);
2826 if (batch > nr_pages)
2827 refill_stock(memcg, batch - nr_pages);
2830 * If the hierarchy is above the normal consumption range, schedule
2831 * reclaim on returning to userland. We can perform reclaim here
2832 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2833 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2834 * not recorded as it most likely matches current's and won't
2835 * change in the meantime. As high limit is checked again before
2836 * reclaim, the cost of mismatch is negligible.
2839 bool mem_high, swap_high;
2841 mem_high = page_counter_read(&memcg->memory) >
2842 READ_ONCE(memcg->memory.high);
2843 swap_high = page_counter_read(&memcg->swap) >
2844 READ_ONCE(memcg->swap.high);
2846 /* Don't bother a random interrupted task */
2847 if (in_interrupt()) {
2849 schedule_work(&memcg->high_work);
2855 if (mem_high || swap_high) {
2857 * The allocating tasks in this cgroup will need to do
2858 * reclaim or be throttled to prevent further growth
2859 * of the memory or swap footprints.
2861 * Target some best-effort fairness between the tasks,
2862 * and distribute reclaim work and delay penalties
2863 * based on how much each task is actually allocating.
2865 current->memcg_nr_pages_over_high += batch;
2866 set_notify_resume(current);
2869 } while ((memcg = parent_mem_cgroup(memcg)));
2874 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2875 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2877 if (mem_cgroup_is_root(memcg))
2880 page_counter_uncharge(&memcg->memory, nr_pages);
2881 if (do_memsw_account())
2882 page_counter_uncharge(&memcg->memsw, nr_pages);
2886 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2888 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2890 * Any of the following ensures page->mem_cgroup stability:
2894 * - lock_page_memcg()
2895 * - exclusive reference
2897 page->mem_cgroup = memcg;
2900 #ifdef CONFIG_MEMCG_KMEM
2902 * The allocated objcg pointers array is not accounted directly.
2903 * Moreover, it should not come from DMA buffer and is not readily
2904 * reclaimable. So those GFP bits should be masked off.
2906 #define OBJCGS_CLEAR_MASK (__GFP_DMA | __GFP_RECLAIMABLE | __GFP_ACCOUNT)
2908 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2911 unsigned int objects = objs_per_slab_page(s, page);
2914 gfp &= ~OBJCGS_CLEAR_MASK;
2915 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2920 if (cmpxchg(&page->obj_cgroups, NULL,
2921 (struct obj_cgroup **) ((unsigned long)vec | 0x1UL)))
2924 kmemleak_not_leak(vec);
2930 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2932 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2933 * cgroup_mutex, etc.
2935 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2939 if (mem_cgroup_disabled())
2942 page = virt_to_head_page(p);
2945 * If page->mem_cgroup is set, it's either a simple mem_cgroup pointer
2946 * or a pointer to obj_cgroup vector. In the latter case the lowest
2947 * bit of the pointer is set.
2948 * The page->mem_cgroup pointer can be asynchronously changed
2949 * from NULL to (obj_cgroup_vec | 0x1UL), but can't be changed
2950 * from a valid memcg pointer to objcg vector or back.
2952 if (!page->mem_cgroup)
2956 * Slab objects are accounted individually, not per-page.
2957 * Memcg membership data for each individual object is saved in
2958 * the page->obj_cgroups.
2960 if (page_has_obj_cgroups(page)) {
2961 struct obj_cgroup *objcg;
2964 off = obj_to_index(page->slab_cache, page, p);
2965 objcg = page_obj_cgroups(page)[off];
2967 return obj_cgroup_memcg(objcg);
2972 /* All other pages use page->mem_cgroup */
2973 return page->mem_cgroup;
2976 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2978 struct obj_cgroup *objcg = NULL;
2979 struct mem_cgroup *memcg;
2981 if (memcg_kmem_bypass())
2985 if (unlikely(active_memcg()))
2986 memcg = active_memcg();
2988 memcg = mem_cgroup_from_task(current);
2990 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2991 objcg = rcu_dereference(memcg->objcg);
2992 if (objcg && obj_cgroup_tryget(objcg))
3001 static int memcg_alloc_cache_id(void)
3006 id = ida_simple_get(&memcg_cache_ida,
3007 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3011 if (id < memcg_nr_cache_ids)
3015 * There's no space for the new id in memcg_caches arrays,
3016 * so we have to grow them.
3018 down_write(&memcg_cache_ids_sem);
3020 size = 2 * (id + 1);
3021 if (size < MEMCG_CACHES_MIN_SIZE)
3022 size = MEMCG_CACHES_MIN_SIZE;
3023 else if (size > MEMCG_CACHES_MAX_SIZE)
3024 size = MEMCG_CACHES_MAX_SIZE;
3026 err = memcg_update_all_list_lrus(size);
3028 memcg_nr_cache_ids = size;
3030 up_write(&memcg_cache_ids_sem);
3033 ida_simple_remove(&memcg_cache_ida, id);
3039 static void memcg_free_cache_id(int id)
3041 ida_simple_remove(&memcg_cache_ida, id);
3045 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
3046 * @memcg: memory cgroup to charge
3047 * @gfp: reclaim mode
3048 * @nr_pages: number of pages to charge
3050 * Returns 0 on success, an error code on failure.
3052 int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
3053 unsigned int nr_pages)
3055 struct page_counter *counter;
3058 ret = try_charge(memcg, gfp, nr_pages);
3062 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3063 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3066 * Enforce __GFP_NOFAIL allocation because callers are not
3067 * prepared to see failures and likely do not have any failure
3070 if (gfp & __GFP_NOFAIL) {
3071 page_counter_charge(&memcg->kmem, nr_pages);
3074 cancel_charge(memcg, nr_pages);
3081 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
3082 * @memcg: memcg to uncharge
3083 * @nr_pages: number of pages to uncharge
3085 void __memcg_kmem_uncharge(struct mem_cgroup *memcg, unsigned int nr_pages)
3087 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3088 page_counter_uncharge(&memcg->kmem, nr_pages);
3090 refill_stock(memcg, nr_pages);
3094 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3095 * @page: page to charge
3096 * @gfp: reclaim mode
3097 * @order: allocation order
3099 * Returns 0 on success, an error code on failure.
3101 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3103 struct mem_cgroup *memcg;
3106 memcg = get_mem_cgroup_from_current();
3107 if (memcg && !mem_cgroup_is_root(memcg)) {
3108 ret = __memcg_kmem_charge(memcg, gfp, 1 << order);
3110 page->mem_cgroup = memcg;
3111 __SetPageKmemcg(page);
3114 css_put(&memcg->css);
3120 * __memcg_kmem_uncharge_page: uncharge a kmem page
3121 * @page: page to uncharge
3122 * @order: allocation order
3124 void __memcg_kmem_uncharge_page(struct page *page, int order)
3126 struct mem_cgroup *memcg = page->mem_cgroup;
3127 unsigned int nr_pages = 1 << order;
3132 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3133 __memcg_kmem_uncharge(memcg, nr_pages);
3134 page->mem_cgroup = NULL;
3135 css_put(&memcg->css);
3137 /* slab pages do not have PageKmemcg flag set */
3138 if (PageKmemcg(page))
3139 __ClearPageKmemcg(page);
3142 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3144 struct memcg_stock_pcp *stock;
3145 unsigned long flags;
3148 local_lock_irqsave(&memcg_stock.lock, flags);
3150 stock = this_cpu_ptr(&memcg_stock);
3151 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3152 stock->nr_bytes -= nr_bytes;
3156 local_unlock_irqrestore(&memcg_stock.lock, flags);
3161 static void drain_obj_stock(struct memcg_stock_pcp *stock)
3163 struct obj_cgroup *old = stock->cached_objcg;
3168 if (stock->nr_bytes) {
3169 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3170 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3173 struct mem_cgroup *memcg;
3177 memcg = obj_cgroup_memcg(old);
3178 if (unlikely(!css_tryget(&memcg->css)))
3182 __memcg_kmem_uncharge(memcg, nr_pages);
3183 css_put(&memcg->css);
3187 * The leftover is flushed to the centralized per-memcg value.
3188 * On the next attempt to refill obj stock it will be moved
3189 * to a per-cpu stock (probably, on an other CPU), see
3190 * refill_obj_stock().
3192 * How often it's flushed is a trade-off between the memory
3193 * limit enforcement accuracy and potential CPU contention,
3194 * so it might be changed in the future.
3196 atomic_add(nr_bytes, &old->nr_charged_bytes);
3197 stock->nr_bytes = 0;
3200 obj_cgroup_put(old);
3201 stock->cached_objcg = NULL;
3204 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3205 struct mem_cgroup *root_memcg)
3207 struct mem_cgroup *memcg;
3209 if (stock->cached_objcg) {
3210 memcg = obj_cgroup_memcg(stock->cached_objcg);
3211 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3218 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3220 struct memcg_stock_pcp *stock;
3221 unsigned long flags;
3223 local_lock_irqsave(&memcg_stock.lock, flags);
3225 stock = this_cpu_ptr(&memcg_stock);
3226 if (stock->cached_objcg != objcg) { /* reset if necessary */
3227 drain_obj_stock(stock);
3228 obj_cgroup_get(objcg);
3229 stock->cached_objcg = objcg;
3230 stock->nr_bytes = atomic_xchg(&objcg->nr_charged_bytes, 0);
3232 stock->nr_bytes += nr_bytes;
3234 if (stock->nr_bytes > PAGE_SIZE)
3235 drain_obj_stock(stock);
3237 local_unlock_irqrestore(&memcg_stock.lock, flags);
3240 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3242 struct mem_cgroup *memcg;
3243 unsigned int nr_pages, nr_bytes;
3246 if (consume_obj_stock(objcg, size))
3250 * In theory, memcg->nr_charged_bytes can have enough
3251 * pre-charged bytes to satisfy the allocation. However,
3252 * flushing memcg->nr_charged_bytes requires two atomic
3253 * operations, and memcg->nr_charged_bytes can't be big,
3254 * so it's better to ignore it and try grab some new pages.
3255 * memcg->nr_charged_bytes will be flushed in
3256 * refill_obj_stock(), called from this function or
3257 * independently later.
3261 memcg = obj_cgroup_memcg(objcg);
3262 if (unlikely(!css_tryget(&memcg->css)))
3266 nr_pages = size >> PAGE_SHIFT;
3267 nr_bytes = size & (PAGE_SIZE - 1);
3272 ret = __memcg_kmem_charge(memcg, gfp, nr_pages);
3273 if (!ret && nr_bytes)
3274 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes);
3276 css_put(&memcg->css);
3280 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3282 refill_obj_stock(objcg, size);
3285 #endif /* CONFIG_MEMCG_KMEM */
3288 * Because head->mem_cgroup is not set on tails, set it now.
3290 void split_page_memcg(struct page *head, unsigned int nr)
3292 struct mem_cgroup *memcg = head->mem_cgroup;
3293 int kmemcg = PageKmemcg(head);
3296 if (mem_cgroup_disabled() || !memcg)
3299 for (i = 1; i < nr; i++) {
3300 head[i].mem_cgroup = memcg;
3302 __SetPageKmemcg(head + i);
3304 css_get_many(&memcg->css, nr - 1);
3307 #ifdef CONFIG_MEMCG_SWAP
3309 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3310 * @entry: swap entry to be moved
3311 * @from: mem_cgroup which the entry is moved from
3312 * @to: mem_cgroup which the entry is moved to
3314 * It succeeds only when the swap_cgroup's record for this entry is the same
3315 * as the mem_cgroup's id of @from.
3317 * Returns 0 on success, -EINVAL on failure.
3319 * The caller must have charged to @to, IOW, called page_counter_charge() about
3320 * both res and memsw, and called css_get().
3322 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3323 struct mem_cgroup *from, struct mem_cgroup *to)
3325 unsigned short old_id, new_id;
3327 old_id = mem_cgroup_id(from);
3328 new_id = mem_cgroup_id(to);
3330 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3331 mod_memcg_state(from, MEMCG_SWAP, -1);
3332 mod_memcg_state(to, MEMCG_SWAP, 1);
3338 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3339 struct mem_cgroup *from, struct mem_cgroup *to)
3345 static DEFINE_MUTEX(memcg_max_mutex);
3347 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3348 unsigned long max, bool memsw)
3350 bool enlarge = false;
3351 bool drained = false;
3353 bool limits_invariant;
3354 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3357 if (signal_pending(current)) {
3362 mutex_lock(&memcg_max_mutex);
3364 * Make sure that the new limit (memsw or memory limit) doesn't
3365 * break our basic invariant rule memory.max <= memsw.max.
3367 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3368 max <= memcg->memsw.max;
3369 if (!limits_invariant) {
3370 mutex_unlock(&memcg_max_mutex);
3374 if (max > counter->max)
3376 ret = page_counter_set_max(counter, max);
3377 mutex_unlock(&memcg_max_mutex);
3383 drain_all_stock(memcg);
3388 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3389 GFP_KERNEL, !memsw)) {
3395 if (!ret && enlarge)
3396 memcg_oom_recover(memcg);
3401 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3403 unsigned long *total_scanned)
3405 unsigned long nr_reclaimed = 0;
3406 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3407 unsigned long reclaimed;
3409 struct mem_cgroup_tree_per_node *mctz;
3410 unsigned long excess;
3411 unsigned long nr_scanned;
3416 mctz = soft_limit_tree_node(pgdat->node_id);
3419 * Do not even bother to check the largest node if the root
3420 * is empty. Do it lockless to prevent lock bouncing. Races
3421 * are acceptable as soft limit is best effort anyway.
