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>
67 #include <linux/uaccess.h>
69 #include <trace/events/vmscan.h>
71 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
72 EXPORT_SYMBOL(memory_cgrp_subsys);
74 struct mem_cgroup *root_mem_cgroup __read_mostly;
76 /* Active memory cgroup to use from an interrupt context */
77 DEFINE_PER_CPU(struct mem_cgroup *, int_active_memcg);
79 /* Socket memory accounting disabled? */
80 static bool cgroup_memory_nosocket;
82 /* Kernel memory accounting disabled? */
83 static bool cgroup_memory_nokmem;
85 /* Whether the swap controller is active */
86 #ifdef CONFIG_MEMCG_SWAP
87 bool cgroup_memory_noswap __read_mostly;
89 #define cgroup_memory_noswap 1
92 #ifdef CONFIG_CGROUP_WRITEBACK
93 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
96 /* Whether legacy memory+swap accounting is active */
97 static bool do_memsw_account(void)
99 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_noswap;
102 #define THRESHOLDS_EVENTS_TARGET 128
103 #define SOFTLIMIT_EVENTS_TARGET 1024
106 * Cgroups above their limits are maintained in a RB-Tree, independent of
107 * their hierarchy representation
110 struct mem_cgroup_tree_per_node {
111 struct rb_root rb_root;
112 struct rb_node *rb_rightmost;
116 struct mem_cgroup_tree {
117 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
120 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
123 struct mem_cgroup_eventfd_list {
124 struct list_head list;
125 struct eventfd_ctx *eventfd;
129 * cgroup_event represents events which userspace want to receive.
131 struct mem_cgroup_event {
133 * memcg which the event belongs to.
135 struct mem_cgroup *memcg;
137 * eventfd to signal userspace about the event.
139 struct eventfd_ctx *eventfd;
141 * Each of these stored in a list by the cgroup.
143 struct list_head list;
145 * register_event() callback will be used to add new userspace
146 * waiter for changes related to this event. Use eventfd_signal()
147 * on eventfd to send notification to userspace.
149 int (*register_event)(struct mem_cgroup *memcg,
150 struct eventfd_ctx *eventfd, const char *args);
152 * unregister_event() callback will be called when userspace closes
153 * the eventfd or on cgroup removing. This callback must be set,
154 * if you want provide notification functionality.
156 void (*unregister_event)(struct mem_cgroup *memcg,
157 struct eventfd_ctx *eventfd);
159 * All fields below needed to unregister event when
160 * userspace closes eventfd.
163 wait_queue_head_t *wqh;
164 wait_queue_entry_t wait;
165 struct work_struct remove;
168 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
169 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
171 /* Stuffs for move charges at task migration. */
173 * Types of charges to be moved.
175 #define MOVE_ANON 0x1U
176 #define MOVE_FILE 0x2U
177 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
179 /* "mc" and its members are protected by cgroup_mutex */
180 static struct move_charge_struct {
181 spinlock_t lock; /* for from, to */
182 struct mm_struct *mm;
183 struct mem_cgroup *from;
184 struct mem_cgroup *to;
186 unsigned long precharge;
187 unsigned long moved_charge;
188 unsigned long moved_swap;
189 struct task_struct *moving_task; /* a task moving charges */
190 wait_queue_head_t waitq; /* a waitq for other context */
192 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
193 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
197 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
198 * limit reclaim to prevent infinite loops, if they ever occur.
200 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
201 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
203 /* for encoding cft->private value on file */
212 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
213 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
214 #define MEMFILE_ATTR(val) ((val) & 0xffff)
215 /* Used for OOM nofiier */
216 #define OOM_CONTROL (0)
219 * Iteration constructs for visiting all cgroups (under a tree). If
220 * loops are exited prematurely (break), mem_cgroup_iter_break() must
221 * be used for reference counting.
223 #define for_each_mem_cgroup_tree(iter, root) \
224 for (iter = mem_cgroup_iter(root, NULL, NULL); \
226 iter = mem_cgroup_iter(root, iter, NULL))
228 #define for_each_mem_cgroup(iter) \
229 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
231 iter = mem_cgroup_iter(NULL, iter, NULL))
233 static inline bool should_force_charge(void)
235 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
236 (current->flags & PF_EXITING);
239 /* Some nice accessors for the vmpressure. */
240 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
243 memcg = root_mem_cgroup;
244 return &memcg->vmpressure;
247 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
249 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
252 #ifdef CONFIG_MEMCG_KMEM
253 extern spinlock_t css_set_lock;
255 static void obj_cgroup_release(struct percpu_ref *ref)
257 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
258 struct mem_cgroup *memcg;
259 unsigned int nr_bytes;
260 unsigned int nr_pages;
264 * At this point all allocated objects are freed, and
265 * objcg->nr_charged_bytes can't have an arbitrary byte value.
266 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
268 * The following sequence can lead to it:
269 * 1) CPU0: objcg == stock->cached_objcg
270 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
271 * PAGE_SIZE bytes are charged
272 * 3) CPU1: a process from another memcg is allocating something,
273 * the stock if flushed,
274 * objcg->nr_charged_bytes = PAGE_SIZE - 92
275 * 5) CPU0: we do release this object,
276 * 92 bytes are added to stock->nr_bytes
277 * 6) CPU0: stock is flushed,
278 * 92 bytes are added to objcg->nr_charged_bytes
280 * In the result, nr_charged_bytes == PAGE_SIZE.
281 * This page will be uncharged in obj_cgroup_release().
283 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
284 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
285 nr_pages = nr_bytes >> PAGE_SHIFT;
287 spin_lock_irqsave(&css_set_lock, flags);
288 memcg = obj_cgroup_memcg(objcg);
290 __memcg_kmem_uncharge(memcg, nr_pages);
291 list_del(&objcg->list);
292 mem_cgroup_put(memcg);
293 spin_unlock_irqrestore(&css_set_lock, flags);
295 percpu_ref_exit(ref);
296 kfree_rcu(objcg, rcu);
299 static struct obj_cgroup *obj_cgroup_alloc(void)
301 struct obj_cgroup *objcg;
304 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
308 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
314 INIT_LIST_HEAD(&objcg->list);
318 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
319 struct mem_cgroup *parent)
321 struct obj_cgroup *objcg, *iter;
323 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
325 spin_lock_irq(&css_set_lock);
327 /* Move active objcg to the parent's list */
328 xchg(&objcg->memcg, parent);
329 css_get(&parent->css);
330 list_add(&objcg->list, &parent->objcg_list);
332 /* Move already reparented objcgs to the parent's list */
333 list_for_each_entry(iter, &memcg->objcg_list, list) {
334 css_get(&parent->css);
335 xchg(&iter->memcg, parent);
336 css_put(&memcg->css);
338 list_splice(&memcg->objcg_list, &parent->objcg_list);
340 spin_unlock_irq(&css_set_lock);
342 percpu_ref_kill(&objcg->refcnt);
346 * This will be used as a shrinker list's index.
347 * The main reason for not using cgroup id for this:
348 * this works better in sparse environments, where we have a lot of memcgs,
349 * but only a few kmem-limited. Or also, if we have, for instance, 200
350 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
351 * 200 entry array for that.
353 * The current size of the caches array is stored in memcg_nr_cache_ids. It
354 * will double each time we have to increase it.
356 static DEFINE_IDA(memcg_cache_ida);
357 int memcg_nr_cache_ids;
359 /* Protects memcg_nr_cache_ids */
360 static DECLARE_RWSEM(memcg_cache_ids_sem);
362 void memcg_get_cache_ids(void)
364 down_read(&memcg_cache_ids_sem);
367 void memcg_put_cache_ids(void)
369 up_read(&memcg_cache_ids_sem);
373 * MIN_SIZE is different than 1, because we would like to avoid going through
374 * the alloc/free process all the time. In a small machine, 4 kmem-limited
375 * cgroups is a reasonable guess. In the future, it could be a parameter or
376 * tunable, but that is strictly not necessary.
378 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
379 * this constant directly from cgroup, but it is understandable that this is
380 * better kept as an internal representation in cgroup.c. In any case, the
381 * cgrp_id space is not getting any smaller, and we don't have to necessarily
382 * increase ours as well if it increases.
384 #define MEMCG_CACHES_MIN_SIZE 4
385 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
388 * A lot of the calls to the cache allocation functions are expected to be
389 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
390 * conditional to this static branch, we'll have to allow modules that does
391 * kmem_cache_alloc and the such to see this symbol as well
393 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
394 EXPORT_SYMBOL(memcg_kmem_enabled_key);
397 static int memcg_shrinker_map_size;
398 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
400 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
402 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
405 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
406 int size, int old_size)
408 struct memcg_shrinker_map *new, *old;
411 lockdep_assert_held(&memcg_shrinker_map_mutex);
414 old = rcu_dereference_protected(
415 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
416 /* Not yet online memcg */
420 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
424 /* Set all old bits, clear all new bits */
425 memset(new->map, (int)0xff, old_size);
426 memset((void *)new->map + old_size, 0, size - old_size);
428 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
429 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
435 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
437 struct mem_cgroup_per_node *pn;
438 struct memcg_shrinker_map *map;
441 if (mem_cgroup_is_root(memcg))
445 pn = mem_cgroup_nodeinfo(memcg, nid);
446 map = rcu_dereference_protected(pn->shrinker_map, true);
449 rcu_assign_pointer(pn->shrinker_map, NULL);
453 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
455 struct memcg_shrinker_map *map;
456 int nid, size, ret = 0;
458 if (mem_cgroup_is_root(memcg))
461 mutex_lock(&memcg_shrinker_map_mutex);
462 size = memcg_shrinker_map_size;
464 map = kvzalloc_node(sizeof(*map) + size, GFP_KERNEL, nid);
466 memcg_free_shrinker_maps(memcg);
470 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
472 mutex_unlock(&memcg_shrinker_map_mutex);
477 int memcg_expand_shrinker_maps(int new_id)
479 int size, old_size, ret = 0;
480 struct mem_cgroup *memcg;
482 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
483 old_size = memcg_shrinker_map_size;
484 if (size <= old_size)
487 mutex_lock(&memcg_shrinker_map_mutex);
488 if (!root_mem_cgroup)
491 for_each_mem_cgroup(memcg) {
492 if (mem_cgroup_is_root(memcg))
494 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
496 mem_cgroup_iter_break(NULL, memcg);
502 memcg_shrinker_map_size = size;
503 mutex_unlock(&memcg_shrinker_map_mutex);
507 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
509 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
510 struct memcg_shrinker_map *map;
513 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
514 /* Pairs with smp mb in shrink_slab() */
515 smp_mb__before_atomic();
516 set_bit(shrinker_id, map->map);
522 * mem_cgroup_css_from_page - css of the memcg associated with a page
523 * @page: page of interest
525 * If memcg is bound to the default hierarchy, css of the memcg associated
526 * with @page is returned. The returned css remains associated with @page
527 * until it is released.
529 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
532 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
534 struct mem_cgroup *memcg;
536 memcg = page->mem_cgroup;
538 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
539 memcg = root_mem_cgroup;
545 * page_cgroup_ino - return inode number of the memcg a page is charged to
548 * Look up the closest online ancestor of the memory cgroup @page is charged to
549 * and return its inode number or 0 if @page is not charged to any cgroup. It
550 * is safe to call this function without holding a reference to @page.
552 * Note, this function is inherently racy, because there is nothing to prevent
553 * the cgroup inode from getting torn down and potentially reallocated a moment
554 * after page_cgroup_ino() returns, so it only should be used by callers that
555 * do not care (such as procfs interfaces).
557 ino_t page_cgroup_ino(struct page *page)
559 struct mem_cgroup *memcg;
560 unsigned long ino = 0;
563 memcg = page->mem_cgroup;
566 * The lowest bit set means that memcg isn't a valid
567 * memcg pointer, but a obj_cgroups pointer.
568 * In this case the page is shared and doesn't belong
569 * to any specific memory cgroup.
571 if ((unsigned long) memcg & 0x1UL)
574 while (memcg && !(memcg->css.flags & CSS_ONLINE))
575 memcg = parent_mem_cgroup(memcg);
577 ino = cgroup_ino(memcg->css.cgroup);
582 static struct mem_cgroup_per_node *
583 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
585 int nid = page_to_nid(page);
587 return memcg->nodeinfo[nid];
590 static struct mem_cgroup_tree_per_node *
591 soft_limit_tree_node(int nid)
593 return soft_limit_tree.rb_tree_per_node[nid];
596 static struct mem_cgroup_tree_per_node *
597 soft_limit_tree_from_page(struct page *page)
599 int nid = page_to_nid(page);
601 return soft_limit_tree.rb_tree_per_node[nid];
604 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
605 struct mem_cgroup_tree_per_node *mctz,
606 unsigned long new_usage_in_excess)
608 struct rb_node **p = &mctz->rb_root.rb_node;
609 struct rb_node *parent = NULL;
610 struct mem_cgroup_per_node *mz_node;
611 bool rightmost = true;
616 mz->usage_in_excess = new_usage_in_excess;
617 if (!mz->usage_in_excess)
621 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
623 if (mz->usage_in_excess < mz_node->usage_in_excess) {
629 * We can't avoid mem cgroups that are over their soft
630 * limit by the same amount
632 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
637 mctz->rb_rightmost = &mz->tree_node;
639 rb_link_node(&mz->tree_node, parent, p);
640 rb_insert_color(&mz->tree_node, &mctz->rb_root);
644 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
645 struct mem_cgroup_tree_per_node *mctz)
650 if (&mz->tree_node == mctz->rb_rightmost)
651 mctz->rb_rightmost = rb_prev(&mz->tree_node);
653 rb_erase(&mz->tree_node, &mctz->rb_root);
657 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
658 struct mem_cgroup_tree_per_node *mctz)
662 spin_lock_irqsave(&mctz->lock, flags);
663 __mem_cgroup_remove_exceeded(mz, mctz);
664 spin_unlock_irqrestore(&mctz->lock, flags);
667 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
669 unsigned long nr_pages = page_counter_read(&memcg->memory);
670 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
671 unsigned long excess = 0;
673 if (nr_pages > soft_limit)
674 excess = nr_pages - soft_limit;
679 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
681 unsigned long excess;
682 struct mem_cgroup_per_node *mz;
683 struct mem_cgroup_tree_per_node *mctz;
685 mctz = soft_limit_tree_from_page(page);
689 * Necessary to update all ancestors when hierarchy is used.
690 * because their event counter is not touched.
692 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
693 mz = mem_cgroup_page_nodeinfo(memcg, page);
694 excess = soft_limit_excess(memcg);
696 * We have to update the tree if mz is on RB-tree or
697 * mem is over its softlimit.
699 if (excess || mz->on_tree) {
702 spin_lock_irqsave(&mctz->lock, flags);
703 /* if on-tree, remove it */
705 __mem_cgroup_remove_exceeded(mz, mctz);
707 * Insert again. mz->usage_in_excess will be updated.
708 * If excess is 0, no tree ops.
710 __mem_cgroup_insert_exceeded(mz, mctz, excess);
711 spin_unlock_irqrestore(&mctz->lock, flags);
716 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
718 struct mem_cgroup_tree_per_node *mctz;
719 struct mem_cgroup_per_node *mz;
723 mz = mem_cgroup_nodeinfo(memcg, nid);
724 mctz = soft_limit_tree_node(nid);
726 mem_cgroup_remove_exceeded(mz, mctz);
730 static struct mem_cgroup_per_node *
731 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
733 struct mem_cgroup_per_node *mz;
737 if (!mctz->rb_rightmost)
738 goto done; /* Nothing to reclaim from */
740 mz = rb_entry(mctz->rb_rightmost,
741 struct mem_cgroup_per_node, tree_node);
743 * Remove the node now but someone else can add it back,
744 * we will to add it back at the end of reclaim to its correct
745 * position in the tree.
747 __mem_cgroup_remove_exceeded(mz, mctz);
748 if (!soft_limit_excess(mz->memcg) ||
749 !css_tryget(&mz->memcg->css))
755 static struct mem_cgroup_per_node *
756 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
758 struct mem_cgroup_per_node *mz;
760 spin_lock_irq(&mctz->lock);
761 mz = __mem_cgroup_largest_soft_limit_node(mctz);
762 spin_unlock_irq(&mctz->lock);
767 * __mod_memcg_state - update cgroup memory statistics
768 * @memcg: the memory cgroup
769 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
770 * @val: delta to add to the counter, can be negative
772 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
774 long x, threshold = MEMCG_CHARGE_BATCH;
776 if (mem_cgroup_disabled())
779 if (memcg_stat_item_in_bytes(idx))
780 threshold <<= PAGE_SHIFT;
782 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
783 if (unlikely(abs(x) > threshold)) {
784 struct mem_cgroup *mi;
787 * Batch local counters to keep them in sync with
788 * the hierarchical ones.
790 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
791 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
792 atomic_long_add(x, &mi->vmstats[idx]);
795 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
798 static struct mem_cgroup_per_node *
799 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
801 struct mem_cgroup *parent;
803 parent = parent_mem_cgroup(pn->memcg);
806 return mem_cgroup_nodeinfo(parent, nid);
809 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
812 struct mem_cgroup_per_node *pn;
813 struct mem_cgroup *memcg;
814 long x, threshold = MEMCG_CHARGE_BATCH;
816 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
820 __mod_memcg_state(memcg, idx, val);
823 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
825 if (vmstat_item_in_bytes(idx))
826 threshold <<= PAGE_SHIFT;
828 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
829 if (unlikely(abs(x) > threshold)) {
830 pg_data_t *pgdat = lruvec_pgdat(lruvec);
831 struct mem_cgroup_per_node *pi;
833 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
834 atomic_long_add(x, &pi->lruvec_stat[idx]);
837 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
841 * __mod_lruvec_state - update lruvec memory statistics
842 * @lruvec: the lruvec
843 * @idx: the stat item
844 * @val: delta to add to the counter, can be negative
846 * The lruvec is the intersection of the NUMA node and a cgroup. This
847 * function updates the all three counters that are affected by a
848 * change of state at this level: per-node, per-cgroup, per-lruvec.
850 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
854 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
856 /* Update memcg and lruvec */
857 if (!mem_cgroup_disabled())
858 __mod_memcg_lruvec_state(lruvec, idx, val);
861 void __mod_lruvec_slab_state(void *p, enum node_stat_item idx, int val)
863 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
864 struct mem_cgroup *memcg;
865 struct lruvec *lruvec;
868 memcg = mem_cgroup_from_obj(p);
871 * Untracked pages have no memcg, no lruvec. Update only the
872 * node. If we reparent the slab objects to the root memcg,
873 * when we free the slab object, we need to update the per-memcg
874 * vmstats to keep it correct for the root memcg.
