1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* memcontrol.c - Memory Controller
4 * Copyright IBM Corporation, 2007
5 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
7 * Copyright 2007 OpenVZ SWsoft Inc
8 * Author: Pavel Emelianov <xemul@openvz.org>
11 * Copyright (C) 2009 Nokia Corporation
12 * Author: Kirill A. Shutemov
14 * Kernel Memory Controller
15 * Copyright (C) 2012 Parallels Inc. and Google Inc.
16 * Authors: Glauber Costa and Suleiman Souhlal
19 * Charge lifetime sanitation
20 * Lockless page tracking & accounting
21 * Unified hierarchy configuration model
22 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
25 #include <linux/page_counter.h>
26 #include <linux/memcontrol.h>
27 #include <linux/cgroup.h>
28 #include <linux/pagewalk.h>
29 #include <linux/sched/mm.h>
30 #include <linux/shmem_fs.h>
31 #include <linux/hugetlb.h>
32 #include <linux/pagemap.h>
33 #include <linux/vm_event_item.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/swap_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
59 #include <linux/tracehook.h>
60 #include <linux/psi.h>
61 #include <linux/seq_buf.h>
67 #include <linux/uaccess.h>
69 #include <trace/events/vmscan.h>
71 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
72 EXPORT_SYMBOL(memory_cgrp_subsys);
74 struct mem_cgroup *root_mem_cgroup __read_mostly;
76 /* Active memory cgroup to use from an interrupt context */
77 DEFINE_PER_CPU(struct mem_cgroup *, int_active_memcg);
79 /* Socket memory accounting disabled? */
80 static bool cgroup_memory_nosocket;
82 /* Kernel memory accounting disabled? */
83 static bool cgroup_memory_nokmem;
85 /* Whether the swap controller is active */
86 #ifdef CONFIG_MEMCG_SWAP
87 bool cgroup_memory_noswap __read_mostly;
89 #define cgroup_memory_noswap 1
92 #ifdef CONFIG_CGROUP_WRITEBACK
93 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
96 /* Whether legacy memory+swap accounting is active */
97 static bool do_memsw_account(void)
99 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_noswap;
102 #define THRESHOLDS_EVENTS_TARGET 128
103 #define SOFTLIMIT_EVENTS_TARGET 1024
106 * Cgroups above their limits are maintained in a RB-Tree, independent of
107 * their hierarchy representation
110 struct mem_cgroup_tree_per_node {
111 struct rb_root rb_root;
112 struct rb_node *rb_rightmost;
116 struct mem_cgroup_tree {
117 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
120 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
123 struct mem_cgroup_eventfd_list {
124 struct list_head list;
125 struct eventfd_ctx *eventfd;
129 * cgroup_event represents events which userspace want to receive.
131 struct mem_cgroup_event {
133 * memcg which the event belongs to.
135 struct mem_cgroup *memcg;
137 * eventfd to signal userspace about the event.
139 struct eventfd_ctx *eventfd;
141 * Each of these stored in a list by the cgroup.
143 struct list_head list;
145 * register_event() callback will be used to add new userspace
146 * waiter for changes related to this event. Use eventfd_signal()
147 * on eventfd to send notification to userspace.
149 int (*register_event)(struct mem_cgroup *memcg,
150 struct eventfd_ctx *eventfd, const char *args);
152 * unregister_event() callback will be called when userspace closes
153 * the eventfd or on cgroup removing. This callback must be set,
154 * if you want provide notification functionality.
156 void (*unregister_event)(struct mem_cgroup *memcg,
157 struct eventfd_ctx *eventfd);
159 * All fields below needed to unregister event when
160 * userspace closes eventfd.
163 wait_queue_head_t *wqh;
164 wait_queue_entry_t wait;
165 struct work_struct remove;
168 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
169 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
171 /* Stuffs for move charges at task migration. */
173 * Types of charges to be moved.
175 #define MOVE_ANON 0x1U
176 #define MOVE_FILE 0x2U
177 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
179 /* "mc" and its members are protected by cgroup_mutex */
180 static struct move_charge_struct {
181 spinlock_t lock; /* for from, to */
182 struct mm_struct *mm;
183 struct mem_cgroup *from;
184 struct mem_cgroup *to;
186 unsigned long precharge;
187 unsigned long moved_charge;
188 unsigned long moved_swap;
189 struct task_struct *moving_task; /* a task moving charges */
190 wait_queue_head_t waitq; /* a waitq for other context */
192 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
193 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
197 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
198 * limit reclaim to prevent infinite loops, if they ever occur.
200 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
201 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
203 /* for encoding cft->private value on file */
212 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
213 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
214 #define MEMFILE_ATTR(val) ((val) & 0xffff)
215 /* Used for OOM nofiier */
216 #define OOM_CONTROL (0)
219 * Iteration constructs for visiting all cgroups (under a tree). If
220 * loops are exited prematurely (break), mem_cgroup_iter_break() must
221 * be used for reference counting.
223 #define for_each_mem_cgroup_tree(iter, root) \
224 for (iter = mem_cgroup_iter(root, NULL, NULL); \
226 iter = mem_cgroup_iter(root, iter, NULL))
228 #define for_each_mem_cgroup(iter) \
229 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
231 iter = mem_cgroup_iter(NULL, iter, NULL))
233 static inline bool should_force_charge(void)
235 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
236 (current->flags & PF_EXITING);
239 /* Some nice accessors for the vmpressure. */
240 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
243 memcg = root_mem_cgroup;
244 return &memcg->vmpressure;
247 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
249 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
252 #ifdef CONFIG_MEMCG_KMEM
253 extern spinlock_t css_set_lock;
255 static void obj_cgroup_release(struct percpu_ref *ref)
257 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
258 struct mem_cgroup *memcg;
259 unsigned int nr_bytes;
260 unsigned int nr_pages;
264 * At this point all allocated objects are freed, and
265 * objcg->nr_charged_bytes can't have an arbitrary byte value.
266 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
268 * The following sequence can lead to it:
269 * 1) CPU0: objcg == stock->cached_objcg
270 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
271 * PAGE_SIZE bytes are charged
272 * 3) CPU1: a process from another memcg is allocating something,
273 * the stock if flushed,
274 * objcg->nr_charged_bytes = PAGE_SIZE - 92
275 * 5) CPU0: we do release this object,
276 * 92 bytes are added to stock->nr_bytes
277 * 6) CPU0: stock is flushed,
278 * 92 bytes are added to objcg->nr_charged_bytes
280 * In the result, nr_charged_bytes == PAGE_SIZE.
281 * This page will be uncharged in obj_cgroup_release().
283 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
284 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
285 nr_pages = nr_bytes >> PAGE_SHIFT;
287 spin_lock_irqsave(&css_set_lock, flags);
288 memcg = obj_cgroup_memcg(objcg);
290 __memcg_kmem_uncharge(memcg, nr_pages);
291 list_del(&objcg->list);
292 mem_cgroup_put(memcg);
293 spin_unlock_irqrestore(&css_set_lock, flags);
295 percpu_ref_exit(ref);
296 kfree_rcu(objcg, rcu);
299 static struct obj_cgroup *obj_cgroup_alloc(void)
301 struct obj_cgroup *objcg;
304 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
308 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
314 INIT_LIST_HEAD(&objcg->list);
318 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
319 struct mem_cgroup *parent)
321 struct obj_cgroup *objcg, *iter;
323 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
325 spin_lock_irq(&css_set_lock);
327 /* Move active objcg to the parent's list */
328 xchg(&objcg->memcg, parent);
329 css_get(&parent->css);
330 list_add(&objcg->list, &parent->objcg_list);
332 /* Move already reparented objcgs to the parent's list */
333 list_for_each_entry(iter, &memcg->objcg_list, list) {
334 css_get(&parent->css);
335 xchg(&iter->memcg, parent);
336 css_put(&memcg->css);
338 list_splice(&memcg->objcg_list, &parent->objcg_list);
340 spin_unlock_irq(&css_set_lock);
342 percpu_ref_kill(&objcg->refcnt);
346 * This will be used as a shrinker list's index.
347 * The main reason for not using cgroup id for this:
348 * this works better in sparse environments, where we have a lot of memcgs,
349 * but only a few kmem-limited. Or also, if we have, for instance, 200
350 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
351 * 200 entry array for that.
353 * The current size of the caches array is stored in memcg_nr_cache_ids. It
354 * will double each time we have to increase it.
356 static DEFINE_IDA(memcg_cache_ida);
357 int memcg_nr_cache_ids;
359 /* Protects memcg_nr_cache_ids */
360 static DECLARE_RWSEM(memcg_cache_ids_sem);
362 void memcg_get_cache_ids(void)
364 down_read(&memcg_cache_ids_sem);
367 void memcg_put_cache_ids(void)
369 up_read(&memcg_cache_ids_sem);
373 * MIN_SIZE is different than 1, because we would like to avoid going through
374 * the alloc/free process all the time. In a small machine, 4 kmem-limited
375 * cgroups is a reasonable guess. In the future, it could be a parameter or
376 * tunable, but that is strictly not necessary.
378 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
379 * this constant directly from cgroup, but it is understandable that this is
380 * better kept as an internal representation in cgroup.c. In any case, the
381 * cgrp_id space is not getting any smaller, and we don't have to necessarily
382 * increase ours as well if it increases.
384 #define MEMCG_CACHES_MIN_SIZE 4
385 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
388 * A lot of the calls to the cache allocation functions are expected to be
389 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
390 * conditional to this static branch, we'll have to allow modules that does
391 * kmem_cache_alloc and the such to see this symbol as well
393 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
394 EXPORT_SYMBOL(memcg_kmem_enabled_key);
397 static int memcg_shrinker_map_size;
398 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
400 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
402 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
405 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
406 int size, int old_size)
408 struct memcg_shrinker_map *new, *old;
411 lockdep_assert_held(&memcg_shrinker_map_mutex);
414 old = rcu_dereference_protected(
415 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
416 /* Not yet online memcg */
420 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
424 /* Set all old bits, clear all new bits */
425 memset(new->map, (int)0xff, old_size);
426 memset((void *)new->map + old_size, 0, size - old_size);
428 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
429 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
435 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
437 struct mem_cgroup_per_node *pn;
438 struct memcg_shrinker_map *map;
441 if (mem_cgroup_is_root(memcg))
445 pn = mem_cgroup_nodeinfo(memcg, nid);
446 map = rcu_dereference_protected(pn->shrinker_map, true);
449 rcu_assign_pointer(pn->shrinker_map, NULL);
453 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
455 struct memcg_shrinker_map *map;
456 int nid, size, ret = 0;
458 if (mem_cgroup_is_root(memcg))
461 mutex_lock(&memcg_shrinker_map_mutex);
462 size = memcg_shrinker_map_size;
464 map = kvzalloc_node(sizeof(*map) + size, GFP_KERNEL, nid);
466 memcg_free_shrinker_maps(memcg);
470 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
472 mutex_unlock(&memcg_shrinker_map_mutex);
477 int memcg_expand_shrinker_maps(int new_id)
479 int size, old_size, ret = 0;
480 struct mem_cgroup *memcg;
482 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
483 old_size = memcg_shrinker_map_size;
484 if (size <= old_size)
487 mutex_lock(&memcg_shrinker_map_mutex);
488 if (!root_mem_cgroup)
491 for_each_mem_cgroup(memcg) {
492 if (mem_cgroup_is_root(memcg))
494 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
496 mem_cgroup_iter_break(NULL, memcg);
502 memcg_shrinker_map_size = size;
503 mutex_unlock(&memcg_shrinker_map_mutex);
507 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
509 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
510 struct memcg_shrinker_map *map;
513 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
514 /* Pairs with smp mb in shrink_slab() */
515 smp_mb__before_atomic();
516 set_bit(shrinker_id, map->map);
522 * mem_cgroup_css_from_page - css of the memcg associated with a page
523 * @page: page of interest
525 * If memcg is bound to the default hierarchy, css of the memcg associated
526 * with @page is returned. The returned css remains associated with @page
527 * until it is released.
529 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
532 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
534 struct mem_cgroup *memcg;
536 memcg = page->mem_cgroup;
538 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
539 memcg = root_mem_cgroup;
545 * page_cgroup_ino - return inode number of the memcg a page is charged to
548 * Look up the closest online ancestor of the memory cgroup @page is charged to
549 * and return its inode number or 0 if @page is not charged to any cgroup. It
550 * is safe to call this function without holding a reference to @page.
552 * Note, this function is inherently racy, because there is nothing to prevent
553 * the cgroup inode from getting torn down and potentially reallocated a moment
554 * after page_cgroup_ino() returns, so it only should be used by callers that
555 * do not care (such as procfs interfaces).
557 ino_t page_cgroup_ino(struct page *page)
559 struct mem_cgroup *memcg;
560 unsigned long ino = 0;
563 memcg = page->mem_cgroup;
566 * The lowest bit set means that memcg isn't a valid
567 * memcg pointer, but a obj_cgroups pointer.
568 * In this case the page is shared and doesn't belong
569 * to any specific memory cgroup.
571 if ((unsigned long) memcg & 0x1UL)
574 while (memcg && !(memcg->css.flags & CSS_ONLINE))
575 memcg = parent_mem_cgroup(memcg);
577 ino = cgroup_ino(memcg->css.cgroup);
582 static struct mem_cgroup_per_node *
583 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
585 int nid = page_to_nid(page);
587 return memcg->nodeinfo[nid];
590 static struct mem_cgroup_tree_per_node *
591 soft_limit_tree_node(int nid)
593 return soft_limit_tree.rb_tree_per_node[nid];
596 static struct mem_cgroup_tree_per_node *
597 soft_limit_tree_from_page(struct page *page)
599 int nid = page_to_nid(page);
601 return soft_limit_tree.rb_tree_per_node[nid];
604 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
605 struct mem_cgroup_tree_per_node *mctz,
606 unsigned long new_usage_in_excess)
608 struct rb_node **p = &mctz->rb_root.rb_node;
609 struct rb_node *parent = NULL;
610 struct mem_cgroup_per_node *mz_node;
611 bool rightmost = true;
616 mz->usage_in_excess = new_usage_in_excess;
617 if (!mz->usage_in_excess)
621 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
623 if (mz->usage_in_excess < mz_node->usage_in_excess) {
632 mctz->rb_rightmost = &mz->tree_node;
634 rb_link_node(&mz->tree_node, parent, p);
635 rb_insert_color(&mz->tree_node, &mctz->rb_root);
639 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
640 struct mem_cgroup_tree_per_node *mctz)
645 if (&mz->tree_node == mctz->rb_rightmost)
646 mctz->rb_rightmost = rb_prev(&mz->tree_node);
648 rb_erase(&mz->tree_node, &mctz->rb_root);
652 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
653 struct mem_cgroup_tree_per_node *mctz)
657 spin_lock_irqsave(&mctz->lock, flags);
658 __mem_cgroup_remove_exceeded(mz, mctz);
659 spin_unlock_irqrestore(&mctz->lock, flags);
662 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
664 unsigned long nr_pages = page_counter_read(&memcg->memory);
665 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
666 unsigned long excess = 0;
668 if (nr_pages > soft_limit)
669 excess = nr_pages - soft_limit;
674 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
676 unsigned long excess;
677 struct mem_cgroup_per_node *mz;
678 struct mem_cgroup_tree_per_node *mctz;
680 mctz = soft_limit_tree_from_page(page);
684 * Necessary to update all ancestors when hierarchy is used.
685 * because their event counter is not touched.
687 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
688 mz = mem_cgroup_page_nodeinfo(memcg, page);
689 excess = soft_limit_excess(memcg);
691 * We have to update the tree if mz is on RB-tree or
692 * mem is over its softlimit.
694 if (excess || mz->on_tree) {
697 spin_lock_irqsave(&mctz->lock, flags);
698 /* if on-tree, remove it */
700 __mem_cgroup_remove_exceeded(mz, mctz);
702 * Insert again. mz->usage_in_excess will be updated.
703 * If excess is 0, no tree ops.
705 __mem_cgroup_insert_exceeded(mz, mctz, excess);
706 spin_unlock_irqrestore(&mctz->lock, flags);
711 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
713 struct mem_cgroup_tree_per_node *mctz;
714 struct mem_cgroup_per_node *mz;
718 mz = mem_cgroup_nodeinfo(memcg, nid);
719 mctz = soft_limit_tree_node(nid);
721 mem_cgroup_remove_exceeded(mz, mctz);
725 static struct mem_cgroup_per_node *
726 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
728 struct mem_cgroup_per_node *mz;
732 if (!mctz->rb_rightmost)
733 goto done; /* Nothing to reclaim from */
735 mz = rb_entry(mctz->rb_rightmost,
736 struct mem_cgroup_per_node, tree_node);
738 * Remove the node now but someone else can add it back,
739 * we will to add it back at the end of reclaim to its correct
740 * position in the tree.
742 __mem_cgroup_remove_exceeded(mz, mctz);
743 if (!soft_limit_excess(mz->memcg) ||
744 !css_tryget(&mz->memcg->css))
750 static struct mem_cgroup_per_node *
751 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
753 struct mem_cgroup_per_node *mz;
755 spin_lock_irq(&mctz->lock);
756 mz = __mem_cgroup_largest_soft_limit_node(mctz);
757 spin_unlock_irq(&mctz->lock);
762 * __mod_memcg_state - update cgroup memory statistics
763 * @memcg: the memory cgroup
764 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
765 * @val: delta to add to the counter, can be negative
767 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
769 long x, threshold = MEMCG_CHARGE_BATCH;
771 if (mem_cgroup_disabled())
774 if (memcg_stat_item_in_bytes(idx))
775 threshold <<= PAGE_SHIFT;
777 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
778 if (unlikely(abs(x) > threshold)) {
779 struct mem_cgroup *mi;
782 * Batch local counters to keep them in sync with
783 * the hierarchical ones.
785 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
786 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
787 atomic_long_add(x, &mi->vmstats[idx]);
790 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
793 static struct mem_cgroup_per_node *
794 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
796 struct mem_cgroup *parent;
798 parent = parent_mem_cgroup(pn->memcg);
801 return mem_cgroup_nodeinfo(parent, nid);
804 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
807 struct mem_cgroup_per_node *pn;
808 struct mem_cgroup *memcg;
809 long x, threshold = MEMCG_CHARGE_BATCH;
811 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
815 __mod_memcg_state(memcg, idx, val);
818 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
820 if (vmstat_item_in_bytes(idx))
821 threshold <<= PAGE_SHIFT;
823 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
824 if (unlikely(abs(x) > threshold)) {
825 pg_data_t *pgdat = lruvec_pgdat(lruvec);
826 struct mem_cgroup_per_node *pi;
828 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
829 atomic_long_add(x, &pi->lruvec_stat[idx]);
832 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
836 * __mod_lruvec_state - update lruvec memory statistics
837 * @lruvec: the lruvec
838 * @idx: the stat item
839 * @val: delta to add to the counter, can be negative
841 * The lruvec is the intersection of the NUMA node and a cgroup. This
842 * function updates the all three counters that are affected by a
843 * change of state at this level: per-node, per-cgroup, per-lruvec.
845 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
849 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
851 /* Update memcg and lruvec */
852 if (!mem_cgroup_disabled())
853 __mod_memcg_lruvec_state(lruvec, idx, val);
856 void __mod_lruvec_slab_state(void *p, enum node_stat_item idx, int val)
858 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
859 struct mem_cgroup *memcg;
860 struct lruvec *lruvec;
863 memcg = mem_cgroup_from_obj(p);
866 * Untracked pages have no memcg, no lruvec. Update only the
867 * node. If we reparent the slab objects to the root memcg,
868 * when we free the slab object, we need to update the per-memcg
869 * vmstats to keep it correct for the root memcg.
872 __mod_node_page_state(pgdat, idx, val);
874 lruvec = mem_cgroup_lruvec(memcg, pgdat);
875 __mod_lruvec_state(lruvec, idx, val);
881 * __count_memcg_events - account VM events in a cgroup
882 * @memcg: the memory cgroup
883 * @idx: the event item
884 * @count: the number of events that occured
886 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
891 if (mem_cgroup_disabled())
894 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
895 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
896 struct mem_cgroup *mi;
899 * Batch local counters to keep them in sync with
900 * the hierarchical ones.
