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_memcg(page);
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_memcg_check(page);
565 while (memcg && !(memcg->css.flags & CSS_ONLINE))
566 memcg = parent_mem_cgroup(memcg);
568 ino = cgroup_ino(memcg->css.cgroup);
573 static struct mem_cgroup_per_node *
574 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
576 int nid = page_to_nid(page);
578 return memcg->nodeinfo[nid];
581 static struct mem_cgroup_tree_per_node *
582 soft_limit_tree_node(int nid)
584 return soft_limit_tree.rb_tree_per_node[nid];
587 static struct mem_cgroup_tree_per_node *
588 soft_limit_tree_from_page(struct page *page)
590 int nid = page_to_nid(page);
592 return soft_limit_tree.rb_tree_per_node[nid];
595 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
596 struct mem_cgroup_tree_per_node *mctz,
597 unsigned long new_usage_in_excess)
599 struct rb_node **p = &mctz->rb_root.rb_node;
600 struct rb_node *parent = NULL;
601 struct mem_cgroup_per_node *mz_node;
602 bool rightmost = true;
607 mz->usage_in_excess = new_usage_in_excess;
608 if (!mz->usage_in_excess)
612 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
614 if (mz->usage_in_excess < mz_node->usage_in_excess) {
620 * We can't avoid mem cgroups that are over their soft
621 * limit by the same amount
623 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
628 mctz->rb_rightmost = &mz->tree_node;
630 rb_link_node(&mz->tree_node, parent, p);
631 rb_insert_color(&mz->tree_node, &mctz->rb_root);
635 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
636 struct mem_cgroup_tree_per_node *mctz)
641 if (&mz->tree_node == mctz->rb_rightmost)
642 mctz->rb_rightmost = rb_prev(&mz->tree_node);
644 rb_erase(&mz->tree_node, &mctz->rb_root);
648 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
649 struct mem_cgroup_tree_per_node *mctz)
653 spin_lock_irqsave(&mctz->lock, flags);
654 __mem_cgroup_remove_exceeded(mz, mctz);
655 spin_unlock_irqrestore(&mctz->lock, flags);
658 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
660 unsigned long nr_pages = page_counter_read(&memcg->memory);
661 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
662 unsigned long excess = 0;
664 if (nr_pages > soft_limit)
665 excess = nr_pages - soft_limit;
670 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
672 unsigned long excess;
673 struct mem_cgroup_per_node *mz;
674 struct mem_cgroup_tree_per_node *mctz;
676 mctz = soft_limit_tree_from_page(page);
680 * Necessary to update all ancestors when hierarchy is used.
681 * because their event counter is not touched.
683 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
684 mz = mem_cgroup_page_nodeinfo(memcg, page);
685 excess = soft_limit_excess(memcg);
687 * We have to update the tree if mz is on RB-tree or
688 * mem is over its softlimit.
690 if (excess || mz->on_tree) {
693 spin_lock_irqsave(&mctz->lock, flags);
694 /* if on-tree, remove it */
696 __mem_cgroup_remove_exceeded(mz, mctz);
698 * Insert again. mz->usage_in_excess will be updated.
699 * If excess is 0, no tree ops.
701 __mem_cgroup_insert_exceeded(mz, mctz, excess);
702 spin_unlock_irqrestore(&mctz->lock, flags);
707 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
709 struct mem_cgroup_tree_per_node *mctz;
710 struct mem_cgroup_per_node *mz;
714 mz = mem_cgroup_nodeinfo(memcg, nid);
715 mctz = soft_limit_tree_node(nid);
717 mem_cgroup_remove_exceeded(mz, mctz);
721 static struct mem_cgroup_per_node *
722 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
724 struct mem_cgroup_per_node *mz;
728 if (!mctz->rb_rightmost)
729 goto done; /* Nothing to reclaim from */
731 mz = rb_entry(mctz->rb_rightmost,
732 struct mem_cgroup_per_node, tree_node);
734 * Remove the node now but someone else can add it back,
735 * we will to add it back at the end of reclaim to its correct
736 * position in the tree.
738 __mem_cgroup_remove_exceeded(mz, mctz);
739 if (!soft_limit_excess(mz->memcg) ||
740 !css_tryget(&mz->memcg->css))
746 static struct mem_cgroup_per_node *
747 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
749 struct mem_cgroup_per_node *mz;
751 spin_lock_irq(&mctz->lock);
752 mz = __mem_cgroup_largest_soft_limit_node(mctz);
753 spin_unlock_irq(&mctz->lock);
758 * __mod_memcg_state - update cgroup memory statistics
759 * @memcg: the memory cgroup
760 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
761 * @val: delta to add to the counter, can be negative
763 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
765 long x, threshold = MEMCG_CHARGE_BATCH;
767 if (mem_cgroup_disabled())
770 if (memcg_stat_item_in_bytes(idx))
771 threshold <<= PAGE_SHIFT;
773 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
774 if (unlikely(abs(x) > threshold)) {
775 struct mem_cgroup *mi;
778 * Batch local counters to keep them in sync with
779 * the hierarchical ones.
781 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
782 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
783 atomic_long_add(x, &mi->vmstats[idx]);
786 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
789 static struct mem_cgroup_per_node *
790 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
792 struct mem_cgroup *parent;
794 parent = parent_mem_cgroup(pn->memcg);
797 return mem_cgroup_nodeinfo(parent, nid);
800 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
803 struct mem_cgroup_per_node *pn;
804 struct mem_cgroup *memcg;
805 long x, threshold = MEMCG_CHARGE_BATCH;
807 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
811 __mod_memcg_state(memcg, idx, val);
814 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
816 if (vmstat_item_in_bytes(idx))
817 threshold <<= PAGE_SHIFT;
819 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
820 if (unlikely(abs(x) > threshold)) {
821 pg_data_t *pgdat = lruvec_pgdat(lruvec);
822 struct mem_cgroup_per_node *pi;
824 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
825 atomic_long_add(x, &pi->lruvec_stat[idx]);
828 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
832 * __mod_lruvec_state - update lruvec memory statistics
833 * @lruvec: the lruvec
834 * @idx: the stat item
835 * @val: delta to add to the counter, can be negative
837 * The lruvec is the intersection of the NUMA node and a cgroup. This
838 * function updates the all three counters that are affected by a
839 * change of state at this level: per-node, per-cgroup, per-lruvec.
841 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
845 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
847 /* Update memcg and lruvec */
848 if (!mem_cgroup_disabled())
849 __mod_memcg_lruvec_state(lruvec, idx, val);
852 void __mod_lruvec_slab_state(void *p, enum node_stat_item idx, int val)
854 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
855 struct mem_cgroup *memcg;
856 struct lruvec *lruvec;
859 memcg = mem_cgroup_from_obj(p);
861 /* Untracked pages have no memcg, no lruvec. Update only the node */
862 if (!memcg || memcg == root_mem_cgroup) {
863 __mod_node_page_state(pgdat, idx, val);
865 lruvec = mem_cgroup_lruvec(memcg, pgdat);
866 __mod_lruvec_state(lruvec, idx, val);
871 void mod_memcg_obj_state(void *p, int idx, int val)
873 struct mem_cgroup *memcg;
876 memcg = mem_cgroup_from_obj(p);
878 mod_memcg_state(memcg, idx, val);
883 * __count_memcg_events - account VM events in a cgroup
884 * @memcg: the memory cgroup
885 * @idx: the event item
886 * @count: the number of events that occured
888 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
893 if (mem_cgroup_disabled())
896 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
897 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
898 struct mem_cgroup *mi;
901 * Batch local counters to keep them in sync with
902 * the hierarchical ones.
904 __this_cpu_add(memcg->vmstats_local->events[idx], x);
905 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
906 atomic_long_add(x, &mi->vmevents[idx]);
909 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
912 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
914 return atomic_long_read(&memcg->vmevents[event]);
917 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
922 for_each_possible_cpu(cpu)
923 x += per_cpu(memcg->vmstats_local->events[event], cpu);
927 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
931 /* pagein of a big page is an event. So, ignore page size */
933 __count_memcg_events(memcg, PGPGIN, 1);
935 __count_memcg_events(memcg, PGPGOUT, 1);
936 nr_pages = -nr_pages; /* for event */
939 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
942 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
943 enum mem_cgroup_events_target target)
945 unsigned long val, next;
947 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
948 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
949 /* from time_after() in jiffies.h */
950 if ((long)(next - val) < 0) {
952 case MEM_CGROUP_TARGET_THRESH:
953 next = val + THRESHOLDS_EVENTS_TARGET;
955 case MEM_CGROUP_TARGET_SOFTLIMIT:
956 next = val + SOFTLIMIT_EVENTS_TARGET;
961 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
968 * Check events in order.
971 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
973 /* threshold event is triggered in finer grain than soft limit */
974 if (unlikely(mem_cgroup_event_ratelimit(memcg,
975 MEM_CGROUP_TARGET_THRESH))) {
978 do_softlimit = mem_cgroup_event_ratelimit(memcg,
979 MEM_CGROUP_TARGET_SOFTLIMIT);
980 mem_cgroup_threshold(memcg);
981 if (unlikely(do_softlimit))
982 mem_cgroup_update_tree(memcg, page);
986 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
989 * mm_update_next_owner() may clear mm->owner to NULL
990 * if it races with swapoff, page migration, etc.
991 * So this can be called with p == NULL.
996 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
998 EXPORT_SYMBOL(mem_cgroup_from_task);
1001 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1002 * @mm: mm from which memcg should be extracted. It can be NULL.
1004 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
1005 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
1008 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1010 struct mem_cgroup *memcg;
1012 if (mem_cgroup_disabled())
1018 * Page cache insertions can happen withou an
1019 * actual mm context, e.g. during disk probing
1020 * on boot, loopback IO, acct() writes etc.
1023 memcg = root_mem_cgroup;
1025 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1026 if (unlikely(!memcg))
1027 memcg = root_mem_cgroup;
1029 } while (!css_tryget(&memcg->css));
1033 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1036 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
1037 * @page: page from which memcg should be extracted.
1039 * Obtain a reference on page->memcg and returns it if successful. Otherwise
1040 * root_mem_cgroup is returned.
1042 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
1044 struct mem_cgroup *memcg = page_memcg(page);
1046 if (mem_cgroup_disabled())
1050 /* Page should not get uncharged and freed memcg under us. */
1051 if (!memcg || WARN_ON_ONCE(!css_tryget(&memcg->css)))
1052 memcg = root_mem_cgroup;
1056 EXPORT_SYMBOL(get_mem_cgroup_from_page);
1058 static __always_inline struct mem_cgroup *active_memcg(void)
1061 return this_cpu_read(int_active_memcg);
1063 return current->active_memcg;
1066 static __always_inline struct mem_cgroup *get_active_memcg(void)
1068 struct mem_cgroup *memcg;
1071 memcg = active_memcg();
1073 /* current->active_memcg must hold a ref. */
1074 if (WARN_ON_ONCE(!css_tryget(&memcg->css)))
1075 memcg = root_mem_cgroup;
1077 memcg = current->active_memcg;
1084 static __always_inline bool memcg_kmem_bypass(void)
1086 /* Allow remote memcg charging from any context. */
1087 if (unlikely(active_memcg()))
1090 /* Memcg to charge can't be determined. */
1091 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
1098 * If active memcg is set, do not fallback to current->mm->memcg.
1100 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
1102 if (memcg_kmem_bypass())
1105 if (unlikely(active_memcg()))
1106 return get_active_memcg();
1108 return get_mem_cgroup_from_mm(current->mm);
1112 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1113 * @root: hierarchy root
1114 * @prev: previously returned memcg, NULL on first invocation
1115 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1117 * Returns references to children of the hierarchy below @root, or
1118 * @root itself, or %NULL after a full round-trip.
1120 * Caller must pass the return value in @prev on subsequent
1121 * invocations for reference counting, or use mem_cgroup_iter_break()
1122 * to cancel a hierarchy walk before the round-trip is complete.
1124 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1125 * in the hierarchy among all concurrent reclaimers operating on the
1128 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1129 struct mem_cgroup *prev,
1130 struct mem_cgroup_reclaim_cookie *reclaim)
1132 struct mem_cgroup_reclaim_iter *iter;
1133 struct cgroup_subsys_state *css = NULL;
1134 struct mem_cgroup *memcg = NULL;
1135 struct mem_cgroup *pos = NULL;
1137 if (mem_cgroup_disabled())
1141 root = root_mem_cgroup;
1143 if (prev && !reclaim)
1146 if (!root->use_hierarchy && root != root_mem_cgroup) {
1155 struct mem_cgroup_per_node *mz;
1157 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1160 if (prev && reclaim->generation != iter->generation)
1164 pos = READ_ONCE(iter->position);
1165 if (!pos || css_tryget(&pos->css))
1168 * css reference reached zero, so iter->position will
1169 * be cleared by ->css_released. However, we should not
1170 * rely on this happening soon, because ->css_released
1171 * is called from a work queue, and by busy-waiting we
1172 * might block it. So we clear iter->position right
1175 (void)cmpxchg(&iter->position, pos, NULL);
1183 css = css_next_descendant_pre(css, &root->css);
1186 * Reclaimers share the hierarchy walk, and a
1187 * new one might jump in right at the end of
1188 * the hierarchy - make sure they see at least
1189 * one group and restart from the beginning.
1197 * Verify the css and acquire a reference. The root
1198 * is provided by the caller, so we know it's alive
1199 * and kicking, and don't take an extra reference.
1201 memcg = mem_cgroup_from_css(css);
1203 if (css == &root->css)
1206 if (css_tryget(css))
1214 * The position could have already been updated by a competing
1215 * thread, so check that the value hasn't changed since we read
1216 * it to avoid reclaiming from the same cgroup twice.
1218 (void)cmpxchg(&iter->position, pos, memcg);
1226 reclaim->generation = iter->generation;
1232 if (prev && prev != root)
1233 css_put(&prev->css);
1239 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1240 * @root: hierarchy root
1241 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1243 void mem_cgroup_iter_break(struct mem_cgroup *root,
1244 struct mem_cgroup *prev)
1247 root = root_mem_cgroup;
1248 if (prev && prev != root)
1249 css_put(&prev->css);
1252 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1253 struct mem_cgroup *dead_memcg)
1255 struct mem_cgroup_reclaim_iter *iter;
1256 struct mem_cgroup_per_node *mz;
1259 for_each_node(nid) {
1260 mz = mem_cgroup_nodeinfo(from, nid);
1262 cmpxchg(&iter->position, dead_memcg, NULL);
1266 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1268 struct mem_cgroup *memcg = dead_memcg;
1269 struct mem_cgroup *last;
1272 __invalidate_reclaim_iterators(memcg, dead_memcg);
1274 } while ((memcg = parent_mem_cgroup(memcg)));
1277 * When cgruop1 non-hierarchy mode is used,
1278 * parent_mem_cgroup() does not walk all the way up to the
1279 * cgroup root (root_mem_cgroup). So we have to handle
1280 * dead_memcg from cgroup root separately.
1282 if (last != root_mem_cgroup)
1283 __invalidate_reclaim_iterators(root_mem_cgroup,
1288 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1289 * @memcg: hierarchy root
1290 * @fn: function to call for each task
1291 * @arg: argument passed to @fn
1293 * This function iterates over tasks attached to @memcg or to any of its
1294 * descendants and calls @fn for each task. If @fn returns a non-zero
1295 * value, the function breaks the iteration loop and returns the value.
1296 * Otherwise, it will iterate over all tasks and return 0.
1298 * This function must not be called for the root memory cgroup.
1300 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1301 int (*fn)(struct task_struct *, void *), void *arg)
1303 struct mem_cgroup *iter;
1306 BUG_ON(memcg == root_mem_cgroup);
1308 for_each_mem_cgroup_tree(iter, memcg) {
1309 struct css_task_iter it;
1310 struct task_struct *task;
1312 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1313 while (!ret && (task = css_task_iter_next(&it)))
1314 ret = fn(task, arg);
1315 css_task_iter_end(&it);
1317 mem_cgroup_iter_break(memcg, iter);
1325 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1327 * @pgdat: pgdat of the page
1329 * This function relies on page->mem_cgroup being stable - see the
1330 * access rules in commit_charge().
1332 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1334 struct mem_cgroup_per_node *mz;
1335 struct mem_cgroup *memcg;
1336 struct lruvec *lruvec;
1338 if (mem_cgroup_disabled()) {
1339 lruvec = &pgdat->__lruvec;
1343 memcg = page_memcg(page);
1345 * Swapcache readahead pages are added to the LRU - and
1346 * possibly migrated - before they are charged.
1349 memcg = root_mem_cgroup;
1351 mz = mem_cgroup_page_nodeinfo(memcg, page);
1352 lruvec = &mz->lruvec;
1355 * Since a node can be onlined after the mem_cgroup was created,
1356 * we have to be prepared to initialize lruvec->zone here;
1357 * and if offlined then reonlined, we need to reinitialize it.
1359 if (unlikely(lruvec->pgdat != pgdat))
1360 lruvec->pgdat = pgdat;
1365 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1366 * @lruvec: mem_cgroup per zone lru vector
1367 * @lru: index of lru list the page is sitting on
1368 * @zid: zone id of the accounted pages
1369 * @nr_pages: positive when adding or negative when removing
1371 * This function must be called under lru_lock, just before a page is added
1372 * to or just after a page is removed from an lru list (that ordering being
1373 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1375 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1376 int zid, int nr_pages)
1378 struct mem_cgroup_per_node *mz;
1379 unsigned long *lru_size;
1382 if (mem_cgroup_disabled())
1385 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1386 lru_size = &mz->lru_zone_size[zid][lru];
1389 *lru_size += nr_pages;
1392 if (WARN_ONCE(size < 0,
1393 "%s(%p, %d, %d): lru_size %ld\n",
1394 __func__, lruvec, lru, nr_pages, size)) {
1400 *lru_size += nr_pages;
1404 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1405 * @memcg: the memory cgroup
1407 * Returns the maximum amount of memory @mem can be charged with, in
1410 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1412 unsigned long margin = 0;
1413 unsigned long count;
1414 unsigned long limit;
1416 count = page_counter_read(&memcg->memory);
1417 limit = READ_ONCE(memcg->memory.max);
1419 margin = limit - count;
1421 if (do_memsw_account()) {
1422 count = page_counter_read(&memcg->memsw);
1423 limit = READ_ONCE(memcg->memsw.max);
1425 margin = min(margin, limit - count);
1434 * A routine for checking "mem" is under move_account() or not.