877 __mod_node_page_state(pgdat, idx, val);
879 lruvec = mem_cgroup_lruvec(memcg, pgdat);
880 __mod_lruvec_state(lruvec, idx, val);
885 void mod_memcg_obj_state(void *p, int idx, int val)
887 struct mem_cgroup *memcg;
890 memcg = mem_cgroup_from_obj(p);
892 mod_memcg_state(memcg, idx, val);
897 * __count_memcg_events - account VM events in a cgroup
898 * @memcg: the memory cgroup
899 * @idx: the event item
900 * @count: the number of events that occured
902 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
907 if (mem_cgroup_disabled())
910 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
911 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
912 struct mem_cgroup *mi;
915 * Batch local counters to keep them in sync with
916 * the hierarchical ones.
918 __this_cpu_add(memcg->vmstats_local->events[idx], x);
919 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
920 atomic_long_add(x, &mi->vmevents[idx]);
923 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
926 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
928 return atomic_long_read(&memcg->vmevents[event]);
931 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
936 for_each_possible_cpu(cpu)
937 x += per_cpu(memcg->vmstats_local->events[event], cpu);
941 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
945 /* pagein of a big page is an event. So, ignore page size */
947 __count_memcg_events(memcg, PGPGIN, 1);
949 __count_memcg_events(memcg, PGPGOUT, 1);
950 nr_pages = -nr_pages; /* for event */
953 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
956 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
957 enum mem_cgroup_events_target target)
959 unsigned long val, next;
961 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
962 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
963 /* from time_after() in jiffies.h */
964 if ((long)(next - val) < 0) {
966 case MEM_CGROUP_TARGET_THRESH:
967 next = val + THRESHOLDS_EVENTS_TARGET;
969 case MEM_CGROUP_TARGET_SOFTLIMIT:
970 next = val + SOFTLIMIT_EVENTS_TARGET;
975 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
982 * Check events in order.
985 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
987 /* threshold event is triggered in finer grain than soft limit */
988 if (unlikely(mem_cgroup_event_ratelimit(memcg,
989 MEM_CGROUP_TARGET_THRESH))) {
992 do_softlimit = mem_cgroup_event_ratelimit(memcg,
993 MEM_CGROUP_TARGET_SOFTLIMIT);
994 mem_cgroup_threshold(memcg);
995 if (unlikely(do_softlimit))
996 mem_cgroup_update_tree(memcg, page);
1000 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1003 * mm_update_next_owner() may clear mm->owner to NULL
1004 * if it races with swapoff, page migration, etc.
1005 * So this can be called with p == NULL.
1010 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1012 EXPORT_SYMBOL(mem_cgroup_from_task);
1015 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1016 * @mm: mm from which memcg should be extracted. It can be NULL.
1018 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
1019 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
1022 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1024 struct mem_cgroup *memcg;
1026 if (mem_cgroup_disabled())
1032 * Page cache insertions can happen withou an
1033 * actual mm context, e.g. during disk probing
1034 * on boot, loopback IO, acct() writes etc.
1037 memcg = root_mem_cgroup;
1039 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1040 if (unlikely(!memcg))
1041 memcg = root_mem_cgroup;
1043 } while (!css_tryget(&memcg->css));
1047 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1050 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
1051 * @page: page from which memcg should be extracted.
1053 * Obtain a reference on page->memcg and returns it if successful. Otherwise
1054 * root_mem_cgroup is returned.
1056 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
1058 struct mem_cgroup *memcg = page->mem_cgroup;
1060 if (mem_cgroup_disabled())
1064 /* Page should not get uncharged and freed memcg under us. */
1065 if (!memcg || WARN_ON_ONCE(!css_tryget(&memcg->css)))
1066 memcg = root_mem_cgroup;
1070 EXPORT_SYMBOL(get_mem_cgroup_from_page);
1072 static __always_inline struct mem_cgroup *active_memcg(void)
1075 return this_cpu_read(int_active_memcg);
1077 return current->active_memcg;
1080 static __always_inline struct mem_cgroup *get_active_memcg(void)
1082 struct mem_cgroup *memcg;
1085 memcg = active_memcg();
1086 /* remote memcg must hold a ref. */
1087 if (memcg && WARN_ON_ONCE(!css_tryget(&memcg->css)))
1088 memcg = root_mem_cgroup;
1094 static __always_inline bool memcg_kmem_bypass(void)
1096 /* Allow remote memcg charging from any context. */
1097 if (unlikely(active_memcg()))
1100 /* Memcg to charge can't be determined. */
1101 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
1108 * If active memcg is set, do not fallback to current->mm->memcg.
1110 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
1112 if (memcg_kmem_bypass())
1115 if (unlikely(active_memcg()))
1116 return get_active_memcg();
1118 return get_mem_cgroup_from_mm(current->mm);
1122 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1123 * @root: hierarchy root
1124 * @prev: previously returned memcg, NULL on first invocation
1125 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1127 * Returns references to children of the hierarchy below @root, or
1128 * @root itself, or %NULL after a full round-trip.
1130 * Caller must pass the return value in @prev on subsequent
1131 * invocations for reference counting, or use mem_cgroup_iter_break()
1132 * to cancel a hierarchy walk before the round-trip is complete.
1134 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1135 * in the hierarchy among all concurrent reclaimers operating on the
1138 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1139 struct mem_cgroup *prev,
1140 struct mem_cgroup_reclaim_cookie *reclaim)
1142 struct mem_cgroup_reclaim_iter *iter;
1143 struct cgroup_subsys_state *css = NULL;
1144 struct mem_cgroup *memcg = NULL;
1145 struct mem_cgroup *pos = NULL;
1147 if (mem_cgroup_disabled())
1151 root = root_mem_cgroup;
1153 if (prev && !reclaim)
1156 if (!root->use_hierarchy && root != root_mem_cgroup) {
1165 struct mem_cgroup_per_node *mz;
1167 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1170 if (prev && reclaim->generation != iter->generation)
1174 pos = READ_ONCE(iter->position);
1175 if (!pos || css_tryget(&pos->css))
1178 * css reference reached zero, so iter->position will
1179 * be cleared by ->css_released. However, we should not
1180 * rely on this happening soon, because ->css_released
1181 * is called from a work queue, and by busy-waiting we
1182 * might block it. So we clear iter->position right
1185 (void)cmpxchg(&iter->position, pos, NULL);
1193 css = css_next_descendant_pre(css, &root->css);
1196 * Reclaimers share the hierarchy walk, and a
1197 * new one might jump in right at the end of
1198 * the hierarchy - make sure they see at least
1199 * one group and restart from the beginning.
1207 * Verify the css and acquire a reference. The root
1208 * is provided by the caller, so we know it's alive
1209 * and kicking, and don't take an extra reference.
1211 memcg = mem_cgroup_from_css(css);
1213 if (css == &root->css)
1216 if (css_tryget(css))
1224 * The position could have already been updated by a competing
1225 * thread, so check that the value hasn't changed since we read
1226 * it to avoid reclaiming from the same cgroup twice.
1228 (void)cmpxchg(&iter->position, pos, memcg);
1236 reclaim->generation = iter->generation;
1242 if (prev && prev != root)
1243 css_put(&prev->css);
1249 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1250 * @root: hierarchy root
1251 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1253 void mem_cgroup_iter_break(struct mem_cgroup *root,
1254 struct mem_cgroup *prev)
1257 root = root_mem_cgroup;
1258 if (prev && prev != root)
1259 css_put(&prev->css);
1262 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1263 struct mem_cgroup *dead_memcg)
1265 struct mem_cgroup_reclaim_iter *iter;
1266 struct mem_cgroup_per_node *mz;
1269 for_each_node(nid) {
1270 mz = mem_cgroup_nodeinfo(from, nid);
1272 cmpxchg(&iter->position, dead_memcg, NULL);
1276 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1278 struct mem_cgroup *memcg = dead_memcg;
1279 struct mem_cgroup *last;
1282 __invalidate_reclaim_iterators(memcg, dead_memcg);
1284 } while ((memcg = parent_mem_cgroup(memcg)));
1287 * When cgruop1 non-hierarchy mode is used,
1288 * parent_mem_cgroup() does not walk all the way up to the
1289 * cgroup root (root_mem_cgroup). So we have to handle
1290 * dead_memcg from cgroup root separately.
1292 if (last != root_mem_cgroup)
1293 __invalidate_reclaim_iterators(root_mem_cgroup,
1298 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1299 * @memcg: hierarchy root
1300 * @fn: function to call for each task
1301 * @arg: argument passed to @fn
1303 * This function iterates over tasks attached to @memcg or to any of its
1304 * descendants and calls @fn for each task. If @fn returns a non-zero
1305 * value, the function breaks the iteration loop and returns the value.
1306 * Otherwise, it will iterate over all tasks and return 0.
1308 * This function must not be called for the root memory cgroup.
1310 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1311 int (*fn)(struct task_struct *, void *), void *arg)
1313 struct mem_cgroup *iter;
1316 BUG_ON(memcg == root_mem_cgroup);
1318 for_each_mem_cgroup_tree(iter, memcg) {
1319 struct css_task_iter it;
1320 struct task_struct *task;
1322 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1323 while (!ret && (task = css_task_iter_next(&it)))
1324 ret = fn(task, arg);
1325 css_task_iter_end(&it);
1327 mem_cgroup_iter_break(memcg, iter);
1335 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1337 * @pgdat: pgdat of the page
1339 * This function relies on page->mem_cgroup being stable - see the
1340 * access rules in commit_charge().
1342 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1344 struct mem_cgroup_per_node *mz;
1345 struct mem_cgroup *memcg;
1346 struct lruvec *lruvec;
1348 if (mem_cgroup_disabled()) {
1349 lruvec = &pgdat->__lruvec;
1353 memcg = page->mem_cgroup;
1355 * Swapcache readahead pages are added to the LRU - and
1356 * possibly migrated - before they are charged.
1359 memcg = root_mem_cgroup;
1361 mz = mem_cgroup_page_nodeinfo(memcg, page);
1362 lruvec = &mz->lruvec;
1365 * Since a node can be onlined after the mem_cgroup was created,
1366 * we have to be prepared to initialize lruvec->zone here;
1367 * and if offlined then reonlined, we need to reinitialize it.
1369 if (unlikely(lruvec->pgdat != pgdat))
1370 lruvec->pgdat = pgdat;
1375 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1376 * @lruvec: mem_cgroup per zone lru vector
1377 * @lru: index of lru list the page is sitting on
1378 * @zid: zone id of the accounted pages
1379 * @nr_pages: positive when adding or negative when removing
1381 * This function must be called under lru_lock, just before a page is added
1382 * to or just after a page is removed from an lru list (that ordering being
1383 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1385 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1386 int zid, int nr_pages)
1388 struct mem_cgroup_per_node *mz;
1389 unsigned long *lru_size;
1392 if (mem_cgroup_disabled())
1395 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1396 lru_size = &mz->lru_zone_size[zid][lru];
1399 *lru_size += nr_pages;
1402 if (WARN_ONCE(size < 0,
1403 "%s(%p, %d, %d): lru_size %ld\n",
1404 __func__, lruvec, lru, nr_pages, size)) {
1410 *lru_size += nr_pages;
1414 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1415 * @memcg: the memory cgroup
1417 * Returns the maximum amount of memory @mem can be charged with, in
1420 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1422 unsigned long margin = 0;
1423 unsigned long count;
1424 unsigned long limit;
1426 count = page_counter_read(&memcg->memory);
1427 limit = READ_ONCE(memcg->memory.max);
1429 margin = limit - count;
1431 if (do_memsw_account()) {
1432 count = page_counter_read(&memcg->memsw);
1433 limit = READ_ONCE(memcg->memsw.max);
1435 margin = min(margin, limit - count);
1444 * A routine for checking "mem" is under move_account() or not.
1446 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1447 * moving cgroups. This is for waiting at high-memory pressure
1450 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1452 struct mem_cgroup *from;
1453 struct mem_cgroup *to;
1456 * Unlike task_move routines, we access mc.to, mc.from not under
1457 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1459 spin_lock(&mc.lock);
1465 ret = mem_cgroup_is_descendant(from, memcg) ||
1466 mem_cgroup_is_descendant(to, memcg);
1468 spin_unlock(&mc.lock);
1472 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1474 if (mc.moving_task && current != mc.moving_task) {
1475 if (mem_cgroup_under_move(memcg)) {
1477 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1478 /* moving charge context might have finished. */
1481 finish_wait(&mc.waitq, &wait);
1488 struct memory_stat {
1494 static struct memory_stat memory_stats[] = {
1495 { "anon", PAGE_SIZE, NR_ANON_MAPPED },
1496 { "file", PAGE_SIZE, NR_FILE_PAGES },
1497 { "kernel_stack", 1024, NR_KERNEL_STACK_KB },
1498 { "percpu", 1, MEMCG_PERCPU_B },
1499 { "sock", PAGE_SIZE, MEMCG_SOCK },
1500 { "shmem", PAGE_SIZE, NR_SHMEM },
1501 { "file_mapped", PAGE_SIZE, NR_FILE_MAPPED },
1502 { "file_dirty", PAGE_SIZE, NR_FILE_DIRTY },
1503 { "file_writeback", PAGE_SIZE, NR_WRITEBACK },
1504 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1506 * The ratio will be initialized in memory_stats_init(). Because
1507 * on some architectures, the macro of HPAGE_PMD_SIZE is not
1508 * constant(e.g. powerpc).
1510 { "anon_thp", 0, NR_ANON_THPS },
1512 { "inactive_anon", PAGE_SIZE, NR_INACTIVE_ANON },
1513 { "active_anon", PAGE_SIZE, NR_ACTIVE_ANON },
1514 { "inactive_file", PAGE_SIZE, NR_INACTIVE_FILE },
1515 { "active_file", PAGE_SIZE, NR_ACTIVE_FILE },
1516 { "unevictable", PAGE_SIZE, NR_UNEVICTABLE },
1519 * Note: The slab_reclaimable and slab_unreclaimable must be
1520 * together and slab_reclaimable must be in front.
1522 { "slab_reclaimable", 1, NR_SLAB_RECLAIMABLE_B },
1523 { "slab_unreclaimable", 1, NR_SLAB_UNRECLAIMABLE_B },
1525 /* The memory events */
1526 { "workingset_refault_anon", 1, WORKINGSET_REFAULT_ANON },
1527 { "workingset_refault_file", 1, WORKINGSET_REFAULT_FILE },
1528 { "workingset_activate_anon", 1, WORKINGSET_ACTIVATE_ANON },
1529 { "workingset_activate_file", 1, WORKINGSET_ACTIVATE_FILE },
1530 { "workingset_restore_anon", 1, WORKINGSET_RESTORE_ANON },
1531 { "workingset_restore_file", 1, WORKINGSET_RESTORE_FILE },
1532 { "workingset_nodereclaim", 1, WORKINGSET_NODERECLAIM },
1535 static int __init memory_stats_init(void)
1539 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1540 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1541 if (memory_stats[i].idx == NR_ANON_THPS)
1542 memory_stats[i].ratio = HPAGE_PMD_SIZE;
1544 VM_BUG_ON(!memory_stats[i].ratio);
1545 VM_BUG_ON(memory_stats[i].idx >= MEMCG_NR_STAT);
1550 pure_initcall(memory_stats_init);
1552 static char *memory_stat_format(struct mem_cgroup *memcg)
1557 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1562 * Provide statistics on the state of the memory subsystem as
1563 * well as cumulative event counters that show past behavior.
1565 * This list is ordered following a combination of these gradients:
1566 * 1) generic big picture -> specifics and details
1567 * 2) reflecting userspace activity -> reflecting kernel heuristics
1569 * Current memory state:
1572 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1575 size = memcg_page_state(memcg, memory_stats[i].idx);
1576 size *= memory_stats[i].ratio;
1577 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1579 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1580 size = memcg_page_state(memcg, NR_SLAB_RECLAIMABLE_B) +
1581 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE_B);
1582 seq_buf_printf(&s, "slab %llu\n", size);
1586 /* Accumulated memory events */
1588 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1589 memcg_events(memcg, PGFAULT));
1590 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1591 memcg_events(memcg, PGMAJFAULT));
1592 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1593 memcg_events(memcg, PGREFILL));
1594 seq_buf_printf(&s, "pgscan %lu\n",
1595 memcg_events(memcg, PGSCAN_KSWAPD) +
1596 memcg_events(memcg, PGSCAN_DIRECT));
1597 seq_buf_printf(&s, "pgsteal %lu\n",
1598 memcg_events(memcg, PGSTEAL_KSWAPD) +
1599 memcg_events(memcg, PGSTEAL_DIRECT));
1600 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1601 memcg_events(memcg, PGACTIVATE));
1602 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1603 memcg_events(memcg, PGDEACTIVATE));
1604 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1605 memcg_events(memcg, PGLAZYFREE));
1606 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1607 memcg_events(memcg, PGLAZYFREED));
1609 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1610 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1611 memcg_events(memcg, THP_FAULT_ALLOC));
1612 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1613 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1614 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1616 /* The above should easily fit into one page */
1617 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1622 #define K(x) ((x) << (PAGE_SHIFT-10))
1624 * mem_cgroup_print_oom_context: Print OOM information relevant to
1625 * memory controller.
1626 * @memcg: The memory cgroup that went over limit
1627 * @p: Task that is going to be killed
1629 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1632 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1637 pr_cont(",oom_memcg=");
1638 pr_cont_cgroup_path(memcg->css.cgroup);
1640 pr_cont(",global_oom");
1642 pr_cont(",task_memcg=");
1643 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1649 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1650 * memory controller.
1651 * @memcg: The memory cgroup that went over limit
1653 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1657 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1658 K((u64)page_counter_read(&memcg->memory)),
1659 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1660 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1661 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1662 K((u64)page_counter_read(&memcg->swap)),
1663 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1665 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1666 K((u64)page_counter_read(&memcg->memsw)),
1667 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1668 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1669 K((u64)page_counter_read(&memcg->kmem)),
1670 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1673 pr_info("Memory cgroup stats for ");
1674 pr_cont_cgroup_path(memcg->css.cgroup);
1676 buf = memory_stat_format(memcg);
1684 * Return the memory (and swap, if configured) limit for a memcg.