902 __this_cpu_add(memcg->vmstats_local->events[idx], x);
903 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
904 atomic_long_add(x, &mi->vmevents[idx]);
907 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
910 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
912 return atomic_long_read(&memcg->vmevents[event]);
915 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
920 for_each_possible_cpu(cpu)
921 x += per_cpu(memcg->vmstats_local->events[event], cpu);
925 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
929 /* pagein of a big page is an event. So, ignore page size */
931 __count_memcg_events(memcg, PGPGIN, 1);
933 __count_memcg_events(memcg, PGPGOUT, 1);
934 nr_pages = -nr_pages; /* for event */
937 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
940 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
941 enum mem_cgroup_events_target target)
943 unsigned long val, next;
945 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
946 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
947 /* from time_after() in jiffies.h */
948 if ((long)(next - val) < 0) {
950 case MEM_CGROUP_TARGET_THRESH:
951 next = val + THRESHOLDS_EVENTS_TARGET;
953 case MEM_CGROUP_TARGET_SOFTLIMIT:
954 next = val + SOFTLIMIT_EVENTS_TARGET;
959 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
966 * Check events in order.
969 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
971 /* threshold event is triggered in finer grain than soft limit */
972 if (unlikely(mem_cgroup_event_ratelimit(memcg,
973 MEM_CGROUP_TARGET_THRESH))) {
976 do_softlimit = mem_cgroup_event_ratelimit(memcg,
977 MEM_CGROUP_TARGET_SOFTLIMIT);
978 mem_cgroup_threshold(memcg);
979 if (unlikely(do_softlimit))
980 mem_cgroup_update_tree(memcg, page);
984 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
987 * mm_update_next_owner() may clear mm->owner to NULL
988 * if it races with swapoff, page migration, etc.
989 * So this can be called with p == NULL.
994 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
996 EXPORT_SYMBOL(mem_cgroup_from_task);
999 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1000 * @mm: mm from which memcg should be extracted. It can be NULL.
1002 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
1003 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
1006 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1008 struct mem_cgroup *memcg;
1010 if (mem_cgroup_disabled())
1016 * Page cache insertions can happen withou an
1017 * actual mm context, e.g. during disk probing
1018 * on boot, loopback IO, acct() writes etc.
1021 memcg = root_mem_cgroup;
1023 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1024 if (unlikely(!memcg))
1025 memcg = root_mem_cgroup;
1027 } while (!css_tryget(&memcg->css));
1031 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1034 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
1035 * @page: page from which memcg should be extracted.
1037 * Obtain a reference on page->memcg and returns it if successful. Otherwise
1038 * root_mem_cgroup is returned.
1040 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
1042 struct mem_cgroup *memcg = page->mem_cgroup;
1044 if (mem_cgroup_disabled())
1048 /* Page should not get uncharged and freed memcg under us. */
1049 if (!memcg || WARN_ON_ONCE(!css_tryget(&memcg->css)))
1050 memcg = root_mem_cgroup;
1054 EXPORT_SYMBOL(get_mem_cgroup_from_page);
1056 static __always_inline struct mem_cgroup *active_memcg(void)
1059 return this_cpu_read(int_active_memcg);
1061 return current->active_memcg;
1064 static __always_inline struct mem_cgroup *get_active_memcg(void)
1066 struct mem_cgroup *memcg;
1069 memcg = active_memcg();
1071 /* current->active_memcg must hold a ref. */
1072 if (WARN_ON_ONCE(!css_tryget(&memcg->css)))
1073 memcg = root_mem_cgroup;
1075 memcg = current->active_memcg;
1082 static __always_inline bool memcg_kmem_bypass(void)
1084 /* Allow remote memcg charging from any context. */
1085 if (unlikely(active_memcg()))
1088 /* Memcg to charge can't be determined. */
1089 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
1096 * If active memcg is set, do not fallback to current->mm->memcg.
1098 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
1100 if (memcg_kmem_bypass())
1103 if (unlikely(active_memcg()))
1104 return get_active_memcg();
1106 return get_mem_cgroup_from_mm(current->mm);
1110 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1111 * @root: hierarchy root
1112 * @prev: previously returned memcg, NULL on first invocation
1113 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1115 * Returns references to children of the hierarchy below @root, or
1116 * @root itself, or %NULL after a full round-trip.
1118 * Caller must pass the return value in @prev on subsequent
1119 * invocations for reference counting, or use mem_cgroup_iter_break()
1120 * to cancel a hierarchy walk before the round-trip is complete.
1122 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1123 * in the hierarchy among all concurrent reclaimers operating on the
1126 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1127 struct mem_cgroup *prev,
1128 struct mem_cgroup_reclaim_cookie *reclaim)
1130 struct mem_cgroup_reclaim_iter *iter;
1131 struct cgroup_subsys_state *css = NULL;
1132 struct mem_cgroup *memcg = NULL;
1133 struct mem_cgroup *pos = NULL;
1135 if (mem_cgroup_disabled())
1139 root = root_mem_cgroup;
1141 if (prev && !reclaim)
1144 if (!root->use_hierarchy && root != root_mem_cgroup) {
1153 struct mem_cgroup_per_node *mz;
1155 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1158 if (prev && reclaim->generation != iter->generation)
1162 pos = READ_ONCE(iter->position);
1163 if (!pos || css_tryget(&pos->css))
1166 * css reference reached zero, so iter->position will
1167 * be cleared by ->css_released. However, we should not
1168 * rely on this happening soon, because ->css_released
1169 * is called from a work queue, and by busy-waiting we
1170 * might block it. So we clear iter->position right
1173 (void)cmpxchg(&iter->position, pos, NULL);
1181 css = css_next_descendant_pre(css, &root->css);
1184 * Reclaimers share the hierarchy walk, and a
1185 * new one might jump in right at the end of
1186 * the hierarchy - make sure they see at least
1187 * one group and restart from the beginning.
1195 * Verify the css and acquire a reference. The root
1196 * is provided by the caller, so we know it's alive
1197 * and kicking, and don't take an extra reference.
1199 memcg = mem_cgroup_from_css(css);
1201 if (css == &root->css)
1204 if (css_tryget(css))
1212 * The position could have already been updated by a competing
1213 * thread, so check that the value hasn't changed since we read
1214 * it to avoid reclaiming from the same cgroup twice.
1216 (void)cmpxchg(&iter->position, pos, memcg);
1224 reclaim->generation = iter->generation;
1230 if (prev && prev != root)
1231 css_put(&prev->css);
1237 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1238 * @root: hierarchy root
1239 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1241 void mem_cgroup_iter_break(struct mem_cgroup *root,
1242 struct mem_cgroup *prev)
1245 root = root_mem_cgroup;
1246 if (prev && prev != root)
1247 css_put(&prev->css);
1250 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1251 struct mem_cgroup *dead_memcg)
1253 struct mem_cgroup_reclaim_iter *iter;
1254 struct mem_cgroup_per_node *mz;
1257 for_each_node(nid) {
1258 mz = mem_cgroup_nodeinfo(from, nid);
1260 cmpxchg(&iter->position, dead_memcg, NULL);
1264 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1266 struct mem_cgroup *memcg = dead_memcg;
1267 struct mem_cgroup *last;
1270 __invalidate_reclaim_iterators(memcg, dead_memcg);
1272 } while ((memcg = parent_mem_cgroup(memcg)));
1275 * When cgruop1 non-hierarchy mode is used,
1276 * parent_mem_cgroup() does not walk all the way up to the
1277 * cgroup root (root_mem_cgroup). So we have to handle
1278 * dead_memcg from cgroup root separately.
1280 if (last != root_mem_cgroup)
1281 __invalidate_reclaim_iterators(root_mem_cgroup,
1286 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1287 * @memcg: hierarchy root
1288 * @fn: function to call for each task
1289 * @arg: argument passed to @fn
1291 * This function iterates over tasks attached to @memcg or to any of its
1292 * descendants and calls @fn for each task. If @fn returns a non-zero
1293 * value, the function breaks the iteration loop and returns the value.
1294 * Otherwise, it will iterate over all tasks and return 0.
1296 * This function must not be called for the root memory cgroup.
1298 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1299 int (*fn)(struct task_struct *, void *), void *arg)
1301 struct mem_cgroup *iter;
1304 BUG_ON(memcg == root_mem_cgroup);
1306 for_each_mem_cgroup_tree(iter, memcg) {
1307 struct css_task_iter it;
1308 struct task_struct *task;
1310 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1311 while (!ret && (task = css_task_iter_next(&it)))
1312 ret = fn(task, arg);
1313 css_task_iter_end(&it);
1315 mem_cgroup_iter_break(memcg, iter);
1323 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1325 * @pgdat: pgdat of the page
1327 * This function relies on page->mem_cgroup being stable - see the
1328 * access rules in commit_charge().
1330 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1332 struct mem_cgroup_per_node *mz;
1333 struct mem_cgroup *memcg;
1334 struct lruvec *lruvec;
1336 if (mem_cgroup_disabled()) {
1337 lruvec = &pgdat->__lruvec;
1341 memcg = page->mem_cgroup;
1343 * Swapcache readahead pages are added to the LRU - and
1344 * possibly migrated - before they are charged.
1347 memcg = root_mem_cgroup;
1349 mz = mem_cgroup_page_nodeinfo(memcg, page);
1350 lruvec = &mz->lruvec;
1353 * Since a node can be onlined after the mem_cgroup was created,
1354 * we have to be prepared to initialize lruvec->zone here;
1355 * and if offlined then reonlined, we need to reinitialize it.
1357 if (unlikely(lruvec->pgdat != pgdat))
1358 lruvec->pgdat = pgdat;
1363 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1364 * @lruvec: mem_cgroup per zone lru vector
1365 * @lru: index of lru list the page is sitting on
1366 * @zid: zone id of the accounted pages
1367 * @nr_pages: positive when adding or negative when removing
1369 * This function must be called under lru_lock, just before a page is added
1370 * to or just after a page is removed from an lru list (that ordering being
1371 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1373 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1374 int zid, int nr_pages)
1376 struct mem_cgroup_per_node *mz;
1377 unsigned long *lru_size;
1380 if (mem_cgroup_disabled())
1383 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1384 lru_size = &mz->lru_zone_size[zid][lru];
1387 *lru_size += nr_pages;
1390 if (WARN_ONCE(size < 0,
1391 "%s(%p, %d, %d): lru_size %ld\n",
1392 __func__, lruvec, lru, nr_pages, size)) {
1398 *lru_size += nr_pages;
1402 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1403 * @memcg: the memory cgroup
1405 * Returns the maximum amount of memory @mem can be charged with, in
1408 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1410 unsigned long margin = 0;
1411 unsigned long count;
1412 unsigned long limit;
1414 count = page_counter_read(&memcg->memory);
1415 limit = READ_ONCE(memcg->memory.max);
1417 margin = limit - count;
1419 if (do_memsw_account()) {
1420 count = page_counter_read(&memcg->memsw);
1421 limit = READ_ONCE(memcg->memsw.max);
1423 margin = min(margin, limit - count);
1432 * A routine for checking "mem" is under move_account() or not.
1434 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1435 * moving cgroups. This is for waiting at high-memory pressure
1438 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1440 struct mem_cgroup *from;
1441 struct mem_cgroup *to;
1444 * Unlike task_move routines, we access mc.to, mc.from not under
1445 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1447 spin_lock(&mc.lock);
1453 ret = mem_cgroup_is_descendant(from, memcg) ||
1454 mem_cgroup_is_descendant(to, memcg);
1456 spin_unlock(&mc.lock);
1460 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1462 if (mc.moving_task && current != mc.moving_task) {
1463 if (mem_cgroup_under_move(memcg)) {
1465 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1466 /* moving charge context might have finished. */
1469 finish_wait(&mc.waitq, &wait);
1476 struct memory_stat {
1482 static struct memory_stat memory_stats[] = {
1483 { "anon", PAGE_SIZE, NR_ANON_MAPPED },
1484 { "file", PAGE_SIZE, NR_FILE_PAGES },
1485 { "kernel_stack", 1024, NR_KERNEL_STACK_KB },
1486 { "percpu", 1, MEMCG_PERCPU_B },
1487 { "sock", PAGE_SIZE, MEMCG_SOCK },
1488 { "shmem", PAGE_SIZE, NR_SHMEM },
1489 { "file_mapped", PAGE_SIZE, NR_FILE_MAPPED },
1490 { "file_dirty", PAGE_SIZE, NR_FILE_DIRTY },
1491 { "file_writeback", PAGE_SIZE, NR_WRITEBACK },
1492 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1494 * The ratio will be initialized in memory_stats_init(). Because
1495 * on some architectures, the macro of HPAGE_PMD_SIZE is not
1496 * constant(e.g. powerpc).
1498 { "anon_thp", 0, NR_ANON_THPS },
1499 { "file_thp", 0, NR_FILE_THPS },
1500 { "shmem_thp", 0, NR_SHMEM_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].idx == NR_FILE_THPS ||
1533 memory_stats[i].idx == NR_SHMEM_THPS)
1534 memory_stats[i].ratio = HPAGE_PMD_SIZE;
1536 VM_BUG_ON(!memory_stats[i].ratio);
1537 VM_BUG_ON(memory_stats[i].idx >= MEMCG_NR_STAT);
1542 pure_initcall(memory_stats_init);
1544 static char *memory_stat_format(struct mem_cgroup *memcg)
1549 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1554 * Provide statistics on the state of the memory subsystem as
1555 * well as cumulative event counters that show past behavior.
1557 * This list is ordered following a combination of these gradients:
1558 * 1) generic big picture -> specifics and details
1559 * 2) reflecting userspace activity -> reflecting kernel heuristics
1561 * Current memory state:
1564 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1567 size = memcg_page_state(memcg, memory_stats[i].idx);
1568 size *= memory_stats[i].ratio;
1569 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1571 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1572 size = memcg_page_state(memcg, NR_SLAB_RECLAIMABLE_B) +
1573 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE_B);
1574 seq_buf_printf(&s, "slab %llu\n", size);
1578 /* Accumulated memory events */
1580 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1581 memcg_events(memcg, PGFAULT));
1582 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1583 memcg_events(memcg, PGMAJFAULT));
1584 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1585 memcg_events(memcg, PGREFILL));
1586 seq_buf_printf(&s, "pgscan %lu\n",
1587 memcg_events(memcg, PGSCAN_KSWAPD) +
1588 memcg_events(memcg, PGSCAN_DIRECT));
1589 seq_buf_printf(&s, "pgsteal %lu\n",
1590 memcg_events(memcg, PGSTEAL_KSWAPD) +
1591 memcg_events(memcg, PGSTEAL_DIRECT));
1592 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1593 memcg_events(memcg, PGACTIVATE));
1594 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1595 memcg_events(memcg, PGDEACTIVATE));
1596 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1597 memcg_events(memcg, PGLAZYFREE));
1598 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1599 memcg_events(memcg, PGLAZYFREED));
1601 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1602 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1603 memcg_events(memcg, THP_FAULT_ALLOC));
1604 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1605 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1606 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1608 /* The above should easily fit into one page */
1609 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1614 #define K(x) ((x) << (PAGE_SHIFT-10))
1616 * mem_cgroup_print_oom_context: Print OOM information relevant to
1617 * memory controller.
1618 * @memcg: The memory cgroup that went over limit
1619 * @p: Task that is going to be killed
1621 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1624 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1629 pr_cont(",oom_memcg=");
1630 pr_cont_cgroup_path(memcg->css.cgroup);
1632 pr_cont(",global_oom");
1634 pr_cont(",task_memcg=");
1635 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1641 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1642 * memory controller.
1643 * @memcg: The memory cgroup that went over limit
1645 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1649 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1650 K((u64)page_counter_read(&memcg->memory)),
1651 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1652 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1653 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1654 K((u64)page_counter_read(&memcg->swap)),
1655 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1657 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1658 K((u64)page_counter_read(&memcg->memsw)),
1659 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1660 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1661 K((u64)page_counter_read(&memcg->kmem)),
1662 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1665 pr_info("Memory cgroup stats for ");
1666 pr_cont_cgroup_path(memcg->css.cgroup);
1668 buf = memory_stat_format(memcg);
1676 * Return the memory (and swap, if configured) limit for a memcg.
1678 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1680 unsigned long max = READ_ONCE(memcg->memory.max);
1682 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1683 if (mem_cgroup_swappiness(memcg))
1684 max += min(READ_ONCE(memcg->swap.max),
1685 (unsigned long)total_swap_pages);
1687 if (mem_cgroup_swappiness(memcg)) {
1688 /* Calculate swap excess capacity from memsw limit */
1689 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1691 max += min(swap, (unsigned long)total_swap_pages);
1697 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1699 return page_counter_read(&memcg->memory);
1702 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1705 struct oom_control oc = {
1709 .gfp_mask = gfp_mask,
1714 if (mutex_lock_killable(&oom_lock))
1717 if (mem_cgroup_margin(memcg) >= (1 << order))
1721 * A few threads which were not waiting at mutex_lock_killable() can
1722 * fail to bail out. Therefore, check again after holding oom_lock.
1724 ret = should_force_charge() || out_of_memory(&oc);
1727 mutex_unlock(&oom_lock);
1731 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1734 unsigned long *total_scanned)
1736 struct mem_cgroup *victim = NULL;
1739 unsigned long excess;
1740 unsigned long nr_scanned;
1741 struct mem_cgroup_reclaim_cookie reclaim = {
1745 excess = soft_limit_excess(root_memcg);
1748 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1753 * If we have not been able to reclaim
1754 * anything, it might because there are
1755 * no reclaimable pages under this hierarchy
1760 * We want to do more targeted reclaim.
1761 * excess >> 2 is not to excessive so as to
1762 * reclaim too much, nor too less that we keep
1763 * coming back to reclaim from this cgroup
1765 if (total >= (excess >> 2) ||
1766 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1771 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1772 pgdat, &nr_scanned);
1773 *total_scanned += nr_scanned;
1774 if (!soft_limit_excess(root_memcg))
1777 mem_cgroup_iter_break(root_memcg, victim);
1781 #ifdef CONFIG_LOCKDEP
1782 static struct lockdep_map memcg_oom_lock_dep_map = {
1783 .name = "memcg_oom_lock",
1787 static DEFINE_SPINLOCK(memcg_oom_lock);
1790 * Check OOM-Killer is already running under our hierarchy.
1791 * If someone is running, return false.
1793 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1795 struct mem_cgroup *iter, *failed = NULL;
1797 spin_lock(&memcg_oom_lock);
1799 for_each_mem_cgroup_tree(iter, memcg) {
1800 if (iter->oom_lock) {
1802 * this subtree of our hierarchy is already locked
1803 * so we cannot give a lock.
1806 mem_cgroup_iter_break(memcg, iter);
1809 iter->oom_lock = true;
1814 * OK, we failed to lock the whole subtree so we have
1815 * to clean up what we set up to the failing subtree
1817 for_each_mem_cgroup_tree(iter, memcg) {
1818 if (iter == failed) {
1819 mem_cgroup_iter_break(memcg, iter);
1822 iter->oom_lock = false;
1825 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1827 spin_unlock(&memcg_oom_lock);
1832 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1834 struct mem_cgroup *iter;
1836 spin_lock(&memcg_oom_lock);
1837 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1838 for_each_mem_cgroup_tree(iter, memcg)
1839 iter->oom_lock = false;
1840 spin_unlock(&memcg_oom_lock);
1843 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1845 struct mem_cgroup *iter;
1847 spin_lock(&memcg_oom_lock);
1848 for_each_mem_cgroup_tree(iter, memcg)
1850 spin_unlock(&memcg_oom_lock);
1853 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1855 struct mem_cgroup *iter;
1858 * Be careful about under_oom underflows becase a child memcg
1859 * could have been added after mem_cgroup_mark_under_oom.
1861 spin_lock(&memcg_oom_lock);
1862 for_each_mem_cgroup_tree(iter, memcg)
1863 if (iter->under_oom > 0)
1865 spin_unlock(&memcg_oom_lock);
1868 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1870 struct oom_wait_info {
1871 struct mem_cgroup *memcg;
1872 wait_queue_entry_t wait;
1875 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1876 unsigned mode, int sync, void *arg)
1878 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1879 struct mem_cgroup *oom_wait_memcg;
1880 struct oom_wait_info *oom_wait_info;
1882 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1883 oom_wait_memcg = oom_wait_info->memcg;
1885 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1886 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1888 return autoremove_wake_function(wait, mode, sync, arg);
1891 static void memcg_oom_recover(struct mem_cgroup *memcg)
1894 * For the following lockless ->under_oom test, the only required
1895 * guarantee is that it must see the state asserted by an OOM when
1896 * this function is called as a result of userland actions
1897 * triggered by the notification of the OOM. This is trivially
1898 * achieved by invoking mem_cgroup_mark_under_oom() before
1899 * triggering notification.