1436 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1437 * moving cgroups. This is for waiting at high-memory pressure
1440 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1442 struct mem_cgroup *from;
1443 struct mem_cgroup *to;
1446 * Unlike task_move routines, we access mc.to, mc.from not under
1447 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1449 spin_lock(&mc.lock);
1455 ret = mem_cgroup_is_descendant(from, memcg) ||
1456 mem_cgroup_is_descendant(to, memcg);
1458 spin_unlock(&mc.lock);
1462 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1464 if (mc.moving_task && current != mc.moving_task) {
1465 if (mem_cgroup_under_move(memcg)) {
1467 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1468 /* moving charge context might have finished. */
1471 finish_wait(&mc.waitq, &wait);
1478 struct memory_stat {
1484 static struct memory_stat memory_stats[] = {
1485 { "anon", PAGE_SIZE, NR_ANON_MAPPED },
1486 { "file", PAGE_SIZE, NR_FILE_PAGES },
1487 { "kernel_stack", 1024, NR_KERNEL_STACK_KB },
1488 { "percpu", 1, MEMCG_PERCPU_B },
1489 { "sock", PAGE_SIZE, MEMCG_SOCK },
1490 { "shmem", PAGE_SIZE, NR_SHMEM },
1491 { "file_mapped", PAGE_SIZE, NR_FILE_MAPPED },
1492 { "file_dirty", PAGE_SIZE, NR_FILE_DIRTY },
1493 { "file_writeback", PAGE_SIZE, NR_WRITEBACK },
1494 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1496 * The ratio will be initialized in memory_stats_init(). Because
1497 * on some architectures, the macro of HPAGE_PMD_SIZE is not
1498 * constant(e.g. powerpc).
1500 { "anon_thp", 0, NR_ANON_THPS },
1502 { "inactive_anon", PAGE_SIZE, NR_INACTIVE_ANON },
1503 { "active_anon", PAGE_SIZE, NR_ACTIVE_ANON },
1504 { "inactive_file", PAGE_SIZE, NR_INACTIVE_FILE },
1505 { "active_file", PAGE_SIZE, NR_ACTIVE_FILE },
1506 { "unevictable", PAGE_SIZE, NR_UNEVICTABLE },
1509 * Note: The slab_reclaimable and slab_unreclaimable must be
1510 * together and slab_reclaimable must be in front.
1512 { "slab_reclaimable", 1, NR_SLAB_RECLAIMABLE_B },
1513 { "slab_unreclaimable", 1, NR_SLAB_UNRECLAIMABLE_B },
1515 /* The memory events */
1516 { "workingset_refault_anon", 1, WORKINGSET_REFAULT_ANON },
1517 { "workingset_refault_file", 1, WORKINGSET_REFAULT_FILE },
1518 { "workingset_activate_anon", 1, WORKINGSET_ACTIVATE_ANON },
1519 { "workingset_activate_file", 1, WORKINGSET_ACTIVATE_FILE },
1520 { "workingset_restore_anon", 1, WORKINGSET_RESTORE_ANON },
1521 { "workingset_restore_file", 1, WORKINGSET_RESTORE_FILE },
1522 { "workingset_nodereclaim", 1, WORKINGSET_NODERECLAIM },
1525 static int __init memory_stats_init(void)
1529 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1530 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1531 if (memory_stats[i].idx == NR_ANON_THPS)
1532 memory_stats[i].ratio = HPAGE_PMD_SIZE;
1534 VM_BUG_ON(!memory_stats[i].ratio);
1535 VM_BUG_ON(memory_stats[i].idx >= MEMCG_NR_STAT);
1540 pure_initcall(memory_stats_init);
1542 static char *memory_stat_format(struct mem_cgroup *memcg)
1547 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1552 * Provide statistics on the state of the memory subsystem as
1553 * well as cumulative event counters that show past behavior.
1555 * This list is ordered following a combination of these gradients:
1556 * 1) generic big picture -> specifics and details
1557 * 2) reflecting userspace activity -> reflecting kernel heuristics
1559 * Current memory state:
1562 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1565 size = memcg_page_state(memcg, memory_stats[i].idx);
1566 size *= memory_stats[i].ratio;
1567 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1569 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1570 size = memcg_page_state(memcg, NR_SLAB_RECLAIMABLE_B) +
1571 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE_B);
1572 seq_buf_printf(&s, "slab %llu\n", size);
1576 /* Accumulated memory events */
1578 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1579 memcg_events(memcg, PGFAULT));
1580 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1581 memcg_events(memcg, PGMAJFAULT));
1582 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1583 memcg_events(memcg, PGREFILL));
1584 seq_buf_printf(&s, "pgscan %lu\n",
1585 memcg_events(memcg, PGSCAN_KSWAPD) +
1586 memcg_events(memcg, PGSCAN_DIRECT));
1587 seq_buf_printf(&s, "pgsteal %lu\n",
1588 memcg_events(memcg, PGSTEAL_KSWAPD) +
1589 memcg_events(memcg, PGSTEAL_DIRECT));
1590 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1591 memcg_events(memcg, PGACTIVATE));
1592 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1593 memcg_events(memcg, PGDEACTIVATE));
1594 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1595 memcg_events(memcg, PGLAZYFREE));
1596 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1597 memcg_events(memcg, PGLAZYFREED));
1599 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1600 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1601 memcg_events(memcg, THP_FAULT_ALLOC));
1602 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1603 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1604 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1606 /* The above should easily fit into one page */
1607 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1612 #define K(x) ((x) << (PAGE_SHIFT-10))
1614 * mem_cgroup_print_oom_context: Print OOM information relevant to
1615 * memory controller.
1616 * @memcg: The memory cgroup that went over limit
1617 * @p: Task that is going to be killed
1619 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1622 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1627 pr_cont(",oom_memcg=");
1628 pr_cont_cgroup_path(memcg->css.cgroup);
1630 pr_cont(",global_oom");
1632 pr_cont(",task_memcg=");
1633 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1639 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1640 * memory controller.
1641 * @memcg: The memory cgroup that went over limit
1643 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1647 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1648 K((u64)page_counter_read(&memcg->memory)),
1649 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1650 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1651 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1652 K((u64)page_counter_read(&memcg->swap)),
1653 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1655 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1656 K((u64)page_counter_read(&memcg->memsw)),
1657 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1658 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1659 K((u64)page_counter_read(&memcg->kmem)),
1660 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1663 pr_info("Memory cgroup stats for ");
1664 pr_cont_cgroup_path(memcg->css.cgroup);
1666 buf = memory_stat_format(memcg);
1674 * Return the memory (and swap, if configured) limit for a memcg.
1676 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1678 unsigned long max = READ_ONCE(memcg->memory.max);
1680 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1681 if (mem_cgroup_swappiness(memcg))
1682 max += min(READ_ONCE(memcg->swap.max),
1683 (unsigned long)total_swap_pages);
1685 if (mem_cgroup_swappiness(memcg)) {
1686 /* Calculate swap excess capacity from memsw limit */
1687 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1689 max += min(swap, (unsigned long)total_swap_pages);
1695 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1697 return page_counter_read(&memcg->memory);
1700 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1703 struct oom_control oc = {
1707 .gfp_mask = gfp_mask,
1712 if (mutex_lock_killable(&oom_lock))
1715 if (mem_cgroup_margin(memcg) >= (1 << order))
1719 * A few threads which were not waiting at mutex_lock_killable() can
1720 * fail to bail out. Therefore, check again after holding oom_lock.
1722 ret = should_force_charge() || out_of_memory(&oc);
1725 mutex_unlock(&oom_lock);
1729 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1732 unsigned long *total_scanned)
1734 struct mem_cgroup *victim = NULL;
1737 unsigned long excess;
1738 unsigned long nr_scanned;
1739 struct mem_cgroup_reclaim_cookie reclaim = {
1743 excess = soft_limit_excess(root_memcg);
1746 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1751 * If we have not been able to reclaim
1752 * anything, it might because there are
1753 * no reclaimable pages under this hierarchy
1758 * We want to do more targeted reclaim.
1759 * excess >> 2 is not to excessive so as to
1760 * reclaim too much, nor too less that we keep
1761 * coming back to reclaim from this cgroup
1763 if (total >= (excess >> 2) ||
1764 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1769 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1770 pgdat, &nr_scanned);
1771 *total_scanned += nr_scanned;
1772 if (!soft_limit_excess(root_memcg))
1775 mem_cgroup_iter_break(root_memcg, victim);
1779 #ifdef CONFIG_LOCKDEP
1780 static struct lockdep_map memcg_oom_lock_dep_map = {
1781 .name = "memcg_oom_lock",
1785 static DEFINE_SPINLOCK(memcg_oom_lock);
1788 * Check OOM-Killer is already running under our hierarchy.
1789 * If someone is running, return false.
1791 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1793 struct mem_cgroup *iter, *failed = NULL;
1795 spin_lock(&memcg_oom_lock);
1797 for_each_mem_cgroup_tree(iter, memcg) {
1798 if (iter->oom_lock) {
1800 * this subtree of our hierarchy is already locked
1801 * so we cannot give a lock.
1804 mem_cgroup_iter_break(memcg, iter);
1807 iter->oom_lock = true;
1812 * OK, we failed to lock the whole subtree so we have
1813 * to clean up what we set up to the failing subtree
1815 for_each_mem_cgroup_tree(iter, memcg) {
1816 if (iter == failed) {
1817 mem_cgroup_iter_break(memcg, iter);
1820 iter->oom_lock = false;
1823 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1825 spin_unlock(&memcg_oom_lock);
1830 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1832 struct mem_cgroup *iter;
1834 spin_lock(&memcg_oom_lock);
1835 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1836 for_each_mem_cgroup_tree(iter, memcg)
1837 iter->oom_lock = false;
1838 spin_unlock(&memcg_oom_lock);
1841 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1843 struct mem_cgroup *iter;
1845 spin_lock(&memcg_oom_lock);
1846 for_each_mem_cgroup_tree(iter, memcg)
1848 spin_unlock(&memcg_oom_lock);
1851 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1853 struct mem_cgroup *iter;
1856 * Be careful about under_oom underflows becase a child memcg
1857 * could have been added after mem_cgroup_mark_under_oom.
1859 spin_lock(&memcg_oom_lock);
1860 for_each_mem_cgroup_tree(iter, memcg)
1861 if (iter->under_oom > 0)
1863 spin_unlock(&memcg_oom_lock);
1866 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1868 struct oom_wait_info {
1869 struct mem_cgroup *memcg;
1870 wait_queue_entry_t wait;
1873 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1874 unsigned mode, int sync, void *arg)
1876 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1877 struct mem_cgroup *oom_wait_memcg;
1878 struct oom_wait_info *oom_wait_info;
1880 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1881 oom_wait_memcg = oom_wait_info->memcg;
1883 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1884 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1886 return autoremove_wake_function(wait, mode, sync, arg);
1889 static void memcg_oom_recover(struct mem_cgroup *memcg)
1892 * For the following lockless ->under_oom test, the only required
1893 * guarantee is that it must see the state asserted by an OOM when
1894 * this function is called as a result of userland actions
1895 * triggered by the notification of the OOM. This is trivially
1896 * achieved by invoking mem_cgroup_mark_under_oom() before
1897 * triggering notification.
1899 if (memcg && memcg->under_oom)
1900 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1910 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1912 enum oom_status ret;
1915 if (order > PAGE_ALLOC_COSTLY_ORDER)
1918 memcg_memory_event(memcg, MEMCG_OOM);
1921 * We are in the middle of the charge context here, so we
1922 * don't want to block when potentially sitting on a callstack
1923 * that holds all kinds of filesystem and mm locks.
1925 * cgroup1 allows disabling the OOM killer and waiting for outside
1926 * handling until the charge can succeed; remember the context and put
1927 * the task to sleep at the end of the page fault when all locks are
1930 * On the other hand, in-kernel OOM killer allows for an async victim
1931 * memory reclaim (oom_reaper) and that means that we are not solely
1932 * relying on the oom victim to make a forward progress and we can
1933 * invoke the oom killer here.
1935 * Please note that mem_cgroup_out_of_memory might fail to find a
1936 * victim and then we have to bail out from the charge path.
1938 if (memcg->oom_kill_disable) {
1939 if (!current->in_user_fault)
1941 css_get(&memcg->css);
1942 current->memcg_in_oom = memcg;
1943 current->memcg_oom_gfp_mask = mask;
1944 current->memcg_oom_order = order;
1949 mem_cgroup_mark_under_oom(memcg);
1951 locked = mem_cgroup_oom_trylock(memcg);
1954 mem_cgroup_oom_notify(memcg);
1956 mem_cgroup_unmark_under_oom(memcg);
1957 if (mem_cgroup_out_of_memory(memcg, mask, order))
1963 mem_cgroup_oom_unlock(memcg);
1969 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1970 * @handle: actually kill/wait or just clean up the OOM state
1972 * This has to be called at the end of a page fault if the memcg OOM
1973 * handler was enabled.
1975 * Memcg supports userspace OOM handling where failed allocations must
1976 * sleep on a waitqueue until the userspace task resolves the
1977 * situation. Sleeping directly in the charge context with all kinds
1978 * of locks held is not a good idea, instead we remember an OOM state
1979 * in the task and mem_cgroup_oom_synchronize() has to be called at
1980 * the end of the page fault to complete the OOM handling.
1982 * Returns %true if an ongoing memcg OOM situation was detected and
1983 * completed, %false otherwise.
1985 bool mem_cgroup_oom_synchronize(bool handle)
1987 struct mem_cgroup *memcg = current->memcg_in_oom;
1988 struct oom_wait_info owait;
1991 /* OOM is global, do not handle */
1998 owait.memcg = memcg;
1999 owait.wait.flags = 0;
2000 owait.wait.func = memcg_oom_wake_function;
2001 owait.wait.private = current;
2002 INIT_LIST_HEAD(&owait.wait.entry);
2004 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2005 mem_cgroup_mark_under_oom(memcg);
2007 locked = mem_cgroup_oom_trylock(memcg);
2010 mem_cgroup_oom_notify(memcg);
2012 if (locked && !memcg->oom_kill_disable) {
2013 mem_cgroup_unmark_under_oom(memcg);
2014 finish_wait(&memcg_oom_waitq, &owait.wait);
2015 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
2016 current->memcg_oom_order);
2019 mem_cgroup_unmark_under_oom(memcg);
2020 finish_wait(&memcg_oom_waitq, &owait.wait);
2024 mem_cgroup_oom_unlock(memcg);
2026 * There is no guarantee that an OOM-lock contender
2027 * sees the wakeups triggered by the OOM kill
2028 * uncharges. Wake any sleepers explicitely.
2030 memcg_oom_recover(memcg);
2033 current->memcg_in_oom = NULL;
2034 css_put(&memcg->css);
2039 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2040 * @victim: task to be killed by the OOM killer
2041 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2043 * Returns a pointer to a memory cgroup, which has to be cleaned up
2044 * by killing all belonging OOM-killable tasks.
2046 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2048 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2049 struct mem_cgroup *oom_domain)
2051 struct mem_cgroup *oom_group = NULL;
2052 struct mem_cgroup *memcg;
2054 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2058 oom_domain = root_mem_cgroup;
2062 memcg = mem_cgroup_from_task(victim);
2063 if (memcg == root_mem_cgroup)
2067 * If the victim task has been asynchronously moved to a different
2068 * memory cgroup, we might end up killing tasks outside oom_domain.
2069 * In this case it's better to ignore memory.group.oom.
2071 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2075 * Traverse the memory cgroup hierarchy from the victim task's
2076 * cgroup up to the OOMing cgroup (or root) to find the
2077 * highest-level memory cgroup with oom.group set.
2079 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2080 if (memcg->oom_group)
2083 if (memcg == oom_domain)
2088 css_get(&oom_group->css);
2095 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2097 pr_info("Tasks in ");
2098 pr_cont_cgroup_path(memcg->css.cgroup);
2099 pr_cont(" are going to be killed due to memory.oom.group set\n");
2103 * lock_page_memcg - lock a page and memcg binding
2106 * This function protects unlocked LRU pages from being moved to
2109 * It ensures lifetime of the returned memcg. Caller is responsible
2110 * for the lifetime of the page; __unlock_page_memcg() is available
2111 * when @page might get freed inside the locked section.
2113 struct mem_cgroup *lock_page_memcg(struct page *page)
2115 struct page *head = compound_head(page); /* rmap on tail pages */
2116 struct mem_cgroup *memcg;
2117 unsigned long flags;
2120 * The RCU lock is held throughout the transaction. The fast
2121 * path can get away without acquiring the memcg->move_lock
2122 * because page moving starts with an RCU grace period.
2124 * The RCU lock also protects the memcg from being freed when
2125 * the page state that is going to change is the only thing
2126 * preventing the page itself from being freed. E.g. writeback
2127 * doesn't hold a page reference and relies on PG_writeback to
2128 * keep off truncation, migration and so forth.
2132 if (mem_cgroup_disabled())
2135 memcg = page_memcg(head);
2136 if (unlikely(!memcg))
2139 if (atomic_read(&memcg->moving_account) <= 0)
2142 spin_lock_irqsave(&memcg->move_lock, flags);
2143 if (memcg != page_memcg(head)) {
2144 spin_unlock_irqrestore(&memcg->move_lock, flags);
2149 * When charge migration first begins, we can have locked and
2150 * unlocked page stat updates happening concurrently. Track
2151 * the task who has the lock for unlock_page_memcg().
2153 memcg->move_lock_task = current;
2154 memcg->move_lock_flags = flags;
2158 EXPORT_SYMBOL(lock_page_memcg);
2161 * __unlock_page_memcg - unlock and unpin a memcg
2164 * Unlock and unpin a memcg returned by lock_page_memcg().
2166 void __unlock_page_memcg(struct mem_cgroup *memcg)
2168 if (memcg && memcg->move_lock_task == current) {
2169 unsigned long flags = memcg->move_lock_flags;
2171 memcg->move_lock_task = NULL;
2172 memcg->move_lock_flags = 0;
2174 spin_unlock_irqrestore(&memcg->move_lock, flags);
2181 * unlock_page_memcg - unlock a page and memcg binding
2184 void unlock_page_memcg(struct page *page)
2186 struct page *head = compound_head(page);
2188 __unlock_page_memcg(page_memcg(head));
2190 EXPORT_SYMBOL(unlock_page_memcg);
2192 struct memcg_stock_pcp {
2193 struct mem_cgroup *cached; /* this never be root cgroup */
2194 unsigned int nr_pages;
2196 #ifdef CONFIG_MEMCG_KMEM
2197 struct obj_cgroup *cached_objcg;
2198 unsigned int nr_bytes;
2201 struct work_struct work;
2202 unsigned long flags;
2203 #define FLUSHING_CACHED_CHARGE 0
2205 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2206 static DEFINE_MUTEX(percpu_charge_mutex);
2208 #ifdef CONFIG_MEMCG_KMEM
2209 static void drain_obj_stock(struct memcg_stock_pcp *stock);
2210 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2211 struct mem_cgroup *root_memcg);
2214 static inline void drain_obj_stock(struct memcg_stock_pcp *stock)
2217 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2218 struct mem_cgroup *root_memcg)
2225 * consume_stock: Try to consume stocked charge on this cpu.
2226 * @memcg: memcg to consume from.
2227 * @nr_pages: how many pages to charge.
