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's memcg 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's memcg 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);
3706 memcg->kmemcg_id = memcg_id;
3707 memcg->kmem_state = KMEM_ONLINE;
3712 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3714 struct cgroup_subsys_state *css;
3715 struct mem_cgroup *parent, *child;
3718 if (memcg->kmem_state != KMEM_ONLINE)
3721 memcg->kmem_state = KMEM_ALLOCATED;
3723 parent = parent_mem_cgroup(memcg);
3725 parent = root_mem_cgroup;
3727 memcg_reparent_objcgs(memcg, parent);
3729 kmemcg_id = memcg->kmemcg_id;
3730 BUG_ON(kmemcg_id < 0);
3733 * Change kmemcg_id of this cgroup and all its descendants to the
3734 * parent's id, and then move all entries from this cgroup's list_lrus
3735 * to ones of the parent. After we have finished, all list_lrus
3736 * corresponding to this cgroup are guaranteed to remain empty. The
3737 * ordering is imposed by list_lru_node->lock taken by
3738 * memcg_drain_all_list_lrus().
3740 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3741 css_for_each_descendant_pre(css, &memcg->css) {
3742 child = mem_cgroup_from_css(css);
3743 BUG_ON(child->kmemcg_id != kmemcg_id);
3744 child->kmemcg_id = parent->kmemcg_id;
3745 if (!memcg->use_hierarchy)
3750 memcg_drain_all_list_lrus(kmemcg_id, parent);
3752 memcg_free_cache_id(kmemcg_id);
3755 static void memcg_free_kmem(struct mem_cgroup *memcg)
3757 /* css_alloc() failed, offlining didn't happen */
3758 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3759 memcg_offline_kmem(memcg);
3762 static int memcg_online_kmem(struct mem_cgroup *memcg)
3766 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3769 static void memcg_free_kmem(struct mem_cgroup *memcg)
3772 #endif /* CONFIG_MEMCG_KMEM */
3774 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3779 mutex_lock(&memcg_max_mutex);
3780 ret = page_counter_set_max(&memcg->kmem, max);
3781 mutex_unlock(&memcg_max_mutex);
3785 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3789 mutex_lock(&memcg_max_mutex);
3791 ret = page_counter_set_max(&memcg->tcpmem, max);
3795 if (!memcg->tcpmem_active) {
3797 * The active flag needs to be written after the static_key
3798 * update. This is what guarantees that the socket activation
3799 * function is the last one to run. See mem_cgroup_sk_alloc()
3800 * for details, and note that we don't mark any socket as
3801 * belonging to this memcg until that flag is up.
3803 * We need to do this, because static_keys will span multiple
3804 * sites, but we can't control their order. If we mark a socket
3805 * as accounted, but the accounting functions are not patched in
3806 * yet, we'll lose accounting.
3808 * We never race with the readers in mem_cgroup_sk_alloc(),
3809 * because when this value change, the code to process it is not
3812 static_branch_inc(&memcg_sockets_enabled_key);
3813 memcg->tcpmem_active = true;
3816 mutex_unlock(&memcg_max_mutex);
3821 * The user of this function is...
3824 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3825 char *buf, size_t nbytes, loff_t off)
3827 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3828 unsigned long nr_pages;
3831 buf = strstrip(buf);
3832 ret = page_counter_memparse(buf, "-1", &nr_pages);
3836 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3838 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3842 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3844 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3847 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3850 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3851 "Please report your usecase to linux-mm@kvack.org if you "
3852 "depend on this functionality.\n");
3853 ret = memcg_update_kmem_max(memcg, nr_pages);
3856 ret = memcg_update_tcp_max(memcg, nr_pages);
3860 case RES_SOFT_LIMIT:
3861 memcg->soft_limit = nr_pages;
3865 return ret ?: nbytes;
3868 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3869 size_t nbytes, loff_t off)
3871 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3872 struct page_counter *counter;
3874 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3876 counter = &memcg->memory;
3879 counter = &memcg->memsw;
3882 counter = &memcg->kmem;
3885 counter = &memcg->tcpmem;
3891 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3893 page_counter_reset_watermark(counter);
3896 counter->failcnt = 0;
3905 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3908 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3912 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3913 struct cftype *cft, u64 val)
3915 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3917 if (val & ~MOVE_MASK)
3921 * No kind of locking is needed in here, because ->can_attach() will
3922 * check this value once in the beginning of the process, and then carry
3923 * on with stale data. This means that changes to this value will only
3924 * affect task migrations starting after the change.
3926 memcg->move_charge_at_immigrate = val;
3930 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3931 struct cftype *cft, u64 val)
3939 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3940 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3941 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3943 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3944 int nid, unsigned int lru_mask, bool tree)
3946 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3947 unsigned long nr = 0;
3950 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3953 if (!(BIT(lru) & lru_mask))
3956 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3958 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3963 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3964 unsigned int lru_mask,
3967 unsigned long nr = 0;
3971 if (!(BIT(lru) & lru_mask))
3974 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3976 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3981 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3985 unsigned int lru_mask;
3988 static const struct numa_stat stats[] = {
3989 { "total", LRU_ALL },
3990 { "file", LRU_ALL_FILE },
3991 { "anon", LRU_ALL_ANON },
3992 { "unevictable", BIT(LRU_UNEVICTABLE) },
3994 const struct numa_stat *stat;
3996 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3998 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3999 seq_printf(m, "%s=%lu", stat->name,
4000 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4002 for_each_node_state(nid, N_MEMORY)
4003 seq_printf(m, " N%d=%lu", nid,
4004 mem_cgroup_node_nr_lru_pages(memcg, nid,
4005 stat->lru_mask, false));
4009 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4011 seq_printf(m, "hierarchical_%s=%lu", stat->name,
4012 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4014 for_each_node_state(nid, N_MEMORY)
4015 seq_printf(m, " N%d=%lu", nid,
4016 mem_cgroup_node_nr_lru_pages(memcg, nid,
4017 stat->lru_mask, true));
4023 #endif /* CONFIG_NUMA */
4025 static const unsigned int memcg1_stats[] = {
4028 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4038 static const char *const memcg1_stat_names[] = {
4041 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4051 /* Universal VM events cgroup1 shows, original sort order */
4052 static const unsigned int memcg1_events[] = {
4059 static int memcg_stat_show(struct seq_file *m, void *v)
4061 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4062 unsigned long memory, memsw;
4063 struct mem_cgroup *mi;
4066 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4068 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4071 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4073 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4074 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4075 if (memcg1_stats[i] == NR_ANON_THPS)
4078 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4081 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4082 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4083 memcg_events_local(memcg, memcg1_events[i]));
4085 for (i = 0; i < NR_LRU_LISTS; i++)
4086 seq_printf(m, "%s %lu\n", lru_list_name(i),
4087 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4090 /* Hierarchical information */
4091 memory = memsw = PAGE_COUNTER_MAX;
4092 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4093 memory = min(memory, READ_ONCE(mi->memory.max));
4094 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4096 seq_printf(m, "hierarchical_memory_limit %llu\n",
4097 (u64)memory * PAGE_SIZE);
4098 if (do_memsw_account())
4099 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4100 (u64)memsw * PAGE_SIZE);
4102 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4105 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4107 nr = memcg_page_state(memcg, memcg1_stats[i]);
4108 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4109 if (memcg1_stats[i] == NR_ANON_THPS)
4112 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4113 (u64)nr * PAGE_SIZE);
4116 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4117 seq_printf(m, "total_%s %llu\n",
4118 vm_event_name(memcg1_events[i]),
4119 (u64)memcg_events(memcg, memcg1_events[i]));
4121 for (i = 0; i < NR_LRU_LISTS; i++)
4122 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4123 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4126 #ifdef CONFIG_DEBUG_VM
4129 struct mem_cgroup_per_node *mz;
4130 unsigned long anon_cost = 0;
4131 unsigned long file_cost = 0;
4133 for_each_online_pgdat(pgdat) {
4134 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
4136 anon_cost += mz->lruvec.anon_cost;
4137 file_cost += mz->lruvec.file_cost;
4139 seq_printf(m, "anon_cost %lu\n", anon_cost);
4140 seq_printf(m, "file_cost %lu\n", file_cost);
4147 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4150 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4152 return mem_cgroup_swappiness(memcg);
4155 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4156 struct cftype *cft, u64 val)
4158 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4164 memcg->swappiness = val;
4166 vm_swappiness = val;
4171 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4173 struct mem_cgroup_threshold_ary *t;
4174 unsigned long usage;
4179 t = rcu_dereference(memcg->thresholds.primary);
4181 t = rcu_dereference(memcg->memsw_thresholds.primary);
4186 usage = mem_cgroup_usage(memcg, swap);
4189 * current_threshold points to threshold just below or equal to usage.
4190 * If it's not true, a threshold was crossed after last
4191 * call of __mem_cgroup_threshold().
4193 i = t->current_threshold;
4196 * Iterate backward over array of thresholds starting from
4197 * current_threshold and check if a threshold is crossed.
4198 * If none of thresholds below usage is crossed, we read
4199 * only one element of the array here.
4201 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4202 eventfd_signal(t->entries[i].eventfd, 1);
4204 /* i = current_threshold + 1 */
4208 * Iterate forward over array of thresholds starting from
4209 * current_threshold+1 and check if a threshold is crossed.
4210 * If none of thresholds above usage is crossed, we read
4211 * only one element of the array here.