1901 if (memcg && memcg->under_oom)
1902 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1912 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1914 enum oom_status ret;
1917 if (order > PAGE_ALLOC_COSTLY_ORDER)
1920 memcg_memory_event(memcg, MEMCG_OOM);
1923 * We are in the middle of the charge context here, so we
1924 * don't want to block when potentially sitting on a callstack
1925 * that holds all kinds of filesystem and mm locks.
1927 * cgroup1 allows disabling the OOM killer and waiting for outside
1928 * handling until the charge can succeed; remember the context and put
1929 * the task to sleep at the end of the page fault when all locks are
1932 * On the other hand, in-kernel OOM killer allows for an async victim
1933 * memory reclaim (oom_reaper) and that means that we are not solely
1934 * relying on the oom victim to make a forward progress and we can
1935 * invoke the oom killer here.
1937 * Please note that mem_cgroup_out_of_memory might fail to find a
1938 * victim and then we have to bail out from the charge path.
1940 if (memcg->oom_kill_disable) {
1941 if (!current->in_user_fault)
1943 css_get(&memcg->css);
1944 current->memcg_in_oom = memcg;
1945 current->memcg_oom_gfp_mask = mask;
1946 current->memcg_oom_order = order;
1951 mem_cgroup_mark_under_oom(memcg);
1953 locked = mem_cgroup_oom_trylock(memcg);
1956 mem_cgroup_oom_notify(memcg);
1958 mem_cgroup_unmark_under_oom(memcg);
1959 if (mem_cgroup_out_of_memory(memcg, mask, order))
1965 mem_cgroup_oom_unlock(memcg);
1971 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1972 * @handle: actually kill/wait or just clean up the OOM state
1974 * This has to be called at the end of a page fault if the memcg OOM
1975 * handler was enabled.
1977 * Memcg supports userspace OOM handling where failed allocations must
1978 * sleep on a waitqueue until the userspace task resolves the
1979 * situation. Sleeping directly in the charge context with all kinds
1980 * of locks held is not a good idea, instead we remember an OOM state
1981 * in the task and mem_cgroup_oom_synchronize() has to be called at
1982 * the end of the page fault to complete the OOM handling.
1984 * Returns %true if an ongoing memcg OOM situation was detected and
1985 * completed, %false otherwise.
1987 bool mem_cgroup_oom_synchronize(bool handle)
1989 struct mem_cgroup *memcg = current->memcg_in_oom;
1990 struct oom_wait_info owait;
1993 /* OOM is global, do not handle */
2000 owait.memcg = memcg;
2001 owait.wait.flags = 0;
2002 owait.wait.func = memcg_oom_wake_function;
2003 owait.wait.private = current;
2004 INIT_LIST_HEAD(&owait.wait.entry);
2006 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2007 mem_cgroup_mark_under_oom(memcg);
2009 locked = mem_cgroup_oom_trylock(memcg);
2012 mem_cgroup_oom_notify(memcg);
2014 if (locked && !memcg->oom_kill_disable) {
2015 mem_cgroup_unmark_under_oom(memcg);
2016 finish_wait(&memcg_oom_waitq, &owait.wait);
2017 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
2018 current->memcg_oom_order);
2021 mem_cgroup_unmark_under_oom(memcg);
2022 finish_wait(&memcg_oom_waitq, &owait.wait);
2026 mem_cgroup_oom_unlock(memcg);
2028 * There is no guarantee that an OOM-lock contender
2029 * sees the wakeups triggered by the OOM kill
2030 * uncharges. Wake any sleepers explicitely.
2032 memcg_oom_recover(memcg);
2035 current->memcg_in_oom = NULL;
2036 css_put(&memcg->css);
2041 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2042 * @victim: task to be killed by the OOM killer
2043 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2045 * Returns a pointer to a memory cgroup, which has to be cleaned up
2046 * by killing all belonging OOM-killable tasks.
2048 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2050 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2051 struct mem_cgroup *oom_domain)
2053 struct mem_cgroup *oom_group = NULL;
2054 struct mem_cgroup *memcg;
2056 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2060 oom_domain = root_mem_cgroup;
2064 memcg = mem_cgroup_from_task(victim);
2065 if (memcg == root_mem_cgroup)
2069 * If the victim task has been asynchronously moved to a different
2070 * memory cgroup, we might end up killing tasks outside oom_domain.
2071 * In this case it's better to ignore memory.group.oom.
2073 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2077 * Traverse the memory cgroup hierarchy from the victim task's
2078 * cgroup up to the OOMing cgroup (or root) to find the
2079 * highest-level memory cgroup with oom.group set.
2081 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2082 if (memcg->oom_group)
2085 if (memcg == oom_domain)
2090 css_get(&oom_group->css);
2097 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2099 pr_info("Tasks in ");
2100 pr_cont_cgroup_path(memcg->css.cgroup);
2101 pr_cont(" are going to be killed due to memory.oom.group set\n");
2105 * lock_page_memcg - lock a page->mem_cgroup binding
2108 * This function protects unlocked LRU pages from being moved to
2111 * It ensures lifetime of the returned memcg. Caller is responsible
2112 * for the lifetime of the page; __unlock_page_memcg() is available
2113 * when @page might get freed inside the locked section.
2115 struct mem_cgroup *lock_page_memcg(struct page *page)
2117 struct page *head = compound_head(page); /* rmap on tail pages */
2118 struct mem_cgroup *memcg;
2119 unsigned long flags;
2122 * The RCU lock is held throughout the transaction. The fast
2123 * path can get away without acquiring the memcg->move_lock
2124 * because page moving starts with an RCU grace period.
2126 * The RCU lock also protects the memcg from being freed when
2127 * the page state that is going to change is the only thing
2128 * preventing the page itself from being freed. E.g. writeback
2129 * doesn't hold a page reference and relies on PG_writeback to
2130 * keep off truncation, migration and so forth.
2134 if (mem_cgroup_disabled())
2137 memcg = head->mem_cgroup;
2138 if (unlikely(!memcg))
2141 if (atomic_read(&memcg->moving_account) <= 0)
2144 spin_lock_irqsave(&memcg->move_lock, flags);
2145 if (memcg != head->mem_cgroup) {
2146 spin_unlock_irqrestore(&memcg->move_lock, flags);
2151 * When charge migration first begins, we can have locked and
2152 * unlocked page stat updates happening concurrently. Track
2153 * the task who has the lock for unlock_page_memcg().
2155 memcg->move_lock_task = current;
2156 memcg->move_lock_flags = flags;
2160 EXPORT_SYMBOL(lock_page_memcg);
2163 * __unlock_page_memcg - unlock and unpin a memcg
2166 * Unlock and unpin a memcg returned by lock_page_memcg().
2168 void __unlock_page_memcg(struct mem_cgroup *memcg)
2170 if (memcg && memcg->move_lock_task == current) {
2171 unsigned long flags = memcg->move_lock_flags;
2173 memcg->move_lock_task = NULL;
2174 memcg->move_lock_flags = 0;
2176 spin_unlock_irqrestore(&memcg->move_lock, flags);
2183 * unlock_page_memcg - unlock a page->mem_cgroup binding
2186 void unlock_page_memcg(struct page *page)
2188 struct page *head = compound_head(page);
2190 __unlock_page_memcg(head->mem_cgroup);
2192 EXPORT_SYMBOL(unlock_page_memcg);
2194 struct memcg_stock_pcp {
2195 struct mem_cgroup *cached; /* this never be root cgroup */
2196 unsigned int nr_pages;
2198 #ifdef CONFIG_MEMCG_KMEM
2199 struct obj_cgroup *cached_objcg;
2200 unsigned int nr_bytes;
2203 struct work_struct work;
2204 unsigned long flags;
2205 #define FLUSHING_CACHED_CHARGE 0
2207 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2208 static DEFINE_MUTEX(percpu_charge_mutex);
2210 #ifdef CONFIG_MEMCG_KMEM
2211 static void drain_obj_stock(struct memcg_stock_pcp *stock);
2212 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2213 struct mem_cgroup *root_memcg);
2216 static inline void drain_obj_stock(struct memcg_stock_pcp *stock)
2219 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2220 struct mem_cgroup *root_memcg)
2227 * consume_stock: Try to consume stocked charge on this cpu.
2228 * @memcg: memcg to consume from.
2229 * @nr_pages: how many pages to charge.
2231 * The charges will only happen if @memcg matches the current cpu's memcg
2232 * stock, and at least @nr_pages are available in that stock. Failure to
2233 * service an allocation will refill the stock.
2235 * returns true if successful, false otherwise.
2237 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2239 struct memcg_stock_pcp *stock;
2240 unsigned long flags;
2243 if (nr_pages > MEMCG_CHARGE_BATCH)
2246 local_irq_save(flags);
2248 stock = this_cpu_ptr(&memcg_stock);
2249 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2250 stock->nr_pages -= nr_pages;
2254 local_irq_restore(flags);
2260 * Returns stocks cached in percpu and reset cached information.
2262 static void drain_stock(struct memcg_stock_pcp *stock)
2264 struct mem_cgroup *old = stock->cached;
2269 if (stock->nr_pages) {
2270 page_counter_uncharge(&old->memory, stock->nr_pages);
2271 if (do_memsw_account())
2272 page_counter_uncharge(&old->memsw, stock->nr_pages);
2273 stock->nr_pages = 0;
2277 stock->cached = NULL;
2280 static void drain_local_stock(struct work_struct *dummy)
2282 struct memcg_stock_pcp *stock;
2283 unsigned long flags;
2286 * The only protection from memory hotplug vs. drain_stock races is
2287 * that we always operate on local CPU stock here with IRQ disabled
2289 local_irq_save(flags);
2291 stock = this_cpu_ptr(&memcg_stock);
2292 drain_obj_stock(stock);
2294 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2296 local_irq_restore(flags);
2300 * Cache charges(val) to local per_cpu area.
2301 * This will be consumed by consume_stock() function, later.
2303 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2305 struct memcg_stock_pcp *stock;
2306 unsigned long flags;
2308 local_irq_save(flags);
2310 stock = this_cpu_ptr(&memcg_stock);
2311 if (stock->cached != memcg) { /* reset if necessary */
2313 css_get(&memcg->css);
2314 stock->cached = memcg;
2316 stock->nr_pages += nr_pages;
2318 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2321 local_irq_restore(flags);
2325 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2326 * of the hierarchy under it.
2328 static void drain_all_stock(struct mem_cgroup *root_memcg)
2332 /* If someone's already draining, avoid adding running more workers. */
2333 if (!mutex_trylock(&percpu_charge_mutex))
2336 * Notify other cpus that system-wide "drain" is running
2337 * We do not care about races with the cpu hotplug because cpu down
2338 * as well as workers from this path always operate on the local
2339 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2342 for_each_online_cpu(cpu) {
2343 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2344 struct mem_cgroup *memcg;
2348 memcg = stock->cached;
2349 if (memcg && stock->nr_pages &&
2350 mem_cgroup_is_descendant(memcg, root_memcg))
2352 if (obj_stock_flush_required(stock, root_memcg))
2357 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2359 drain_local_stock(&stock->work);
2361 schedule_work_on(cpu, &stock->work);
2365 mutex_unlock(&percpu_charge_mutex);
2368 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2370 struct memcg_stock_pcp *stock;
2371 struct mem_cgroup *memcg, *mi;
2373 stock = &per_cpu(memcg_stock, cpu);
2376 for_each_mem_cgroup(memcg) {
2379 for (i = 0; i < MEMCG_NR_STAT; i++) {
2383 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2385 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2386 atomic_long_add(x, &memcg->vmstats[i]);
2388 if (i >= NR_VM_NODE_STAT_ITEMS)
2391 for_each_node(nid) {
2392 struct mem_cgroup_per_node *pn;
2394 pn = mem_cgroup_nodeinfo(memcg, nid);
2395 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2398 atomic_long_add(x, &pn->lruvec_stat[i]);
2399 } while ((pn = parent_nodeinfo(pn, nid)));
2403 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2406 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2408 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2409 atomic_long_add(x, &memcg->vmevents[i]);
2416 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2417 unsigned int nr_pages,
2420 unsigned long nr_reclaimed = 0;
2423 unsigned long pflags;
2425 if (page_counter_read(&memcg->memory) <=
2426 READ_ONCE(memcg->memory.high))
2429 memcg_memory_event(memcg, MEMCG_HIGH);
2431 psi_memstall_enter(&pflags);
2432 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2434 psi_memstall_leave(&pflags);
2435 } while ((memcg = parent_mem_cgroup(memcg)) &&
2436 !mem_cgroup_is_root(memcg));
2438 return nr_reclaimed;
2441 static void high_work_func(struct work_struct *work)
2443 struct mem_cgroup *memcg;
2445 memcg = container_of(work, struct mem_cgroup, high_work);
2446 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2450 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2451 * enough to still cause a significant slowdown in most cases, while still
2452 * allowing diagnostics and tracing to proceed without becoming stuck.
2454 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2457 * When calculating the delay, we use these either side of the exponentiation to
2458 * maintain precision and scale to a reasonable number of jiffies (see the table
2461 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2462 * overage ratio to a delay.
2463 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2464 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2465 * to produce a reasonable delay curve.
2467 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2468 * reasonable delay curve compared to precision-adjusted overage, not
2469 * penalising heavily at first, but still making sure that growth beyond the
2470 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2471 * example, with a high of 100 megabytes:
2473 * +-------+------------------------+
2474 * | usage | time to allocate in ms |
2475 * +-------+------------------------+
2497 * +-------+------------------------+
2499 #define MEMCG_DELAY_PRECISION_SHIFT 20
2500 #define MEMCG_DELAY_SCALING_SHIFT 14
2502 static u64 calculate_overage(unsigned long usage, unsigned long high)
2510 * Prevent division by 0 in overage calculation by acting as if
2511 * it was a threshold of 1 page
2513 high = max(high, 1UL);
2515 overage = usage - high;
2516 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2517 return div64_u64(overage, high);
2520 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2522 u64 overage, max_overage = 0;
2525 overage = calculate_overage(page_counter_read(&memcg->memory),
2526 READ_ONCE(memcg->memory.high));
2527 max_overage = max(overage, max_overage);
2528 } while ((memcg = parent_mem_cgroup(memcg)) &&
2529 !mem_cgroup_is_root(memcg));
2534 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2536 u64 overage, max_overage = 0;
2539 overage = calculate_overage(page_counter_read(&memcg->swap),
2540 READ_ONCE(memcg->swap.high));
2542 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2543 max_overage = max(overage, max_overage);
2544 } while ((memcg = parent_mem_cgroup(memcg)) &&
2545 !mem_cgroup_is_root(memcg));
2551 * Get the number of jiffies that we should penalise a mischievous cgroup which
2552 * is exceeding its memory.high by checking both it and its ancestors.
2554 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2555 unsigned int nr_pages,
2558 unsigned long penalty_jiffies;
2564 * We use overage compared to memory.high to calculate the number of
2565 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2566 * fairly lenient on small overages, and increasingly harsh when the
2567 * memcg in question makes it clear that it has no intention of stopping
2568 * its crazy behaviour, so we exponentially increase the delay based on
2571 penalty_jiffies = max_overage * max_overage * HZ;
2572 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2573 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2576 * Factor in the task's own contribution to the overage, such that four
2577 * N-sized allocations are throttled approximately the same as one
2578 * 4N-sized allocation.
2580 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2581 * larger the current charge patch is than that.
2583 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2587 * Scheduled by try_charge() to be executed from the userland return path
2588 * and reclaims memory over the high limit.
2590 void mem_cgroup_handle_over_high(void)
2592 unsigned long penalty_jiffies;
2593 unsigned long pflags;
2594 unsigned long nr_reclaimed;
2595 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2596 int nr_retries = MAX_RECLAIM_RETRIES;
2597 struct mem_cgroup *memcg;
2598 bool in_retry = false;
2600 if (likely(!nr_pages))
2603 memcg = get_mem_cgroup_from_mm(current->mm);
2604 current->memcg_nr_pages_over_high = 0;
2608 * The allocating task should reclaim at least the batch size, but for
2609 * subsequent retries we only want to do what's necessary to prevent oom
2610 * or breaching resource isolation.
2612 * This is distinct from memory.max or page allocator behaviour because
2613 * memory.high is currently batched, whereas memory.max and the page
2614 * allocator run every time an allocation is made.
2616 nr_reclaimed = reclaim_high(memcg,
2617 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2621 * memory.high is breached and reclaim is unable to keep up. Throttle
2622 * allocators proactively to slow down excessive growth.
2624 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2625 mem_find_max_overage(memcg));
2627 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2628 swap_find_max_overage(memcg));
2631 * Clamp the max delay per usermode return so as to still keep the
2632 * application moving forwards and also permit diagnostics, albeit
2635 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2638 * Don't sleep if the amount of jiffies this memcg owes us is so low
2639 * that it's not even worth doing, in an attempt to be nice to those who
2640 * go only a small amount over their memory.high value and maybe haven't
2641 * been aggressively reclaimed enough yet.
2643 if (penalty_jiffies <= HZ / 100)
2647 * If reclaim is making forward progress but we're still over
2648 * memory.high, we want to encourage that rather than doing allocator
2651 if (nr_reclaimed || nr_retries--) {
2657 * If we exit early, we're guaranteed to die (since
2658 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2659 * need to account for any ill-begotten jiffies to pay them off later.
2661 psi_memstall_enter(&pflags);
2662 schedule_timeout_killable(penalty_jiffies);
2663 psi_memstall_leave(&pflags);
2666 css_put(&memcg->css);
2669 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2670 unsigned int nr_pages)
2672 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2673 int nr_retries = MAX_RECLAIM_RETRIES;
2674 struct mem_cgroup *mem_over_limit;
2675 struct page_counter *counter;
2676 enum oom_status oom_status;
2677 unsigned long nr_reclaimed;
2678 bool may_swap = true;
2679 bool drained = false;
2680 unsigned long pflags;
2682 if (mem_cgroup_is_root(memcg))
2685 if (consume_stock(memcg, nr_pages))
2688 if (!do_memsw_account() ||
2689 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2690 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2692 if (do_memsw_account())
2693 page_counter_uncharge(&memcg->memsw, batch);
2694 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2696 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2700 if (batch > nr_pages) {
2706 * Memcg doesn't have a dedicated reserve for atomic
2707 * allocations. But like the global atomic pool, we need to
2708 * put the burden of reclaim on regular allocation requests
2709 * and let these go through as privileged allocations.
2711 if (gfp_mask & __GFP_ATOMIC)
2715 * Unlike in global OOM situations, memcg is not in a physical
2716 * memory shortage. Allow dying and OOM-killed tasks to
2717 * bypass the last charges so that they can exit quickly and
2718 * free their memory.
2720 if (unlikely(should_force_charge()))
2724 * Prevent unbounded recursion when reclaim operations need to
2725 * allocate memory. This might exceed the limits temporarily,
2726 * but we prefer facilitating memory reclaim and getting back
2727 * under the limit over triggering OOM kills in these cases.
2729 if (unlikely(current->flags & PF_MEMALLOC))
2732 if (unlikely(task_in_memcg_oom(current)))
2735 if (!gfpflags_allow_blocking(gfp_mask))
2738 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2740 psi_memstall_enter(&pflags);
2741 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2742 gfp_mask, may_swap);
2743 psi_memstall_leave(&pflags);
2745 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2749 drain_all_stock(mem_over_limit);
2754 if (gfp_mask & __GFP_NORETRY)
2757 * Even though the limit is exceeded at this point, reclaim
2758 * may have been able to free some pages. Retry the charge
2759 * before killing the task.
2761 * Only for regular pages, though: huge pages are rather
2762 * unlikely to succeed so close to the limit, and we fall back
2763 * to regular pages anyway in case of failure.
2765 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2768 * At task move, charge accounts can be doubly counted. So, it's
2769 * better to wait until the end of task_move if something is going on.
2771 if (mem_cgroup_wait_acct_move(mem_over_limit))
2777 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2780 if (gfp_mask & __GFP_NOFAIL)
2783 if (fatal_signal_pending(current))
2787 * keep retrying as long as the memcg oom killer is able to make
2788 * a forward progress or bypass the charge if the oom killer
2789 * couldn't make any progress.