2229 * The charges will only happen if @memcg matches the current cpu's memcg
2230 * stock, and at least @nr_pages are available in that stock. Failure to
2231 * service an allocation will refill the stock.
2233 * returns true if successful, false otherwise.
2235 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2237 struct memcg_stock_pcp *stock;
2238 unsigned long flags;
2241 if (nr_pages > MEMCG_CHARGE_BATCH)
2244 local_irq_save(flags);
2246 stock = this_cpu_ptr(&memcg_stock);
2247 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2248 stock->nr_pages -= nr_pages;
2252 local_irq_restore(flags);
2258 * Returns stocks cached in percpu and reset cached information.
2260 static void drain_stock(struct memcg_stock_pcp *stock)
2262 struct mem_cgroup *old = stock->cached;
2267 if (stock->nr_pages) {
2268 page_counter_uncharge(&old->memory, stock->nr_pages);
2269 if (do_memsw_account())
2270 page_counter_uncharge(&old->memsw, stock->nr_pages);
2271 stock->nr_pages = 0;
2275 stock->cached = NULL;
2278 static void drain_local_stock(struct work_struct *dummy)
2280 struct memcg_stock_pcp *stock;
2281 unsigned long flags;
2284 * The only protection from memory hotplug vs. drain_stock races is
2285 * that we always operate on local CPU stock here with IRQ disabled
2287 local_irq_save(flags);
2289 stock = this_cpu_ptr(&memcg_stock);
2290 drain_obj_stock(stock);
2292 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2294 local_irq_restore(flags);
2298 * Cache charges(val) to local per_cpu area.
2299 * This will be consumed by consume_stock() function, later.
2301 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2303 struct memcg_stock_pcp *stock;
2304 unsigned long flags;
2306 local_irq_save(flags);
2308 stock = this_cpu_ptr(&memcg_stock);
2309 if (stock->cached != memcg) { /* reset if necessary */
2311 css_get(&memcg->css);
2312 stock->cached = memcg;
2314 stock->nr_pages += nr_pages;
2316 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2319 local_irq_restore(flags);
2323 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2324 * of the hierarchy under it.
2326 static void drain_all_stock(struct mem_cgroup *root_memcg)
2330 /* If someone's already draining, avoid adding running more workers. */
2331 if (!mutex_trylock(&percpu_charge_mutex))
2334 * Notify other cpus that system-wide "drain" is running
2335 * We do not care about races with the cpu hotplug because cpu down
2336 * as well as workers from this path always operate on the local
2337 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2340 for_each_online_cpu(cpu) {
2341 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2342 struct mem_cgroup *memcg;
2346 memcg = stock->cached;
2347 if (memcg && stock->nr_pages &&
2348 mem_cgroup_is_descendant(memcg, root_memcg))
2350 if (obj_stock_flush_required(stock, root_memcg))
2355 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2357 drain_local_stock(&stock->work);
2359 schedule_work_on(cpu, &stock->work);
2363 mutex_unlock(&percpu_charge_mutex);
2366 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2368 struct memcg_stock_pcp *stock;
2369 struct mem_cgroup *memcg, *mi;
2371 stock = &per_cpu(memcg_stock, cpu);
2374 for_each_mem_cgroup(memcg) {
2377 for (i = 0; i < MEMCG_NR_STAT; i++) {
2381 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2383 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2384 atomic_long_add(x, &memcg->vmstats[i]);
2386 if (i >= NR_VM_NODE_STAT_ITEMS)
2389 for_each_node(nid) {
2390 struct mem_cgroup_per_node *pn;
2392 pn = mem_cgroup_nodeinfo(memcg, nid);
2393 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2396 atomic_long_add(x, &pn->lruvec_stat[i]);
2397 } while ((pn = parent_nodeinfo(pn, nid)));
2401 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2404 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2406 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2407 atomic_long_add(x, &memcg->vmevents[i]);
2414 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2415 unsigned int nr_pages,
2418 unsigned long nr_reclaimed = 0;
2421 unsigned long pflags;
2423 if (page_counter_read(&memcg->memory) <=
2424 READ_ONCE(memcg->memory.high))
2427 memcg_memory_event(memcg, MEMCG_HIGH);
2429 psi_memstall_enter(&pflags);
2430 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2432 psi_memstall_leave(&pflags);
2433 } while ((memcg = parent_mem_cgroup(memcg)) &&
2434 !mem_cgroup_is_root(memcg));
2436 return nr_reclaimed;
2439 static void high_work_func(struct work_struct *work)
2441 struct mem_cgroup *memcg;
2443 memcg = container_of(work, struct mem_cgroup, high_work);
2444 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2448 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2449 * enough to still cause a significant slowdown in most cases, while still
2450 * allowing diagnostics and tracing to proceed without becoming stuck.
2452 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2455 * When calculating the delay, we use these either side of the exponentiation to
2456 * maintain precision and scale to a reasonable number of jiffies (see the table
2459 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2460 * overage ratio to a delay.
2461 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2462 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2463 * to produce a reasonable delay curve.
2465 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2466 * reasonable delay curve compared to precision-adjusted overage, not
2467 * penalising heavily at first, but still making sure that growth beyond the
2468 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2469 * example, with a high of 100 megabytes:
2471 * +-------+------------------------+
2472 * | usage | time to allocate in ms |
2473 * +-------+------------------------+
2495 * +-------+------------------------+
2497 #define MEMCG_DELAY_PRECISION_SHIFT 20
2498 #define MEMCG_DELAY_SCALING_SHIFT 14
2500 static u64 calculate_overage(unsigned long usage, unsigned long high)
2508 * Prevent division by 0 in overage calculation by acting as if
2509 * it was a threshold of 1 page
2511 high = max(high, 1UL);
2513 overage = usage - high;
2514 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2515 return div64_u64(overage, high);
2518 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2520 u64 overage, max_overage = 0;
2523 overage = calculate_overage(page_counter_read(&memcg->memory),
2524 READ_ONCE(memcg->memory.high));
2525 max_overage = max(overage, max_overage);
2526 } while ((memcg = parent_mem_cgroup(memcg)) &&
2527 !mem_cgroup_is_root(memcg));
2532 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2534 u64 overage, max_overage = 0;
2537 overage = calculate_overage(page_counter_read(&memcg->swap),
2538 READ_ONCE(memcg->swap.high));
2540 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2541 max_overage = max(overage, max_overage);
2542 } while ((memcg = parent_mem_cgroup(memcg)) &&
2543 !mem_cgroup_is_root(memcg));
2549 * Get the number of jiffies that we should penalise a mischievous cgroup which
2550 * is exceeding its memory.high by checking both it and its ancestors.
2552 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2553 unsigned int nr_pages,
2556 unsigned long penalty_jiffies;
2562 * We use overage compared to memory.high to calculate the number of
2563 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2564 * fairly lenient on small overages, and increasingly harsh when the
2565 * memcg in question makes it clear that it has no intention of stopping
2566 * its crazy behaviour, so we exponentially increase the delay based on
2569 penalty_jiffies = max_overage * max_overage * HZ;
2570 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2571 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2574 * Factor in the task's own contribution to the overage, such that four
2575 * N-sized allocations are throttled approximately the same as one
2576 * 4N-sized allocation.
2578 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2579 * larger the current charge patch is than that.
2581 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2585 * Scheduled by try_charge() to be executed from the userland return path
2586 * and reclaims memory over the high limit.
2588 void mem_cgroup_handle_over_high(void)
2590 unsigned long penalty_jiffies;
2591 unsigned long pflags;
2592 unsigned long nr_reclaimed;
2593 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2594 int nr_retries = MAX_RECLAIM_RETRIES;
2595 struct mem_cgroup *memcg;
2596 bool in_retry = false;
2598 if (likely(!nr_pages))
2601 memcg = get_mem_cgroup_from_mm(current->mm);
2602 current->memcg_nr_pages_over_high = 0;
2606 * The allocating task should reclaim at least the batch size, but for
2607 * subsequent retries we only want to do what's necessary to prevent oom
2608 * or breaching resource isolation.
2610 * This is distinct from memory.max or page allocator behaviour because
2611 * memory.high is currently batched, whereas memory.max and the page
2612 * allocator run every time an allocation is made.
2614 nr_reclaimed = reclaim_high(memcg,
2615 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2619 * memory.high is breached and reclaim is unable to keep up. Throttle
2620 * allocators proactively to slow down excessive growth.
2622 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2623 mem_find_max_overage(memcg));
2625 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2626 swap_find_max_overage(memcg));
2629 * Clamp the max delay per usermode return so as to still keep the
2630 * application moving forwards and also permit diagnostics, albeit
2633 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2636 * Don't sleep if the amount of jiffies this memcg owes us is so low
2637 * that it's not even worth doing, in an attempt to be nice to those who
2638 * go only a small amount over their memory.high value and maybe haven't
2639 * been aggressively reclaimed enough yet.
2641 if (penalty_jiffies <= HZ / 100)
2645 * If reclaim is making forward progress but we're still over
2646 * memory.high, we want to encourage that rather than doing allocator
2649 if (nr_reclaimed || nr_retries--) {
2655 * If we exit early, we're guaranteed to die (since
2656 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2657 * need to account for any ill-begotten jiffies to pay them off later.
2659 psi_memstall_enter(&pflags);
2660 schedule_timeout_killable(penalty_jiffies);
2661 psi_memstall_leave(&pflags);
2664 css_put(&memcg->css);
2667 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2668 unsigned int nr_pages)
2670 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2671 int nr_retries = MAX_RECLAIM_RETRIES;
2672 struct mem_cgroup *mem_over_limit;
2673 struct page_counter *counter;
2674 enum oom_status oom_status;
2675 unsigned long nr_reclaimed;
2676 bool may_swap = true;
2677 bool drained = false;
2678 unsigned long pflags;
2680 if (mem_cgroup_is_root(memcg))
2683 if (consume_stock(memcg, nr_pages))
2686 if (!do_memsw_account() ||
2687 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2688 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2690 if (do_memsw_account())
2691 page_counter_uncharge(&memcg->memsw, batch);
2692 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2694 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2698 if (batch > nr_pages) {
2704 * Memcg doesn't have a dedicated reserve for atomic
2705 * allocations. But like the global atomic pool, we need to
2706 * put the burden of reclaim on regular allocation requests
2707 * and let these go through as privileged allocations.
2709 if (gfp_mask & __GFP_ATOMIC)
2713 * Unlike in global OOM situations, memcg is not in a physical
2714 * memory shortage. Allow dying and OOM-killed tasks to
2715 * bypass the last charges so that they can exit quickly and
2716 * free their memory.
2718 if (unlikely(should_force_charge()))
2722 * Prevent unbounded recursion when reclaim operations need to
2723 * allocate memory. This might exceed the limits temporarily,
2724 * but we prefer facilitating memory reclaim and getting back
2725 * under the limit over triggering OOM kills in these cases.
2727 if (unlikely(current->flags & PF_MEMALLOC))
2730 if (unlikely(task_in_memcg_oom(current)))
2733 if (!gfpflags_allow_blocking(gfp_mask))
2736 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2738 psi_memstall_enter(&pflags);
2739 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2740 gfp_mask, may_swap);
2741 psi_memstall_leave(&pflags);
2743 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2747 drain_all_stock(mem_over_limit);
2752 if (gfp_mask & __GFP_NORETRY)
2755 * Even though the limit is exceeded at this point, reclaim
2756 * may have been able to free some pages. Retry the charge
2757 * before killing the task.
2759 * Only for regular pages, though: huge pages are rather
2760 * unlikely to succeed so close to the limit, and we fall back
2761 * to regular pages anyway in case of failure.
2763 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2766 * At task move, charge accounts can be doubly counted. So, it's
2767 * better to wait until the end of task_move if something is going on.
2769 if (mem_cgroup_wait_acct_move(mem_over_limit))
2775 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2778 if (gfp_mask & __GFP_NOFAIL)
2781 if (fatal_signal_pending(current))
2785 * keep retrying as long as the memcg oom killer is able to make
2786 * a forward progress or bypass the charge if the oom killer
2787 * couldn't make any progress.
2789 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2790 get_order(nr_pages * PAGE_SIZE));
2791 switch (oom_status) {
2793 nr_retries = MAX_RECLAIM_RETRIES;
2801 if (!(gfp_mask & __GFP_NOFAIL))
2805 * The allocation either can't fail or will lead to more memory
2806 * being freed very soon. Allow memory usage go over the limit
2807 * temporarily by force charging it.
2809 page_counter_charge(&memcg->memory, nr_pages);
2810 if (do_memsw_account())
2811 page_counter_charge(&memcg->memsw, nr_pages);
2816 if (batch > nr_pages)
2817 refill_stock(memcg, batch - nr_pages);
2820 * If the hierarchy is above the normal consumption range, schedule
2821 * reclaim on returning to userland. We can perform reclaim here
2822 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2823 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2824 * not recorded as it most likely matches current's and won't
2825 * change in the meantime. As high limit is checked again before
2826 * reclaim, the cost of mismatch is negligible.
2829 bool mem_high, swap_high;
2831 mem_high = page_counter_read(&memcg->memory) >
2832 READ_ONCE(memcg->memory.high);
2833 swap_high = page_counter_read(&memcg->swap) >
2834 READ_ONCE(memcg->swap.high);
2836 /* Don't bother a random interrupted task */
2837 if (in_interrupt()) {
2839 schedule_work(&memcg->high_work);
2845 if (mem_high || swap_high) {
2847 * The allocating tasks in this cgroup will need to do
2848 * reclaim or be throttled to prevent further growth
2849 * of the memory or swap footprints.
2851 * Target some best-effort fairness between the tasks,
2852 * and distribute reclaim work and delay penalties
2853 * based on how much each task is actually allocating.
2855 current->memcg_nr_pages_over_high += batch;
2856 set_notify_resume(current);
2859 } while ((memcg = parent_mem_cgroup(memcg)));
2864 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2865 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2867 if (mem_cgroup_is_root(memcg))
2870 page_counter_uncharge(&memcg->memory, nr_pages);
2871 if (do_memsw_account())
2872 page_counter_uncharge(&memcg->memsw, nr_pages);
2876 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2878 VM_BUG_ON_PAGE(page_memcg(page), page);
2880 * Any of the following ensures page->mem_cgroup stability:
2884 * - lock_page_memcg()
2885 * - exclusive reference
2887 page->memcg_data = (unsigned long)memcg;
2890 #ifdef CONFIG_MEMCG_KMEM
2891 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2894 unsigned int objects = objs_per_slab_page(s, page);
2897 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2902 if (!set_page_objcgs(page, vec))
2905 kmemleak_not_leak(vec);
2911 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2913 * A passed kernel object can be a slab object or a generic kernel page, so
2914 * different mechanisms for getting the memory cgroup pointer should be used.
2915 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2916 * can not know for sure how the kernel object is implemented.
2917 * mem_cgroup_from_obj() can be safely used in such cases.
2919 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2920 * cgroup_mutex, etc.
2922 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2926 if (mem_cgroup_disabled())
2929 page = virt_to_head_page(p);
2932 * Slab objects are accounted individually, not per-page.
2933 * Memcg membership data for each individual object is saved in
2934 * the page->obj_cgroups.
2936 if (page_objcgs_check(page)) {
2937 struct obj_cgroup *objcg;
2940 off = obj_to_index(page->slab_cache, page, p);
2941 objcg = page_objcgs(page)[off];
2943 return obj_cgroup_memcg(objcg);
2949 * page_memcg_check() is used here, because page_has_obj_cgroups()
2950 * check above could fail because the object cgroups vector wasn't set
2951 * at that moment, but it can be set concurrently.
2952 * page_memcg_check(page) will guarantee that a proper memory
2953 * cgroup pointer or NULL will be returned.
2955 return page_memcg_check(page);
2958 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2960 struct obj_cgroup *objcg = NULL;
2961 struct mem_cgroup *memcg;
2963 if (memcg_kmem_bypass())
2967 if (unlikely(active_memcg()))
2968 memcg = active_memcg();
2970 memcg = mem_cgroup_from_task(current);
2972 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2973 objcg = rcu_dereference(memcg->objcg);
2974 if (objcg && obj_cgroup_tryget(objcg))
2982 static int memcg_alloc_cache_id(void)
2987 id = ida_simple_get(&memcg_cache_ida,
2988 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2992 if (id < memcg_nr_cache_ids)
2996 * There's no space for the new id in memcg_caches arrays,
2997 * so we have to grow them.
2999 down_write(&memcg_cache_ids_sem);
3001 size = 2 * (id + 1);
3002 if (size < MEMCG_CACHES_MIN_SIZE)
3003 size = MEMCG_CACHES_MIN_SIZE;
3004 else if (size > MEMCG_CACHES_MAX_SIZE)
3005 size = MEMCG_CACHES_MAX_SIZE;
3007 err = memcg_update_all_list_lrus(size);
3009 memcg_nr_cache_ids = size;
3011 up_write(&memcg_cache_ids_sem);
3014 ida_simple_remove(&memcg_cache_ida, id);
3020 static void memcg_free_cache_id(int id)
3022 ida_simple_remove(&memcg_cache_ida, id);
3026 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
3027 * @memcg: memory cgroup to charge
3028 * @gfp: reclaim mode
3029 * @nr_pages: number of pages to charge
3031 * Returns 0 on success, an error code on failure.
3033 int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
3034 unsigned int nr_pages)
3036 struct page_counter *counter;
3039 ret = try_charge(memcg, gfp, nr_pages);
3043 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3044 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3047 * Enforce __GFP_NOFAIL allocation because callers are not
3048 * prepared to see failures and likely do not have any failure
3051 if (gfp & __GFP_NOFAIL) {
3052 page_counter_charge(&memcg->kmem, nr_pages);
3055 cancel_charge(memcg, nr_pages);
3062 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
3063 * @memcg: memcg to uncharge
3064 * @nr_pages: number of pages to uncharge
3066 void __memcg_kmem_uncharge(struct mem_cgroup *memcg, unsigned int nr_pages)
3068 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3069 page_counter_uncharge(&memcg->kmem, nr_pages);
3071 page_counter_uncharge(&memcg->memory, nr_pages);
3072 if (do_memsw_account())
3073 page_counter_uncharge(&memcg->memsw, nr_pages);
3077 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3078 * @page: page to charge
3079 * @gfp: reclaim mode
3080 * @order: allocation order
3082 * Returns 0 on success, an error code on failure.
