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 #define MEM_CGROUP_RECLAIM_RETRIES 5
78 /* Socket memory accounting disabled? */
79 static bool cgroup_memory_nosocket;
81 /* Kernel memory accounting disabled? */
82 static bool cgroup_memory_nokmem;
84 /* Whether the swap controller is active */
85 #ifdef CONFIG_MEMCG_SWAP
86 int do_swap_account __read_mostly;
88 #define do_swap_account 0
91 #ifdef CONFIG_CGROUP_WRITEBACK
92 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
95 /* Whether legacy memory+swap accounting is active */
96 static bool do_memsw_account(void)
98 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && do_swap_account;
101 static const char *const mem_cgroup_lru_names[] = {
109 #define THRESHOLDS_EVENTS_TARGET 128
110 #define SOFTLIMIT_EVENTS_TARGET 1024
111 #define NUMAINFO_EVENTS_TARGET 1024
114 * Cgroups above their limits are maintained in a RB-Tree, independent of
115 * their hierarchy representation
118 struct mem_cgroup_tree_per_node {
119 struct rb_root rb_root;
120 struct rb_node *rb_rightmost;
124 struct mem_cgroup_tree {
125 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
128 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
131 struct mem_cgroup_eventfd_list {
132 struct list_head list;
133 struct eventfd_ctx *eventfd;
137 * cgroup_event represents events which userspace want to receive.
139 struct mem_cgroup_event {
141 * memcg which the event belongs to.
143 struct mem_cgroup *memcg;
145 * eventfd to signal userspace about the event.
147 struct eventfd_ctx *eventfd;
149 * Each of these stored in a list by the cgroup.
151 struct list_head list;
153 * register_event() callback will be used to add new userspace
154 * waiter for changes related to this event. Use eventfd_signal()
155 * on eventfd to send notification to userspace.
157 int (*register_event)(struct mem_cgroup *memcg,
158 struct eventfd_ctx *eventfd, const char *args);
160 * unregister_event() callback will be called when userspace closes
161 * the eventfd or on cgroup removing. This callback must be set,
162 * if you want provide notification functionality.
164 void (*unregister_event)(struct mem_cgroup *memcg,
165 struct eventfd_ctx *eventfd);
167 * All fields below needed to unregister event when
168 * userspace closes eventfd.
171 wait_queue_head_t *wqh;
172 wait_queue_entry_t wait;
173 struct work_struct remove;
176 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
177 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
179 /* Stuffs for move charges at task migration. */
181 * Types of charges to be moved.
183 #define MOVE_ANON 0x1U
184 #define MOVE_FILE 0x2U
185 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
187 /* "mc" and its members are protected by cgroup_mutex */
188 static struct move_charge_struct {
189 spinlock_t lock; /* for from, to */
190 struct mm_struct *mm;
191 struct mem_cgroup *from;
192 struct mem_cgroup *to;
194 unsigned long precharge;
195 unsigned long moved_charge;
196 unsigned long moved_swap;
197 struct task_struct *moving_task; /* a task moving charges */
198 wait_queue_head_t waitq; /* a waitq for other context */
200 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
201 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
205 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
206 * limit reclaim to prevent infinite loops, if they ever occur.
208 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
209 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
212 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
213 MEM_CGROUP_CHARGE_TYPE_ANON,
214 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
215 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
219 /* for encoding cft->private value on file */
228 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
229 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
230 #define MEMFILE_ATTR(val) ((val) & 0xffff)
231 /* Used for OOM nofiier */
232 #define OOM_CONTROL (0)
235 * Iteration constructs for visiting all cgroups (under a tree). If
236 * loops are exited prematurely (break), mem_cgroup_iter_break() must
237 * be used for reference counting.
239 #define for_each_mem_cgroup_tree(iter, root) \
240 for (iter = mem_cgroup_iter(root, NULL, NULL); \
242 iter = mem_cgroup_iter(root, iter, NULL))
244 #define for_each_mem_cgroup(iter) \
245 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
247 iter = mem_cgroup_iter(NULL, iter, NULL))
249 static inline bool should_force_charge(void)
251 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
252 (current->flags & PF_EXITING);
255 /* Some nice accessors for the vmpressure. */
256 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
259 memcg = root_mem_cgroup;
260 return &memcg->vmpressure;
263 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
265 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
268 #ifdef CONFIG_MEMCG_KMEM
270 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
271 * The main reason for not using cgroup id for this:
272 * this works better in sparse environments, where we have a lot of memcgs,
273 * but only a few kmem-limited. Or also, if we have, for instance, 200
274 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
275 * 200 entry array for that.
277 * The current size of the caches array is stored in memcg_nr_cache_ids. It
278 * will double each time we have to increase it.
280 static DEFINE_IDA(memcg_cache_ida);
281 int memcg_nr_cache_ids;
283 /* Protects memcg_nr_cache_ids */
284 static DECLARE_RWSEM(memcg_cache_ids_sem);
286 void memcg_get_cache_ids(void)
288 down_read(&memcg_cache_ids_sem);
291 void memcg_put_cache_ids(void)
293 up_read(&memcg_cache_ids_sem);
297 * MIN_SIZE is different than 1, because we would like to avoid going through
298 * the alloc/free process all the time. In a small machine, 4 kmem-limited
299 * cgroups is a reasonable guess. In the future, it could be a parameter or
300 * tunable, but that is strictly not necessary.
302 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
303 * this constant directly from cgroup, but it is understandable that this is
304 * better kept as an internal representation in cgroup.c. In any case, the
305 * cgrp_id space is not getting any smaller, and we don't have to necessarily
306 * increase ours as well if it increases.
308 #define MEMCG_CACHES_MIN_SIZE 4
309 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
312 * A lot of the calls to the cache allocation functions are expected to be
313 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
314 * conditional to this static branch, we'll have to allow modules that does
315 * kmem_cache_alloc and the such to see this symbol as well
317 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
318 EXPORT_SYMBOL(memcg_kmem_enabled_key);
320 struct workqueue_struct *memcg_kmem_cache_wq;
322 static int memcg_shrinker_map_size;
323 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
325 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
327 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
330 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
331 int size, int old_size)
333 struct memcg_shrinker_map *new, *old;
336 lockdep_assert_held(&memcg_shrinker_map_mutex);
339 old = rcu_dereference_protected(
340 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
341 /* Not yet online memcg */
345 new = kvmalloc(sizeof(*new) + size, GFP_KERNEL);
349 /* Set all old bits, clear all new bits */
350 memset(new->map, (int)0xff, old_size);
351 memset((void *)new->map + old_size, 0, size - old_size);
353 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
354 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
360 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
362 struct mem_cgroup_per_node *pn;
363 struct memcg_shrinker_map *map;
366 if (mem_cgroup_is_root(memcg))
370 pn = mem_cgroup_nodeinfo(memcg, nid);
371 map = rcu_dereference_protected(pn->shrinker_map, true);
374 rcu_assign_pointer(pn->shrinker_map, NULL);
378 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
380 struct memcg_shrinker_map *map;
381 int nid, size, ret = 0;
383 if (mem_cgroup_is_root(memcg))
386 mutex_lock(&memcg_shrinker_map_mutex);
387 size = memcg_shrinker_map_size;
389 map = kvzalloc(sizeof(*map) + size, GFP_KERNEL);
391 memcg_free_shrinker_maps(memcg);
395 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
397 mutex_unlock(&memcg_shrinker_map_mutex);
402 int memcg_expand_shrinker_maps(int new_id)
404 int size, old_size, ret = 0;
405 struct mem_cgroup *memcg;
407 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
408 old_size = memcg_shrinker_map_size;
409 if (size <= old_size)
412 mutex_lock(&memcg_shrinker_map_mutex);
413 if (!root_mem_cgroup)
416 for_each_mem_cgroup(memcg) {
417 if (mem_cgroup_is_root(memcg))
419 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
425 memcg_shrinker_map_size = size;
426 mutex_unlock(&memcg_shrinker_map_mutex);
430 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
432 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
433 struct memcg_shrinker_map *map;
436 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
437 /* Pairs with smp mb in shrink_slab() */
438 smp_mb__before_atomic();
439 set_bit(shrinker_id, map->map);
444 #else /* CONFIG_MEMCG_KMEM */
445 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
449 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg) { }
450 #endif /* CONFIG_MEMCG_KMEM */
453 * mem_cgroup_css_from_page - css of the memcg associated with a page
454 * @page: page of interest
456 * If memcg is bound to the default hierarchy, css of the memcg associated
457 * with @page is returned. The returned css remains associated with @page
458 * until it is released.
460 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
463 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
465 struct mem_cgroup *memcg;
467 memcg = page->mem_cgroup;
469 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
470 memcg = root_mem_cgroup;
476 * page_cgroup_ino - return inode number of the memcg a page is charged to
479 * Look up the closest online ancestor of the memory cgroup @page is charged to
480 * and return its inode number or 0 if @page is not charged to any cgroup. It
481 * is safe to call this function without holding a reference to @page.
483 * Note, this function is inherently racy, because there is nothing to prevent
484 * the cgroup inode from getting torn down and potentially reallocated a moment
485 * after page_cgroup_ino() returns, so it only should be used by callers that
486 * do not care (such as procfs interfaces).
488 ino_t page_cgroup_ino(struct page *page)
490 struct mem_cgroup *memcg;
491 unsigned long ino = 0;
494 if (PageHead(page) && PageSlab(page))
495 memcg = memcg_from_slab_page(page);
497 memcg = READ_ONCE(page->mem_cgroup);
498 while (memcg && !(memcg->css.flags & CSS_ONLINE))
499 memcg = parent_mem_cgroup(memcg);
501 ino = cgroup_ino(memcg->css.cgroup);
506 static struct mem_cgroup_per_node *
507 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
509 int nid = page_to_nid(page);
511 return memcg->nodeinfo[nid];
514 static struct mem_cgroup_tree_per_node *
515 soft_limit_tree_node(int nid)
517 return soft_limit_tree.rb_tree_per_node[nid];
520 static struct mem_cgroup_tree_per_node *
521 soft_limit_tree_from_page(struct page *page)
523 int nid = page_to_nid(page);
525 return soft_limit_tree.rb_tree_per_node[nid];
528 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
529 struct mem_cgroup_tree_per_node *mctz,
530 unsigned long new_usage_in_excess)
532 struct rb_node **p = &mctz->rb_root.rb_node;
533 struct rb_node *parent = NULL;
534 struct mem_cgroup_per_node *mz_node;
535 bool rightmost = true;
540 mz->usage_in_excess = new_usage_in_excess;
541 if (!mz->usage_in_excess)
545 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
547 if (mz->usage_in_excess < mz_node->usage_in_excess) {
553 * We can't avoid mem cgroups that are over their soft
554 * limit by the same amount
556 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
561 mctz->rb_rightmost = &mz->tree_node;
563 rb_link_node(&mz->tree_node, parent, p);
564 rb_insert_color(&mz->tree_node, &mctz->rb_root);
568 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
569 struct mem_cgroup_tree_per_node *mctz)
574 if (&mz->tree_node == mctz->rb_rightmost)
575 mctz->rb_rightmost = rb_prev(&mz->tree_node);
577 rb_erase(&mz->tree_node, &mctz->rb_root);
581 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
582 struct mem_cgroup_tree_per_node *mctz)
586 spin_lock_irqsave(&mctz->lock, flags);
587 __mem_cgroup_remove_exceeded(mz, mctz);
588 spin_unlock_irqrestore(&mctz->lock, flags);
591 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
593 unsigned long nr_pages = page_counter_read(&memcg->memory);
594 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
595 unsigned long excess = 0;
597 if (nr_pages > soft_limit)
598 excess = nr_pages - soft_limit;
603 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
605 unsigned long excess;
606 struct mem_cgroup_per_node *mz;
607 struct mem_cgroup_tree_per_node *mctz;
609 mctz = soft_limit_tree_from_page(page);
613 * Necessary to update all ancestors when hierarchy is used.