2791 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2792 get_order(nr_pages * PAGE_SIZE));
2793 switch (oom_status) {
2795 nr_retries = MAX_RECLAIM_RETRIES;
2803 if (!(gfp_mask & __GFP_NOFAIL))
2807 * The allocation either can't fail or will lead to more memory
2808 * being freed very soon. Allow memory usage go over the limit
2809 * temporarily by force charging it.
2811 page_counter_charge(&memcg->memory, nr_pages);
2812 if (do_memsw_account())
2813 page_counter_charge(&memcg->memsw, nr_pages);
2818 if (batch > nr_pages)
2819 refill_stock(memcg, batch - nr_pages);
2822 * If the hierarchy is above the normal consumption range, schedule
2823 * reclaim on returning to userland. We can perform reclaim here
2824 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2825 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2826 * not recorded as it most likely matches current's and won't
2827 * change in the meantime. As high limit is checked again before
2828 * reclaim, the cost of mismatch is negligible.
2831 bool mem_high, swap_high;
2833 mem_high = page_counter_read(&memcg->memory) >
2834 READ_ONCE(memcg->memory.high);
2835 swap_high = page_counter_read(&memcg->swap) >
2836 READ_ONCE(memcg->swap.high);
2838 /* Don't bother a random interrupted task */
2839 if (in_interrupt()) {
2841 schedule_work(&memcg->high_work);
2847 if (mem_high || swap_high) {
2849 * The allocating tasks in this cgroup will need to do
2850 * reclaim or be throttled to prevent further growth
2851 * of the memory or swap footprints.
2853 * Target some best-effort fairness between the tasks,
2854 * and distribute reclaim work and delay penalties
2855 * based on how much each task is actually allocating.
2857 current->memcg_nr_pages_over_high += batch;
2858 set_notify_resume(current);
2861 } while ((memcg = parent_mem_cgroup(memcg)));
2866 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2867 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2869 if (mem_cgroup_is_root(memcg))
2872 page_counter_uncharge(&memcg->memory, nr_pages);
2873 if (do_memsw_account())
2874 page_counter_uncharge(&memcg->memsw, nr_pages);
2878 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2880 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2882 * Any of the following ensures page->mem_cgroup stability:
2886 * - lock_page_memcg()
2887 * - exclusive reference
2889 page->mem_cgroup = memcg;
2892 #ifdef CONFIG_MEMCG_KMEM
2893 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2896 unsigned int objects = objs_per_slab_page(s, page);
2899 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2904 if (cmpxchg(&page->obj_cgroups, NULL,
2905 (struct obj_cgroup **) ((unsigned long)vec | 0x1UL)))
2908 kmemleak_not_leak(vec);
2914 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2916 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2917 * cgroup_mutex, etc.
2919 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2923 if (mem_cgroup_disabled())
2926 page = virt_to_head_page(p);
2929 * If page->mem_cgroup is set, it's either a simple mem_cgroup pointer
2930 * or a pointer to obj_cgroup vector. In the latter case the lowest
2931 * bit of the pointer is set.
2932 * The page->mem_cgroup pointer can be asynchronously changed
2933 * from NULL to (obj_cgroup_vec | 0x1UL), but can't be changed
2934 * from a valid memcg pointer to objcg vector or back.
2936 if (!page->mem_cgroup)
2940 * Slab objects are accounted individually, not per-page.
2941 * Memcg membership data for each individual object is saved in
2942 * the page->obj_cgroups.
2944 if (page_has_obj_cgroups(page)) {
2945 struct obj_cgroup *objcg;
2948 off = obj_to_index(page->slab_cache, page, p);
2949 objcg = page_obj_cgroups(page)[off];
2951 return obj_cgroup_memcg(objcg);
2956 /* All other pages use page->mem_cgroup */
2957 return page->mem_cgroup;
2960 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2962 struct obj_cgroup *objcg = NULL;
2963 struct mem_cgroup *memcg;
2965 if (memcg_kmem_bypass())
2969 if (unlikely(active_memcg()))
2970 memcg = active_memcg();
2972 memcg = mem_cgroup_from_task(current);
2974 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2975 objcg = rcu_dereference(memcg->objcg);
2976 if (objcg && obj_cgroup_tryget(objcg))
2985 static int memcg_alloc_cache_id(void)
2990 id = ida_simple_get(&memcg_cache_ida,
2991 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2995 if (id < memcg_nr_cache_ids)
2999 * There's no space for the new id in memcg_caches arrays,
3000 * so we have to grow them.
3002 down_write(&memcg_cache_ids_sem);
3004 size = 2 * (id + 1);
3005 if (size < MEMCG_CACHES_MIN_SIZE)
3006 size = MEMCG_CACHES_MIN_SIZE;
3007 else if (size > MEMCG_CACHES_MAX_SIZE)
3008 size = MEMCG_CACHES_MAX_SIZE;
3010 err = memcg_update_all_list_lrus(size);
3012 memcg_nr_cache_ids = size;
3014 up_write(&memcg_cache_ids_sem);
3017 ida_simple_remove(&memcg_cache_ida, id);
3023 static void memcg_free_cache_id(int id)
3025 ida_simple_remove(&memcg_cache_ida, id);
3029 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
3030 * @memcg: memory cgroup to charge
3031 * @gfp: reclaim mode
3032 * @nr_pages: number of pages to charge
3034 * Returns 0 on success, an error code on failure.
3036 int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
3037 unsigned int nr_pages)
3039 struct page_counter *counter;
3042 ret = try_charge(memcg, gfp, nr_pages);
3046 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3047 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3050 * Enforce __GFP_NOFAIL allocation because callers are not
3051 * prepared to see failures and likely do not have any failure
3054 if (gfp & __GFP_NOFAIL) {
3055 page_counter_charge(&memcg->kmem, nr_pages);
3058 cancel_charge(memcg, nr_pages);
3065 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
3066 * @memcg: memcg to uncharge
3067 * @nr_pages: number of pages to uncharge
3069 void __memcg_kmem_uncharge(struct mem_cgroup *memcg, unsigned int nr_pages)
3071 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3072 page_counter_uncharge(&memcg->kmem, nr_pages);
3074 page_counter_uncharge(&memcg->memory, nr_pages);
3075 if (do_memsw_account())
3076 page_counter_uncharge(&memcg->memsw, nr_pages);
3080 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3081 * @page: page to charge
3082 * @gfp: reclaim mode
3083 * @order: allocation order
3085 * Returns 0 on success, an error code on failure.
3087 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3089 struct mem_cgroup *memcg;
3092 memcg = get_mem_cgroup_from_current();
3093 if (memcg && !mem_cgroup_is_root(memcg)) {
3094 ret = __memcg_kmem_charge(memcg, gfp, 1 << order);
3096 page->mem_cgroup = memcg;
3097 __SetPageKmemcg(page);
3100 css_put(&memcg->css);
3106 * __memcg_kmem_uncharge_page: uncharge a kmem page
3107 * @page: page to uncharge
3108 * @order: allocation order
3110 void __memcg_kmem_uncharge_page(struct page *page, int order)
3112 struct mem_cgroup *memcg = page->mem_cgroup;
3113 unsigned int nr_pages = 1 << order;
3118 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3119 __memcg_kmem_uncharge(memcg, nr_pages);
3120 page->mem_cgroup = NULL;
3121 css_put(&memcg->css);
3123 /* slab pages do not have PageKmemcg flag set */
3124 if (PageKmemcg(page))
3125 __ClearPageKmemcg(page);
3128 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3130 struct memcg_stock_pcp *stock;
3131 unsigned long flags;
3134 local_irq_save(flags);
3136 stock = this_cpu_ptr(&memcg_stock);
3137 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3138 stock->nr_bytes -= nr_bytes;
3142 local_irq_restore(flags);
3147 static void drain_obj_stock(struct memcg_stock_pcp *stock)
3149 struct obj_cgroup *old = stock->cached_objcg;
3154 if (stock->nr_bytes) {
3155 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3156 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3160 __memcg_kmem_uncharge(obj_cgroup_memcg(old), nr_pages);
3165 * The leftover is flushed to the centralized per-memcg value.
3166 * On the next attempt to refill obj stock it will be moved
3167 * to a per-cpu stock (probably, on an other CPU), see
3168 * refill_obj_stock().
3170 * How often it's flushed is a trade-off between the memory
3171 * limit enforcement accuracy and potential CPU contention,
3172 * so it might be changed in the future.
3174 atomic_add(nr_bytes, &old->nr_charged_bytes);
3175 stock->nr_bytes = 0;
3178 obj_cgroup_put(old);
3179 stock->cached_objcg = NULL;
3182 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3183 struct mem_cgroup *root_memcg)
3185 struct mem_cgroup *memcg;
3187 if (stock->cached_objcg) {
3188 memcg = obj_cgroup_memcg(stock->cached_objcg);
3189 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3196 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3198 struct memcg_stock_pcp *stock;
3199 unsigned long flags;
3201 local_irq_save(flags);
3203 stock = this_cpu_ptr(&memcg_stock);
3204 if (stock->cached_objcg != objcg) { /* reset if necessary */
3205 drain_obj_stock(stock);
3206 obj_cgroup_get(objcg);
3207 stock->cached_objcg = objcg;
3208 stock->nr_bytes = atomic_xchg(&objcg->nr_charged_bytes, 0);
3210 stock->nr_bytes += nr_bytes;
3212 if (stock->nr_bytes > PAGE_SIZE)
3213 drain_obj_stock(stock);
3215 local_irq_restore(flags);
3218 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3220 struct mem_cgroup *memcg;
3221 unsigned int nr_pages, nr_bytes;
3224 if (consume_obj_stock(objcg, size))
3228 * In theory, memcg->nr_charged_bytes can have enough
3229 * pre-charged bytes to satisfy the allocation. However,
3230 * flushing memcg->nr_charged_bytes requires two atomic
3231 * operations, and memcg->nr_charged_bytes can't be big,
3232 * so it's better to ignore it and try grab some new pages.
3233 * memcg->nr_charged_bytes will be flushed in
3234 * refill_obj_stock(), called from this function or
3235 * independently later.
3239 memcg = obj_cgroup_memcg(objcg);
3240 if (unlikely(!css_tryget(&memcg->css)))
3244 nr_pages = size >> PAGE_SHIFT;
3245 nr_bytes = size & (PAGE_SIZE - 1);
3250 ret = __memcg_kmem_charge(memcg, gfp, nr_pages);
3251 if (!ret && nr_bytes)
3252 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes);
3254 css_put(&memcg->css);
3258 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3260 refill_obj_stock(objcg, size);
3263 #endif /* CONFIG_MEMCG_KMEM */
3265 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3268 * Because tail pages are not marked as "used", set it. We're under
3269 * pgdat->lru_lock and migration entries setup in all page mappings.
3271 void mem_cgroup_split_huge_fixup(struct page *head)
3273 struct mem_cgroup *memcg = head->mem_cgroup;
3276 if (mem_cgroup_disabled())
3279 for (i = 1; i < HPAGE_PMD_NR; i++) {
3280 css_get(&memcg->css);
3281 head[i].mem_cgroup = memcg;
3284 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3286 #ifdef CONFIG_MEMCG_SWAP
3288 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3289 * @entry: swap entry to be moved
3290 * @from: mem_cgroup which the entry is moved from
3291 * @to: mem_cgroup which the entry is moved to
3293 * It succeeds only when the swap_cgroup's record for this entry is the same
3294 * as the mem_cgroup's id of @from.
3296 * Returns 0 on success, -EINVAL on failure.
3298 * The caller must have charged to @to, IOW, called page_counter_charge() about
3299 * both res and memsw, and called css_get().
3301 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3302 struct mem_cgroup *from, struct mem_cgroup *to)
3304 unsigned short old_id, new_id;
3306 old_id = mem_cgroup_id(from);
3307 new_id = mem_cgroup_id(to);
3309 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3310 mod_memcg_state(from, MEMCG_SWAP, -1);
3311 mod_memcg_state(to, MEMCG_SWAP, 1);
3317 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3318 struct mem_cgroup *from, struct mem_cgroup *to)
3324 static DEFINE_MUTEX(memcg_max_mutex);
3326 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3327 unsigned long max, bool memsw)
3329 bool enlarge = false;
3330 bool drained = false;
3332 bool limits_invariant;
3333 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3336 if (signal_pending(current)) {
3341 mutex_lock(&memcg_max_mutex);
3343 * Make sure that the new limit (memsw or memory limit) doesn't
3344 * break our basic invariant rule memory.max <= memsw.max.
3346 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3347 max <= memcg->memsw.max;
3348 if (!limits_invariant) {
3349 mutex_unlock(&memcg_max_mutex);
3353 if (max > counter->max)
3355 ret = page_counter_set_max(counter, max);
3356 mutex_unlock(&memcg_max_mutex);
3362 drain_all_stock(memcg);
3367 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3368 GFP_KERNEL, !memsw)) {
3374 if (!ret && enlarge)
3375 memcg_oom_recover(memcg);
3380 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3382 unsigned long *total_scanned)
3384 unsigned long nr_reclaimed = 0;
3385 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3386 unsigned long reclaimed;
3388 struct mem_cgroup_tree_per_node *mctz;
3389 unsigned long excess;
3390 unsigned long nr_scanned;
3395 mctz = soft_limit_tree_node(pgdat->node_id);
3398 * Do not even bother to check the largest node if the root
3399 * is empty. Do it lockless to prevent lock bouncing. Races
3400 * are acceptable as soft limit is best effort anyway.
3402 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3406 * This loop can run a while, specially if mem_cgroup's continuously
3407 * keep exceeding their soft limit and putting the system under
3414 mz = mem_cgroup_largest_soft_limit_node(mctz);
3419 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3420 gfp_mask, &nr_scanned);
3421 nr_reclaimed += reclaimed;
3422 *total_scanned += nr_scanned;
3423 spin_lock_irq(&mctz->lock);
3424 __mem_cgroup_remove_exceeded(mz, mctz);
3427 * If we failed to reclaim anything from this memory cgroup
3428 * it is time to move on to the next cgroup
3432 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3434 excess = soft_limit_excess(mz->memcg);
3436 * One school of thought says that we should not add
3437 * back the node to the tree if reclaim returns 0.
3438 * But our reclaim could return 0, simply because due
3439 * to priority we are exposing a smaller subset of
3440 * memory to reclaim from. Consider this as a longer
3443 /* If excess == 0, no tree ops */
3444 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3445 spin_unlock_irq(&mctz->lock);
3446 css_put(&mz->memcg->css);
3449 * Could not reclaim anything and there are no more
3450 * mem cgroups to try or we seem to be looping without
3451 * reclaiming anything.
3453 if (!nr_reclaimed &&
3455 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3457 } while (!nr_reclaimed);
3459 css_put(&next_mz->memcg->css);
3460 return nr_reclaimed;
3464 * Test whether @memcg has children, dead or alive. Note that this
3465 * function doesn't care whether @memcg has use_hierarchy enabled and
3466 * returns %true if there are child csses according to the cgroup
3467 * hierarchy. Testing use_hierarchy is the caller's responsibility.
3469 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3474 ret = css_next_child(NULL, &memcg->css);
3480 * Reclaims as many pages from the given memcg as possible.
3482 * Caller is responsible for holding css reference for memcg.
3484 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3486 int nr_retries = MAX_RECLAIM_RETRIES;
3488 /* we call try-to-free pages for make this cgroup empty */
3489 lru_add_drain_all();
3491 drain_all_stock(memcg);
3493 /* try to free all pages in this cgroup */
3494 while (nr_retries && page_counter_read(&memcg->memory)) {
3497 if (signal_pending(current))
3500 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3504 /* maybe some writeback is necessary */
3505 congestion_wait(BLK_RW_ASYNC, HZ/10);
3513 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3514 char *buf, size_t nbytes,
3517 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3519 if (mem_cgroup_is_root(memcg))
3521 return mem_cgroup_force_empty(memcg) ?: nbytes;
3524 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3527 return mem_cgroup_from_css(css)->use_hierarchy;
3530 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3531 struct cftype *cft, u64 val)
3534 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3535 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3537 if (memcg->use_hierarchy == val)
3541 * If parent's use_hierarchy is set, we can't make any modifications
3542 * in the child subtrees. If it is unset, then the change can
3543 * occur, provided the current cgroup has no children.
3545 * For the root cgroup, parent_mem is NULL, we allow value to be
3546 * set if there are no children.
3548 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3549 (val == 1 || val == 0)) {
3550 if (!memcg_has_children(memcg))
3551 memcg->use_hierarchy = val;
3560 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3564 if (mem_cgroup_is_root(memcg)) {
3565 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3566 memcg_page_state(memcg, NR_ANON_MAPPED);
3568 val += memcg_page_state(memcg, MEMCG_SWAP);
3571 val = page_counter_read(&memcg->memory);
3573 val = page_counter_read(&memcg->memsw);
3586 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3589 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3590 struct page_counter *counter;
3592 switch (MEMFILE_TYPE(cft->private)) {
3594 counter = &memcg->memory;
3597 counter = &memcg->memsw;
3600 counter = &memcg->kmem;
3603 counter = &memcg->tcpmem;
3609 switch (MEMFILE_ATTR(cft->private)) {
3611 if (counter == &memcg->memory)
3612 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3613 if (counter == &memcg->memsw)
3614 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3615 return (u64)page_counter_read(counter) * PAGE_SIZE;
3617 return (u64)counter->max * PAGE_SIZE;
3619 return (u64)counter->watermark * PAGE_SIZE;
3621 return counter->failcnt;
3622 case RES_SOFT_LIMIT:
3623 return (u64)memcg->soft_limit * PAGE_SIZE;
3629 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3631 unsigned long stat[MEMCG_NR_STAT] = {0};
3632 struct mem_cgroup *mi;
3635 for_each_online_cpu(cpu)
3636 for (i = 0; i < MEMCG_NR_STAT; i++)
3637 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3639 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3640 for (i = 0; i < MEMCG_NR_STAT; i++)
3641 atomic_long_add(stat[i], &mi->vmstats[i]);
3643 for_each_node(node) {
3644 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3645 struct mem_cgroup_per_node *pi;
3647 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3650 for_each_online_cpu(cpu)
3651 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3653 pn->lruvec_stat_cpu->count[i], cpu);
3655 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3656 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3657 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3661 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3663 unsigned long events[NR_VM_EVENT_ITEMS];
3664 struct mem_cgroup *mi;
3667 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3670 for_each_online_cpu(cpu)
3671 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3672 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3675 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3676 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3677 atomic_long_add(events[i], &mi->vmevents[i]);
3680 #ifdef CONFIG_MEMCG_KMEM
3681 static int memcg_online_kmem(struct mem_cgroup *memcg)
3683 struct obj_cgroup *objcg;
3686 if (cgroup_memory_nokmem)
3689 BUG_ON(memcg->kmemcg_id >= 0);
3690 BUG_ON(memcg->kmem_state);
3692 memcg_id = memcg_alloc_cache_id();
3696 objcg = obj_cgroup_alloc();
3698 memcg_free_cache_id(memcg_id);
3701 objcg->memcg = memcg;
3702 rcu_assign_pointer(memcg->objcg, objcg);
3704 static_branch_enable(&memcg_kmem_enabled_key);
3707 * A memory cgroup is considered kmem-online as soon as it gets
3708 * kmemcg_id. Setting the id after enabling static branching will
3709 * guarantee no one starts accounting before all call sites are
3712 memcg->kmemcg_id = memcg_id;
3713 memcg->kmem_state = KMEM_ONLINE;
3718 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3720 struct cgroup_subsys_state *css;
3721 struct mem_cgroup *parent, *child;
3724 if (memcg->kmem_state != KMEM_ONLINE)
3727 memcg->kmem_state = KMEM_ALLOCATED;
3729 parent = parent_mem_cgroup(memcg);
3731 parent = root_mem_cgroup;
3733 memcg_reparent_objcgs(memcg, parent);
3735 kmemcg_id = memcg->kmemcg_id;
3736 BUG_ON(kmemcg_id < 0);
3739 * Change kmemcg_id of this cgroup and all its descendants to the
3740 * parent's id, and then move all entries from this cgroup's list_lrus
3741 * to ones of the parent. After we have finished, all list_lrus
3742 * corresponding to this cgroup are guaranteed to remain empty. The
3743 * ordering is imposed by list_lru_node->lock taken by
3744 * memcg_drain_all_list_lrus().