3084 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3086 struct mem_cgroup *memcg;
3089 memcg = get_mem_cgroup_from_current();
3090 if (memcg && !mem_cgroup_is_root(memcg)) {
3091 ret = __memcg_kmem_charge(memcg, gfp, 1 << order);
3093 page->memcg_data = (unsigned long)memcg |
3097 css_put(&memcg->css);
3103 * __memcg_kmem_uncharge_page: uncharge a kmem page
3104 * @page: page to uncharge
3105 * @order: allocation order
3107 void __memcg_kmem_uncharge_page(struct page *page, int order)
3109 struct mem_cgroup *memcg = page_memcg(page);
3110 unsigned int nr_pages = 1 << order;
3115 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3116 __memcg_kmem_uncharge(memcg, nr_pages);
3117 page->memcg_data = 0;
3118 css_put(&memcg->css);
3121 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3123 struct memcg_stock_pcp *stock;
3124 unsigned long flags;
3127 local_irq_save(flags);
3129 stock = this_cpu_ptr(&memcg_stock);
3130 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3131 stock->nr_bytes -= nr_bytes;
3135 local_irq_restore(flags);
3140 static void drain_obj_stock(struct memcg_stock_pcp *stock)
3142 struct obj_cgroup *old = stock->cached_objcg;
3147 if (stock->nr_bytes) {
3148 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3149 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3153 __memcg_kmem_uncharge(obj_cgroup_memcg(old), nr_pages);
3158 * The leftover is flushed to the centralized per-memcg value.
3159 * On the next attempt to refill obj stock it will be moved
3160 * to a per-cpu stock (probably, on an other CPU), see
3161 * refill_obj_stock().
3163 * How often it's flushed is a trade-off between the memory
3164 * limit enforcement accuracy and potential CPU contention,
3165 * so it might be changed in the future.
3167 atomic_add(nr_bytes, &old->nr_charged_bytes);
3168 stock->nr_bytes = 0;
3171 obj_cgroup_put(old);
3172 stock->cached_objcg = NULL;
3175 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3176 struct mem_cgroup *root_memcg)
3178 struct mem_cgroup *memcg;
3180 if (stock->cached_objcg) {
3181 memcg = obj_cgroup_memcg(stock->cached_objcg);
3182 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3189 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3191 struct memcg_stock_pcp *stock;
3192 unsigned long flags;
3194 local_irq_save(flags);
3196 stock = this_cpu_ptr(&memcg_stock);
3197 if (stock->cached_objcg != objcg) { /* reset if necessary */
3198 drain_obj_stock(stock);
3199 obj_cgroup_get(objcg);
3200 stock->cached_objcg = objcg;
3201 stock->nr_bytes = atomic_xchg(&objcg->nr_charged_bytes, 0);
3203 stock->nr_bytes += nr_bytes;
3205 if (stock->nr_bytes > PAGE_SIZE)
3206 drain_obj_stock(stock);
3208 local_irq_restore(flags);
3211 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3213 struct mem_cgroup *memcg;
3214 unsigned int nr_pages, nr_bytes;
3217 if (consume_obj_stock(objcg, size))
3221 * In theory, memcg->nr_charged_bytes can have enough
3222 * pre-charged bytes to satisfy the allocation. However,
3223 * flushing memcg->nr_charged_bytes requires two atomic
3224 * operations, and memcg->nr_charged_bytes can't be big,
3225 * so it's better to ignore it and try grab some new pages.
3226 * memcg->nr_charged_bytes will be flushed in
3227 * refill_obj_stock(), called from this function or
3228 * independently later.
3231 memcg = obj_cgroup_memcg(objcg);
3232 css_get(&memcg->css);
3235 nr_pages = size >> PAGE_SHIFT;
3236 nr_bytes = size & (PAGE_SIZE - 1);
3241 ret = __memcg_kmem_charge(memcg, gfp, nr_pages);
3242 if (!ret && nr_bytes)
3243 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes);
3245 css_put(&memcg->css);
3249 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3251 refill_obj_stock(objcg, size);
3254 #endif /* CONFIG_MEMCG_KMEM */
3256 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3259 * Because tail pages are not marked as "used", set it. We're under
3260 * pgdat->lru_lock and migration entries setup in all page mappings.
3262 void mem_cgroup_split_huge_fixup(struct page *head)
3264 struct mem_cgroup *memcg = page_memcg(head);
3267 if (mem_cgroup_disabled())
3270 for (i = 1; i < HPAGE_PMD_NR; i++) {
3271 css_get(&memcg->css);
3272 head[i].memcg_data = (unsigned long)memcg;
3275 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3277 #ifdef CONFIG_MEMCG_SWAP
3279 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3280 * @entry: swap entry to be moved
3281 * @from: mem_cgroup which the entry is moved from
3282 * @to: mem_cgroup which the entry is moved to
3284 * It succeeds only when the swap_cgroup's record for this entry is the same
3285 * as the mem_cgroup's id of @from.
3287 * Returns 0 on success, -EINVAL on failure.
3289 * The caller must have charged to @to, IOW, called page_counter_charge() about
3290 * both res and memsw, and called css_get().
3292 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3293 struct mem_cgroup *from, struct mem_cgroup *to)
3295 unsigned short old_id, new_id;
3297 old_id = mem_cgroup_id(from);
3298 new_id = mem_cgroup_id(to);
3300 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3301 mod_memcg_state(from, MEMCG_SWAP, -1);
3302 mod_memcg_state(to, MEMCG_SWAP, 1);
3308 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3309 struct mem_cgroup *from, struct mem_cgroup *to)
3315 static DEFINE_MUTEX(memcg_max_mutex);
3317 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3318 unsigned long max, bool memsw)
3320 bool enlarge = false;
3321 bool drained = false;
3323 bool limits_invariant;
3324 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3327 if (signal_pending(current)) {
3332 mutex_lock(&memcg_max_mutex);
3334 * Make sure that the new limit (memsw or memory limit) doesn't
3335 * break our basic invariant rule memory.max <= memsw.max.
3337 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3338 max <= memcg->memsw.max;
3339 if (!limits_invariant) {
3340 mutex_unlock(&memcg_max_mutex);
3344 if (max > counter->max)
3346 ret = page_counter_set_max(counter, max);
3347 mutex_unlock(&memcg_max_mutex);
3353 drain_all_stock(memcg);
3358 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3359 GFP_KERNEL, !memsw)) {
3365 if (!ret && enlarge)
3366 memcg_oom_recover(memcg);
3371 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3373 unsigned long *total_scanned)
3375 unsigned long nr_reclaimed = 0;
3376 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3377 unsigned long reclaimed;
3379 struct mem_cgroup_tree_per_node *mctz;
3380 unsigned long excess;
3381 unsigned long nr_scanned;
3386 mctz = soft_limit_tree_node(pgdat->node_id);
3389 * Do not even bother to check the largest node if the root
3390 * is empty. Do it lockless to prevent lock bouncing. Races
3391 * are acceptable as soft limit is best effort anyway.
3393 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3397 * This loop can run a while, specially if mem_cgroup's continuously
3398 * keep exceeding their soft limit and putting the system under
3405 mz = mem_cgroup_largest_soft_limit_node(mctz);
3410 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3411 gfp_mask, &nr_scanned);
3412 nr_reclaimed += reclaimed;
3413 *total_scanned += nr_scanned;
3414 spin_lock_irq(&mctz->lock);
3415 __mem_cgroup_remove_exceeded(mz, mctz);
3418 * If we failed to reclaim anything from this memory cgroup
3419 * it is time to move on to the next cgroup
3423 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3425 excess = soft_limit_excess(mz->memcg);
3427 * One school of thought says that we should not add
3428 * back the node to the tree if reclaim returns 0.
3429 * But our reclaim could return 0, simply because due
3430 * to priority we are exposing a smaller subset of
3431 * memory to reclaim from. Consider this as a longer
3434 /* If excess == 0, no tree ops */
3435 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3436 spin_unlock_irq(&mctz->lock);
3437 css_put(&mz->memcg->css);
3440 * Could not reclaim anything and there are no more
3441 * mem cgroups to try or we seem to be looping without
3442 * reclaiming anything.
3444 if (!nr_reclaimed &&
3446 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3448 } while (!nr_reclaimed);
3450 css_put(&next_mz->memcg->css);
3451 return nr_reclaimed;
3455 * Test whether @memcg has children, dead or alive. Note that this
3456 * function doesn't care whether @memcg has use_hierarchy enabled and
3457 * returns %true if there are child csses according to the cgroup
3458 * hierarchy. Testing use_hierarchy is the caller's responsibility.
3460 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3465 ret = css_next_child(NULL, &memcg->css);
3471 * Reclaims as many pages from the given memcg as possible.
3473 * Caller is responsible for holding css reference for memcg.
3475 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3477 int nr_retries = MAX_RECLAIM_RETRIES;
3479 /* we call try-to-free pages for make this cgroup empty */
3480 lru_add_drain_all();
3482 drain_all_stock(memcg);
3484 /* try to free all pages in this cgroup */
3485 while (nr_retries && page_counter_read(&memcg->memory)) {
3488 if (signal_pending(current))
3491 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3495 /* maybe some writeback is necessary */
3496 congestion_wait(BLK_RW_ASYNC, HZ/10);
3504 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3505 char *buf, size_t nbytes,
3508 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3510 if (mem_cgroup_is_root(memcg))
3512 return mem_cgroup_force_empty(memcg) ?: nbytes;
3515 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3518 return mem_cgroup_from_css(css)->use_hierarchy;
3521 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3522 struct cftype *cft, u64 val)
3525 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3526 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3528 if (memcg->use_hierarchy == val)
3532 * If parent's use_hierarchy is set, we can't make any modifications
3533 * in the child subtrees. If it is unset, then the change can
3534 * occur, provided the current cgroup has no children.
3536 * For the root cgroup, parent_mem is NULL, we allow value to be
3537 * set if there are no children.
3539 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3540 (val == 1 || val == 0)) {
3541 if (!memcg_has_children(memcg))
3542 memcg->use_hierarchy = val;
3551 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3555 if (mem_cgroup_is_root(memcg)) {
3556 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3557 memcg_page_state(memcg, NR_ANON_MAPPED);
3559 val += memcg_page_state(memcg, MEMCG_SWAP);
3562 val = page_counter_read(&memcg->memory);
3564 val = page_counter_read(&memcg->memsw);
3577 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3580 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3581 struct page_counter *counter;
3583 switch (MEMFILE_TYPE(cft->private)) {
3585 counter = &memcg->memory;
3588 counter = &memcg->memsw;
3591 counter = &memcg->kmem;
3594 counter = &memcg->tcpmem;
3600 switch (MEMFILE_ATTR(cft->private)) {
3602 if (counter == &memcg->memory)
3603 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3604 if (counter == &memcg->memsw)
3605 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3606 return (u64)page_counter_read(counter) * PAGE_SIZE;
3608 return (u64)counter->max * PAGE_SIZE;
3610 return (u64)counter->watermark * PAGE_SIZE;
3612 return counter->failcnt;
3613 case RES_SOFT_LIMIT:
3614 return (u64)memcg->soft_limit * PAGE_SIZE;
3620 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3622 unsigned long stat[MEMCG_NR_STAT] = {0};
3623 struct mem_cgroup *mi;
3626 for_each_online_cpu(cpu)
3627 for (i = 0; i < MEMCG_NR_STAT; i++)
3628 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3630 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3631 for (i = 0; i < MEMCG_NR_STAT; i++)
3632 atomic_long_add(stat[i], &mi->vmstats[i]);
3634 for_each_node(node) {
3635 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3636 struct mem_cgroup_per_node *pi;
3638 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3641 for_each_online_cpu(cpu)
3642 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3644 pn->lruvec_stat_cpu->count[i], cpu);
3646 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3647 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3648 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3652 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3654 unsigned long events[NR_VM_EVENT_ITEMS];
3655 struct mem_cgroup *mi;
3658 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3661 for_each_online_cpu(cpu)
3662 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3663 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3666 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3667 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3668 atomic_long_add(events[i], &mi->vmevents[i]);
3671 #ifdef CONFIG_MEMCG_KMEM
3672 static int memcg_online_kmem(struct mem_cgroup *memcg)
3674 struct obj_cgroup *objcg;
3677 if (cgroup_memory_nokmem)
3680 BUG_ON(memcg->kmemcg_id >= 0);
3681 BUG_ON(memcg->kmem_state);
3683 memcg_id = memcg_alloc_cache_id();
3687 objcg = obj_cgroup_alloc();
3689 memcg_free_cache_id(memcg_id);
3692 objcg->memcg = memcg;
3693 rcu_assign_pointer(memcg->objcg, objcg);
3695 static_branch_enable(&memcg_kmem_enabled_key);
3698 * A memory cgroup is considered kmem-online as soon as it gets
3699 * kmemcg_id. Setting the id after enabling static branching will
3700 * guarantee no one starts accounting before all call sites are
3703 memcg->kmemcg_id = memcg_id;
3704 memcg->kmem_state = KMEM_ONLINE;
3709 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3711 struct cgroup_subsys_state *css;
3712 struct mem_cgroup *parent, *child;
3715 if (memcg->kmem_state != KMEM_ONLINE)
3718 memcg->kmem_state = KMEM_ALLOCATED;
3720 parent = parent_mem_cgroup(memcg);
3722 parent = root_mem_cgroup;
3724 memcg_reparent_objcgs(memcg, parent);
3726 kmemcg_id = memcg->kmemcg_id;
3727 BUG_ON(kmemcg_id < 0);
3730 * Change kmemcg_id of this cgroup and all its descendants to the
3731 * parent's id, and then move all entries from this cgroup's list_lrus
3732 * to ones of the parent. After we have finished, all list_lrus
3733 * corresponding to this cgroup are guaranteed to remain empty. The
3734 * ordering is imposed by list_lru_node->lock taken by
3735 * memcg_drain_all_list_lrus().
3737 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3738 css_for_each_descendant_pre(css, &memcg->css) {
3739 child = mem_cgroup_from_css(css);
3740 BUG_ON(child->kmemcg_id != kmemcg_id);
3741 child->kmemcg_id = parent->kmemcg_id;
3742 if (!memcg->use_hierarchy)
3747 memcg_drain_all_list_lrus(kmemcg_id, parent);
3749 memcg_free_cache_id(kmemcg_id);
3752 static void memcg_free_kmem(struct mem_cgroup *memcg)
3754 /* css_alloc() failed, offlining didn't happen */
3755 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3756 memcg_offline_kmem(memcg);
3759 static int memcg_online_kmem(struct mem_cgroup *memcg)
3763 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3766 static void memcg_free_kmem(struct mem_cgroup *memcg)
3769 #endif /* CONFIG_MEMCG_KMEM */
3771 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3776 mutex_lock(&memcg_max_mutex);
3777 ret = page_counter_set_max(&memcg->kmem, max);
3778 mutex_unlock(&memcg_max_mutex);
3782 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3786 mutex_lock(&memcg_max_mutex);
3788 ret = page_counter_set_max(&memcg->tcpmem, max);
3792 if (!memcg->tcpmem_active) {
3794 * The active flag needs to be written after the static_key
3795 * update. This is what guarantees that the socket activation
3796 * function is the last one to run. See mem_cgroup_sk_alloc()
3797 * for details, and note that we don't mark any socket as
3798 * belonging to this memcg until that flag is up.
3800 * We need to do this, because static_keys will span multiple
3801 * sites, but we can't control their order. If we mark a socket
3802 * as accounted, but the accounting functions are not patched in
3803 * yet, we'll lose accounting.
3805 * We never race with the readers in mem_cgroup_sk_alloc(),
3806 * because when this value change, the code to process it is not
3809 static_branch_inc(&memcg_sockets_enabled_key);
3810 memcg->tcpmem_active = true;
3813 mutex_unlock(&memcg_max_mutex);
3818 * The user of this function is...
3821 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3822 char *buf, size_t nbytes, loff_t off)
3824 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3825 unsigned long nr_pages;
3828 buf = strstrip(buf);
3829 ret = page_counter_memparse(buf, "-1", &nr_pages);
3833 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3835 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3839 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3841 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3844 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3847 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3848 "Please report your usecase to linux-mm@kvack.org if you "
3849 "depend on this functionality.\n");
3850 ret = memcg_update_kmem_max(memcg, nr_pages);
3853 ret = memcg_update_tcp_max(memcg, nr_pages);
3857 case RES_SOFT_LIMIT:
3858 memcg->soft_limit = nr_pages;
3862 return ret ?: nbytes;
3865 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3866 size_t nbytes, loff_t off)
3868 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3869 struct page_counter *counter;
3871 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3873 counter = &memcg->memory;
3876 counter = &memcg->memsw;
3879 counter = &memcg->kmem;
3882 counter = &memcg->tcpmem;
3888 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3890 page_counter_reset_watermark(counter);
3893 counter->failcnt = 0;
3902 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3905 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3909 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3910 struct cftype *cft, u64 val)
3912 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3914 if (val & ~MOVE_MASK)
3918 * No kind of locking is needed in here, because ->can_attach() will
3919 * check this value once in the beginning of the process, and then carry
3920 * on with stale data. This means that changes to this value will only
3921 * affect task migrations starting after the change.
3923 memcg->move_charge_at_immigrate = val;
3927 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3928 struct cftype *cft, u64 val)
3936 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3937 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3938 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3940 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3941 int nid, unsigned int lru_mask, bool tree)
3943 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3944 unsigned long nr = 0;
3947 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3950 if (!(BIT(lru) & lru_mask))
3953 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3955 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3960 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3961 unsigned int lru_mask,
3964 unsigned long nr = 0;
3968 if (!(BIT(lru) & lru_mask))
3971 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3973 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3978 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3982 unsigned int lru_mask;
3985 static const struct numa_stat stats[] = {
3986 { "total", LRU_ALL },
3987 { "file", LRU_ALL_FILE },
3988 { "anon", LRU_ALL_ANON },
3989 { "unevictable", BIT(LRU_UNEVICTABLE) },
3991 const struct numa_stat *stat;
3993 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3995 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3996 seq_printf(m, "%s=%lu", stat->name,
3997 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3999 for_each_node_state(nid, N_MEMORY)
4000 seq_printf(m, " N%d=%lu", nid,
4001 mem_cgroup_node_nr_lru_pages(memcg, nid,
4002 stat->lru_mask, false));
4006 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4008 seq_printf(m, "hierarchical_%s=%lu", stat->name,
4009 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4011 for_each_node_state(nid, N_MEMORY)
4012 seq_printf(m, " N%d=%lu", nid,
4013 mem_cgroup_node_nr_lru_pages(memcg, nid,
4014 stat->lru_mask, true));
4020 #endif /* CONFIG_NUMA */
4022 static const unsigned int memcg1_stats[] = {
4025 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4035 static const char *const memcg1_stat_names[] = {
4038 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4048 /* Universal VM events cgroup1 shows, original sort order */
4049 static const unsigned int memcg1_events[] = {
4056 static int memcg_stat_show(struct seq_file *m, void *v)
4058 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4059 unsigned long memory, memsw;
4060 struct mem_cgroup *mi;
4063 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4065 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4068 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4070 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4071 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4072 if (memcg1_stats[i] == NR_ANON_THPS)
4075 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4078 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4079 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4080 memcg_events_local(memcg, memcg1_events[i]));
4082 for (i = 0; i < NR_LRU_LISTS; i++)
4083 seq_printf(m, "%s %lu\n", lru_list_name(i),
4084 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4087 /* Hierarchical information */
4088 memory = memsw = PAGE_COUNTER_MAX;
4089 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4090 memory = min(memory, READ_ONCE(mi->memory.max));
4091 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4093 seq_printf(m, "hierarchical_memory_limit %llu\n",
4094 (u64)memory * PAGE_SIZE);
4095 if (do_memsw_account())
4096 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4097 (u64)memsw * PAGE_SIZE);
4099 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4102 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4104 nr = memcg_page_state(memcg, memcg1_stats[i]);
4105 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4106 if (memcg1_stats[i] == NR_ANON_THPS)
4109 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4110 (u64)nr * PAGE_SIZE);
4113 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4114 seq_printf(m, "total_%s %llu\n",
4115 vm_event_name(memcg1_events[i]),
4116 (u64)memcg_events(memcg, memcg1_events[i]));
4118 for (i = 0; i < NR_LRU_LISTS; i++)
4119 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4120 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4123 #ifdef CONFIG_DEBUG_VM
4126 struct mem_cgroup_per_node *mz;
4127 unsigned long anon_cost = 0;
4128 unsigned long file_cost = 0;
4130 for_each_online_pgdat(pgdat) {
4131 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
4133 anon_cost += mz->lruvec.anon_cost;
4134 file_cost += mz->lruvec.file_cost;
4136 seq_printf(m, "anon_cost %lu\n", anon_cost);
4137 seq_printf(m, "file_cost %lu\n", file_cost);
4144 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4147 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4149 return mem_cgroup_swappiness(memcg);
4152 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4153 struct cftype *cft, u64 val)
4155 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4161 memcg->swappiness = val;
4163 vm_swappiness = val;
4168 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4170 struct mem_cgroup_threshold_ary *t;
4171 unsigned long usage;
4176 t = rcu_dereference(memcg->thresholds.primary);
4178 t = rcu_dereference(memcg->memsw_thresholds.primary);
4183 usage = mem_cgroup_usage(memcg, swap);
4186 * current_threshold points to threshold just below or equal to usage.