3423 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3427 * This loop can run a while, specially if mem_cgroup's continuously
3428 * keep exceeding their soft limit and putting the system under
3435 mz = mem_cgroup_largest_soft_limit_node(mctz);
3440 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3441 gfp_mask, &nr_scanned);
3442 nr_reclaimed += reclaimed;
3443 *total_scanned += nr_scanned;
3444 spin_lock_irq(&mctz->lock);
3445 __mem_cgroup_remove_exceeded(mz, mctz);
3448 * If we failed to reclaim anything from this memory cgroup
3449 * it is time to move on to the next cgroup
3453 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3455 excess = soft_limit_excess(mz->memcg);
3457 * One school of thought says that we should not add
3458 * back the node to the tree if reclaim returns 0.
3459 * But our reclaim could return 0, simply because due
3460 * to priority we are exposing a smaller subset of
3461 * memory to reclaim from. Consider this as a longer
3464 /* If excess == 0, no tree ops */
3465 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3466 spin_unlock_irq(&mctz->lock);
3467 css_put(&mz->memcg->css);
3470 * Could not reclaim anything and there are no more
3471 * mem cgroups to try or we seem to be looping without
3472 * reclaiming anything.
3474 if (!nr_reclaimed &&
3476 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3478 } while (!nr_reclaimed);
3480 css_put(&next_mz->memcg->css);
3481 return nr_reclaimed;
3485 * Test whether @memcg has children, dead or alive. Note that this
3486 * function doesn't care whether @memcg has use_hierarchy enabled and
3487 * returns %true if there are child csses according to the cgroup
3488 * hierarchy. Testing use_hierarchy is the caller's responsibility.
3490 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3495 ret = css_next_child(NULL, &memcg->css);
3501 * Reclaims as many pages from the given memcg as possible.
3503 * Caller is responsible for holding css reference for memcg.
3505 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3507 int nr_retries = MAX_RECLAIM_RETRIES;
3509 /* we call try-to-free pages for make this cgroup empty */
3510 lru_add_drain_all();
3512 drain_all_stock(memcg);
3514 /* try to free all pages in this cgroup */
3515 while (nr_retries && page_counter_read(&memcg->memory)) {
3518 if (signal_pending(current))
3521 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3525 /* maybe some writeback is necessary */
3526 congestion_wait(BLK_RW_ASYNC, HZ/10);
3534 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3535 char *buf, size_t nbytes,
3538 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3540 if (mem_cgroup_is_root(memcg))
3542 return mem_cgroup_force_empty(memcg) ?: nbytes;
3545 #ifdef CONFIG_MEMCG_SWAP
3546 static int mem_cgroup_force_reclaim(struct cgroup_subsys_state *css,
3547 struct cftype *cft, u64 val)
3549 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3550 unsigned long nr_to_reclaim = val;
3551 unsigned long total = 0;
3554 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
3555 total += try_to_free_mem_cgroup_pages(memcg, nr_to_reclaim,
3559 * If nothing was reclaimed after two attempts, there
3560 * may be no reclaimable pages in this hierarchy.
3561 * If more than nr_to_reclaim pages were already reclaimed,
3562 * finish force reclaim.
3564 if (loop && (!total || total > nr_to_reclaim))
3568 pr_info("%s: [Mem_reclaim] Loop: %d - Total_reclaimed: %lu - nr_to_reclaim: %lu\n",
3569 __func__, loop, total, nr_to_reclaim);
3575 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3578 return mem_cgroup_from_css(css)->use_hierarchy;
3581 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3582 struct cftype *cft, u64 val)
3585 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3586 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3588 if (memcg->use_hierarchy == val)
3592 * If parent's use_hierarchy is set, we can't make any modifications
3593 * in the child subtrees. If it is unset, then the change can
3594 * occur, provided the current cgroup has no children.
3596 * For the root cgroup, parent_mem is NULL, we allow value to be
3597 * set if there are no children.
3599 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3600 (val == 1 || val == 0)) {
3601 if (!memcg_has_children(memcg))
3602 memcg->use_hierarchy = val;
3611 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3615 if (mem_cgroup_is_root(memcg)) {
3616 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3617 memcg_page_state(memcg, NR_ANON_MAPPED);
3619 val += memcg_page_state(memcg, MEMCG_SWAP);
3622 val = page_counter_read(&memcg->memory);
3624 val = page_counter_read(&memcg->memsw);
3637 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3640 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3641 struct page_counter *counter;
3643 switch (MEMFILE_TYPE(cft->private)) {
3645 counter = &memcg->memory;
3648 counter = &memcg->memsw;
3651 counter = &memcg->kmem;
3654 counter = &memcg->tcpmem;
3660 switch (MEMFILE_ATTR(cft->private)) {
3662 if (counter == &memcg->memory)
3663 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3664 if (counter == &memcg->memsw)
3665 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3666 return (u64)page_counter_read(counter) * PAGE_SIZE;
3668 return (u64)counter->max * PAGE_SIZE;
3670 return (u64)counter->watermark * PAGE_SIZE;
3672 return counter->failcnt;
3673 case RES_SOFT_LIMIT:
3674 return (u64)memcg->soft_limit * PAGE_SIZE;
3680 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3682 unsigned long stat[MEMCG_NR_STAT] = {0};
3683 struct mem_cgroup *mi;
3686 for_each_online_cpu(cpu)
3687 for (i = 0; i < MEMCG_NR_STAT; i++)
3688 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3690 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3691 for (i = 0; i < MEMCG_NR_STAT; i++)
3692 atomic_long_add(stat[i], &mi->vmstats[i]);
3694 for_each_node(node) {
3695 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3696 struct mem_cgroup_per_node *pi;
3698 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3701 for_each_online_cpu(cpu)
3702 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3704 pn->lruvec_stat_cpu->count[i], cpu);
3706 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3707 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3708 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3712 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3714 unsigned long events[NR_VM_EVENT_ITEMS];
3715 struct mem_cgroup *mi;
3718 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3721 for_each_online_cpu(cpu)
3722 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3723 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3726 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3727 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3728 atomic_long_add(events[i], &mi->vmevents[i]);
3731 #ifdef CONFIG_MEMCG_KMEM
3732 static int memcg_online_kmem(struct mem_cgroup *memcg)
3734 struct obj_cgroup *objcg;
3737 if (cgroup_memory_nokmem)
3740 BUG_ON(memcg->kmemcg_id >= 0);
3741 BUG_ON(memcg->kmem_state);
3743 memcg_id = memcg_alloc_cache_id();
3747 objcg = obj_cgroup_alloc();
3749 memcg_free_cache_id(memcg_id);
3752 objcg->memcg = memcg;
3753 rcu_assign_pointer(memcg->objcg, objcg);
3755 static_branch_enable(&memcg_kmem_enabled_key);
3758 * A memory cgroup is considered kmem-online as soon as it gets
3759 * kmemcg_id. Setting the id after enabling static branching will
3760 * guarantee no one starts accounting before all call sites are
3763 memcg->kmemcg_id = memcg_id;
3764 memcg->kmem_state = KMEM_ONLINE;
3769 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3771 struct cgroup_subsys_state *css;
3772 struct mem_cgroup *parent, *child;
3775 if (memcg->kmem_state != KMEM_ONLINE)
3778 memcg->kmem_state = KMEM_ALLOCATED;
3780 parent = parent_mem_cgroup(memcg);
3782 parent = root_mem_cgroup;
3784 memcg_reparent_objcgs(memcg, parent);
3786 kmemcg_id = memcg->kmemcg_id;
3787 BUG_ON(kmemcg_id < 0);
3790 * Change kmemcg_id of this cgroup and all its descendants to the
3791 * parent's id, and then move all entries from this cgroup's list_lrus
3792 * to ones of the parent. After we have finished, all list_lrus
3793 * corresponding to this cgroup are guaranteed to remain empty. The
3794 * ordering is imposed by list_lru_node->lock taken by
3795 * memcg_drain_all_list_lrus().
3797 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3798 css_for_each_descendant_pre(css, &memcg->css) {
3799 child = mem_cgroup_from_css(css);
3800 BUG_ON(child->kmemcg_id != kmemcg_id);
3801 child->kmemcg_id = parent->kmemcg_id;
3802 if (!memcg->use_hierarchy)
3807 memcg_drain_all_list_lrus(kmemcg_id, parent);
3809 memcg_free_cache_id(kmemcg_id);
3812 static void memcg_free_kmem(struct mem_cgroup *memcg)
3814 /* css_alloc() failed, offlining didn't happen */
3815 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3816 memcg_offline_kmem(memcg);
3819 static int memcg_online_kmem(struct mem_cgroup *memcg)
3823 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3826 static void memcg_free_kmem(struct mem_cgroup *memcg)
3829 #endif /* CONFIG_MEMCG_KMEM */
3831 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3836 mutex_lock(&memcg_max_mutex);
3837 ret = page_counter_set_max(&memcg->kmem, max);
3838 mutex_unlock(&memcg_max_mutex);
3842 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3846 mutex_lock(&memcg_max_mutex);
3848 ret = page_counter_set_max(&memcg->tcpmem, max);
3852 if (!memcg->tcpmem_active) {
3854 * The active flag needs to be written after the static_key
3855 * update. This is what guarantees that the socket activation
3856 * function is the last one to run. See mem_cgroup_sk_alloc()
3857 * for details, and note that we don't mark any socket as
3858 * belonging to this memcg until that flag is up.
3860 * We need to do this, because static_keys will span multiple
3861 * sites, but we can't control their order. If we mark a socket
3862 * as accounted, but the accounting functions are not patched in
3863 * yet, we'll lose accounting.
3865 * We never race with the readers in mem_cgroup_sk_alloc(),
3866 * because when this value change, the code to process it is not
3869 static_branch_inc(&memcg_sockets_enabled_key);
3870 memcg->tcpmem_active = true;
3873 mutex_unlock(&memcg_max_mutex);
3878 * The user of this function is...
3881 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3882 char *buf, size_t nbytes, loff_t off)
3884 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3885 unsigned long nr_pages;
3888 buf = strstrip(buf);
3889 ret = page_counter_memparse(buf, "-1", &nr_pages);
3893 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3895 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3899 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3901 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3904 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3907 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3908 "Please report your usecase to linux-mm@kvack.org if you "
3909 "depend on this functionality.\n");
3910 ret = memcg_update_kmem_max(memcg, nr_pages);
3913 ret = memcg_update_tcp_max(memcg, nr_pages);
3917 case RES_SOFT_LIMIT:
3918 memcg->soft_limit = nr_pages;
3922 return ret ?: nbytes;
3925 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3926 size_t nbytes, loff_t off)
3928 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3929 struct page_counter *counter;
3931 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3933 counter = &memcg->memory;
3936 counter = &memcg->memsw;
3939 counter = &memcg->kmem;
3942 counter = &memcg->tcpmem;
3948 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3950 page_counter_reset_watermark(counter);
3953 counter->failcnt = 0;
3962 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3965 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3969 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3970 struct cftype *cft, u64 val)
3972 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3974 if (val & ~MOVE_MASK)
3978 * No kind of locking is needed in here, because ->can_attach() will
3979 * check this value once in the beginning of the process, and then carry
3980 * on with stale data. This means that changes to this value will only
3981 * affect task migrations starting after the change.