614 * because their event counter is not touched.
616 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
617 mz = mem_cgroup_page_nodeinfo(memcg, page);
618 excess = soft_limit_excess(memcg);
620 * We have to update the tree if mz is on RB-tree or
621 * mem is over its softlimit.
623 if (excess || mz->on_tree) {
626 spin_lock_irqsave(&mctz->lock, flags);
627 /* if on-tree, remove it */
629 __mem_cgroup_remove_exceeded(mz, mctz);
631 * Insert again. mz->usage_in_excess will be updated.
632 * If excess is 0, no tree ops.
634 __mem_cgroup_insert_exceeded(mz, mctz, excess);
635 spin_unlock_irqrestore(&mctz->lock, flags);
640 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
642 struct mem_cgroup_tree_per_node *mctz;
643 struct mem_cgroup_per_node *mz;
647 mz = mem_cgroup_nodeinfo(memcg, nid);
648 mctz = soft_limit_tree_node(nid);
650 mem_cgroup_remove_exceeded(mz, mctz);
654 static struct mem_cgroup_per_node *
655 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
657 struct mem_cgroup_per_node *mz;
661 if (!mctz->rb_rightmost)
662 goto done; /* Nothing to reclaim from */
664 mz = rb_entry(mctz->rb_rightmost,
665 struct mem_cgroup_per_node, tree_node);
667 * Remove the node now but someone else can add it back,
668 * we will to add it back at the end of reclaim to its correct
669 * position in the tree.
671 __mem_cgroup_remove_exceeded(mz, mctz);
672 if (!soft_limit_excess(mz->memcg) ||
673 !css_tryget_online(&mz->memcg->css))
679 static struct mem_cgroup_per_node *
680 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
682 struct mem_cgroup_per_node *mz;
684 spin_lock_irq(&mctz->lock);
685 mz = __mem_cgroup_largest_soft_limit_node(mctz);
686 spin_unlock_irq(&mctz->lock);
691 * __mod_memcg_state - update cgroup memory statistics
692 * @memcg: the memory cgroup
693 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
694 * @val: delta to add to the counter, can be negative
696 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
700 if (mem_cgroup_disabled())
703 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
704 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
705 struct mem_cgroup *mi;
708 * Batch local counters to keep them in sync with
709 * the hierarchical ones.
711 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
712 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
713 atomic_long_add(x, &mi->vmstats[idx]);
716 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
719 static struct mem_cgroup_per_node *
720 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
722 struct mem_cgroup *parent;
724 parent = parent_mem_cgroup(pn->memcg);
727 return mem_cgroup_nodeinfo(parent, nid);
731 * __mod_lruvec_state - update lruvec memory statistics
732 * @lruvec: the lruvec
733 * @idx: the stat item
734 * @val: delta to add to the counter, can be negative
736 * The lruvec is the intersection of the NUMA node and a cgroup. This
737 * function updates the all three counters that are affected by a
738 * change of state at this level: per-node, per-cgroup, per-lruvec.
740 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
743 pg_data_t *pgdat = lruvec_pgdat(lruvec);
744 struct mem_cgroup_per_node *pn;
745 struct mem_cgroup *memcg;
749 __mod_node_page_state(pgdat, idx, val);
751 if (mem_cgroup_disabled())
754 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
758 __mod_memcg_state(memcg, idx, val);
761 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
763 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
764 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
765 struct mem_cgroup_per_node *pi;
767 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
768 atomic_long_add(x, &pi->lruvec_stat[idx]);
771 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
774 void __mod_lruvec_slab_state(void *p, enum node_stat_item idx, int val)
776 struct page *page = virt_to_head_page(p);
777 pg_data_t *pgdat = page_pgdat(page);
778 struct mem_cgroup *memcg;
779 struct lruvec *lruvec;
782 memcg = memcg_from_slab_page(page);
784 /* Untracked pages have no memcg, no lruvec. Update only the node */
785 if (!memcg || memcg == root_mem_cgroup) {
786 __mod_node_page_state(pgdat, idx, val);
788 lruvec = mem_cgroup_lruvec(pgdat, memcg);
789 __mod_lruvec_state(lruvec, idx, val);
795 * __count_memcg_events - account VM events in a cgroup
796 * @memcg: the memory cgroup
797 * @idx: the event item
798 * @count: the number of events that occured
800 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
805 if (mem_cgroup_disabled())
808 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
809 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
810 struct mem_cgroup *mi;
813 * Batch local counters to keep them in sync with
814 * the hierarchical ones.
816 __this_cpu_add(memcg->vmstats_local->events[idx], x);
817 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
818 atomic_long_add(x, &mi->vmevents[idx]);
821 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
824 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
826 return atomic_long_read(&memcg->vmevents[event]);
829 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
834 for_each_possible_cpu(cpu)
835 x += per_cpu(memcg->vmstats_local->events[event], cpu);
839 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
841 bool compound, int nr_pages)
844 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
845 * counted as CACHE even if it's on ANON LRU.
848 __mod_memcg_state(memcg, MEMCG_RSS, nr_pages);
850 __mod_memcg_state(memcg, MEMCG_CACHE, nr_pages);
851 if (PageSwapBacked(page))
852 __mod_memcg_state(memcg, NR_SHMEM, nr_pages);
856 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
857 __mod_memcg_state(memcg, MEMCG_RSS_HUGE, nr_pages);
860 /* pagein of a big page is an event. So, ignore page size */
862 __count_memcg_events(memcg, PGPGIN, 1);
864 __count_memcg_events(memcg, PGPGOUT, 1);
865 nr_pages = -nr_pages; /* for event */
868 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
871 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
872 enum mem_cgroup_events_target target)
874 unsigned long val, next;
876 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
877 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
878 /* from time_after() in jiffies.h */
879 if ((long)(next - val) < 0) {
881 case MEM_CGROUP_TARGET_THRESH:
882 next = val + THRESHOLDS_EVENTS_TARGET;
884 case MEM_CGROUP_TARGET_SOFTLIMIT:
885 next = val + SOFTLIMIT_EVENTS_TARGET;
887 case MEM_CGROUP_TARGET_NUMAINFO:
888 next = val + NUMAINFO_EVENTS_TARGET;
893 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
900 * Check events in order.
903 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
905 /* threshold event is triggered in finer grain than soft limit */
906 if (unlikely(mem_cgroup_event_ratelimit(memcg,
907 MEM_CGROUP_TARGET_THRESH))) {
909 bool do_numainfo __maybe_unused;
911 do_softlimit = mem_cgroup_event_ratelimit(memcg,
912 MEM_CGROUP_TARGET_SOFTLIMIT);
914 do_numainfo = mem_cgroup_event_ratelimit(memcg,
915 MEM_CGROUP_TARGET_NUMAINFO);
917 mem_cgroup_threshold(memcg);
918 if (unlikely(do_softlimit))
919 mem_cgroup_update_tree(memcg, page);
921 if (unlikely(do_numainfo))
922 atomic_inc(&memcg->numainfo_events);
927 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
930 * mm_update_next_owner() may clear mm->owner to NULL
931 * if it races with swapoff, page migration, etc.
932 * So this can be called with p == NULL.
937 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
939 EXPORT_SYMBOL(mem_cgroup_from_task);
942 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
943 * @mm: mm from which memcg should be extracted. It can be NULL.
945 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
946 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
949 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
951 struct mem_cgroup *memcg;
953 if (mem_cgroup_disabled())
959 * Page cache insertions can happen withou an
960 * actual mm context, e.g. during disk probing
961 * on boot, loopback IO, acct() writes etc.
964 memcg = root_mem_cgroup;
966 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
967 if (unlikely(!memcg))
968 memcg = root_mem_cgroup;
970 } while (!css_tryget_online(&memcg->css));
974 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
977 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
978 * @page: page from which memcg should be extracted.
980 * Obtain a reference on page->memcg and returns it if successful. Otherwise
981 * root_mem_cgroup is returned.
983 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
985 struct mem_cgroup *memcg = page->mem_cgroup;
987 if (mem_cgroup_disabled())
991 if (!memcg || !css_tryget_online(&memcg->css))
992 memcg = root_mem_cgroup;
996 EXPORT_SYMBOL(get_mem_cgroup_from_page);
999 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
1001 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
1003 if (unlikely(current->active_memcg)) {
1004 struct mem_cgroup *memcg = root_mem_cgroup;
1007 if (css_tryget_online(¤t->active_memcg->css))
1008 memcg = current->active_memcg;
1012 return get_mem_cgroup_from_mm(current->mm);
1016 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1017 * @root: hierarchy root
1018 * @prev: previously returned memcg, NULL on first invocation
1019 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1021 * Returns references to children of the hierarchy below @root, or
1022 * @root itself, or %NULL after a full round-trip.
1024 * Caller must pass the return value in @prev on subsequent
1025 * invocations for reference counting, or use mem_cgroup_iter_break()
1026 * to cancel a hierarchy walk before the round-trip is complete.
1028 * Reclaimers can specify a node and a priority level in @reclaim to
1029 * divide up the memcgs in the hierarchy among all concurrent
1030 * reclaimers operating on the same node and priority.
1032 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1033 struct mem_cgroup *prev,
1034 struct mem_cgroup_reclaim_cookie *reclaim)
1036 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1037 struct cgroup_subsys_state *css = NULL;
1038 struct mem_cgroup *memcg = NULL;
1039 struct mem_cgroup *pos = NULL;
1041 if (mem_cgroup_disabled())
1045 root = root_mem_cgroup;
1047 if (prev && !reclaim)
1050 if (!root->use_hierarchy && root != root_mem_cgroup) {
1059 struct mem_cgroup_per_node *mz;
1061 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1062 iter = &mz->iter[reclaim->priority];
1064 if (prev && reclaim->generation != iter->generation)
1068 pos = READ_ONCE(iter->position);
1069 if (!pos || css_tryget(&pos->css))
1072 * css reference reached zero, so iter->position will
1073 * be cleared by ->css_released. However, we should not
1074 * rely on this happening soon, because ->css_released
1075 * is called from a work queue, and by busy-waiting we
1076 * might block it. So we clear iter->position right
1079 (void)cmpxchg(&iter->position, pos, NULL);
1087 css = css_next_descendant_pre(css, &root->css);
1090 * Reclaimers share the hierarchy walk, and a
1091 * new one might jump in right at the end of
1092 * the hierarchy - make sure they see at least
1093 * one group and restart from the beginning.
1101 * Verify the css and acquire a reference. The root
1102 * is provided by the caller, so we know it's alive
1103 * and kicking, and don't take an extra reference.