1686 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1688 unsigned long max = READ_ONCE(memcg->memory.max);
1690 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1691 if (mem_cgroup_swappiness(memcg))
1692 max += min(READ_ONCE(memcg->swap.max),
1693 (unsigned long)total_swap_pages);
1695 if (mem_cgroup_swappiness(memcg)) {
1696 /* Calculate swap excess capacity from memsw limit */
1697 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1699 max += min(swap, (unsigned long)total_swap_pages);
1705 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1707 return page_counter_read(&memcg->memory);
1710 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1713 struct oom_control oc = {
1717 .gfp_mask = gfp_mask,
1722 if (mutex_lock_killable(&oom_lock))
1725 if (mem_cgroup_margin(memcg) >= (1 << order))
1729 * A few threads which were not waiting at mutex_lock_killable() can
1730 * fail to bail out. Therefore, check again after holding oom_lock.
1732 ret = should_force_charge() || out_of_memory(&oc);
1735 mutex_unlock(&oom_lock);
1739 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1742 unsigned long *total_scanned)
1744 struct mem_cgroup *victim = NULL;
1747 unsigned long excess;
1748 unsigned long nr_scanned;
1749 struct mem_cgroup_reclaim_cookie reclaim = {
1753 excess = soft_limit_excess(root_memcg);
1756 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1761 * If we have not been able to reclaim
1762 * anything, it might because there are
1763 * no reclaimable pages under this hierarchy
1768 * We want to do more targeted reclaim.
1769 * excess >> 2 is not to excessive so as to
1770 * reclaim too much, nor too less that we keep
1771 * coming back to reclaim from this cgroup
1773 if (total >= (excess >> 2) ||
1774 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1779 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1780 pgdat, &nr_scanned);
1781 *total_scanned += nr_scanned;
1782 if (!soft_limit_excess(root_memcg))
1785 mem_cgroup_iter_break(root_memcg, victim);
1789 #ifdef CONFIG_LOCKDEP
1790 static struct lockdep_map memcg_oom_lock_dep_map = {
1791 .name = "memcg_oom_lock",
1795 static DEFINE_SPINLOCK(memcg_oom_lock);
1798 * Check OOM-Killer is already running under our hierarchy.
1799 * If someone is running, return false.
1801 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1803 struct mem_cgroup *iter, *failed = NULL;
1805 spin_lock(&memcg_oom_lock);
1807 for_each_mem_cgroup_tree(iter, memcg) {
1808 if (iter->oom_lock) {
1810 * this subtree of our hierarchy is already locked
1811 * so we cannot give a lock.
1814 mem_cgroup_iter_break(memcg, iter);
1817 iter->oom_lock = true;
1822 * OK, we failed to lock the whole subtree so we have
1823 * to clean up what we set up to the failing subtree
1825 for_each_mem_cgroup_tree(iter, memcg) {
1826 if (iter == failed) {
1827 mem_cgroup_iter_break(memcg, iter);
1830 iter->oom_lock = false;
1833 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1835 spin_unlock(&memcg_oom_lock);
1840 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1842 struct mem_cgroup *iter;
1844 spin_lock(&memcg_oom_lock);
1845 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1846 for_each_mem_cgroup_tree(iter, memcg)
1847 iter->oom_lock = false;
1848 spin_unlock(&memcg_oom_lock);
1851 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1853 struct mem_cgroup *iter;
1855 spin_lock(&memcg_oom_lock);
1856 for_each_mem_cgroup_tree(iter, memcg)
1858 spin_unlock(&memcg_oom_lock);
1861 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1863 struct mem_cgroup *iter;
1866 * Be careful about under_oom underflows becase a child memcg
1867 * could have been added after mem_cgroup_mark_under_oom.
1869 spin_lock(&memcg_oom_lock);
1870 for_each_mem_cgroup_tree(iter, memcg)
1871 if (iter->under_oom > 0)
1873 spin_unlock(&memcg_oom_lock);
1876 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1878 struct oom_wait_info {
1879 struct mem_cgroup *memcg;
1880 wait_queue_entry_t wait;
1883 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1884 unsigned mode, int sync, void *arg)
1886 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1887 struct mem_cgroup *oom_wait_memcg;
1888 struct oom_wait_info *oom_wait_info;
1890 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1891 oom_wait_memcg = oom_wait_info->memcg;
1893 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1894 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1896 return autoremove_wake_function(wait, mode, sync, arg);
1899 static void memcg_oom_recover(struct mem_cgroup *memcg)
1902 * For the following lockless ->under_oom test, the only required
1903 * guarantee is that it must see the state asserted by an OOM when
1904 * this function is called as a result of userland actions
1905 * triggered by the notification of the OOM. This is trivially
1906 * achieved by invoking mem_cgroup_mark_under_oom() before
1907 * triggering notification.
1909 if (memcg && memcg->under_oom)
1910 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1920 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1922 enum oom_status ret;
1925 if (order > PAGE_ALLOC_COSTLY_ORDER)
1928 memcg_memory_event(memcg, MEMCG_OOM);
1931 * We are in the middle of the charge context here, so we
1932 * don't want to block when potentially sitting on a callstack
1933 * that holds all kinds of filesystem and mm locks.
1935 * cgroup1 allows disabling the OOM killer and waiting for outside
1936 * handling until the charge can succeed; remember the context and put
1937 * the task to sleep at the end of the page fault when all locks are
1940 * On the other hand, in-kernel OOM killer allows for an async victim
1941 * memory reclaim (oom_reaper) and that means that we are not solely
1942 * relying on the oom victim to make a forward progress and we can
1943 * invoke the oom killer here.
1945 * Please note that mem_cgroup_out_of_memory might fail to find a
1946 * victim and then we have to bail out from the charge path.
1948 if (memcg->oom_kill_disable) {
1949 if (!current->in_user_fault)
1951 css_get(&memcg->css);
1952 current->memcg_in_oom = memcg;
1953 current->memcg_oom_gfp_mask = mask;
1954 current->memcg_oom_order = order;
1959 mem_cgroup_mark_under_oom(memcg);
1961 locked = mem_cgroup_oom_trylock(memcg);
1964 mem_cgroup_oom_notify(memcg);
1966 mem_cgroup_unmark_under_oom(memcg);
1967 if (mem_cgroup_out_of_memory(memcg, mask, order))
1973 mem_cgroup_oom_unlock(memcg);
1979 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1980 * @handle: actually kill/wait or just clean up the OOM state
1982 * This has to be called at the end of a page fault if the memcg OOM
1983 * handler was enabled.
1985 * Memcg supports userspace OOM handling where failed allocations must
1986 * sleep on a waitqueue until the userspace task resolves the
1987 * situation. Sleeping directly in the charge context with all kinds
1988 * of locks held is not a good idea, instead we remember an OOM state
1989 * in the task and mem_cgroup_oom_synchronize() has to be called at
1990 * the end of the page fault to complete the OOM handling.
1992 * Returns %true if an ongoing memcg OOM situation was detected and
1993 * completed, %false otherwise.
1995 bool mem_cgroup_oom_synchronize(bool handle)
1997 struct mem_cgroup *memcg = current->memcg_in_oom;
1998 struct oom_wait_info owait;
2001 /* OOM is global, do not handle */
2008 owait.memcg = memcg;
2009 owait.wait.flags = 0;
2010 owait.wait.func = memcg_oom_wake_function;
2011 owait.wait.private = current;
2012 INIT_LIST_HEAD(&owait.wait.entry);
2014 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2015 mem_cgroup_mark_under_oom(memcg);
2017 locked = mem_cgroup_oom_trylock(memcg);
2020 mem_cgroup_oom_notify(memcg);
2022 if (locked && !memcg->oom_kill_disable) {
2023 mem_cgroup_unmark_under_oom(memcg);
2024 finish_wait(&memcg_oom_waitq, &owait.wait);
2025 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
2026 current->memcg_oom_order);
2029 mem_cgroup_unmark_under_oom(memcg);
2030 finish_wait(&memcg_oom_waitq, &owait.wait);
2034 mem_cgroup_oom_unlock(memcg);
2036 * There is no guarantee that an OOM-lock contender
2037 * sees the wakeups triggered by the OOM kill
2038 * uncharges. Wake any sleepers explicitely.
2040 memcg_oom_recover(memcg);
2043 current->memcg_in_oom = NULL;
2044 css_put(&memcg->css);
2049 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2050 * @victim: task to be killed by the OOM killer
2051 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2053 * Returns a pointer to a memory cgroup, which has to be cleaned up
2054 * by killing all belonging OOM-killable tasks.
2056 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2058 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2059 struct mem_cgroup *oom_domain)
2061 struct mem_cgroup *oom_group = NULL;
2062 struct mem_cgroup *memcg;
2064 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2068 oom_domain = root_mem_cgroup;
2072 memcg = mem_cgroup_from_task(victim);
2073 if (memcg == root_mem_cgroup)
2077 * If the victim task has been asynchronously moved to a different
2078 * memory cgroup, we might end up killing tasks outside oom_domain.
2079 * In this case it's better to ignore memory.group.oom.
2081 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2085 * Traverse the memory cgroup hierarchy from the victim task's
2086 * cgroup up to the OOMing cgroup (or root) to find the
2087 * highest-level memory cgroup with oom.group set.
2089 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2090 if (memcg->oom_group)
2093 if (memcg == oom_domain)
2098 css_get(&oom_group->css);
2105 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2107 pr_info("Tasks in ");
2108 pr_cont_cgroup_path(memcg->css.cgroup);
2109 pr_cont(" are going to be killed due to memory.oom.group set\n");
2113 * lock_page_memcg - lock a page->mem_cgroup binding
2116 * This function protects unlocked LRU pages from being moved to
2119 * It ensures lifetime of the returned memcg. Caller is responsible
2120 * for the lifetime of the page; __unlock_page_memcg() is available
2121 * when @page might get freed inside the locked section.
2123 struct mem_cgroup *lock_page_memcg(struct page *page)
2125 struct page *head = compound_head(page); /* rmap on tail pages */
2126 struct mem_cgroup *memcg;
2127 unsigned long flags;
2130 * The RCU lock is held throughout the transaction. The fast
2131 * path can get away without acquiring the memcg->move_lock
2132 * because page moving starts with an RCU grace period.
2134 * The RCU lock also protects the memcg from being freed when
2135 * the page state that is going to change is the only thing
2136 * preventing the page itself from being freed. E.g. writeback
2137 * doesn't hold a page reference and relies on PG_writeback to
2138 * keep off truncation, migration and so forth.
2142 if (mem_cgroup_disabled())
2145 memcg = head->mem_cgroup;
2146 if (unlikely(!memcg))
2149 if (atomic_read(&memcg->moving_account) <= 0)
2152 spin_lock_irqsave(&memcg->move_lock, flags);
2153 if (memcg != head->mem_cgroup) {
2154 spin_unlock_irqrestore(&memcg->move_lock, flags);
2159 * When charge migration first begins, we can have locked and
2160 * unlocked page stat updates happening concurrently. Track
2161 * the task who has the lock for unlock_page_memcg().
2163 memcg->move_lock_task = current;
2164 memcg->move_lock_flags = flags;
2168 EXPORT_SYMBOL(lock_page_memcg);
2171 * __unlock_page_memcg - unlock and unpin a memcg
2174 * Unlock and unpin a memcg returned by lock_page_memcg().
2176 void __unlock_page_memcg(struct mem_cgroup *memcg)
2178 if (memcg && memcg->move_lock_task == current) {
2179 unsigned long flags = memcg->move_lock_flags;
2181 memcg->move_lock_task = NULL;
2182 memcg->move_lock_flags = 0;
2184 spin_unlock_irqrestore(&memcg->move_lock, flags);
2191 * unlock_page_memcg - unlock a page->mem_cgroup binding
2194 void unlock_page_memcg(struct page *page)
2196 struct page *head = compound_head(page);
2198 __unlock_page_memcg(head->mem_cgroup);
2200 EXPORT_SYMBOL(unlock_page_memcg);
2202 struct memcg_stock_pcp {
2203 struct mem_cgroup *cached; /* this never be root cgroup */
2204 unsigned int nr_pages;
2206 #ifdef CONFIG_MEMCG_KMEM
2207 struct obj_cgroup *cached_objcg;
2208 unsigned int nr_bytes;
2211 struct work_struct work;
2212 unsigned long flags;
2213 #define FLUSHING_CACHED_CHARGE 0
2215 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2216 static DEFINE_MUTEX(percpu_charge_mutex);
2218 #ifdef CONFIG_MEMCG_KMEM
2219 static void drain_obj_stock(struct memcg_stock_pcp *stock);
2220 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2221 struct mem_cgroup *root_memcg);
2224 static inline void drain_obj_stock(struct memcg_stock_pcp *stock)
2227 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2228 struct mem_cgroup *root_memcg)
2235 * consume_stock: Try to consume stocked charge on this cpu.
2236 * @memcg: memcg to consume from.
2237 * @nr_pages: how many pages to charge.
2239 * The charges will only happen if @memcg matches the current cpu's memcg
2240 * stock, and at least @nr_pages are available in that stock. Failure to
2241 * service an allocation will refill the stock.
2243 * returns true if successful, false otherwise.
2245 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2247 struct memcg_stock_pcp *stock;
2248 unsigned long flags;
2251 if (nr_pages > MEMCG_CHARGE_BATCH)
2254 local_irq_save(flags);
2256 stock = this_cpu_ptr(&memcg_stock);
2257 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2258 stock->nr_pages -= nr_pages;
2262 local_irq_restore(flags);
2268 * Returns stocks cached in percpu and reset cached information.
2270 static void drain_stock(struct memcg_stock_pcp *stock)
2272 struct mem_cgroup *old = stock->cached;
2277 if (stock->nr_pages) {
2278 page_counter_uncharge(&old->memory, stock->nr_pages);
2279 if (do_memsw_account())
2280 page_counter_uncharge(&old->memsw, stock->nr_pages);
2281 stock->nr_pages = 0;
2285 stock->cached = NULL;
2288 static void drain_local_stock(struct work_struct *dummy)
2290 struct memcg_stock_pcp *stock;
2291 unsigned long flags;
2294 * The only protection from memory hotplug vs. drain_stock races is
2295 * that we always operate on local CPU stock here with IRQ disabled
2297 local_irq_save(flags);
2299 stock = this_cpu_ptr(&memcg_stock);
2300 drain_obj_stock(stock);
2302 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2304 local_irq_restore(flags);
2308 * Cache charges(val) to local per_cpu area.
2309 * This will be consumed by consume_stock() function, later.
2311 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2313 struct memcg_stock_pcp *stock;
2314 unsigned long flags;
2316 local_irq_save(flags);
2318 stock = this_cpu_ptr(&memcg_stock);
2319 if (stock->cached != memcg) { /* reset if necessary */
2321 css_get(&memcg->css);
2322 stock->cached = memcg;
2324 stock->nr_pages += nr_pages;
2326 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2329 local_irq_restore(flags);
2333 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2334 * of the hierarchy under it.
2336 static void drain_all_stock(struct mem_cgroup *root_memcg)
2340 /* If someone's already draining, avoid adding running more workers. */
2341 if (!mutex_trylock(&percpu_charge_mutex))
2344 * Notify other cpus that system-wide "drain" is running
2345 * We do not care about races with the cpu hotplug because cpu down
2346 * as well as workers from this path always operate on the local
2347 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2350 for_each_online_cpu(cpu) {
2351 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2352 struct mem_cgroup *memcg;
2356 memcg = stock->cached;
2357 if (memcg && stock->nr_pages &&
2358 mem_cgroup_is_descendant(memcg, root_memcg))
2360 if (obj_stock_flush_required(stock, root_memcg))
2365 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2367 drain_local_stock(&stock->work);
2369 schedule_work_on(cpu, &stock->work);
2373 mutex_unlock(&percpu_charge_mutex);
2376 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2378 struct memcg_stock_pcp *stock;
2379 struct mem_cgroup *memcg, *mi;
2381 stock = &per_cpu(memcg_stock, cpu);
2384 for_each_mem_cgroup(memcg) {
2387 for (i = 0; i < MEMCG_NR_STAT; i++) {
2391 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2393 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2394 atomic_long_add(x, &memcg->vmstats[i]);
2396 if (i >= NR_VM_NODE_STAT_ITEMS)
2399 for_each_node(nid) {
2400 struct mem_cgroup_per_node *pn;
2402 pn = mem_cgroup_nodeinfo(memcg, nid);
2403 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2406 atomic_long_add(x, &pn->lruvec_stat[i]);
2407 } while ((pn = parent_nodeinfo(pn, nid)));
2411 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2414 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2416 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2417 atomic_long_add(x, &memcg->vmevents[i]);
2424 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2425 unsigned int nr_pages,
2428 unsigned long nr_reclaimed = 0;
2431 unsigned long pflags;
2433 if (page_counter_read(&memcg->memory) <=
2434 READ_ONCE(memcg->memory.high))
2437 memcg_memory_event(memcg, MEMCG_HIGH);
2439 psi_memstall_enter(&pflags);
2440 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2442 psi_memstall_leave(&pflags);
2443 } while ((memcg = parent_mem_cgroup(memcg)) &&
2444 !mem_cgroup_is_root(memcg));
2446 return nr_reclaimed;
2449 static void high_work_func(struct work_struct *work)
2451 struct mem_cgroup *memcg;
2453 memcg = container_of(work, struct mem_cgroup, high_work);
2454 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2458 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2459 * enough to still cause a significant slowdown in most cases, while still
2460 * allowing diagnostics and tracing to proceed without becoming stuck.
2462 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2465 * When calculating the delay, we use these either side of the exponentiation to
2466 * maintain precision and scale to a reasonable number of jiffies (see the table
2469 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2470 * overage ratio to a delay.
2471 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2472 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2473 * to produce a reasonable delay curve.
2475 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2476 * reasonable delay curve compared to precision-adjusted overage, not
2477 * penalising heavily at first, but still making sure that growth beyond the
2478 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2479 * example, with a high of 100 megabytes:
2481 * +-------+------------------------+
2482 * | usage | time to allocate in ms |
2483 * +-------+------------------------+
2505 * +-------+------------------------+
2507 #define MEMCG_DELAY_PRECISION_SHIFT 20
2508 #define MEMCG_DELAY_SCALING_SHIFT 14
2510 static u64 calculate_overage(unsigned long usage, unsigned long high)
2518 * Prevent division by 0 in overage calculation by acting as if
2519 * it was a threshold of 1 page
2521 high = max(high, 1UL);
2523 overage = usage - high;
2524 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2525 return div64_u64(overage, high);
2528 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2530 u64 overage, max_overage = 0;
2533 overage = calculate_overage(page_counter_read(&memcg->memory),
2534 READ_ONCE(memcg->memory.high));
2535 max_overage = max(overage, max_overage);
2536 } while ((memcg = parent_mem_cgroup(memcg)) &&
2537 !mem_cgroup_is_root(memcg));
2542 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2544 u64 overage, max_overage = 0;
2547 overage = calculate_overage(page_counter_read(&memcg->swap),
2548 READ_ONCE(memcg->swap.high));
2550 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2551 max_overage = max(overage, max_overage);
2552 } while ((memcg = parent_mem_cgroup(memcg)) &&
2553 !mem_cgroup_is_root(memcg));
2559 * Get the number of jiffies that we should penalise a mischievous cgroup which
2560 * is exceeding its memory.high by checking both it and its ancestors.