3983 memcg->move_charge_at_immigrate = val;
3987 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3988 struct cftype *cft, u64 val)
3996 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3997 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3998 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
4000 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
4001 int nid, unsigned int lru_mask, bool tree)
4003 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
4004 unsigned long nr = 0;
4007 VM_BUG_ON((unsigned)nid >= nr_node_ids);
4010 if (!(BIT(lru) & lru_mask))
4013 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
4015 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
4020 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
4021 unsigned int lru_mask,
4024 unsigned long nr = 0;
4028 if (!(BIT(lru) & lru_mask))
4031 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
4033 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
4038 static int memcg_numa_stat_show(struct seq_file *m, void *v)
4042 unsigned int lru_mask;
4045 static const struct numa_stat stats[] = {
4046 { "total", LRU_ALL },
4047 { "file", LRU_ALL_FILE },
4048 { "anon", LRU_ALL_ANON },
4049 { "unevictable", BIT(LRU_UNEVICTABLE) },
4051 const struct numa_stat *stat;
4053 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4055 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4056 seq_printf(m, "%s=%lu", stat->name,
4057 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4059 for_each_node_state(nid, N_MEMORY)
4060 seq_printf(m, " N%d=%lu", nid,
4061 mem_cgroup_node_nr_lru_pages(memcg, nid,
4062 stat->lru_mask, false));
4066 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4068 seq_printf(m, "hierarchical_%s=%lu", stat->name,
4069 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4071 for_each_node_state(nid, N_MEMORY)
4072 seq_printf(m, " N%d=%lu", nid,
4073 mem_cgroup_node_nr_lru_pages(memcg, nid,
4074 stat->lru_mask, true));
4080 #endif /* CONFIG_NUMA */
4082 static const unsigned int memcg1_stats[] = {
4085 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4095 static const char *const memcg1_stat_names[] = {
4098 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4108 /* Universal VM events cgroup1 shows, original sort order */
4109 static const unsigned int memcg1_events[] = {
4116 static int memcg_stat_show(struct seq_file *m, void *v)
4118 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4119 unsigned long memory, memsw;
4120 struct mem_cgroup *mi;
4123 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4125 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4128 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4130 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4131 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4132 if (memcg1_stats[i] == NR_ANON_THPS)
4135 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4138 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4139 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4140 memcg_events_local(memcg, memcg1_events[i]));
4142 for (i = 0; i < NR_LRU_LISTS; i++)
4143 seq_printf(m, "%s %lu\n", lru_list_name(i),
4144 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4147 /* Hierarchical information */
4148 memory = memsw = PAGE_COUNTER_MAX;
4149 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4150 memory = min(memory, READ_ONCE(mi->memory.max));
4151 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4153 seq_printf(m, "hierarchical_memory_limit %llu\n",
4154 (u64)memory * PAGE_SIZE);
4155 if (do_memsw_account())
4156 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4157 (u64)memsw * PAGE_SIZE);
4159 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4162 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4164 nr = memcg_page_state(memcg, memcg1_stats[i]);
4165 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4166 if (memcg1_stats[i] == NR_ANON_THPS)
4169 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4170 (u64)nr * PAGE_SIZE);
4173 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4174 seq_printf(m, "total_%s %llu\n",
4175 vm_event_name(memcg1_events[i]),
4176 (u64)memcg_events(memcg, memcg1_events[i]));
4178 for (i = 0; i < NR_LRU_LISTS; i++)
4179 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4180 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4183 #ifdef CONFIG_DEBUG_VM
4186 struct mem_cgroup_per_node *mz;
4187 unsigned long anon_cost = 0;
4188 unsigned long file_cost = 0;
4190 for_each_online_pgdat(pgdat) {
4191 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
4193 anon_cost += mz->lruvec.anon_cost;
4194 file_cost += mz->lruvec.file_cost;
4196 seq_printf(m, "anon_cost %lu\n", anon_cost);
4197 seq_printf(m, "file_cost %lu\n", file_cost);
4204 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4207 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4209 return mem_cgroup_swappiness(memcg);
4212 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4213 struct cftype *cft, u64 val)
4215 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4221 memcg->swappiness = val;
4223 vm_swappiness = val;
4228 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4230 struct mem_cgroup_threshold_ary *t;
4231 unsigned long usage;
4236 t = rcu_dereference(memcg->thresholds.primary);
4238 t = rcu_dereference(memcg->memsw_thresholds.primary);
4243 usage = mem_cgroup_usage(memcg, swap);
4246 * current_threshold points to threshold just below or equal to usage.
4247 * If it's not true, a threshold was crossed after last
4248 * call of __mem_cgroup_threshold().
4250 i = t->current_threshold;
4253 * Iterate backward over array of thresholds starting from
4254 * current_threshold and check if a threshold is crossed.
4255 * If none of thresholds below usage is crossed, we read
4256 * only one element of the array here.
4258 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4259 eventfd_signal(t->entries[i].eventfd, 1);
4261 /* i = current_threshold + 1 */
4265 * Iterate forward over array of thresholds starting from
4266 * current_threshold+1 and check if a threshold is crossed.
4267 * If none of thresholds above usage is crossed, we read
4268 * only one element of the array here.
4270 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4271 eventfd_signal(t->entries[i].eventfd, 1);
4273 /* Update current_threshold */
4274 t->current_threshold = i - 1;
4279 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4282 __mem_cgroup_threshold(memcg, false);
4283 if (do_memsw_account())
4284 __mem_cgroup_threshold(memcg, true);
4286 memcg = parent_mem_cgroup(memcg);
4290 static int compare_thresholds(const void *a, const void *b)
4292 const struct mem_cgroup_threshold *_a = a;
4293 const struct mem_cgroup_threshold *_b = b;
4295 if (_a->threshold > _b->threshold)
4298 if (_a->threshold < _b->threshold)
4304 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4306 struct mem_cgroup_eventfd_list *ev;
4308 spin_lock(&memcg_oom_lock);
4310 list_for_each_entry(ev, &memcg->oom_notify, list)
4311 eventfd_signal(ev->eventfd, 1);
4313 spin_unlock(&memcg_oom_lock);
4317 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4319 struct mem_cgroup *iter;
4321 for_each_mem_cgroup_tree(iter, memcg)
4322 mem_cgroup_oom_notify_cb(iter);
4325 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4326 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4328 struct mem_cgroup_thresholds *thresholds;
4329 struct mem_cgroup_threshold_ary *new;
4330 unsigned long threshold;
4331 unsigned long usage;
4334 ret = page_counter_memparse(args, "-1", &threshold);
4338 mutex_lock(&memcg->thresholds_lock);
4341 thresholds = &memcg->thresholds;
4342 usage = mem_cgroup_usage(memcg, false);
4343 } else if (type == _MEMSWAP) {
4344 thresholds = &memcg->memsw_thresholds;
4345 usage = mem_cgroup_usage(memcg, true);
4349 /* Check if a threshold crossed before adding a new one */
4350 if (thresholds->primary)
4351 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4353 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4355 /* Allocate memory for new array of thresholds */
4356 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4363 /* Copy thresholds (if any) to new array */
4364 if (thresholds->primary)
4365 memcpy(new->entries, thresholds->primary->entries,
4366 flex_array_size(new, entries, size - 1));
4368 /* Add new threshold */
4369 new->entries[size - 1].eventfd = eventfd;
4370 new->entries[size - 1].threshold = threshold;
4372 /* Sort thresholds. Registering of new threshold isn't time-critical */
4373 sort(new->entries, size, sizeof(*new->entries),
4374 compare_thresholds, NULL);
4376 /* Find current threshold */
4377 new->current_threshold = -1;
4378 for (i = 0; i < size; i++) {
4379 if (new->entries[i].threshold <= usage) {
4381 * new->current_threshold will not be used until
4382 * rcu_assign_pointer(), so it's safe to increment
4385 ++new->current_threshold;
4390 /* Free old spare buffer and save old primary buffer as spare */
4391 kfree(thresholds->spare);
4392 thresholds->spare = thresholds->primary;
4394 rcu_assign_pointer(thresholds->primary, new);
4396 /* To be sure that nobody uses thresholds */
4400 mutex_unlock(&memcg->thresholds_lock);
4405 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4406 struct eventfd_ctx *eventfd, const char *args)
4408 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4411 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4412 struct eventfd_ctx *eventfd, const char *args)
4414 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4417 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4418 struct eventfd_ctx *eventfd, enum res_type type)
4420 struct mem_cgroup_thresholds *thresholds;
4421 struct mem_cgroup_threshold_ary *new;
4422 unsigned long usage;
4423 int i, j, size, entries;
4425 mutex_lock(&memcg->thresholds_lock);
4428 thresholds = &memcg->thresholds;
4429 usage = mem_cgroup_usage(memcg, false);
4430 } else if (type == _MEMSWAP) {
4431 thresholds = &memcg->memsw_thresholds;
4432 usage = mem_cgroup_usage(memcg, true);
4436 if (!thresholds->primary)
4439 /* Check if a threshold crossed before removing */
4440 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4442 /* Calculate new number of threshold */
4444 for (i = 0; i < thresholds->primary->size; i++) {
4445 if (thresholds->primary->entries[i].eventfd != eventfd)
4451 new = thresholds->spare;
4453 /* If no items related to eventfd have been cleared, nothing to do */
4457 /* Set thresholds array to NULL if we don't have thresholds */
4466 /* Copy thresholds and find current threshold */
4467 new->current_threshold = -1;
4468 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4469 if (thresholds->primary->entries[i].eventfd == eventfd)
4472 new->entries[j] = thresholds->primary->entries[i];
4473 if (new->entries[j].threshold <= usage) {
4475 * new->current_threshold will not be used
4476 * until rcu_assign_pointer(), so it's safe to increment
4479 ++new->current_threshold;
4485 /* Swap primary and spare array */
4486 thresholds->spare = thresholds->primary;
4488 rcu_assign_pointer(thresholds->primary, new);
4490 /* To be sure that nobody uses thresholds */
4493 /* If all events are unregistered, free the spare array */
4495 kfree(thresholds->spare);
4496 thresholds->spare = NULL;
4499 mutex_unlock(&memcg->thresholds_lock);
4502 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4503 struct eventfd_ctx *eventfd)
4505 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4508 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4509 struct eventfd_ctx *eventfd)
4511 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4514 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4515 struct eventfd_ctx *eventfd, const char *args)
4517 struct mem_cgroup_eventfd_list *event;
4519 event = kmalloc(sizeof(*event), GFP_KERNEL);
4523 spin_lock(&memcg_oom_lock);
4525 event->eventfd = eventfd;
4526 list_add(&event->list, &memcg->oom_notify);
4528 /* already in OOM ? */
4529 if (memcg->under_oom)
4530 eventfd_signal(eventfd, 1);
4531 spin_unlock(&memcg_oom_lock);
4536 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4537 struct eventfd_ctx *eventfd)
4539 struct mem_cgroup_eventfd_list *ev, *tmp;
4541 spin_lock(&memcg_oom_lock);
4543 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4544 if (ev->eventfd == eventfd) {
4545 list_del(&ev->list);
4550 spin_unlock(&memcg_oom_lock);
4553 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4555 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4557 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4558 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4559 seq_printf(sf, "oom_kill %lu\n",
4560 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4564 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4565 struct cftype *cft, u64 val)
4567 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4569 /* cannot set to root cgroup and only 0 and 1 are allowed */
4570 if (!css->parent || !((val == 0) || (val == 1)))
4573 memcg->oom_kill_disable = val;
4575 memcg_oom_recover(memcg);
4580 #ifdef CONFIG_CGROUP_WRITEBACK
4582 #include <trace/events/writeback.h>
4584 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4586 return wb_domain_init(&memcg->cgwb_domain, gfp);
4589 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4591 wb_domain_exit(&memcg->cgwb_domain);
4594 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4596 wb_domain_size_changed(&memcg->cgwb_domain);
4599 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4601 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4603 if (!memcg->css.parent)
4606 return &memcg->cgwb_domain;
4610 * idx can be of type enum memcg_stat_item or node_stat_item.
4611 * Keep in sync with memcg_exact_page().
4613 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4615 long x = atomic_long_read(&memcg->vmstats[idx]);
4618 for_each_online_cpu(cpu)
4619 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4626 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4627 * @wb: bdi_writeback in question
4628 * @pfilepages: out parameter for number of file pages
4629 * @pheadroom: out parameter for number of allocatable pages according to memcg
4630 * @pdirty: out parameter for number of dirty pages
4631 * @pwriteback: out parameter for number of pages under writeback
4633 * Determine the numbers of file, headroom, dirty, and writeback pages in
4634 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4635 * is a bit more involved.
4637 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4638 * headroom is calculated as the lowest headroom of itself and the
4639 * ancestors. Note that this doesn't consider the actual amount of
4640 * available memory in the system. The caller should further cap
4641 * *@pheadroom accordingly.
4643 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4644 unsigned long *pheadroom, unsigned long *pdirty,
4645 unsigned long *pwriteback)
4647 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4648 struct mem_cgroup *parent;
4650 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4652 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4653 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4654 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4655 *pheadroom = PAGE_COUNTER_MAX;
4657 while ((parent = parent_mem_cgroup(memcg))) {
4658 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4659 READ_ONCE(memcg->memory.high));
4660 unsigned long used = page_counter_read(&memcg->memory);
4662 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4668 * Foreign dirty flushing
4670 * There's an inherent mismatch between memcg and writeback. The former
4671 * trackes ownership per-page while the latter per-inode. This was a
4672 * deliberate design decision because honoring per-page ownership in the
4673 * writeback path is complicated, may lead to higher CPU and IO overheads
4674 * and deemed unnecessary given that write-sharing an inode across
4675 * different cgroups isn't a common use-case.
4677 * Combined with inode majority-writer ownership switching, this works well
4678 * enough in most cases but there are some pathological cases. For
4679 * example, let's say there are two cgroups A and B which keep writing to
4680 * different but confined parts of the same inode. B owns the inode and
4681 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4682 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4683 * triggering background writeback. A will be slowed down without a way to
4684 * make writeback of the dirty pages happen.
4686 * Conditions like the above can lead to a cgroup getting repatedly and
4687 * severely throttled after making some progress after each
4688 * dirty_expire_interval while the underyling IO device is almost
4691 * Solving this problem completely requires matching the ownership tracking
4692 * granularities between memcg and writeback in either direction. However,
4693 * the more egregious behaviors can be avoided by simply remembering the
4694 * most recent foreign dirtying events and initiating remote flushes on
4695 * them when local writeback isn't enough to keep the memory clean enough.