4213 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4214 eventfd_signal(t->entries[i].eventfd, 1);
4216 /* Update current_threshold */
4217 t->current_threshold = i - 1;
4222 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4225 __mem_cgroup_threshold(memcg, false);
4226 if (do_memsw_account())
4227 __mem_cgroup_threshold(memcg, true);
4229 memcg = parent_mem_cgroup(memcg);
4233 static int compare_thresholds(const void *a, const void *b)
4235 const struct mem_cgroup_threshold *_a = a;
4236 const struct mem_cgroup_threshold *_b = b;
4238 if (_a->threshold > _b->threshold)
4241 if (_a->threshold < _b->threshold)
4247 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4249 struct mem_cgroup_eventfd_list *ev;
4251 spin_lock(&memcg_oom_lock);
4253 list_for_each_entry(ev, &memcg->oom_notify, list)
4254 eventfd_signal(ev->eventfd, 1);
4256 spin_unlock(&memcg_oom_lock);
4260 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4262 struct mem_cgroup *iter;
4264 for_each_mem_cgroup_tree(iter, memcg)
4265 mem_cgroup_oom_notify_cb(iter);
4268 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4269 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4271 struct mem_cgroup_thresholds *thresholds;
4272 struct mem_cgroup_threshold_ary *new;
4273 unsigned long threshold;
4274 unsigned long usage;
4277 ret = page_counter_memparse(args, "-1", &threshold);
4281 mutex_lock(&memcg->thresholds_lock);
4284 thresholds = &memcg->thresholds;
4285 usage = mem_cgroup_usage(memcg, false);
4286 } else if (type == _MEMSWAP) {
4287 thresholds = &memcg->memsw_thresholds;
4288 usage = mem_cgroup_usage(memcg, true);
4292 /* Check if a threshold crossed before adding a new one */
4293 if (thresholds->primary)
4294 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4296 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4298 /* Allocate memory for new array of thresholds */
4299 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4306 /* Copy thresholds (if any) to new array */
4307 if (thresholds->primary)
4308 memcpy(new->entries, thresholds->primary->entries,
4309 flex_array_size(new, entries, size - 1));
4311 /* Add new threshold */
4312 new->entries[size - 1].eventfd = eventfd;
4313 new->entries[size - 1].threshold = threshold;
4315 /* Sort thresholds. Registering of new threshold isn't time-critical */
4316 sort(new->entries, size, sizeof(*new->entries),
4317 compare_thresholds, NULL);
4319 /* Find current threshold */
4320 new->current_threshold = -1;
4321 for (i = 0; i < size; i++) {
4322 if (new->entries[i].threshold <= usage) {
4324 * new->current_threshold will not be used until
4325 * rcu_assign_pointer(), so it's safe to increment
4328 ++new->current_threshold;
4333 /* Free old spare buffer and save old primary buffer as spare */
4334 kfree(thresholds->spare);
4335 thresholds->spare = thresholds->primary;
4337 rcu_assign_pointer(thresholds->primary, new);
4339 /* To be sure that nobody uses thresholds */
4343 mutex_unlock(&memcg->thresholds_lock);
4348 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4349 struct eventfd_ctx *eventfd, const char *args)
4351 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4354 static int memsw_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, _MEMSWAP);
4360 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4361 struct eventfd_ctx *eventfd, enum res_type type)
4363 struct mem_cgroup_thresholds *thresholds;
4364 struct mem_cgroup_threshold_ary *new;
4365 unsigned long usage;
4366 int i, j, size, entries;
4368 mutex_lock(&memcg->thresholds_lock);
4371 thresholds = &memcg->thresholds;
4372 usage = mem_cgroup_usage(memcg, false);
4373 } else if (type == _MEMSWAP) {
4374 thresholds = &memcg->memsw_thresholds;
4375 usage = mem_cgroup_usage(memcg, true);
4379 if (!thresholds->primary)
4382 /* Check if a threshold crossed before removing */
4383 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4385 /* Calculate new number of threshold */
4387 for (i = 0; i < thresholds->primary->size; i++) {
4388 if (thresholds->primary->entries[i].eventfd != eventfd)
4394 new = thresholds->spare;
4396 /* If no items related to eventfd have been cleared, nothing to do */
4400 /* Set thresholds array to NULL if we don't have thresholds */
4409 /* Copy thresholds and find current threshold */
4410 new->current_threshold = -1;
4411 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4412 if (thresholds->primary->entries[i].eventfd == eventfd)
4415 new->entries[j] = thresholds->primary->entries[i];
4416 if (new->entries[j].threshold <= usage) {
4418 * new->current_threshold will not be used
4419 * until rcu_assign_pointer(), so it's safe to increment
4422 ++new->current_threshold;
4428 /* Swap primary and spare array */
4429 thresholds->spare = thresholds->primary;
4431 rcu_assign_pointer(thresholds->primary, new);
4433 /* To be sure that nobody uses thresholds */
4436 /* If all events are unregistered, free the spare array */
4438 kfree(thresholds->spare);
4439 thresholds->spare = NULL;
4442 mutex_unlock(&memcg->thresholds_lock);
4445 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4446 struct eventfd_ctx *eventfd)
4448 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4451 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4452 struct eventfd_ctx *eventfd)
4454 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4457 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4458 struct eventfd_ctx *eventfd, const char *args)
4460 struct mem_cgroup_eventfd_list *event;
4462 event = kmalloc(sizeof(*event), GFP_KERNEL);
4466 spin_lock(&memcg_oom_lock);
4468 event->eventfd = eventfd;
4469 list_add(&event->list, &memcg->oom_notify);
4471 /* already in OOM ? */
4472 if (memcg->under_oom)
4473 eventfd_signal(eventfd, 1);
4474 spin_unlock(&memcg_oom_lock);
4479 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4480 struct eventfd_ctx *eventfd)
4482 struct mem_cgroup_eventfd_list *ev, *tmp;
4484 spin_lock(&memcg_oom_lock);
4486 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4487 if (ev->eventfd == eventfd) {
4488 list_del(&ev->list);
4493 spin_unlock(&memcg_oom_lock);
4496 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4498 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4500 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4501 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4502 seq_printf(sf, "oom_kill %lu\n",
4503 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4507 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4508 struct cftype *cft, u64 val)
4510 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4512 /* cannot set to root cgroup and only 0 and 1 are allowed */
4513 if (!css->parent || !((val == 0) || (val == 1)))
4516 memcg->oom_kill_disable = val;
4518 memcg_oom_recover(memcg);
4523 #ifdef CONFIG_CGROUP_WRITEBACK
4525 #include <trace/events/writeback.h>
4527 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4529 return wb_domain_init(&memcg->cgwb_domain, gfp);
4532 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4534 wb_domain_exit(&memcg->cgwb_domain);
4537 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4539 wb_domain_size_changed(&memcg->cgwb_domain);
4542 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4544 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4546 if (!memcg->css.parent)
4549 return &memcg->cgwb_domain;
4553 * idx can be of type enum memcg_stat_item or node_stat_item.
4554 * Keep in sync with memcg_exact_page().
4556 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4558 long x = atomic_long_read(&memcg->vmstats[idx]);
4561 for_each_online_cpu(cpu)
4562 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4569 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4570 * @wb: bdi_writeback in question
4571 * @pfilepages: out parameter for number of file pages
4572 * @pheadroom: out parameter for number of allocatable pages according to memcg
4573 * @pdirty: out parameter for number of dirty pages
4574 * @pwriteback: out parameter for number of pages under writeback
4576 * Determine the numbers of file, headroom, dirty, and writeback pages in
4577 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4578 * is a bit more involved.
4580 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4581 * headroom is calculated as the lowest headroom of itself and the
4582 * ancestors. Note that this doesn't consider the actual amount of
4583 * available memory in the system. The caller should further cap
4584 * *@pheadroom accordingly.
4586 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4587 unsigned long *pheadroom, unsigned long *pdirty,
4588 unsigned long *pwriteback)
4590 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4591 struct mem_cgroup *parent;
4593 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4595 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4596 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4597 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4598 *pheadroom = PAGE_COUNTER_MAX;
4600 while ((parent = parent_mem_cgroup(memcg))) {
4601 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4602 READ_ONCE(memcg->memory.high));
4603 unsigned long used = page_counter_read(&memcg->memory);
4605 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4611 * Foreign dirty flushing
4613 * There's an inherent mismatch between memcg and writeback. The former
4614 * trackes ownership per-page while the latter per-inode. This was a
4615 * deliberate design decision because honoring per-page ownership in the
4616 * writeback path is complicated, may lead to higher CPU and IO overheads
4617 * and deemed unnecessary given that write-sharing an inode across
4618 * different cgroups isn't a common use-case.
4620 * Combined with inode majority-writer ownership switching, this works well
4621 * enough in most cases but there are some pathological cases. For
4622 * example, let's say there are two cgroups A and B which keep writing to
4623 * different but confined parts of the same inode. B owns the inode and
4624 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4625 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4626 * triggering background writeback. A will be slowed down without a way to
4627 * make writeback of the dirty pages happen.
4629 * Conditions like the above can lead to a cgroup getting repatedly and
4630 * severely throttled after making some progress after each
4631 * dirty_expire_interval while the underyling IO device is almost
4634 * Solving this problem completely requires matching the ownership tracking
4635 * granularities between memcg and writeback in either direction. However,
4636 * the more egregious behaviors can be avoided by simply remembering the
4637 * most recent foreign dirtying events and initiating remote flushes on
4638 * them when local writeback isn't enough to keep the memory clean enough.
4640 * The following two functions implement such mechanism. When a foreign
4641 * page - a page whose memcg and writeback ownerships don't match - is
4642 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4643 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4644 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4645 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4646 * foreign bdi_writebacks which haven't expired. Both the numbers of
4647 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4648 * limited to MEMCG_CGWB_FRN_CNT.
4650 * The mechanism only remembers IDs and doesn't hold any object references.
4651 * As being wrong occasionally doesn't matter, updates and accesses to the
4652 * records are lockless and racy.
4654 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4655 struct bdi_writeback *wb)
4657 struct mem_cgroup *memcg = page->mem_cgroup;
4658 struct memcg_cgwb_frn *frn;
4659 u64 now = get_jiffies_64();
4660 u64 oldest_at = now;
4664 trace_track_foreign_dirty(page, wb);
4667 * Pick the slot to use. If there is already a slot for @wb, keep
4668 * using it. If not replace the oldest one which isn't being
4671 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4672 frn = &memcg->cgwb_frn[i];
4673 if (frn->bdi_id == wb->bdi->id &&
4674 frn->memcg_id == wb->memcg_css->id)
4676 if (time_before64(frn->at, oldest_at) &&
4677 atomic_read(&frn->done.cnt) == 1) {
4679 oldest_at = frn->at;
4683 if (i < MEMCG_CGWB_FRN_CNT) {
4685 * Re-using an existing one. Update timestamp lazily to
4686 * avoid making the cacheline hot. We want them to be
4687 * reasonably up-to-date and significantly shorter than
4688 * dirty_expire_interval as that's what expires the record.
4689 * Use the shorter of 1s and dirty_expire_interval / 8.