2562 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2563 unsigned int nr_pages,
2566 unsigned long penalty_jiffies;
2572 * We use overage compared to memory.high to calculate the number of
2573 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2574 * fairly lenient on small overages, and increasingly harsh when the
2575 * memcg in question makes it clear that it has no intention of stopping
2576 * its crazy behaviour, so we exponentially increase the delay based on
2579 penalty_jiffies = max_overage * max_overage * HZ;
2580 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2581 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2584 * Factor in the task's own contribution to the overage, such that four
2585 * N-sized allocations are throttled approximately the same as one
2586 * 4N-sized allocation.
2588 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2589 * larger the current charge patch is than that.
2591 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2595 * Scheduled by try_charge() to be executed from the userland return path
2596 * and reclaims memory over the high limit.
2598 void mem_cgroup_handle_over_high(void)
2600 unsigned long penalty_jiffies;
2601 unsigned long pflags;
2602 unsigned long nr_reclaimed;
2603 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2604 int nr_retries = MAX_RECLAIM_RETRIES;
2605 struct mem_cgroup *memcg;
2606 bool in_retry = false;
2608 if (likely(!nr_pages))
2611 memcg = get_mem_cgroup_from_mm(current->mm);
2612 current->memcg_nr_pages_over_high = 0;
2616 * The allocating task should reclaim at least the batch size, but for
2617 * subsequent retries we only want to do what's necessary to prevent oom
2618 * or breaching resource isolation.
2620 * This is distinct from memory.max or page allocator behaviour because
2621 * memory.high is currently batched, whereas memory.max and the page
2622 * allocator run every time an allocation is made.
2624 nr_reclaimed = reclaim_high(memcg,
2625 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2629 * memory.high is breached and reclaim is unable to keep up. Throttle
2630 * allocators proactively to slow down excessive growth.
2632 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2633 mem_find_max_overage(memcg));
2635 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2636 swap_find_max_overage(memcg));
2639 * Clamp the max delay per usermode return so as to still keep the
2640 * application moving forwards and also permit diagnostics, albeit
2643 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2646 * Don't sleep if the amount of jiffies this memcg owes us is so low
2647 * that it's not even worth doing, in an attempt to be nice to those who
2648 * go only a small amount over their memory.high value and maybe haven't
2649 * been aggressively reclaimed enough yet.
2651 if (penalty_jiffies <= HZ / 100)
2655 * If reclaim is making forward progress but we're still over
2656 * memory.high, we want to encourage that rather than doing allocator
2659 if (nr_reclaimed || nr_retries--) {
2665 * If we exit early, we're guaranteed to die (since
2666 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2667 * need to account for any ill-begotten jiffies to pay them off later.
2669 psi_memstall_enter(&pflags);
2670 schedule_timeout_killable(penalty_jiffies);
2671 psi_memstall_leave(&pflags);
2674 css_put(&memcg->css);
2677 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2678 unsigned int nr_pages)
2680 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2681 int nr_retries = MAX_RECLAIM_RETRIES;
2682 struct mem_cgroup *mem_over_limit;
2683 struct page_counter *counter;
2684 enum oom_status oom_status;
2685 unsigned long nr_reclaimed;
2686 bool may_swap = true;
2687 bool drained = false;
2688 unsigned long pflags;
2690 if (mem_cgroup_is_root(memcg))
2693 if (consume_stock(memcg, nr_pages))
2696 if (!do_memsw_account() ||
2697 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2698 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2700 if (do_memsw_account())
2701 page_counter_uncharge(&memcg->memsw, batch);
2702 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2704 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2708 if (batch > nr_pages) {
2714 * Memcg doesn't have a dedicated reserve for atomic
2715 * allocations. But like the global atomic pool, we need to
2716 * put the burden of reclaim on regular allocation requests
2717 * and let these go through as privileged allocations.
2719 if (gfp_mask & __GFP_ATOMIC)
2723 * Unlike in global OOM situations, memcg is not in a physical
2724 * memory shortage. Allow dying and OOM-killed tasks to
2725 * bypass the last charges so that they can exit quickly and
2726 * free their memory.
2728 if (unlikely(should_force_charge()))
2732 * Prevent unbounded recursion when reclaim operations need to
2733 * allocate memory. This might exceed the limits temporarily,
2734 * but we prefer facilitating memory reclaim and getting back
2735 * under the limit over triggering OOM kills in these cases.
2737 if (unlikely(current->flags & PF_MEMALLOC))
2740 if (unlikely(task_in_memcg_oom(current)))
2743 if (!gfpflags_allow_blocking(gfp_mask))
2746 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2748 psi_memstall_enter(&pflags);
2749 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2750 gfp_mask, may_swap);
2751 psi_memstall_leave(&pflags);
2753 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2757 drain_all_stock(mem_over_limit);
2762 if (gfp_mask & __GFP_NORETRY)
2765 * Even though the limit is exceeded at this point, reclaim
2766 * may have been able to free some pages. Retry the charge
2767 * before killing the task.
2769 * Only for regular pages, though: huge pages are rather
2770 * unlikely to succeed so close to the limit, and we fall back
2771 * to regular pages anyway in case of failure.
2773 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2776 * At task move, charge accounts can be doubly counted. So, it's
2777 * better to wait until the end of task_move if something is going on.
2779 if (mem_cgroup_wait_acct_move(mem_over_limit))
2785 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2788 if (gfp_mask & __GFP_NOFAIL)
2791 if (fatal_signal_pending(current))
2795 * keep retrying as long as the memcg oom killer is able to make
2796 * a forward progress or bypass the charge if the oom killer
2797 * couldn't make any progress.
2799 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2800 get_order(nr_pages * PAGE_SIZE));
2801 switch (oom_status) {
2803 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_irq_save(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_irq_restore(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_irq_save(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_irq_restore(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 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3548 return mem_cgroup_from_css(css)->use_hierarchy;
3551 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3552 struct cftype *cft, u64 val)
3555 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3556 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3558 if (memcg->use_hierarchy == val)
3562 * If parent's use_hierarchy is set, we can't make any modifications
3563 * in the child subtrees. If it is unset, then the change can
3564 * occur, provided the current cgroup has no children.
3566 * For the root cgroup, parent_mem is NULL, we allow value to be
3567 * set if there are no children.
3569 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3570 (val == 1 || val == 0)) {
3571 if (!memcg_has_children(memcg))
3572 memcg->use_hierarchy = val;
3581 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3585 if (mem_cgroup_is_root(memcg)) {
3586 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3587 memcg_page_state(memcg, NR_ANON_MAPPED);
3589 val += memcg_page_state(memcg, MEMCG_SWAP);
3592 val = page_counter_read(&memcg->memory);
3594 val = page_counter_read(&memcg->memsw);
3607 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3610 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3611 struct page_counter *counter;
3613 switch (MEMFILE_TYPE(cft->private)) {
3615 counter = &memcg->memory;
3618 counter = &memcg->memsw;
3621 counter = &memcg->kmem;
3624 counter = &memcg->tcpmem;
3630 switch (MEMFILE_ATTR(cft->private)) {
3632 if (counter == &memcg->memory)
3633 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3634 if (counter == &memcg->memsw)
3635 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3636 return (u64)page_counter_read(counter) * PAGE_SIZE;
3638 return (u64)counter->max * PAGE_SIZE;
3640 return (u64)counter->watermark * PAGE_SIZE;
3642 return counter->failcnt;
3643 case RES_SOFT_LIMIT:
3644 return (u64)memcg->soft_limit * PAGE_SIZE;
3650 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3652 unsigned long stat[MEMCG_NR_STAT] = {0};
3653 struct mem_cgroup *mi;
3656 for_each_online_cpu(cpu)
3657 for (i = 0; i < MEMCG_NR_STAT; i++)
3658 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3660 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3661 for (i = 0; i < MEMCG_NR_STAT; i++)
3662 atomic_long_add(stat[i], &mi->vmstats[i]);
3664 for_each_node(node) {
3665 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3666 struct mem_cgroup_per_node *pi;
3668 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3671 for_each_online_cpu(cpu)
3672 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3674 pn->lruvec_stat_cpu->count[i], cpu);
3676 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3677 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3678 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3682 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3684 unsigned long events[NR_VM_EVENT_ITEMS];
3685 struct mem_cgroup *mi;
3688 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3691 for_each_online_cpu(cpu)
3692 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3693 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3696 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3697 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3698 atomic_long_add(events[i], &mi->vmevents[i]);
3701 #ifdef CONFIG_MEMCG_KMEM
3702 static int memcg_online_kmem(struct mem_cgroup *memcg)
3704 struct obj_cgroup *objcg;
3707 if (cgroup_memory_nokmem)
3710 BUG_ON(memcg->kmemcg_id >= 0);
3711 BUG_ON(memcg->kmem_state);
3713 memcg_id = memcg_alloc_cache_id();
3717 objcg = obj_cgroup_alloc();
3719 memcg_free_cache_id(memcg_id);
3722 objcg->memcg = memcg;
3723 rcu_assign_pointer(memcg->objcg, objcg);
3725 static_branch_enable(&memcg_kmem_enabled_key);
3728 * A memory cgroup is considered kmem-online as soon as it gets
3729 * kmemcg_id. Setting the id after enabling static branching will
3730 * guarantee no one starts accounting before all call sites are
3733 memcg->kmemcg_id = memcg_id;
3734 memcg->kmem_state = KMEM_ONLINE;
3739 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3741 struct cgroup_subsys_state *css;
3742 struct mem_cgroup *parent, *child;
3745 if (memcg->kmem_state != KMEM_ONLINE)
3748 memcg->kmem_state = KMEM_ALLOCATED;
3750 parent = parent_mem_cgroup(memcg);
3752 parent = root_mem_cgroup;
3754 memcg_reparent_objcgs(memcg, parent);
3756 kmemcg_id = memcg->kmemcg_id;
3757 BUG_ON(kmemcg_id < 0);
3760 * Change kmemcg_id of this cgroup and all its descendants to the
3761 * parent's id, and then move all entries from this cgroup's list_lrus
3762 * to ones of the parent. After we have finished, all list_lrus
3763 * corresponding to this cgroup are guaranteed to remain empty. The
3764 * ordering is imposed by list_lru_node->lock taken by
3765 * memcg_drain_all_list_lrus().
3767 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3768 css_for_each_descendant_pre(css, &memcg->css) {
3769 child = mem_cgroup_from_css(css);
3770 BUG_ON(child->kmemcg_id != kmemcg_id);
3771 child->kmemcg_id = parent->kmemcg_id;
3772 if (!memcg->use_hierarchy)
3777 memcg_drain_all_list_lrus(kmemcg_id, parent);
3779 memcg_free_cache_id(kmemcg_id);
3782 static void memcg_free_kmem(struct mem_cgroup *memcg)
3784 /* css_alloc() failed, offlining didn't happen */
3785 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3786 memcg_offline_kmem(memcg);
3789 static int memcg_online_kmem(struct mem_cgroup *memcg)
3793 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3796 static void memcg_free_kmem(struct mem_cgroup *memcg)
3799 #endif /* CONFIG_MEMCG_KMEM */
3801 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3806 mutex_lock(&memcg_max_mutex);
3807 ret = page_counter_set_max(&memcg->kmem, max);
3808 mutex_unlock(&memcg_max_mutex);
3812 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3816 mutex_lock(&memcg_max_mutex);
3818 ret = page_counter_set_max(&memcg->tcpmem, max);
3822 if (!memcg->tcpmem_active) {
3824 * The active flag needs to be written after the static_key
3825 * update. This is what guarantees that the socket activation
3826 * function is the last one to run. See mem_cgroup_sk_alloc()
3827 * for details, and note that we don't mark any socket as
3828 * belonging to this memcg until that flag is up.
3830 * We need to do this, because static_keys will span multiple
3831 * sites, but we can't control their order. If we mark a socket
3832 * as accounted, but the accounting functions are not patched in
3833 * yet, we'll lose accounting.
3835 * We never race with the readers in mem_cgroup_sk_alloc(),
3836 * because when this value change, the code to process it is not
3839 static_branch_inc(&memcg_sockets_enabled_key);
3840 memcg->tcpmem_active = true;
3843 mutex_unlock(&memcg_max_mutex);
3848 * The user of this function is...
3851 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3852 char *buf, size_t nbytes, loff_t off)
3854 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3855 unsigned long nr_pages;
3858 buf = strstrip(buf);
3859 ret = page_counter_memparse(buf, "-1", &nr_pages);
3863 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3865 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3869 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3871 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3874 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3877 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3878 "Please report your usecase to linux-mm@kvack.org if you "
3879 "depend on this functionality.\n");
3880 ret = memcg_update_kmem_max(memcg, nr_pages);
3883 ret = memcg_update_tcp_max(memcg, nr_pages);
3887 case RES_SOFT_LIMIT:
3888 memcg->soft_limit = nr_pages;
3892 return ret ?: nbytes;
3895 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3896 size_t nbytes, loff_t off)
3898 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3899 struct page_counter *counter;
3901 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3903 counter = &memcg->memory;
3906 counter = &memcg->memsw;
3909 counter = &memcg->kmem;
3912 counter = &memcg->tcpmem;
3918 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3920 page_counter_reset_watermark(counter);
3923 counter->failcnt = 0;
3932 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3935 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3939 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3940 struct cftype *cft, u64 val)
3942 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3944 if (val & ~MOVE_MASK)
3948 * No kind of locking is needed in here, because ->can_attach() will
3949 * check this value once in the beginning of the process, and then carry
3950 * on with stale data. This means that changes to this value will only
3951 * affect task migrations starting after the change.
3953 memcg->move_charge_at_immigrate = val;
3957 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3958 struct cftype *cft, u64 val)
3966 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3967 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3968 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3970 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3971 int nid, unsigned int lru_mask, bool tree)
3973 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3974 unsigned long nr = 0;
3977 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3980 if (!(BIT(lru) & lru_mask))
3983 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3985 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3990 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3991 unsigned int lru_mask,
3994 unsigned long nr = 0;
3998 if (!(BIT(lru) & lru_mask))
4001 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
4003 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
4008 static int memcg_numa_stat_show(struct seq_file *m, void *v)
4012 unsigned int lru_mask;
4015 static const struct numa_stat stats[] = {
4016 { "total", LRU_ALL },
4017 { "file", LRU_ALL_FILE },
4018 { "anon", LRU_ALL_ANON },
4019 { "unevictable", BIT(LRU_UNEVICTABLE) },
4021 const struct numa_stat *stat;
4023 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4025 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4026 seq_printf(m, "%s=%lu", stat->name,
4027 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4029 for_each_node_state(nid, N_MEMORY)
4030 seq_printf(m, " N%d=%lu", nid,
4031 mem_cgroup_node_nr_lru_pages(memcg, nid,
4032 stat->lru_mask, false));
4036 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4038 seq_printf(m, "hierarchical_%s=%lu", stat->name,
4039 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4041 for_each_node_state(nid, N_MEMORY)
4042 seq_printf(m, " N%d=%lu", nid,
4043 mem_cgroup_node_nr_lru_pages(memcg, nid,
4044 stat->lru_mask, true));
4050 #endif /* CONFIG_NUMA */
4052 static const unsigned int memcg1_stats[] = {
4055 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4065 static const char *const memcg1_stat_names[] = {
4068 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4078 /* Universal VM events cgroup1 shows, original sort order */
4079 static const unsigned int memcg1_events[] = {
4086 static int memcg_stat_show(struct seq_file *m, void *v)
4088 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4089 unsigned long memory, memsw;
4090 struct mem_cgroup *mi;
4093 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4095 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4098 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4100 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4101 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4102 if (memcg1_stats[i] == NR_ANON_THPS)
4105 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4108 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4109 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4110 memcg_events_local(memcg, memcg1_events[i]));
4112 for (i = 0; i < NR_LRU_LISTS; i++)
4113 seq_printf(m, "%s %lu\n", lru_list_name(i),
4114 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4117 /* Hierarchical information */
4118 memory = memsw = PAGE_COUNTER_MAX;
4119 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4120 memory = min(memory, READ_ONCE(mi->memory.max));
4121 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4123 seq_printf(m, "hierarchical_memory_limit %llu\n",
4124 (u64)memory * PAGE_SIZE);
4125 if (do_memsw_account())
4126 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4127 (u64)memsw * PAGE_SIZE);
4129 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4132 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4134 nr = memcg_page_state(memcg, memcg1_stats[i]);
4135 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4136 if (memcg1_stats[i] == NR_ANON_THPS)
4139 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4140 (u64)nr * PAGE_SIZE);
4143 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4144 seq_printf(m, "total_%s %llu\n",
4145 vm_event_name(memcg1_events[i]),
4146 (u64)memcg_events(memcg, memcg1_events[i]));
4148 for (i = 0; i < NR_LRU_LISTS; i++)
4149 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4150 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4153 #ifdef CONFIG_DEBUG_VM
4156 struct mem_cgroup_per_node *mz;
4157 unsigned long anon_cost = 0;
4158 unsigned long file_cost = 0;
4160 for_each_online_pgdat(pgdat) {
4161 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
4163 anon_cost += mz->lruvec.anon_cost;
4164 file_cost += mz->lruvec.file_cost;
4166 seq_printf(m, "anon_cost %lu\n", anon_cost);
4167 seq_printf(m, "file_cost %lu\n", file_cost);
4174 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4177 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4179 return mem_cgroup_swappiness(memcg);
4182 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4183 struct cftype *cft, u64 val)
4185 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4191 memcg->swappiness = val;
4193 vm_swappiness = val;
4198 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4200 struct mem_cgroup_threshold_ary *t;
4201 unsigned long usage;
4206 t = rcu_dereference(memcg->thresholds.primary);
4208 t = rcu_dereference(memcg->memsw_thresholds.primary);
4213 usage = mem_cgroup_usage(memcg, swap);
4216 * current_threshold points to threshold just below or equal to usage.
4217 * If it's not true, a threshold was crossed after last
4218 * call of __mem_cgroup_threshold().
4220 i = t->current_threshold;
4223 * Iterate backward over array of thresholds starting from
4224 * current_threshold and check if a threshold is crossed.