4187 * If it's not true, a threshold was crossed after last
4188 * call of __mem_cgroup_threshold().
4190 i = t->current_threshold;
4193 * Iterate backward over array of thresholds starting from
4194 * current_threshold and check if a threshold is crossed.
4195 * If none of thresholds below usage is crossed, we read
4196 * only one element of the array here.
4198 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4199 eventfd_signal(t->entries[i].eventfd, 1);
4201 /* i = current_threshold + 1 */
4205 * Iterate forward over array of thresholds starting from
4206 * current_threshold+1 and check if a threshold is crossed.
4207 * If none of thresholds above usage is crossed, we read
4208 * only one element of the array here.
4210 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4211 eventfd_signal(t->entries[i].eventfd, 1);
4213 /* Update current_threshold */
4214 t->current_threshold = i - 1;
4219 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4222 __mem_cgroup_threshold(memcg, false);
4223 if (do_memsw_account())
4224 __mem_cgroup_threshold(memcg, true);
4226 memcg = parent_mem_cgroup(memcg);
4230 static int compare_thresholds(const void *a, const void *b)
4232 const struct mem_cgroup_threshold *_a = a;
4233 const struct mem_cgroup_threshold *_b = b;
4235 if (_a->threshold > _b->threshold)
4238 if (_a->threshold < _b->threshold)
4244 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4246 struct mem_cgroup_eventfd_list *ev;
4248 spin_lock(&memcg_oom_lock);
4250 list_for_each_entry(ev, &memcg->oom_notify, list)
4251 eventfd_signal(ev->eventfd, 1);
4253 spin_unlock(&memcg_oom_lock);
4257 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4259 struct mem_cgroup *iter;
4261 for_each_mem_cgroup_tree(iter, memcg)
4262 mem_cgroup_oom_notify_cb(iter);
4265 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4266 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4268 struct mem_cgroup_thresholds *thresholds;
4269 struct mem_cgroup_threshold_ary *new;
4270 unsigned long threshold;
4271 unsigned long usage;
4274 ret = page_counter_memparse(args, "-1", &threshold);
4278 mutex_lock(&memcg->thresholds_lock);
4281 thresholds = &memcg->thresholds;
4282 usage = mem_cgroup_usage(memcg, false);
4283 } else if (type == _MEMSWAP) {
4284 thresholds = &memcg->memsw_thresholds;
4285 usage = mem_cgroup_usage(memcg, true);
4289 /* Check if a threshold crossed before adding a new one */
4290 if (thresholds->primary)
4291 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4293 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4295 /* Allocate memory for new array of thresholds */
4296 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4303 /* Copy thresholds (if any) to new array */
4304 if (thresholds->primary)
4305 memcpy(new->entries, thresholds->primary->entries,
4306 flex_array_size(new, entries, size - 1));
4308 /* Add new threshold */
4309 new->entries[size - 1].eventfd = eventfd;
4310 new->entries[size - 1].threshold = threshold;
4312 /* Sort thresholds. Registering of new threshold isn't time-critical */
4313 sort(new->entries, size, sizeof(*new->entries),
4314 compare_thresholds, NULL);
4316 /* Find current threshold */
4317 new->current_threshold = -1;
4318 for (i = 0; i < size; i++) {
4319 if (new->entries[i].threshold <= usage) {
4321 * new->current_threshold will not be used until
4322 * rcu_assign_pointer(), so it's safe to increment
4325 ++new->current_threshold;
4330 /* Free old spare buffer and save old primary buffer as spare */
4331 kfree(thresholds->spare);
4332 thresholds->spare = thresholds->primary;
4334 rcu_assign_pointer(thresholds->primary, new);
4336 /* To be sure that nobody uses thresholds */
4340 mutex_unlock(&memcg->thresholds_lock);
4345 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4346 struct eventfd_ctx *eventfd, const char *args)
4348 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4351 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4352 struct eventfd_ctx *eventfd, const char *args)
4354 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4357 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4358 struct eventfd_ctx *eventfd, enum res_type type)
4360 struct mem_cgroup_thresholds *thresholds;
4361 struct mem_cgroup_threshold_ary *new;
4362 unsigned long usage;
4363 int i, j, size, entries;
4365 mutex_lock(&memcg->thresholds_lock);
4368 thresholds = &memcg->thresholds;
4369 usage = mem_cgroup_usage(memcg, false);
4370 } else if (type == _MEMSWAP) {
4371 thresholds = &memcg->memsw_thresholds;
4372 usage = mem_cgroup_usage(memcg, true);
4376 if (!thresholds->primary)
4379 /* Check if a threshold crossed before removing */
4380 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4382 /* Calculate new number of threshold */
4384 for (i = 0; i < thresholds->primary->size; i++) {
4385 if (thresholds->primary->entries[i].eventfd != eventfd)
4391 new = thresholds->spare;
4393 /* If no items related to eventfd have been cleared, nothing to do */
4397 /* Set thresholds array to NULL if we don't have thresholds */
4406 /* Copy thresholds and find current threshold */
4407 new->current_threshold = -1;
4408 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4409 if (thresholds->primary->entries[i].eventfd == eventfd)
4412 new->entries[j] = thresholds->primary->entries[i];
4413 if (new->entries[j].threshold <= usage) {
4415 * new->current_threshold will not be used
4416 * until rcu_assign_pointer(), so it's safe to increment
4419 ++new->current_threshold;
4425 /* Swap primary and spare array */
4426 thresholds->spare = thresholds->primary;
4428 rcu_assign_pointer(thresholds->primary, new);
4430 /* To be sure that nobody uses thresholds */
4433 /* If all events are unregistered, free the spare array */
4435 kfree(thresholds->spare);
4436 thresholds->spare = NULL;
4439 mutex_unlock(&memcg->thresholds_lock);
4442 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4443 struct eventfd_ctx *eventfd)
4445 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4448 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4449 struct eventfd_ctx *eventfd)
4451 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4454 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4455 struct eventfd_ctx *eventfd, const char *args)
4457 struct mem_cgroup_eventfd_list *event;
4459 event = kmalloc(sizeof(*event), GFP_KERNEL);
4463 spin_lock(&memcg_oom_lock);
4465 event->eventfd = eventfd;
4466 list_add(&event->list, &memcg->oom_notify);
4468 /* already in OOM ? */
4469 if (memcg->under_oom)
4470 eventfd_signal(eventfd, 1);
4471 spin_unlock(&memcg_oom_lock);
4476 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4477 struct eventfd_ctx *eventfd)
4479 struct mem_cgroup_eventfd_list *ev, *tmp;
4481 spin_lock(&memcg_oom_lock);
4483 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4484 if (ev->eventfd == eventfd) {
4485 list_del(&ev->list);
4490 spin_unlock(&memcg_oom_lock);
4493 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4495 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4497 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4498 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4499 seq_printf(sf, "oom_kill %lu\n",
4500 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4504 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4505 struct cftype *cft, u64 val)
4507 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4509 /* cannot set to root cgroup and only 0 and 1 are allowed */
4510 if (!css->parent || !((val == 0) || (val == 1)))
4513 memcg->oom_kill_disable = val;
4515 memcg_oom_recover(memcg);
4520 #ifdef CONFIG_CGROUP_WRITEBACK
4522 #include <trace/events/writeback.h>
4524 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4526 return wb_domain_init(&memcg->cgwb_domain, gfp);
4529 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4531 wb_domain_exit(&memcg->cgwb_domain);
4534 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4536 wb_domain_size_changed(&memcg->cgwb_domain);
4539 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4541 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4543 if (!memcg->css.parent)
4546 return &memcg->cgwb_domain;
4550 * idx can be of type enum memcg_stat_item or node_stat_item.
4551 * Keep in sync with memcg_exact_page().
4553 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4555 long x = atomic_long_read(&memcg->vmstats[idx]);
4558 for_each_online_cpu(cpu)
4559 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4566 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4567 * @wb: bdi_writeback in question
4568 * @pfilepages: out parameter for number of file pages
4569 * @pheadroom: out parameter for number of allocatable pages according to memcg
4570 * @pdirty: out parameter for number of dirty pages
4571 * @pwriteback: out parameter for number of pages under writeback
4573 * Determine the numbers of file, headroom, dirty, and writeback pages in
4574 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4575 * is a bit more involved.
4577 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4578 * headroom is calculated as the lowest headroom of itself and the
4579 * ancestors. Note that this doesn't consider the actual amount of
4580 * available memory in the system. The caller should further cap
4581 * *@pheadroom accordingly.
4583 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4584 unsigned long *pheadroom, unsigned long *pdirty,
4585 unsigned long *pwriteback)
4587 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4588 struct mem_cgroup *parent;
4590 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4592 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4593 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4594 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4595 *pheadroom = PAGE_COUNTER_MAX;
4597 while ((parent = parent_mem_cgroup(memcg))) {
4598 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4599 READ_ONCE(memcg->memory.high));
4600 unsigned long used = page_counter_read(&memcg->memory);
4602 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4608 * Foreign dirty flushing
4610 * There's an inherent mismatch between memcg and writeback. The former
4611 * trackes ownership per-page while the latter per-inode. This was a
4612 * deliberate design decision because honoring per-page ownership in the
4613 * writeback path is complicated, may lead to higher CPU and IO overheads
4614 * and deemed unnecessary given that write-sharing an inode across
4615 * different cgroups isn't a common use-case.
4617 * Combined with inode majority-writer ownership switching, this works well
4618 * enough in most cases but there are some pathological cases. For
4619 * example, let's say there are two cgroups A and B which keep writing to
4620 * different but confined parts of the same inode. B owns the inode and
4621 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4622 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4623 * triggering background writeback. A will be slowed down without a way to
4624 * make writeback of the dirty pages happen.
4626 * Conditions like the above can lead to a cgroup getting repatedly and
4627 * severely throttled after making some progress after each
4628 * dirty_expire_interval while the underyling IO device is almost
4631 * Solving this problem completely requires matching the ownership tracking
4632 * granularities between memcg and writeback in either direction. However,
4633 * the more egregious behaviors can be avoided by simply remembering the
4634 * most recent foreign dirtying events and initiating remote flushes on
4635 * them when local writeback isn't enough to keep the memory clean enough.
4637 * The following two functions implement such mechanism. When a foreign
4638 * page - a page whose memcg and writeback ownerships don't match - is
4639 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4640 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4641 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4642 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4643 * foreign bdi_writebacks which haven't expired. Both the numbers of
4644 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4645 * limited to MEMCG_CGWB_FRN_CNT.
4647 * The mechanism only remembers IDs and doesn't hold any object references.
4648 * As being wrong occasionally doesn't matter, updates and accesses to the
4649 * records are lockless and racy.
4651 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4652 struct bdi_writeback *wb)
4654 struct mem_cgroup *memcg = page_memcg(page);
4655 struct memcg_cgwb_frn *frn;
4656 u64 now = get_jiffies_64();
4657 u64 oldest_at = now;
4661 trace_track_foreign_dirty(page, wb);
4664 * Pick the slot to use. If there is already a slot for @wb, keep
4665 * using it. If not replace the oldest one which isn't being
4668 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4669 frn = &memcg->cgwb_frn[i];
4670 if (frn->bdi_id == wb->bdi->id &&
4671 frn->memcg_id == wb->memcg_css->id)
4673 if (time_before64(frn->at, oldest_at) &&
4674 atomic_read(&frn->done.cnt) == 1) {
4676 oldest_at = frn->at;
4680 if (i < MEMCG_CGWB_FRN_CNT) {
4682 * Re-using an existing one. Update timestamp lazily to
4683 * avoid making the cacheline hot. We want them to be
4684 * reasonably up-to-date and significantly shorter than
4685 * dirty_expire_interval as that's what expires the record.
4686 * Use the shorter of 1s and dirty_expire_interval / 8.
4688 unsigned long update_intv =
4689 min_t(unsigned long, HZ,
4690 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4692 if (time_before64(frn->at, now - update_intv))
4694 } else if (oldest >= 0) {
4695 /* replace the oldest free one */
4696 frn = &memcg->cgwb_frn[oldest];
4697 frn->bdi_id = wb->bdi->id;
4698 frn->memcg_id = wb->memcg_css->id;
4703 /* issue foreign writeback flushes for recorded foreign dirtying events */
4704 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4706 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4707 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4708 u64 now = jiffies_64;
4711 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4712 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4715 * If the record is older than dirty_expire_interval,
4716 * writeback on it has already started. No need to kick it
4717 * off again. Also, don't start a new one if there's
4718 * already one in flight.
4720 if (time_after64(frn->at, now - intv) &&
4721 atomic_read(&frn->done.cnt) == 1) {
4723 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4724 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4725 WB_REASON_FOREIGN_FLUSH,
4731 #else /* CONFIG_CGROUP_WRITEBACK */
4733 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4738 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4742 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4746 #endif /* CONFIG_CGROUP_WRITEBACK */
4749 * DO NOT USE IN NEW FILES.
4751 * "cgroup.event_control" implementation.
4753 * This is way over-engineered. It tries to support fully configurable
4754 * events for each user. Such level of flexibility is completely
4755 * unnecessary especially in the light of the planned unified hierarchy.
4757 * Please deprecate this and replace with something simpler if at all
4762 * Unregister event and free resources.
4764 * Gets called from workqueue.
4766 static void memcg_event_remove(struct work_struct *work)
4768 struct mem_cgroup_event *event =
4769 container_of(work, struct mem_cgroup_event, remove);
4770 struct mem_cgroup *memcg = event->memcg;
4772 remove_wait_queue(event->wqh, &event->wait);
4774 event->unregister_event(memcg, event->eventfd);
4776 /* Notify userspace the event is going away. */
4777 eventfd_signal(event->eventfd, 1);
4779 eventfd_ctx_put(event->eventfd);
4781 css_put(&memcg->css);
4785 * Gets called on EPOLLHUP on eventfd when user closes it.
4787 * Called with wqh->lock held and interrupts disabled.
4789 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4790 int sync, void *key)
4792 struct mem_cgroup_event *event =
4793 container_of(wait, struct mem_cgroup_event, wait);
4794 struct mem_cgroup *memcg = event->memcg;
4795 __poll_t flags = key_to_poll(key);
4797 if (flags & EPOLLHUP) {
4799 * If the event has been detached at cgroup removal, we
4800 * can simply return knowing the other side will cleanup
4803 * We can't race against event freeing since the other
4804 * side will require wqh->lock via remove_wait_queue(),
4807 spin_lock(&memcg->event_list_lock);
4808 if (!list_empty(&event->list)) {
4809 list_del_init(&event->list);
4811 * We are in atomic context, but cgroup_event_remove()
4812 * may sleep, so we have to call it in workqueue.
4814 schedule_work(&event->remove);
4816 spin_unlock(&memcg->event_list_lock);
4822 static void memcg_event_ptable_queue_proc(struct file *file,
4823 wait_queue_head_t *wqh, poll_table *pt)
4825 struct mem_cgroup_event *event =
4826 container_of(pt, struct mem_cgroup_event, pt);
4829 add_wait_queue(wqh, &event->wait);
4833 * DO NOT USE IN NEW FILES.
4835 * Parse input and register new cgroup event handler.
4837 * Input must be in format '<event_fd> <control_fd> <args>'.
4838 * Interpretation of args is defined by control file implementation.
4840 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4841 char *buf, size_t nbytes, loff_t off)
4843 struct cgroup_subsys_state *css = of_css(of);
4844 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4845 struct mem_cgroup_event *event;
4846 struct cgroup_subsys_state *cfile_css;
4847 unsigned int efd, cfd;
4854 buf = strstrip(buf);
4856 efd = simple_strtoul(buf, &endp, 10);
4861 cfd = simple_strtoul(buf, &endp, 10);
4862 if ((*endp != ' ') && (*endp != '\0'))
4866 event = kzalloc(sizeof(*event), GFP_KERNEL);
4870 event->memcg = memcg;
4871 INIT_LIST_HEAD(&event->list);
4872 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4873 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4874 INIT_WORK(&event->remove, memcg_event_remove);
4882 event->eventfd = eventfd_ctx_fileget(efile.file);
4883 if (IS_ERR(event->eventfd)) {
4884 ret = PTR_ERR(event->eventfd);
4891 goto out_put_eventfd;
4894 /* the process need read permission on control file */
4895 /* AV: shouldn't we check that it's been opened for read instead? */
4896 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4901 * Determine the event callbacks and set them in @event. This used
4902 * to be done via struct cftype but cgroup core no longer knows
4903 * about these events. The following is crude but the whole thing
4904 * is for compatibility anyway.
4906 * DO NOT ADD NEW FILES.