3746 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3747 css_for_each_descendant_pre(css, &memcg->css) {
3748 child = mem_cgroup_from_css(css);
3749 BUG_ON(child->kmemcg_id != kmemcg_id);
3750 child->kmemcg_id = parent->kmemcg_id;
3751 if (!memcg->use_hierarchy)
3756 memcg_drain_all_list_lrus(kmemcg_id, parent);
3758 memcg_free_cache_id(kmemcg_id);
3761 static void memcg_free_kmem(struct mem_cgroup *memcg)
3763 /* css_alloc() failed, offlining didn't happen */
3764 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3765 memcg_offline_kmem(memcg);
3768 static int memcg_online_kmem(struct mem_cgroup *memcg)
3772 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3775 static void memcg_free_kmem(struct mem_cgroup *memcg)
3778 #endif /* CONFIG_MEMCG_KMEM */
3780 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3785 mutex_lock(&memcg_max_mutex);
3786 ret = page_counter_set_max(&memcg->kmem, max);
3787 mutex_unlock(&memcg_max_mutex);
3791 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3795 mutex_lock(&memcg_max_mutex);
3797 ret = page_counter_set_max(&memcg->tcpmem, max);
3801 if (!memcg->tcpmem_active) {
3803 * The active flag needs to be written after the static_key
3804 * update. This is what guarantees that the socket activation
3805 * function is the last one to run. See mem_cgroup_sk_alloc()
3806 * for details, and note that we don't mark any socket as
3807 * belonging to this memcg until that flag is up.
3809 * We need to do this, because static_keys will span multiple
3810 * sites, but we can't control their order. If we mark a socket
3811 * as accounted, but the accounting functions are not patched in
3812 * yet, we'll lose accounting.
3814 * We never race with the readers in mem_cgroup_sk_alloc(),
3815 * because when this value change, the code to process it is not
3818 static_branch_inc(&memcg_sockets_enabled_key);
3819 memcg->tcpmem_active = true;
3822 mutex_unlock(&memcg_max_mutex);
3827 * The user of this function is...
3830 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3831 char *buf, size_t nbytes, loff_t off)
3833 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3834 unsigned long nr_pages;
3837 buf = strstrip(buf);
3838 ret = page_counter_memparse(buf, "-1", &nr_pages);
3842 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3844 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3848 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3850 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3853 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3856 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3857 "Please report your usecase to linux-mm@kvack.org if you "
3858 "depend on this functionality.\n");
3859 ret = memcg_update_kmem_max(memcg, nr_pages);
3862 ret = memcg_update_tcp_max(memcg, nr_pages);
3866 case RES_SOFT_LIMIT:
3867 memcg->soft_limit = nr_pages;
3871 return ret ?: nbytes;
3874 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3875 size_t nbytes, loff_t off)
3877 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3878 struct page_counter *counter;
3880 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3882 counter = &memcg->memory;
3885 counter = &memcg->memsw;
3888 counter = &memcg->kmem;
3891 counter = &memcg->tcpmem;
3897 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3899 page_counter_reset_watermark(counter);
3902 counter->failcnt = 0;
3911 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3914 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3918 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3919 struct cftype *cft, u64 val)
3921 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3923 if (val & ~MOVE_MASK)
3927 * No kind of locking is needed in here, because ->can_attach() will
3928 * check this value once in the beginning of the process, and then carry
3929 * on with stale data. This means that changes to this value will only
3930 * affect task migrations starting after the change.
3932 memcg->move_charge_at_immigrate = val;
3936 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3937 struct cftype *cft, u64 val)
3945 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3946 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3947 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3949 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3950 int nid, unsigned int lru_mask, bool tree)
3952 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3953 unsigned long nr = 0;
3956 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3959 if (!(BIT(lru) & lru_mask))
3962 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3964 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3969 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3970 unsigned int lru_mask,
3973 unsigned long nr = 0;
3977 if (!(BIT(lru) & lru_mask))
3980 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3982 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3987 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3991 unsigned int lru_mask;
3994 static const struct numa_stat stats[] = {
3995 { "total", LRU_ALL },
3996 { "file", LRU_ALL_FILE },
3997 { "anon", LRU_ALL_ANON },
3998 { "unevictable", BIT(LRU_UNEVICTABLE) },
4000 const struct numa_stat *stat;
4002 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4004 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4005 seq_printf(m, "%s=%lu", stat->name,
4006 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4008 for_each_node_state(nid, N_MEMORY)
4009 seq_printf(m, " N%d=%lu", nid,
4010 mem_cgroup_node_nr_lru_pages(memcg, nid,
4011 stat->lru_mask, false));
4015 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4017 seq_printf(m, "hierarchical_%s=%lu", stat->name,
4018 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4020 for_each_node_state(nid, N_MEMORY)
4021 seq_printf(m, " N%d=%lu", nid,
4022 mem_cgroup_node_nr_lru_pages(memcg, nid,
4023 stat->lru_mask, true));
4029 #endif /* CONFIG_NUMA */
4031 static const unsigned int memcg1_stats[] = {
4034 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4044 static const char *const memcg1_stat_names[] = {
4047 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4057 /* Universal VM events cgroup1 shows, original sort order */
4058 static const unsigned int memcg1_events[] = {
4065 static int memcg_stat_show(struct seq_file *m, void *v)
4067 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4068 unsigned long memory, memsw;
4069 struct mem_cgroup *mi;
4072 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4074 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4077 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4079 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4080 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4081 if (memcg1_stats[i] == NR_ANON_THPS)
4084 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4087 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4088 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4089 memcg_events_local(memcg, memcg1_events[i]));
4091 for (i = 0; i < NR_LRU_LISTS; i++)
4092 seq_printf(m, "%s %lu\n", lru_list_name(i),
4093 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4096 /* Hierarchical information */
4097 memory = memsw = PAGE_COUNTER_MAX;
4098 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4099 memory = min(memory, READ_ONCE(mi->memory.max));
4100 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4102 seq_printf(m, "hierarchical_memory_limit %llu\n",
4103 (u64)memory * PAGE_SIZE);
4104 if (do_memsw_account())
4105 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4106 (u64)memsw * PAGE_SIZE);
4108 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4111 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4113 nr = memcg_page_state(memcg, memcg1_stats[i]);
4114 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4115 if (memcg1_stats[i] == NR_ANON_THPS)
4118 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4119 (u64)nr * PAGE_SIZE);
4122 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4123 seq_printf(m, "total_%s %llu\n",
4124 vm_event_name(memcg1_events[i]),
4125 (u64)memcg_events(memcg, memcg1_events[i]));
4127 for (i = 0; i < NR_LRU_LISTS; i++)
4128 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4129 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4132 #ifdef CONFIG_DEBUG_VM
4135 struct mem_cgroup_per_node *mz;
4136 unsigned long anon_cost = 0;
4137 unsigned long file_cost = 0;
4139 for_each_online_pgdat(pgdat) {
4140 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
4142 anon_cost += mz->lruvec.anon_cost;
4143 file_cost += mz->lruvec.file_cost;
4145 seq_printf(m, "anon_cost %lu\n", anon_cost);
4146 seq_printf(m, "file_cost %lu\n", file_cost);
4153 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4156 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4158 return mem_cgroup_swappiness(memcg);
4161 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4162 struct cftype *cft, u64 val)
4164 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4170 memcg->swappiness = val;
4172 vm_swappiness = val;
4177 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4179 struct mem_cgroup_threshold_ary *t;
4180 unsigned long usage;
4185 t = rcu_dereference(memcg->thresholds.primary);
4187 t = rcu_dereference(memcg->memsw_thresholds.primary);
4192 usage = mem_cgroup_usage(memcg, swap);
4195 * current_threshold points to threshold just below or equal to usage.
4196 * If it's not true, a threshold was crossed after last
4197 * call of __mem_cgroup_threshold().
4199 i = t->current_threshold;
4202 * Iterate backward over array of thresholds starting from
4203 * current_threshold and check if a threshold is crossed.
4204 * If none of thresholds below usage is crossed, we read
4205 * only one element of the array here.
4207 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4208 eventfd_signal(t->entries[i].eventfd, 1);
4210 /* i = current_threshold + 1 */
4214 * Iterate forward over array of thresholds starting from
4215 * current_threshold+1 and check if a threshold is crossed.
4216 * If none of thresholds above usage is crossed, we read
4217 * only one element of the array here.
4219 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4220 eventfd_signal(t->entries[i].eventfd, 1);
4222 /* Update current_threshold */
4223 t->current_threshold = i - 1;
4228 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4231 __mem_cgroup_threshold(memcg, false);
4232 if (do_memsw_account())
4233 __mem_cgroup_threshold(memcg, true);
4235 memcg = parent_mem_cgroup(memcg);
4239 static int compare_thresholds(const void *a, const void *b)
4241 const struct mem_cgroup_threshold *_a = a;
4242 const struct mem_cgroup_threshold *_b = b;
4244 if (_a->threshold > _b->threshold)
4247 if (_a->threshold < _b->threshold)
4253 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4255 struct mem_cgroup_eventfd_list *ev;
4257 spin_lock(&memcg_oom_lock);
4259 list_for_each_entry(ev, &memcg->oom_notify, list)
4260 eventfd_signal(ev->eventfd, 1);
4262 spin_unlock(&memcg_oom_lock);
4266 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4268 struct mem_cgroup *iter;
4270 for_each_mem_cgroup_tree(iter, memcg)
4271 mem_cgroup_oom_notify_cb(iter);
4274 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4275 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4277 struct mem_cgroup_thresholds *thresholds;
4278 struct mem_cgroup_threshold_ary *new;
4279 unsigned long threshold;
4280 unsigned long usage;
4283 ret = page_counter_memparse(args, "-1", &threshold);
4287 mutex_lock(&memcg->thresholds_lock);
4290 thresholds = &memcg->thresholds;
4291 usage = mem_cgroup_usage(memcg, false);
4292 } else if (type == _MEMSWAP) {
4293 thresholds = &memcg->memsw_thresholds;
4294 usage = mem_cgroup_usage(memcg, true);
4298 /* Check if a threshold crossed before adding a new one */
4299 if (thresholds->primary)
4300 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4302 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4304 /* Allocate memory for new array of thresholds */
4305 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4312 /* Copy thresholds (if any) to new array */
4313 if (thresholds->primary)
4314 memcpy(new->entries, thresholds->primary->entries,
4315 flex_array_size(new, entries, size - 1));
4317 /* Add new threshold */
4318 new->entries[size - 1].eventfd = eventfd;
4319 new->entries[size - 1].threshold = threshold;
4321 /* Sort thresholds. Registering of new threshold isn't time-critical */
4322 sort(new->entries, size, sizeof(*new->entries),
4323 compare_thresholds, NULL);
4325 /* Find current threshold */
4326 new->current_threshold = -1;
4327 for (i = 0; i < size; i++) {
4328 if (new->entries[i].threshold <= usage) {
4330 * new->current_threshold will not be used until
4331 * rcu_assign_pointer(), so it's safe to increment
4334 ++new->current_threshold;
4339 /* Free old spare buffer and save old primary buffer as spare */
4340 kfree(thresholds->spare);
4341 thresholds->spare = thresholds->primary;
4343 rcu_assign_pointer(thresholds->primary, new);
4345 /* To be sure that nobody uses thresholds */
4349 mutex_unlock(&memcg->thresholds_lock);
4354 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4355 struct eventfd_ctx *eventfd, const char *args)
4357 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4360 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4361 struct eventfd_ctx *eventfd, const char *args)
4363 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4366 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4367 struct eventfd_ctx *eventfd, enum res_type type)
4369 struct mem_cgroup_thresholds *thresholds;
4370 struct mem_cgroup_threshold_ary *new;
4371 unsigned long usage;
4372 int i, j, size, entries;
4374 mutex_lock(&memcg->thresholds_lock);
4377 thresholds = &memcg->thresholds;
4378 usage = mem_cgroup_usage(memcg, false);
4379 } else if (type == _MEMSWAP) {
4380 thresholds = &memcg->memsw_thresholds;
4381 usage = mem_cgroup_usage(memcg, true);
4385 if (!thresholds->primary)
4388 /* Check if a threshold crossed before removing */
4389 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4391 /* Calculate new number of threshold */
4393 for (i = 0; i < thresholds->primary->size; i++) {
4394 if (thresholds->primary->entries[i].eventfd != eventfd)
4400 new = thresholds->spare;
4402 /* If no items related to eventfd have been cleared, nothing to do */
4406 /* Set thresholds array to NULL if we don't have thresholds */
4415 /* Copy thresholds and find current threshold */
4416 new->current_threshold = -1;
4417 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4418 if (thresholds->primary->entries[i].eventfd == eventfd)
4421 new->entries[j] = thresholds->primary->entries[i];
4422 if (new->entries[j].threshold <= usage) {
4424 * new->current_threshold will not be used
4425 * until rcu_assign_pointer(), so it's safe to increment
4428 ++new->current_threshold;
4434 /* Swap primary and spare array */
4435 thresholds->spare = thresholds->primary;
4437 rcu_assign_pointer(thresholds->primary, new);
4439 /* To be sure that nobody uses thresholds */
4442 /* If all events are unregistered, free the spare array */
4444 kfree(thresholds->spare);
4445 thresholds->spare = NULL;
4448 mutex_unlock(&memcg->thresholds_lock);
4451 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4452 struct eventfd_ctx *eventfd)
4454 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4457 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4458 struct eventfd_ctx *eventfd)
4460 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4463 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4464 struct eventfd_ctx *eventfd, const char *args)
4466 struct mem_cgroup_eventfd_list *event;
4468 event = kmalloc(sizeof(*event), GFP_KERNEL);
4472 spin_lock(&memcg_oom_lock);
4474 event->eventfd = eventfd;
4475 list_add(&event->list, &memcg->oom_notify);
4477 /* already in OOM ? */
4478 if (memcg->under_oom)
4479 eventfd_signal(eventfd, 1);
4480 spin_unlock(&memcg_oom_lock);
4485 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4486 struct eventfd_ctx *eventfd)
4488 struct mem_cgroup_eventfd_list *ev, *tmp;
4490 spin_lock(&memcg_oom_lock);
4492 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4493 if (ev->eventfd == eventfd) {
4494 list_del(&ev->list);
4499 spin_unlock(&memcg_oom_lock);
4502 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4504 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4506 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4507 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4508 seq_printf(sf, "oom_kill %lu\n",
4509 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4513 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4514 struct cftype *cft, u64 val)
4516 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4518 /* cannot set to root cgroup and only 0 and 1 are allowed */
4519 if (!css->parent || !((val == 0) || (val == 1)))
4522 memcg->oom_kill_disable = val;
4524 memcg_oom_recover(memcg);
4529 #ifdef CONFIG_CGROUP_WRITEBACK
4531 #include <trace/events/writeback.h>
4533 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4535 return wb_domain_init(&memcg->cgwb_domain, gfp);
4538 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4540 wb_domain_exit(&memcg->cgwb_domain);
4543 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4545 wb_domain_size_changed(&memcg->cgwb_domain);
4548 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4550 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4552 if (!memcg->css.parent)
4555 return &memcg->cgwb_domain;
4559 * idx can be of type enum memcg_stat_item or node_stat_item.
4560 * Keep in sync with memcg_exact_page().
4562 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4564 long x = atomic_long_read(&memcg->vmstats[idx]);
4567 for_each_online_cpu(cpu)
4568 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4575 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4576 * @wb: bdi_writeback in question
4577 * @pfilepages: out parameter for number of file pages
4578 * @pheadroom: out parameter for number of allocatable pages according to memcg
4579 * @pdirty: out parameter for number of dirty pages
4580 * @pwriteback: out parameter for number of pages under writeback
4582 * Determine the numbers of file, headroom, dirty, and writeback pages in
4583 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4584 * is a bit more involved.
4586 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4587 * headroom is calculated as the lowest headroom of itself and the
4588 * ancestors. Note that this doesn't consider the actual amount of
4589 * available memory in the system. The caller should further cap
4590 * *@pheadroom accordingly.
4592 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4593 unsigned long *pheadroom, unsigned long *pdirty,
4594 unsigned long *pwriteback)
4596 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4597 struct mem_cgroup *parent;
4599 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4601 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4602 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4603 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4604 *pheadroom = PAGE_COUNTER_MAX;
4606 while ((parent = parent_mem_cgroup(memcg))) {
4607 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4608 READ_ONCE(memcg->memory.high));
4609 unsigned long used = page_counter_read(&memcg->memory);
4611 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4617 * Foreign dirty flushing
4619 * There's an inherent mismatch between memcg and writeback. The former
4620 * trackes ownership per-page while the latter per-inode. This was a
4621 * deliberate design decision because honoring per-page ownership in the
4622 * writeback path is complicated, may lead to higher CPU and IO overheads
4623 * and deemed unnecessary given that write-sharing an inode across
4624 * different cgroups isn't a common use-case.
4626 * Combined with inode majority-writer ownership switching, this works well
4627 * enough in most cases but there are some pathological cases. For
4628 * example, let's say there are two cgroups A and B which keep writing to
4629 * different but confined parts of the same inode. B owns the inode and
4630 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4631 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4632 * triggering background writeback. A will be slowed down without a way to
4633 * make writeback of the dirty pages happen.
4635 * Conditions like the above can lead to a cgroup getting repatedly and
4636 * severely throttled after making some progress after each
4637 * dirty_expire_interval while the underyling IO device is almost
4640 * Solving this problem completely requires matching the ownership tracking
4641 * granularities between memcg and writeback in either direction. However,
4642 * the more egregious behaviors can be avoided by simply remembering the
4643 * most recent foreign dirtying events and initiating remote flushes on
4644 * them when local writeback isn't enough to keep the memory clean enough.
4646 * The following two functions implement such mechanism. When a foreign
4647 * page - a page whose memcg and writeback ownerships don't match - is
4648 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4649 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4650 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4651 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4652 * foreign bdi_writebacks which haven't expired. Both the numbers of
4653 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4654 * limited to MEMCG_CGWB_FRN_CNT.
4656 * The mechanism only remembers IDs and doesn't hold any object references.
4657 * As being wrong occasionally doesn't matter, updates and accesses to the
4658 * records are lockless and racy.
4660 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4661 struct bdi_writeback *wb)
4663 struct mem_cgroup *memcg = page->mem_cgroup;
4664 struct memcg_cgwb_frn *frn;
4665 u64 now = get_jiffies_64();
4666 u64 oldest_at = now;
4670 trace_track_foreign_dirty(page, wb);
4673 * Pick the slot to use. If there is already a slot for @wb, keep
4674 * using it. If not replace the oldest one which isn't being
4677 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4678 frn = &memcg->cgwb_frn[i];
4679 if (frn->bdi_id == wb->bdi->id &&
4680 frn->memcg_id == wb->memcg_css->id)
4682 if (time_before64(frn->at, oldest_at) &&
4683 atomic_read(&frn->done.cnt) == 1) {
4685 oldest_at = frn->at;
4689 if (i < MEMCG_CGWB_FRN_CNT) {
4691 * Re-using an existing one. Update timestamp lazily to
4692 * avoid making the cacheline hot. We want them to be
4693 * reasonably up-to-date and significantly shorter than
4694 * dirty_expire_interval as that's what expires the record.
4695 * Use the shorter of 1s and dirty_expire_interval / 8.
4697 unsigned long update_intv =
4698 min_t(unsigned long, HZ,
4699 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4701 if (time_before64(frn->at, now - update_intv))
4703 } else if (oldest >= 0) {
4704 /* replace the oldest free one */
4705 frn = &memcg->cgwb_frn[oldest];
4706 frn->bdi_id = wb->bdi->id;
4707 frn->memcg_id = wb->memcg_css->id;
4712 /* issue foreign writeback flushes for recorded foreign dirtying events */
4713 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4715 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4716 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4717 u64 now = jiffies_64;
4720 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4721 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4724 * If the record is older than dirty_expire_interval,
4725 * writeback on it has already started. No need to kick it
4726 * off again. Also, don't start a new one if there's
4727 * already one in flight.
4729 if (time_after64(frn->at, now - intv) &&
4730 atomic_read(&frn->done.cnt) == 1) {
4732 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4733 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4734 WB_REASON_FOREIGN_FLUSH,
4740 #else /* CONFIG_CGROUP_WRITEBACK */
4742 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4747 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4751 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4755 #endif /* CONFIG_CGROUP_WRITEBACK */
4758 * DO NOT USE IN NEW FILES.