4225 * If none of thresholds below usage is crossed, we read
4226 * only one element of the array here.
4228 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4229 eventfd_signal(t->entries[i].eventfd, 1);
4231 /* i = current_threshold + 1 */
4235 * Iterate forward over array of thresholds starting from
4236 * current_threshold+1 and check if a threshold is crossed.
4237 * If none of thresholds above usage is crossed, we read
4238 * only one element of the array here.
4240 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4241 eventfd_signal(t->entries[i].eventfd, 1);
4243 /* Update current_threshold */
4244 t->current_threshold = i - 1;
4249 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4252 __mem_cgroup_threshold(memcg, false);
4253 if (do_memsw_account())
4254 __mem_cgroup_threshold(memcg, true);
4256 memcg = parent_mem_cgroup(memcg);
4260 static int compare_thresholds(const void *a, const void *b)
4262 const struct mem_cgroup_threshold *_a = a;
4263 const struct mem_cgroup_threshold *_b = b;
4265 if (_a->threshold > _b->threshold)
4268 if (_a->threshold < _b->threshold)
4274 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4276 struct mem_cgroup_eventfd_list *ev;
4278 spin_lock(&memcg_oom_lock);
4280 list_for_each_entry(ev, &memcg->oom_notify, list)
4281 eventfd_signal(ev->eventfd, 1);
4283 spin_unlock(&memcg_oom_lock);
4287 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4289 struct mem_cgroup *iter;
4291 for_each_mem_cgroup_tree(iter, memcg)
4292 mem_cgroup_oom_notify_cb(iter);
4295 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4296 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4298 struct mem_cgroup_thresholds *thresholds;
4299 struct mem_cgroup_threshold_ary *new;
4300 unsigned long threshold;
4301 unsigned long usage;
4304 ret = page_counter_memparse(args, "-1", &threshold);
4308 mutex_lock(&memcg->thresholds_lock);
4311 thresholds = &memcg->thresholds;
4312 usage = mem_cgroup_usage(memcg, false);
4313 } else if (type == _MEMSWAP) {
4314 thresholds = &memcg->memsw_thresholds;
4315 usage = mem_cgroup_usage(memcg, true);
4319 /* Check if a threshold crossed before adding a new one */
4320 if (thresholds->primary)
4321 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4323 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4325 /* Allocate memory for new array of thresholds */
4326 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4333 /* Copy thresholds (if any) to new array */
4334 if (thresholds->primary)
4335 memcpy(new->entries, thresholds->primary->entries,
4336 flex_array_size(new, entries, size - 1));
4338 /* Add new threshold */
4339 new->entries[size - 1].eventfd = eventfd;
4340 new->entries[size - 1].threshold = threshold;
4342 /* Sort thresholds. Registering of new threshold isn't time-critical */
4343 sort(new->entries, size, sizeof(*new->entries),
4344 compare_thresholds, NULL);
4346 /* Find current threshold */
4347 new->current_threshold = -1;
4348 for (i = 0; i < size; i++) {
4349 if (new->entries[i].threshold <= usage) {
4351 * new->current_threshold will not be used until
4352 * rcu_assign_pointer(), so it's safe to increment
4355 ++new->current_threshold;
4360 /* Free old spare buffer and save old primary buffer as spare */
4361 kfree(thresholds->spare);
4362 thresholds->spare = thresholds->primary;
4364 rcu_assign_pointer(thresholds->primary, new);
4366 /* To be sure that nobody uses thresholds */
4370 mutex_unlock(&memcg->thresholds_lock);
4375 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4376 struct eventfd_ctx *eventfd, const char *args)
4378 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4381 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4382 struct eventfd_ctx *eventfd, const char *args)
4384 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4387 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4388 struct eventfd_ctx *eventfd, enum res_type type)
4390 struct mem_cgroup_thresholds *thresholds;
4391 struct mem_cgroup_threshold_ary *new;
4392 unsigned long usage;
4393 int i, j, size, entries;
4395 mutex_lock(&memcg->thresholds_lock);
4398 thresholds = &memcg->thresholds;
4399 usage = mem_cgroup_usage(memcg, false);
4400 } else if (type == _MEMSWAP) {
4401 thresholds = &memcg->memsw_thresholds;
4402 usage = mem_cgroup_usage(memcg, true);
4406 if (!thresholds->primary)
4409 /* Check if a threshold crossed before removing */
4410 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4412 /* Calculate new number of threshold */
4414 for (i = 0; i < thresholds->primary->size; i++) {
4415 if (thresholds->primary->entries[i].eventfd != eventfd)
4421 new = thresholds->spare;
4423 /* If no items related to eventfd have been cleared, nothing to do */
4427 /* Set thresholds array to NULL if we don't have thresholds */
4436 /* Copy thresholds and find current threshold */
4437 new->current_threshold = -1;
4438 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4439 if (thresholds->primary->entries[i].eventfd == eventfd)
4442 new->entries[j] = thresholds->primary->entries[i];
4443 if (new->entries[j].threshold <= usage) {
4445 * new->current_threshold will not be used
4446 * until rcu_assign_pointer(), so it's safe to increment
4449 ++new->current_threshold;
4455 /* Swap primary and spare array */
4456 thresholds->spare = thresholds->primary;
4458 rcu_assign_pointer(thresholds->primary, new);
4460 /* To be sure that nobody uses thresholds */
4463 /* If all events are unregistered, free the spare array */
4465 kfree(thresholds->spare);
4466 thresholds->spare = NULL;
4469 mutex_unlock(&memcg->thresholds_lock);
4472 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4473 struct eventfd_ctx *eventfd)
4475 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4478 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4479 struct eventfd_ctx *eventfd)
4481 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4484 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4485 struct eventfd_ctx *eventfd, const char *args)
4487 struct mem_cgroup_eventfd_list *event;
4489 event = kmalloc(sizeof(*event), GFP_KERNEL);
4493 spin_lock(&memcg_oom_lock);
4495 event->eventfd = eventfd;
4496 list_add(&event->list, &memcg->oom_notify);
4498 /* already in OOM ? */
4499 if (memcg->under_oom)
4500 eventfd_signal(eventfd, 1);
4501 spin_unlock(&memcg_oom_lock);
4506 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4507 struct eventfd_ctx *eventfd)
4509 struct mem_cgroup_eventfd_list *ev, *tmp;
4511 spin_lock(&memcg_oom_lock);
4513 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4514 if (ev->eventfd == eventfd) {
4515 list_del(&ev->list);
4520 spin_unlock(&memcg_oom_lock);
4523 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4525 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4527 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4528 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4529 seq_printf(sf, "oom_kill %lu\n",
4530 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4534 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4535 struct cftype *cft, u64 val)
4537 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4539 /* cannot set to root cgroup and only 0 and 1 are allowed */
4540 if (!css->parent || !((val == 0) || (val == 1)))
4543 memcg->oom_kill_disable = val;
4545 memcg_oom_recover(memcg);
4550 #ifdef CONFIG_CGROUP_WRITEBACK
4552 #include <trace/events/writeback.h>
4554 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4556 return wb_domain_init(&memcg->cgwb_domain, gfp);
4559 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4561 wb_domain_exit(&memcg->cgwb_domain);
4564 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4566 wb_domain_size_changed(&memcg->cgwb_domain);
4569 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4571 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4573 if (!memcg->css.parent)
4576 return &memcg->cgwb_domain;
4580 * idx can be of type enum memcg_stat_item or node_stat_item.
4581 * Keep in sync with memcg_exact_page().
4583 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4585 long x = atomic_long_read(&memcg->vmstats[idx]);
4588 for_each_online_cpu(cpu)
4589 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4596 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4597 * @wb: bdi_writeback in question
4598 * @pfilepages: out parameter for number of file pages
4599 * @pheadroom: out parameter for number of allocatable pages according to memcg
4600 * @pdirty: out parameter for number of dirty pages
4601 * @pwriteback: out parameter for number of pages under writeback
4603 * Determine the numbers of file, headroom, dirty, and writeback pages in
4604 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4605 * is a bit more involved.
4607 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4608 * headroom is calculated as the lowest headroom of itself and the
4609 * ancestors. Note that this doesn't consider the actual amount of
4610 * available memory in the system. The caller should further cap
4611 * *@pheadroom accordingly.
4613 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4614 unsigned long *pheadroom, unsigned long *pdirty,
4615 unsigned long *pwriteback)
4617 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4618 struct mem_cgroup *parent;
4620 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4622 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4623 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4624 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4625 *pheadroom = PAGE_COUNTER_MAX;
4627 while ((parent = parent_mem_cgroup(memcg))) {
4628 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4629 READ_ONCE(memcg->memory.high));
4630 unsigned long used = page_counter_read(&memcg->memory);
4632 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4638 * Foreign dirty flushing
4640 * There's an inherent mismatch between memcg and writeback. The former
4641 * trackes ownership per-page while the latter per-inode. This was a
4642 * deliberate design decision because honoring per-page ownership in the
4643 * writeback path is complicated, may lead to higher CPU and IO overheads
4644 * and deemed unnecessary given that write-sharing an inode across
4645 * different cgroups isn't a common use-case.
4647 * Combined with inode majority-writer ownership switching, this works well
4648 * enough in most cases but there are some pathological cases. For
4649 * example, let's say there are two cgroups A and B which keep writing to
4650 * different but confined parts of the same inode. B owns the inode and
4651 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4652 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4653 * triggering background writeback. A will be slowed down without a way to
4654 * make writeback of the dirty pages happen.
4656 * Conditions like the above can lead to a cgroup getting repatedly and
4657 * severely throttled after making some progress after each
4658 * dirty_expire_interval while the underyling IO device is almost
4661 * Solving this problem completely requires matching the ownership tracking
4662 * granularities between memcg and writeback in either direction. However,
4663 * the more egregious behaviors can be avoided by simply remembering the
4664 * most recent foreign dirtying events and initiating remote flushes on
4665 * them when local writeback isn't enough to keep the memory clean enough.
4667 * The following two functions implement such mechanism. When a foreign
4668 * page - a page whose memcg and writeback ownerships don't match - is
4669 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4670 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4671 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4672 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4673 * foreign bdi_writebacks which haven't expired. Both the numbers of
4674 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4675 * limited to MEMCG_CGWB_FRN_CNT.
4677 * The mechanism only remembers IDs and doesn't hold any object references.
4678 * As being wrong occasionally doesn't matter, updates and accesses to the
4679 * records are lockless and racy.
4681 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4682 struct bdi_writeback *wb)
4684 struct mem_cgroup *memcg = page->mem_cgroup;
4685 struct memcg_cgwb_frn *frn;
4686 u64 now = get_jiffies_64();
4687 u64 oldest_at = now;
4691 trace_track_foreign_dirty(page, wb);
4694 * Pick the slot to use. If there is already a slot for @wb, keep
4695 * using it. If not replace the oldest one which isn't being
4698 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4699 frn = &memcg->cgwb_frn[i];
4700 if (frn->bdi_id == wb->bdi->id &&
4701 frn->memcg_id == wb->memcg_css->id)
4703 if (time_before64(frn->at, oldest_at) &&
4704 atomic_read(&frn->done.cnt) == 1) {
4706 oldest_at = frn->at;
4710 if (i < MEMCG_CGWB_FRN_CNT) {
4712 * Re-using an existing one. Update timestamp lazily to
4713 * avoid making the cacheline hot. We want them to be
4714 * reasonably up-to-date and significantly shorter than
4715 * dirty_expire_interval as that's what expires the record.
4716 * Use the shorter of 1s and dirty_expire_interval / 8.
4718 unsigned long update_intv =
4719 min_t(unsigned long, HZ,
4720 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4722 if (time_before64(frn->at, now - update_intv))
4724 } else if (oldest >= 0) {
4725 /* replace the oldest free one */
4726 frn = &memcg->cgwb_frn[oldest];
4727 frn->bdi_id = wb->bdi->id;
4728 frn->memcg_id = wb->memcg_css->id;
4733 /* issue foreign writeback flushes for recorded foreign dirtying events */
4734 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4736 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4737 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4738 u64 now = jiffies_64;
4741 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4742 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4745 * If the record is older than dirty_expire_interval,
4746 * writeback on it has already started. No need to kick it
4747 * off again. Also, don't start a new one if there's
4748 * already one in flight.
4750 if (time_after64(frn->at, now - intv) &&
4751 atomic_read(&frn->done.cnt) == 1) {
4753 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4754 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4755 WB_REASON_FOREIGN_FLUSH,
4761 #else /* CONFIG_CGROUP_WRITEBACK */
4763 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4768 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4772 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4776 #endif /* CONFIG_CGROUP_WRITEBACK */
4779 * DO NOT USE IN NEW FILES.
4781 * "cgroup.event_control" implementation.
4783 * This is way over-engineered. It tries to support fully configurable
4784 * events for each user. Such level of flexibility is completely
4785 * unnecessary especially in the light of the planned unified hierarchy.
4787 * Please deprecate this and replace with something simpler if at all
4792 * Unregister event and free resources.
4794 * Gets called from workqueue.
4796 static void memcg_event_remove(struct work_struct *work)
4798 struct mem_cgroup_event *event =
4799 container_of(work, struct mem_cgroup_event, remove);
4800 struct mem_cgroup *memcg = event->memcg;
4802 remove_wait_queue(event->wqh, &event->wait);
4804 event->unregister_event(memcg, event->eventfd);
4806 /* Notify userspace the event is going away. */
4807 eventfd_signal(event->eventfd, 1);
4809 eventfd_ctx_put(event->eventfd);
4811 css_put(&memcg->css);
4815 * Gets called on EPOLLHUP on eventfd when user closes it.
4817 * Called with wqh->lock held and interrupts disabled.
4819 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4820 int sync, void *key)
4822 struct mem_cgroup_event *event =
4823 container_of(wait, struct mem_cgroup_event, wait);
4824 struct mem_cgroup *memcg = event->memcg;
4825 __poll_t flags = key_to_poll(key);
4827 if (flags & EPOLLHUP) {
4829 * If the event has been detached at cgroup removal, we
4830 * can simply return knowing the other side will cleanup
4833 * We can't race against event freeing since the other
4834 * side will require wqh->lock via remove_wait_queue(),
4837 spin_lock(&memcg->event_list_lock);
4838 if (!list_empty(&event->list)) {
4839 list_del_init(&event->list);
4841 * We are in atomic context, but cgroup_event_remove()
4842 * may sleep, so we have to call it in workqueue.
4844 schedule_work(&event->remove);
4846 spin_unlock(&memcg->event_list_lock);
4852 static void memcg_event_ptable_queue_proc(struct file *file,
4853 wait_queue_head_t *wqh, poll_table *pt)
4855 struct mem_cgroup_event *event =
4856 container_of(pt, struct mem_cgroup_event, pt);
4859 add_wait_queue(wqh, &event->wait);
4863 * DO NOT USE IN NEW FILES.
4865 * Parse input and register new cgroup event handler.
4867 * Input must be in format '<event_fd> <control_fd> <args>'.
4868 * Interpretation of args is defined by control file implementation.
4870 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4871 char *buf, size_t nbytes, loff_t off)
4873 struct cgroup_subsys_state *css = of_css(of);
4874 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4875 struct mem_cgroup_event *event;
4876 struct cgroup_subsys_state *cfile_css;
4877 unsigned int efd, cfd;
4884 buf = strstrip(buf);
4886 efd = simple_strtoul(buf, &endp, 10);
4891 cfd = simple_strtoul(buf, &endp, 10);
4892 if ((*endp != ' ') && (*endp != '\0'))
4896 event = kzalloc(sizeof(*event), GFP_KERNEL);
4900 event->memcg = memcg;
4901 INIT_LIST_HEAD(&event->list);
4902 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4903 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4904 INIT_WORK(&event->remove, memcg_event_remove);
4912 event->eventfd = eventfd_ctx_fileget(efile.file);
4913 if (IS_ERR(event->eventfd)) {
4914 ret = PTR_ERR(event->eventfd);
4921 goto out_put_eventfd;
4924 /* the process need read permission on control file */
4925 /* AV: shouldn't we check that it's been opened for read instead? */
4926 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4931 * Determine the event callbacks and set them in @event. This used
4932 * to be done via struct cftype but cgroup core no longer knows
4933 * about these events. The following is crude but the whole thing
4934 * is for compatibility anyway.
4936 * DO NOT ADD NEW FILES.
4938 name = cfile.file->f_path.dentry->d_name.name;
4940 if (!strcmp(name, "memory.usage_in_bytes")) {
4941 event->register_event = mem_cgroup_usage_register_event;
4942 event->unregister_event = mem_cgroup_usage_unregister_event;
4943 } else if (!strcmp(name, "memory.oom_control")) {
4944 event->register_event = mem_cgroup_oom_register_event;
4945 event->unregister_event = mem_cgroup_oom_unregister_event;
4946 } else if (!strcmp(name, "memory.pressure_level")) {
4947 event->register_event = vmpressure_register_event;
4948 event->unregister_event = vmpressure_unregister_event;
4949 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4950 event->register_event = memsw_cgroup_usage_register_event;
4951 event->unregister_event = memsw_cgroup_usage_unregister_event;
4958 * Verify @cfile should belong to @css. Also, remaining events are
4959 * automatically removed on cgroup destruction but the removal is
4960 * asynchronous, so take an extra ref on @css.