4691 unsigned long update_intv =
4692 min_t(unsigned long, HZ,
4693 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4695 if (time_before64(frn->at, now - update_intv))
4697 } else if (oldest >= 0) {
4698 /* replace the oldest free one */
4699 frn = &memcg->cgwb_frn[oldest];
4700 frn->bdi_id = wb->bdi->id;
4701 frn->memcg_id = wb->memcg_css->id;
4706 /* issue foreign writeback flushes for recorded foreign dirtying events */
4707 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4709 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4710 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4711 u64 now = jiffies_64;
4714 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4715 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4718 * If the record is older than dirty_expire_interval,
4719 * writeback on it has already started. No need to kick it
4720 * off again. Also, don't start a new one if there's
4721 * already one in flight.
4723 if (time_after64(frn->at, now - intv) &&
4724 atomic_read(&frn->done.cnt) == 1) {
4726 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4727 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4728 WB_REASON_FOREIGN_FLUSH,
4734 #else /* CONFIG_CGROUP_WRITEBACK */
4736 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4741 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4745 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4749 #endif /* CONFIG_CGROUP_WRITEBACK */
4752 * DO NOT USE IN NEW FILES.
4754 * "cgroup.event_control" implementation.
4756 * This is way over-engineered. It tries to support fully configurable
4757 * events for each user. Such level of flexibility is completely
4758 * unnecessary especially in the light of the planned unified hierarchy.
4760 * Please deprecate this and replace with something simpler if at all
4765 * Unregister event and free resources.
4767 * Gets called from workqueue.
4769 static void memcg_event_remove(struct work_struct *work)
4771 struct mem_cgroup_event *event =
4772 container_of(work, struct mem_cgroup_event, remove);
4773 struct mem_cgroup *memcg = event->memcg;
4775 remove_wait_queue(event->wqh, &event->wait);
4777 event->unregister_event(memcg, event->eventfd);
4779 /* Notify userspace the event is going away. */
4780 eventfd_signal(event->eventfd, 1);
4782 eventfd_ctx_put(event->eventfd);
4784 css_put(&memcg->css);
4788 * Gets called on EPOLLHUP on eventfd when user closes it.
4790 * Called with wqh->lock held and interrupts disabled.
4792 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4793 int sync, void *key)
4795 struct mem_cgroup_event *event =
4796 container_of(wait, struct mem_cgroup_event, wait);
4797 struct mem_cgroup *memcg = event->memcg;
4798 __poll_t flags = key_to_poll(key);
4800 if (flags & EPOLLHUP) {
4802 * If the event has been detached at cgroup removal, we
4803 * can simply return knowing the other side will cleanup
4806 * We can't race against event freeing since the other
4807 * side will require wqh->lock via remove_wait_queue(),
4810 spin_lock(&memcg->event_list_lock);
4811 if (!list_empty(&event->list)) {
4812 list_del_init(&event->list);
4814 * We are in atomic context, but cgroup_event_remove()
4815 * may sleep, so we have to call it in workqueue.
4817 schedule_work(&event->remove);
4819 spin_unlock(&memcg->event_list_lock);
4825 static void memcg_event_ptable_queue_proc(struct file *file,
4826 wait_queue_head_t *wqh, poll_table *pt)
4828 struct mem_cgroup_event *event =
4829 container_of(pt, struct mem_cgroup_event, pt);
4832 add_wait_queue(wqh, &event->wait);
4836 * DO NOT USE IN NEW FILES.
4838 * Parse input and register new cgroup event handler.
4840 * Input must be in format '<event_fd> <control_fd> <args>'.
4841 * Interpretation of args is defined by control file implementation.
4843 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4844 char *buf, size_t nbytes, loff_t off)
4846 struct cgroup_subsys_state *css = of_css(of);
4847 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4848 struct mem_cgroup_event *event;
4849 struct cgroup_subsys_state *cfile_css;
4850 unsigned int efd, cfd;
4857 buf = strstrip(buf);
4859 efd = simple_strtoul(buf, &endp, 10);
4864 cfd = simple_strtoul(buf, &endp, 10);
4865 if ((*endp != ' ') && (*endp != '\0'))
4869 event = kzalloc(sizeof(*event), GFP_KERNEL);
4873 event->memcg = memcg;
4874 INIT_LIST_HEAD(&event->list);
4875 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4876 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4877 INIT_WORK(&event->remove, memcg_event_remove);
4885 event->eventfd = eventfd_ctx_fileget(efile.file);
4886 if (IS_ERR(event->eventfd)) {
4887 ret = PTR_ERR(event->eventfd);
4894 goto out_put_eventfd;
4897 /* the process need read permission on control file */
4898 /* AV: shouldn't we check that it's been opened for read instead? */
4899 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4904 * Determine the event callbacks and set them in @event. This used
4905 * to be done via struct cftype but cgroup core no longer knows
4906 * about these events. The following is crude but the whole thing
4907 * is for compatibility anyway.
4909 * DO NOT ADD NEW FILES.
4911 name = cfile.file->f_path.dentry->d_name.name;
4913 if (!strcmp(name, "memory.usage_in_bytes")) {
4914 event->register_event = mem_cgroup_usage_register_event;
4915 event->unregister_event = mem_cgroup_usage_unregister_event;
4916 } else if (!strcmp(name, "memory.oom_control")) {
4917 event->register_event = mem_cgroup_oom_register_event;
4918 event->unregister_event = mem_cgroup_oom_unregister_event;
4919 } else if (!strcmp(name, "memory.pressure_level")) {
4920 event->register_event = vmpressure_register_event;
4921 event->unregister_event = vmpressure_unregister_event;
4922 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4923 event->register_event = memsw_cgroup_usage_register_event;
4924 event->unregister_event = memsw_cgroup_usage_unregister_event;
4931 * Verify @cfile should belong to @css. Also, remaining events are
4932 * automatically removed on cgroup destruction but the removal is
4933 * asynchronous, so take an extra ref on @css.
4935 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4936 &memory_cgrp_subsys);
4938 if (IS_ERR(cfile_css))
4940 if (cfile_css != css) {
4945 ret = event->register_event(memcg, event->eventfd, buf);
4949 vfs_poll(efile.file, &event->pt);
4951 spin_lock(&memcg->event_list_lock);
4952 list_add(&event->list, &memcg->event_list);
4953 spin_unlock(&memcg->event_list_lock);
4965 eventfd_ctx_put(event->eventfd);
4974 static struct cftype mem_cgroup_legacy_files[] = {
4976 .name = "usage_in_bytes",
4977 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4978 .read_u64 = mem_cgroup_read_u64,
4981 .name = "max_usage_in_bytes",
4982 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4983 .write = mem_cgroup_reset,
4984 .read_u64 = mem_cgroup_read_u64,
4987 .name = "limit_in_bytes",
4988 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4989 .write = mem_cgroup_write,
4990 .read_u64 = mem_cgroup_read_u64,
4993 .name = "soft_limit_in_bytes",
4994 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4995 .write = mem_cgroup_write,
4996 .read_u64 = mem_cgroup_read_u64,
5000 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5001 .write = mem_cgroup_reset,
5002 .read_u64 = mem_cgroup_read_u64,
5006 .seq_show = memcg_stat_show,
5009 .name = "force_empty",
5010 .write = mem_cgroup_force_empty_write,
5013 .name = "use_hierarchy",
5014 .write_u64 = mem_cgroup_hierarchy_write,
5015 .read_u64 = mem_cgroup_hierarchy_read,
5018 .name = "cgroup.event_control", /* XXX: for compat */
5019 .write = memcg_write_event_control,
5020 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
5023 .name = "swappiness",
5024 .read_u64 = mem_cgroup_swappiness_read,
5025 .write_u64 = mem_cgroup_swappiness_write,
5028 .name = "move_charge_at_immigrate",
5029 .read_u64 = mem_cgroup_move_charge_read,
5030 .write_u64 = mem_cgroup_move_charge_write,
5033 .name = "oom_control",
5034 .seq_show = mem_cgroup_oom_control_read,
5035 .write_u64 = mem_cgroup_oom_control_write,
5036 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5039 .name = "pressure_level",
5043 .name = "numa_stat",
5044 .seq_show = memcg_numa_stat_show,
5048 .name = "kmem.limit_in_bytes",
5049 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5050 .write = mem_cgroup_write,
5051 .read_u64 = mem_cgroup_read_u64,
5054 .name = "kmem.usage_in_bytes",
5055 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5056 .read_u64 = mem_cgroup_read_u64,
5059 .name = "kmem.failcnt",
5060 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5061 .write = mem_cgroup_reset,
5062 .read_u64 = mem_cgroup_read_u64,
5065 .name = "kmem.max_usage_in_bytes",
5066 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5067 .write = mem_cgroup_reset,
5068 .read_u64 = mem_cgroup_read_u64,
5070 #if defined(CONFIG_MEMCG_KMEM) && \
5071 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5073 .name = "kmem.slabinfo",
5074 .seq_show = memcg_slab_show,
5078 .name = "kmem.tcp.limit_in_bytes",
5079 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5080 .write = mem_cgroup_write,
5081 .read_u64 = mem_cgroup_read_u64,
5084 .name = "kmem.tcp.usage_in_bytes",
5085 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5086 .read_u64 = mem_cgroup_read_u64,
5089 .name = "kmem.tcp.failcnt",
5090 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5091 .write = mem_cgroup_reset,
5092 .read_u64 = mem_cgroup_read_u64,
5095 .name = "kmem.tcp.max_usage_in_bytes",
5096 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5097 .write = mem_cgroup_reset,
5098 .read_u64 = mem_cgroup_read_u64,
5100 { }, /* terminate */
5104 * Private memory cgroup IDR
5106 * Swap-out records and page cache shadow entries need to store memcg
5107 * references in constrained space, so we maintain an ID space that is
5108 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5109 * memory-controlled cgroups to 64k.
5111 * However, there usually are many references to the offline CSS after
5112 * the cgroup has been destroyed, such as page cache or reclaimable
5113 * slab objects, that don't need to hang on to the ID. We want to keep
5114 * those dead CSS from occupying IDs, or we might quickly exhaust the
5115 * relatively small ID space and prevent the creation of new cgroups
5116 * even when there are much fewer than 64k cgroups - possibly none.
5118 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5119 * be freed and recycled when it's no longer needed, which is usually
5120 * when the CSS is offlined.
5122 * The only exception to that are records of swapped out tmpfs/shmem
5123 * pages that need to be attributed to live ancestors on swapin. But
5124 * those references are manageable from userspace.