4697 * The following two functions implement such mechanism. When a foreign
4698 * page - a page whose memcg and writeback ownerships don't match - is
4699 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4700 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4701 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4702 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4703 * foreign bdi_writebacks which haven't expired. Both the numbers of
4704 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4705 * limited to MEMCG_CGWB_FRN_CNT.
4707 * The mechanism only remembers IDs and doesn't hold any object references.
4708 * As being wrong occasionally doesn't matter, updates and accesses to the
4709 * records are lockless and racy.
4711 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4712 struct bdi_writeback *wb)
4714 struct mem_cgroup *memcg = page->mem_cgroup;
4715 struct memcg_cgwb_frn *frn;
4716 u64 now = get_jiffies_64();
4717 u64 oldest_at = now;
4721 trace_track_foreign_dirty(page, wb);
4724 * Pick the slot to use. If there is already a slot for @wb, keep
4725 * using it. If not replace the oldest one which isn't being
4728 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4729 frn = &memcg->cgwb_frn[i];
4730 if (frn->bdi_id == wb->bdi->id &&
4731 frn->memcg_id == wb->memcg_css->id)
4733 if (time_before64(frn->at, oldest_at) &&
4734 atomic_read(&frn->done.cnt) == 1) {
4736 oldest_at = frn->at;
4740 if (i < MEMCG_CGWB_FRN_CNT) {
4742 * Re-using an existing one. Update timestamp lazily to
4743 * avoid making the cacheline hot. We want them to be
4744 * reasonably up-to-date and significantly shorter than
4745 * dirty_expire_interval as that's what expires the record.
4746 * Use the shorter of 1s and dirty_expire_interval / 8.
4748 unsigned long update_intv =
4749 min_t(unsigned long, HZ,
4750 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4752 if (time_before64(frn->at, now - update_intv))
4754 } else if (oldest >= 0) {
4755 /* replace the oldest free one */
4756 frn = &memcg->cgwb_frn[oldest];
4757 frn->bdi_id = wb->bdi->id;
4758 frn->memcg_id = wb->memcg_css->id;
4763 /* issue foreign writeback flushes for recorded foreign dirtying events */
4764 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4766 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4767 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4768 u64 now = jiffies_64;
4771 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4772 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4775 * If the record is older than dirty_expire_interval,
4776 * writeback on it has already started. No need to kick it
4777 * off again. Also, don't start a new one if there's
4778 * already one in flight.
4780 if (time_after64(frn->at, now - intv) &&
4781 atomic_read(&frn->done.cnt) == 1) {
4783 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4784 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4785 WB_REASON_FOREIGN_FLUSH,
4791 #else /* CONFIG_CGROUP_WRITEBACK */
4793 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4798 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4802 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4806 #endif /* CONFIG_CGROUP_WRITEBACK */
4809 * DO NOT USE IN NEW FILES.
4811 * "cgroup.event_control" implementation.
4813 * This is way over-engineered. It tries to support fully configurable
4814 * events for each user. Such level of flexibility is completely
4815 * unnecessary especially in the light of the planned unified hierarchy.
4817 * Please deprecate this and replace with something simpler if at all
4822 * Unregister event and free resources.
4824 * Gets called from workqueue.
4826 static void memcg_event_remove(struct work_struct *work)
4828 struct mem_cgroup_event *event =
4829 container_of(work, struct mem_cgroup_event, remove);
4830 struct mem_cgroup *memcg = event->memcg;
4832 remove_wait_queue(event->wqh, &event->wait);
4834 event->unregister_event(memcg, event->eventfd);
4836 /* Notify userspace the event is going away. */
4837 eventfd_signal(event->eventfd, 1);
4839 eventfd_ctx_put(event->eventfd);
4841 css_put(&memcg->css);
4845 * Gets called on EPOLLHUP on eventfd when user closes it.
4847 * Called with wqh->lock held and interrupts disabled.
4849 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4850 int sync, void *key)
4852 struct mem_cgroup_event *event =
4853 container_of(wait, struct mem_cgroup_event, wait);
4854 struct mem_cgroup *memcg = event->memcg;
4855 __poll_t flags = key_to_poll(key);
4857 if (flags & EPOLLHUP) {
4859 * If the event has been detached at cgroup removal, we
4860 * can simply return knowing the other side will cleanup
4863 * We can't race against event freeing since the other
4864 * side will require wqh->lock via remove_wait_queue(),
4867 spin_lock(&memcg->event_list_lock);
4868 if (!list_empty(&event->list)) {
4869 list_del_init(&event->list);
4871 * We are in atomic context, but cgroup_event_remove()
4872 * may sleep, so we have to call it in workqueue.
4874 schedule_work(&event->remove);
4876 spin_unlock(&memcg->event_list_lock);
4882 static void memcg_event_ptable_queue_proc(struct file *file,
4883 wait_queue_head_t *wqh, poll_table *pt)
4885 struct mem_cgroup_event *event =
4886 container_of(pt, struct mem_cgroup_event, pt);
4889 add_wait_queue(wqh, &event->wait);
4893 * DO NOT USE IN NEW FILES.
4895 * Parse input and register new cgroup event handler.
4897 * Input must be in format '<event_fd> <control_fd> <args>'.
4898 * Interpretation of args is defined by control file implementation.
4900 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4901 char *buf, size_t nbytes, loff_t off)
4903 struct cgroup_subsys_state *css = of_css(of);
4904 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4905 struct mem_cgroup_event *event;
4906 struct cgroup_subsys_state *cfile_css;
4907 unsigned int efd, cfd;
4914 buf = strstrip(buf);
4916 efd = simple_strtoul(buf, &endp, 10);
4921 cfd = simple_strtoul(buf, &endp, 10);
4922 if ((*endp != ' ') && (*endp != '\0'))
4926 event = kzalloc(sizeof(*event), GFP_KERNEL);
4930 event->memcg = memcg;
4931 INIT_LIST_HEAD(&event->list);
4932 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4933 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4934 INIT_WORK(&event->remove, memcg_event_remove);
4942 event->eventfd = eventfd_ctx_fileget(efile.file);
4943 if (IS_ERR(event->eventfd)) {
4944 ret = PTR_ERR(event->eventfd);
4951 goto out_put_eventfd;
4954 /* the process need read permission on control file */
4955 /* AV: shouldn't we check that it's been opened for read instead? */
4956 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4961 * Determine the event callbacks and set them in @event. This used
4962 * to be done via struct cftype but cgroup core no longer knows
4963 * about these events. The following is crude but the whole thing
4964 * is for compatibility anyway.
4966 * DO NOT ADD NEW FILES.
4968 name = cfile.file->f_path.dentry->d_name.name;
4970 if (!strcmp(name, "memory.usage_in_bytes")) {
4971 event->register_event = mem_cgroup_usage_register_event;
4972 event->unregister_event = mem_cgroup_usage_unregister_event;
4973 } else if (!strcmp(name, "memory.oom_control")) {
4974 event->register_event = mem_cgroup_oom_register_event;
4975 event->unregister_event = mem_cgroup_oom_unregister_event;
4976 } else if (!strcmp(name, "memory.pressure_level")) {
4977 event->register_event = vmpressure_register_event;
4978 event->unregister_event = vmpressure_unregister_event;
4979 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4980 event->register_event = memsw_cgroup_usage_register_event;
4981 event->unregister_event = memsw_cgroup_usage_unregister_event;
4988 * Verify @cfile should belong to @css. Also, remaining events are
4989 * automatically removed on cgroup destruction but the removal is
4990 * asynchronous, so take an extra ref on @css.
4992 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4993 &memory_cgrp_subsys);
4995 if (IS_ERR(cfile_css))
4997 if (cfile_css != css) {
5002 ret = event->register_event(memcg, event->eventfd, buf);
5006 vfs_poll(efile.file, &event->pt);
5008 spin_lock(&memcg->event_list_lock);
5009 list_add(&event->list, &memcg->event_list);
5010 spin_unlock(&memcg->event_list_lock);
5022 eventfd_ctx_put(event->eventfd);
5031 static struct cftype mem_cgroup_legacy_files[] = {
5033 .name = "usage_in_bytes",
5034 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5035 .read_u64 = mem_cgroup_read_u64,
5038 .name = "max_usage_in_bytes",
5039 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5040 .write = mem_cgroup_reset,
5041 .read_u64 = mem_cgroup_read_u64,
5044 .name = "limit_in_bytes",
5045 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5046 .write = mem_cgroup_write,
5047 .read_u64 = mem_cgroup_read_u64,
5050 .name = "soft_limit_in_bytes",
5051 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5052 .write = mem_cgroup_write,
5053 .read_u64 = mem_cgroup_read_u64,
5057 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5058 .write = mem_cgroup_reset,
5059 .read_u64 = mem_cgroup_read_u64,
5063 .seq_show = memcg_stat_show,
5066 .name = "force_empty",
5067 .write = mem_cgroup_force_empty_write,
5070 .name = "use_hierarchy",
5071 .write_u64 = mem_cgroup_hierarchy_write,
5072 .read_u64 = mem_cgroup_hierarchy_read,
5075 .name = "cgroup.event_control", /* XXX: for compat */
5076 .write = memcg_write_event_control,
5077 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
5080 .name = "swappiness",
5081 .read_u64 = mem_cgroup_swappiness_read,
5082 .write_u64 = mem_cgroup_swappiness_write,
5085 .name = "move_charge_at_immigrate",
5086 .read_u64 = mem_cgroup_move_charge_read,
5087 .write_u64 = mem_cgroup_move_charge_write,
5090 .name = "oom_control",
5091 .seq_show = mem_cgroup_oom_control_read,
5092 .write_u64 = mem_cgroup_oom_control_write,
5093 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5096 .name = "pressure_level",
5100 .name = "numa_stat",
5101 .seq_show = memcg_numa_stat_show,
5105 .name = "kmem.limit_in_bytes",
5106 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5107 .write = mem_cgroup_write,
5108 .read_u64 = mem_cgroup_read_u64,
5111 .name = "kmem.usage_in_bytes",
5112 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5113 .read_u64 = mem_cgroup_read_u64,
5116 .name = "kmem.failcnt",
5117 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5118 .write = mem_cgroup_reset,
5119 .read_u64 = mem_cgroup_read_u64,
5122 .name = "kmem.max_usage_in_bytes",
5123 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5124 .write = mem_cgroup_reset,
5125 .read_u64 = mem_cgroup_read_u64,
5127 #if defined(CONFIG_MEMCG_KMEM) && \
5128 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5130 .name = "kmem.slabinfo",
5131 .seq_show = memcg_slab_show,
5135 .name = "kmem.tcp.limit_in_bytes",
5136 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5137 .write = mem_cgroup_write,
5138 .read_u64 = mem_cgroup_read_u64,
5141 .name = "kmem.tcp.usage_in_bytes",
5142 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5143 .read_u64 = mem_cgroup_read_u64,
5146 .name = "kmem.tcp.failcnt",
5147 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5148 .write = mem_cgroup_reset,
5149 .read_u64 = mem_cgroup_read_u64,
5152 .name = "kmem.tcp.max_usage_in_bytes",
5153 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5154 .write = mem_cgroup_reset,
5155 .read_u64 = mem_cgroup_read_u64,
5157 { }, /* terminate */
5161 * Private memory cgroup IDR
5163 * Swap-out records and page cache shadow entries need to store memcg
5164 * references in constrained space, so we maintain an ID space that is
5165 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5166 * memory-controlled cgroups to 64k.
5168 * However, there usually are many references to the offline CSS after
5169 * the cgroup has been destroyed, such as page cache or reclaimable
5170 * slab objects, that don't need to hang on to the ID. We want to keep
5171 * those dead CSS from occupying IDs, or we might quickly exhaust the
5172 * relatively small ID space and prevent the creation of new cgroups
5173 * even when there are much fewer than 64k cgroups - possibly none.
5175 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5176 * be freed and recycled when it's no longer needed, which is usually
5177 * when the CSS is offlined.
5179 * The only exception to that are records of swapped out tmpfs/shmem
5180 * pages that need to be attributed to live ancestors on swapin. But
5181 * those references are manageable from userspace.
5184 static DEFINE_IDR(mem_cgroup_idr);
5186 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5188 if (memcg->id.id > 0) {
5189 idr_remove(&mem_cgroup_idr, memcg->id.id);
5194 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5197 refcount_add(n, &memcg->id.ref);
5200 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5202 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5203 mem_cgroup_id_remove(memcg);
5205 /* Memcg ID pins CSS */
5206 css_put(&memcg->css);
5210 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5212 mem_cgroup_id_put_many(memcg, 1);
5216 * mem_cgroup_from_id - look up a memcg from a memcg id
5217 * @id: the memcg id to look up
5219 * Caller must hold rcu_read_lock().
5221 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5223 WARN_ON_ONCE(!rcu_read_lock_held());
5224 return idr_find(&mem_cgroup_idr, id);
5227 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5229 struct mem_cgroup_per_node *pn;
5232 * This routine is called against possible nodes.
5233 * But it's BUG to call kmalloc() against offline node.