1105 memcg = mem_cgroup_from_css(css);
1107 if (css == &root->css)
1110 if (css_tryget(css))
1118 * The position could have already been updated by a competing
1119 * thread, so check that the value hasn't changed since we read
1120 * it to avoid reclaiming from the same cgroup twice.
1122 (void)cmpxchg(&iter->position, pos, memcg);
1130 reclaim->generation = iter->generation;
1136 if (prev && prev != root)
1137 css_put(&prev->css);
1143 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1144 * @root: hierarchy root
1145 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1147 void mem_cgroup_iter_break(struct mem_cgroup *root,
1148 struct mem_cgroup *prev)
1151 root = root_mem_cgroup;
1152 if (prev && prev != root)
1153 css_put(&prev->css);
1156 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1157 struct mem_cgroup *dead_memcg)
1159 struct mem_cgroup_reclaim_iter *iter;
1160 struct mem_cgroup_per_node *mz;
1164 for_each_node(nid) {
1165 mz = mem_cgroup_nodeinfo(from, nid);
1166 for (i = 0; i <= DEF_PRIORITY; i++) {
1167 iter = &mz->iter[i];
1168 cmpxchg(&iter->position,
1174 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1176 struct mem_cgroup *memcg = dead_memcg;
1177 struct mem_cgroup *last;
1180 __invalidate_reclaim_iterators(memcg, dead_memcg);
1182 } while ((memcg = parent_mem_cgroup(memcg)));
1185 * When cgruop1 non-hierarchy mode is used,
1186 * parent_mem_cgroup() does not walk all the way up to the
1187 * cgroup root (root_mem_cgroup). So we have to handle
1188 * dead_memcg from cgroup root separately.
1190 if (last != root_mem_cgroup)
1191 __invalidate_reclaim_iterators(root_mem_cgroup,
1196 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1197 * @memcg: hierarchy root
1198 * @fn: function to call for each task
1199 * @arg: argument passed to @fn
1201 * This function iterates over tasks attached to @memcg or to any of its
1202 * descendants and calls @fn for each task. If @fn returns a non-zero
1203 * value, the function breaks the iteration loop and returns the value.
1204 * Otherwise, it will iterate over all tasks and return 0.
1206 * This function must not be called for the root memory cgroup.
1208 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1209 int (*fn)(struct task_struct *, void *), void *arg)
1211 struct mem_cgroup *iter;
1214 BUG_ON(memcg == root_mem_cgroup);
1216 for_each_mem_cgroup_tree(iter, memcg) {
1217 struct css_task_iter it;
1218 struct task_struct *task;
1220 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1221 while (!ret && (task = css_task_iter_next(&it)))
1222 ret = fn(task, arg);
1223 css_task_iter_end(&it);
1225 mem_cgroup_iter_break(memcg, iter);
1233 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1235 * @pgdat: pgdat of the page
1237 * This function is only safe when following the LRU page isolation
1238 * and putback protocol: the LRU lock must be held, and the page must
1239 * either be PageLRU() or the caller must have isolated/allocated it.
1241 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1243 struct mem_cgroup_per_node *mz;
1244 struct mem_cgroup *memcg;
1245 struct lruvec *lruvec;
1247 if (mem_cgroup_disabled()) {
1248 lruvec = &pgdat->lruvec;
1252 memcg = page->mem_cgroup;
1254 * Swapcache readahead pages are added to the LRU - and
1255 * possibly migrated - before they are charged.
1258 memcg = root_mem_cgroup;
1260 mz = mem_cgroup_page_nodeinfo(memcg, page);
1261 lruvec = &mz->lruvec;
1264 * Since a node can be onlined after the mem_cgroup was created,
1265 * we have to be prepared to initialize lruvec->zone here;
1266 * and if offlined then reonlined, we need to reinitialize it.
1268 if (unlikely(lruvec->pgdat != pgdat))
1269 lruvec->pgdat = pgdat;
1274 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1275 * @lruvec: mem_cgroup per zone lru vector
1276 * @lru: index of lru list the page is sitting on
1277 * @zid: zone id of the accounted pages
1278 * @nr_pages: positive when adding or negative when removing
1280 * This function must be called under lru_lock, just before a page is added
1281 * to or just after a page is removed from an lru list (that ordering being
1282 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1284 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1285 int zid, int nr_pages)
1287 struct mem_cgroup_per_node *mz;
1288 unsigned long *lru_size;
1291 if (mem_cgroup_disabled())
1294 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1295 lru_size = &mz->lru_zone_size[zid][lru];
1298 *lru_size += nr_pages;
1301 if (WARN_ONCE(size < 0,
1302 "%s(%p, %d, %d): lru_size %ld\n",
1303 __func__, lruvec, lru, nr_pages, size)) {
1309 *lru_size += nr_pages;
1313 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1314 * @memcg: the memory cgroup
1316 * Returns the maximum amount of memory @mem can be charged with, in
1319 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1321 unsigned long margin = 0;
1322 unsigned long count;
1323 unsigned long limit;
1325 count = page_counter_read(&memcg->memory);
1326 limit = READ_ONCE(memcg->memory.max);
1328 margin = limit - count;
1330 if (do_memsw_account()) {
1331 count = page_counter_read(&memcg->memsw);
1332 limit = READ_ONCE(memcg->memsw.max);
1334 margin = min(margin, limit - count);
1343 * A routine for checking "mem" is under move_account() or not.
1345 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1346 * moving cgroups. This is for waiting at high-memory pressure
1349 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1351 struct mem_cgroup *from;
1352 struct mem_cgroup *to;
1355 * Unlike task_move routines, we access mc.to, mc.from not under
1356 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1358 spin_lock(&mc.lock);
1364 ret = mem_cgroup_is_descendant(from, memcg) ||
1365 mem_cgroup_is_descendant(to, memcg);
1367 spin_unlock(&mc.lock);
1371 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1373 if (mc.moving_task && current != mc.moving_task) {
1374 if (mem_cgroup_under_move(memcg)) {
1376 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1377 /* moving charge context might have finished. */
1380 finish_wait(&mc.waitq, &wait);
1387 static char *memory_stat_format(struct mem_cgroup *memcg)
1392 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1397 * Provide statistics on the state of the memory subsystem as
1398 * well as cumulative event counters that show past behavior.
1400 * This list is ordered following a combination of these gradients:
1401 * 1) generic big picture -> specifics and details
1402 * 2) reflecting userspace activity -> reflecting kernel heuristics
1404 * Current memory state:
1407 seq_buf_printf(&s, "anon %llu\n",
1408 (u64)memcg_page_state(memcg, MEMCG_RSS) *
1410 seq_buf_printf(&s, "file %llu\n",
1411 (u64)memcg_page_state(memcg, MEMCG_CACHE) *
1413 seq_buf_printf(&s, "kernel_stack %llu\n",
1414 (u64)memcg_page_state(memcg, MEMCG_KERNEL_STACK_KB) *
1416 seq_buf_printf(&s, "slab %llu\n",
1417 (u64)(memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) +
1418 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE)) *
1420 seq_buf_printf(&s, "sock %llu\n",
1421 (u64)memcg_page_state(memcg, MEMCG_SOCK) *
1424 seq_buf_printf(&s, "shmem %llu\n",
1425 (u64)memcg_page_state(memcg, NR_SHMEM) *
1427 seq_buf_printf(&s, "file_mapped %llu\n",
1428 (u64)memcg_page_state(memcg, NR_FILE_MAPPED) *
1430 seq_buf_printf(&s, "file_dirty %llu\n",
1431 (u64)memcg_page_state(memcg, NR_FILE_DIRTY) *
1433 seq_buf_printf(&s, "file_writeback %llu\n",
1434 (u64)memcg_page_state(memcg, NR_WRITEBACK) *
1438 * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter
1439 * with the NR_ANON_THP vm counter, but right now it's a pain in the
1440 * arse because it requires migrating the work out of rmap to a place
1441 * where the page->mem_cgroup is set up and stable.
1443 seq_buf_printf(&s, "anon_thp %llu\n",
1444 (u64)memcg_page_state(memcg, MEMCG_RSS_HUGE) *
1447 for (i = 0; i < NR_LRU_LISTS; i++)
1448 seq_buf_printf(&s, "%s %llu\n", mem_cgroup_lru_names[i],
1449 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
1452 seq_buf_printf(&s, "slab_reclaimable %llu\n",
1453 (u64)memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) *
1455 seq_buf_printf(&s, "slab_unreclaimable %llu\n",
1456 (u64)memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE) *
1459 /* Accumulated memory events */
1461 seq_buf_printf(&s, "pgfault %lu\n", memcg_events(memcg, PGFAULT));
1462 seq_buf_printf(&s, "pgmajfault %lu\n", memcg_events(memcg, PGMAJFAULT));
1464 seq_buf_printf(&s, "workingset_refault %lu\n",
1465 memcg_page_state(memcg, WORKINGSET_REFAULT));
1466 seq_buf_printf(&s, "workingset_activate %lu\n",
1467 memcg_page_state(memcg, WORKINGSET_ACTIVATE));
1468 seq_buf_printf(&s, "workingset_nodereclaim %lu\n",
1469 memcg_page_state(memcg, WORKINGSET_NODERECLAIM));
1471 seq_buf_printf(&s, "pgrefill %lu\n", memcg_events(memcg, PGREFILL));
1472 seq_buf_printf(&s, "pgscan %lu\n",
1473 memcg_events(memcg, PGSCAN_KSWAPD) +
1474 memcg_events(memcg, PGSCAN_DIRECT));
1475 seq_buf_printf(&s, "pgsteal %lu\n",
1476 memcg_events(memcg, PGSTEAL_KSWAPD) +
1477 memcg_events(memcg, PGSTEAL_DIRECT));
1478 seq_buf_printf(&s, "pgactivate %lu\n", memcg_events(memcg, PGACTIVATE));
1479 seq_buf_printf(&s, "pgdeactivate %lu\n", memcg_events(memcg, PGDEACTIVATE));
1480 seq_buf_printf(&s, "pglazyfree %lu\n", memcg_events(memcg, PGLAZYFREE));
1481 seq_buf_printf(&s, "pglazyfreed %lu\n", memcg_events(memcg, PGLAZYFREED));
1483 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1484 seq_buf_printf(&s, "thp_fault_alloc %lu\n",
1485 memcg_events(memcg, THP_FAULT_ALLOC));
1486 seq_buf_printf(&s, "thp_collapse_alloc %lu\n",
1487 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1488 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1490 /* The above should easily fit into one page */
1491 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1496 #define K(x) ((x) << (PAGE_SHIFT-10))
1498 * mem_cgroup_print_oom_context: Print OOM information relevant to
1499 * memory controller.
1500 * @memcg: The memory cgroup that went over limit
1501 * @p: Task that is going to be killed
1503 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1506 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1511 pr_cont(",oom_memcg=");
1512 pr_cont_cgroup_path(memcg->css.cgroup);
1514 pr_cont(",global_oom");
1516 pr_cont(",task_memcg=");
1517 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1523 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1524 * memory controller.