4908 name = cfile.file->f_path.dentry->d_name.name;
4910 if (!strcmp(name, "memory.usage_in_bytes")) {
4911 event->register_event = mem_cgroup_usage_register_event;
4912 event->unregister_event = mem_cgroup_usage_unregister_event;
4913 } else if (!strcmp(name, "memory.oom_control")) {
4914 event->register_event = mem_cgroup_oom_register_event;
4915 event->unregister_event = mem_cgroup_oom_unregister_event;
4916 } else if (!strcmp(name, "memory.pressure_level")) {
4917 event->register_event = vmpressure_register_event;
4918 event->unregister_event = vmpressure_unregister_event;
4919 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4920 event->register_event = memsw_cgroup_usage_register_event;
4921 event->unregister_event = memsw_cgroup_usage_unregister_event;
4928 * Verify @cfile should belong to @css. Also, remaining events are
4929 * automatically removed on cgroup destruction but the removal is
4930 * asynchronous, so take an extra ref on @css.
4932 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4933 &memory_cgrp_subsys);
4935 if (IS_ERR(cfile_css))
4937 if (cfile_css != css) {
4942 ret = event->register_event(memcg, event->eventfd, buf);
4946 vfs_poll(efile.file, &event->pt);
4948 spin_lock(&memcg->event_list_lock);
4949 list_add(&event->list, &memcg->event_list);
4950 spin_unlock(&memcg->event_list_lock);
4962 eventfd_ctx_put(event->eventfd);
4971 static struct cftype mem_cgroup_legacy_files[] = {
4973 .name = "usage_in_bytes",
4974 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4975 .read_u64 = mem_cgroup_read_u64,
4978 .name = "max_usage_in_bytes",
4979 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4980 .write = mem_cgroup_reset,
4981 .read_u64 = mem_cgroup_read_u64,
4984 .name = "limit_in_bytes",
4985 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4986 .write = mem_cgroup_write,
4987 .read_u64 = mem_cgroup_read_u64,
4990 .name = "soft_limit_in_bytes",
4991 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4992 .write = mem_cgroup_write,
4993 .read_u64 = mem_cgroup_read_u64,
4997 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4998 .write = mem_cgroup_reset,
4999 .read_u64 = mem_cgroup_read_u64,
5003 .seq_show = memcg_stat_show,
5006 .name = "force_empty",
5007 .write = mem_cgroup_force_empty_write,
5010 .name = "use_hierarchy",
5011 .write_u64 = mem_cgroup_hierarchy_write,
5012 .read_u64 = mem_cgroup_hierarchy_read,
5015 .name = "cgroup.event_control", /* XXX: for compat */
5016 .write = memcg_write_event_control,
5017 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
5020 .name = "swappiness",
5021 .read_u64 = mem_cgroup_swappiness_read,
5022 .write_u64 = mem_cgroup_swappiness_write,
5025 .name = "move_charge_at_immigrate",
5026 .read_u64 = mem_cgroup_move_charge_read,
5027 .write_u64 = mem_cgroup_move_charge_write,
5030 .name = "oom_control",
5031 .seq_show = mem_cgroup_oom_control_read,
5032 .write_u64 = mem_cgroup_oom_control_write,
5033 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5036 .name = "pressure_level",
5040 .name = "numa_stat",
5041 .seq_show = memcg_numa_stat_show,
5045 .name = "kmem.limit_in_bytes",
5046 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5047 .write = mem_cgroup_write,
5048 .read_u64 = mem_cgroup_read_u64,
5051 .name = "kmem.usage_in_bytes",
5052 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5053 .read_u64 = mem_cgroup_read_u64,
5056 .name = "kmem.failcnt",
5057 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5058 .write = mem_cgroup_reset,
5059 .read_u64 = mem_cgroup_read_u64,
5062 .name = "kmem.max_usage_in_bytes",
5063 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5064 .write = mem_cgroup_reset,
5065 .read_u64 = mem_cgroup_read_u64,
5067 #if defined(CONFIG_MEMCG_KMEM) && \
5068 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5070 .name = "kmem.slabinfo",
5071 .seq_show = memcg_slab_show,
5075 .name = "kmem.tcp.limit_in_bytes",
5076 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5077 .write = mem_cgroup_write,
5078 .read_u64 = mem_cgroup_read_u64,
5081 .name = "kmem.tcp.usage_in_bytes",
5082 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5083 .read_u64 = mem_cgroup_read_u64,
5086 .name = "kmem.tcp.failcnt",
5087 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5088 .write = mem_cgroup_reset,
5089 .read_u64 = mem_cgroup_read_u64,
5092 .name = "kmem.tcp.max_usage_in_bytes",
5093 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5094 .write = mem_cgroup_reset,
5095 .read_u64 = mem_cgroup_read_u64,
5097 { }, /* terminate */
5101 * Private memory cgroup IDR
5103 * Swap-out records and page cache shadow entries need to store memcg
5104 * references in constrained space, so we maintain an ID space that is
5105 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5106 * memory-controlled cgroups to 64k.
5108 * However, there usually are many references to the offline CSS after
5109 * the cgroup has been destroyed, such as page cache or reclaimable
5110 * slab objects, that don't need to hang on to the ID. We want to keep
5111 * those dead CSS from occupying IDs, or we might quickly exhaust the
5112 * relatively small ID space and prevent the creation of new cgroups
5113 * even when there are much fewer than 64k cgroups - possibly none.
5115 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5116 * be freed and recycled when it's no longer needed, which is usually
5117 * when the CSS is offlined.
5119 * The only exception to that are records of swapped out tmpfs/shmem
5120 * pages that need to be attributed to live ancestors on swapin. But
5121 * those references are manageable from userspace.
5124 static DEFINE_IDR(mem_cgroup_idr);
5126 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5128 if (memcg->id.id > 0) {
5129 idr_remove(&mem_cgroup_idr, memcg->id.id);
5134 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5137 refcount_add(n, &memcg->id.ref);
5140 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5142 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5143 mem_cgroup_id_remove(memcg);
5145 /* Memcg ID pins CSS */
5146 css_put(&memcg->css);
5150 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5152 mem_cgroup_id_put_many(memcg, 1);
5156 * mem_cgroup_from_id - look up a memcg from a memcg id
5157 * @id: the memcg id to look up
5159 * Caller must hold rcu_read_lock().
5161 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5163 WARN_ON_ONCE(!rcu_read_lock_held());
5164 return idr_find(&mem_cgroup_idr, id);
5167 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5169 struct mem_cgroup_per_node *pn;
5172 * This routine is called against possible nodes.
5173 * But it's BUG to call kmalloc() against offline node.
5175 * TODO: this routine can waste much memory for nodes which will
5176 * never be onlined. It's better to use memory hotplug callback
5179 if (!node_state(node, N_NORMAL_MEMORY))
5181 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5185 pn->lruvec_stat_local = alloc_percpu_gfp(struct lruvec_stat,
5186 GFP_KERNEL_ACCOUNT);
5187 if (!pn->lruvec_stat_local) {
5192 pn->lruvec_stat_cpu = alloc_percpu_gfp(struct lruvec_stat,
5193 GFP_KERNEL_ACCOUNT);
5194 if (!pn->lruvec_stat_cpu) {
5195 free_percpu(pn->lruvec_stat_local);
5200 lruvec_init(&pn->lruvec);
5201 pn->usage_in_excess = 0;
5202 pn->on_tree = false;
5205 memcg->nodeinfo[node] = pn;
5209 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5211 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5216 free_percpu(pn->lruvec_stat_cpu);
5217 free_percpu(pn->lruvec_stat_local);
5221 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5226 free_mem_cgroup_per_node_info(memcg, node);
5227 free_percpu(memcg->vmstats_percpu);
5228 free_percpu(memcg->vmstats_local);
5232 static void mem_cgroup_free(struct mem_cgroup *memcg)
5234 memcg_wb_domain_exit(memcg);
5236 * Flush percpu vmstats and vmevents to guarantee the value correctness
5237 * on parent's and all ancestor levels.
5239 memcg_flush_percpu_vmstats(memcg);
5240 memcg_flush_percpu_vmevents(memcg);
5241 __mem_cgroup_free(memcg);
5244 static struct mem_cgroup *mem_cgroup_alloc(void)
5246 struct mem_cgroup *memcg;
5249 int __maybe_unused i;
5250 long error = -ENOMEM;
5252 size = sizeof(struct mem_cgroup);
5253 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5255 memcg = kzalloc(size, GFP_KERNEL);
5257 return ERR_PTR(error);
5259 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5260 1, MEM_CGROUP_ID_MAX,
5262 if (memcg->id.id < 0) {
5263 error = memcg->id.id;
5267 memcg->vmstats_local = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5268 GFP_KERNEL_ACCOUNT);
5269 if (!memcg->vmstats_local)
5272 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5273 GFP_KERNEL_ACCOUNT);
5274 if (!memcg->vmstats_percpu)
5278 if (alloc_mem_cgroup_per_node_info(memcg, node))
5281 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5284 INIT_WORK(&memcg->high_work, high_work_func);
5285 INIT_LIST_HEAD(&memcg->oom_notify);
5286 mutex_init(&memcg->thresholds_lock);
5287 spin_lock_init(&memcg->move_lock);
5288 vmpressure_init(&memcg->vmpressure);
5289 INIT_LIST_HEAD(&memcg->event_list);
5290 spin_lock_init(&memcg->event_list_lock);
5291 memcg->socket_pressure = jiffies;
5292 #ifdef CONFIG_MEMCG_KMEM
5293 memcg->kmemcg_id = -1;
5294 INIT_LIST_HEAD(&memcg->objcg_list);
5296 #ifdef CONFIG_CGROUP_WRITEBACK
5297 INIT_LIST_HEAD(&memcg->cgwb_list);
5298 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5299 memcg->cgwb_frn[i].done =
5300 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5302 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5303 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5304 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5305 memcg->deferred_split_queue.split_queue_len = 0;
5307 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5310 mem_cgroup_id_remove(memcg);
5311 __mem_cgroup_free(memcg);
5312 return ERR_PTR(error);
5315 static struct cgroup_subsys_state * __ref
5316 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5318 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5319 struct mem_cgroup *memcg, *old_memcg;
5320 long error = -ENOMEM;
5322 old_memcg = set_active_memcg(parent);
5323 memcg = mem_cgroup_alloc();
5324 set_active_memcg(old_memcg);
5326 return ERR_CAST(memcg);
5328 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5329 memcg->soft_limit = PAGE_COUNTER_MAX;
5330 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5332 memcg->swappiness = mem_cgroup_swappiness(parent);
5333 memcg->oom_kill_disable = parent->oom_kill_disable;
5336 page_counter_init(&memcg->memory, NULL);
5337 page_counter_init(&memcg->swap, NULL);
5338 page_counter_init(&memcg->kmem, NULL);
5339 page_counter_init(&memcg->tcpmem, NULL);
5340 } else if (parent->use_hierarchy) {
5341 memcg->use_hierarchy = true;
5342 page_counter_init(&memcg->memory, &parent->memory);
5343 page_counter_init(&memcg->swap, &parent->swap);
5344 page_counter_init(&memcg->kmem, &parent->kmem);
5345 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5347 page_counter_init(&memcg->memory, &root_mem_cgroup->memory);
5348 page_counter_init(&memcg->swap, &root_mem_cgroup->swap);
5349 page_counter_init(&memcg->kmem, &root_mem_cgroup->kmem);
5350 page_counter_init(&memcg->tcpmem, &root_mem_cgroup->tcpmem);
5352 * Deeper hierachy with use_hierarchy == false doesn't make
5353 * much sense so let cgroup subsystem know about this
5354 * unfortunate state in our controller.
5356 if (parent != root_mem_cgroup)
5357 memory_cgrp_subsys.broken_hierarchy = true;
5360 /* The following stuff does not apply to the root */
5362 root_mem_cgroup = memcg;
5366 error = memcg_online_kmem(memcg);
5370 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5371 static_branch_inc(&memcg_sockets_enabled_key);
5375 mem_cgroup_id_remove(memcg);
5376 mem_cgroup_free(memcg);
5377 return ERR_PTR(error);
5380 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5382 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5385 * A memcg must be visible for memcg_expand_shrinker_maps()
5386 * by the time the maps are allocated. So, we allocate maps
5387 * here, when for_each_mem_cgroup() can't skip it.
5389 if (memcg_alloc_shrinker_maps(memcg)) {
5390 mem_cgroup_id_remove(memcg);
5394 /* Online state pins memcg ID, memcg ID pins CSS */
5395 refcount_set(&memcg->id.ref, 1);
5400 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5402 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5403 struct mem_cgroup_event *event, *tmp;
5406 * Unregister events and notify userspace.
5407 * Notify userspace about cgroup removing only after rmdir of cgroup
5408 * directory to avoid race between userspace and kernelspace.
5410 spin_lock(&memcg->event_list_lock);
5411 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5412 list_del_init(&event->list);
5413 schedule_work(&event->remove);
5415 spin_unlock(&memcg->event_list_lock);
5417 page_counter_set_min(&memcg->memory, 0);
5418 page_counter_set_low(&memcg->memory, 0);
5420 memcg_offline_kmem(memcg);
5421 wb_memcg_offline(memcg);
5423 drain_all_stock(memcg);
5425 mem_cgroup_id_put(memcg);
5428 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5430 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5432 invalidate_reclaim_iterators(memcg);
5435 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5437 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5438 int __maybe_unused i;
5440 #ifdef CONFIG_CGROUP_WRITEBACK
5441 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5442 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5444 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5445 static_branch_dec(&memcg_sockets_enabled_key);
5447 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5448 static_branch_dec(&memcg_sockets_enabled_key);
5450 vmpressure_cleanup(&memcg->vmpressure);
5451 cancel_work_sync(&memcg->high_work);
5452 mem_cgroup_remove_from_trees(memcg);
5453 memcg_free_shrinker_maps(memcg);
5454 memcg_free_kmem(memcg);
5455 mem_cgroup_free(memcg);
5459 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5460 * @css: the target css
5462 * Reset the states of the mem_cgroup associated with @css. This is
5463 * invoked when the userland requests disabling on the default hierarchy
5464 * but the memcg is pinned through dependency. The memcg should stop
5465 * applying policies and should revert to the vanilla state as it may be
5466 * made visible again.
5468 * The current implementation only resets the essential configurations.
5469 * This needs to be expanded to cover all the visible parts.
5471 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5473 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5475 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5476 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5477 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5478 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5479 page_counter_set_min(&memcg->memory, 0);
5480 page_counter_set_low(&memcg->memory, 0);
5481 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5482 memcg->soft_limit = PAGE_COUNTER_MAX;
5483 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5484 memcg_wb_domain_size_changed(memcg);
5488 /* Handlers for move charge at task migration. */
5489 static int mem_cgroup_do_precharge(unsigned long count)
5493 /* Try a single bulk charge without reclaim first, kswapd may wake */
5494 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5496 mc.precharge += count;
5500 /* Try charges one by one with reclaim, but do not retry */
5502 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5516 enum mc_target_type {
5523 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5524 unsigned long addr, pte_t ptent)
5526 struct page *page = vm_normal_page(vma, addr, ptent);
5528 if (!page || !page_mapped(page))
5530 if (PageAnon(page)) {
5531 if (!(mc.flags & MOVE_ANON))
5534 if (!(mc.flags & MOVE_FILE))
5537 if (!get_page_unless_zero(page))
5543 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5544 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5545 pte_t ptent, swp_entry_t *entry)
5547 struct page *page = NULL;
5548 swp_entry_t ent = pte_to_swp_entry(ptent);
5550 if (!(mc.flags & MOVE_ANON))
5554 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5555 * a device and because they are not accessible by CPU they are store
5556 * as special swap entry in the CPU page table.
5558 if (is_device_private_entry(ent)) {
5559 page = device_private_entry_to_page(ent);
5561 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5562 * a refcount of 1 when free (unlike normal page)
5564 if (!page_ref_add_unless(page, 1, 1))
5569 if (non_swap_entry(ent))
5573 * Because lookup_swap_cache() updates some statistics counter,
5574 * we call find_get_page() with swapper_space directly.
5576 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5577 entry->val = ent.val;
5582 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5583 pte_t ptent, swp_entry_t *entry)
5589 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5590 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5592 if (!vma->vm_file) /* anonymous vma */
5594 if (!(mc.flags & MOVE_FILE))
5597 /* page is moved even if it's not RSS of this task(page-faulted). */
5598 /* shmem/tmpfs may report page out on swap: account for that too. */
5599 return find_get_incore_page(vma->vm_file->f_mapping,
5600 linear_page_index(vma, addr));
5604 * mem_cgroup_move_account - move account of the page
5606 * @compound: charge the page as compound or small page
5607 * @from: mem_cgroup which the page is moved from.
5608 * @to: mem_cgroup which the page is moved to. @from != @to.
5610 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5612 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5615 static int mem_cgroup_move_account(struct page *page,
5617 struct mem_cgroup *from,
5618 struct mem_cgroup *to)
5620 struct lruvec *from_vec, *to_vec;
5621 struct pglist_data *pgdat;
5622 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5625 VM_BUG_ON(from == to);
5626 VM_BUG_ON_PAGE(PageLRU(page), page);
5627 VM_BUG_ON(compound && !PageTransHuge(page));
5630 * Prevent mem_cgroup_migrate() from looking at
5631 * page's memory cgroup of its source page while we change it.
5634 if (!trylock_page(page))
5638 if (page_memcg(page) != from)
5641 pgdat = page_pgdat(page);
5642 from_vec = mem_cgroup_lruvec(from, pgdat);
5643 to_vec = mem_cgroup_lruvec(to, pgdat);
5645 lock_page_memcg(page);
5647 if (PageAnon(page)) {
5648 if (page_mapped(page)) {
5649 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5650 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5651 if (PageTransHuge(page)) {
5652 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5654 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5660 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5661 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5663 if (PageSwapBacked(page)) {
5664 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5665 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5668 if (page_mapped(page)) {
5669 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5670 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5673 if (PageDirty(page)) {
5674 struct address_space *mapping = page_mapping(page);
5676 if (mapping_can_writeback(mapping)) {
5677 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5679 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5685 if (PageWriteback(page)) {
5686 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5687 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5691 * All state has been migrated, let's switch to the new memcg.
5693 * It is safe to change page's memcg here because the page
5694 * is referenced, charged, isolated, and locked: we can't race
5695 * with (un)charging, migration, LRU putback, or anything else
5696 * that would rely on a stable page's memory cgroup.
5698 * Note that lock_page_memcg is a memcg lock, not a page lock,
5699 * to save space. As soon as we switch page's memory cgroup to a
5700 * new memcg that isn't locked, the above state can change
5701 * concurrently again. Make sure we're truly done with it.
5706 css_put(&from->css);
5708 page->memcg_data = (unsigned long)to;
5710 __unlock_page_memcg(from);
5714 local_irq_disable();
5715 mem_cgroup_charge_statistics(to, page, nr_pages);
5716 memcg_check_events(to, page);
5717 mem_cgroup_charge_statistics(from, page, -nr_pages);
5718 memcg_check_events(from, page);
5727 * get_mctgt_type - get target type of moving charge
5728 * @vma: the vma the pte to be checked belongs
5729 * @addr: the address corresponding to the pte to be checked
5730 * @ptent: the pte to be checked
5731 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5734 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5735 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5736 * move charge. if @target is not NULL, the page is stored in target->page
5737 * with extra refcnt got(Callers should handle it).