5235 * TODO: this routine can waste much memory for nodes which will
5236 * never be onlined. It's better to use memory hotplug callback
5239 if (!node_state(node, N_NORMAL_MEMORY))
5241 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5245 pn->lruvec_stat_local = alloc_percpu_gfp(struct lruvec_stat,
5246 GFP_KERNEL_ACCOUNT);
5247 if (!pn->lruvec_stat_local) {
5252 pn->lruvec_stat_cpu = alloc_percpu_gfp(struct lruvec_stat,
5253 GFP_KERNEL_ACCOUNT);
5254 if (!pn->lruvec_stat_cpu) {
5255 free_percpu(pn->lruvec_stat_local);
5260 lruvec_init(&pn->lruvec);
5261 pn->usage_in_excess = 0;
5262 pn->on_tree = false;
5265 memcg->nodeinfo[node] = pn;
5269 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5271 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5276 free_percpu(pn->lruvec_stat_cpu);
5277 free_percpu(pn->lruvec_stat_local);
5281 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5286 free_mem_cgroup_per_node_info(memcg, node);
5287 free_percpu(memcg->vmstats_percpu);
5288 free_percpu(memcg->vmstats_local);
5292 static void mem_cgroup_free(struct mem_cgroup *memcg)
5294 memcg_wb_domain_exit(memcg);
5296 * Flush percpu vmstats and vmevents to guarantee the value correctness
5297 * on parent's and all ancestor levels.
5299 memcg_flush_percpu_vmstats(memcg);
5300 memcg_flush_percpu_vmevents(memcg);
5301 __mem_cgroup_free(memcg);
5304 static struct mem_cgroup *mem_cgroup_alloc(void)
5306 struct mem_cgroup *memcg;
5309 int __maybe_unused i;
5310 long error = -ENOMEM;
5312 size = sizeof(struct mem_cgroup);
5313 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5315 memcg = kzalloc(size, GFP_KERNEL);
5317 return ERR_PTR(error);
5319 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5320 1, MEM_CGROUP_ID_MAX,
5322 if (memcg->id.id < 0) {
5323 error = memcg->id.id;
5327 memcg->vmstats_local = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5328 GFP_KERNEL_ACCOUNT);
5329 if (!memcg->vmstats_local)
5332 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5333 GFP_KERNEL_ACCOUNT);
5334 if (!memcg->vmstats_percpu)
5338 if (alloc_mem_cgroup_per_node_info(memcg, node))
5341 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5344 INIT_WORK(&memcg->high_work, high_work_func);
5345 INIT_LIST_HEAD(&memcg->oom_notify);
5346 mutex_init(&memcg->thresholds_lock);
5347 spin_lock_init(&memcg->move_lock);
5348 vmpressure_init(&memcg->vmpressure);
5349 INIT_LIST_HEAD(&memcg->event_list);
5350 spin_lock_init(&memcg->event_list_lock);
5351 memcg->socket_pressure = jiffies;
5352 #ifdef CONFIG_MEMCG_KMEM
5353 memcg->kmemcg_id = -1;
5354 INIT_LIST_HEAD(&memcg->objcg_list);
5356 #ifdef CONFIG_CGROUP_WRITEBACK
5357 INIT_LIST_HEAD(&memcg->cgwb_list);
5358 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5359 memcg->cgwb_frn[i].done =
5360 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5362 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5363 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5364 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5365 memcg->deferred_split_queue.split_queue_len = 0;
5367 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5370 mem_cgroup_id_remove(memcg);
5371 __mem_cgroup_free(memcg);
5372 return ERR_PTR(error);
5375 static struct cgroup_subsys_state * __ref
5376 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5378 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5379 struct mem_cgroup *memcg, *old_memcg;
5380 long error = -ENOMEM;
5382 old_memcg = set_active_memcg(parent);
5383 memcg = mem_cgroup_alloc();
5384 set_active_memcg(old_memcg);
5386 return ERR_CAST(memcg);
5388 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5389 memcg->soft_limit = PAGE_COUNTER_MAX;
5390 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5392 memcg->swappiness = mem_cgroup_swappiness(parent);
5393 memcg->oom_kill_disable = parent->oom_kill_disable;
5396 page_counter_init(&memcg->memory, NULL);
5397 page_counter_init(&memcg->swap, NULL);
5398 page_counter_init(&memcg->kmem, NULL);
5399 page_counter_init(&memcg->tcpmem, NULL);
5400 } else if (parent->use_hierarchy) {
5401 memcg->use_hierarchy = true;
5402 page_counter_init(&memcg->memory, &parent->memory);
5403 page_counter_init(&memcg->swap, &parent->swap);
5404 page_counter_init(&memcg->kmem, &parent->kmem);
5405 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5407 page_counter_init(&memcg->memory, &root_mem_cgroup->memory);
5408 page_counter_init(&memcg->swap, &root_mem_cgroup->swap);
5409 page_counter_init(&memcg->kmem, &root_mem_cgroup->kmem);
5410 page_counter_init(&memcg->tcpmem, &root_mem_cgroup->tcpmem);
5412 * Deeper hierachy with use_hierarchy == false doesn't make
5413 * much sense so let cgroup subsystem know about this
5414 * unfortunate state in our controller.
5416 if (parent != root_mem_cgroup)
5417 memory_cgrp_subsys.broken_hierarchy = true;
5420 /* The following stuff does not apply to the root */
5422 root_mem_cgroup = memcg;
5426 error = memcg_online_kmem(memcg);
5430 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5431 static_branch_inc(&memcg_sockets_enabled_key);
5435 mem_cgroup_id_remove(memcg);
5436 mem_cgroup_free(memcg);
5437 return ERR_PTR(error);
5440 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5442 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5445 * A memcg must be visible for memcg_expand_shrinker_maps()
5446 * by the time the maps are allocated. So, we allocate maps
5447 * here, when for_each_mem_cgroup() can't skip it.
5449 if (memcg_alloc_shrinker_maps(memcg)) {
5450 mem_cgroup_id_remove(memcg);
5454 /* Online state pins memcg ID, memcg ID pins CSS */
5455 refcount_set(&memcg->id.ref, 1);
5460 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5462 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5463 struct mem_cgroup_event *event, *tmp;
5466 * Unregister events and notify userspace.
5467 * Notify userspace about cgroup removing only after rmdir of cgroup
5468 * directory to avoid race between userspace and kernelspace.
5470 spin_lock(&memcg->event_list_lock);
5471 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5472 list_del_init(&event->list);
5473 schedule_work(&event->remove);
5475 spin_unlock(&memcg->event_list_lock);
5477 page_counter_set_min(&memcg->memory, 0);
5478 page_counter_set_low(&memcg->memory, 0);
5480 memcg_offline_kmem(memcg);
5481 wb_memcg_offline(memcg);
5483 drain_all_stock(memcg);
5485 mem_cgroup_id_put(memcg);
5488 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5490 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5492 invalidate_reclaim_iterators(memcg);
5495 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5497 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5498 int __maybe_unused i;
5500 #ifdef CONFIG_CGROUP_WRITEBACK
5501 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5502 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5504 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5505 static_branch_dec(&memcg_sockets_enabled_key);
5507 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5508 static_branch_dec(&memcg_sockets_enabled_key);
5510 vmpressure_cleanup(&memcg->vmpressure);
5511 cancel_work_sync(&memcg->high_work);
5512 mem_cgroup_remove_from_trees(memcg);
5513 memcg_free_shrinker_maps(memcg);
5514 memcg_free_kmem(memcg);
5515 mem_cgroup_free(memcg);
5519 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5520 * @css: the target css
5522 * Reset the states of the mem_cgroup associated with @css. This is
5523 * invoked when the userland requests disabling on the default hierarchy
5524 * but the memcg is pinned through dependency. The memcg should stop
5525 * applying policies and should revert to the vanilla state as it may be
5526 * made visible again.
5528 * The current implementation only resets the essential configurations.
5529 * This needs to be expanded to cover all the visible parts.
5531 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5533 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5535 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5536 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5537 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5538 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5539 page_counter_set_min(&memcg->memory, 0);
5540 page_counter_set_low(&memcg->memory, 0);
5541 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5542 memcg->soft_limit = PAGE_COUNTER_MAX;
5543 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5544 memcg_wb_domain_size_changed(memcg);
5548 /* Handlers for move charge at task migration. */
5549 static int mem_cgroup_do_precharge(unsigned long count)
5553 /* Try a single bulk charge without reclaim first, kswapd may wake */
5554 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5556 mc.precharge += count;
5560 /* Try charges one by one with reclaim, but do not retry */
5562 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5576 enum mc_target_type {
5583 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5584 unsigned long addr, pte_t ptent)
5586 struct page *page = vm_normal_page(vma, addr, ptent);
5588 if (!page || !page_mapped(page))
5590 if (PageAnon(page)) {
5591 if (!(mc.flags & MOVE_ANON))
5594 if (!(mc.flags & MOVE_FILE))
5597 if (!get_page_unless_zero(page))
5603 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5604 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5605 pte_t ptent, swp_entry_t *entry)
5607 struct page *page = NULL;
5608 swp_entry_t ent = pte_to_swp_entry(ptent);
5610 if (!(mc.flags & MOVE_ANON))
5614 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5615 * a device and because they are not accessible by CPU they are store
5616 * as special swap entry in the CPU page table.
5618 if (is_device_private_entry(ent)) {
5619 page = device_private_entry_to_page(ent);
5621 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5622 * a refcount of 1 when free (unlike normal page)
5624 if (!page_ref_add_unless(page, 1, 1))
5629 if (non_swap_entry(ent))
5633 * Because lookup_swap_cache() updates some statistics counter,
5634 * we call find_get_page() with swapper_space directly.
5636 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5637 entry->val = ent.val;
5642 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5643 pte_t ptent, swp_entry_t *entry)
5649 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5650 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5652 if (!vma->vm_file) /* anonymous vma */
5654 if (!(mc.flags & MOVE_FILE))
5657 /* page is moved even if it's not RSS of this task(page-faulted). */
5658 /* shmem/tmpfs may report page out on swap: account for that too. */
5659 return find_get_incore_page(vma->vm_file->f_mapping,
5660 linear_page_index(vma, addr));
5664 * mem_cgroup_move_account - move account of the page
5666 * @compound: charge the page as compound or small page
5667 * @from: mem_cgroup which the page is moved from.
5668 * @to: mem_cgroup which the page is moved to. @from != @to.
5670 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5672 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5675 static int mem_cgroup_move_account(struct page *page,
5677 struct mem_cgroup *from,
5678 struct mem_cgroup *to)
5680 struct lruvec *from_vec, *to_vec;
5681 struct pglist_data *pgdat;
5682 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5685 VM_BUG_ON(from == to);
5686 VM_BUG_ON_PAGE(PageLRU(page), page);
5687 VM_BUG_ON(compound && !PageTransHuge(page));
5690 * Prevent mem_cgroup_migrate() from looking at
5691 * page->mem_cgroup of its source page while we change it.
5694 if (!trylock_page(page))
5698 if (page->mem_cgroup != from)
5701 pgdat = page_pgdat(page);
5702 from_vec = mem_cgroup_lruvec(from, pgdat);
5703 to_vec = mem_cgroup_lruvec(to, pgdat);
5705 lock_page_memcg(page);
5707 if (PageAnon(page)) {
5708 if (page_mapped(page)) {
5709 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5710 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5711 if (PageTransHuge(page)) {
5712 __dec_lruvec_state(from_vec, NR_ANON_THPS);
5713 __inc_lruvec_state(to_vec, NR_ANON_THPS);
5718 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5719 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5721 if (PageSwapBacked(page)) {
5722 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5723 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5726 if (page_mapped(page)) {
5727 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5728 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5731 if (PageDirty(page)) {
5732 struct address_space *mapping = page_mapping(page);
5734 if (mapping_can_writeback(mapping)) {
5735 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5737 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5743 if (PageWriteback(page)) {
5744 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5745 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5749 * All state has been migrated, let's switch to the new memcg.
5751 * It is safe to change page->mem_cgroup here because the page
5752 * is referenced, charged, isolated, and locked: we can't race
5753 * with (un)charging, migration, LRU putback, or anything else
5754 * that would rely on a stable page->mem_cgroup.
5756 * Note that lock_page_memcg is a memcg lock, not a page lock,
5757 * to save space. As soon as we switch page->mem_cgroup to a
5758 * new memcg that isn't locked, the above state can change
5759 * concurrently again. Make sure we're truly done with it.
5764 css_put(&from->css);
5766 page->mem_cgroup = to;
5768 __unlock_page_memcg(from);
5772 local_lock_irq(&event_lock.l);
5773 mem_cgroup_charge_statistics(to, page, nr_pages);
5774 memcg_check_events(to, page);
5775 mem_cgroup_charge_statistics(from, page, -nr_pages);
5776 memcg_check_events(from, page);
5777 local_unlock_irq(&event_lock.l);
5785 * get_mctgt_type - get target type of moving charge
5786 * @vma: the vma the pte to be checked belongs
5787 * @addr: the address corresponding to the pte to be checked
5788 * @ptent: the pte to be checked
5789 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5792 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5793 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5794 * move charge. if @target is not NULL, the page is stored in target->page
5795 * with extra refcnt got(Callers should handle it).