1525 * @memcg: The memory cgroup that went over limit
1527 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1531 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1532 K((u64)page_counter_read(&memcg->memory)),
1533 K((u64)memcg->memory.max), memcg->memory.failcnt);
1534 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1535 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1536 K((u64)page_counter_read(&memcg->swap)),
1537 K((u64)memcg->swap.max), memcg->swap.failcnt);
1539 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1540 K((u64)page_counter_read(&memcg->memsw)),
1541 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1542 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1543 K((u64)page_counter_read(&memcg->kmem)),
1544 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1547 pr_info("Memory cgroup stats for ");
1548 pr_cont_cgroup_path(memcg->css.cgroup);
1550 buf = memory_stat_format(memcg);
1558 * Return the memory (and swap, if configured) limit for a memcg.
1560 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1564 max = memcg->memory.max;
1565 if (mem_cgroup_swappiness(memcg)) {
1566 unsigned long memsw_max;
1567 unsigned long swap_max;
1569 memsw_max = memcg->memsw.max;
1570 swap_max = memcg->swap.max;
1571 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1572 max = min(max + swap_max, memsw_max);
1577 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1580 struct oom_control oc = {
1584 .gfp_mask = gfp_mask,
1589 if (mutex_lock_killable(&oom_lock))
1592 * A few threads which were not waiting at mutex_lock_killable() can
1593 * fail to bail out. Therefore, check again after holding oom_lock.
1595 ret = should_force_charge() || out_of_memory(&oc);
1596 mutex_unlock(&oom_lock);
1600 #if MAX_NUMNODES > 1
1603 * test_mem_cgroup_node_reclaimable
1604 * @memcg: the target memcg
1605 * @nid: the node ID to be checked.
1606 * @noswap : specify true here if the user wants flle only information.
1608 * This function returns whether the specified memcg contains any
1609 * reclaimable pages on a node. Returns true if there are any reclaimable
1610 * pages in the node.
1612 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1613 int nid, bool noswap)
1615 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
1617 if (lruvec_page_state(lruvec, NR_INACTIVE_FILE) ||
1618 lruvec_page_state(lruvec, NR_ACTIVE_FILE))
1620 if (noswap || !total_swap_pages)
1622 if (lruvec_page_state(lruvec, NR_INACTIVE_ANON) ||
1623 lruvec_page_state(lruvec, NR_ACTIVE_ANON))
1630 * Always updating the nodemask is not very good - even if we have an empty
1631 * list or the wrong list here, we can start from some node and traverse all
1632 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1635 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1639 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1640 * pagein/pageout changes since the last update.
1642 if (!atomic_read(&memcg->numainfo_events))
1644 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1647 /* make a nodemask where this memcg uses memory from */
1648 memcg->scan_nodes = node_states[N_MEMORY];
1650 for_each_node_mask(nid, node_states[N_MEMORY]) {
1652 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1653 node_clear(nid, memcg->scan_nodes);
1656 atomic_set(&memcg->numainfo_events, 0);
1657 atomic_set(&memcg->numainfo_updating, 0);
1661 * Selecting a node where we start reclaim from. Because what we need is just
1662 * reducing usage counter, start from anywhere is O,K. Considering
1663 * memory reclaim from current node, there are pros. and cons.
1665 * Freeing memory from current node means freeing memory from a node which
1666 * we'll use or we've used. So, it may make LRU bad. And if several threads
1667 * hit limits, it will see a contention on a node. But freeing from remote
1668 * node means more costs for memory reclaim because of memory latency.
1670 * Now, we use round-robin. Better algorithm is welcomed.
1672 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1676 mem_cgroup_may_update_nodemask(memcg);
1677 node = memcg->last_scanned_node;
1679 node = next_node_in(node, memcg->scan_nodes);
1681 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1682 * last time it really checked all the LRUs due to rate limiting.
1683 * Fallback to the current node in that case for simplicity.
1685 if (unlikely(node == MAX_NUMNODES))
1686 node = numa_node_id();
1688 memcg->last_scanned_node = node;
1692 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1698 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1701 unsigned long *total_scanned)
1703 struct mem_cgroup *victim = NULL;
1706 unsigned long excess;
1707 unsigned long nr_scanned;
1708 struct mem_cgroup_reclaim_cookie reclaim = {
1713 excess = soft_limit_excess(root_memcg);
1716 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1721 * If we have not been able to reclaim
1722 * anything, it might because there are
1723 * no reclaimable pages under this hierarchy
1728 * We want to do more targeted reclaim.
1729 * excess >> 2 is not to excessive so as to
1730 * reclaim too much, nor too less that we keep
1731 * coming back to reclaim from this cgroup
1733 if (total >= (excess >> 2) ||
1734 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1739 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1740 pgdat, &nr_scanned);
1741 *total_scanned += nr_scanned;
1742 if (!soft_limit_excess(root_memcg))
1745 mem_cgroup_iter_break(root_memcg, victim);
1749 #ifdef CONFIG_LOCKDEP
1750 static struct lockdep_map memcg_oom_lock_dep_map = {
1751 .name = "memcg_oom_lock",
1755 static DEFINE_SPINLOCK(memcg_oom_lock);
1758 * Check OOM-Killer is already running under our hierarchy.
1759 * If someone is running, return false.
1761 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1763 struct mem_cgroup *iter, *failed = NULL;
1765 spin_lock(&memcg_oom_lock);
1767 for_each_mem_cgroup_tree(iter, memcg) {
1768 if (iter->oom_lock) {
1770 * this subtree of our hierarchy is already locked
1771 * so we cannot give a lock.
1774 mem_cgroup_iter_break(memcg, iter);
1777 iter->oom_lock = true;
1782 * OK, we failed to lock the whole subtree so we have
1783 * to clean up what we set up to the failing subtree
1785 for_each_mem_cgroup_tree(iter, memcg) {
1786 if (iter == failed) {
1787 mem_cgroup_iter_break(memcg, iter);
1790 iter->oom_lock = false;
1793 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1795 spin_unlock(&memcg_oom_lock);
1800 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1802 struct mem_cgroup *iter;
1804 spin_lock(&memcg_oom_lock);
1805 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1806 for_each_mem_cgroup_tree(iter, memcg)
1807 iter->oom_lock = false;
1808 spin_unlock(&memcg_oom_lock);
1811 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1813 struct mem_cgroup *iter;
1815 spin_lock(&memcg_oom_lock);
1816 for_each_mem_cgroup_tree(iter, memcg)
1818 spin_unlock(&memcg_oom_lock);
1821 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1823 struct mem_cgroup *iter;
1826 * When a new child is created while the hierarchy is under oom,
1827 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1829 spin_lock(&memcg_oom_lock);
1830 for_each_mem_cgroup_tree(iter, memcg)
1831 if (iter->under_oom > 0)
1833 spin_unlock(&memcg_oom_lock);
1836 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1838 struct oom_wait_info {
1839 struct mem_cgroup *memcg;
1840 wait_queue_entry_t wait;
1843 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1844 unsigned mode, int sync, void *arg)
1846 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1847 struct mem_cgroup *oom_wait_memcg;
1848 struct oom_wait_info *oom_wait_info;
1850 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1851 oom_wait_memcg = oom_wait_info->memcg;
1853 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1854 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1856 return autoremove_wake_function(wait, mode, sync, arg);
1859 static void memcg_oom_recover(struct mem_cgroup *memcg)
1862 * For the following lockless ->under_oom test, the only required
1863 * guarantee is that it must see the state asserted by an OOM when
1864 * this function is called as a result of userland actions
1865 * triggered by the notification of the OOM. This is trivially
1866 * achieved by invoking mem_cgroup_mark_under_oom() before
1867 * triggering notification.
1869 if (memcg && memcg->under_oom)
1870 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1880 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1882 enum oom_status ret;
1885 if (order > PAGE_ALLOC_COSTLY_ORDER)
1888 memcg_memory_event(memcg, MEMCG_OOM);
1891 * We are in the middle of the charge context here, so we
1892 * don't want to block when potentially sitting on a callstack
1893 * that holds all kinds of filesystem and mm locks.
1895 * cgroup1 allows disabling the OOM killer and waiting for outside
1896 * handling until the charge can succeed; remember the context and put
1897 * the task to sleep at the end of the page fault when all locks are
1900 * On the other hand, in-kernel OOM killer allows for an async victim
1901 * memory reclaim (oom_reaper) and that means that we are not solely
1902 * relying on the oom victim to make a forward progress and we can
1903 * invoke the oom killer here.
1905 * Please note that mem_cgroup_out_of_memory might fail to find a
1906 * victim and then we have to bail out from the charge path.
1908 if (memcg->oom_kill_disable) {
1909 if (!current->in_user_fault)
1911 css_get(&memcg->css);
1912 current->memcg_in_oom = memcg;
1913 current->memcg_oom_gfp_mask = mask;
1914 current->memcg_oom_order = order;
1919 mem_cgroup_mark_under_oom(memcg);
1921 locked = mem_cgroup_oom_trylock(memcg);
1924 mem_cgroup_oom_notify(memcg);
1926 mem_cgroup_unmark_under_oom(memcg);
1927 if (mem_cgroup_out_of_memory(memcg, mask, order))
1933 mem_cgroup_oom_unlock(memcg);
1939 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1940 * @handle: actually kill/wait or just clean up the OOM state
1942 * This has to be called at the end of a page fault if the memcg OOM
1943 * handler was enabled.
1945 * Memcg supports userspace OOM handling where failed allocations must
1946 * sleep on a waitqueue until the userspace task resolves the
1947 * situation. Sleeping directly in the charge context with all kinds
1948 * of locks held is not a good idea, instead we remember an OOM state
1949 * in the task and mem_cgroup_oom_synchronize() has to be called at
1950 * the end of the page fault to complete the OOM handling.
1952 * Returns %true if an ongoing memcg OOM situation was detected and
1953 * completed, %false otherwise.
1955 bool mem_cgroup_oom_synchronize(bool handle)
1957 struct mem_cgroup *memcg = current->memcg_in_oom;
1958 struct oom_wait_info owait;
1961 /* OOM is global, do not handle */
1968 owait.memcg = memcg;
1969 owait.wait.flags = 0;
1970 owait.wait.func = memcg_oom_wake_function;
1971 owait.wait.private = current;
1972 INIT_LIST_HEAD(&owait.wait.entry);
1974 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1975 mem_cgroup_mark_under_oom(memcg);
1977 locked = mem_cgroup_oom_trylock(memcg);
1980 mem_cgroup_oom_notify(memcg);
1982 if (locked && !memcg->oom_kill_disable) {
1983 mem_cgroup_unmark_under_oom(memcg);
1984 finish_wait(&memcg_oom_waitq, &owait.wait);
1985 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1986 current->memcg_oom_order);
1989 mem_cgroup_unmark_under_oom(memcg);
1990 finish_wait(&memcg_oom_waitq, &owait.wait);
1994 mem_cgroup_oom_unlock(memcg);
1996 * There is no guarantee that an OOM-lock contender
1997 * sees the wakeups triggered by the OOM kill
1998 * uncharges. Wake any sleepers explicitely.
2000 memcg_oom_recover(memcg);
2003 current->memcg_in_oom = NULL;
2004 css_put(&memcg->css);
2009 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2010 * @victim: task to be killed by the OOM killer
2011 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2013 * Returns a pointer to a memory cgroup, which has to be cleaned up
2014 * by killing all belonging OOM-killable tasks.