5127 static DEFINE_IDR(mem_cgroup_idr);
5129 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5131 if (memcg->id.id > 0) {
5132 idr_remove(&mem_cgroup_idr, memcg->id.id);
5137 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5140 refcount_add(n, &memcg->id.ref);
5143 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5145 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5146 mem_cgroup_id_remove(memcg);
5148 /* Memcg ID pins CSS */
5149 css_put(&memcg->css);
5153 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5155 mem_cgroup_id_put_many(memcg, 1);
5159 * mem_cgroup_from_id - look up a memcg from a memcg id
5160 * @id: the memcg id to look up
5162 * Caller must hold rcu_read_lock().
5164 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5166 WARN_ON_ONCE(!rcu_read_lock_held());
5167 return idr_find(&mem_cgroup_idr, id);
5170 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5172 struct mem_cgroup_per_node *pn;
5175 * This routine is called against possible nodes.
5176 * But it's BUG to call kmalloc() against offline node.
5178 * TODO: this routine can waste much memory for nodes which will
5179 * never be onlined. It's better to use memory hotplug callback
5182 if (!node_state(node, N_NORMAL_MEMORY))
5184 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5188 pn->lruvec_stat_local = alloc_percpu_gfp(struct lruvec_stat,
5189 GFP_KERNEL_ACCOUNT);
5190 if (!pn->lruvec_stat_local) {
5195 pn->lruvec_stat_cpu = alloc_percpu_gfp(struct lruvec_stat,
5196 GFP_KERNEL_ACCOUNT);
5197 if (!pn->lruvec_stat_cpu) {
5198 free_percpu(pn->lruvec_stat_local);
5203 lruvec_init(&pn->lruvec);
5204 pn->usage_in_excess = 0;
5205 pn->on_tree = false;
5208 memcg->nodeinfo[node] = pn;
5212 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5214 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5219 free_percpu(pn->lruvec_stat_cpu);
5220 free_percpu(pn->lruvec_stat_local);
5224 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5229 free_mem_cgroup_per_node_info(memcg, node);
5230 free_percpu(memcg->vmstats_percpu);
5231 free_percpu(memcg->vmstats_local);
5235 static void mem_cgroup_free(struct mem_cgroup *memcg)
5237 memcg_wb_domain_exit(memcg);
5239 * Flush percpu vmstats and vmevents to guarantee the value correctness
5240 * on parent's and all ancestor levels.
5242 memcg_flush_percpu_vmstats(memcg);
5243 memcg_flush_percpu_vmevents(memcg);
5244 __mem_cgroup_free(memcg);
5247 static struct mem_cgroup *mem_cgroup_alloc(void)
5249 struct mem_cgroup *memcg;
5252 int __maybe_unused i;
5253 long error = -ENOMEM;
5255 size = sizeof(struct mem_cgroup);
5256 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5258 memcg = kzalloc(size, GFP_KERNEL);
5260 return ERR_PTR(error);
5262 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5263 1, MEM_CGROUP_ID_MAX,
5265 if (memcg->id.id < 0) {
5266 error = memcg->id.id;
5270 memcg->vmstats_local = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5271 GFP_KERNEL_ACCOUNT);
5272 if (!memcg->vmstats_local)
5275 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5276 GFP_KERNEL_ACCOUNT);
5277 if (!memcg->vmstats_percpu)
5281 if (alloc_mem_cgroup_per_node_info(memcg, node))
5284 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5287 INIT_WORK(&memcg->high_work, high_work_func);
5288 INIT_LIST_HEAD(&memcg->oom_notify);
5289 mutex_init(&memcg->thresholds_lock);
5290 spin_lock_init(&memcg->move_lock);
5291 vmpressure_init(&memcg->vmpressure);
5292 INIT_LIST_HEAD(&memcg->event_list);
5293 spin_lock_init(&memcg->event_list_lock);
5294 memcg->socket_pressure = jiffies;
5295 #ifdef CONFIG_MEMCG_KMEM
5296 memcg->kmemcg_id = -1;
5297 INIT_LIST_HEAD(&memcg->objcg_list);
5299 #ifdef CONFIG_CGROUP_WRITEBACK
5300 INIT_LIST_HEAD(&memcg->cgwb_list);
5301 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5302 memcg->cgwb_frn[i].done =
5303 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5305 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5306 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5307 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5308 memcg->deferred_split_queue.split_queue_len = 0;
5310 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5313 mem_cgroup_id_remove(memcg);
5314 __mem_cgroup_free(memcg);
5315 return ERR_PTR(error);
5318 static struct cgroup_subsys_state * __ref
5319 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5321 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5322 struct mem_cgroup *memcg, *old_memcg;
5323 long error = -ENOMEM;
5325 old_memcg = set_active_memcg(parent);
5326 memcg = mem_cgroup_alloc();
5327 set_active_memcg(old_memcg);
5329 return ERR_CAST(memcg);
5331 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5332 memcg->soft_limit = PAGE_COUNTER_MAX;
5333 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5335 memcg->swappiness = mem_cgroup_swappiness(parent);
5336 memcg->oom_kill_disable = parent->oom_kill_disable;
5339 page_counter_init(&memcg->memory, NULL);
5340 page_counter_init(&memcg->swap, NULL);
5341 page_counter_init(&memcg->kmem, NULL);
5342 page_counter_init(&memcg->tcpmem, NULL);
5343 } else if (parent->use_hierarchy) {
5344 memcg->use_hierarchy = true;
5345 page_counter_init(&memcg->memory, &parent->memory);
5346 page_counter_init(&memcg->swap, &parent->swap);
5347 page_counter_init(&memcg->kmem, &parent->kmem);
5348 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5350 page_counter_init(&memcg->memory, &root_mem_cgroup->memory);
5351 page_counter_init(&memcg->swap, &root_mem_cgroup->swap);
5352 page_counter_init(&memcg->kmem, &root_mem_cgroup->kmem);
5353 page_counter_init(&memcg->tcpmem, &root_mem_cgroup->tcpmem);
5355 * Deeper hierachy with use_hierarchy == false doesn't make
5356 * much sense so let cgroup subsystem know about this
5357 * unfortunate state in our controller.
5359 if (parent != root_mem_cgroup)
5360 memory_cgrp_subsys.broken_hierarchy = true;
5363 /* The following stuff does not apply to the root */
5365 root_mem_cgroup = memcg;
5369 error = memcg_online_kmem(memcg);
5373 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5374 static_branch_inc(&memcg_sockets_enabled_key);
5378 mem_cgroup_id_remove(memcg);
5379 mem_cgroup_free(memcg);
5380 return ERR_PTR(error);
5383 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5385 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5388 * A memcg must be visible for memcg_expand_shrinker_maps()
5389 * by the time the maps are allocated. So, we allocate maps
5390 * here, when for_each_mem_cgroup() can't skip it.
5392 if (memcg_alloc_shrinker_maps(memcg)) {
5393 mem_cgroup_id_remove(memcg);
5397 /* Online state pins memcg ID, memcg ID pins CSS */
5398 refcount_set(&memcg->id.ref, 1);
5403 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5405 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5406 struct mem_cgroup_event *event, *tmp;
5409 * Unregister events and notify userspace.
5410 * Notify userspace about cgroup removing only after rmdir of cgroup
5411 * directory to avoid race between userspace and kernelspace.
5413 spin_lock(&memcg->event_list_lock);
5414 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5415 list_del_init(&event->list);
5416 schedule_work(&event->remove);
5418 spin_unlock(&memcg->event_list_lock);
5420 page_counter_set_min(&memcg->memory, 0);
5421 page_counter_set_low(&memcg->memory, 0);
5423 memcg_offline_kmem(memcg);
5424 wb_memcg_offline(memcg);
5426 drain_all_stock(memcg);
5428 mem_cgroup_id_put(memcg);
5431 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5433 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5435 invalidate_reclaim_iterators(memcg);
5438 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5440 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5441 int __maybe_unused i;
5443 #ifdef CONFIG_CGROUP_WRITEBACK
5444 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5445 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5447 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5448 static_branch_dec(&memcg_sockets_enabled_key);
5450 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5451 static_branch_dec(&memcg_sockets_enabled_key);
5453 vmpressure_cleanup(&memcg->vmpressure);
5454 cancel_work_sync(&memcg->high_work);
5455 mem_cgroup_remove_from_trees(memcg);
5456 memcg_free_shrinker_maps(memcg);
5457 memcg_free_kmem(memcg);
5458 mem_cgroup_free(memcg);
5462 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5463 * @css: the target css
5465 * Reset the states of the mem_cgroup associated with @css. This is
5466 * invoked when the userland requests disabling on the default hierarchy
5467 * but the memcg is pinned through dependency. The memcg should stop
5468 * applying policies and should revert to the vanilla state as it may be
5469 * made visible again.
5471 * The current implementation only resets the essential configurations.
5472 * This needs to be expanded to cover all the visible parts.
5474 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5476 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5478 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5479 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5480 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5481 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5482 page_counter_set_min(&memcg->memory, 0);
5483 page_counter_set_low(&memcg->memory, 0);
5484 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5485 memcg->soft_limit = PAGE_COUNTER_MAX;
5486 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5487 memcg_wb_domain_size_changed(memcg);
5491 /* Handlers for move charge at task migration. */
5492 static int mem_cgroup_do_precharge(unsigned long count)
5496 /* Try a single bulk charge without reclaim first, kswapd may wake */
5497 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5499 mc.precharge += count;
5503 /* Try charges one by one with reclaim, but do not retry */
5505 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5519 enum mc_target_type {
5526 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5527 unsigned long addr, pte_t ptent)
5529 struct page *page = vm_normal_page(vma, addr, ptent);
5531 if (!page || !page_mapped(page))
5533 if (PageAnon(page)) {
5534 if (!(mc.flags & MOVE_ANON))
5537 if (!(mc.flags & MOVE_FILE))
5540 if (!get_page_unless_zero(page))
5546 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5547 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5548 pte_t ptent, swp_entry_t *entry)
5550 struct page *page = NULL;
5551 swp_entry_t ent = pte_to_swp_entry(ptent);
5553 if (!(mc.flags & MOVE_ANON))
5557 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5558 * a device and because they are not accessible by CPU they are store
5559 * as special swap entry in the CPU page table.
5561 if (is_device_private_entry(ent)) {
5562 page = device_private_entry_to_page(ent);
5564 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5565 * a refcount of 1 when free (unlike normal page)
5567 if (!page_ref_add_unless(page, 1, 1))
5572 if (non_swap_entry(ent))
5576 * Because lookup_swap_cache() updates some statistics counter,
5577 * we call find_get_page() with swapper_space directly.