5796 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5797 * target for charge migration. if @target is not NULL, the entry is stored
5799 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5800 * (so ZONE_DEVICE page and thus not on the lru).
5801 * For now we such page is charge like a regular page would be as for all
5802 * intent and purposes it is just special memory taking the place of a
5805 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5807 * Called with pte lock held.
5810 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5811 unsigned long addr, pte_t ptent, union mc_target *target)
5813 struct page *page = NULL;
5814 enum mc_target_type ret = MC_TARGET_NONE;
5815 swp_entry_t ent = { .val = 0 };
5817 if (pte_present(ptent))
5818 page = mc_handle_present_pte(vma, addr, ptent);
5819 else if (is_swap_pte(ptent))
5820 page = mc_handle_swap_pte(vma, ptent, &ent);
5821 else if (pte_none(ptent))
5822 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5824 if (!page && !ent.val)
5828 * Do only loose check w/o serialization.
5829 * mem_cgroup_move_account() checks the page is valid or
5830 * not under LRU exclusion.
5832 if (page->mem_cgroup == mc.from) {
5833 ret = MC_TARGET_PAGE;
5834 if (is_device_private_page(page))
5835 ret = MC_TARGET_DEVICE;
5837 target->page = page;
5839 if (!ret || !target)
5843 * There is a swap entry and a page doesn't exist or isn't charged.
5844 * But we cannot move a tail-page in a THP.
5846 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5847 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5848 ret = MC_TARGET_SWAP;
5855 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5857 * We don't consider PMD mapped swapping or file mapped pages because THP does
5858 * not support them for now.
5859 * Caller should make sure that pmd_trans_huge(pmd) is true.
5861 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5862 unsigned long addr, pmd_t pmd, union mc_target *target)
5864 struct page *page = NULL;
5865 enum mc_target_type ret = MC_TARGET_NONE;
5867 if (unlikely(is_swap_pmd(pmd))) {
5868 VM_BUG_ON(thp_migration_supported() &&
5869 !is_pmd_migration_entry(pmd));
5872 page = pmd_page(pmd);
5873 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5874 if (!(mc.flags & MOVE_ANON))
5876 if (page->mem_cgroup == mc.from) {
5877 ret = MC_TARGET_PAGE;
5880 target->page = page;
5886 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5887 unsigned long addr, pmd_t pmd, union mc_target *target)
5889 return MC_TARGET_NONE;
5893 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5894 unsigned long addr, unsigned long end,
5895 struct mm_walk *walk)
5897 struct vm_area_struct *vma = walk->vma;
5901 ptl = pmd_trans_huge_lock(pmd, vma);
5904 * Note their can not be MC_TARGET_DEVICE for now as we do not
5905 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5906 * this might change.
5908 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5909 mc.precharge += HPAGE_PMD_NR;
5914 if (pmd_trans_unstable(pmd))
5916 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5917 for (; addr != end; pte++, addr += PAGE_SIZE)
5918 if (get_mctgt_type(vma, addr, *pte, NULL))
5919 mc.precharge++; /* increment precharge temporarily */
5920 pte_unmap_unlock(pte - 1, ptl);
5926 static const struct mm_walk_ops precharge_walk_ops = {
5927 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5930 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5932 unsigned long precharge;
5935 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5936 mmap_read_unlock(mm);
5938 precharge = mc.precharge;
5944 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5946 unsigned long precharge = mem_cgroup_count_precharge(mm);
5948 VM_BUG_ON(mc.moving_task);
5949 mc.moving_task = current;
5950 return mem_cgroup_do_precharge(precharge);
5953 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5954 static void __mem_cgroup_clear_mc(void)
5956 struct mem_cgroup *from = mc.from;
5957 struct mem_cgroup *to = mc.to;
5959 /* we must uncharge all the leftover precharges from mc.to */
5961 cancel_charge(mc.to, mc.precharge);
5965 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5966 * we must uncharge here.
5968 if (mc.moved_charge) {
5969 cancel_charge(mc.from, mc.moved_charge);
5970 mc.moved_charge = 0;
5972 /* we must fixup refcnts and charges */
5973 if (mc.moved_swap) {
5974 /* uncharge swap account from the old cgroup */
5975 if (!mem_cgroup_is_root(mc.from))
5976 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5978 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5981 * we charged both to->memory and to->memsw, so we
5982 * should uncharge to->memory.
5984 if (!mem_cgroup_is_root(mc.to))
5985 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5989 memcg_oom_recover(from);
5990 memcg_oom_recover(to);
5991 wake_up_all(&mc.waitq);
5994 static void mem_cgroup_clear_mc(void)
5996 struct mm_struct *mm = mc.mm;
5999 * we must clear moving_task before waking up waiters at the end of
6002 mc.moving_task = NULL;
6003 __mem_cgroup_clear_mc();
6004 spin_lock(&mc.lock);
6008 spin_unlock(&mc.lock);
6013 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6015 struct cgroup_subsys_state *css;
6016 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
6017 struct mem_cgroup *from;
6018 struct task_struct *leader, *p;
6019 struct mm_struct *mm;
6020 unsigned long move_flags;
6023 /* charge immigration isn't supported on the default hierarchy */
6024 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6028 * Multi-process migrations only happen on the default hierarchy
6029 * where charge immigration is not used. Perform charge
6030 * immigration if @tset contains a leader and whine if there are
6034 cgroup_taskset_for_each_leader(leader, css, tset) {
6037 memcg = mem_cgroup_from_css(css);
6043 * We are now commited to this value whatever it is. Changes in this
6044 * tunable will only affect upcoming migrations, not the current one.
6045 * So we need to save it, and keep it going.
6047 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
6051 from = mem_cgroup_from_task(p);
6053 VM_BUG_ON(from == memcg);
6055 mm = get_task_mm(p);
6058 /* We move charges only when we move a owner of the mm */
6059 if (mm->owner == p) {
6062 VM_BUG_ON(mc.precharge);
6063 VM_BUG_ON(mc.moved_charge);
6064 VM_BUG_ON(mc.moved_swap);
6066 spin_lock(&mc.lock);
6070 mc.flags = move_flags;
6071 spin_unlock(&mc.lock);
6072 /* We set mc.moving_task later */
6074 ret = mem_cgroup_precharge_mc(mm);
6076 mem_cgroup_clear_mc();
6083 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6086 mem_cgroup_clear_mc();
6089 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6090 unsigned long addr, unsigned long end,
6091 struct mm_walk *walk)
6094 struct vm_area_struct *vma = walk->vma;
6097 enum mc_target_type target_type;
6098 union mc_target target;
6101 ptl = pmd_trans_huge_lock(pmd, vma);
6103 if (mc.precharge < HPAGE_PMD_NR) {
6107 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6108 if (target_type == MC_TARGET_PAGE) {
6110 if (!isolate_lru_page(page)) {
6111 if (!mem_cgroup_move_account(page, true,
6113 mc.precharge -= HPAGE_PMD_NR;
6114 mc.moved_charge += HPAGE_PMD_NR;
6116 putback_lru_page(page);
6119 } else if (target_type == MC_TARGET_DEVICE) {
6121 if (!mem_cgroup_move_account(page, true,
6123 mc.precharge -= HPAGE_PMD_NR;
6124 mc.moved_charge += HPAGE_PMD_NR;
6132 if (pmd_trans_unstable(pmd))
6135 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6136 for (; addr != end; addr += PAGE_SIZE) {
6137 pte_t ptent = *(pte++);
6138 bool device = false;
6144 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6145 case MC_TARGET_DEVICE:
6148 case MC_TARGET_PAGE:
6151 * We can have a part of the split pmd here. Moving it
6152 * can be done but it would be too convoluted so simply
6153 * ignore such a partial THP and keep it in original
6154 * memcg. There should be somebody mapping the head.
6156 if (PageTransCompound(page))
6158 if (!device && isolate_lru_page(page))
6160 if (!mem_cgroup_move_account(page, false,
6163 /* we uncharge from mc.from later. */
6167 putback_lru_page(page);
6168 put: /* get_mctgt_type() gets the page */
6171 case MC_TARGET_SWAP:
6173 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6175 mem_cgroup_id_get_many(mc.to, 1);
6176 /* we fixup other refcnts and charges later. */
6184 pte_unmap_unlock(pte - 1, ptl);
6189 * We have consumed all precharges we got in can_attach().
6190 * We try charge one by one, but don't do any additional
6191 * charges to mc.to if we have failed in charge once in attach()
6194 ret = mem_cgroup_do_precharge(1);
6202 static const struct mm_walk_ops charge_walk_ops = {
6203 .pmd_entry = mem_cgroup_move_charge_pte_range,
6206 static void mem_cgroup_move_charge(void)
6208 lru_add_drain_all();
6210 * Signal lock_page_memcg() to take the memcg's move_lock
6211 * while we're moving its pages to another memcg. Then wait
6212 * for already started RCU-only updates to finish.
6214 atomic_inc(&mc.from->moving_account);
6217 if (unlikely(!mmap_read_trylock(mc.mm))) {
6219 * Someone who are holding the mmap_lock might be waiting in
6220 * waitq. So we cancel all extra charges, wake up all waiters,
6221 * and retry. Because we cancel precharges, we might not be able
6222 * to move enough charges, but moving charge is a best-effort
6223 * feature anyway, so it wouldn't be a big problem.
6225 __mem_cgroup_clear_mc();
6230 * When we have consumed all precharges and failed in doing
6231 * additional charge, the page walk just aborts.
6233 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6236 mmap_read_unlock(mc.mm);
6237 atomic_dec(&mc.from->moving_account);
6240 static void mem_cgroup_move_task(void)
6243 mem_cgroup_move_charge();
6244 mem_cgroup_clear_mc();
6247 #else /* !CONFIG_MMU */
6248 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6252 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6255 static void mem_cgroup_move_task(void)
6261 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6262 * to verify whether we're attached to the default hierarchy on each mount
6265 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6268 * use_hierarchy is forced on the default hierarchy. cgroup core
6269 * guarantees that @root doesn't have any children, so turning it
6270 * on for the root memcg is enough.