2016 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2018 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2019 struct mem_cgroup *oom_domain)
2021 struct mem_cgroup *oom_group = NULL;
2022 struct mem_cgroup *memcg;
2024 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2028 oom_domain = root_mem_cgroup;
2032 memcg = mem_cgroup_from_task(victim);
2033 if (memcg == root_mem_cgroup)
2037 * Traverse the memory cgroup hierarchy from the victim task's
2038 * cgroup up to the OOMing cgroup (or root) to find the
2039 * highest-level memory cgroup with oom.group set.
2041 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2042 if (memcg->oom_group)
2045 if (memcg == oom_domain)
2050 css_get(&oom_group->css);
2057 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2059 pr_info("Tasks in ");
2060 pr_cont_cgroup_path(memcg->css.cgroup);
2061 pr_cont(" are going to be killed due to memory.oom.group set\n");
2065 * lock_page_memcg - lock a page->mem_cgroup binding
2068 * This function protects unlocked LRU pages from being moved to
2071 * It ensures lifetime of the returned memcg. Caller is responsible
2072 * for the lifetime of the page; __unlock_page_memcg() is available
2073 * when @page might get freed inside the locked section.
2075 struct mem_cgroup *lock_page_memcg(struct page *page)
2077 struct mem_cgroup *memcg;
2078 unsigned long flags;
2081 * The RCU lock is held throughout the transaction. The fast
2082 * path can get away without acquiring the memcg->move_lock
2083 * because page moving starts with an RCU grace period.
2085 * The RCU lock also protects the memcg from being freed when
2086 * the page state that is going to change is the only thing
2087 * preventing the page itself from being freed. E.g. writeback
2088 * doesn't hold a page reference and relies on PG_writeback to
2089 * keep off truncation, migration and so forth.
2093 if (mem_cgroup_disabled())
2096 memcg = page->mem_cgroup;
2097 if (unlikely(!memcg))
2100 if (atomic_read(&memcg->moving_account) <= 0)
2103 spin_lock_irqsave(&memcg->move_lock, flags);
2104 if (memcg != page->mem_cgroup) {
2105 spin_unlock_irqrestore(&memcg->move_lock, flags);
2110 * When charge migration first begins, we can have locked and
2111 * unlocked page stat updates happening concurrently. Track
2112 * the task who has the lock for unlock_page_memcg().
2114 memcg->move_lock_task = current;
2115 memcg->move_lock_flags = flags;
2119 EXPORT_SYMBOL(lock_page_memcg);
2122 * __unlock_page_memcg - unlock and unpin a memcg
2125 * Unlock and unpin a memcg returned by lock_page_memcg().
2127 void __unlock_page_memcg(struct mem_cgroup *memcg)
2129 if (memcg && memcg->move_lock_task == current) {
2130 unsigned long flags = memcg->move_lock_flags;
2132 memcg->move_lock_task = NULL;
2133 memcg->move_lock_flags = 0;
2135 spin_unlock_irqrestore(&memcg->move_lock, flags);
2142 * unlock_page_memcg - unlock a page->mem_cgroup binding
2145 void unlock_page_memcg(struct page *page)
2147 __unlock_page_memcg(page->mem_cgroup);
2149 EXPORT_SYMBOL(unlock_page_memcg);
2151 struct memcg_stock_pcp {
2152 struct mem_cgroup *cached; /* this never be root cgroup */
2153 unsigned int nr_pages;
2154 struct work_struct work;
2155 unsigned long flags;
2156 #define FLUSHING_CACHED_CHARGE 0
2158 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2159 static DEFINE_MUTEX(percpu_charge_mutex);
2162 * consume_stock: Try to consume stocked charge on this cpu.
2163 * @memcg: memcg to consume from.
2164 * @nr_pages: how many pages to charge.
2166 * The charges will only happen if @memcg matches the current cpu's memcg
2167 * stock, and at least @nr_pages are available in that stock. Failure to
2168 * service an allocation will refill the stock.
2170 * returns true if successful, false otherwise.
2172 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2174 struct memcg_stock_pcp *stock;
2175 unsigned long flags;
2178 if (nr_pages > MEMCG_CHARGE_BATCH)
2181 local_irq_save(flags);
2183 stock = this_cpu_ptr(&memcg_stock);
2184 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2185 stock->nr_pages -= nr_pages;
2189 local_irq_restore(flags);
2195 * Returns stocks cached in percpu and reset cached information.
2197 static void drain_stock(struct memcg_stock_pcp *stock)
2199 struct mem_cgroup *old = stock->cached;
2201 if (stock->nr_pages) {
2202 page_counter_uncharge(&old->memory, stock->nr_pages);
2203 if (do_memsw_account())
2204 page_counter_uncharge(&old->memsw, stock->nr_pages);
2205 css_put_many(&old->css, stock->nr_pages);
2206 stock->nr_pages = 0;
2208 stock->cached = NULL;
2211 static void drain_local_stock(struct work_struct *dummy)
2213 struct memcg_stock_pcp *stock;
2214 unsigned long flags;
2217 * The only protection from memory hotplug vs. drain_stock races is
2218 * that we always operate on local CPU stock here with IRQ disabled
2220 local_irq_save(flags);
2222 stock = this_cpu_ptr(&memcg_stock);
2224 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2226 local_irq_restore(flags);
2230 * Cache charges(val) to local per_cpu area.
2231 * This will be consumed by consume_stock() function, later.
2233 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2235 struct memcg_stock_pcp *stock;
2236 unsigned long flags;
2238 local_irq_save(flags);
2240 stock = this_cpu_ptr(&memcg_stock);
2241 if (stock->cached != memcg) { /* reset if necessary */
2243 stock->cached = memcg;
2245 stock->nr_pages += nr_pages;
2247 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2250 local_irq_restore(flags);
2254 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2255 * of the hierarchy under it.
2257 static void drain_all_stock(struct mem_cgroup *root_memcg)
2261 /* If someone's already draining, avoid adding running more workers. */
2262 if (!mutex_trylock(&percpu_charge_mutex))
2265 * Notify other cpus that system-wide "drain" is running
2266 * We do not care about races with the cpu hotplug because cpu down
2267 * as well as workers from this path always operate on the local
2268 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2271 for_each_online_cpu(cpu) {
2272 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2273 struct mem_cgroup *memcg;
2277 memcg = stock->cached;
2278 if (memcg && stock->nr_pages &&
2279 mem_cgroup_is_descendant(memcg, root_memcg))
2284 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2286 drain_local_stock(&stock->work);
2288 schedule_work_on(cpu, &stock->work);
2292 mutex_unlock(&percpu_charge_mutex);
2295 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2297 struct memcg_stock_pcp *stock;
2298 struct mem_cgroup *memcg, *mi;
2300 stock = &per_cpu(memcg_stock, cpu);
2303 for_each_mem_cgroup(memcg) {
2306 for (i = 0; i < MEMCG_NR_STAT; i++) {
2310 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2312 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2313 atomic_long_add(x, &memcg->vmstats[i]);
2315 if (i >= NR_VM_NODE_STAT_ITEMS)
2318 for_each_node(nid) {
2319 struct mem_cgroup_per_node *pn;
2321 pn = mem_cgroup_nodeinfo(memcg, nid);
2322 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2325 atomic_long_add(x, &pn->lruvec_stat[i]);
2326 } while ((pn = parent_nodeinfo(pn, nid)));
2330 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2333 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2335 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2336 atomic_long_add(x, &memcg->vmevents[i]);
2343 static void reclaim_high(struct mem_cgroup *memcg,
2344 unsigned int nr_pages,
2348 if (page_counter_read(&memcg->memory) <= memcg->high)
2350 memcg_memory_event(memcg, MEMCG_HIGH);
2351 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2352 } while ((memcg = parent_mem_cgroup(memcg)));
2355 static void high_work_func(struct work_struct *work)
2357 struct mem_cgroup *memcg;
2359 memcg = container_of(work, struct mem_cgroup, high_work);
2360 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2364 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2365 * enough to still cause a significant slowdown in most cases, while still
2366 * allowing diagnostics and tracing to proceed without becoming stuck.
2368 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2371 * When calculating the delay, we use these either side of the exponentiation to
2372 * maintain precision and scale to a reasonable number of jiffies (see the table
2375 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2376 * overage ratio to a delay.
2377 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down down the
2378 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2379 * to produce a reasonable delay curve.
2381 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2382 * reasonable delay curve compared to precision-adjusted overage, not
2383 * penalising heavily at first, but still making sure that growth beyond the
2384 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2385 * example, with a high of 100 megabytes:
2387 * +-------+------------------------+
2388 * | usage | time to allocate in ms |
2389 * +-------+------------------------+
2411 * +-------+------------------------+
2413 #define MEMCG_DELAY_PRECISION_SHIFT 20
2414 #define MEMCG_DELAY_SCALING_SHIFT 14
2417 * Scheduled by try_charge() to be executed from the userland return path
2418 * and reclaims memory over the high limit.
2420 void mem_cgroup_handle_over_high(void)
2422 unsigned long usage, high, clamped_high;
2423 unsigned long pflags;
2424 unsigned long penalty_jiffies, overage;
2425 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2426 struct mem_cgroup *memcg;
2428 if (likely(!nr_pages))
2431 memcg = get_mem_cgroup_from_mm(current->mm);
2432 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2433 current->memcg_nr_pages_over_high = 0;
2436 * memory.high is breached and reclaim is unable to keep up. Throttle
2437 * allocators proactively to slow down excessive growth.
2439 * We use overage compared to memory.high to calculate the number of
2440 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2441 * fairly lenient on small overages, and increasingly harsh when the
2442 * memcg in question makes it clear that it has no intention of stopping
2443 * its crazy behaviour, so we exponentially increase the delay based on
2447 usage = page_counter_read(&memcg->memory);
2448 high = READ_ONCE(memcg->high);
2454 * Prevent division by 0 in overage calculation by acting as if it was a
2455 * threshold of 1 page
2457 clamped_high = max(high, 1UL);
2459 overage = div_u64((u64)(usage - high) << MEMCG_DELAY_PRECISION_SHIFT,
2462 penalty_jiffies = ((u64)overage * overage * HZ)
2463 >> (MEMCG_DELAY_PRECISION_SHIFT + MEMCG_DELAY_SCALING_SHIFT);
2466 * Factor in the task's own contribution to the overage, such that four
2467 * N-sized allocations are throttled approximately the same as one
2468 * 4N-sized allocation.
2470 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2471 * larger the current charge patch is than that.
2473 penalty_jiffies = penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2476 * Clamp the max delay per usermode return so as to still keep the
2477 * application moving forwards and also permit diagnostics, albeit
2480 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2483 * Don't sleep if the amount of jiffies this memcg owes us is so low
2484 * that it's not even worth doing, in an attempt to be nice to those who
2485 * go only a small amount over their memory.high value and maybe haven't
2486 * been aggressively reclaimed enough yet.
2488 if (penalty_jiffies <= HZ / 100)
2492 * If we exit early, we're guaranteed to die (since
2493 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2494 * need to account for any ill-begotten jiffies to pay them off later.