4962 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4963 &memory_cgrp_subsys);
4965 if (IS_ERR(cfile_css))
4967 if (cfile_css != css) {
4972 ret = event->register_event(memcg, event->eventfd, buf);
4976 vfs_poll(efile.file, &event->pt);
4978 spin_lock(&memcg->event_list_lock);
4979 list_add(&event->list, &memcg->event_list);
4980 spin_unlock(&memcg->event_list_lock);
4992 eventfd_ctx_put(event->eventfd);
5001 static struct cftype mem_cgroup_legacy_files[] = {
5003 .name = "usage_in_bytes",
5004 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5005 .read_u64 = mem_cgroup_read_u64,
5008 .name = "max_usage_in_bytes",
5009 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5010 .write = mem_cgroup_reset,
5011 .read_u64 = mem_cgroup_read_u64,
5014 .name = "limit_in_bytes",
5015 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5016 .write = mem_cgroup_write,
5017 .read_u64 = mem_cgroup_read_u64,
5020 .name = "soft_limit_in_bytes",
5021 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5022 .write = mem_cgroup_write,
5023 .read_u64 = mem_cgroup_read_u64,
5027 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5028 .write = mem_cgroup_reset,
5029 .read_u64 = mem_cgroup_read_u64,
5033 .seq_show = memcg_stat_show,
5036 .name = "force_empty",
5037 .write = mem_cgroup_force_empty_write,
5040 .name = "use_hierarchy",
5041 .write_u64 = mem_cgroup_hierarchy_write,
5042 .read_u64 = mem_cgroup_hierarchy_read,
5045 .name = "cgroup.event_control", /* XXX: for compat */
5046 .write = memcg_write_event_control,
5047 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
5050 .name = "swappiness",
5051 .read_u64 = mem_cgroup_swappiness_read,
5052 .write_u64 = mem_cgroup_swappiness_write,
5055 .name = "move_charge_at_immigrate",
5056 .read_u64 = mem_cgroup_move_charge_read,
5057 .write_u64 = mem_cgroup_move_charge_write,
5060 .name = "oom_control",
5061 .seq_show = mem_cgroup_oom_control_read,
5062 .write_u64 = mem_cgroup_oom_control_write,
5063 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5066 .name = "pressure_level",
5070 .name = "numa_stat",
5071 .seq_show = memcg_numa_stat_show,
5075 .name = "kmem.limit_in_bytes",
5076 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5077 .write = mem_cgroup_write,
5078 .read_u64 = mem_cgroup_read_u64,
5081 .name = "kmem.usage_in_bytes",
5082 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5083 .read_u64 = mem_cgroup_read_u64,
5086 .name = "kmem.failcnt",
5087 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5088 .write = mem_cgroup_reset,
5089 .read_u64 = mem_cgroup_read_u64,
5092 .name = "kmem.max_usage_in_bytes",
5093 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5094 .write = mem_cgroup_reset,
5095 .read_u64 = mem_cgroup_read_u64,
5097 #if defined(CONFIG_MEMCG_KMEM) && \
5098 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5100 .name = "kmem.slabinfo",
5101 .seq_show = memcg_slab_show,
5105 .name = "kmem.tcp.limit_in_bytes",
5106 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5107 .write = mem_cgroup_write,
5108 .read_u64 = mem_cgroup_read_u64,
5111 .name = "kmem.tcp.usage_in_bytes",
5112 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5113 .read_u64 = mem_cgroup_read_u64,
5116 .name = "kmem.tcp.failcnt",
5117 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5118 .write = mem_cgroup_reset,
5119 .read_u64 = mem_cgroup_read_u64,
5122 .name = "kmem.tcp.max_usage_in_bytes",
5123 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5124 .write = mem_cgroup_reset,
5125 .read_u64 = mem_cgroup_read_u64,
5127 { }, /* terminate */
5131 * Private memory cgroup IDR
5133 * Swap-out records and page cache shadow entries need to store memcg
5134 * references in constrained space, so we maintain an ID space that is
5135 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5136 * memory-controlled cgroups to 64k.
5138 * However, there usually are many references to the offline CSS after
5139 * the cgroup has been destroyed, such as page cache or reclaimable
5140 * slab objects, that don't need to hang on to the ID. We want to keep
5141 * those dead CSS from occupying IDs, or we might quickly exhaust the
5142 * relatively small ID space and prevent the creation of new cgroups
5143 * even when there are much fewer than 64k cgroups - possibly none.
5145 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5146 * be freed and recycled when it's no longer needed, which is usually
5147 * when the CSS is offlined.
5149 * The only exception to that are records of swapped out tmpfs/shmem
5150 * pages that need to be attributed to live ancestors on swapin. But
5151 * those references are manageable from userspace.
5154 static DEFINE_IDR(mem_cgroup_idr);
5156 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5158 if (memcg->id.id > 0) {
5159 idr_remove(&mem_cgroup_idr, memcg->id.id);
5164 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5167 refcount_add(n, &memcg->id.ref);
5170 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5172 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5173 mem_cgroup_id_remove(memcg);
5175 /* Memcg ID pins CSS */
5176 css_put(&memcg->css);
5180 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5182 mem_cgroup_id_put_many(memcg, 1);
5186 * mem_cgroup_from_id - look up a memcg from a memcg id
5187 * @id: the memcg id to look up
5189 * Caller must hold rcu_read_lock().
5191 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5193 WARN_ON_ONCE(!rcu_read_lock_held());
5194 return idr_find(&mem_cgroup_idr, id);
5197 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5199 struct mem_cgroup_per_node *pn;
5202 * This routine is called against possible nodes.
5203 * But it's BUG to call kmalloc() against offline node.
5205 * TODO: this routine can waste much memory for nodes which will
5206 * never be onlined. It's better to use memory hotplug callback
5209 if (!node_state(node, N_NORMAL_MEMORY))
5211 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5215 pn->lruvec_stat_local = alloc_percpu_gfp(struct lruvec_stat,
5216 GFP_KERNEL_ACCOUNT);
5217 if (!pn->lruvec_stat_local) {
5222 pn->lruvec_stat_cpu = alloc_percpu_gfp(struct lruvec_stat,
5223 GFP_KERNEL_ACCOUNT);
5224 if (!pn->lruvec_stat_cpu) {
5225 free_percpu(pn->lruvec_stat_local);
5230 lruvec_init(&pn->lruvec);
5231 pn->usage_in_excess = 0;
5232 pn->on_tree = false;
5235 memcg->nodeinfo[node] = pn;
5239 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5241 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5246 free_percpu(pn->lruvec_stat_cpu);
5247 free_percpu(pn->lruvec_stat_local);
5251 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5256 free_mem_cgroup_per_node_info(memcg, node);
5257 free_percpu(memcg->vmstats_percpu);
5258 free_percpu(memcg->vmstats_local);
5262 static void mem_cgroup_free(struct mem_cgroup *memcg)
5264 memcg_wb_domain_exit(memcg);
5266 * Flush percpu vmstats and vmevents to guarantee the value correctness
5267 * on parent's and all ancestor levels.
5269 memcg_flush_percpu_vmstats(memcg);
5270 memcg_flush_percpu_vmevents(memcg);
5271 __mem_cgroup_free(memcg);
5274 static struct mem_cgroup *mem_cgroup_alloc(void)
5276 struct mem_cgroup *memcg;
5279 int __maybe_unused i;
5280 long error = -ENOMEM;
5282 size = sizeof(struct mem_cgroup);
5283 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5285 memcg = kzalloc(size, GFP_KERNEL);
5287 return ERR_PTR(error);
5289 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5290 1, MEM_CGROUP_ID_MAX,
5292 if (memcg->id.id < 0) {
5293 error = memcg->id.id;
5297 memcg->vmstats_local = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5298 GFP_KERNEL_ACCOUNT);
5299 if (!memcg->vmstats_local)
5302 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5303 GFP_KERNEL_ACCOUNT);
5304 if (!memcg->vmstats_percpu)
5308 if (alloc_mem_cgroup_per_node_info(memcg, node))
5311 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5314 INIT_WORK(&memcg->high_work, high_work_func);
5315 INIT_LIST_HEAD(&memcg->oom_notify);
5316 mutex_init(&memcg->thresholds_lock);
5317 spin_lock_init(&memcg->move_lock);
5318 vmpressure_init(&memcg->vmpressure);
5319 INIT_LIST_HEAD(&memcg->event_list);
5320 spin_lock_init(&memcg->event_list_lock);
5321 memcg->socket_pressure = jiffies;
5322 #ifdef CONFIG_MEMCG_KMEM
5323 memcg->kmemcg_id = -1;
5324 INIT_LIST_HEAD(&memcg->objcg_list);
5326 #ifdef CONFIG_CGROUP_WRITEBACK
5327 INIT_LIST_HEAD(&memcg->cgwb_list);
5328 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5329 memcg->cgwb_frn[i].done =
5330 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5332 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5333 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5334 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5335 memcg->deferred_split_queue.split_queue_len = 0;
5337 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5340 mem_cgroup_id_remove(memcg);
5341 __mem_cgroup_free(memcg);
5342 return ERR_PTR(error);
5345 static struct cgroup_subsys_state * __ref
5346 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5348 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5349 struct mem_cgroup *memcg, *old_memcg;
5350 long error = -ENOMEM;
5352 old_memcg = set_active_memcg(parent);
5353 memcg = mem_cgroup_alloc();
5354 set_active_memcg(old_memcg);
5356 return ERR_CAST(memcg);
5358 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5359 memcg->soft_limit = PAGE_COUNTER_MAX;
5360 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5362 memcg->swappiness = mem_cgroup_swappiness(parent);
5363 memcg->oom_kill_disable = parent->oom_kill_disable;
5366 page_counter_init(&memcg->memory, NULL);
5367 page_counter_init(&memcg->swap, NULL);
5368 page_counter_init(&memcg->kmem, NULL);
5369 page_counter_init(&memcg->tcpmem, NULL);
5370 } else if (parent->use_hierarchy) {
5371 memcg->use_hierarchy = true;
5372 page_counter_init(&memcg->memory, &parent->memory);
5373 page_counter_init(&memcg->swap, &parent->swap);
5374 page_counter_init(&memcg->kmem, &parent->kmem);
5375 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5377 page_counter_init(&memcg->memory, &root_mem_cgroup->memory);
5378 page_counter_init(&memcg->swap, &root_mem_cgroup->swap);
5379 page_counter_init(&memcg->kmem, &root_mem_cgroup->kmem);
5380 page_counter_init(&memcg->tcpmem, &root_mem_cgroup->tcpmem);
5382 * Deeper hierachy with use_hierarchy == false doesn't make
5383 * much sense so let cgroup subsystem know about this
5384 * unfortunate state in our controller.
5386 if (parent != root_mem_cgroup)
5387 memory_cgrp_subsys.broken_hierarchy = true;
5390 /* The following stuff does not apply to the root */
5392 root_mem_cgroup = memcg;
5396 error = memcg_online_kmem(memcg);
5400 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5401 static_branch_inc(&memcg_sockets_enabled_key);
5405 mem_cgroup_id_remove(memcg);
5406 mem_cgroup_free(memcg);
5407 return ERR_PTR(error);
5410 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5412 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5415 * A memcg must be visible for memcg_expand_shrinker_maps()
5416 * by the time the maps are allocated. So, we allocate maps
5417 * here, when for_each_mem_cgroup() can't skip it.
5419 if (memcg_alloc_shrinker_maps(memcg)) {
5420 mem_cgroup_id_remove(memcg);
5424 /* Online state pins memcg ID, memcg ID pins CSS */
5425 refcount_set(&memcg->id.ref, 1);
5430 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5432 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5433 struct mem_cgroup_event *event, *tmp;
5436 * Unregister events and notify userspace.
5437 * Notify userspace about cgroup removing only after rmdir of cgroup
5438 * directory to avoid race between userspace and kernelspace.
5440 spin_lock(&memcg->event_list_lock);
5441 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5442 list_del_init(&event->list);
5443 schedule_work(&event->remove);
5445 spin_unlock(&memcg->event_list_lock);
5447 page_counter_set_min(&memcg->memory, 0);
5448 page_counter_set_low(&memcg->memory, 0);
5450 memcg_offline_kmem(memcg);
5451 wb_memcg_offline(memcg);
5453 drain_all_stock(memcg);
5455 mem_cgroup_id_put(memcg);
5458 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5460 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5462 invalidate_reclaim_iterators(memcg);
5465 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5467 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5468 int __maybe_unused i;
5470 #ifdef CONFIG_CGROUP_WRITEBACK
5471 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5472 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5474 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5475 static_branch_dec(&memcg_sockets_enabled_key);
5477 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5478 static_branch_dec(&memcg_sockets_enabled_key);
5480 vmpressure_cleanup(&memcg->vmpressure);
5481 cancel_work_sync(&memcg->high_work);
5482 mem_cgroup_remove_from_trees(memcg);
5483 memcg_free_shrinker_maps(memcg);
5484 memcg_free_kmem(memcg);
5485 mem_cgroup_free(memcg);
5489 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5490 * @css: the target css
5492 * Reset the states of the mem_cgroup associated with @css. This is
5493 * invoked when the userland requests disabling on the default hierarchy
5494 * but the memcg is pinned through dependency. The memcg should stop
5495 * applying policies and should revert to the vanilla state as it may be
5496 * made visible again.
5498 * The current implementation only resets the essential configurations.
5499 * This needs to be expanded to cover all the visible parts.
5501 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5503 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5505 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5506 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5507 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5508 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5509 page_counter_set_min(&memcg->memory, 0);
5510 page_counter_set_low(&memcg->memory, 0);
5511 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5512 memcg->soft_limit = PAGE_COUNTER_MAX;
5513 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5514 memcg_wb_domain_size_changed(memcg);
5518 /* Handlers for move charge at task migration. */
5519 static int mem_cgroup_do_precharge(unsigned long count)
5523 /* Try a single bulk charge without reclaim first, kswapd may wake */
5524 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5526 mc.precharge += count;
5530 /* Try charges one by one with reclaim, but do not retry */
5532 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5546 enum mc_target_type {
5553 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5554 unsigned long addr, pte_t ptent)
5556 struct page *page = vm_normal_page(vma, addr, ptent);
5558 if (!page || !page_mapped(page))
5560 if (PageAnon(page)) {
5561 if (!(mc.flags & MOVE_ANON))
5564 if (!(mc.flags & MOVE_FILE))
5567 if (!get_page_unless_zero(page))
5573 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5574 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5575 pte_t ptent, swp_entry_t *entry)
5577 struct page *page = NULL;
5578 swp_entry_t ent = pte_to_swp_entry(ptent);
5580 if (!(mc.flags & MOVE_ANON))
5584 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5585 * a device and because they are not accessible by CPU they are store
5586 * as special swap entry in the CPU page table.
5588 if (is_device_private_entry(ent)) {
5589 page = device_private_entry_to_page(ent);
5591 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5592 * a refcount of 1 when free (unlike normal page)
5594 if (!page_ref_add_unless(page, 1, 1))
5599 if (non_swap_entry(ent))
5603 * Because lookup_swap_cache() updates some statistics counter,
5604 * we call find_get_page() with swapper_space directly.
5606 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5607 entry->val = ent.val;
5612 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5613 pte_t ptent, swp_entry_t *entry)
5619 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5620 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5622 if (!vma->vm_file) /* anonymous vma */
5624 if (!(mc.flags & MOVE_FILE))
5627 /* page is moved even if it's not RSS of this task(page-faulted). */
5628 /* shmem/tmpfs may report page out on swap: account for that too. */
5629 return find_get_incore_page(vma->vm_file->f_mapping,
5630 linear_page_index(vma, addr));
5634 * mem_cgroup_move_account - move account of the page
5636 * @compound: charge the page as compound or small page
5637 * @from: mem_cgroup which the page is moved from.
5638 * @to: mem_cgroup which the page is moved to. @from != @to.
5640 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5642 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5645 static int mem_cgroup_move_account(struct page *page,
5647 struct mem_cgroup *from,
5648 struct mem_cgroup *to)
5650 struct lruvec *from_vec, *to_vec;
5651 struct pglist_data *pgdat;
5652 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5655 VM_BUG_ON(from == to);
5656 VM_BUG_ON_PAGE(PageLRU(page), page);
5657 VM_BUG_ON(compound && !PageTransHuge(page));
5660 * Prevent mem_cgroup_migrate() from looking at
5661 * page->mem_cgroup of its source page while we change it.
5664 if (!trylock_page(page))
5668 if (page->mem_cgroup != from)
5671 pgdat = page_pgdat(page);
5672 from_vec = mem_cgroup_lruvec(from, pgdat);
5673 to_vec = mem_cgroup_lruvec(to, pgdat);
5675 lock_page_memcg(page);
5677 if (PageAnon(page)) {
5678 if (page_mapped(page)) {
5679 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5680 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5681 if (PageTransHuge(page)) {
5682 __dec_lruvec_state(from_vec, NR_ANON_THPS);
5683 __inc_lruvec_state(to_vec, NR_ANON_THPS);
5688 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5689 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5691 if (PageSwapBacked(page)) {
5692 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5693 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5696 if (page_mapped(page)) {
5697 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5698 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5701 if (PageDirty(page)) {
5702 struct address_space *mapping = page_mapping(page);
5704 if (mapping_can_writeback(mapping)) {
5705 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5707 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5713 if (PageWriteback(page)) {
5714 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5715 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5719 * All state has been migrated, let's switch to the new memcg.
5721 * It is safe to change page->mem_cgroup here because the page
5722 * is referenced, charged, isolated, and locked: we can't race
5723 * with (un)charging, migration, LRU putback, or anything else
5724 * that would rely on a stable page->mem_cgroup.
5726 * Note that lock_page_memcg is a memcg lock, not a page lock,
5727 * to save space. As soon as we switch page->mem_cgroup to a
5728 * new memcg that isn't locked, the above state can change
5729 * concurrently again. Make sure we're truly done with it.
5734 css_put(&from->css);
5736 page->mem_cgroup = to;
5738 __unlock_page_memcg(from);
5742 local_irq_disable();
5743 mem_cgroup_charge_statistics(to, page, nr_pages);
5744 memcg_check_events(to, page);
5745 mem_cgroup_charge_statistics(from, page, -nr_pages);
5746 memcg_check_events(from, page);
5755 * get_mctgt_type - get target type of moving charge
5756 * @vma: the vma the pte to be checked belongs
5757 * @addr: the address corresponding to the pte to be checked
5758 * @ptent: the pte to be checked
5759 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5762 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5763 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5764 * move charge. if @target is not NULL, the page is stored in target->page
5765 * with extra refcnt got(Callers should handle it).
5766 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5767 * target for charge migration. if @target is not NULL, the entry is stored
5769 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5770 * (so ZONE_DEVICE page and thus not on the lru).
5771 * For now we such page is charge like a regular page would be as for all
5772 * intent and purposes it is just special memory taking the place of a
5775 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5777 * Called with pte lock held.
5780 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5781 unsigned long addr, pte_t ptent, union mc_target *target)
5783 struct page *page = NULL;
5784 enum mc_target_type ret = MC_TARGET_NONE;
5785 swp_entry_t ent = { .val = 0 };
5787 if (pte_present(ptent))
5788 page = mc_handle_present_pte(vma, addr, ptent);
5789 else if (is_swap_pte(ptent))
5790 page = mc_handle_swap_pte(vma, ptent, &ent);
5791 else if (pte_none(ptent))
5792 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5794 if (!page && !ent.val)
5798 * Do only loose check w/o serialization.
5799 * mem_cgroup_move_account() checks the page is valid or
5800 * not under LRU exclusion.
5802 if (page->mem_cgroup == mc.from) {
5803 ret = MC_TARGET_PAGE;
5804 if (is_device_private_page(page))
5805 ret = MC_TARGET_DEVICE;
5807 target->page = page;
5809 if (!ret || !target)
5813 * There is a swap entry and a page doesn't exist or isn't charged.