6272 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6273 root_mem_cgroup->use_hierarchy = true;
6275 root_mem_cgroup->use_hierarchy = false;
6278 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6280 if (value == PAGE_COUNTER_MAX)
6281 seq_puts(m, "max\n");
6283 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6288 static u64 memory_current_read(struct cgroup_subsys_state *css,
6291 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6293 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6296 static int memory_min_show(struct seq_file *m, void *v)
6298 return seq_puts_memcg_tunable(m,
6299 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6302 static ssize_t memory_min_write(struct kernfs_open_file *of,
6303 char *buf, size_t nbytes, loff_t off)
6305 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6309 buf = strstrip(buf);
6310 err = page_counter_memparse(buf, "max", &min);
6314 page_counter_set_min(&memcg->memory, min);
6319 static int memory_low_show(struct seq_file *m, void *v)
6321 return seq_puts_memcg_tunable(m,
6322 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6325 static ssize_t memory_low_write(struct kernfs_open_file *of,
6326 char *buf, size_t nbytes, loff_t off)
6328 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6332 buf = strstrip(buf);
6333 err = page_counter_memparse(buf, "max", &low);
6337 page_counter_set_low(&memcg->memory, low);
6342 static int memory_high_show(struct seq_file *m, void *v)
6344 return seq_puts_memcg_tunable(m,
6345 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6348 static ssize_t memory_high_write(struct kernfs_open_file *of,
6349 char *buf, size_t nbytes, loff_t off)
6351 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6352 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6353 bool drained = false;
6357 buf = strstrip(buf);
6358 err = page_counter_memparse(buf, "max", &high);
6362 page_counter_set_high(&memcg->memory, high);
6365 unsigned long nr_pages = page_counter_read(&memcg->memory);
6366 unsigned long reclaimed;
6368 if (nr_pages <= high)
6371 if (signal_pending(current))
6375 drain_all_stock(memcg);
6380 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6383 if (!reclaimed && !nr_retries--)
6387 memcg_wb_domain_size_changed(memcg);
6391 static int memory_max_show(struct seq_file *m, void *v)
6393 return seq_puts_memcg_tunable(m,
6394 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6397 static ssize_t memory_max_write(struct kernfs_open_file *of,
6398 char *buf, size_t nbytes, loff_t off)
6400 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6401 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6402 bool drained = false;
6406 buf = strstrip(buf);
6407 err = page_counter_memparse(buf, "max", &max);
6411 xchg(&memcg->memory.max, max);
6414 unsigned long nr_pages = page_counter_read(&memcg->memory);
6416 if (nr_pages <= max)
6419 if (signal_pending(current))
6423 drain_all_stock(memcg);
6429 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6435 memcg_memory_event(memcg, MEMCG_OOM);
6436 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6440 memcg_wb_domain_size_changed(memcg);
6444 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6446 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6447 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6448 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6449 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6450 seq_printf(m, "oom_kill %lu\n",
6451 atomic_long_read(&events[MEMCG_OOM_KILL]));
6454 static int memory_events_show(struct seq_file *m, void *v)
6456 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6458 __memory_events_show(m, memcg->memory_events);
6462 static int memory_events_local_show(struct seq_file *m, void *v)
6464 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6466 __memory_events_show(m, memcg->memory_events_local);
6470 static int memory_stat_show(struct seq_file *m, void *v)
6472 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6475 buf = memory_stat_format(memcg);
6484 static int memory_numa_stat_show(struct seq_file *m, void *v)
6487 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6489 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6492 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6495 seq_printf(m, "%s", memory_stats[i].name);
6496 for_each_node_state(nid, N_MEMORY) {
6498 struct lruvec *lruvec;
6500 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6501 size = lruvec_page_state(lruvec, memory_stats[i].idx);
6502 size *= memory_stats[i].ratio;
6503 seq_printf(m, " N%d=%llu", nid, size);
6512 static int memory_oom_group_show(struct seq_file *m, void *v)
6514 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6516 seq_printf(m, "%d\n", memcg->oom_group);
6521 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6522 char *buf, size_t nbytes, loff_t off)
6524 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6527 buf = strstrip(buf);
6531 ret = kstrtoint(buf, 0, &oom_group);
6535 if (oom_group != 0 && oom_group != 1)
6538 memcg->oom_group = oom_group;
6543 static struct cftype memory_files[] = {
6546 .flags = CFTYPE_NOT_ON_ROOT,
6547 .read_u64 = memory_current_read,
6551 .flags = CFTYPE_NOT_ON_ROOT,
6552 .seq_show = memory_min_show,
6553 .write = memory_min_write,
6557 .flags = CFTYPE_NOT_ON_ROOT,
6558 .seq_show = memory_low_show,
6559 .write = memory_low_write,
6563 .flags = CFTYPE_NOT_ON_ROOT,
6564 .seq_show = memory_high_show,
6565 .write = memory_high_write,
6569 .flags = CFTYPE_NOT_ON_ROOT,
6570 .seq_show = memory_max_show,
6571 .write = memory_max_write,
6575 .flags = CFTYPE_NOT_ON_ROOT,
6576 .file_offset = offsetof(struct mem_cgroup, events_file),
6577 .seq_show = memory_events_show,
6580 .name = "events.local",
6581 .flags = CFTYPE_NOT_ON_ROOT,
6582 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6583 .seq_show = memory_events_local_show,
6587 .seq_show = memory_stat_show,
6591 .name = "numa_stat",
6592 .seq_show = memory_numa_stat_show,
6596 .name = "oom.group",
6597 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6598 .seq_show = memory_oom_group_show,
6599 .write = memory_oom_group_write,
6604 struct cgroup_subsys memory_cgrp_subsys = {
6605 .css_alloc = mem_cgroup_css_alloc,
6606 .css_online = mem_cgroup_css_online,
6607 .css_offline = mem_cgroup_css_offline,
6608 .css_released = mem_cgroup_css_released,
6609 .css_free = mem_cgroup_css_free,
6610 .css_reset = mem_cgroup_css_reset,
6611 .can_attach = mem_cgroup_can_attach,
6612 .cancel_attach = mem_cgroup_cancel_attach,
6613 .post_attach = mem_cgroup_move_task,
6614 .bind = mem_cgroup_bind,
6615 .dfl_cftypes = memory_files,
6616 .legacy_cftypes = mem_cgroup_legacy_files,
6621 * This function calculates an individual cgroup's effective
6622 * protection which is derived from its own memory.min/low, its
6623 * parent's and siblings' settings, as well as the actual memory
6624 * distribution in the tree.
6626 * The following rules apply to the effective protection values:
6628 * 1. At the first level of reclaim, effective protection is equal to
6629 * the declared protection in memory.min and memory.low.
6631 * 2. To enable safe delegation of the protection configuration, at
6632 * subsequent levels the effective protection is capped to the
6633 * parent's effective protection.
6635 * 3. To make complex and dynamic subtrees easier to configure, the
6636 * user is allowed to overcommit the declared protection at a given
6637 * level. If that is the case, the parent's effective protection is
6638 * distributed to the children in proportion to how much protection
6639 * they have declared and how much of it they are utilizing.
6641 * This makes distribution proportional, but also work-conserving:
6642 * if one cgroup claims much more protection than it uses memory,
6643 * the unused remainder is available to its siblings.
6645 * 4. Conversely, when the declared protection is undercommitted at a
6646 * given level, the distribution of the larger parental protection
6647 * budget is NOT proportional. A cgroup's protection from a sibling
6648 * is capped to its own memory.min/low setting.
6650 * 5. However, to allow protecting recursive subtrees from each other
6651 * without having to declare each individual cgroup's fixed share
6652 * of the ancestor's claim to protection, any unutilized -
6653 * "floating" - protection from up the tree is distributed in
6654 * proportion to each cgroup's *usage*. This makes the protection
6655 * neutral wrt sibling cgroups and lets them compete freely over
6656 * the shared parental protection budget, but it protects the
6657 * subtree as a whole from neighboring subtrees.
6659 * Note that 4. and 5. are not in conflict: 4. is about protecting
6660 * against immediate siblings whereas 5. is about protecting against
6661 * neighboring subtrees.
6663 static unsigned long effective_protection(unsigned long usage,
6664 unsigned long parent_usage,
6665 unsigned long setting,
6666 unsigned long parent_effective,
6667 unsigned long siblings_protected)
6669 unsigned long protected;
6672 protected = min(usage, setting);
6674 * If all cgroups at this level combined claim and use more
6675 * protection then what the parent affords them, distribute
6676 * shares in proportion to utilization.
6678 * We are using actual utilization rather than the statically
6679 * claimed protection in order to be work-conserving: claimed
6680 * but unused protection is available to siblings that would
6681 * otherwise get a smaller chunk than what they claimed.
6683 if (siblings_protected > parent_effective)
6684 return protected * parent_effective / siblings_protected;
6687 * Ok, utilized protection of all children is within what the
6688 * parent affords them, so we know whatever this child claims
6689 * and utilizes is effectively protected.
6691 * If there is unprotected usage beyond this value, reclaim
6692 * will apply pressure in proportion to that amount.
6694 * If there is unutilized protection, the cgroup will be fully
6695 * shielded from reclaim, but we do return a smaller value for
6696 * protection than what the group could enjoy in theory. This
6697 * is okay. With the overcommit distribution above, effective
6698 * protection is always dependent on how memory is actually
6699 * consumed among the siblings anyway.
6704 * If the children aren't claiming (all of) the protection
6705 * afforded to them by the parent, distribute the remainder in
6706 * proportion to the (unprotected) memory of each cgroup. That
6707 * way, cgroups that aren't explicitly prioritized wrt each
6708 * other compete freely over the allowance, but they are
6709 * collectively protected from neighboring trees.
6711 * We're using unprotected memory for the weight so that if
6712 * some cgroups DO claim explicit protection, we don't protect
6713 * the same bytes twice.
6715 * Check both usage and parent_usage against the respective
6716 * protected values. One should imply the other, but they
6717 * aren't read atomically - make sure the division is sane.
6719 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6721 if (parent_effective > siblings_protected &&
6722 parent_usage > siblings_protected &&
6723 usage > protected) {
6724 unsigned long unclaimed;
6726 unclaimed = parent_effective - siblings_protected;
6727 unclaimed *= usage - protected;
6728 unclaimed /= parent_usage - siblings_protected;
6737 * mem_cgroup_protected - check if memory consumption is in the normal range
6738 * @root: the top ancestor of the sub-tree being checked
6739 * @memcg: the memory cgroup to check
6741 * WARNING: This function is not stateless! It can only be used as part
6742 * of a top-down tree iteration, not for isolated queries.
6744 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6745 struct mem_cgroup *memcg)
6747 unsigned long usage, parent_usage;
6748 struct mem_cgroup *parent;
6750 if (mem_cgroup_disabled())
6754 root = root_mem_cgroup;
6757 * Effective values of the reclaim targets are ignored so they
6758 * can be stale. Have a look at mem_cgroup_protection for more
6760 * TODO: calculation should be more robust so that we do not need
6761 * that special casing.
6766 usage = page_counter_read(&memcg->memory);
6770 parent = parent_mem_cgroup(memcg);
6771 /* No parent means a non-hierarchical mode on v1 memcg */
6775 if (parent == root) {
6776 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6777 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6781 parent_usage = page_counter_read(&parent->memory);
6783 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6784 READ_ONCE(memcg->memory.min),
6785 READ_ONCE(parent->memory.emin),
6786 atomic_long_read(&parent->memory.children_min_usage)));
6788 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6789 READ_ONCE(memcg->memory.low),
6790 READ_ONCE(parent->memory.elow),
6791 atomic_long_read(&parent->memory.children_low_usage)));
6795 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6796 * @page: page to charge
6797 * @mm: mm context of the victim
6798 * @gfp_mask: reclaim mode
6800 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6801 * pages according to @gfp_mask if necessary.
6803 * Returns 0 on success. Otherwise, an error code is returned.
6805 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6807 unsigned int nr_pages = thp_nr_pages(page);
6808 struct mem_cgroup *memcg = NULL;
6811 if (mem_cgroup_disabled())
6814 if (PageSwapCache(page)) {
6815 swp_entry_t ent = { .val = page_private(page), };
6819 * Every swap fault against a single page tries to charge the
6820 * page, bail as early as possible. shmem_unuse() encounters
6821 * already charged pages, too. page->mem_cgroup is protected
6822 * by the page lock, which serializes swap cache removal, which
6823 * in turn serializes uncharging.
6825 VM_BUG_ON_PAGE(!PageLocked(page), page);
6826 if (compound_head(page)->mem_cgroup)
6829 id = lookup_swap_cgroup_id(ent);
6831 memcg = mem_cgroup_from_id(id);
6832 if (memcg && !css_tryget_online(&memcg->css))
6838 memcg = get_mem_cgroup_from_mm(mm);
6840 ret = try_charge(memcg, gfp_mask, nr_pages);
6844 css_get(&memcg->css);
6845 commit_charge(page, memcg);
6847 local_lock_irq(&event_lock.l);
6848 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6849 memcg_check_events(memcg, page);
6850 local_unlock_irq(&event_lock.l);
6853 * Cgroup1's unified memory+swap counter has been charged with the
6854 * new swapcache page, finish the transfer by uncharging the swap
6855 * slot. The swap slot would also get uncharged when it dies, but
6856 * it can stick around indefinitely and we'd count the page twice
6859 * Cgroup2 has separate resource counters for memory and swap,
6860 * so this is a non-issue here. Memory and swap charge lifetimes
6861 * correspond 1:1 to page and swap slot lifetimes: we charge the
6862 * page to memory here, and uncharge swap when the slot is freed.
6864 if (do_memsw_account() && PageSwapCache(page)) {
6865 swp_entry_t entry = { .val = page_private(page) };
6867 * The swap entry might not get freed for a long time,
6868 * let's not wait for it. The page already received a
6869 * memory+swap charge, drop the swap entry duplicate.
6871 mem_cgroup_uncharge_swap(entry, nr_pages);
6875 css_put(&memcg->css);
6880 struct uncharge_gather {
6881 struct mem_cgroup *memcg;
6882 unsigned long nr_pages;
6883 unsigned long pgpgout;
6884 unsigned long nr_kmem;
6885 struct page *dummy_page;
6888 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6890 memset(ug, 0, sizeof(*ug));
6893 static void uncharge_batch(const struct uncharge_gather *ug)
6895 unsigned long flags;
6897 if (!mem_cgroup_is_root(ug->memcg)) {
6898 page_counter_uncharge(&ug->memcg->memory, ug->nr_pages);
6899 if (do_memsw_account())
6900 page_counter_uncharge(&ug->memcg->memsw, ug->nr_pages);
6901 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6902 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6903 memcg_oom_recover(ug->memcg);
6906 local_lock_irqsave(&event_lock.l, flags);
6907 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6908 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_pages);
6909 memcg_check_events(ug->memcg, ug->dummy_page);
6910 local_unlock_irqrestore(&event_lock.l, flags);
6912 /* drop reference from uncharge_page */
6913 css_put(&ug->memcg->css);
6916 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6918 unsigned long nr_pages;
6920 VM_BUG_ON_PAGE(PageLRU(page), page);
6922 if (!page->mem_cgroup)
6926 * Nobody should be changing or seriously looking at
6927 * page->mem_cgroup at this point, we have fully
6928 * exclusive access to the page.
6931 if (ug->memcg != page->mem_cgroup) {
6934 uncharge_gather_clear(ug);
6936 ug->memcg = page->mem_cgroup;
6938 /* pairs with css_put in uncharge_batch */
6939 css_get(&ug->memcg->css);
6942 nr_pages = compound_nr(page);
6943 ug->nr_pages += nr_pages;
6945 if (!PageKmemcg(page)) {
6948 ug->nr_kmem += nr_pages;
6949 __ClearPageKmemcg(page);
6952 ug->dummy_page = page;
6953 page->mem_cgroup = NULL;
6954 css_put(&ug->memcg->css);
6957 static void uncharge_list(struct list_head *page_list)
6959 struct uncharge_gather ug;
6960 struct list_head *next;
6962 uncharge_gather_clear(&ug);
6965 * Note that the list can be a single page->lru; hence the
6966 * do-while loop instead of a simple list_for_each_entry().