5579 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5580 entry->val = ent.val;
5585 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5586 pte_t ptent, swp_entry_t *entry)
5592 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5593 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5595 if (!vma->vm_file) /* anonymous vma */
5597 if (!(mc.flags & MOVE_FILE))
5600 /* page is moved even if it's not RSS of this task(page-faulted). */
5601 /* shmem/tmpfs may report page out on swap: account for that too. */
5602 return find_get_incore_page(vma->vm_file->f_mapping,
5603 linear_page_index(vma, addr));
5607 * mem_cgroup_move_account - move account of the page
5609 * @compound: charge the page as compound or small page
5610 * @from: mem_cgroup which the page is moved from.
5611 * @to: mem_cgroup which the page is moved to. @from != @to.
5613 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5615 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5618 static int mem_cgroup_move_account(struct page *page,
5620 struct mem_cgroup *from,
5621 struct mem_cgroup *to)
5623 struct lruvec *from_vec, *to_vec;
5624 struct pglist_data *pgdat;
5625 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5628 VM_BUG_ON(from == to);
5629 VM_BUG_ON_PAGE(PageLRU(page), page);
5630 VM_BUG_ON(compound && !PageTransHuge(page));
5633 * Prevent mem_cgroup_migrate() from looking at
5634 * page->mem_cgroup of its source page while we change it.
5637 if (!trylock_page(page))
5641 if (page->mem_cgroup != from)
5644 pgdat = page_pgdat(page);
5645 from_vec = mem_cgroup_lruvec(from, pgdat);
5646 to_vec = mem_cgroup_lruvec(to, pgdat);
5648 lock_page_memcg(page);
5650 if (PageAnon(page)) {
5651 if (page_mapped(page)) {
5652 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5653 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5654 if (PageTransHuge(page)) {
5655 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5657 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5663 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5664 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5666 if (PageSwapBacked(page)) {
5667 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5668 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5671 if (page_mapped(page)) {
5672 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5673 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5676 if (PageDirty(page)) {
5677 struct address_space *mapping = page_mapping(page);
5679 if (mapping_can_writeback(mapping)) {
5680 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5682 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5688 if (PageWriteback(page)) {
5689 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5690 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5694 * All state has been migrated, let's switch to the new memcg.
5696 * It is safe to change page->mem_cgroup here because the page
5697 * is referenced, charged, isolated, and locked: we can't race
5698 * with (un)charging, migration, LRU putback, or anything else
5699 * that would rely on a stable page->mem_cgroup.
5701 * Note that lock_page_memcg is a memcg lock, not a page lock,
5702 * to save space. As soon as we switch page->mem_cgroup to a
5703 * new memcg that isn't locked, the above state can change
5704 * concurrently again. Make sure we're truly done with it.
5709 css_put(&from->css);
5711 page->mem_cgroup = to;
5713 __unlock_page_memcg(from);
5717 local_irq_disable();
5718 mem_cgroup_charge_statistics(to, page, nr_pages);
5719 memcg_check_events(to, page);
5720 mem_cgroup_charge_statistics(from, page, -nr_pages);
5721 memcg_check_events(from, page);
5730 * get_mctgt_type - get target type of moving charge
5731 * @vma: the vma the pte to be checked belongs
5732 * @addr: the address corresponding to the pte to be checked
5733 * @ptent: the pte to be checked
5734 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5737 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5738 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5739 * move charge. if @target is not NULL, the page is stored in target->page
5740 * with extra refcnt got(Callers should handle it).
5741 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5742 * target for charge migration. if @target is not NULL, the entry is stored
5744 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5745 * (so ZONE_DEVICE page and thus not on the lru).
5746 * For now we such page is charge like a regular page would be as for all
5747 * intent and purposes it is just special memory taking the place of a
5750 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5752 * Called with pte lock held.
5755 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5756 unsigned long addr, pte_t ptent, union mc_target *target)
5758 struct page *page = NULL;
5759 enum mc_target_type ret = MC_TARGET_NONE;
5760 swp_entry_t ent = { .val = 0 };
5762 if (pte_present(ptent))
5763 page = mc_handle_present_pte(vma, addr, ptent);
5764 else if (is_swap_pte(ptent))
5765 page = mc_handle_swap_pte(vma, ptent, &ent);
5766 else if (pte_none(ptent))
5767 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5769 if (!page && !ent.val)
5773 * Do only loose check w/o serialization.
5774 * mem_cgroup_move_account() checks the page is valid or
5775 * not under LRU exclusion.
5777 if (page->mem_cgroup == mc.from) {
5778 ret = MC_TARGET_PAGE;
5779 if (is_device_private_page(page))
5780 ret = MC_TARGET_DEVICE;
5782 target->page = page;
5784 if (!ret || !target)
5788 * There is a swap entry and a page doesn't exist or isn't charged.
5789 * But we cannot move a tail-page in a THP.
5791 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5792 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5793 ret = MC_TARGET_SWAP;
5800 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5802 * We don't consider PMD mapped swapping or file mapped pages because THP does
5803 * not support them for now.
5804 * Caller should make sure that pmd_trans_huge(pmd) is true.
5806 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5807 unsigned long addr, pmd_t pmd, union mc_target *target)
5809 struct page *page = NULL;
5810 enum mc_target_type ret = MC_TARGET_NONE;
5812 if (unlikely(is_swap_pmd(pmd))) {
5813 VM_BUG_ON(thp_migration_supported() &&
5814 !is_pmd_migration_entry(pmd));
5817 page = pmd_page(pmd);
5818 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5819 if (!(mc.flags & MOVE_ANON))
5821 if (page->mem_cgroup == mc.from) {
5822 ret = MC_TARGET_PAGE;
5825 target->page = page;
5831 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5832 unsigned long addr, pmd_t pmd, union mc_target *target)
5834 return MC_TARGET_NONE;
5838 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5839 unsigned long addr, unsigned long end,
5840 struct mm_walk *walk)
5842 struct vm_area_struct *vma = walk->vma;
5846 ptl = pmd_trans_huge_lock(pmd, vma);
5849 * Note their can not be MC_TARGET_DEVICE for now as we do not
5850 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5851 * this might change.
5853 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5854 mc.precharge += HPAGE_PMD_NR;
5859 if (pmd_trans_unstable(pmd))
5861 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5862 for (; addr != end; pte++, addr += PAGE_SIZE)
5863 if (get_mctgt_type(vma, addr, *pte, NULL))
5864 mc.precharge++; /* increment precharge temporarily */
5865 pte_unmap_unlock(pte - 1, ptl);
5871 static const struct mm_walk_ops precharge_walk_ops = {
5872 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5875 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5877 unsigned long precharge;
5880 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5881 mmap_read_unlock(mm);
5883 precharge = mc.precharge;
5889 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5891 unsigned long precharge = mem_cgroup_count_precharge(mm);
5893 VM_BUG_ON(mc.moving_task);
5894 mc.moving_task = current;
5895 return mem_cgroup_do_precharge(precharge);
5898 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5899 static void __mem_cgroup_clear_mc(void)
5901 struct mem_cgroup *from = mc.from;
5902 struct mem_cgroup *to = mc.to;
5904 /* we must uncharge all the leftover precharges from mc.to */
5906 cancel_charge(mc.to, mc.precharge);
5910 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5911 * we must uncharge here.
5913 if (mc.moved_charge) {
5914 cancel_charge(mc.from, mc.moved_charge);
5915 mc.moved_charge = 0;
5917 /* we must fixup refcnts and charges */
5918 if (mc.moved_swap) {
5919 /* uncharge swap account from the old cgroup */
5920 if (!mem_cgroup_is_root(mc.from))
5921 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5923 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5926 * we charged both to->memory and to->memsw, so we
5927 * should uncharge to->memory.
5929 if (!mem_cgroup_is_root(mc.to))
5930 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5934 memcg_oom_recover(from);
5935 memcg_oom_recover(to);
5936 wake_up_all(&mc.waitq);
5939 static void mem_cgroup_clear_mc(void)
5941 struct mm_struct *mm = mc.mm;
5944 * we must clear moving_task before waking up waiters at the end of
5947 mc.moving_task = NULL;
5948 __mem_cgroup_clear_mc();
5949 spin_lock(&mc.lock);
5953 spin_unlock(&mc.lock);
5958 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5960 struct cgroup_subsys_state *css;
5961 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5962 struct mem_cgroup *from;
5963 struct task_struct *leader, *p;
5964 struct mm_struct *mm;
5965 unsigned long move_flags;
5968 /* charge immigration isn't supported on the default hierarchy */
5969 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5973 * Multi-process migrations only happen on the default hierarchy
5974 * where charge immigration is not used. Perform charge
5975 * immigration if @tset contains a leader and whine if there are
5979 cgroup_taskset_for_each_leader(leader, css, tset) {
5982 memcg = mem_cgroup_from_css(css);
5988 * We are now commited to this value whatever it is. Changes in this
5989 * tunable will only affect upcoming migrations, not the current one.
5990 * So we need to save it, and keep it going.
5992 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5996 from = mem_cgroup_from_task(p);
5998 VM_BUG_ON(from == memcg);
6000 mm = get_task_mm(p);
6003 /* We move charges only when we move a owner of the mm */
6004 if (mm->owner == p) {
6007 VM_BUG_ON(mc.precharge);
6008 VM_BUG_ON(mc.moved_charge);
6009 VM_BUG_ON(mc.moved_swap);
6011 spin_lock(&mc.lock);
6015 mc.flags = move_flags;
6016 spin_unlock(&mc.lock);
6017 /* We set mc.moving_task later */
6019 ret = mem_cgroup_precharge_mc(mm);
6021 mem_cgroup_clear_mc();
6028 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6031 mem_cgroup_clear_mc();
6034 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6035 unsigned long addr, unsigned long end,
6036 struct mm_walk *walk)
6039 struct vm_area_struct *vma = walk->vma;
6042 enum mc_target_type target_type;
6043 union mc_target target;
6046 ptl = pmd_trans_huge_lock(pmd, vma);
6048 if (mc.precharge < HPAGE_PMD_NR) {
6052 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6053 if (target_type == MC_TARGET_PAGE) {
6055 if (!isolate_lru_page(page)) {
6056 if (!mem_cgroup_move_account(page, true,
6058 mc.precharge -= HPAGE_PMD_NR;
6059 mc.moved_charge += HPAGE_PMD_NR;
6061 putback_lru_page(page);
6064 } else if (target_type == MC_TARGET_DEVICE) {
6066 if (!mem_cgroup_move_account(page, true,
6068 mc.precharge -= HPAGE_PMD_NR;
6069 mc.moved_charge += HPAGE_PMD_NR;
6077 if (pmd_trans_unstable(pmd))
6080 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6081 for (; addr != end; addr += PAGE_SIZE) {
6082 pte_t ptent = *(pte++);
6083 bool device = false;
6089 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6090 case MC_TARGET_DEVICE:
6093 case MC_TARGET_PAGE:
6096 * We can have a part of the split pmd here. Moving it
6097 * can be done but it would be too convoluted so simply
6098 * ignore such a partial THP and keep it in original
6099 * memcg. There should be somebody mapping the head.