5814 * But we cannot move a tail-page in a THP.
5816 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5817 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5818 ret = MC_TARGET_SWAP;
5825 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5827 * We don't consider PMD mapped swapping or file mapped pages because THP does
5828 * not support them for now.
5829 * Caller should make sure that pmd_trans_huge(pmd) is true.
5831 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5832 unsigned long addr, pmd_t pmd, union mc_target *target)
5834 struct page *page = NULL;
5835 enum mc_target_type ret = MC_TARGET_NONE;
5837 if (unlikely(is_swap_pmd(pmd))) {
5838 VM_BUG_ON(thp_migration_supported() &&
5839 !is_pmd_migration_entry(pmd));
5842 page = pmd_page(pmd);
5843 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5844 if (!(mc.flags & MOVE_ANON))
5846 if (page->mem_cgroup == mc.from) {
5847 ret = MC_TARGET_PAGE;
5850 target->page = page;
5856 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5857 unsigned long addr, pmd_t pmd, union mc_target *target)
5859 return MC_TARGET_NONE;
5863 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5864 unsigned long addr, unsigned long end,
5865 struct mm_walk *walk)
5867 struct vm_area_struct *vma = walk->vma;
5871 ptl = pmd_trans_huge_lock(pmd, vma);
5874 * Note their can not be MC_TARGET_DEVICE for now as we do not
5875 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5876 * this might change.
5878 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5879 mc.precharge += HPAGE_PMD_NR;
5884 if (pmd_trans_unstable(pmd))
5886 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5887 for (; addr != end; pte++, addr += PAGE_SIZE)
5888 if (get_mctgt_type(vma, addr, *pte, NULL))
5889 mc.precharge++; /* increment precharge temporarily */
5890 pte_unmap_unlock(pte - 1, ptl);
5896 static const struct mm_walk_ops precharge_walk_ops = {
5897 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5900 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5902 unsigned long precharge;
5905 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5906 mmap_read_unlock(mm);
5908 precharge = mc.precharge;
5914 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5916 unsigned long precharge = mem_cgroup_count_precharge(mm);
5918 VM_BUG_ON(mc.moving_task);
5919 mc.moving_task = current;
5920 return mem_cgroup_do_precharge(precharge);
5923 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5924 static void __mem_cgroup_clear_mc(void)
5926 struct mem_cgroup *from = mc.from;
5927 struct mem_cgroup *to = mc.to;
5929 /* we must uncharge all the leftover precharges from mc.to */
5931 cancel_charge(mc.to, mc.precharge);
5935 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5936 * we must uncharge here.
5938 if (mc.moved_charge) {
5939 cancel_charge(mc.from, mc.moved_charge);
5940 mc.moved_charge = 0;
5942 /* we must fixup refcnts and charges */
5943 if (mc.moved_swap) {
5944 /* uncharge swap account from the old cgroup */
5945 if (!mem_cgroup_is_root(mc.from))
5946 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5948 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5951 * we charged both to->memory and to->memsw, so we
5952 * should uncharge to->memory.
5954 if (!mem_cgroup_is_root(mc.to))
5955 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5959 memcg_oom_recover(from);
5960 memcg_oom_recover(to);
5961 wake_up_all(&mc.waitq);
5964 static void mem_cgroup_clear_mc(void)
5966 struct mm_struct *mm = mc.mm;
5969 * we must clear moving_task before waking up waiters at the end of
5972 mc.moving_task = NULL;
5973 __mem_cgroup_clear_mc();
5974 spin_lock(&mc.lock);
5978 spin_unlock(&mc.lock);
5983 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5985 struct cgroup_subsys_state *css;
5986 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5987 struct mem_cgroup *from;
5988 struct task_struct *leader, *p;
5989 struct mm_struct *mm;
5990 unsigned long move_flags;
5993 /* charge immigration isn't supported on the default hierarchy */
5994 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5998 * Multi-process migrations only happen on the default hierarchy
5999 * where charge immigration is not used. Perform charge
6000 * immigration if @tset contains a leader and whine if there are
6004 cgroup_taskset_for_each_leader(leader, css, tset) {
6007 memcg = mem_cgroup_from_css(css);
6013 * We are now commited to this value whatever it is. Changes in this
6014 * tunable will only affect upcoming migrations, not the current one.
6015 * So we need to save it, and keep it going.
6017 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
6021 from = mem_cgroup_from_task(p);
6023 VM_BUG_ON(from == memcg);
6025 mm = get_task_mm(p);
6028 /* We move charges only when we move a owner of the mm */
6029 if (mm->owner == p) {
6032 VM_BUG_ON(mc.precharge);
6033 VM_BUG_ON(mc.moved_charge);
6034 VM_BUG_ON(mc.moved_swap);
6036 spin_lock(&mc.lock);
6040 mc.flags = move_flags;
6041 spin_unlock(&mc.lock);
6042 /* We set mc.moving_task later */
6044 ret = mem_cgroup_precharge_mc(mm);
6046 mem_cgroup_clear_mc();
6053 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6056 mem_cgroup_clear_mc();
6059 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6060 unsigned long addr, unsigned long end,
6061 struct mm_walk *walk)
6064 struct vm_area_struct *vma = walk->vma;
6067 enum mc_target_type target_type;
6068 union mc_target target;
6071 ptl = pmd_trans_huge_lock(pmd, vma);
6073 if (mc.precharge < HPAGE_PMD_NR) {
6077 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6078 if (target_type == MC_TARGET_PAGE) {
6080 if (!isolate_lru_page(page)) {
6081 if (!mem_cgroup_move_account(page, true,
6083 mc.precharge -= HPAGE_PMD_NR;
6084 mc.moved_charge += HPAGE_PMD_NR;
6086 putback_lru_page(page);
6089 } else if (target_type == MC_TARGET_DEVICE) {
6091 if (!mem_cgroup_move_account(page, true,
6093 mc.precharge -= HPAGE_PMD_NR;
6094 mc.moved_charge += HPAGE_PMD_NR;
6102 if (pmd_trans_unstable(pmd))
6105 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6106 for (; addr != end; addr += PAGE_SIZE) {
6107 pte_t ptent = *(pte++);
6108 bool device = false;
6114 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6115 case MC_TARGET_DEVICE:
6118 case MC_TARGET_PAGE:
6121 * We can have a part of the split pmd here. Moving it
6122 * can be done but it would be too convoluted so simply
6123 * ignore such a partial THP and keep it in original
6124 * memcg. There should be somebody mapping the head.
6126 if (PageTransCompound(page))
6128 if (!device && isolate_lru_page(page))
6130 if (!mem_cgroup_move_account(page, false,
6133 /* we uncharge from mc.from later. */
6137 putback_lru_page(page);
6138 put: /* get_mctgt_type() gets the page */
6141 case MC_TARGET_SWAP:
6143 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6145 mem_cgroup_id_get_many(mc.to, 1);
6146 /* we fixup other refcnts and charges later. */
6154 pte_unmap_unlock(pte - 1, ptl);
6159 * We have consumed all precharges we got in can_attach().
6160 * We try charge one by one, but don't do any additional
6161 * charges to mc.to if we have failed in charge once in attach()
6164 ret = mem_cgroup_do_precharge(1);
6172 static const struct mm_walk_ops charge_walk_ops = {
6173 .pmd_entry = mem_cgroup_move_charge_pte_range,
6176 static void mem_cgroup_move_charge(void)
6178 lru_add_drain_all();
6180 * Signal lock_page_memcg() to take the memcg's move_lock
6181 * while we're moving its pages to another memcg. Then wait
6182 * for already started RCU-only updates to finish.
6184 atomic_inc(&mc.from->moving_account);
6187 if (unlikely(!mmap_read_trylock(mc.mm))) {
6189 * Someone who are holding the mmap_lock might be waiting in
6190 * waitq. So we cancel all extra charges, wake up all waiters,
6191 * and retry. Because we cancel precharges, we might not be able
6192 * to move enough charges, but moving charge is a best-effort
6193 * feature anyway, so it wouldn't be a big problem.
6195 __mem_cgroup_clear_mc();
6200 * When we have consumed all precharges and failed in doing
6201 * additional charge, the page walk just aborts.
6203 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6206 mmap_read_unlock(mc.mm);
6207 atomic_dec(&mc.from->moving_account);
6210 static void mem_cgroup_move_task(void)
6213 mem_cgroup_move_charge();
6214 mem_cgroup_clear_mc();
6217 #else /* !CONFIG_MMU */
6218 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6222 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6225 static void mem_cgroup_move_task(void)
6231 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6232 * to verify whether we're attached to the default hierarchy on each mount
6235 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6238 * use_hierarchy is forced on the default hierarchy. cgroup core
6239 * guarantees that @root doesn't have any children, so turning it
6240 * on for the root memcg is enough.
6242 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6243 root_mem_cgroup->use_hierarchy = true;
6245 root_mem_cgroup->use_hierarchy = false;
6248 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6250 if (value == PAGE_COUNTER_MAX)
6251 seq_puts(m, "max\n");
6253 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6258 static u64 memory_current_read(struct cgroup_subsys_state *css,
6261 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6263 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6266 static int memory_min_show(struct seq_file *m, void *v)
6268 return seq_puts_memcg_tunable(m,
6269 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6272 static ssize_t memory_min_write(struct kernfs_open_file *of,
6273 char *buf, size_t nbytes, loff_t off)
6275 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6279 buf = strstrip(buf);
6280 err = page_counter_memparse(buf, "max", &min);
6284 page_counter_set_min(&memcg->memory, min);
6289 static int memory_low_show(struct seq_file *m, void *v)
6291 return seq_puts_memcg_tunable(m,
6292 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6295 static ssize_t memory_low_write(struct kernfs_open_file *of,
6296 char *buf, size_t nbytes, loff_t off)
6298 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6302 buf = strstrip(buf);
6303 err = page_counter_memparse(buf, "max", &low);
6307 page_counter_set_low(&memcg->memory, low);
6312 static int memory_high_show(struct seq_file *m, void *v)
6314 return seq_puts_memcg_tunable(m,
6315 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6318 static ssize_t memory_high_write(struct kernfs_open_file *of,
6319 char *buf, size_t nbytes, loff_t off)
6321 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6322 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6323 bool drained = false;
6327 buf = strstrip(buf);
6328 err = page_counter_memparse(buf, "max", &high);
6332 page_counter_set_high(&memcg->memory, high);
6335 unsigned long nr_pages = page_counter_read(&memcg->memory);
6336 unsigned long reclaimed;
6338 if (nr_pages <= high)
6341 if (signal_pending(current))
6345 drain_all_stock(memcg);
6350 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6353 if (!reclaimed && !nr_retries--)
6357 memcg_wb_domain_size_changed(memcg);
6361 static int memory_max_show(struct seq_file *m, void *v)
6363 return seq_puts_memcg_tunable(m,
6364 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6367 static ssize_t memory_max_write(struct kernfs_open_file *of,
6368 char *buf, size_t nbytes, loff_t off)
6370 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6371 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6372 bool drained = false;
6376 buf = strstrip(buf);
6377 err = page_counter_memparse(buf, "max", &max);
6381 xchg(&memcg->memory.max, max);
6384 unsigned long nr_pages = page_counter_read(&memcg->memory);
6386 if (nr_pages <= max)
6389 if (signal_pending(current))
6393 drain_all_stock(memcg);
6399 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6405 memcg_memory_event(memcg, MEMCG_OOM);
6406 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6410 memcg_wb_domain_size_changed(memcg);
6414 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6416 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6417 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6418 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6419 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6420 seq_printf(m, "oom_kill %lu\n",
6421 atomic_long_read(&events[MEMCG_OOM_KILL]));
6424 static int memory_events_show(struct seq_file *m, void *v)
6426 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6428 __memory_events_show(m, memcg->memory_events);
6432 static int memory_events_local_show(struct seq_file *m, void *v)
6434 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6436 __memory_events_show(m, memcg->memory_events_local);
6440 static int memory_stat_show(struct seq_file *m, void *v)
6442 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6445 buf = memory_stat_format(memcg);
6454 static int memory_numa_stat_show(struct seq_file *m, void *v)
6457 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6459 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6462 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6465 seq_printf(m, "%s", memory_stats[i].name);
6466 for_each_node_state(nid, N_MEMORY) {
6468 struct lruvec *lruvec;
6470 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6471 size = lruvec_page_state(lruvec, memory_stats[i].idx);
6472 size *= memory_stats[i].ratio;
6473 seq_printf(m, " N%d=%llu", nid, size);
6482 static int memory_oom_group_show(struct seq_file *m, void *v)
6484 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6486 seq_printf(m, "%d\n", memcg->oom_group);
6491 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6492 char *buf, size_t nbytes, loff_t off)
6494 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6497 buf = strstrip(buf);
6501 ret = kstrtoint(buf, 0, &oom_group);
6505 if (oom_group != 0 && oom_group != 1)
6508 memcg->oom_group = oom_group;
6513 static struct cftype memory_files[] = {
6516 .flags = CFTYPE_NOT_ON_ROOT,
6517 .read_u64 = memory_current_read,
6521 .flags = CFTYPE_NOT_ON_ROOT,
6522 .seq_show = memory_min_show,
6523 .write = memory_min_write,
6527 .flags = CFTYPE_NOT_ON_ROOT,
6528 .seq_show = memory_low_show,
6529 .write = memory_low_write,
6533 .flags = CFTYPE_NOT_ON_ROOT,
6534 .seq_show = memory_high_show,
6535 .write = memory_high_write,
6539 .flags = CFTYPE_NOT_ON_ROOT,
6540 .seq_show = memory_max_show,
6541 .write = memory_max_write,
6545 .flags = CFTYPE_NOT_ON_ROOT,
6546 .file_offset = offsetof(struct mem_cgroup, events_file),
6547 .seq_show = memory_events_show,
6550 .name = "events.local",
6551 .flags = CFTYPE_NOT_ON_ROOT,
6552 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6553 .seq_show = memory_events_local_show,
6557 .seq_show = memory_stat_show,
6561 .name = "numa_stat",
6562 .seq_show = memory_numa_stat_show,
6566 .name = "oom.group",
6567 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6568 .seq_show = memory_oom_group_show,
6569 .write = memory_oom_group_write,
6574 struct cgroup_subsys memory_cgrp_subsys = {
6575 .css_alloc = mem_cgroup_css_alloc,
6576 .css_online = mem_cgroup_css_online,
6577 .css_offline = mem_cgroup_css_offline,
6578 .css_released = mem_cgroup_css_released,
6579 .css_free = mem_cgroup_css_free,
6580 .css_reset = mem_cgroup_css_reset,
6581 .can_attach = mem_cgroup_can_attach,
6582 .cancel_attach = mem_cgroup_cancel_attach,
6583 .post_attach = mem_cgroup_move_task,
6584 .bind = mem_cgroup_bind,
6585 .dfl_cftypes = memory_files,
6586 .legacy_cftypes = mem_cgroup_legacy_files,
6591 * This function calculates an individual cgroup's effective
6592 * protection which is derived from its own memory.min/low, its
6593 * parent's and siblings' settings, as well as the actual memory
6594 * distribution in the tree.
6596 * The following rules apply to the effective protection values:
6598 * 1. At the first level of reclaim, effective protection is equal to
6599 * the declared protection in memory.min and memory.low.
6601 * 2. To enable safe delegation of the protection configuration, at
6602 * subsequent levels the effective protection is capped to the
6603 * parent's effective protection.
6605 * 3. To make complex and dynamic subtrees easier to configure, the
6606 * user is allowed to overcommit the declared protection at a given
6607 * level. If that is the case, the parent's effective protection is
6608 * distributed to the children in proportion to how much protection
6609 * they have declared and how much of it they are utilizing.
6611 * This makes distribution proportional, but also work-conserving:
6612 * if one cgroup claims much more protection than it uses memory,
6613 * the unused remainder is available to its siblings.
6615 * 4. Conversely, when the declared protection is undercommitted at a
6616 * given level, the distribution of the larger parental protection
6617 * budget is NOT proportional. A cgroup's protection from a sibling
6618 * is capped to its own memory.min/low setting.
6620 * 5. However, to allow protecting recursive subtrees from each other
6621 * without having to declare each individual cgroup's fixed share
6622 * of the ancestor's claim to protection, any unutilized -
6623 * "floating" - protection from up the tree is distributed in
6624 * proportion to each cgroup's *usage*. This makes the protection
6625 * neutral wrt sibling cgroups and lets them compete freely over
6626 * the shared parental protection budget, but it protects the
6627 * subtree as a whole from neighboring subtrees.
6629 * Note that 4. and 5. are not in conflict: 4. is about protecting
6630 * against immediate siblings whereas 5. is about protecting against
6631 * neighboring subtrees.
6633 static unsigned long effective_protection(unsigned long usage,
6634 unsigned long parent_usage,
6635 unsigned long setting,
6636 unsigned long parent_effective,
6637 unsigned long siblings_protected)
6639 unsigned long protected;
6642 protected = min(usage, setting);
6644 * If all cgroups at this level combined claim and use more
6645 * protection then what the parent affords them, distribute
6646 * shares in proportion to utilization.
6648 * We are using actual utilization rather than the statically
6649 * claimed protection in order to be work-conserving: claimed
6650 * but unused protection is available to siblings that would
6651 * otherwise get a smaller chunk than what they claimed.
6653 if (siblings_protected > parent_effective)
6654 return protected * parent_effective / siblings_protected;
6657 * Ok, utilized protection of all children is within what the
6658 * parent affords them, so we know whatever this child claims
6659 * and utilizes is effectively protected.
6661 * If there is unprotected usage beyond this value, reclaim
6662 * will apply pressure in proportion to that amount.
6664 * If there is unutilized protection, the cgroup will be fully
6665 * shielded from reclaim, but we do return a smaller value for
6666 * protection than what the group could enjoy in theory. This
6667 * is okay. With the overcommit distribution above, effective
6668 * protection is always dependent on how memory is actually
6669 * consumed among the siblings anyway.
6674 * If the children aren't claiming (all of) the protection
6675 * afforded to them by the parent, distribute the remainder in
6676 * proportion to the (unprotected) memory of each cgroup. That
6677 * way, cgroups that aren't explicitly prioritized wrt each
6678 * other compete freely over the allowance, but they are
6679 * collectively protected from neighboring trees.
6681 * We're using unprotected memory for the weight so that if
6682 * some cgroups DO claim explicit protection, we don't protect
6683 * the same bytes twice.
6685 * Check both usage and parent_usage against the respective
6686 * protected values. One should imply the other, but they
6687 * aren't read atomically - make sure the division is sane.