4760 * "cgroup.event_control" implementation.
4762 * This is way over-engineered. It tries to support fully configurable
4763 * events for each user. Such level of flexibility is completely
4764 * unnecessary especially in the light of the planned unified hierarchy.
4766 * Please deprecate this and replace with something simpler if at all
4771 * Unregister event and free resources.
4773 * Gets called from workqueue.
4775 static void memcg_event_remove(struct work_struct *work)
4777 struct mem_cgroup_event *event =
4778 container_of(work, struct mem_cgroup_event, remove);
4779 struct mem_cgroup *memcg = event->memcg;
4781 remove_wait_queue(event->wqh, &event->wait);
4783 event->unregister_event(memcg, event->eventfd);
4785 /* Notify userspace the event is going away. */
4786 eventfd_signal(event->eventfd, 1);
4788 eventfd_ctx_put(event->eventfd);
4790 css_put(&memcg->css);
4794 * Gets called on EPOLLHUP on eventfd when user closes it.
4796 * Called with wqh->lock held and interrupts disabled.
4798 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4799 int sync, void *key)
4801 struct mem_cgroup_event *event =
4802 container_of(wait, struct mem_cgroup_event, wait);
4803 struct mem_cgroup *memcg = event->memcg;
4804 __poll_t flags = key_to_poll(key);
4806 if (flags & EPOLLHUP) {
4808 * If the event has been detached at cgroup removal, we
4809 * can simply return knowing the other side will cleanup
4812 * We can't race against event freeing since the other
4813 * side will require wqh->lock via remove_wait_queue(),
4816 spin_lock(&memcg->event_list_lock);
4817 if (!list_empty(&event->list)) {
4818 list_del_init(&event->list);
4820 * We are in atomic context, but cgroup_event_remove()
4821 * may sleep, so we have to call it in workqueue.
4823 schedule_work(&event->remove);
4825 spin_unlock(&memcg->event_list_lock);
4831 static void memcg_event_ptable_queue_proc(struct file *file,
4832 wait_queue_head_t *wqh, poll_table *pt)
4834 struct mem_cgroup_event *event =
4835 container_of(pt, struct mem_cgroup_event, pt);
4838 add_wait_queue(wqh, &event->wait);
4842 * DO NOT USE IN NEW FILES.
4844 * Parse input and register new cgroup event handler.
4846 * Input must be in format '<event_fd> <control_fd> <args>'.
4847 * Interpretation of args is defined by control file implementation.
4849 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4850 char *buf, size_t nbytes, loff_t off)
4852 struct cgroup_subsys_state *css = of_css(of);
4853 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4854 struct mem_cgroup_event *event;
4855 struct cgroup_subsys_state *cfile_css;
4856 unsigned int efd, cfd;
4863 buf = strstrip(buf);
4865 efd = simple_strtoul(buf, &endp, 10);
4870 cfd = simple_strtoul(buf, &endp, 10);
4871 if ((*endp != ' ') && (*endp != '\0'))
4875 event = kzalloc(sizeof(*event), GFP_KERNEL);
4879 event->memcg = memcg;
4880 INIT_LIST_HEAD(&event->list);
4881 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4882 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4883 INIT_WORK(&event->remove, memcg_event_remove);
4891 event->eventfd = eventfd_ctx_fileget(efile.file);
4892 if (IS_ERR(event->eventfd)) {
4893 ret = PTR_ERR(event->eventfd);
4900 goto out_put_eventfd;
4903 /* the process need read permission on control file */
4904 /* AV: shouldn't we check that it's been opened for read instead? */
4905 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4910 * Determine the event callbacks and set them in @event. This used
4911 * to be done via struct cftype but cgroup core no longer knows
4912 * about these events. The following is crude but the whole thing
4913 * is for compatibility anyway.
4915 * DO NOT ADD NEW FILES.
4917 name = cfile.file->f_path.dentry->d_name.name;
4919 if (!strcmp(name, "memory.usage_in_bytes")) {
4920 event->register_event = mem_cgroup_usage_register_event;
4921 event->unregister_event = mem_cgroup_usage_unregister_event;
4922 } else if (!strcmp(name, "memory.oom_control")) {
4923 event->register_event = mem_cgroup_oom_register_event;
4924 event->unregister_event = mem_cgroup_oom_unregister_event;
4925 } else if (!strcmp(name, "memory.pressure_level")) {
4926 event->register_event = vmpressure_register_event;
4927 event->unregister_event = vmpressure_unregister_event;
4928 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4929 event->register_event = memsw_cgroup_usage_register_event;
4930 event->unregister_event = memsw_cgroup_usage_unregister_event;
4937 * Verify @cfile should belong to @css. Also, remaining events are
4938 * automatically removed on cgroup destruction but the removal is
4939 * asynchronous, so take an extra ref on @css.
4941 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4942 &memory_cgrp_subsys);
4944 if (IS_ERR(cfile_css))
4946 if (cfile_css != css) {
4951 ret = event->register_event(memcg, event->eventfd, buf);
4955 vfs_poll(efile.file, &event->pt);
4957 spin_lock(&memcg->event_list_lock);
4958 list_add(&event->list, &memcg->event_list);
4959 spin_unlock(&memcg->event_list_lock);
4971 eventfd_ctx_put(event->eventfd);
4980 static struct cftype mem_cgroup_legacy_files[] = {
4982 .name = "usage_in_bytes",
4983 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4984 .read_u64 = mem_cgroup_read_u64,
4987 .name = "max_usage_in_bytes",
4988 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4989 .write = mem_cgroup_reset,
4990 .read_u64 = mem_cgroup_read_u64,
4993 .name = "limit_in_bytes",
4994 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4995 .write = mem_cgroup_write,
4996 .read_u64 = mem_cgroup_read_u64,
4999 .name = "soft_limit_in_bytes",
5000 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5001 .write = mem_cgroup_write,
5002 .read_u64 = mem_cgroup_read_u64,
5006 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5007 .write = mem_cgroup_reset,
5008 .read_u64 = mem_cgroup_read_u64,
5012 .seq_show = memcg_stat_show,
5015 .name = "force_empty",
5016 .write = mem_cgroup_force_empty_write,
5019 .name = "use_hierarchy",
5020 .write_u64 = mem_cgroup_hierarchy_write,
5021 .read_u64 = mem_cgroup_hierarchy_read,
5024 .name = "cgroup.event_control", /* XXX: for compat */
5025 .write = memcg_write_event_control,
5026 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
5029 .name = "swappiness",
5030 .read_u64 = mem_cgroup_swappiness_read,
5031 .write_u64 = mem_cgroup_swappiness_write,
5034 .name = "move_charge_at_immigrate",
5035 .read_u64 = mem_cgroup_move_charge_read,
5036 .write_u64 = mem_cgroup_move_charge_write,
5039 .name = "oom_control",
5040 .seq_show = mem_cgroup_oom_control_read,
5041 .write_u64 = mem_cgroup_oom_control_write,
5042 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5045 .name = "pressure_level",
5049 .name = "numa_stat",
5050 .seq_show = memcg_numa_stat_show,
5054 .name = "kmem.limit_in_bytes",
5055 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5056 .write = mem_cgroup_write,
5057 .read_u64 = mem_cgroup_read_u64,
5060 .name = "kmem.usage_in_bytes",
5061 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5062 .read_u64 = mem_cgroup_read_u64,
5065 .name = "kmem.failcnt",
5066 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5067 .write = mem_cgroup_reset,
5068 .read_u64 = mem_cgroup_read_u64,
5071 .name = "kmem.max_usage_in_bytes",
5072 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5073 .write = mem_cgroup_reset,
5074 .read_u64 = mem_cgroup_read_u64,
5076 #if defined(CONFIG_MEMCG_KMEM) && \
5077 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5079 .name = "kmem.slabinfo",
5080 .seq_show = memcg_slab_show,
5084 .name = "kmem.tcp.limit_in_bytes",
5085 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5086 .write = mem_cgroup_write,
5087 .read_u64 = mem_cgroup_read_u64,
5090 .name = "kmem.tcp.usage_in_bytes",
5091 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5092 .read_u64 = mem_cgroup_read_u64,
5095 .name = "kmem.tcp.failcnt",
5096 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5097 .write = mem_cgroup_reset,
5098 .read_u64 = mem_cgroup_read_u64,
5101 .name = "kmem.tcp.max_usage_in_bytes",
5102 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5103 .write = mem_cgroup_reset,
5104 .read_u64 = mem_cgroup_read_u64,
5106 { }, /* terminate */
5110 * Private memory cgroup IDR
5112 * Swap-out records and page cache shadow entries need to store memcg
5113 * references in constrained space, so we maintain an ID space that is
5114 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5115 * memory-controlled cgroups to 64k.
5117 * However, there usually are many references to the offline CSS after
5118 * the cgroup has been destroyed, such as page cache or reclaimable
5119 * slab objects, that don't need to hang on to the ID. We want to keep
5120 * those dead CSS from occupying IDs, or we might quickly exhaust the
5121 * relatively small ID space and prevent the creation of new cgroups
5122 * even when there are much fewer than 64k cgroups - possibly none.
5124 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5125 * be freed and recycled when it's no longer needed, which is usually
5126 * when the CSS is offlined.
5128 * The only exception to that are records of swapped out tmpfs/shmem
5129 * pages that need to be attributed to live ancestors on swapin. But
5130 * those references are manageable from userspace.
5133 static DEFINE_IDR(mem_cgroup_idr);
5135 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5137 if (memcg->id.id > 0) {
5138 idr_remove(&mem_cgroup_idr, memcg->id.id);
5143 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5146 refcount_add(n, &memcg->id.ref);
5149 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5151 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5152 mem_cgroup_id_remove(memcg);
5154 /* Memcg ID pins CSS */
5155 css_put(&memcg->css);
5159 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5161 mem_cgroup_id_put_many(memcg, 1);
5165 * mem_cgroup_from_id - look up a memcg from a memcg id
5166 * @id: the memcg id to look up
5168 * Caller must hold rcu_read_lock().
5170 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5172 WARN_ON_ONCE(!rcu_read_lock_held());
5173 return idr_find(&mem_cgroup_idr, id);
5176 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5178 struct mem_cgroup_per_node *pn;
5181 * This routine is called against possible nodes.
5182 * But it's BUG to call kmalloc() against offline node.
5184 * TODO: this routine can waste much memory for nodes which will
5185 * never be onlined. It's better to use memory hotplug callback
5188 if (!node_state(node, N_NORMAL_MEMORY))
5190 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5194 pn->lruvec_stat_local = alloc_percpu_gfp(struct lruvec_stat,
5195 GFP_KERNEL_ACCOUNT);
5196 if (!pn->lruvec_stat_local) {
5201 pn->lruvec_stat_cpu = alloc_percpu_gfp(struct lruvec_stat,
5202 GFP_KERNEL_ACCOUNT);
5203 if (!pn->lruvec_stat_cpu) {
5204 free_percpu(pn->lruvec_stat_local);
5209 lruvec_init(&pn->lruvec);
5210 pn->usage_in_excess = 0;
5211 pn->on_tree = false;
5214 memcg->nodeinfo[node] = pn;
5218 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5220 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5225 free_percpu(pn->lruvec_stat_cpu);
5226 free_percpu(pn->lruvec_stat_local);
5230 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5235 free_mem_cgroup_per_node_info(memcg, node);
5236 free_percpu(memcg->vmstats_percpu);
5237 free_percpu(memcg->vmstats_local);
5241 static void mem_cgroup_free(struct mem_cgroup *memcg)
5243 memcg_wb_domain_exit(memcg);
5245 * Flush percpu vmstats and vmevents to guarantee the value correctness
5246 * on parent's and all ancestor levels.
5248 memcg_flush_percpu_vmstats(memcg);
5249 memcg_flush_percpu_vmevents(memcg);
5250 __mem_cgroup_free(memcg);
5253 static struct mem_cgroup *mem_cgroup_alloc(void)
5255 struct mem_cgroup *memcg;
5258 int __maybe_unused i;
5259 long error = -ENOMEM;
5261 size = sizeof(struct mem_cgroup);
5262 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5264 memcg = kzalloc(size, GFP_KERNEL);
5266 return ERR_PTR(error);
5268 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5269 1, MEM_CGROUP_ID_MAX,
5271 if (memcg->id.id < 0) {
5272 error = memcg->id.id;
5276 memcg->vmstats_local = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5277 GFP_KERNEL_ACCOUNT);
5278 if (!memcg->vmstats_local)
5281 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5282 GFP_KERNEL_ACCOUNT);
5283 if (!memcg->vmstats_percpu)
5287 if (alloc_mem_cgroup_per_node_info(memcg, node))
5290 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5293 INIT_WORK(&memcg->high_work, high_work_func);
5294 INIT_LIST_HEAD(&memcg->oom_notify);
5295 mutex_init(&memcg->thresholds_lock);
5296 spin_lock_init(&memcg->move_lock);
5297 vmpressure_init(&memcg->vmpressure);
5298 INIT_LIST_HEAD(&memcg->event_list);
5299 spin_lock_init(&memcg->event_list_lock);
5300 memcg->socket_pressure = jiffies;
5301 #ifdef CONFIG_MEMCG_KMEM
5302 memcg->kmemcg_id = -1;
5303 INIT_LIST_HEAD(&memcg->objcg_list);
5305 #ifdef CONFIG_CGROUP_WRITEBACK
5306 INIT_LIST_HEAD(&memcg->cgwb_list);
5307 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5308 memcg->cgwb_frn[i].done =
5309 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5311 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5312 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5313 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5314 memcg->deferred_split_queue.split_queue_len = 0;
5316 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5319 mem_cgroup_id_remove(memcg);
5320 __mem_cgroup_free(memcg);
5321 return ERR_PTR(error);
5324 static struct cgroup_subsys_state * __ref
5325 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5327 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5328 struct mem_cgroup *memcg, *old_memcg;
5329 long error = -ENOMEM;
5331 old_memcg = set_active_memcg(parent);
5332 memcg = mem_cgroup_alloc();
5333 set_active_memcg(old_memcg);
5335 return ERR_CAST(memcg);
5337 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5338 memcg->soft_limit = PAGE_COUNTER_MAX;
5339 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5341 memcg->swappiness = mem_cgroup_swappiness(parent);
5342 memcg->oom_kill_disable = parent->oom_kill_disable;
5345 page_counter_init(&memcg->memory, NULL);
5346 page_counter_init(&memcg->swap, NULL);
5347 page_counter_init(&memcg->kmem, NULL);
5348 page_counter_init(&memcg->tcpmem, NULL);
5349 } else if (parent->use_hierarchy) {
5350 memcg->use_hierarchy = true;
5351 page_counter_init(&memcg->memory, &parent->memory);
5352 page_counter_init(&memcg->swap, &parent->swap);
5353 page_counter_init(&memcg->kmem, &parent->kmem);
5354 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5356 page_counter_init(&memcg->memory, &root_mem_cgroup->memory);
5357 page_counter_init(&memcg->swap, &root_mem_cgroup->swap);
5358 page_counter_init(&memcg->kmem, &root_mem_cgroup->kmem);
5359 page_counter_init(&memcg->tcpmem, &root_mem_cgroup->tcpmem);
5361 * Deeper hierachy with use_hierarchy == false doesn't make
5362 * much sense so let cgroup subsystem know about this
5363 * unfortunate state in our controller.
5365 if (parent != root_mem_cgroup)
5366 memory_cgrp_subsys.broken_hierarchy = true;
5369 /* The following stuff does not apply to the root */
5371 root_mem_cgroup = memcg;
5375 error = memcg_online_kmem(memcg);
5379 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5380 static_branch_inc(&memcg_sockets_enabled_key);
5384 mem_cgroup_id_remove(memcg);
5385 mem_cgroup_free(memcg);
5386 return ERR_PTR(error);
5389 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5391 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5394 * A memcg must be visible for memcg_expand_shrinker_maps()
5395 * by the time the maps are allocated. So, we allocate maps
5396 * here, when for_each_mem_cgroup() can't skip it.
5398 if (memcg_alloc_shrinker_maps(memcg)) {
5399 mem_cgroup_id_remove(memcg);
5403 /* Online state pins memcg ID, memcg ID pins CSS */
5404 refcount_set(&memcg->id.ref, 1);
5409 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5411 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5412 struct mem_cgroup_event *event, *tmp;
5415 * Unregister events and notify userspace.
5416 * Notify userspace about cgroup removing only after rmdir of cgroup
5417 * directory to avoid race between userspace and kernelspace.
5419 spin_lock(&memcg->event_list_lock);
5420 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5421 list_del_init(&event->list);
5422 schedule_work(&event->remove);
5424 spin_unlock(&memcg->event_list_lock);
5426 page_counter_set_min(&memcg->memory, 0);
5427 page_counter_set_low(&memcg->memory, 0);
5429 memcg_offline_kmem(memcg);
5430 wb_memcg_offline(memcg);
5432 drain_all_stock(memcg);
5434 mem_cgroup_id_put(memcg);
5437 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5439 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5441 invalidate_reclaim_iterators(memcg);
5444 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5446 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5447 int __maybe_unused i;
5449 #ifdef CONFIG_CGROUP_WRITEBACK
5450 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5451 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5453 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5454 static_branch_dec(&memcg_sockets_enabled_key);
5456 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5457 static_branch_dec(&memcg_sockets_enabled_key);
5459 vmpressure_cleanup(&memcg->vmpressure);
5460 cancel_work_sync(&memcg->high_work);
5461 mem_cgroup_remove_from_trees(memcg);
5462 memcg_free_shrinker_maps(memcg);
5463 memcg_free_kmem(memcg);
5464 mem_cgroup_free(memcg);
5468 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5469 * @css: the target css
5471 * Reset the states of the mem_cgroup associated with @css. This is
5472 * invoked when the userland requests disabling on the default hierarchy
5473 * but the memcg is pinned through dependency. The memcg should stop
5474 * applying policies and should revert to the vanilla state as it may be
5475 * made visible again.
5477 * The current implementation only resets the essential configurations.
5478 * This needs to be expanded to cover all the visible parts.
5480 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5482 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5484 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5485 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5486 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5487 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5488 page_counter_set_min(&memcg->memory, 0);
5489 page_counter_set_low(&memcg->memory, 0);
5490 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5491 memcg->soft_limit = PAGE_COUNTER_MAX;
5492 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5493 memcg_wb_domain_size_changed(memcg);
5497 /* Handlers for move charge at task migration. */
5498 static int mem_cgroup_do_precharge(unsigned long count)
5502 /* Try a single bulk charge without reclaim first, kswapd may wake */
5503 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5505 mc.precharge += count;
5509 /* Try charges one by one with reclaim, but do not retry */
5511 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5525 enum mc_target_type {
5532 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5533 unsigned long addr, pte_t ptent)
5535 struct page *page = vm_normal_page(vma, addr, ptent);
5537 if (!page || !page_mapped(page))
5539 if (PageAnon(page)) {
5540 if (!(mc.flags & MOVE_ANON))
5543 if (!(mc.flags & MOVE_FILE))
5546 if (!get_page_unless_zero(page))
5552 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5553 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5554 pte_t ptent, swp_entry_t *entry)
5556 struct page *page = NULL;
5557 swp_entry_t ent = pte_to_swp_entry(ptent);
5559 if (!(mc.flags & MOVE_ANON))
5563 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5564 * a device and because they are not accessible by CPU they are store
5565 * as special swap entry in the CPU page table.
5567 if (is_device_private_entry(ent)) {
5568 page = device_private_entry_to_page(ent);
5570 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5571 * a refcount of 1 when free (unlike normal page)
5573 if (!page_ref_add_unless(page, 1, 1))
5578 if (non_swap_entry(ent))
5582 * Because lookup_swap_cache() updates some statistics counter,
5583 * we call find_get_page() with swapper_space directly.
5585 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5586 entry->val = ent.val;
5591 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5592 pte_t ptent, swp_entry_t *entry)
5598 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5599 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5601 if (!vma->vm_file) /* anonymous vma */
5603 if (!(mc.flags & MOVE_FILE))
5606 /* page is moved even if it's not RSS of this task(page-faulted). */
5607 /* shmem/tmpfs may report page out on swap: account for that too. */
5608 return find_get_incore_page(vma->vm_file->f_mapping,
5609 linear_page_index(vma, addr));
5613 * mem_cgroup_move_account - move account of the page
5615 * @compound: charge the page as compound or small page
5616 * @from: mem_cgroup which the page is moved from.