2496 psi_memstall_enter(&pflags);
2497 schedule_timeout_killable(penalty_jiffies);
2498 psi_memstall_leave(&pflags);
2501 css_put(&memcg->css);
2504 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2505 unsigned int nr_pages)
2507 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2508 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2509 struct mem_cgroup *mem_over_limit;
2510 struct page_counter *counter;
2511 unsigned long nr_reclaimed;
2512 bool may_swap = true;
2513 bool drained = false;
2514 enum oom_status oom_status;
2516 if (mem_cgroup_is_root(memcg))
2519 if (consume_stock(memcg, nr_pages))
2522 if (!do_memsw_account() ||
2523 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2524 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2526 if (do_memsw_account())
2527 page_counter_uncharge(&memcg->memsw, batch);
2528 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2530 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2534 if (batch > nr_pages) {
2540 * Unlike in global OOM situations, memcg is not in a physical
2541 * memory shortage. Allow dying and OOM-killed tasks to
2542 * bypass the last charges so that they can exit quickly and
2543 * free their memory.
2545 if (unlikely(should_force_charge()))
2549 * Prevent unbounded recursion when reclaim operations need to
2550 * allocate memory. This might exceed the limits temporarily,
2551 * but we prefer facilitating memory reclaim and getting back
2552 * under the limit over triggering OOM kills in these cases.
2554 if (unlikely(current->flags & PF_MEMALLOC))
2557 if (unlikely(task_in_memcg_oom(current)))
2560 if (!gfpflags_allow_blocking(gfp_mask))
2563 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2565 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2566 gfp_mask, may_swap);
2568 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2572 drain_all_stock(mem_over_limit);
2577 if (gfp_mask & __GFP_NORETRY)
2580 * Even though the limit is exceeded at this point, reclaim
2581 * may have been able to free some pages. Retry the charge
2582 * before killing the task.
2584 * Only for regular pages, though: huge pages are rather
2585 * unlikely to succeed so close to the limit, and we fall back
2586 * to regular pages anyway in case of failure.
2588 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2591 * At task move, charge accounts can be doubly counted. So, it's
2592 * better to wait until the end of task_move if something is going on.
2594 if (mem_cgroup_wait_acct_move(mem_over_limit))
2600 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2603 if (gfp_mask & __GFP_NOFAIL)
2606 if (fatal_signal_pending(current))
2610 * keep retrying as long as the memcg oom killer is able to make
2611 * a forward progress or bypass the charge if the oom killer
2612 * couldn't make any progress.
2614 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2615 get_order(nr_pages * PAGE_SIZE));
2616 switch (oom_status) {
2618 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2626 if (!(gfp_mask & __GFP_NOFAIL))
2630 * The allocation either can't fail or will lead to more memory
2631 * being freed very soon. Allow memory usage go over the limit
2632 * temporarily by force charging it.
2634 page_counter_charge(&memcg->memory, nr_pages);
2635 if (do_memsw_account())
2636 page_counter_charge(&memcg->memsw, nr_pages);
2637 css_get_many(&memcg->css, nr_pages);
2642 css_get_many(&memcg->css, batch);
2643 if (batch > nr_pages)
2644 refill_stock(memcg, batch - nr_pages);
2647 * If the hierarchy is above the normal consumption range, schedule
2648 * reclaim on returning to userland. We can perform reclaim here
2649 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2650 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2651 * not recorded as it most likely matches current's and won't
2652 * change in the meantime. As high limit is checked again before
2653 * reclaim, the cost of mismatch is negligible.
2656 if (page_counter_read(&memcg->memory) > memcg->high) {
2657 /* Don't bother a random interrupted task */
2658 if (in_interrupt()) {
2659 schedule_work(&memcg->high_work);
2662 current->memcg_nr_pages_over_high += batch;
2663 set_notify_resume(current);
2666 } while ((memcg = parent_mem_cgroup(memcg)));
2671 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2673 if (mem_cgroup_is_root(memcg))
2676 page_counter_uncharge(&memcg->memory, nr_pages);
2677 if (do_memsw_account())
2678 page_counter_uncharge(&memcg->memsw, nr_pages);
2680 css_put_many(&memcg->css, nr_pages);
2683 static void lock_page_lru(struct page *page, int *isolated)
2685 pg_data_t *pgdat = page_pgdat(page);
2687 spin_lock_irq(&pgdat->lru_lock);
2688 if (PageLRU(page)) {
2689 struct lruvec *lruvec;
2691 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2693 del_page_from_lru_list(page, lruvec, page_lru(page));
2699 static void unlock_page_lru(struct page *page, int isolated)
2701 pg_data_t *pgdat = page_pgdat(page);
2704 struct lruvec *lruvec;
2706 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2707 VM_BUG_ON_PAGE(PageLRU(page), page);
2709 add_page_to_lru_list(page, lruvec, page_lru(page));
2711 spin_unlock_irq(&pgdat->lru_lock);
2714 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2719 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2722 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2723 * may already be on some other mem_cgroup's LRU. Take care of it.
2726 lock_page_lru(page, &isolated);
2729 * Nobody should be changing or seriously looking at
2730 * page->mem_cgroup at this point:
2732 * - the page is uncharged
2734 * - the page is off-LRU
2736 * - an anonymous fault has exclusive page access, except for
2737 * a locked page table
2739 * - a page cache insertion, a swapin fault, or a migration
2740 * have the page locked
2742 page->mem_cgroup = memcg;
2745 unlock_page_lru(page, isolated);
2748 #ifdef CONFIG_MEMCG_KMEM
2749 static int memcg_alloc_cache_id(void)
2754 id = ida_simple_get(&memcg_cache_ida,
2755 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2759 if (id < memcg_nr_cache_ids)
2763 * There's no space for the new id in memcg_caches arrays,
2764 * so we have to grow them.
2766 down_write(&memcg_cache_ids_sem);
2768 size = 2 * (id + 1);
2769 if (size < MEMCG_CACHES_MIN_SIZE)
2770 size = MEMCG_CACHES_MIN_SIZE;
2771 else if (size > MEMCG_CACHES_MAX_SIZE)
2772 size = MEMCG_CACHES_MAX_SIZE;
2774 err = memcg_update_all_caches(size);
2776 err = memcg_update_all_list_lrus(size);
2778 memcg_nr_cache_ids = size;
2780 up_write(&memcg_cache_ids_sem);
2783 ida_simple_remove(&memcg_cache_ida, id);
2789 static void memcg_free_cache_id(int id)
2791 ida_simple_remove(&memcg_cache_ida, id);
2794 struct memcg_kmem_cache_create_work {
2795 struct mem_cgroup *memcg;
2796 struct kmem_cache *cachep;
2797 struct work_struct work;
2800 static void memcg_kmem_cache_create_func(struct work_struct *w)
2802 struct memcg_kmem_cache_create_work *cw =
2803 container_of(w, struct memcg_kmem_cache_create_work, work);
2804 struct mem_cgroup *memcg = cw->memcg;
2805 struct kmem_cache *cachep = cw->cachep;
2807 memcg_create_kmem_cache(memcg, cachep);
2809 css_put(&memcg->css);
2814 * Enqueue the creation of a per-memcg kmem_cache.
2816 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2817 struct kmem_cache *cachep)
2819 struct memcg_kmem_cache_create_work *cw;
2821 if (!css_tryget_online(&memcg->css))
2824 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2829 cw->cachep = cachep;
2830 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2832 queue_work(memcg_kmem_cache_wq, &cw->work);
2835 static inline bool memcg_kmem_bypass(void)
2837 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2843 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2844 * @cachep: the original global kmem cache
2846 * Return the kmem_cache we're supposed to use for a slab allocation.
2847 * We try to use the current memcg's version of the cache.
2849 * If the cache does not exist yet, if we are the first user of it, we
2850 * create it asynchronously in a workqueue and let the current allocation
2851 * go through with the original cache.
2853 * This function takes a reference to the cache it returns to assure it
2854 * won't get destroyed while we are working with it. Once the caller is
2855 * done with it, memcg_kmem_put_cache() must be called to release the
2858 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2860 struct mem_cgroup *memcg;
2861 struct kmem_cache *memcg_cachep;
2862 struct memcg_cache_array *arr;
2865 VM_BUG_ON(!is_root_cache(cachep));
2867 if (memcg_kmem_bypass())
2872 if (unlikely(current->active_memcg))
2873 memcg = current->active_memcg;
2875 memcg = mem_cgroup_from_task(current);
2877 if (!memcg || memcg == root_mem_cgroup)
2880 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2884 arr = rcu_dereference(cachep->memcg_params.memcg_caches);
2887 * Make sure we will access the up-to-date value. The code updating
2888 * memcg_caches issues a write barrier to match the data dependency
2889 * barrier inside READ_ONCE() (see memcg_create_kmem_cache()).
2891 memcg_cachep = READ_ONCE(arr->entries[kmemcg_id]);
2894 * If we are in a safe context (can wait, and not in interrupt
2895 * context), we could be be predictable and return right away.
2896 * This would guarantee that the allocation being performed
2897 * already belongs in the new cache.
2899 * However, there are some clashes that can arrive from locking.
2900 * For instance, because we acquire the slab_mutex while doing
2901 * memcg_create_kmem_cache, this means no further allocation
2902 * could happen with the slab_mutex held. So it's better to
2905 * If the memcg is dying or memcg_cache is about to be released,
2906 * don't bother creating new kmem_caches. Because memcg_cachep
2907 * is ZEROed as the fist step of kmem offlining, we don't need
2908 * percpu_ref_tryget_live() here. css_tryget_online() check in
2909 * memcg_schedule_kmem_cache_create() will prevent us from
2910 * creation of a new kmem_cache.
2912 if (unlikely(!memcg_cachep))
2913 memcg_schedule_kmem_cache_create(memcg, cachep);
2914 else if (percpu_ref_tryget(&memcg_cachep->memcg_params.refcnt))
2915 cachep = memcg_cachep;
2922 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2923 * @cachep: the cache returned by memcg_kmem_get_cache
2925 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2927 if (!is_root_cache(cachep))
2928 percpu_ref_put(&cachep->memcg_params.refcnt);
2932 * __memcg_kmem_charge_memcg: charge a kmem page
2933 * @page: page to charge
2934 * @gfp: reclaim mode
2935 * @order: allocation order
2936 * @memcg: memory cgroup to charge
2938 * Returns 0 on success, an error code on failure.
2940 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2941 struct mem_cgroup *memcg)
2943 unsigned int nr_pages = 1 << order;
2944 struct page_counter *counter;
2947 ret = try_charge(memcg, gfp, nr_pages);
2951 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2952 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2953 cancel_charge(memcg, nr_pages);
2960 * __memcg_kmem_charge: charge a kmem page to the current memory cgroup
2961 * @page: page to charge
2962 * @gfp: reclaim mode
2963 * @order: allocation order
2965 * Returns 0 on success, an error code on failure.