5738 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5739 * target for charge migration. if @target is not NULL, the entry is stored
5741 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5742 * (so ZONE_DEVICE page and thus not on the lru).
5743 * For now we such page is charge like a regular page would be as for all
5744 * intent and purposes it is just special memory taking the place of a
5747 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5749 * Called with pte lock held.
5752 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5753 unsigned long addr, pte_t ptent, union mc_target *target)
5755 struct page *page = NULL;
5756 enum mc_target_type ret = MC_TARGET_NONE;
5757 swp_entry_t ent = { .val = 0 };
5759 if (pte_present(ptent))
5760 page = mc_handle_present_pte(vma, addr, ptent);
5761 else if (is_swap_pte(ptent))
5762 page = mc_handle_swap_pte(vma, ptent, &ent);
5763 else if (pte_none(ptent))
5764 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5766 if (!page && !ent.val)
5770 * Do only loose check w/o serialization.
5771 * mem_cgroup_move_account() checks the page is valid or
5772 * not under LRU exclusion.
5774 if (page_memcg(page) == mc.from) {
5775 ret = MC_TARGET_PAGE;
5776 if (is_device_private_page(page))
5777 ret = MC_TARGET_DEVICE;
5779 target->page = page;
5781 if (!ret || !target)
5785 * There is a swap entry and a page doesn't exist or isn't charged.
5786 * But we cannot move a tail-page in a THP.
5788 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5789 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5790 ret = MC_TARGET_SWAP;
5797 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5799 * We don't consider PMD mapped swapping or file mapped pages because THP does
5800 * not support them for now.
5801 * Caller should make sure that pmd_trans_huge(pmd) is true.
5803 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5804 unsigned long addr, pmd_t pmd, union mc_target *target)
5806 struct page *page = NULL;
5807 enum mc_target_type ret = MC_TARGET_NONE;
5809 if (unlikely(is_swap_pmd(pmd))) {
5810 VM_BUG_ON(thp_migration_supported() &&
5811 !is_pmd_migration_entry(pmd));
5814 page = pmd_page(pmd);
5815 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5816 if (!(mc.flags & MOVE_ANON))
5818 if (page_memcg(page) == mc.from) {
5819 ret = MC_TARGET_PAGE;
5822 target->page = page;
5828 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5829 unsigned long addr, pmd_t pmd, union mc_target *target)
5831 return MC_TARGET_NONE;
5835 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5836 unsigned long addr, unsigned long end,
5837 struct mm_walk *walk)
5839 struct vm_area_struct *vma = walk->vma;
5843 ptl = pmd_trans_huge_lock(pmd, vma);
5846 * Note their can not be MC_TARGET_DEVICE for now as we do not
5847 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5848 * this might change.
5850 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5851 mc.precharge += HPAGE_PMD_NR;
5856 if (pmd_trans_unstable(pmd))
5858 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5859 for (; addr != end; pte++, addr += PAGE_SIZE)
5860 if (get_mctgt_type(vma, addr, *pte, NULL))
5861 mc.precharge++; /* increment precharge temporarily */
5862 pte_unmap_unlock(pte - 1, ptl);
5868 static const struct mm_walk_ops precharge_walk_ops = {
5869 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5872 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5874 unsigned long precharge;
5877 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5878 mmap_read_unlock(mm);
5880 precharge = mc.precharge;
5886 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5888 unsigned long precharge = mem_cgroup_count_precharge(mm);
5890 VM_BUG_ON(mc.moving_task);
5891 mc.moving_task = current;
5892 return mem_cgroup_do_precharge(precharge);
5895 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5896 static void __mem_cgroup_clear_mc(void)
5898 struct mem_cgroup *from = mc.from;
5899 struct mem_cgroup *to = mc.to;
5901 /* we must uncharge all the leftover precharges from mc.to */
5903 cancel_charge(mc.to, mc.precharge);
5907 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5908 * we must uncharge here.
5910 if (mc.moved_charge) {
5911 cancel_charge(mc.from, mc.moved_charge);
5912 mc.moved_charge = 0;
5914 /* we must fixup refcnts and charges */
5915 if (mc.moved_swap) {
5916 /* uncharge swap account from the old cgroup */
5917 if (!mem_cgroup_is_root(mc.from))
5918 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5920 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5923 * we charged both to->memory and to->memsw, so we
5924 * should uncharge to->memory.
5926 if (!mem_cgroup_is_root(mc.to))
5927 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5931 memcg_oom_recover(from);
5932 memcg_oom_recover(to);
5933 wake_up_all(&mc.waitq);
5936 static void mem_cgroup_clear_mc(void)
5938 struct mm_struct *mm = mc.mm;
5941 * we must clear moving_task before waking up waiters at the end of
5944 mc.moving_task = NULL;
5945 __mem_cgroup_clear_mc();
5946 spin_lock(&mc.lock);
5950 spin_unlock(&mc.lock);
5955 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5957 struct cgroup_subsys_state *css;
5958 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5959 struct mem_cgroup *from;
5960 struct task_struct *leader, *p;
5961 struct mm_struct *mm;
5962 unsigned long move_flags;
5965 /* charge immigration isn't supported on the default hierarchy */
5966 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5970 * Multi-process migrations only happen on the default hierarchy
5971 * where charge immigration is not used. Perform charge
5972 * immigration if @tset contains a leader and whine if there are
5976 cgroup_taskset_for_each_leader(leader, css, tset) {
5979 memcg = mem_cgroup_from_css(css);
5985 * We are now commited to this value whatever it is. Changes in this
5986 * tunable will only affect upcoming migrations, not the current one.
5987 * So we need to save it, and keep it going.
5989 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5993 from = mem_cgroup_from_task(p);
5995 VM_BUG_ON(from == memcg);
5997 mm = get_task_mm(p);
6000 /* We move charges only when we move a owner of the mm */
6001 if (mm->owner == p) {
6004 VM_BUG_ON(mc.precharge);
6005 VM_BUG_ON(mc.moved_charge);
6006 VM_BUG_ON(mc.moved_swap);
6008 spin_lock(&mc.lock);
6012 mc.flags = move_flags;
6013 spin_unlock(&mc.lock);
6014 /* We set mc.moving_task later */
6016 ret = mem_cgroup_precharge_mc(mm);
6018 mem_cgroup_clear_mc();
6025 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6028 mem_cgroup_clear_mc();
6031 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6032 unsigned long addr, unsigned long end,
6033 struct mm_walk *walk)
6036 struct vm_area_struct *vma = walk->vma;
6039 enum mc_target_type target_type;
6040 union mc_target target;
6043 ptl = pmd_trans_huge_lock(pmd, vma);
6045 if (mc.precharge < HPAGE_PMD_NR) {
6049 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6050 if (target_type == MC_TARGET_PAGE) {
6052 if (!isolate_lru_page(page)) {
6053 if (!mem_cgroup_move_account(page, true,
6055 mc.precharge -= HPAGE_PMD_NR;
6056 mc.moved_charge += HPAGE_PMD_NR;
6058 putback_lru_page(page);
6061 } else if (target_type == MC_TARGET_DEVICE) {
6063 if (!mem_cgroup_move_account(page, true,
6065 mc.precharge -= HPAGE_PMD_NR;
6066 mc.moved_charge += HPAGE_PMD_NR;
6074 if (pmd_trans_unstable(pmd))
6077 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6078 for (; addr != end; addr += PAGE_SIZE) {
6079 pte_t ptent = *(pte++);
6080 bool device = false;
6086 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6087 case MC_TARGET_DEVICE:
6090 case MC_TARGET_PAGE:
6093 * We can have a part of the split pmd here. Moving it
6094 * can be done but it would be too convoluted so simply
6095 * ignore such a partial THP and keep it in original
6096 * memcg. There should be somebody mapping the head.
6098 if (PageTransCompound(page))
6100 if (!device && isolate_lru_page(page))
6102 if (!mem_cgroup_move_account(page, false,
6105 /* we uncharge from mc.from later. */
6109 putback_lru_page(page);
6110 put: /* get_mctgt_type() gets the page */
6113 case MC_TARGET_SWAP:
6115 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6117 mem_cgroup_id_get_many(mc.to, 1);
6118 /* we fixup other refcnts and charges later. */
6126 pte_unmap_unlock(pte - 1, ptl);
6131 * We have consumed all precharges we got in can_attach().
6132 * We try charge one by one, but don't do any additional
6133 * charges to mc.to if we have failed in charge once in attach()
6136 ret = mem_cgroup_do_precharge(1);
6144 static const struct mm_walk_ops charge_walk_ops = {
6145 .pmd_entry = mem_cgroup_move_charge_pte_range,
6148 static void mem_cgroup_move_charge(void)
6150 lru_add_drain_all();
6152 * Signal lock_page_memcg() to take the memcg's move_lock
6153 * while we're moving its pages to another memcg. Then wait
6154 * for already started RCU-only updates to finish.
6156 atomic_inc(&mc.from->moving_account);
6159 if (unlikely(!mmap_read_trylock(mc.mm))) {
6161 * Someone who are holding the mmap_lock might be waiting in
6162 * waitq. So we cancel all extra charges, wake up all waiters,
6163 * and retry. Because we cancel precharges, we might not be able
6164 * to move enough charges, but moving charge is a best-effort
6165 * feature anyway, so it wouldn't be a big problem.
6167 __mem_cgroup_clear_mc();
6172 * When we have consumed all precharges and failed in doing
6173 * additional charge, the page walk just aborts.
6175 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6178 mmap_read_unlock(mc.mm);
6179 atomic_dec(&mc.from->moving_account);
6182 static void mem_cgroup_move_task(void)
6185 mem_cgroup_move_charge();
6186 mem_cgroup_clear_mc();
6189 #else /* !CONFIG_MMU */
6190 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6194 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6197 static void mem_cgroup_move_task(void)
6203 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6204 * to verify whether we're attached to the default hierarchy on each mount
6207 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6210 * use_hierarchy is forced on the default hierarchy. cgroup core
6211 * guarantees that @root doesn't have any children, so turning it
6212 * on for the root memcg is enough.
6214 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6215 root_mem_cgroup->use_hierarchy = true;
6217 root_mem_cgroup->use_hierarchy = false;
6220 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6222 if (value == PAGE_COUNTER_MAX)
6223 seq_puts(m, "max\n");
6225 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6230 static u64 memory_current_read(struct cgroup_subsys_state *css,
6233 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6235 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6238 static int memory_min_show(struct seq_file *m, void *v)
6240 return seq_puts_memcg_tunable(m,
6241 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6244 static ssize_t memory_min_write(struct kernfs_open_file *of,
6245 char *buf, size_t nbytes, loff_t off)
6247 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6251 buf = strstrip(buf);
6252 err = page_counter_memparse(buf, "max", &min);
6256 page_counter_set_min(&memcg->memory, min);
6261 static int memory_low_show(struct seq_file *m, void *v)
6263 return seq_puts_memcg_tunable(m,
6264 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6267 static ssize_t memory_low_write(struct kernfs_open_file *of,
6268 char *buf, size_t nbytes, loff_t off)
6270 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6274 buf = strstrip(buf);
6275 err = page_counter_memparse(buf, "max", &low);
6279 page_counter_set_low(&memcg->memory, low);
6284 static int memory_high_show(struct seq_file *m, void *v)
6286 return seq_puts_memcg_tunable(m,
6287 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6290 static ssize_t memory_high_write(struct kernfs_open_file *of,
6291 char *buf, size_t nbytes, loff_t off)
6293 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6294 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6295 bool drained = false;
6299 buf = strstrip(buf);
6300 err = page_counter_memparse(buf, "max", &high);
6305 unsigned long nr_pages = page_counter_read(&memcg->memory);
6306 unsigned long reclaimed;
6308 if (nr_pages <= high)
6311 if (signal_pending(current))
6315 drain_all_stock(memcg);
6320 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6323 if (!reclaimed && !nr_retries--)
6327 page_counter_set_high(&memcg->memory, high);
6329 memcg_wb_domain_size_changed(memcg);
6334 static int memory_max_show(struct seq_file *m, void *v)
6336 return seq_puts_memcg_tunable(m,
6337 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6340 static ssize_t memory_max_write(struct kernfs_open_file *of,
6341 char *buf, size_t nbytes, loff_t off)
6343 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6344 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6345 bool drained = false;
6349 buf = strstrip(buf);
6350 err = page_counter_memparse(buf, "max", &max);
6354 xchg(&memcg->memory.max, max);
6357 unsigned long nr_pages = page_counter_read(&memcg->memory);
6359 if (nr_pages <= max)
6362 if (signal_pending(current))
6366 drain_all_stock(memcg);
6372 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6378 memcg_memory_event(memcg, MEMCG_OOM);
6379 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6383 memcg_wb_domain_size_changed(memcg);
6387 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6389 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6390 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6391 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6392 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6393 seq_printf(m, "oom_kill %lu\n",
6394 atomic_long_read(&events[MEMCG_OOM_KILL]));
6397 static int memory_events_show(struct seq_file *m, void *v)
6399 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6401 __memory_events_show(m, memcg->memory_events);
6405 static int memory_events_local_show(struct seq_file *m, void *v)
6407 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6409 __memory_events_show(m, memcg->memory_events_local);
6413 static int memory_stat_show(struct seq_file *m, void *v)
6415 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6418 buf = memory_stat_format(memcg);
6427 static int memory_numa_stat_show(struct seq_file *m, void *v)
6430 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6432 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6435 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6438 seq_printf(m, "%s", memory_stats[i].name);
6439 for_each_node_state(nid, N_MEMORY) {
6441 struct lruvec *lruvec;
6443 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6444 size = lruvec_page_state(lruvec, memory_stats[i].idx);
6445 size *= memory_stats[i].ratio;
6446 seq_printf(m, " N%d=%llu", nid, size);
6455 static int memory_oom_group_show(struct seq_file *m, void *v)
6457 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6459 seq_printf(m, "%d\n", memcg->oom_group);
6464 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6465 char *buf, size_t nbytes, loff_t off)
6467 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6470 buf = strstrip(buf);
6474 ret = kstrtoint(buf, 0, &oom_group);
6478 if (oom_group != 0 && oom_group != 1)
6481 memcg->oom_group = oom_group;
6486 static struct cftype memory_files[] = {
6489 .flags = CFTYPE_NOT_ON_ROOT,
6490 .read_u64 = memory_current_read,
6494 .flags = CFTYPE_NOT_ON_ROOT,
6495 .seq_show = memory_min_show,
6496 .write = memory_min_write,
6500 .flags = CFTYPE_NOT_ON_ROOT,
6501 .seq_show = memory_low_show,
6502 .write = memory_low_write,
6506 .flags = CFTYPE_NOT_ON_ROOT,
6507 .seq_show = memory_high_show,
6508 .write = memory_high_write,
6512 .flags = CFTYPE_NOT_ON_ROOT,
6513 .seq_show = memory_max_show,
6514 .write = memory_max_write,
6518 .flags = CFTYPE_NOT_ON_ROOT,
6519 .file_offset = offsetof(struct mem_cgroup, events_file),
6520 .seq_show = memory_events_show,
6523 .name = "events.local",
6524 .flags = CFTYPE_NOT_ON_ROOT,
6525 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6526 .seq_show = memory_events_local_show,
6530 .seq_show = memory_stat_show,
6534 .name = "numa_stat",
6535 .seq_show = memory_numa_stat_show,
6539 .name = "oom.group",
6540 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6541 .seq_show = memory_oom_group_show,
6542 .write = memory_oom_group_write,
6547 struct cgroup_subsys memory_cgrp_subsys = {
6548 .css_alloc = mem_cgroup_css_alloc,
6549 .css_online = mem_cgroup_css_online,
6550 .css_offline = mem_cgroup_css_offline,
6551 .css_released = mem_cgroup_css_released,
6552 .css_free = mem_cgroup_css_free,
6553 .css_reset = mem_cgroup_css_reset,
6554 .can_attach = mem_cgroup_can_attach,
6555 .cancel_attach = mem_cgroup_cancel_attach,
6556 .post_attach = mem_cgroup_move_task,
6557 .bind = mem_cgroup_bind,
6558 .dfl_cftypes = memory_files,
6559 .legacy_cftypes = mem_cgroup_legacy_files,
6564 * This function calculates an individual cgroup's effective
6565 * protection which is derived from its own memory.min/low, its
6566 * parent's and siblings' settings, as well as the actual memory
6567 * distribution in the tree.
6569 * The following rules apply to the effective protection values:
6571 * 1. At the first level of reclaim, effective protection is equal to
6572 * the declared protection in memory.min and memory.low.
6574 * 2. To enable safe delegation of the protection configuration, at
6575 * subsequent levels the effective protection is capped to the
6576 * parent's effective protection.
6578 * 3. To make complex and dynamic subtrees easier to configure, the
6579 * user is allowed to overcommit the declared protection at a given
6580 * level. If that is the case, the parent's effective protection is
6581 * distributed to the children in proportion to how much protection
6582 * they have declared and how much of it they are utilizing.
6584 * This makes distribution proportional, but also work-conserving:
6585 * if one cgroup claims much more protection than it uses memory,
6586 * the unused remainder is available to its siblings.
6588 * 4. Conversely, when the declared protection is undercommitted at a
6589 * given level, the distribution of the larger parental protection
6590 * budget is NOT proportional. A cgroup's protection from a sibling
6591 * is capped to its own memory.min/low setting.
6593 * 5. However, to allow protecting recursive subtrees from each other
6594 * without having to declare each individual cgroup's fixed share
6595 * of the ancestor's claim to protection, any unutilized -
6596 * "floating" - protection from up the tree is distributed in
6597 * proportion to each cgroup's *usage*. This makes the protection
6598 * neutral wrt sibling cgroups and lets them compete freely over
6599 * the shared parental protection budget, but it protects the
6600 * subtree as a whole from neighboring subtrees.
6602 * Note that 4. and 5. are not in conflict: 4. is about protecting
6603 * against immediate siblings whereas 5. is about protecting against
6604 * neighboring subtrees.
6606 static unsigned long effective_protection(unsigned long usage,
6607 unsigned long parent_usage,
6608 unsigned long setting,
6609 unsigned long parent_effective,
6610 unsigned long siblings_protected)
6612 unsigned long protected;
6615 protected = min(usage, setting);
6617 * If all cgroups at this level combined claim and use more
6618 * protection then what the parent affords them, distribute
6619 * shares in proportion to utilization.
6621 * We are using actual utilization rather than the statically
6622 * claimed protection in order to be work-conserving: claimed
6623 * but unused protection is available to siblings that would
6624 * otherwise get a smaller chunk than what they claimed.
6626 if (siblings_protected > parent_effective)
6627 return protected * parent_effective / siblings_protected;
6630 * Ok, utilized protection of all children is within what the
6631 * parent affords them, so we know whatever this child claims
6632 * and utilizes is effectively protected.