5617 * @to: mem_cgroup which the page is moved to. @from != @to.
5619 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5621 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5624 static int mem_cgroup_move_account(struct page *page,
5626 struct mem_cgroup *from,
5627 struct mem_cgroup *to)
5629 struct lruvec *from_vec, *to_vec;
5630 struct pglist_data *pgdat;
5631 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5634 VM_BUG_ON(from == to);
5635 VM_BUG_ON_PAGE(PageLRU(page), page);
5636 VM_BUG_ON(compound && !PageTransHuge(page));
5639 * Prevent mem_cgroup_migrate() from looking at
5640 * page->mem_cgroup of its source page while we change it.
5643 if (!trylock_page(page))
5647 if (page->mem_cgroup != from)
5650 pgdat = page_pgdat(page);
5651 from_vec = mem_cgroup_lruvec(from, pgdat);
5652 to_vec = mem_cgroup_lruvec(to, pgdat);
5654 lock_page_memcg(page);
5656 if (PageAnon(page)) {
5657 if (page_mapped(page)) {
5658 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5659 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5660 if (PageTransHuge(page)) {
5661 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5663 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5669 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5670 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5672 if (PageSwapBacked(page)) {
5673 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5674 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5677 if (page_mapped(page)) {
5678 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5679 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5682 if (PageDirty(page)) {
5683 struct address_space *mapping = page_mapping(page);
5685 if (mapping_can_writeback(mapping)) {
5686 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5688 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5694 if (PageWriteback(page)) {
5695 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5696 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5700 * All state has been migrated, let's switch to the new memcg.
5702 * It is safe to change page->mem_cgroup here because the page
5703 * is referenced, charged, isolated, and locked: we can't race
5704 * with (un)charging, migration, LRU putback, or anything else
5705 * that would rely on a stable page->mem_cgroup.
5707 * Note that lock_page_memcg is a memcg lock, not a page lock,
5708 * to save space. As soon as we switch page->mem_cgroup to a
5709 * new memcg that isn't locked, the above state can change
5710 * concurrently again. Make sure we're truly done with it.
5715 css_put(&from->css);
5717 page->mem_cgroup = to;
5719 __unlock_page_memcg(from);
5723 local_irq_disable();
5724 mem_cgroup_charge_statistics(to, page, nr_pages);
5725 memcg_check_events(to, page);
5726 mem_cgroup_charge_statistics(from, page, -nr_pages);
5727 memcg_check_events(from, page);
5736 * get_mctgt_type - get target type of moving charge
5737 * @vma: the vma the pte to be checked belongs
5738 * @addr: the address corresponding to the pte to be checked
5739 * @ptent: the pte to be checked
5740 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5743 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5744 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5745 * move charge. if @target is not NULL, the page is stored in target->page
5746 * with extra refcnt got(Callers should handle it).
5747 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5748 * target for charge migration. if @target is not NULL, the entry is stored
5750 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5751 * (so ZONE_DEVICE page and thus not on the lru).
5752 * For now we such page is charge like a regular page would be as for all
5753 * intent and purposes it is just special memory taking the place of a
5756 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5758 * Called with pte lock held.
5761 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5762 unsigned long addr, pte_t ptent, union mc_target *target)
5764 struct page *page = NULL;
5765 enum mc_target_type ret = MC_TARGET_NONE;
5766 swp_entry_t ent = { .val = 0 };
5768 if (pte_present(ptent))
5769 page = mc_handle_present_pte(vma, addr, ptent);
5770 else if (is_swap_pte(ptent))
5771 page = mc_handle_swap_pte(vma, ptent, &ent);
5772 else if (pte_none(ptent))
5773 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5775 if (!page && !ent.val)
5779 * Do only loose check w/o serialization.
5780 * mem_cgroup_move_account() checks the page is valid or
5781 * not under LRU exclusion.
5783 if (page->mem_cgroup == mc.from) {
5784 ret = MC_TARGET_PAGE;
5785 if (is_device_private_page(page))
5786 ret = MC_TARGET_DEVICE;
5788 target->page = page;
5790 if (!ret || !target)
5794 * There is a swap entry and a page doesn't exist or isn't charged.
5795 * But we cannot move a tail-page in a THP.
5797 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5798 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5799 ret = MC_TARGET_SWAP;
5806 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5808 * We don't consider PMD mapped swapping or file mapped pages because THP does
5809 * not support them for now.
5810 * Caller should make sure that pmd_trans_huge(pmd) is true.
5812 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5813 unsigned long addr, pmd_t pmd, union mc_target *target)
5815 struct page *page = NULL;
5816 enum mc_target_type ret = MC_TARGET_NONE;
5818 if (unlikely(is_swap_pmd(pmd))) {
5819 VM_BUG_ON(thp_migration_supported() &&
5820 !is_pmd_migration_entry(pmd));
5823 page = pmd_page(pmd);
5824 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5825 if (!(mc.flags & MOVE_ANON))
5827 if (page->mem_cgroup == mc.from) {
5828 ret = MC_TARGET_PAGE;
5831 target->page = page;
5837 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5838 unsigned long addr, pmd_t pmd, union mc_target *target)
5840 return MC_TARGET_NONE;
5844 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5845 unsigned long addr, unsigned long end,
5846 struct mm_walk *walk)
5848 struct vm_area_struct *vma = walk->vma;
5852 ptl = pmd_trans_huge_lock(pmd, vma);
5855 * Note their can not be MC_TARGET_DEVICE for now as we do not
5856 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5857 * this might change.
5859 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5860 mc.precharge += HPAGE_PMD_NR;
5865 if (pmd_trans_unstable(pmd))
5867 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5868 for (; addr != end; pte++, addr += PAGE_SIZE)
5869 if (get_mctgt_type(vma, addr, *pte, NULL))
5870 mc.precharge++; /* increment precharge temporarily */
5871 pte_unmap_unlock(pte - 1, ptl);
5877 static const struct mm_walk_ops precharge_walk_ops = {
5878 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5881 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5883 unsigned long precharge;
5886 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5887 mmap_read_unlock(mm);
5889 precharge = mc.precharge;
5895 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5897 unsigned long precharge = mem_cgroup_count_precharge(mm);
5899 VM_BUG_ON(mc.moving_task);
5900 mc.moving_task = current;
5901 return mem_cgroup_do_precharge(precharge);
5904 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5905 static void __mem_cgroup_clear_mc(void)
5907 struct mem_cgroup *from = mc.from;
5908 struct mem_cgroup *to = mc.to;
5910 /* we must uncharge all the leftover precharges from mc.to */
5912 cancel_charge(mc.to, mc.precharge);
5916 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5917 * we must uncharge here.
5919 if (mc.moved_charge) {
5920 cancel_charge(mc.from, mc.moved_charge);
5921 mc.moved_charge = 0;
5923 /* we must fixup refcnts and charges */
5924 if (mc.moved_swap) {
5925 /* uncharge swap account from the old cgroup */
5926 if (!mem_cgroup_is_root(mc.from))
5927 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5929 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5932 * we charged both to->memory and to->memsw, so we
5933 * should uncharge to->memory.
5935 if (!mem_cgroup_is_root(mc.to))
5936 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5940 memcg_oom_recover(from);
5941 memcg_oom_recover(to);
5942 wake_up_all(&mc.waitq);
5945 static void mem_cgroup_clear_mc(void)
5947 struct mm_struct *mm = mc.mm;
5950 * we must clear moving_task before waking up waiters at the end of
5953 mc.moving_task = NULL;
5954 __mem_cgroup_clear_mc();
5955 spin_lock(&mc.lock);
5959 spin_unlock(&mc.lock);
5964 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5966 struct cgroup_subsys_state *css;
5967 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5968 struct mem_cgroup *from;
5969 struct task_struct *leader, *p;
5970 struct mm_struct *mm;
5971 unsigned long move_flags;
5974 /* charge immigration isn't supported on the default hierarchy */
5975 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5979 * Multi-process migrations only happen on the default hierarchy
5980 * where charge immigration is not used. Perform charge
5981 * immigration if @tset contains a leader and whine if there are
5985 cgroup_taskset_for_each_leader(leader, css, tset) {
5988 memcg = mem_cgroup_from_css(css);
5994 * We are now commited to this value whatever it is. Changes in this
5995 * tunable will only affect upcoming migrations, not the current one.
5996 * So we need to save it, and keep it going.
5998 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
6002 from = mem_cgroup_from_task(p);
6004 VM_BUG_ON(from == memcg);
6006 mm = get_task_mm(p);
6009 /* We move charges only when we move a owner of the mm */
6010 if (mm->owner == p) {
6013 VM_BUG_ON(mc.precharge);
6014 VM_BUG_ON(mc.moved_charge);
6015 VM_BUG_ON(mc.moved_swap);
6017 spin_lock(&mc.lock);
6021 mc.flags = move_flags;
6022 spin_unlock(&mc.lock);
6023 /* We set mc.moving_task later */
6025 ret = mem_cgroup_precharge_mc(mm);
6027 mem_cgroup_clear_mc();
6034 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6037 mem_cgroup_clear_mc();
6040 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6041 unsigned long addr, unsigned long end,
6042 struct mm_walk *walk)
6045 struct vm_area_struct *vma = walk->vma;
6048 enum mc_target_type target_type;
6049 union mc_target target;
6052 ptl = pmd_trans_huge_lock(pmd, vma);
6054 if (mc.precharge < HPAGE_PMD_NR) {
6058 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6059 if (target_type == MC_TARGET_PAGE) {
6061 if (!isolate_lru_page(page)) {
6062 if (!mem_cgroup_move_account(page, true,
6064 mc.precharge -= HPAGE_PMD_NR;
6065 mc.moved_charge += HPAGE_PMD_NR;
6067 putback_lru_page(page);
6070 } else if (target_type == MC_TARGET_DEVICE) {
6072 if (!mem_cgroup_move_account(page, true,
6074 mc.precharge -= HPAGE_PMD_NR;
6075 mc.moved_charge += HPAGE_PMD_NR;
6083 if (pmd_trans_unstable(pmd))
6086 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6087 for (; addr != end; addr += PAGE_SIZE) {
6088 pte_t ptent = *(pte++);
6089 bool device = false;
6095 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6096 case MC_TARGET_DEVICE:
6099 case MC_TARGET_PAGE:
6102 * We can have a part of the split pmd here. Moving it
6103 * can be done but it would be too convoluted so simply
6104 * ignore such a partial THP and keep it in original
6105 * memcg. There should be somebody mapping the head.
6107 if (PageTransCompound(page))
6109 if (!device && isolate_lru_page(page))
6111 if (!mem_cgroup_move_account(page, false,
6114 /* we uncharge from mc.from later. */
6118 putback_lru_page(page);
6119 put: /* get_mctgt_type() gets the page */
6122 case MC_TARGET_SWAP:
6124 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6126 mem_cgroup_id_get_many(mc.to, 1);
6127 /* we fixup other refcnts and charges later. */
6135 pte_unmap_unlock(pte - 1, ptl);
6140 * We have consumed all precharges we got in can_attach().
6141 * We try charge one by one, but don't do any additional
6142 * charges to mc.to if we have failed in charge once in attach()
6145 ret = mem_cgroup_do_precharge(1);
6153 static const struct mm_walk_ops charge_walk_ops = {
6154 .pmd_entry = mem_cgroup_move_charge_pte_range,
6157 static void mem_cgroup_move_charge(void)
6159 lru_add_drain_all();
6161 * Signal lock_page_memcg() to take the memcg's move_lock
6162 * while we're moving its pages to another memcg. Then wait
6163 * for already started RCU-only updates to finish.
6165 atomic_inc(&mc.from->moving_account);
6168 if (unlikely(!mmap_read_trylock(mc.mm))) {
6170 * Someone who are holding the mmap_lock might be waiting in
6171 * waitq. So we cancel all extra charges, wake up all waiters,
6172 * and retry. Because we cancel precharges, we might not be able
6173 * to move enough charges, but moving charge is a best-effort
6174 * feature anyway, so it wouldn't be a big problem.
6176 __mem_cgroup_clear_mc();
6181 * When we have consumed all precharges and failed in doing
6182 * additional charge, the page walk just aborts.
6184 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6187 mmap_read_unlock(mc.mm);
6188 atomic_dec(&mc.from->moving_account);
6191 static void mem_cgroup_move_task(void)
6194 mem_cgroup_move_charge();
6195 mem_cgroup_clear_mc();
6198 #else /* !CONFIG_MMU */
6199 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6203 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6206 static void mem_cgroup_move_task(void)
6212 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6213 * to verify whether we're attached to the default hierarchy on each mount
6216 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6219 * use_hierarchy is forced on the default hierarchy. cgroup core
6220 * guarantees that @root doesn't have any children, so turning it
6221 * on for the root memcg is enough.
6223 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6224 root_mem_cgroup->use_hierarchy = true;
6226 root_mem_cgroup->use_hierarchy = false;
6229 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6231 if (value == PAGE_COUNTER_MAX)
6232 seq_puts(m, "max\n");
6234 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6239 static u64 memory_current_read(struct cgroup_subsys_state *css,
6242 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6244 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6247 static int memory_min_show(struct seq_file *m, void *v)
6249 return seq_puts_memcg_tunable(m,
6250 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6253 static ssize_t memory_min_write(struct kernfs_open_file *of,
6254 char *buf, size_t nbytes, loff_t off)
6256 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6260 buf = strstrip(buf);
6261 err = page_counter_memparse(buf, "max", &min);
6265 page_counter_set_min(&memcg->memory, min);
6270 static int memory_low_show(struct seq_file *m, void *v)
6272 return seq_puts_memcg_tunable(m,
6273 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6276 static ssize_t memory_low_write(struct kernfs_open_file *of,
6277 char *buf, size_t nbytes, loff_t off)
6279 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6283 buf = strstrip(buf);
6284 err = page_counter_memparse(buf, "max", &low);
6288 page_counter_set_low(&memcg->memory, low);
6293 static int memory_high_show(struct seq_file *m, void *v)
6295 return seq_puts_memcg_tunable(m,
6296 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6299 static ssize_t memory_high_write(struct kernfs_open_file *of,
6300 char *buf, size_t nbytes, loff_t off)
6302 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6303 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6304 bool drained = false;
6308 buf = strstrip(buf);
6309 err = page_counter_memparse(buf, "max", &high);
6314 unsigned long nr_pages = page_counter_read(&memcg->memory);
6315 unsigned long reclaimed;
6317 if (nr_pages <= high)
6320 if (signal_pending(current))
6324 drain_all_stock(memcg);
6329 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6332 if (!reclaimed && !nr_retries--)
6336 page_counter_set_high(&memcg->memory, high);
6338 memcg_wb_domain_size_changed(memcg);
6343 static int memory_max_show(struct seq_file *m, void *v)
6345 return seq_puts_memcg_tunable(m,
6346 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6349 static ssize_t memory_max_write(struct kernfs_open_file *of,
6350 char *buf, size_t nbytes, loff_t off)
6352 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6353 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6354 bool drained = false;
6358 buf = strstrip(buf);
6359 err = page_counter_memparse(buf, "max", &max);
6363 xchg(&memcg->memory.max, max);
6366 unsigned long nr_pages = page_counter_read(&memcg->memory);
6368 if (nr_pages <= max)
6371 if (signal_pending(current))
6375 drain_all_stock(memcg);
6381 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6387 memcg_memory_event(memcg, MEMCG_OOM);
6388 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6392 memcg_wb_domain_size_changed(memcg);
6396 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6398 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6399 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6400 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6401 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6402 seq_printf(m, "oom_kill %lu\n",
6403 atomic_long_read(&events[MEMCG_OOM_KILL]));
6406 static int memory_events_show(struct seq_file *m, void *v)
6408 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6410 __memory_events_show(m, memcg->memory_events);
6414 static int memory_events_local_show(struct seq_file *m, void *v)
6416 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6418 __memory_events_show(m, memcg->memory_events_local);
6422 static int memory_stat_show(struct seq_file *m, void *v)
6424 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6427 buf = memory_stat_format(memcg);
6436 static int memory_numa_stat_show(struct seq_file *m, void *v)
6439 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6441 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6444 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6447 seq_printf(m, "%s", memory_stats[i].name);
6448 for_each_node_state(nid, N_MEMORY) {
6450 struct lruvec *lruvec;
6452 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6453 size = lruvec_page_state(lruvec, memory_stats[i].idx);
6454 size *= memory_stats[i].ratio;
6455 seq_printf(m, " N%d=%llu", nid, size);
6464 static int memory_oom_group_show(struct seq_file *m, void *v)
6466 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6468 seq_printf(m, "%d\n", memcg->oom_group);
6473 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6474 char *buf, size_t nbytes, loff_t off)
6476 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6479 buf = strstrip(buf);
6483 ret = kstrtoint(buf, 0, &oom_group);
6487 if (oom_group != 0 && oom_group != 1)
6490 memcg->oom_group = oom_group;
6495 static struct cftype memory_files[] = {
6498 .flags = CFTYPE_NOT_ON_ROOT,
6499 .read_u64 = memory_current_read,
6503 .flags = CFTYPE_NOT_ON_ROOT,
6504 .seq_show = memory_min_show,
6505 .write = memory_min_write,
6509 .flags = CFTYPE_NOT_ON_ROOT,
6510 .seq_show = memory_low_show,
6511 .write = memory_low_write,
6515 .flags = CFTYPE_NOT_ON_ROOT,
6516 .seq_show = memory_high_show,
6517 .write = memory_high_write,
6521 .flags = CFTYPE_NOT_ON_ROOT,
6522 .seq_show = memory_max_show,
6523 .write = memory_max_write,
6527 .flags = CFTYPE_NOT_ON_ROOT,
6528 .file_offset = offsetof(struct mem_cgroup, events_file),
6529 .seq_show = memory_events_show,
6532 .name = "events.local",
6533 .flags = CFTYPE_NOT_ON_ROOT,
6534 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6535 .seq_show = memory_events_local_show,
6539 .seq_show = memory_stat_show,
6543 .name = "numa_stat",
6544 .seq_show = memory_numa_stat_show,
6548 .name = "oom.group",
6549 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6550 .seq_show = memory_oom_group_show,
6551 .write = memory_oom_group_write,
6556 struct cgroup_subsys memory_cgrp_subsys = {
6557 .css_alloc = mem_cgroup_css_alloc,
6558 .css_online = mem_cgroup_css_online,
6559 .css_offline = mem_cgroup_css_offline,
6560 .css_released = mem_cgroup_css_released,
6561 .css_free = mem_cgroup_css_free,
6562 .css_reset = mem_cgroup_css_reset,
6563 .can_attach = mem_cgroup_can_attach,
6564 .cancel_attach = mem_cgroup_cancel_attach,
6565 .post_attach = mem_cgroup_move_task,
6566 .bind = mem_cgroup_bind,
6567 .dfl_cftypes = memory_files,
6568 .legacy_cftypes = mem_cgroup_legacy_files,
6573 * This function calculates an individual cgroup's effective
6574 * protection which is derived from its own memory.min/low, its
6575 * parent's and siblings' settings, as well as the actual memory
6576 * distribution in the tree.
6578 * The following rules apply to the effective protection values:
6580 * 1. At the first level of reclaim, effective protection is equal to
6581 * the declared protection in memory.min and memory.low.
6583 * 2. To enable safe delegation of the protection configuration, at
6584 * subsequent levels the effective protection is capped to the
6585 * parent's effective protection.
6587 * 3. To make complex and dynamic subtrees easier to configure, the
6588 * user is allowed to overcommit the declared protection at a given
6589 * level. If that is the case, the parent's effective protection is
6590 * distributed to the children in proportion to how much protection
6591 * they have declared and how much of it they are utilizing.
6593 * This makes distribution proportional, but also work-conserving:
6594 * if one cgroup claims much more protection than it uses memory,
6595 * the unused remainder is available to its siblings.
6597 * 4. Conversely, when the declared protection is undercommitted at a
6598 * given level, the distribution of the larger parental protection
6599 * budget is NOT proportional. A cgroup's protection from a sibling
6600 * is capped to its own memory.min/low setting.