2967 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2969 struct mem_cgroup *memcg;
2972 if (memcg_kmem_bypass())
2975 memcg = get_mem_cgroup_from_current();
2976 if (!mem_cgroup_is_root(memcg)) {
2977 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2979 page->mem_cgroup = memcg;
2980 __SetPageKmemcg(page);
2983 css_put(&memcg->css);
2988 * __memcg_kmem_uncharge_memcg: uncharge a kmem page
2989 * @memcg: memcg to uncharge
2990 * @nr_pages: number of pages to uncharge
2992 void __memcg_kmem_uncharge_memcg(struct mem_cgroup *memcg,
2993 unsigned int nr_pages)
2995 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2996 page_counter_uncharge(&memcg->kmem, nr_pages);
2998 page_counter_uncharge(&memcg->memory, nr_pages);
2999 if (do_memsw_account())
3000 page_counter_uncharge(&memcg->memsw, nr_pages);
3003 * __memcg_kmem_uncharge: uncharge a kmem page
3004 * @page: page to uncharge
3005 * @order: allocation order
3007 void __memcg_kmem_uncharge(struct page *page, int order)
3009 struct mem_cgroup *memcg = page->mem_cgroup;
3010 unsigned int nr_pages = 1 << order;
3015 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3016 __memcg_kmem_uncharge_memcg(memcg, nr_pages);
3017 page->mem_cgroup = NULL;
3019 /* slab pages do not have PageKmemcg flag set */
3020 if (PageKmemcg(page))
3021 __ClearPageKmemcg(page);
3023 css_put_many(&memcg->css, nr_pages);
3025 #endif /* CONFIG_MEMCG_KMEM */
3027 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3030 * Because tail pages are not marked as "used", set it. We're under
3031 * pgdat->lru_lock and migration entries setup in all page mappings.
3033 void mem_cgroup_split_huge_fixup(struct page *head)
3037 if (mem_cgroup_disabled())
3040 for (i = 1; i < HPAGE_PMD_NR; i++)
3041 head[i].mem_cgroup = head->mem_cgroup;
3043 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
3045 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3047 #ifdef CONFIG_MEMCG_SWAP
3049 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3050 * @entry: swap entry to be moved
3051 * @from: mem_cgroup which the entry is moved from
3052 * @to: mem_cgroup which the entry is moved to
3054 * It succeeds only when the swap_cgroup's record for this entry is the same
3055 * as the mem_cgroup's id of @from.
3057 * Returns 0 on success, -EINVAL on failure.
3059 * The caller must have charged to @to, IOW, called page_counter_charge() about
3060 * both res and memsw, and called css_get().
3062 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3063 struct mem_cgroup *from, struct mem_cgroup *to)
3065 unsigned short old_id, new_id;
3067 old_id = mem_cgroup_id(from);
3068 new_id = mem_cgroup_id(to);
3070 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3071 mod_memcg_state(from, MEMCG_SWAP, -1);
3072 mod_memcg_state(to, MEMCG_SWAP, 1);
3078 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3079 struct mem_cgroup *from, struct mem_cgroup *to)
3085 static DEFINE_MUTEX(memcg_max_mutex);
3087 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3088 unsigned long max, bool memsw)
3090 bool enlarge = false;
3091 bool drained = false;
3093 bool limits_invariant;
3094 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3097 if (signal_pending(current)) {
3102 mutex_lock(&memcg_max_mutex);
3104 * Make sure that the new limit (memsw or memory limit) doesn't
3105 * break our basic invariant rule memory.max <= memsw.max.
3107 limits_invariant = memsw ? max >= memcg->memory.max :
3108 max <= memcg->memsw.max;
3109 if (!limits_invariant) {
3110 mutex_unlock(&memcg_max_mutex);
3114 if (max > counter->max)
3116 ret = page_counter_set_max(counter, max);
3117 mutex_unlock(&memcg_max_mutex);
3123 drain_all_stock(memcg);
3128 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3129 GFP_KERNEL, !memsw)) {
3135 if (!ret && enlarge)
3136 memcg_oom_recover(memcg);
3141 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3143 unsigned long *total_scanned)
3145 unsigned long nr_reclaimed = 0;
3146 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3147 unsigned long reclaimed;
3149 struct mem_cgroup_tree_per_node *mctz;
3150 unsigned long excess;
3151 unsigned long nr_scanned;
3156 mctz = soft_limit_tree_node(pgdat->node_id);
3159 * Do not even bother to check the largest node if the root
3160 * is empty. Do it lockless to prevent lock bouncing. Races
3161 * are acceptable as soft limit is best effort anyway.
3163 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3167 * This loop can run a while, specially if mem_cgroup's continuously
3168 * keep exceeding their soft limit and putting the system under
3175 mz = mem_cgroup_largest_soft_limit_node(mctz);
3180 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3181 gfp_mask, &nr_scanned);
3182 nr_reclaimed += reclaimed;
3183 *total_scanned += nr_scanned;
3184 spin_lock_irq(&mctz->lock);
3185 __mem_cgroup_remove_exceeded(mz, mctz);
3188 * If we failed to reclaim anything from this memory cgroup
3189 * it is time to move on to the next cgroup
3193 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3195 excess = soft_limit_excess(mz->memcg);
3197 * One school of thought says that we should not add
3198 * back the node to the tree if reclaim returns 0.
3199 * But our reclaim could return 0, simply because due
3200 * to priority we are exposing a smaller subset of
3201 * memory to reclaim from. Consider this as a longer
3204 /* If excess == 0, no tree ops */
3205 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3206 spin_unlock_irq(&mctz->lock);
3207 css_put(&mz->memcg->css);
3210 * Could not reclaim anything and there are no more
3211 * mem cgroups to try or we seem to be looping without
3212 * reclaiming anything.
3214 if (!nr_reclaimed &&
3216 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3218 } while (!nr_reclaimed);
3220 css_put(&next_mz->memcg->css);
3221 return nr_reclaimed;
3225 * Test whether @memcg has children, dead or alive. Note that this
3226 * function doesn't care whether @memcg has use_hierarchy enabled and
3227 * returns %true if there are child csses according to the cgroup
3228 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3230 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3235 ret = css_next_child(NULL, &memcg->css);
3241 * Reclaims as many pages from the given memcg as possible.
3243 * Caller is responsible for holding css reference for memcg.
3245 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3247 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3249 /* we call try-to-free pages for make this cgroup empty */
3250 lru_add_drain_all();
3252 drain_all_stock(memcg);
3254 /* try to free all pages in this cgroup */
3255 while (nr_retries && page_counter_read(&memcg->memory)) {
3258 if (signal_pending(current))
3261 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3265 /* maybe some writeback is necessary */
3266 congestion_wait(BLK_RW_ASYNC, HZ/10);
3274 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3275 char *buf, size_t nbytes,
3278 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3280 if (mem_cgroup_is_root(memcg))
3282 return mem_cgroup_force_empty(memcg) ?: nbytes;
3285 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3288 return mem_cgroup_from_css(css)->use_hierarchy;
3291 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3292 struct cftype *cft, u64 val)
3295 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3296 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3298 if (memcg->use_hierarchy == val)
3302 * If parent's use_hierarchy is set, we can't make any modifications
3303 * in the child subtrees. If it is unset, then the change can
3304 * occur, provided the current cgroup has no children.
3306 * For the root cgroup, parent_mem is NULL, we allow value to be
3307 * set if there are no children.
3309 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3310 (val == 1 || val == 0)) {
3311 if (!memcg_has_children(memcg))
3312 memcg->use_hierarchy = val;
3321 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3325 if (mem_cgroup_is_root(memcg)) {
3326 val = memcg_page_state(memcg, MEMCG_CACHE) +
3327 memcg_page_state(memcg, MEMCG_RSS);
3329 val += memcg_page_state(memcg, MEMCG_SWAP);
3332 val = page_counter_read(&memcg->memory);
3334 val = page_counter_read(&memcg->memsw);
3347 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3350 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3351 struct page_counter *counter;
3353 switch (MEMFILE_TYPE(cft->private)) {
3355 counter = &memcg->memory;
3358 counter = &memcg->memsw;
3361 counter = &memcg->kmem;
3364 counter = &memcg->tcpmem;
3370 switch (MEMFILE_ATTR(cft->private)) {
3372 if (counter == &memcg->memory)
3373 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3374 if (counter == &memcg->memsw)
3375 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3376 return (u64)page_counter_read(counter) * PAGE_SIZE;
3378 return (u64)counter->max * PAGE_SIZE;
3380 return (u64)counter->watermark * PAGE_SIZE;
3382 return counter->failcnt;
3383 case RES_SOFT_LIMIT:
3384 return (u64)memcg->soft_limit * PAGE_SIZE;
3390 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg, bool slab_only)
3392 unsigned long stat[MEMCG_NR_STAT];
3393 struct mem_cgroup *mi;
3395 int min_idx, max_idx;
3398 min_idx = NR_SLAB_RECLAIMABLE;
3399 max_idx = NR_SLAB_UNRECLAIMABLE;
3402 max_idx = MEMCG_NR_STAT;
3405 for (i = min_idx; i < max_idx; i++)
3408 for_each_online_cpu(cpu)
3409 for (i = min_idx; i < max_idx; i++)
3410 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3412 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3413 for (i = min_idx; i < max_idx; i++)
3414 atomic_long_add(stat[i], &mi->vmstats[i]);
3417 max_idx = NR_VM_NODE_STAT_ITEMS;
3419 for_each_node(node) {
3420 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3421 struct mem_cgroup_per_node *pi;
3423 for (i = min_idx; i < max_idx; i++)
3426 for_each_online_cpu(cpu)
3427 for (i = min_idx; i < max_idx; i++)
3429 pn->lruvec_stat_cpu->count[i], cpu);
3431 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3432 for (i = min_idx; i < max_idx; i++)
3433 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3437 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3439 unsigned long events[NR_VM_EVENT_ITEMS];
3440 struct mem_cgroup *mi;
3443 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3446 for_each_online_cpu(cpu)
3447 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3448 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3451 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3452 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3453 atomic_long_add(events[i], &mi->vmevents[i]);
3456 #ifdef CONFIG_MEMCG_KMEM
3457 static int memcg_online_kmem(struct mem_cgroup *memcg)
3461 if (cgroup_memory_nokmem)
3464 BUG_ON(memcg->kmemcg_id >= 0);
3465 BUG_ON(memcg->kmem_state);
3467 memcg_id = memcg_alloc_cache_id();
3471 static_branch_inc(&memcg_kmem_enabled_key);
3473 * A memory cgroup is considered kmem-online as soon as it gets
3474 * kmemcg_id. Setting the id after enabling static branching will
3475 * guarantee no one starts accounting before all call sites are
3478 memcg->kmemcg_id = memcg_id;
3479 memcg->kmem_state = KMEM_ONLINE;
3480 INIT_LIST_HEAD(&memcg->kmem_caches);
3485 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3487 struct cgroup_subsys_state *css;
3488 struct mem_cgroup *parent, *child;
3491 if (memcg->kmem_state != KMEM_ONLINE)
3494 * Clear the online state before clearing memcg_caches array
3495 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3496 * guarantees that no cache will be created for this cgroup
3497 * after we are done (see memcg_create_kmem_cache()).
3499 memcg->kmem_state = KMEM_ALLOCATED;
3501 parent = parent_mem_cgroup(memcg);
3503 parent = root_mem_cgroup;
3506 * Deactivate and reparent kmem_caches. Then flush percpu
3507 * slab statistics to have precise values at the parent and
3508 * all ancestor levels. It's required to keep slab stats
3509 * accurate after the reparenting of kmem_caches.
3511 memcg_deactivate_kmem_caches(memcg, parent);
3512 memcg_flush_percpu_vmstats(memcg, true);
3514 kmemcg_id = memcg->kmemcg_id;
3515 BUG_ON(kmemcg_id < 0);
3518 * Change kmemcg_id of this cgroup and all its descendants to the
3519 * parent's id, and then move all entries from this cgroup's list_lrus
3520 * to ones of the parent. After we have finished, all list_lrus
3521 * corresponding to this cgroup are guaranteed to remain empty. The
3522 * ordering is imposed by list_lru_node->lock taken by
3523 * memcg_drain_all_list_lrus().