6689 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6691 if (parent_effective > siblings_protected &&
6692 parent_usage > siblings_protected &&
6693 usage > protected) {
6694 unsigned long unclaimed;
6696 unclaimed = parent_effective - siblings_protected;
6697 unclaimed *= usage - protected;
6698 unclaimed /= parent_usage - siblings_protected;
6707 * mem_cgroup_protected - check if memory consumption is in the normal range
6708 * @root: the top ancestor of the sub-tree being checked
6709 * @memcg: the memory cgroup to check
6711 * WARNING: This function is not stateless! It can only be used as part
6712 * of a top-down tree iteration, not for isolated queries.
6714 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6715 struct mem_cgroup *memcg)
6717 unsigned long usage, parent_usage;
6718 struct mem_cgroup *parent;
6720 if (mem_cgroup_disabled())
6724 root = root_mem_cgroup;
6727 * Effective values of the reclaim targets are ignored so they
6728 * can be stale. Have a look at mem_cgroup_protection for more
6730 * TODO: calculation should be more robust so that we do not need
6731 * that special casing.
6736 usage = page_counter_read(&memcg->memory);
6740 parent = parent_mem_cgroup(memcg);
6741 /* No parent means a non-hierarchical mode on v1 memcg */
6745 if (parent == root) {
6746 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6747 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6751 parent_usage = page_counter_read(&parent->memory);
6753 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6754 READ_ONCE(memcg->memory.min),
6755 READ_ONCE(parent->memory.emin),
6756 atomic_long_read(&parent->memory.children_min_usage)));
6758 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6759 READ_ONCE(memcg->memory.low),
6760 READ_ONCE(parent->memory.elow),
6761 atomic_long_read(&parent->memory.children_low_usage)));
6765 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6766 * @page: page to charge
6767 * @mm: mm context of the victim
6768 * @gfp_mask: reclaim mode
6770 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6771 * pages according to @gfp_mask if necessary.
6773 * Returns 0 on success. Otherwise, an error code is returned.
6775 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6777 unsigned int nr_pages = thp_nr_pages(page);
6778 struct mem_cgroup *memcg = NULL;
6781 if (mem_cgroup_disabled())
6784 if (PageSwapCache(page)) {
6785 swp_entry_t ent = { .val = page_private(page), };
6789 * Every swap fault against a single page tries to charge the
6790 * page, bail as early as possible. shmem_unuse() encounters
6791 * already charged pages, too. page->mem_cgroup is protected
6792 * by the page lock, which serializes swap cache removal, which
6793 * in turn serializes uncharging.
6795 VM_BUG_ON_PAGE(!PageLocked(page), page);
6796 if (compound_head(page)->mem_cgroup)
6799 id = lookup_swap_cgroup_id(ent);
6801 memcg = mem_cgroup_from_id(id);
6802 if (memcg && !css_tryget_online(&memcg->css))
6808 memcg = get_mem_cgroup_from_mm(mm);
6810 ret = try_charge(memcg, gfp_mask, nr_pages);
6814 css_get(&memcg->css);
6815 commit_charge(page, memcg);
6817 local_irq_disable();
6818 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6819 memcg_check_events(memcg, page);
6823 * Cgroup1's unified memory+swap counter has been charged with the
6824 * new swapcache page, finish the transfer by uncharging the swap
6825 * slot. The swap slot would also get uncharged when it dies, but
6826 * it can stick around indefinitely and we'd count the page twice
6829 * Cgroup2 has separate resource counters for memory and swap,
6830 * so this is a non-issue here. Memory and swap charge lifetimes
6831 * correspond 1:1 to page and swap slot lifetimes: we charge the
6832 * page to memory here, and uncharge swap when the slot is freed.
6834 if (do_memsw_account() && PageSwapCache(page)) {
6835 swp_entry_t entry = { .val = page_private(page) };
6837 * The swap entry might not get freed for a long time,
6838 * let's not wait for it. The page already received a
6839 * memory+swap charge, drop the swap entry duplicate.
6841 mem_cgroup_uncharge_swap(entry, nr_pages);
6845 css_put(&memcg->css);
6850 struct uncharge_gather {
6851 struct mem_cgroup *memcg;
6852 unsigned long nr_pages;
6853 unsigned long pgpgout;
6854 unsigned long nr_kmem;
6855 struct page *dummy_page;
6858 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6860 memset(ug, 0, sizeof(*ug));
6863 static void uncharge_batch(const struct uncharge_gather *ug)
6865 unsigned long flags;
6867 if (!mem_cgroup_is_root(ug->memcg)) {
6868 page_counter_uncharge(&ug->memcg->memory, ug->nr_pages);
6869 if (do_memsw_account())
6870 page_counter_uncharge(&ug->memcg->memsw, ug->nr_pages);
6871 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6872 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6873 memcg_oom_recover(ug->memcg);
6876 local_irq_save(flags);
6877 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6878 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_pages);
6879 memcg_check_events(ug->memcg, ug->dummy_page);
6880 local_irq_restore(flags);
6882 /* drop reference from uncharge_page */
6883 css_put(&ug->memcg->css);
6886 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6888 unsigned long nr_pages;
6890 VM_BUG_ON_PAGE(PageLRU(page), page);
6892 if (!page->mem_cgroup)
6896 * Nobody should be changing or seriously looking at
6897 * page->mem_cgroup at this point, we have fully
6898 * exclusive access to the page.
6901 if (ug->memcg != page->mem_cgroup) {
6904 uncharge_gather_clear(ug);
6906 ug->memcg = page->mem_cgroup;
6908 /* pairs with css_put in uncharge_batch */
6909 css_get(&ug->memcg->css);
6912 nr_pages = compound_nr(page);
6913 ug->nr_pages += nr_pages;
6915 if (!PageKmemcg(page)) {
6918 ug->nr_kmem += nr_pages;
6919 __ClearPageKmemcg(page);
6922 ug->dummy_page = page;
6923 page->mem_cgroup = NULL;
6924 css_put(&ug->memcg->css);
6927 static void uncharge_list(struct list_head *page_list)
6929 struct uncharge_gather ug;
6930 struct list_head *next;
6932 uncharge_gather_clear(&ug);
6935 * Note that the list can be a single page->lru; hence the
6936 * do-while loop instead of a simple list_for_each_entry().
6938 next = page_list->next;
6942 page = list_entry(next, struct page, lru);
6943 next = page->lru.next;
6945 uncharge_page(page, &ug);
6946 } while (next != page_list);
6949 uncharge_batch(&ug);
6953 * mem_cgroup_uncharge - uncharge a page
6954 * @page: page to uncharge
6956 * Uncharge a page previously charged with mem_cgroup_charge().
6958 void mem_cgroup_uncharge(struct page *page)
6960 struct uncharge_gather ug;
6962 if (mem_cgroup_disabled())
6965 /* Don't touch page->lru of any random page, pre-check: */
6966 if (!page->mem_cgroup)
6969 uncharge_gather_clear(&ug);
6970 uncharge_page(page, &ug);
6971 uncharge_batch(&ug);
6975 * mem_cgroup_uncharge_list - uncharge a list of page
6976 * @page_list: list of pages to uncharge
6978 * Uncharge a list of pages previously charged with
6979 * mem_cgroup_charge().
6981 void mem_cgroup_uncharge_list(struct list_head *page_list)
6983 if (mem_cgroup_disabled())
6986 if (!list_empty(page_list))
6987 uncharge_list(page_list);
6991 * mem_cgroup_migrate - charge a page's replacement
6992 * @oldpage: currently circulating page
6993 * @newpage: replacement page
6995 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6996 * be uncharged upon free.
6998 * Both pages must be locked, @newpage->mapping must be set up.
7000 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
7002 struct mem_cgroup *memcg;
7003 unsigned int nr_pages;
7004 unsigned long flags;
7006 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
7007 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
7008 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
7009 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
7012 if (mem_cgroup_disabled())
7015 /* Page cache replacement: new page already charged? */
7016 if (newpage->mem_cgroup)
7019 /* Swapcache readahead pages can get replaced before being charged */
7020 memcg = oldpage->mem_cgroup;
7024 /* Force-charge the new page. The old one will be freed soon */
7025 nr_pages = thp_nr_pages(newpage);
7027 page_counter_charge(&memcg->memory, nr_pages);
7028 if (do_memsw_account())
7029 page_counter_charge(&memcg->memsw, nr_pages);
7031 css_get(&memcg->css);
7032 commit_charge(newpage, memcg);
7034 local_irq_save(flags);
7035 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
7036 memcg_check_events(memcg, newpage);
7037 local_irq_restore(flags);
7040 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
7041 EXPORT_SYMBOL(memcg_sockets_enabled_key);
7043 void mem_cgroup_sk_alloc(struct sock *sk)
7045 struct mem_cgroup *memcg;
7047 if (!mem_cgroup_sockets_enabled)
7050 /* Do not associate the sock with unrelated interrupted task's memcg. */
7055 memcg = mem_cgroup_from_task(current);
7056 if (memcg == root_mem_cgroup)
7058 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
7060 if (css_tryget(&memcg->css))
7061 sk->sk_memcg = memcg;
7066 void mem_cgroup_sk_free(struct sock *sk)
7069 css_put(&sk->sk_memcg->css);
7073 * mem_cgroup_charge_skmem - charge socket memory
7074 * @memcg: memcg to charge
7075 * @nr_pages: number of pages to charge
7077 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7078 * @memcg's configured limit, %false if the charge had to be forced.
7080 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7082 gfp_t gfp_mask = GFP_KERNEL;
7084 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7085 struct page_counter *fail;
7087 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7088 memcg->tcpmem_pressure = 0;
7091 page_counter_charge(&memcg->tcpmem, nr_pages);
7092 memcg->tcpmem_pressure = 1;
7096 /* Don't block in the packet receive path */
7098 gfp_mask = GFP_NOWAIT;
7100 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7102 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
7105 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
7110 * mem_cgroup_uncharge_skmem - uncharge socket memory
7111 * @memcg: memcg to uncharge
7112 * @nr_pages: number of pages to uncharge
7114 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7116 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7117 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7121 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7123 refill_stock(memcg, nr_pages);
7126 static int __init cgroup_memory(char *s)
7130 while ((token = strsep(&s, ",")) != NULL) {
7133 if (!strcmp(token, "nosocket"))
7134 cgroup_memory_nosocket = true;
7135 if (!strcmp(token, "nokmem"))
7136 cgroup_memory_nokmem = true;
7140 __setup("cgroup.memory=", cgroup_memory);
7143 * subsys_initcall() for memory controller.
7145 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7146 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7147 * basically everything that doesn't depend on a specific mem_cgroup structure
7148 * should be initialized from here.
7150 static int __init mem_cgroup_init(void)
7154 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7155 memcg_hotplug_cpu_dead);
7157 for_each_possible_cpu(cpu)
7158 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7161 for_each_node(node) {
7162 struct mem_cgroup_tree_per_node *rtpn;
7164 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7165 node_online(node) ? node : NUMA_NO_NODE);
7167 rtpn->rb_root = RB_ROOT;
7168 rtpn->rb_rightmost = NULL;
7169 spin_lock_init(&rtpn->lock);
7170 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7175 subsys_initcall(mem_cgroup_init);
7177 #ifdef CONFIG_MEMCG_SWAP
7178 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7180 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7182 * The root cgroup cannot be destroyed, so it's refcount must
7185 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7189 memcg = parent_mem_cgroup(memcg);
7191 memcg = root_mem_cgroup;
7197 * mem_cgroup_swapout - transfer a memsw charge to swap
7198 * @page: page whose memsw charge to transfer
7199 * @entry: swap entry to move the charge to
7201 * Transfer the memsw charge of @page to @entry.
7203 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7205 struct mem_cgroup *memcg, *swap_memcg;
7206 unsigned int nr_entries;
7207 unsigned short oldid;
7209 VM_BUG_ON_PAGE(PageLRU(page), page);
7210 VM_BUG_ON_PAGE(page_count(page), page);
7212 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7215 memcg = page->mem_cgroup;
7217 /* Readahead page, never charged */
7222 * In case the memcg owning these pages has been offlined and doesn't
7223 * have an ID allocated to it anymore, charge the closest online
7224 * ancestor for the swap instead and transfer the memory+swap charge.
7226 swap_memcg = mem_cgroup_id_get_online(memcg);
7227 nr_entries = thp_nr_pages(page);
7228 /* Get references for the tail pages, too */
7230 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7231 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7233 VM_BUG_ON_PAGE(oldid, page);
7234 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7236 page->mem_cgroup = NULL;
7238 if (!mem_cgroup_is_root(memcg))
7239 page_counter_uncharge(&memcg->memory, nr_entries);
7241 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7242 if (!mem_cgroup_is_root(swap_memcg))
7243 page_counter_charge(&swap_memcg->memsw, nr_entries);
7244 page_counter_uncharge(&memcg->memsw, nr_entries);
7248 * Interrupts should be disabled here because the caller holds the
7249 * i_pages lock which is taken with interrupts-off. It is
7250 * important here to have the interrupts disabled because it is the
7251 * only synchronisation we have for updating the per-CPU variables.
7253 VM_BUG_ON(!irqs_disabled());
7254 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7255 memcg_check_events(memcg, page);
7257 css_put(&memcg->css);
7261 * mem_cgroup_try_charge_swap - try charging swap space for a page
7262 * @page: page being added to swap
7263 * @entry: swap entry to charge
7265 * Try to charge @page's memcg for the swap space at @entry.
7267 * Returns 0 on success, -ENOMEM on failure.
7269 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7271 unsigned int nr_pages = thp_nr_pages(page);
7272 struct page_counter *counter;
7273 struct mem_cgroup *memcg;
7274 unsigned short oldid;
7276 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7279 memcg = page->mem_cgroup;
7281 /* Readahead page, never charged */
7286 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7290 memcg = mem_cgroup_id_get_online(memcg);
7292 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7293 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7294 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7295 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7296 mem_cgroup_id_put(memcg);
7300 /* Get references for the tail pages, too */
7302 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7303 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7304 VM_BUG_ON_PAGE(oldid, page);
7305 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7311 * mem_cgroup_uncharge_swap - uncharge swap space
7312 * @entry: swap entry to uncharge
7313 * @nr_pages: the amount of swap space to uncharge
7315 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7317 struct mem_cgroup *memcg;
7320 id = swap_cgroup_record(entry, 0, nr_pages);
7322 memcg = mem_cgroup_from_id(id);
7324 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7325 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7326 page_counter_uncharge(&memcg->swap, nr_pages);
7328 page_counter_uncharge(&memcg->memsw, nr_pages);
7330 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7331 mem_cgroup_id_put_many(memcg, nr_pages);
7336 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7338 long nr_swap_pages = get_nr_swap_pages();
7340 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7341 return nr_swap_pages;
7342 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7343 nr_swap_pages = min_t(long, nr_swap_pages,
7344 READ_ONCE(memcg->swap.max) -
7345 page_counter_read(&memcg->swap));
7346 return nr_swap_pages;
7349 bool mem_cgroup_swap_full(struct page *page)
7351 struct mem_cgroup *memcg;
7353 VM_BUG_ON_PAGE(!PageLocked(page), page);
7357 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7360 memcg = page->mem_cgroup;
7364 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7365 unsigned long usage = page_counter_read(&memcg->swap);
7367 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7368 usage * 2 >= READ_ONCE(memcg->swap.max))
7375 static int __init setup_swap_account(char *s)
7377 if (!strcmp(s, "1"))
7378 cgroup_memory_noswap = 0;
7379 else if (!strcmp(s, "0"))
7380 cgroup_memory_noswap = 1;
7383 __setup("swapaccount=", setup_swap_account);
7385 static u64 swap_current_read(struct cgroup_subsys_state *css,
7388 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7390 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7393 static int swap_high_show(struct seq_file *m, void *v)
7395 return seq_puts_memcg_tunable(m,
7396 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7399 static ssize_t swap_high_write(struct kernfs_open_file *of,
7400 char *buf, size_t nbytes, loff_t off)
7402 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7406 buf = strstrip(buf);
7407 err = page_counter_memparse(buf, "max", &high);
7411 page_counter_set_high(&memcg->swap, high);
7416 static int swap_max_show(struct seq_file *m, void *v)
7418 return seq_puts_memcg_tunable(m,
7419 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7422 static ssize_t swap_max_write(struct kernfs_open_file *of,
7423 char *buf, size_t nbytes, loff_t off)
7425 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7429 buf = strstrip(buf);
7430 err = page_counter_memparse(buf, "max", &max);
7434 xchg(&memcg->swap.max, max);
7439 static int swap_events_show(struct seq_file *m, void *v)
7441 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7443 seq_printf(m, "high %lu\n",
7444 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7445 seq_printf(m, "max %lu\n",
7446 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7447 seq_printf(m, "fail %lu\n",
7448 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7453 static struct cftype swap_files[] = {
7455 .name = "swap.current",
7456 .flags = CFTYPE_NOT_ON_ROOT,
7457 .read_u64 = swap_current_read,
7460 .name = "swap.high",
7461 .flags = CFTYPE_NOT_ON_ROOT,
7462 .seq_show = swap_high_show,
7463 .write = swap_high_write,
7467 .flags = CFTYPE_NOT_ON_ROOT,
7468 .seq_show = swap_max_show,
7469 .write = swap_max_write,
7472 .name = "swap.events",
7473 .flags = CFTYPE_NOT_ON_ROOT,
7474 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7475 .seq_show = swap_events_show,
7480 static struct cftype memsw_files[] = {
7482 .name = "memsw.usage_in_bytes",
7483 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7484 .read_u64 = mem_cgroup_read_u64,
7487 .name = "memsw.max_usage_in_bytes",
7488 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7489 .write = mem_cgroup_reset,
7490 .read_u64 = mem_cgroup_read_u64,
7493 .name = "memsw.limit_in_bytes",
7494 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7495 .write = mem_cgroup_write,
7496 .read_u64 = mem_cgroup_read_u64,
7499 .name = "memsw.failcnt",
7500 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7501 .write = mem_cgroup_reset,
7502 .read_u64 = mem_cgroup_read_u64,
7504 { }, /* terminate */
7508 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7509 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7510 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7511 * boot parameter. This may result in premature OOPS inside
7512 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7514 static int __init mem_cgroup_swap_init(void)
7516 /* No memory control -> no swap control */
7517 if (mem_cgroup_disabled())
7518 cgroup_memory_noswap = true;
7520 if (cgroup_memory_noswap)
7523 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7524 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7528 core_initcall(mem_cgroup_swap_init);
7530 #endif /* CONFIG_MEMCG_SWAP */