6101 if (PageTransCompound(page))
6103 if (!device && isolate_lru_page(page))
6105 if (!mem_cgroup_move_account(page, false,
6108 /* we uncharge from mc.from later. */
6112 putback_lru_page(page);
6113 put: /* get_mctgt_type() gets the page */
6116 case MC_TARGET_SWAP:
6118 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6120 mem_cgroup_id_get_many(mc.to, 1);
6121 /* we fixup other refcnts and charges later. */
6129 pte_unmap_unlock(pte - 1, ptl);
6134 * We have consumed all precharges we got in can_attach().
6135 * We try charge one by one, but don't do any additional
6136 * charges to mc.to if we have failed in charge once in attach()
6139 ret = mem_cgroup_do_precharge(1);
6147 static const struct mm_walk_ops charge_walk_ops = {
6148 .pmd_entry = mem_cgroup_move_charge_pte_range,
6151 static void mem_cgroup_move_charge(void)
6153 lru_add_drain_all();
6155 * Signal lock_page_memcg() to take the memcg's move_lock
6156 * while we're moving its pages to another memcg. Then wait
6157 * for already started RCU-only updates to finish.
6159 atomic_inc(&mc.from->moving_account);
6162 if (unlikely(!mmap_read_trylock(mc.mm))) {
6164 * Someone who are holding the mmap_lock might be waiting in
6165 * waitq. So we cancel all extra charges, wake up all waiters,
6166 * and retry. Because we cancel precharges, we might not be able
6167 * to move enough charges, but moving charge is a best-effort
6168 * feature anyway, so it wouldn't be a big problem.
6170 __mem_cgroup_clear_mc();
6175 * When we have consumed all precharges and failed in doing
6176 * additional charge, the page walk just aborts.
6178 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6181 mmap_read_unlock(mc.mm);
6182 atomic_dec(&mc.from->moving_account);
6185 static void mem_cgroup_move_task(void)
6188 mem_cgroup_move_charge();
6189 mem_cgroup_clear_mc();
6192 #else /* !CONFIG_MMU */
6193 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6197 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6200 static void mem_cgroup_move_task(void)
6206 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6207 * to verify whether we're attached to the default hierarchy on each mount
6210 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6213 * use_hierarchy is forced on the default hierarchy. cgroup core
6214 * guarantees that @root doesn't have any children, so turning it
6215 * on for the root memcg is enough.
6217 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6218 root_mem_cgroup->use_hierarchy = true;
6220 root_mem_cgroup->use_hierarchy = false;
6223 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6225 if (value == PAGE_COUNTER_MAX)
6226 seq_puts(m, "max\n");
6228 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6233 static u64 memory_current_read(struct cgroup_subsys_state *css,
6236 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6238 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6241 static int memory_min_show(struct seq_file *m, void *v)
6243 return seq_puts_memcg_tunable(m,
6244 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6247 static ssize_t memory_min_write(struct kernfs_open_file *of,
6248 char *buf, size_t nbytes, loff_t off)
6250 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6254 buf = strstrip(buf);
6255 err = page_counter_memparse(buf, "max", &min);
6259 page_counter_set_min(&memcg->memory, min);
6264 static int memory_low_show(struct seq_file *m, void *v)
6266 return seq_puts_memcg_tunable(m,
6267 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6270 static ssize_t memory_low_write(struct kernfs_open_file *of,
6271 char *buf, size_t nbytes, loff_t off)
6273 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6277 buf = strstrip(buf);
6278 err = page_counter_memparse(buf, "max", &low);
6282 page_counter_set_low(&memcg->memory, low);
6287 static int memory_high_show(struct seq_file *m, void *v)
6289 return seq_puts_memcg_tunable(m,
6290 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6293 static ssize_t memory_high_write(struct kernfs_open_file *of,
6294 char *buf, size_t nbytes, loff_t off)
6296 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6297 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6298 bool drained = false;
6302 buf = strstrip(buf);
6303 err = page_counter_memparse(buf, "max", &high);
6308 unsigned long nr_pages = page_counter_read(&memcg->memory);
6309 unsigned long reclaimed;
6311 if (nr_pages <= high)
6314 if (signal_pending(current))
6318 drain_all_stock(memcg);
6323 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6326 if (!reclaimed && !nr_retries--)
6330 page_counter_set_high(&memcg->memory, high);
6332 memcg_wb_domain_size_changed(memcg);
6337 static int memory_max_show(struct seq_file *m, void *v)
6339 return seq_puts_memcg_tunable(m,
6340 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6343 static ssize_t memory_max_write(struct kernfs_open_file *of,
6344 char *buf, size_t nbytes, loff_t off)
6346 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6347 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6348 bool drained = false;
6352 buf = strstrip(buf);
6353 err = page_counter_memparse(buf, "max", &max);
6357 xchg(&memcg->memory.max, max);
6360 unsigned long nr_pages = page_counter_read(&memcg->memory);
6362 if (nr_pages <= max)
6365 if (signal_pending(current))
6369 drain_all_stock(memcg);
6375 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6381 memcg_memory_event(memcg, MEMCG_OOM);
6382 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6386 memcg_wb_domain_size_changed(memcg);
6390 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6392 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6393 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6394 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6395 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6396 seq_printf(m, "oom_kill %lu\n",
6397 atomic_long_read(&events[MEMCG_OOM_KILL]));
6400 static int memory_events_show(struct seq_file *m, void *v)
6402 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6404 __memory_events_show(m, memcg->memory_events);
6408 static int memory_events_local_show(struct seq_file *m, void *v)
6410 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6412 __memory_events_show(m, memcg->memory_events_local);
6416 static int memory_stat_show(struct seq_file *m, void *v)
6418 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6421 buf = memory_stat_format(memcg);
6430 static int memory_numa_stat_show(struct seq_file *m, void *v)
6433 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6435 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6438 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6441 seq_printf(m, "%s", memory_stats[i].name);
6442 for_each_node_state(nid, N_MEMORY) {
6444 struct lruvec *lruvec;
6446 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6447 size = lruvec_page_state(lruvec, memory_stats[i].idx);
6448 size *= memory_stats[i].ratio;
6449 seq_printf(m, " N%d=%llu", nid, size);
6458 static int memory_oom_group_show(struct seq_file *m, void *v)
6460 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6462 seq_printf(m, "%d\n", memcg->oom_group);
6467 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6468 char *buf, size_t nbytes, loff_t off)
6470 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6473 buf = strstrip(buf);
6477 ret = kstrtoint(buf, 0, &oom_group);
6481 if (oom_group != 0 && oom_group != 1)
6484 memcg->oom_group = oom_group;
6489 static struct cftype memory_files[] = {
6492 .flags = CFTYPE_NOT_ON_ROOT,
6493 .read_u64 = memory_current_read,
6497 .flags = CFTYPE_NOT_ON_ROOT,
6498 .seq_show = memory_min_show,
6499 .write = memory_min_write,
6503 .flags = CFTYPE_NOT_ON_ROOT,
6504 .seq_show = memory_low_show,
6505 .write = memory_low_write,
6509 .flags = CFTYPE_NOT_ON_ROOT,
6510 .seq_show = memory_high_show,
6511 .write = memory_high_write,
6515 .flags = CFTYPE_NOT_ON_ROOT,
6516 .seq_show = memory_max_show,
6517 .write = memory_max_write,
6521 .flags = CFTYPE_NOT_ON_ROOT,
6522 .file_offset = offsetof(struct mem_cgroup, events_file),
6523 .seq_show = memory_events_show,
6526 .name = "events.local",
6527 .flags = CFTYPE_NOT_ON_ROOT,
6528 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6529 .seq_show = memory_events_local_show,
6533 .seq_show = memory_stat_show,
6537 .name = "numa_stat",
6538 .seq_show = memory_numa_stat_show,
6542 .name = "oom.group",
6543 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6544 .seq_show = memory_oom_group_show,
6545 .write = memory_oom_group_write,
6550 struct cgroup_subsys memory_cgrp_subsys = {
6551 .css_alloc = mem_cgroup_css_alloc,
6552 .css_online = mem_cgroup_css_online,
6553 .css_offline = mem_cgroup_css_offline,
6554 .css_released = mem_cgroup_css_released,
6555 .css_free = mem_cgroup_css_free,
6556 .css_reset = mem_cgroup_css_reset,
6557 .can_attach = mem_cgroup_can_attach,
6558 .cancel_attach = mem_cgroup_cancel_attach,
6559 .post_attach = mem_cgroup_move_task,
6560 .bind = mem_cgroup_bind,
6561 .dfl_cftypes = memory_files,
6562 .legacy_cftypes = mem_cgroup_legacy_files,
6567 * This function calculates an individual cgroup's effective
6568 * protection which is derived from its own memory.min/low, its
6569 * parent's and siblings' settings, as well as the actual memory
6570 * distribution in the tree.
6572 * The following rules apply to the effective protection values:
6574 * 1. At the first level of reclaim, effective protection is equal to
6575 * the declared protection in memory.min and memory.low.
6577 * 2. To enable safe delegation of the protection configuration, at
6578 * subsequent levels the effective protection is capped to the
6579 * parent's effective protection.
6581 * 3. To make complex and dynamic subtrees easier to configure, the
6582 * user is allowed to overcommit the declared protection at a given
6583 * level. If that is the case, the parent's effective protection is
6584 * distributed to the children in proportion to how much protection
6585 * they have declared and how much of it they are utilizing.
6587 * This makes distribution proportional, but also work-conserving:
6588 * if one cgroup claims much more protection than it uses memory,
6589 * the unused remainder is available to its siblings.
6591 * 4. Conversely, when the declared protection is undercommitted at a
6592 * given level, the distribution of the larger parental protection
6593 * budget is NOT proportional. A cgroup's protection from a sibling
6594 * is capped to its own memory.min/low setting.