6634 * If there is unprotected usage beyond this value, reclaim
6635 * will apply pressure in proportion to that amount.
6637 * If there is unutilized protection, the cgroup will be fully
6638 * shielded from reclaim, but we do return a smaller value for
6639 * protection than what the group could enjoy in theory. This
6640 * is okay. With the overcommit distribution above, effective
6641 * protection is always dependent on how memory is actually
6642 * consumed among the siblings anyway.
6647 * If the children aren't claiming (all of) the protection
6648 * afforded to them by the parent, distribute the remainder in
6649 * proportion to the (unprotected) memory of each cgroup. That
6650 * way, cgroups that aren't explicitly prioritized wrt each
6651 * other compete freely over the allowance, but they are
6652 * collectively protected from neighboring trees.
6654 * We're using unprotected memory for the weight so that if
6655 * some cgroups DO claim explicit protection, we don't protect
6656 * the same bytes twice.
6658 * Check both usage and parent_usage against the respective
6659 * protected values. One should imply the other, but they
6660 * aren't read atomically - make sure the division is sane.
6662 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6664 if (parent_effective > siblings_protected &&
6665 parent_usage > siblings_protected &&
6666 usage > protected) {
6667 unsigned long unclaimed;
6669 unclaimed = parent_effective - siblings_protected;
6670 unclaimed *= usage - protected;
6671 unclaimed /= parent_usage - siblings_protected;
6680 * mem_cgroup_protected - check if memory consumption is in the normal range
6681 * @root: the top ancestor of the sub-tree being checked
6682 * @memcg: the memory cgroup to check
6684 * WARNING: This function is not stateless! It can only be used as part
6685 * of a top-down tree iteration, not for isolated queries.
6687 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6688 struct mem_cgroup *memcg)
6690 unsigned long usage, parent_usage;
6691 struct mem_cgroup *parent;
6693 if (mem_cgroup_disabled())
6697 root = root_mem_cgroup;
6700 * Effective values of the reclaim targets are ignored so they
6701 * can be stale. Have a look at mem_cgroup_protection for more
6703 * TODO: calculation should be more robust so that we do not need
6704 * that special casing.
6709 usage = page_counter_read(&memcg->memory);
6713 parent = parent_mem_cgroup(memcg);
6714 /* No parent means a non-hierarchical mode on v1 memcg */
6718 if (parent == root) {
6719 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6720 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6724 parent_usage = page_counter_read(&parent->memory);
6726 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6727 READ_ONCE(memcg->memory.min),
6728 READ_ONCE(parent->memory.emin),
6729 atomic_long_read(&parent->memory.children_min_usage)));
6731 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6732 READ_ONCE(memcg->memory.low),
6733 READ_ONCE(parent->memory.elow),
6734 atomic_long_read(&parent->memory.children_low_usage)));
6738 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6739 * @page: page to charge
6740 * @mm: mm context of the victim
6741 * @gfp_mask: reclaim mode
6743 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6744 * pages according to @gfp_mask if necessary.
6746 * Returns 0 on success. Otherwise, an error code is returned.
6748 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6750 unsigned int nr_pages = thp_nr_pages(page);
6751 struct mem_cgroup *memcg = NULL;
6754 if (mem_cgroup_disabled())
6757 if (PageSwapCache(page)) {
6758 swp_entry_t ent = { .val = page_private(page), };
6762 * Every swap fault against a single page tries to charge the
6763 * page, bail as early as possible. shmem_unuse() encounters
6764 * already charged pages, too. page and memcg binding is
6765 * protected by the page lock, which serializes swap cache
6766 * removal, which in turn serializes uncharging.
6768 VM_BUG_ON_PAGE(!PageLocked(page), page);
6769 if (page_memcg(compound_head(page)))
6772 id = lookup_swap_cgroup_id(ent);
6774 memcg = mem_cgroup_from_id(id);
6775 if (memcg && !css_tryget_online(&memcg->css))
6781 memcg = get_mem_cgroup_from_mm(mm);
6783 ret = try_charge(memcg, gfp_mask, nr_pages);
6787 css_get(&memcg->css);
6788 commit_charge(page, memcg);
6790 local_irq_disable();
6791 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6792 memcg_check_events(memcg, page);
6795 if (PageSwapCache(page)) {
6796 swp_entry_t entry = { .val = page_private(page) };
6798 * The swap entry might not get freed for a long time,
6799 * let's not wait for it. The page already received a
6800 * memory+swap charge, drop the swap entry duplicate.
6802 mem_cgroup_uncharge_swap(entry, nr_pages);
6806 css_put(&memcg->css);
6811 struct uncharge_gather {
6812 struct mem_cgroup *memcg;
6813 unsigned long nr_pages;
6814 unsigned long pgpgout;
6815 unsigned long nr_kmem;
6816 struct page *dummy_page;
6819 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6821 memset(ug, 0, sizeof(*ug));
6824 static void uncharge_batch(const struct uncharge_gather *ug)
6826 unsigned long flags;
6828 if (!mem_cgroup_is_root(ug->memcg)) {
6829 page_counter_uncharge(&ug->memcg->memory, ug->nr_pages);
6830 if (do_memsw_account())
6831 page_counter_uncharge(&ug->memcg->memsw, ug->nr_pages);
6832 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6833 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6834 memcg_oom_recover(ug->memcg);
6837 local_irq_save(flags);
6838 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6839 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_pages);
6840 memcg_check_events(ug->memcg, ug->dummy_page);
6841 local_irq_restore(flags);
6843 /* drop reference from uncharge_page */
6844 css_put(&ug->memcg->css);
6847 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6849 unsigned long nr_pages;
6851 VM_BUG_ON_PAGE(PageLRU(page), page);
6853 if (!page_memcg(page))
6857 * Nobody should be changing or seriously looking at
6858 * page_memcg(page) at this point, we have fully
6859 * exclusive access to the page.
6862 if (ug->memcg != page_memcg(page)) {
6865 uncharge_gather_clear(ug);
6867 ug->memcg = page_memcg(page);
6869 /* pairs with css_put in uncharge_batch */
6870 css_get(&ug->memcg->css);
6873 nr_pages = compound_nr(page);
6874 ug->nr_pages += nr_pages;
6876 if (PageMemcgKmem(page))
6877 ug->nr_kmem += nr_pages;
6881 ug->dummy_page = page;
6882 page->memcg_data = 0;
6883 css_put(&ug->memcg->css);
6886 static void uncharge_list(struct list_head *page_list)
6888 struct uncharge_gather ug;
6889 struct list_head *next;
6891 uncharge_gather_clear(&ug);
6894 * Note that the list can be a single page->lru; hence the
6895 * do-while loop instead of a simple list_for_each_entry().
6897 next = page_list->next;
6901 page = list_entry(next, struct page, lru);
6902 next = page->lru.next;
6904 uncharge_page(page, &ug);
6905 } while (next != page_list);
6908 uncharge_batch(&ug);
6912 * mem_cgroup_uncharge - uncharge a page
6913 * @page: page to uncharge
6915 * Uncharge a page previously charged with mem_cgroup_charge().
6917 void mem_cgroup_uncharge(struct page *page)
6919 struct uncharge_gather ug;
6921 if (mem_cgroup_disabled())
6924 /* Don't touch page->lru of any random page, pre-check: */
6925 if (!page_memcg(page))
6928 uncharge_gather_clear(&ug);
6929 uncharge_page(page, &ug);
6930 uncharge_batch(&ug);
6934 * mem_cgroup_uncharge_list - uncharge a list of page
6935 * @page_list: list of pages to uncharge
6937 * Uncharge a list of pages previously charged with
6938 * mem_cgroup_charge().
6940 void mem_cgroup_uncharge_list(struct list_head *page_list)
6942 if (mem_cgroup_disabled())
6945 if (!list_empty(page_list))
6946 uncharge_list(page_list);
6950 * mem_cgroup_migrate - charge a page's replacement
6951 * @oldpage: currently circulating page
6952 * @newpage: replacement page
6954 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6955 * be uncharged upon free.
6957 * Both pages must be locked, @newpage->mapping must be set up.
6959 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6961 struct mem_cgroup *memcg;
6962 unsigned int nr_pages;
6963 unsigned long flags;
6965 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6966 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6967 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6968 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6971 if (mem_cgroup_disabled())
6974 /* Page cache replacement: new page already charged? */
6975 if (page_memcg(newpage))
6978 /* Swapcache readahead pages can get replaced before being charged */
6979 memcg = page_memcg(oldpage);
6983 /* Force-charge the new page. The old one will be freed soon */
6984 nr_pages = thp_nr_pages(newpage);
6986 page_counter_charge(&memcg->memory, nr_pages);
6987 if (do_memsw_account())
6988 page_counter_charge(&memcg->memsw, nr_pages);
6990 css_get(&memcg->css);
6991 commit_charge(newpage, memcg);
6993 local_irq_save(flags);
6994 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
6995 memcg_check_events(memcg, newpage);
6996 local_irq_restore(flags);
6999 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
7000 EXPORT_SYMBOL(memcg_sockets_enabled_key);
7002 void mem_cgroup_sk_alloc(struct sock *sk)
7004 struct mem_cgroup *memcg;
7006 if (!mem_cgroup_sockets_enabled)
7009 /* Do not associate the sock with unrelated interrupted task's memcg. */
7014 memcg = mem_cgroup_from_task(current);
7015 if (memcg == root_mem_cgroup)
7017 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
7019 if (css_tryget(&memcg->css))
7020 sk->sk_memcg = memcg;
7025 void mem_cgroup_sk_free(struct sock *sk)
7028 css_put(&sk->sk_memcg->css);
7032 * mem_cgroup_charge_skmem - charge socket memory
7033 * @memcg: memcg to charge
7034 * @nr_pages: number of pages to charge
7036 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7037 * @memcg's configured limit, %false if the charge had to be forced.
7039 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7041 gfp_t gfp_mask = GFP_KERNEL;
7043 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7044 struct page_counter *fail;
7046 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7047 memcg->tcpmem_pressure = 0;
7050 page_counter_charge(&memcg->tcpmem, nr_pages);
7051 memcg->tcpmem_pressure = 1;
7055 /* Don't block in the packet receive path */
7057 gfp_mask = GFP_NOWAIT;
7059 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7061 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
7064 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
7069 * mem_cgroup_uncharge_skmem - uncharge socket memory
7070 * @memcg: memcg to uncharge
7071 * @nr_pages: number of pages to uncharge
7073 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7075 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7076 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7080 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7082 refill_stock(memcg, nr_pages);
7085 static int __init cgroup_memory(char *s)
7089 while ((token = strsep(&s, ",")) != NULL) {
7092 if (!strcmp(token, "nosocket"))
7093 cgroup_memory_nosocket = true;
7094 if (!strcmp(token, "nokmem"))
7095 cgroup_memory_nokmem = true;
7099 __setup("cgroup.memory=", cgroup_memory);
7102 * subsys_initcall() for memory controller.
7104 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7105 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7106 * basically everything that doesn't depend on a specific mem_cgroup structure
7107 * should be initialized from here.
7109 static int __init mem_cgroup_init(void)
7113 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7114 memcg_hotplug_cpu_dead);
7116 for_each_possible_cpu(cpu)
7117 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7120 for_each_node(node) {
7121 struct mem_cgroup_tree_per_node *rtpn;
7123 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7124 node_online(node) ? node : NUMA_NO_NODE);
7126 rtpn->rb_root = RB_ROOT;
7127 rtpn->rb_rightmost = NULL;
7128 spin_lock_init(&rtpn->lock);
7129 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7134 subsys_initcall(mem_cgroup_init);
7136 #ifdef CONFIG_MEMCG_SWAP
7137 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7139 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7141 * The root cgroup cannot be destroyed, so it's refcount must
7144 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7148 memcg = parent_mem_cgroup(memcg);
7150 memcg = root_mem_cgroup;
7156 * mem_cgroup_swapout - transfer a memsw charge to swap
7157 * @page: page whose memsw charge to transfer
7158 * @entry: swap entry to move the charge to
7160 * Transfer the memsw charge of @page to @entry.
7162 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7164 struct mem_cgroup *memcg, *swap_memcg;
7165 unsigned int nr_entries;
7166 unsigned short oldid;
7168 VM_BUG_ON_PAGE(PageLRU(page), page);
7169 VM_BUG_ON_PAGE(page_count(page), page);
7171 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7174 memcg = page_memcg(page);
7176 /* Readahead page, never charged */
7181 * In case the memcg owning these pages has been offlined and doesn't
7182 * have an ID allocated to it anymore, charge the closest online
7183 * ancestor for the swap instead and transfer the memory+swap charge.
7185 swap_memcg = mem_cgroup_id_get_online(memcg);
7186 nr_entries = thp_nr_pages(page);
7187 /* Get references for the tail pages, too */
7189 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7190 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7192 VM_BUG_ON_PAGE(oldid, page);
7193 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7195 page->memcg_data = 0;
7197 if (!mem_cgroup_is_root(memcg))
7198 page_counter_uncharge(&memcg->memory, nr_entries);
7200 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7201 if (!mem_cgroup_is_root(swap_memcg))
7202 page_counter_charge(&swap_memcg->memsw, nr_entries);
7203 page_counter_uncharge(&memcg->memsw, nr_entries);
7207 * Interrupts should be disabled here because the caller holds the
7208 * i_pages lock which is taken with interrupts-off. It is
7209 * important here to have the interrupts disabled because it is the
7210 * only synchronisation we have for updating the per-CPU variables.
7212 VM_BUG_ON(!irqs_disabled());
7213 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7214 memcg_check_events(memcg, page);
7216 css_put(&memcg->css);
7220 * mem_cgroup_try_charge_swap - try charging swap space for a page
7221 * @page: page being added to swap
7222 * @entry: swap entry to charge
7224 * Try to charge @page's memcg for the swap space at @entry.
7226 * Returns 0 on success, -ENOMEM on failure.
7228 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7230 unsigned int nr_pages = thp_nr_pages(page);
7231 struct page_counter *counter;
7232 struct mem_cgroup *memcg;
7233 unsigned short oldid;
7235 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7238 memcg = page_memcg(page);
7240 /* Readahead page, never charged */
7245 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7249 memcg = mem_cgroup_id_get_online(memcg);
7251 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7252 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7253 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7254 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7255 mem_cgroup_id_put(memcg);
7259 /* Get references for the tail pages, too */
7261 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7262 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7263 VM_BUG_ON_PAGE(oldid, page);
7264 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7270 * mem_cgroup_uncharge_swap - uncharge swap space
7271 * @entry: swap entry to uncharge
7272 * @nr_pages: the amount of swap space to uncharge
7274 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7276 struct mem_cgroup *memcg;
7279 id = swap_cgroup_record(entry, 0, nr_pages);
7281 memcg = mem_cgroup_from_id(id);
7283 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7284 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7285 page_counter_uncharge(&memcg->swap, nr_pages);
7287 page_counter_uncharge(&memcg->memsw, nr_pages);
7289 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7290 mem_cgroup_id_put_many(memcg, nr_pages);
7295 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7297 long nr_swap_pages = get_nr_swap_pages();
7299 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7300 return nr_swap_pages;
7301 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7302 nr_swap_pages = min_t(long, nr_swap_pages,
7303 READ_ONCE(memcg->swap.max) -
7304 page_counter_read(&memcg->swap));
7305 return nr_swap_pages;
7308 bool mem_cgroup_swap_full(struct page *page)
7310 struct mem_cgroup *memcg;
7312 VM_BUG_ON_PAGE(!PageLocked(page), page);
7316 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7319 memcg = page_memcg(page);
7323 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7324 unsigned long usage = page_counter_read(&memcg->swap);
7326 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7327 usage * 2 >= READ_ONCE(memcg->swap.max))
7334 static int __init setup_swap_account(char *s)
7336 if (!strcmp(s, "1"))
7337 cgroup_memory_noswap = 0;
7338 else if (!strcmp(s, "0"))
7339 cgroup_memory_noswap = 1;
7342 __setup("swapaccount=", setup_swap_account);
7344 static u64 swap_current_read(struct cgroup_subsys_state *css,
7347 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7349 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7352 static int swap_high_show(struct seq_file *m, void *v)
7354 return seq_puts_memcg_tunable(m,
7355 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7358 static ssize_t swap_high_write(struct kernfs_open_file *of,
7359 char *buf, size_t nbytes, loff_t off)
7361 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7365 buf = strstrip(buf);
7366 err = page_counter_memparse(buf, "max", &high);
7370 page_counter_set_high(&memcg->swap, high);
7375 static int swap_max_show(struct seq_file *m, void *v)
7377 return seq_puts_memcg_tunable(m,
7378 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7381 static ssize_t swap_max_write(struct kernfs_open_file *of,
7382 char *buf, size_t nbytes, loff_t off)
7384 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7388 buf = strstrip(buf);
7389 err = page_counter_memparse(buf, "max", &max);
7393 xchg(&memcg->swap.max, max);
7398 static int swap_events_show(struct seq_file *m, void *v)
7400 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7402 seq_printf(m, "high %lu\n",
7403 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7404 seq_printf(m, "max %lu\n",
7405 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7406 seq_printf(m, "fail %lu\n",
7407 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7412 static struct cftype swap_files[] = {
7414 .name = "swap.current",
7415 .flags = CFTYPE_NOT_ON_ROOT,
7416 .read_u64 = swap_current_read,
7419 .name = "swap.high",
7420 .flags = CFTYPE_NOT_ON_ROOT,
7421 .seq_show = swap_high_show,
7422 .write = swap_high_write,
7426 .flags = CFTYPE_NOT_ON_ROOT,
7427 .seq_show = swap_max_show,
7428 .write = swap_max_write,
7431 .name = "swap.events",
7432 .flags = CFTYPE_NOT_ON_ROOT,
7433 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7434 .seq_show = swap_events_show,
7439 static struct cftype memsw_files[] = {
7441 .name = "memsw.usage_in_bytes",
7442 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7443 .read_u64 = mem_cgroup_read_u64,
7446 .name = "memsw.max_usage_in_bytes",
7447 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7448 .write = mem_cgroup_reset,
7449 .read_u64 = mem_cgroup_read_u64,
7452 .name = "memsw.limit_in_bytes",
7453 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7454 .write = mem_cgroup_write,
7455 .read_u64 = mem_cgroup_read_u64,
7458 .name = "memsw.failcnt",
7459 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7460 .write = mem_cgroup_reset,
7461 .read_u64 = mem_cgroup_read_u64,
7463 { }, /* terminate */
7467 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7468 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7469 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7470 * boot parameter. This may result in premature OOPS inside
7471 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7473 static int __init mem_cgroup_swap_init(void)
7475 /* No memory control -> no swap control */
7476 if (mem_cgroup_disabled())
7477 cgroup_memory_noswap = true;
7479 if (cgroup_memory_noswap)
7482 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7483 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7487 core_initcall(mem_cgroup_swap_init);
7489 #endif /* CONFIG_MEMCG_SWAP */