6602 * 5. However, to allow protecting recursive subtrees from each other
6603 * without having to declare each individual cgroup's fixed share
6604 * of the ancestor's claim to protection, any unutilized -
6605 * "floating" - protection from up the tree is distributed in
6606 * proportion to each cgroup's *usage*. This makes the protection
6607 * neutral wrt sibling cgroups and lets them compete freely over
6608 * the shared parental protection budget, but it protects the
6609 * subtree as a whole from neighboring subtrees.
6611 * Note that 4. and 5. are not in conflict: 4. is about protecting
6612 * against immediate siblings whereas 5. is about protecting against
6613 * neighboring subtrees.
6615 static unsigned long effective_protection(unsigned long usage,
6616 unsigned long parent_usage,
6617 unsigned long setting,
6618 unsigned long parent_effective,
6619 unsigned long siblings_protected)
6621 unsigned long protected;
6624 protected = min(usage, setting);
6626 * If all cgroups at this level combined claim and use more
6627 * protection then what the parent affords them, distribute
6628 * shares in proportion to utilization.
6630 * We are using actual utilization rather than the statically
6631 * claimed protection in order to be work-conserving: claimed
6632 * but unused protection is available to siblings that would
6633 * otherwise get a smaller chunk than what they claimed.
6635 if (siblings_protected > parent_effective)
6636 return protected * parent_effective / siblings_protected;
6639 * Ok, utilized protection of all children is within what the
6640 * parent affords them, so we know whatever this child claims
6641 * and utilizes is effectively protected.
6643 * If there is unprotected usage beyond this value, reclaim
6644 * will apply pressure in proportion to that amount.
6646 * If there is unutilized protection, the cgroup will be fully
6647 * shielded from reclaim, but we do return a smaller value for
6648 * protection than what the group could enjoy in theory. This
6649 * is okay. With the overcommit distribution above, effective
6650 * protection is always dependent on how memory is actually
6651 * consumed among the siblings anyway.
6656 * If the children aren't claiming (all of) the protection
6657 * afforded to them by the parent, distribute the remainder in
6658 * proportion to the (unprotected) memory of each cgroup. That
6659 * way, cgroups that aren't explicitly prioritized wrt each
6660 * other compete freely over the allowance, but they are
6661 * collectively protected from neighboring trees.
6663 * We're using unprotected memory for the weight so that if
6664 * some cgroups DO claim explicit protection, we don't protect
6665 * the same bytes twice.
6667 * Check both usage and parent_usage against the respective
6668 * protected values. One should imply the other, but they
6669 * aren't read atomically - make sure the division is sane.
6671 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6673 if (parent_effective > siblings_protected &&
6674 parent_usage > siblings_protected &&
6675 usage > protected) {
6676 unsigned long unclaimed;
6678 unclaimed = parent_effective - siblings_protected;
6679 unclaimed *= usage - protected;
6680 unclaimed /= parent_usage - siblings_protected;
6689 * mem_cgroup_protected - check if memory consumption is in the normal range
6690 * @root: the top ancestor of the sub-tree being checked
6691 * @memcg: the memory cgroup to check
6693 * WARNING: This function is not stateless! It can only be used as part
6694 * of a top-down tree iteration, not for isolated queries.
6696 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6697 struct mem_cgroup *memcg)
6699 unsigned long usage, parent_usage;
6700 struct mem_cgroup *parent;
6702 if (mem_cgroup_disabled())
6706 root = root_mem_cgroup;
6709 * Effective values of the reclaim targets are ignored so they
6710 * can be stale. Have a look at mem_cgroup_protection for more
6712 * TODO: calculation should be more robust so that we do not need
6713 * that special casing.
6718 usage = page_counter_read(&memcg->memory);
6722 parent = parent_mem_cgroup(memcg);
6723 /* No parent means a non-hierarchical mode on v1 memcg */
6727 if (parent == root) {
6728 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6729 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6733 parent_usage = page_counter_read(&parent->memory);
6735 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6736 READ_ONCE(memcg->memory.min),
6737 READ_ONCE(parent->memory.emin),
6738 atomic_long_read(&parent->memory.children_min_usage)));
6740 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6741 READ_ONCE(memcg->memory.low),
6742 READ_ONCE(parent->memory.elow),
6743 atomic_long_read(&parent->memory.children_low_usage)));
6747 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6748 * @page: page to charge
6749 * @mm: mm context of the victim
6750 * @gfp_mask: reclaim mode
6752 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6753 * pages according to @gfp_mask if necessary.
6755 * Returns 0 on success. Otherwise, an error code is returned.
6757 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6759 unsigned int nr_pages = thp_nr_pages(page);
6760 struct mem_cgroup *memcg = NULL;
6763 if (mem_cgroup_disabled())
6766 if (PageSwapCache(page)) {
6767 swp_entry_t ent = { .val = page_private(page), };
6771 * Every swap fault against a single page tries to charge the
6772 * page, bail as early as possible. shmem_unuse() encounters
6773 * already charged pages, too. page->mem_cgroup is protected
6774 * by the page lock, which serializes swap cache removal, which
6775 * in turn serializes uncharging.
6777 VM_BUG_ON_PAGE(!PageLocked(page), page);
6778 if (compound_head(page)->mem_cgroup)
6781 id = lookup_swap_cgroup_id(ent);
6783 memcg = mem_cgroup_from_id(id);
6784 if (memcg && !css_tryget_online(&memcg->css))
6790 memcg = get_mem_cgroup_from_mm(mm);
6792 ret = try_charge(memcg, gfp_mask, nr_pages);
6796 css_get(&memcg->css);
6797 commit_charge(page, memcg);
6799 local_irq_disable();
6800 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6801 memcg_check_events(memcg, page);
6804 if (PageSwapCache(page)) {
6805 swp_entry_t entry = { .val = page_private(page) };
6807 * The swap entry might not get freed for a long time,
6808 * let's not wait for it. The page already received a
6809 * memory+swap charge, drop the swap entry duplicate.
6811 mem_cgroup_uncharge_swap(entry, nr_pages);
6815 css_put(&memcg->css);
6820 struct uncharge_gather {
6821 struct mem_cgroup *memcg;
6822 unsigned long nr_pages;
6823 unsigned long pgpgout;
6824 unsigned long nr_kmem;
6825 struct page *dummy_page;
6828 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6830 memset(ug, 0, sizeof(*ug));
6833 static void uncharge_batch(const struct uncharge_gather *ug)
6835 unsigned long flags;
6837 if (!mem_cgroup_is_root(ug->memcg)) {
6838 page_counter_uncharge(&ug->memcg->memory, ug->nr_pages);
6839 if (do_memsw_account())
6840 page_counter_uncharge(&ug->memcg->memsw, ug->nr_pages);
6841 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6842 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6843 memcg_oom_recover(ug->memcg);
6846 local_irq_save(flags);
6847 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6848 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_pages);
6849 memcg_check_events(ug->memcg, ug->dummy_page);
6850 local_irq_restore(flags);
6852 /* drop reference from uncharge_page */
6853 css_put(&ug->memcg->css);
6856 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6858 unsigned long nr_pages;
6860 VM_BUG_ON_PAGE(PageLRU(page), page);
6862 if (!page->mem_cgroup)
6866 * Nobody should be changing or seriously looking at
6867 * page->mem_cgroup at this point, we have fully
6868 * exclusive access to the page.
6871 if (ug->memcg != page->mem_cgroup) {
6874 uncharge_gather_clear(ug);
6876 ug->memcg = page->mem_cgroup;
6878 /* pairs with css_put in uncharge_batch */
6879 css_get(&ug->memcg->css);
6882 nr_pages = compound_nr(page);
6883 ug->nr_pages += nr_pages;
6885 if (!PageKmemcg(page)) {
6888 ug->nr_kmem += nr_pages;
6889 __ClearPageKmemcg(page);
6892 ug->dummy_page = page;
6893 page->mem_cgroup = NULL;
6894 css_put(&ug->memcg->css);
6897 static void uncharge_list(struct list_head *page_list)
6899 struct uncharge_gather ug;
6900 struct list_head *next;
6902 uncharge_gather_clear(&ug);
6905 * Note that the list can be a single page->lru; hence the
6906 * do-while loop instead of a simple list_for_each_entry().
6908 next = page_list->next;
6912 page = list_entry(next, struct page, lru);
6913 next = page->lru.next;
6915 uncharge_page(page, &ug);
6916 } while (next != page_list);
6919 uncharge_batch(&ug);
6923 * mem_cgroup_uncharge - uncharge a page
6924 * @page: page to uncharge
6926 * Uncharge a page previously charged with mem_cgroup_charge().
6928 void mem_cgroup_uncharge(struct page *page)
6930 struct uncharge_gather ug;
6932 if (mem_cgroup_disabled())
6935 /* Don't touch page->lru of any random page, pre-check: */
6936 if (!page->mem_cgroup)
6939 uncharge_gather_clear(&ug);
6940 uncharge_page(page, &ug);
6941 uncharge_batch(&ug);
6945 * mem_cgroup_uncharge_list - uncharge a list of page
6946 * @page_list: list of pages to uncharge
6948 * Uncharge a list of pages previously charged with
6949 * mem_cgroup_charge().
6951 void mem_cgroup_uncharge_list(struct list_head *page_list)
6953 if (mem_cgroup_disabled())
6956 if (!list_empty(page_list))
6957 uncharge_list(page_list);
6961 * mem_cgroup_migrate - charge a page's replacement
6962 * @oldpage: currently circulating page
6963 * @newpage: replacement page
6965 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6966 * be uncharged upon free.
6968 * Both pages must be locked, @newpage->mapping must be set up.
6970 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6972 struct mem_cgroup *memcg;
6973 unsigned int nr_pages;
6974 unsigned long flags;
6976 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6977 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6978 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6979 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6982 if (mem_cgroup_disabled())
6985 /* Page cache replacement: new page already charged? */
6986 if (newpage->mem_cgroup)
6989 /* Swapcache readahead pages can get replaced before being charged */
6990 memcg = oldpage->mem_cgroup;
6994 /* Force-charge the new page. The old one will be freed soon */
6995 nr_pages = thp_nr_pages(newpage);
6997 page_counter_charge(&memcg->memory, nr_pages);
6998 if (do_memsw_account())
6999 page_counter_charge(&memcg->memsw, nr_pages);
7001 css_get(&memcg->css);
7002 commit_charge(newpage, memcg);
7004 local_irq_save(flags);
7005 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
7006 memcg_check_events(memcg, newpage);
7007 local_irq_restore(flags);
7010 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
7011 EXPORT_SYMBOL(memcg_sockets_enabled_key);
7013 void mem_cgroup_sk_alloc(struct sock *sk)
7015 struct mem_cgroup *memcg;
7017 if (!mem_cgroup_sockets_enabled)
7020 /* Do not associate the sock with unrelated interrupted task's memcg. */
7025 memcg = mem_cgroup_from_task(current);
7026 if (memcg == root_mem_cgroup)
7028 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
7030 if (css_tryget(&memcg->css))
7031 sk->sk_memcg = memcg;
7036 void mem_cgroup_sk_free(struct sock *sk)
7039 css_put(&sk->sk_memcg->css);
7043 * mem_cgroup_charge_skmem - charge socket memory
7044 * @memcg: memcg to charge
7045 * @nr_pages: number of pages to charge
7047 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7048 * @memcg's configured limit, %false if the charge had to be forced.
7050 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7052 gfp_t gfp_mask = GFP_KERNEL;
7054 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7055 struct page_counter *fail;
7057 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7058 memcg->tcpmem_pressure = 0;
7061 page_counter_charge(&memcg->tcpmem, nr_pages);
7062 memcg->tcpmem_pressure = 1;
7066 /* Don't block in the packet receive path */
7068 gfp_mask = GFP_NOWAIT;
7070 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7072 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
7075 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
7080 * mem_cgroup_uncharge_skmem - uncharge socket memory
7081 * @memcg: memcg to uncharge
7082 * @nr_pages: number of pages to uncharge
7084 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7086 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7087 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7091 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7093 refill_stock(memcg, nr_pages);
7096 static int __init cgroup_memory(char *s)
7100 while ((token = strsep(&s, ",")) != NULL) {
7103 if (!strcmp(token, "nosocket"))
7104 cgroup_memory_nosocket = true;
7105 if (!strcmp(token, "nokmem"))
7106 cgroup_memory_nokmem = true;
7110 __setup("cgroup.memory=", cgroup_memory);
7113 * subsys_initcall() for memory controller.
7115 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7116 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7117 * basically everything that doesn't depend on a specific mem_cgroup structure
7118 * should be initialized from here.
7120 static int __init mem_cgroup_init(void)
7124 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7125 memcg_hotplug_cpu_dead);
7127 for_each_possible_cpu(cpu)
7128 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7131 for_each_node(node) {
7132 struct mem_cgroup_tree_per_node *rtpn;
7134 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7135 node_online(node) ? node : NUMA_NO_NODE);
7137 rtpn->rb_root = RB_ROOT;
7138 rtpn->rb_rightmost = NULL;
7139 spin_lock_init(&rtpn->lock);
7140 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7145 subsys_initcall(mem_cgroup_init);
7147 #ifdef CONFIG_MEMCG_SWAP
7148 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7150 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7152 * The root cgroup cannot be destroyed, so it's refcount must
7155 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7159 memcg = parent_mem_cgroup(memcg);
7161 memcg = root_mem_cgroup;
7167 * mem_cgroup_swapout - transfer a memsw charge to swap
7168 * @page: page whose memsw charge to transfer
7169 * @entry: swap entry to move the charge to
7171 * Transfer the memsw charge of @page to @entry.
7173 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7175 struct mem_cgroup *memcg, *swap_memcg;
7176 unsigned int nr_entries;
7177 unsigned short oldid;
7179 VM_BUG_ON_PAGE(PageLRU(page), page);
7180 VM_BUG_ON_PAGE(page_count(page), page);
7182 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7185 memcg = page->mem_cgroup;
7187 /* Readahead page, never charged */
7192 * In case the memcg owning these pages has been offlined and doesn't
7193 * have an ID allocated to it anymore, charge the closest online
7194 * ancestor for the swap instead and transfer the memory+swap charge.
7196 swap_memcg = mem_cgroup_id_get_online(memcg);
7197 nr_entries = thp_nr_pages(page);
7198 /* Get references for the tail pages, too */
7200 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7201 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7203 VM_BUG_ON_PAGE(oldid, page);
7204 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7206 page->mem_cgroup = NULL;
7208 if (!mem_cgroup_is_root(memcg))
7209 page_counter_uncharge(&memcg->memory, nr_entries);
7211 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7212 if (!mem_cgroup_is_root(swap_memcg))
7213 page_counter_charge(&swap_memcg->memsw, nr_entries);
7214 page_counter_uncharge(&memcg->memsw, nr_entries);
7218 * Interrupts should be disabled here because the caller holds the
7219 * i_pages lock which is taken with interrupts-off. It is
7220 * important here to have the interrupts disabled because it is the
7221 * only synchronisation we have for updating the per-CPU variables.
7223 VM_BUG_ON(!irqs_disabled());
7224 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7225 memcg_check_events(memcg, page);
7227 css_put(&memcg->css);
7231 * mem_cgroup_try_charge_swap - try charging swap space for a page
7232 * @page: page being added to swap
7233 * @entry: swap entry to charge
7235 * Try to charge @page's memcg for the swap space at @entry.
7237 * Returns 0 on success, -ENOMEM on failure.
7239 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7241 unsigned int nr_pages = thp_nr_pages(page);
7242 struct page_counter *counter;
7243 struct mem_cgroup *memcg;
7244 unsigned short oldid;
7246 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7249 memcg = page->mem_cgroup;
7251 /* Readahead page, never charged */
7256 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7260 memcg = mem_cgroup_id_get_online(memcg);
7262 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7263 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7264 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7265 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7266 mem_cgroup_id_put(memcg);
7270 /* Get references for the tail pages, too */
7272 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7273 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7274 VM_BUG_ON_PAGE(oldid, page);
7275 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7281 * mem_cgroup_uncharge_swap - uncharge swap space
7282 * @entry: swap entry to uncharge
7283 * @nr_pages: the amount of swap space to uncharge
7285 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7287 struct mem_cgroup *memcg;
7290 id = swap_cgroup_record(entry, 0, nr_pages);
7292 memcg = mem_cgroup_from_id(id);
7294 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7295 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7296 page_counter_uncharge(&memcg->swap, nr_pages);
7298 page_counter_uncharge(&memcg->memsw, nr_pages);
7300 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7301 mem_cgroup_id_put_many(memcg, nr_pages);
7306 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7308 long nr_swap_pages = get_nr_swap_pages();
7310 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7311 return nr_swap_pages;
7312 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7313 nr_swap_pages = min_t(long, nr_swap_pages,
7314 READ_ONCE(memcg->swap.max) -
7315 page_counter_read(&memcg->swap));
7316 return nr_swap_pages;
7319 bool mem_cgroup_swap_full(struct page *page)
7321 struct mem_cgroup *memcg;
7323 VM_BUG_ON_PAGE(!PageLocked(page), page);
7327 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7330 memcg = page->mem_cgroup;
7334 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7335 unsigned long usage = page_counter_read(&memcg->swap);
7337 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7338 usage * 2 >= READ_ONCE(memcg->swap.max))
7345 static int __init setup_swap_account(char *s)
7347 if (!strcmp(s, "1"))
7348 cgroup_memory_noswap = 0;
7349 else if (!strcmp(s, "0"))
7350 cgroup_memory_noswap = 1;
7353 __setup("swapaccount=", setup_swap_account);
7355 static u64 swap_current_read(struct cgroup_subsys_state *css,
7358 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7360 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7363 static int swap_high_show(struct seq_file *m, void *v)
7365 return seq_puts_memcg_tunable(m,
7366 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7369 static ssize_t swap_high_write(struct kernfs_open_file *of,
7370 char *buf, size_t nbytes, loff_t off)
7372 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7376 buf = strstrip(buf);
7377 err = page_counter_memparse(buf, "max", &high);
7381 page_counter_set_high(&memcg->swap, high);
7386 static int swap_max_show(struct seq_file *m, void *v)
7388 return seq_puts_memcg_tunable(m,
7389 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7392 static ssize_t swap_max_write(struct kernfs_open_file *of,
7393 char *buf, size_t nbytes, loff_t off)
7395 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7399 buf = strstrip(buf);
7400 err = page_counter_memparse(buf, "max", &max);
7404 xchg(&memcg->swap.max, max);
7409 static int swap_events_show(struct seq_file *m, void *v)
7411 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7413 seq_printf(m, "high %lu\n",
7414 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7415 seq_printf(m, "max %lu\n",
7416 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7417 seq_printf(m, "fail %lu\n",
7418 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7423 static struct cftype swap_files[] = {
7425 .name = "swap.current",
7426 .flags = CFTYPE_NOT_ON_ROOT,
7427 .read_u64 = swap_current_read,
7430 .name = "swap.high",
7431 .flags = CFTYPE_NOT_ON_ROOT,
7432 .seq_show = swap_high_show,
7433 .write = swap_high_write,
7437 .flags = CFTYPE_NOT_ON_ROOT,
7438 .seq_show = swap_max_show,
7439 .write = swap_max_write,
7442 .name = "swap.events",
7443 .flags = CFTYPE_NOT_ON_ROOT,
7444 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7445 .seq_show = swap_events_show,
7450 static struct cftype memsw_files[] = {
7452 .name = "memsw.usage_in_bytes",
7453 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7454 .read_u64 = mem_cgroup_read_u64,
7457 .name = "memsw.max_usage_in_bytes",
7458 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7459 .write = mem_cgroup_reset,
7460 .read_u64 = mem_cgroup_read_u64,
7463 .name = "memsw.limit_in_bytes",
7464 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7465 .write = mem_cgroup_write,
7466 .read_u64 = mem_cgroup_read_u64,
7469 .name = "memsw.failcnt",
7470 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7471 .write = mem_cgroup_reset,
7472 .read_u64 = mem_cgroup_read_u64,
7474 { }, /* terminate */
7478 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7479 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7480 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7481 * boot parameter. This may result in premature OOPS inside
7482 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7484 static int __init mem_cgroup_swap_init(void)
7486 /* No memory control -> no swap control */
7487 if (mem_cgroup_disabled())
7488 cgroup_memory_noswap = true;
7490 if (cgroup_memory_noswap)
7493 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7494 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7498 core_initcall(mem_cgroup_swap_init);
7500 #endif /* CONFIG_MEMCG_SWAP */