6596 * 5. However, to allow protecting recursive subtrees from each other
6597 * without having to declare each individual cgroup's fixed share
6598 * of the ancestor's claim to protection, any unutilized -
6599 * "floating" - protection from up the tree is distributed in
6600 * proportion to each cgroup's *usage*. This makes the protection
6601 * neutral wrt sibling cgroups and lets them compete freely over
6602 * the shared parental protection budget, but it protects the
6603 * subtree as a whole from neighboring subtrees.
6605 * Note that 4. and 5. are not in conflict: 4. is about protecting
6606 * against immediate siblings whereas 5. is about protecting against
6607 * neighboring subtrees.
6609 static unsigned long effective_protection(unsigned long usage,
6610 unsigned long parent_usage,
6611 unsigned long setting,
6612 unsigned long parent_effective,
6613 unsigned long siblings_protected)
6615 unsigned long protected;
6618 protected = min(usage, setting);
6620 * If all cgroups at this level combined claim and use more
6621 * protection then what the parent affords them, distribute
6622 * shares in proportion to utilization.
6624 * We are using actual utilization rather than the statically
6625 * claimed protection in order to be work-conserving: claimed
6626 * but unused protection is available to siblings that would
6627 * otherwise get a smaller chunk than what they claimed.
6629 if (siblings_protected > parent_effective)
6630 return protected * parent_effective / siblings_protected;
6633 * Ok, utilized protection of all children is within what the
6634 * parent affords them, so we know whatever this child claims
6635 * and utilizes is effectively protected.
6637 * If there is unprotected usage beyond this value, reclaim
6638 * will apply pressure in proportion to that amount.
6640 * If there is unutilized protection, the cgroup will be fully
6641 * shielded from reclaim, but we do return a smaller value for
6642 * protection than what the group could enjoy in theory. This
6643 * is okay. With the overcommit distribution above, effective
6644 * protection is always dependent on how memory is actually
6645 * consumed among the siblings anyway.
6650 * If the children aren't claiming (all of) the protection
6651 * afforded to them by the parent, distribute the remainder in
6652 * proportion to the (unprotected) memory of each cgroup. That
6653 * way, cgroups that aren't explicitly prioritized wrt each
6654 * other compete freely over the allowance, but they are
6655 * collectively protected from neighboring trees.
6657 * We're using unprotected memory for the weight so that if
6658 * some cgroups DO claim explicit protection, we don't protect
6659 * the same bytes twice.
6661 * Check both usage and parent_usage against the respective
6662 * protected values. One should imply the other, but they
6663 * aren't read atomically - make sure the division is sane.
6665 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6667 if (parent_effective > siblings_protected &&
6668 parent_usage > siblings_protected &&
6669 usage > protected) {
6670 unsigned long unclaimed;
6672 unclaimed = parent_effective - siblings_protected;
6673 unclaimed *= usage - protected;
6674 unclaimed /= parent_usage - siblings_protected;
6683 * mem_cgroup_protected - check if memory consumption is in the normal range
6684 * @root: the top ancestor of the sub-tree being checked
6685 * @memcg: the memory cgroup to check
6687 * WARNING: This function is not stateless! It can only be used as part
6688 * of a top-down tree iteration, not for isolated queries.
6690 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6691 struct mem_cgroup *memcg)
6693 unsigned long usage, parent_usage;
6694 struct mem_cgroup *parent;
6696 if (mem_cgroup_disabled())
6700 root = root_mem_cgroup;
6703 * Effective values of the reclaim targets are ignored so they
6704 * can be stale. Have a look at mem_cgroup_protection for more
6706 * TODO: calculation should be more robust so that we do not need
6707 * that special casing.
6712 usage = page_counter_read(&memcg->memory);
6716 parent = parent_mem_cgroup(memcg);
6717 /* No parent means a non-hierarchical mode on v1 memcg */
6721 if (parent == root) {
6722 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6723 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6727 parent_usage = page_counter_read(&parent->memory);
6729 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6730 READ_ONCE(memcg->memory.min),
6731 READ_ONCE(parent->memory.emin),
6732 atomic_long_read(&parent->memory.children_min_usage)));
6734 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6735 READ_ONCE(memcg->memory.low),
6736 READ_ONCE(parent->memory.elow),
6737 atomic_long_read(&parent->memory.children_low_usage)));
6741 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6742 * @page: page to charge
6743 * @mm: mm context of the victim
6744 * @gfp_mask: reclaim mode
6746 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6747 * pages according to @gfp_mask if necessary.
6749 * Returns 0 on success. Otherwise, an error code is returned.
6751 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6753 unsigned int nr_pages = thp_nr_pages(page);
6754 struct mem_cgroup *memcg = NULL;
6757 if (mem_cgroup_disabled())
6760 if (PageSwapCache(page)) {
6761 swp_entry_t ent = { .val = page_private(page), };
6765 * Every swap fault against a single page tries to charge the
6766 * page, bail as early as possible. shmem_unuse() encounters
6767 * already charged pages, too. page->mem_cgroup is protected
6768 * by the page lock, which serializes swap cache removal, which
6769 * in turn serializes uncharging.
6771 VM_BUG_ON_PAGE(!PageLocked(page), page);
6772 if (compound_head(page)->mem_cgroup)
6775 id = lookup_swap_cgroup_id(ent);
6777 memcg = mem_cgroup_from_id(id);
6778 if (memcg && !css_tryget_online(&memcg->css))
6784 memcg = get_mem_cgroup_from_mm(mm);
6786 ret = try_charge(memcg, gfp_mask, nr_pages);
6790 css_get(&memcg->css);
6791 commit_charge(page, memcg);
6793 local_irq_disable();
6794 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6795 memcg_check_events(memcg, page);
6798 if (PageSwapCache(page)) {
6799 swp_entry_t entry = { .val = page_private(page) };
6801 * The swap entry might not get freed for a long time,
6802 * let's not wait for it. The page already received a
6803 * memory+swap charge, drop the swap entry duplicate.
6805 mem_cgroup_uncharge_swap(entry, nr_pages);
6809 css_put(&memcg->css);
6814 struct uncharge_gather {
6815 struct mem_cgroup *memcg;
6816 unsigned long nr_pages;
6817 unsigned long pgpgout;
6818 unsigned long nr_kmem;
6819 struct page *dummy_page;
6822 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6824 memset(ug, 0, sizeof(*ug));
6827 static void uncharge_batch(const struct uncharge_gather *ug)
6829 unsigned long flags;
6831 if (!mem_cgroup_is_root(ug->memcg)) {
6832 page_counter_uncharge(&ug->memcg->memory, ug->nr_pages);
6833 if (do_memsw_account())
6834 page_counter_uncharge(&ug->memcg->memsw, ug->nr_pages);
6835 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6836 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6837 memcg_oom_recover(ug->memcg);
6840 local_irq_save(flags);
6841 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6842 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_pages);
6843 memcg_check_events(ug->memcg, ug->dummy_page);
6844 local_irq_restore(flags);
6846 /* drop reference from uncharge_page */
6847 css_put(&ug->memcg->css);
6850 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6852 unsigned long nr_pages;
6854 VM_BUG_ON_PAGE(PageLRU(page), page);
6856 if (!page->mem_cgroup)
6860 * Nobody should be changing or seriously looking at
6861 * page->mem_cgroup at this point, we have fully
6862 * exclusive access to the page.
6865 if (ug->memcg != page->mem_cgroup) {
6868 uncharge_gather_clear(ug);
6870 ug->memcg = page->mem_cgroup;
6872 /* pairs with css_put in uncharge_batch */
6873 css_get(&ug->memcg->css);
6876 nr_pages = compound_nr(page);
6877 ug->nr_pages += nr_pages;
6879 if (!PageKmemcg(page)) {
6882 ug->nr_kmem += nr_pages;
6883 __ClearPageKmemcg(page);
6886 ug->dummy_page = page;
6887 page->mem_cgroup = NULL;
6888 css_put(&ug->memcg->css);
6891 static void uncharge_list(struct list_head *page_list)
6893 struct uncharge_gather ug;
6894 struct list_head *next;
6896 uncharge_gather_clear(&ug);
6899 * Note that the list can be a single page->lru; hence the
6900 * do-while loop instead of a simple list_for_each_entry().
6902 next = page_list->next;
6906 page = list_entry(next, struct page, lru);
6907 next = page->lru.next;
6909 uncharge_page(page, &ug);
6910 } while (next != page_list);
6913 uncharge_batch(&ug);
6917 * mem_cgroup_uncharge - uncharge a page
6918 * @page: page to uncharge
6920 * Uncharge a page previously charged with mem_cgroup_charge().
6922 void mem_cgroup_uncharge(struct page *page)
6924 struct uncharge_gather ug;
6926 if (mem_cgroup_disabled())
6929 /* Don't touch page->lru of any random page, pre-check: */
6930 if (!page->mem_cgroup)
6933 uncharge_gather_clear(&ug);
6934 uncharge_page(page, &ug);
6935 uncharge_batch(&ug);
6939 * mem_cgroup_uncharge_list - uncharge a list of page
6940 * @page_list: list of pages to uncharge
6942 * Uncharge a list of pages previously charged with
6943 * mem_cgroup_charge().
6945 void mem_cgroup_uncharge_list(struct list_head *page_list)
6947 if (mem_cgroup_disabled())
6950 if (!list_empty(page_list))
6951 uncharge_list(page_list);
6955 * mem_cgroup_migrate - charge a page's replacement
6956 * @oldpage: currently circulating page
6957 * @newpage: replacement page
6959 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6960 * be uncharged upon free.
6962 * Both pages must be locked, @newpage->mapping must be set up.