6968 next = page_list->next;
6972 page = list_entry(next, struct page, lru);
6973 next = page->lru.next;
6975 uncharge_page(page, &ug);
6976 } while (next != page_list);
6979 uncharge_batch(&ug);
6983 * mem_cgroup_uncharge - uncharge a page
6984 * @page: page to uncharge
6986 * Uncharge a page previously charged with mem_cgroup_charge().
6988 void mem_cgroup_uncharge(struct page *page)
6990 struct uncharge_gather ug;
6992 if (mem_cgroup_disabled())
6995 /* Don't touch page->lru of any random page, pre-check: */
6996 if (!page->mem_cgroup)
6999 uncharge_gather_clear(&ug);
7000 uncharge_page(page, &ug);
7001 uncharge_batch(&ug);
7005 * mem_cgroup_uncharge_list - uncharge a list of page
7006 * @page_list: list of pages to uncharge
7008 * Uncharge a list of pages previously charged with
7009 * mem_cgroup_charge().
7011 void mem_cgroup_uncharge_list(struct list_head *page_list)
7013 if (mem_cgroup_disabled())
7016 if (!list_empty(page_list))
7017 uncharge_list(page_list);
7021 * mem_cgroup_migrate - charge a page's replacement
7022 * @oldpage: currently circulating page
7023 * @newpage: replacement page
7025 * Charge @newpage as a replacement page for @oldpage. @oldpage will
7026 * be uncharged upon free.
7028 * Both pages must be locked, @newpage->mapping must be set up.
7030 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
7032 struct mem_cgroup *memcg;
7033 unsigned int nr_pages;
7034 unsigned long flags;
7036 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
7037 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
7038 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
7039 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
7042 if (mem_cgroup_disabled())
7045 /* Page cache replacement: new page already charged? */
7046 if (newpage->mem_cgroup)
7049 /* Swapcache readahead pages can get replaced before being charged */
7050 memcg = oldpage->mem_cgroup;
7054 /* Force-charge the new page. The old one will be freed soon */
7055 nr_pages = thp_nr_pages(newpage);
7057 page_counter_charge(&memcg->memory, nr_pages);
7058 if (do_memsw_account())
7059 page_counter_charge(&memcg->memsw, nr_pages);
7061 css_get(&memcg->css);
7062 commit_charge(newpage, memcg);
7064 local_lock_irqsave(&event_lock.l, flags);
7065 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
7066 memcg_check_events(memcg, newpage);
7067 local_unlock_irqrestore(&event_lock.l, flags);
7070 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
7071 EXPORT_SYMBOL(memcg_sockets_enabled_key);
7073 void mem_cgroup_sk_alloc(struct sock *sk)
7075 struct mem_cgroup *memcg;
7077 if (!mem_cgroup_sockets_enabled)
7080 /* Do not associate the sock with unrelated interrupted task's memcg. */
7085 memcg = mem_cgroup_from_task(current);
7086 if (memcg == root_mem_cgroup)
7088 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
7090 if (css_tryget(&memcg->css))
7091 sk->sk_memcg = memcg;
7096 void mem_cgroup_sk_free(struct sock *sk)
7099 css_put(&sk->sk_memcg->css);
7103 * mem_cgroup_charge_skmem - charge socket memory
7104 * @memcg: memcg to charge
7105 * @nr_pages: number of pages to charge
7107 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7108 * @memcg's configured limit, %false if the charge had to be forced.
7110 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7112 gfp_t gfp_mask = GFP_KERNEL;
7114 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7115 struct page_counter *fail;
7117 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7118 memcg->tcpmem_pressure = 0;
7121 page_counter_charge(&memcg->tcpmem, nr_pages);
7122 memcg->tcpmem_pressure = 1;
7126 /* Don't block in the packet receive path */
7128 gfp_mask = GFP_NOWAIT;
7130 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7132 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
7135 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
7140 * mem_cgroup_uncharge_skmem - uncharge socket memory
7141 * @memcg: memcg to uncharge
7142 * @nr_pages: number of pages to uncharge
7144 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7146 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7147 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7151 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7153 refill_stock(memcg, nr_pages);
7156 static int __init cgroup_memory(char *s)
7160 while ((token = strsep(&s, ",")) != NULL) {
7163 if (!strcmp(token, "nosocket"))
7164 cgroup_memory_nosocket = true;
7165 if (!strcmp(token, "nokmem"))
7166 cgroup_memory_nokmem = true;
7170 __setup("cgroup.memory=", cgroup_memory);
7173 * subsys_initcall() for memory controller.
7175 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7176 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7177 * basically everything that doesn't depend on a specific mem_cgroup structure
7178 * should be initialized from here.
7180 static int __init mem_cgroup_init(void)
7184 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7185 memcg_hotplug_cpu_dead);
7187 for_each_possible_cpu(cpu) {
7188 struct memcg_stock_pcp *stock;
7190 stock = per_cpu_ptr(&memcg_stock, cpu);
7191 INIT_WORK(&stock->work, drain_local_stock);
7192 local_lock_init(&stock->lock);
7195 for_each_node(node) {
7196 struct mem_cgroup_tree_per_node *rtpn;
7198 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7199 node_online(node) ? node : NUMA_NO_NODE);
7201 rtpn->rb_root = RB_ROOT;
7202 rtpn->rb_rightmost = NULL;
7203 spin_lock_init(&rtpn->lock);
7204 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7209 subsys_initcall(mem_cgroup_init);
7211 #ifdef CONFIG_MEMCG_SWAP
7212 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7214 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7216 * The root cgroup cannot be destroyed, so it's refcount must
7219 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7223 memcg = parent_mem_cgroup(memcg);
7225 memcg = root_mem_cgroup;
7231 * mem_cgroup_swapout - transfer a memsw charge to swap
7232 * @page: page whose memsw charge to transfer
7233 * @entry: swap entry to move the charge to
7235 * Transfer the memsw charge of @page to @entry.
7237 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7239 struct mem_cgroup *memcg, *swap_memcg;
7240 unsigned int nr_entries;
7241 unsigned short oldid;
7242 unsigned long flags;
7244 VM_BUG_ON_PAGE(PageLRU(page), page);
7245 VM_BUG_ON_PAGE(page_count(page), page);
7247 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7250 memcg = page->mem_cgroup;
7252 /* Readahead page, never charged */
7257 * In case the memcg owning these pages has been offlined and doesn't
7258 * have an ID allocated to it anymore, charge the closest online
7259 * ancestor for the swap instead and transfer the memory+swap charge.
7261 swap_memcg = mem_cgroup_id_get_online(memcg);
7262 nr_entries = thp_nr_pages(page);
7263 /* Get references for the tail pages, too */
7265 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7266 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7268 VM_BUG_ON_PAGE(oldid, page);
7269 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7271 page->mem_cgroup = NULL;
7273 if (!mem_cgroup_is_root(memcg))
7274 page_counter_uncharge(&memcg->memory, nr_entries);
7276 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7277 if (!mem_cgroup_is_root(swap_memcg))
7278 page_counter_charge(&swap_memcg->memsw, nr_entries);
7279 page_counter_uncharge(&memcg->memsw, nr_entries);
7283 * Interrupts should be disabled here because the caller holds the
7284 * i_pages lock which is taken with interrupts-off. It is
7285 * important here to have the interrupts disabled because it is the
7286 * only synchronisation we have for updating the per-CPU variables.
7288 local_lock_irqsave(&event_lock.l, flags);
7289 #ifndef CONFIG_PREEMPT_RT
7290 VM_BUG_ON(!irqs_disabled());
7292 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7293 memcg_check_events(memcg, page);
7294 local_unlock_irqrestore(&event_lock.l, flags);
7296 css_put(&memcg->css);
7300 * mem_cgroup_try_charge_swap - try charging swap space for a page
7301 * @page: page being added to swap
7302 * @entry: swap entry to charge
7304 * Try to charge @page's memcg for the swap space at @entry.
7306 * Returns 0 on success, -ENOMEM on failure.
7308 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7310 unsigned int nr_pages = thp_nr_pages(page);
7311 struct page_counter *counter;
7312 struct mem_cgroup *memcg;
7313 unsigned short oldid;
7315 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7318 memcg = page->mem_cgroup;
7320 /* Readahead page, never charged */
7325 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7329 memcg = mem_cgroup_id_get_online(memcg);
7331 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7332 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7333 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7334 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7335 mem_cgroup_id_put(memcg);
7339 /* Get references for the tail pages, too */
7341 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7342 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7343 VM_BUG_ON_PAGE(oldid, page);
7344 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7350 * mem_cgroup_uncharge_swap - uncharge swap space
7351 * @entry: swap entry to uncharge
7352 * @nr_pages: the amount of swap space to uncharge
7354 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7356 struct mem_cgroup *memcg;
7359 id = swap_cgroup_record(entry, 0, nr_pages);
7361 memcg = mem_cgroup_from_id(id);
7363 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7364 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7365 page_counter_uncharge(&memcg->swap, nr_pages);
7367 page_counter_uncharge(&memcg->memsw, nr_pages);
7369 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7370 mem_cgroup_id_put_many(memcg, nr_pages);
7375 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7377 long nr_swap_pages = get_nr_swap_pages();
7379 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7380 return nr_swap_pages;
7381 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7382 nr_swap_pages = min_t(long, nr_swap_pages,
7383 READ_ONCE(memcg->swap.max) -
7384 page_counter_read(&memcg->swap));
7385 return nr_swap_pages;
7388 bool mem_cgroup_swap_full(struct page *page)
7390 struct mem_cgroup *memcg;
7392 VM_BUG_ON_PAGE(!PageLocked(page), page);
7396 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7399 memcg = page->mem_cgroup;
7403 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7404 unsigned long usage = page_counter_read(&memcg->swap);
7406 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7407 usage * 2 >= READ_ONCE(memcg->swap.max))
7414 static int __init setup_swap_account(char *s)
7416 if (!strcmp(s, "1"))
7417 cgroup_memory_noswap = 0;
7418 else if (!strcmp(s, "0"))
7419 cgroup_memory_noswap = 1;
7422 __setup("swapaccount=", setup_swap_account);
7424 static u64 swap_current_read(struct cgroup_subsys_state *css,
7427 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7429 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7432 static int swap_high_show(struct seq_file *m, void *v)
7434 return seq_puts_memcg_tunable(m,
7435 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7438 static ssize_t swap_high_write(struct kernfs_open_file *of,
7439 char *buf, size_t nbytes, loff_t off)
7441 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7445 buf = strstrip(buf);
7446 err = page_counter_memparse(buf, "max", &high);
7450 page_counter_set_high(&memcg->swap, high);
7455 static int swap_max_show(struct seq_file *m, void *v)
7457 return seq_puts_memcg_tunable(m,
7458 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7461 static ssize_t swap_max_write(struct kernfs_open_file *of,
7462 char *buf, size_t nbytes, loff_t off)
7464 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7468 buf = strstrip(buf);
7469 err = page_counter_memparse(buf, "max", &max);
7473 xchg(&memcg->swap.max, max);
7478 static int swap_events_show(struct seq_file *m, void *v)
7480 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7482 seq_printf(m, "high %lu\n",
7483 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7484 seq_printf(m, "max %lu\n",
7485 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7486 seq_printf(m, "fail %lu\n",
7487 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7492 static struct cftype swap_files[] = {
7494 .name = "swap.current",
7495 .flags = CFTYPE_NOT_ON_ROOT,
7496 .read_u64 = swap_current_read,
7499 .name = "swap.high",
7500 .flags = CFTYPE_NOT_ON_ROOT,
7501 .seq_show = swap_high_show,
7502 .write = swap_high_write,
7506 .flags = CFTYPE_NOT_ON_ROOT,
7507 .seq_show = swap_max_show,
7508 .write = swap_max_write,
7511 .name = "swap.events",
7512 .flags = CFTYPE_NOT_ON_ROOT,
7513 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7514 .seq_show = swap_events_show,
7519 static struct cftype memsw_files[] = {
7521 .name = "memsw.usage_in_bytes",
7522 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7523 .read_u64 = mem_cgroup_read_u64,
7526 .name = "memsw.max_usage_in_bytes",
7527 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7528 .write = mem_cgroup_reset,
7529 .read_u64 = mem_cgroup_read_u64,
7532 .name = "memsw.limit_in_bytes",
7533 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7534 .write = mem_cgroup_write,
7535 .read_u64 = mem_cgroup_read_u64,
7538 .name = "memsw.failcnt",
7539 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7540 .write = mem_cgroup_reset,
7541 .read_u64 = mem_cgroup_read_u64,
7544 .name = "force_reclaim",
7545 .write_u64 = mem_cgroup_force_reclaim,
7547 { }, /* terminate */
7551 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7552 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7553 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7554 * boot parameter. This may result in premature OOPS inside
7555 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7557 static int __init mem_cgroup_swap_init(void)
7559 /* No memory control -> no swap control */
7560 if (mem_cgroup_disabled())
7561 cgroup_memory_noswap = true;
7563 if (cgroup_memory_noswap)
7566 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7567 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7571 core_initcall(mem_cgroup_swap_init);
7573 #endif /* CONFIG_MEMCG_SWAP */