6964 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6966 struct mem_cgroup *memcg;
6967 unsigned int nr_pages;
6968 unsigned long flags;
6970 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6971 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6972 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6973 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6976 if (mem_cgroup_disabled())
6979 /* Page cache replacement: new page already charged? */
6980 if (newpage->mem_cgroup)
6983 /* Swapcache readahead pages can get replaced before being charged */
6984 memcg = oldpage->mem_cgroup;
6988 /* Force-charge the new page. The old one will be freed soon */
6989 nr_pages = thp_nr_pages(newpage);
6991 page_counter_charge(&memcg->memory, nr_pages);
6992 if (do_memsw_account())
6993 page_counter_charge(&memcg->memsw, nr_pages);
6995 css_get(&memcg->css);
6996 commit_charge(newpage, memcg);
6998 local_irq_save(flags);
6999 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
7000 memcg_check_events(memcg, newpage);
7001 local_irq_restore(flags);
7004 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
7005 EXPORT_SYMBOL(memcg_sockets_enabled_key);
7007 void mem_cgroup_sk_alloc(struct sock *sk)
7009 struct mem_cgroup *memcg;
7011 if (!mem_cgroup_sockets_enabled)
7014 /* Do not associate the sock with unrelated interrupted task's memcg. */
7019 memcg = mem_cgroup_from_task(current);
7020 if (memcg == root_mem_cgroup)
7022 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
7024 if (css_tryget(&memcg->css))
7025 sk->sk_memcg = memcg;
7030 void mem_cgroup_sk_free(struct sock *sk)
7033 css_put(&sk->sk_memcg->css);
7037 * mem_cgroup_charge_skmem - charge socket memory
7038 * @memcg: memcg to charge
7039 * @nr_pages: number of pages to charge
7041 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7042 * @memcg's configured limit, %false if the charge had to be forced.
7044 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7046 gfp_t gfp_mask = GFP_KERNEL;
7048 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7049 struct page_counter *fail;
7051 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7052 memcg->tcpmem_pressure = 0;
7055 page_counter_charge(&memcg->tcpmem, nr_pages);
7056 memcg->tcpmem_pressure = 1;
7060 /* Don't block in the packet receive path */
7062 gfp_mask = GFP_NOWAIT;
7064 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7066 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
7069 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
7074 * mem_cgroup_uncharge_skmem - uncharge socket memory
7075 * @memcg: memcg to uncharge
7076 * @nr_pages: number of pages to uncharge
7078 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7080 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7081 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7085 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7087 refill_stock(memcg, nr_pages);
7090 static int __init cgroup_memory(char *s)
7094 while ((token = strsep(&s, ",")) != NULL) {
7097 if (!strcmp(token, "nosocket"))
7098 cgroup_memory_nosocket = true;
7099 if (!strcmp(token, "nokmem"))
7100 cgroup_memory_nokmem = true;
7104 __setup("cgroup.memory=", cgroup_memory);
7107 * subsys_initcall() for memory controller.
7109 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7110 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7111 * basically everything that doesn't depend on a specific mem_cgroup structure
7112 * should be initialized from here.
7114 static int __init mem_cgroup_init(void)
7118 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7119 memcg_hotplug_cpu_dead);
7121 for_each_possible_cpu(cpu)
7122 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7125 for_each_node(node) {
7126 struct mem_cgroup_tree_per_node *rtpn;
7128 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7129 node_online(node) ? node : NUMA_NO_NODE);
7131 rtpn->rb_root = RB_ROOT;
7132 rtpn->rb_rightmost = NULL;
7133 spin_lock_init(&rtpn->lock);
7134 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7139 subsys_initcall(mem_cgroup_init);
7141 #ifdef CONFIG_MEMCG_SWAP
7142 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7144 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7146 * The root cgroup cannot be destroyed, so it's refcount must
7149 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7153 memcg = parent_mem_cgroup(memcg);
7155 memcg = root_mem_cgroup;
7161 * mem_cgroup_swapout - transfer a memsw charge to swap
7162 * @page: page whose memsw charge to transfer
7163 * @entry: swap entry to move the charge to
7165 * Transfer the memsw charge of @page to @entry.
7167 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7169 struct mem_cgroup *memcg, *swap_memcg;
7170 unsigned int nr_entries;
7171 unsigned short oldid;
7173 VM_BUG_ON_PAGE(PageLRU(page), page);
7174 VM_BUG_ON_PAGE(page_count(page), page);
7176 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7179 memcg = page->mem_cgroup;
7181 /* Readahead page, never charged */
7186 * In case the memcg owning these pages has been offlined and doesn't
7187 * have an ID allocated to it anymore, charge the closest online
7188 * ancestor for the swap instead and transfer the memory+swap charge.
7190 swap_memcg = mem_cgroup_id_get_online(memcg);
7191 nr_entries = thp_nr_pages(page);
7192 /* Get references for the tail pages, too */
7194 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7195 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7197 VM_BUG_ON_PAGE(oldid, page);
7198 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7200 page->mem_cgroup = NULL;
7202 if (!mem_cgroup_is_root(memcg))
7203 page_counter_uncharge(&memcg->memory, nr_entries);
7205 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7206 if (!mem_cgroup_is_root(swap_memcg))
7207 page_counter_charge(&swap_memcg->memsw, nr_entries);
7208 page_counter_uncharge(&memcg->memsw, nr_entries);
7212 * Interrupts should be disabled here because the caller holds the
7213 * i_pages lock which is taken with interrupts-off. It is
7214 * important here to have the interrupts disabled because it is the
7215 * only synchronisation we have for updating the per-CPU variables.
7217 VM_BUG_ON(!irqs_disabled());
7218 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7219 memcg_check_events(memcg, page);
7221 css_put(&memcg->css);
7225 * mem_cgroup_try_charge_swap - try charging swap space for a page
7226 * @page: page being added to swap
7227 * @entry: swap entry to charge
7229 * Try to charge @page's memcg for the swap space at @entry.
7231 * Returns 0 on success, -ENOMEM on failure.
7233 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7235 unsigned int nr_pages = thp_nr_pages(page);
7236 struct page_counter *counter;
7237 struct mem_cgroup *memcg;
7238 unsigned short oldid;
7240 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7243 memcg = page->mem_cgroup;
7245 /* Readahead page, never charged */
7250 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7254 memcg = mem_cgroup_id_get_online(memcg);
7256 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7257 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7258 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7259 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7260 mem_cgroup_id_put(memcg);
7264 /* Get references for the tail pages, too */
7266 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7267 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7268 VM_BUG_ON_PAGE(oldid, page);
7269 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7275 * mem_cgroup_uncharge_swap - uncharge swap space
7276 * @entry: swap entry to uncharge
7277 * @nr_pages: the amount of swap space to uncharge
7279 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7281 struct mem_cgroup *memcg;
7284 id = swap_cgroup_record(entry, 0, nr_pages);
7286 memcg = mem_cgroup_from_id(id);
7288 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7289 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7290 page_counter_uncharge(&memcg->swap, nr_pages);
7292 page_counter_uncharge(&memcg->memsw, nr_pages);
7294 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7295 mem_cgroup_id_put_many(memcg, nr_pages);
7300 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7302 long nr_swap_pages = get_nr_swap_pages();
7304 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7305 return nr_swap_pages;
7306 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7307 nr_swap_pages = min_t(long, nr_swap_pages,
7308 READ_ONCE(memcg->swap.max) -
7309 page_counter_read(&memcg->swap));
7310 return nr_swap_pages;
7313 bool mem_cgroup_swap_full(struct page *page)
7315 struct mem_cgroup *memcg;
7317 VM_BUG_ON_PAGE(!PageLocked(page), page);
7321 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7324 memcg = page->mem_cgroup;
7328 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7329 unsigned long usage = page_counter_read(&memcg->swap);
7331 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7332 usage * 2 >= READ_ONCE(memcg->swap.max))
7339 static int __init setup_swap_account(char *s)
7341 if (!strcmp(s, "1"))
7342 cgroup_memory_noswap = 0;
7343 else if (!strcmp(s, "0"))
7344 cgroup_memory_noswap = 1;
7347 __setup("swapaccount=", setup_swap_account);
7349 static u64 swap_current_read(struct cgroup_subsys_state *css,
7352 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7354 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7357 static int swap_high_show(struct seq_file *m, void *v)
7359 return seq_puts_memcg_tunable(m,
7360 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7363 static ssize_t swap_high_write(struct kernfs_open_file *of,
7364 char *buf, size_t nbytes, loff_t off)
7366 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7370 buf = strstrip(buf);
7371 err = page_counter_memparse(buf, "max", &high);
7375 page_counter_set_high(&memcg->swap, high);
7380 static int swap_max_show(struct seq_file *m, void *v)
7382 return seq_puts_memcg_tunable(m,
7383 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7386 static ssize_t swap_max_write(struct kernfs_open_file *of,
7387 char *buf, size_t nbytes, loff_t off)
7389 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7393 buf = strstrip(buf);
7394 err = page_counter_memparse(buf, "max", &max);
7398 xchg(&memcg->swap.max, max);
7403 static int swap_events_show(struct seq_file *m, void *v)
7405 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7407 seq_printf(m, "high %lu\n",
7408 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7409 seq_printf(m, "max %lu\n",
7410 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7411 seq_printf(m, "fail %lu\n",
7412 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7417 static struct cftype swap_files[] = {
7419 .name = "swap.current",
7420 .flags = CFTYPE_NOT_ON_ROOT,
7421 .read_u64 = swap_current_read,
7424 .name = "swap.high",
7425 .flags = CFTYPE_NOT_ON_ROOT,
7426 .seq_show = swap_high_show,
7427 .write = swap_high_write,
7431 .flags = CFTYPE_NOT_ON_ROOT,
7432 .seq_show = swap_max_show,
7433 .write = swap_max_write,
7436 .name = "swap.events",
7437 .flags = CFTYPE_NOT_ON_ROOT,
7438 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7439 .seq_show = swap_events_show,
7444 static struct cftype memsw_files[] = {
7446 .name = "memsw.usage_in_bytes",
7447 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7448 .read_u64 = mem_cgroup_read_u64,
7451 .name = "memsw.max_usage_in_bytes",
7452 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7453 .write = mem_cgroup_reset,
7454 .read_u64 = mem_cgroup_read_u64,
7457 .name = "memsw.limit_in_bytes",
7458 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7459 .write = mem_cgroup_write,
7460 .read_u64 = mem_cgroup_read_u64,
7463 .name = "memsw.failcnt",
7464 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7465 .write = mem_cgroup_reset,
7466 .read_u64 = mem_cgroup_read_u64,
7468 { }, /* terminate */
7472 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7473 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7474 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7475 * boot parameter. This may result in premature OOPS inside
7476 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7478 static int __init mem_cgroup_swap_init(void)
7480 /* No memory control -> no swap control */
7481 if (mem_cgroup_disabled())
7482 cgroup_memory_noswap = true;
7484 if (cgroup_memory_noswap)
7487 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7488 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7492 core_initcall(mem_cgroup_swap_init);
7494 #endif /* CONFIG_MEMCG_SWAP */