3525 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3526 css_for_each_descendant_pre(css, &memcg->css) {
3527 child = mem_cgroup_from_css(css);
3528 BUG_ON(child->kmemcg_id != kmemcg_id);
3529 child->kmemcg_id = parent->kmemcg_id;
3530 if (!memcg->use_hierarchy)
3535 memcg_drain_all_list_lrus(kmemcg_id, parent);
3537 memcg_free_cache_id(kmemcg_id);
3540 static void memcg_free_kmem(struct mem_cgroup *memcg)
3542 /* css_alloc() failed, offlining didn't happen */
3543 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3544 memcg_offline_kmem(memcg);
3546 if (memcg->kmem_state == KMEM_ALLOCATED) {
3547 WARN_ON(!list_empty(&memcg->kmem_caches));
3548 static_branch_dec(&memcg_kmem_enabled_key);
3552 static int memcg_online_kmem(struct mem_cgroup *memcg)
3556 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3559 static void memcg_free_kmem(struct mem_cgroup *memcg)
3562 #endif /* CONFIG_MEMCG_KMEM */
3564 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3569 mutex_lock(&memcg_max_mutex);
3570 ret = page_counter_set_max(&memcg->kmem, max);
3571 mutex_unlock(&memcg_max_mutex);
3575 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3579 mutex_lock(&memcg_max_mutex);
3581 ret = page_counter_set_max(&memcg->tcpmem, max);
3585 if (!memcg->tcpmem_active) {
3587 * The active flag needs to be written after the static_key
3588 * update. This is what guarantees that the socket activation
3589 * function is the last one to run. See mem_cgroup_sk_alloc()
3590 * for details, and note that we don't mark any socket as
3591 * belonging to this memcg until that flag is up.
3593 * We need to do this, because static_keys will span multiple
3594 * sites, but we can't control their order. If we mark a socket
3595 * as accounted, but the accounting functions are not patched in
3596 * yet, we'll lose accounting.
3598 * We never race with the readers in mem_cgroup_sk_alloc(),
3599 * because when this value change, the code to process it is not
3602 static_branch_inc(&memcg_sockets_enabled_key);
3603 memcg->tcpmem_active = true;
3606 mutex_unlock(&memcg_max_mutex);
3611 * The user of this function is...
3614 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3615 char *buf, size_t nbytes, loff_t off)
3617 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3618 unsigned long nr_pages;
3621 buf = strstrip(buf);
3622 ret = page_counter_memparse(buf, "-1", &nr_pages);
3626 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3628 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3632 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3634 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3637 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3640 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3641 "Please report your usecase to linux-mm@kvack.org if you "
3642 "depend on this functionality.\n");
3643 ret = memcg_update_kmem_max(memcg, nr_pages);
3646 ret = memcg_update_tcp_max(memcg, nr_pages);
3650 case RES_SOFT_LIMIT:
3651 memcg->soft_limit = nr_pages;
3655 return ret ?: nbytes;
3658 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3659 size_t nbytes, loff_t off)
3661 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3662 struct page_counter *counter;
3664 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3666 counter = &memcg->memory;
3669 counter = &memcg->memsw;
3672 counter = &memcg->kmem;
3675 counter = &memcg->tcpmem;
3681 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3683 page_counter_reset_watermark(counter);
3686 counter->failcnt = 0;
3695 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3698 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3702 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3703 struct cftype *cft, u64 val)
3705 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3707 if (val & ~MOVE_MASK)
3711 * No kind of locking is needed in here, because ->can_attach() will
3712 * check this value once in the beginning of the process, and then carry
3713 * on with stale data. This means that changes to this value will only
3714 * affect task migrations starting after the change.
3716 memcg->move_charge_at_immigrate = val;
3720 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3721 struct cftype *cft, u64 val)
3729 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3730 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3731 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3733 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3734 int nid, unsigned int lru_mask)
3736 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
3737 unsigned long nr = 0;
3740 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3743 if (!(BIT(lru) & lru_mask))
3745 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3750 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3751 unsigned int lru_mask)
3753 unsigned long nr = 0;
3757 if (!(BIT(lru) & lru_mask))
3759 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3764 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3768 unsigned int lru_mask;
3771 static const struct numa_stat stats[] = {
3772 { "total", LRU_ALL },
3773 { "file", LRU_ALL_FILE },
3774 { "anon", LRU_ALL_ANON },
3775 { "unevictable", BIT(LRU_UNEVICTABLE) },
3777 const struct numa_stat *stat;
3780 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3782 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3783 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3784 seq_printf(m, "%s=%lu", stat->name, nr);
3785 for_each_node_state(nid, N_MEMORY) {
3786 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3788 seq_printf(m, " N%d=%lu", nid, nr);
3793 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3794 struct mem_cgroup *iter;
3797 for_each_mem_cgroup_tree(iter, memcg)
3798 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3799 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3800 for_each_node_state(nid, N_MEMORY) {
3802 for_each_mem_cgroup_tree(iter, memcg)
3803 nr += mem_cgroup_node_nr_lru_pages(
3804 iter, nid, stat->lru_mask);
3805 seq_printf(m, " N%d=%lu", nid, nr);
3812 #endif /* CONFIG_NUMA */
3814 static const unsigned int memcg1_stats[] = {
3825 static const char *const memcg1_stat_names[] = {
3836 /* Universal VM events cgroup1 shows, original sort order */
3837 static const unsigned int memcg1_events[] = {
3844 static const char *const memcg1_event_names[] = {
3851 static int memcg_stat_show(struct seq_file *m, void *v)
3853 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3854 unsigned long memory, memsw;
3855 struct mem_cgroup *mi;
3858 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3859 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3861 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3862 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3864 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3865 memcg_page_state_local(memcg, memcg1_stats[i]) *
3869 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3870 seq_printf(m, "%s %lu\n", memcg1_event_names[i],
3871 memcg_events_local(memcg, memcg1_events[i]));
3873 for (i = 0; i < NR_LRU_LISTS; i++)
3874 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3875 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3878 /* Hierarchical information */
3879 memory = memsw = PAGE_COUNTER_MAX;
3880 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3881 memory = min(memory, mi->memory.max);
3882 memsw = min(memsw, mi->memsw.max);
3884 seq_printf(m, "hierarchical_memory_limit %llu\n",
3885 (u64)memory * PAGE_SIZE);
3886 if (do_memsw_account())
3887 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3888 (u64)memsw * PAGE_SIZE);
3890 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3891 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3893 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3894 (u64)memcg_page_state(memcg, memcg1_stats[i]) *
3898 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3899 seq_printf(m, "total_%s %llu\n", memcg1_event_names[i],
3900 (u64)memcg_events(memcg, memcg1_events[i]));
3902 for (i = 0; i < NR_LRU_LISTS; i++)
3903 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i],
3904 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
3907 #ifdef CONFIG_DEBUG_VM
3910 struct mem_cgroup_per_node *mz;
3911 struct zone_reclaim_stat *rstat;
3912 unsigned long recent_rotated[2] = {0, 0};
3913 unsigned long recent_scanned[2] = {0, 0};
3915 for_each_online_pgdat(pgdat) {
3916 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3917 rstat = &mz->lruvec.reclaim_stat;
3919 recent_rotated[0] += rstat->recent_rotated[0];
3920 recent_rotated[1] += rstat->recent_rotated[1];
3921 recent_scanned[0] += rstat->recent_scanned[0];
3922 recent_scanned[1] += rstat->recent_scanned[1];
3924 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3925 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3926 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3927 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3934 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3937 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3939 return mem_cgroup_swappiness(memcg);
3942 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3943 struct cftype *cft, u64 val)
3945 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3951 memcg->swappiness = val;
3953 vm_swappiness = val;
3958 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3960 struct mem_cgroup_threshold_ary *t;
3961 unsigned long usage;
3966 t = rcu_dereference(memcg->thresholds.primary);
3968 t = rcu_dereference(memcg->memsw_thresholds.primary);
3973 usage = mem_cgroup_usage(memcg, swap);
3976 * current_threshold points to threshold just below or equal to usage.
3977 * If it's not true, a threshold was crossed after last
3978 * call of __mem_cgroup_threshold().
3980 i = t->current_threshold;
3983 * Iterate backward over array of thresholds starting from
3984 * current_threshold and check if a threshold is crossed.
3985 * If none of thresholds below usage is crossed, we read
3986 * only one element of the array here.
3988 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3989 eventfd_signal(t->entries[i].eventfd, 1);
3991 /* i = current_threshold + 1 */
3995 * Iterate forward over array of thresholds starting from
3996 * current_threshold+1 and check if a threshold is crossed.
3997 * If none of thresholds above usage is crossed, we read
3998 * only one element of the array here.
4000 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4001 eventfd_signal(t->entries[i].eventfd, 1);
4003 /* Update current_threshold */
4004 t->current_threshold = i - 1;
4009 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4012 __mem_cgroup_threshold(memcg, false);
4013 if (do_memsw_account())
4014 __mem_cgroup_threshold(memcg, true);
4016 memcg = parent_mem_cgroup(memcg);
4020 static int compare_thresholds(const void *a, const void *b)
4022 const struct mem_cgroup_threshold *_a = a;
4023 const struct mem_cgroup_threshold *_b = b;
4025 if (_a->threshold > _b->threshold)
4028 if (_a->threshold < _b->threshold)
4034 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4036 struct mem_cgroup_eventfd_list *ev;
4038 spin_lock(&memcg_oom_lock);
4040 list_for_each_entry(ev, &memcg->oom_notify, list)
4041 eventfd_signal(ev->eventfd, 1);
4043 spin_unlock(&memcg_oom_lock);
4047 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4049 struct mem_cgroup *iter;
4051 for_each_mem_cgroup_tree(iter, memcg)
4052 mem_cgroup_oom_notify_cb(iter);
4055 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4056 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4058 struct mem_cgroup_thresholds *thresholds;
4059 struct mem_cgroup_threshold_ary *new;
4060 unsigned long threshold;
4061 unsigned long usage;
4064 ret = page_counter_memparse(args, "-1", &threshold);
4068 mutex_lock(&memcg->thresholds_lock);
4071 thresholds = &memcg->thresholds;
4072 usage = mem_cgroup_usage(memcg, false);
4073 } else if (type == _MEMSWAP) {
4074 thresholds = &memcg->memsw_thresholds;
4075 usage = mem_cgroup_usage(memcg, true);
4079 /* Check if a threshold crossed before adding a new one */
4080 if (thresholds->primary)
4081 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4083 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4085 /* Allocate memory for new array of thresholds */
4086 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4093 /* Copy thresholds (if any) to new array */
4094 if (thresholds->primary) {
4095 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4096 sizeof(struct mem_cgroup_threshold));
4099 /* Add new threshold */
4100 new->entries[size - 1].eventfd = eventfd;
4101 new->entries[size - 1].threshold = threshold;
4103 /* Sort thresholds. Registering of new threshold isn't time-critical */
4104 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4105 compare_thresholds, NULL);
4107 /* Find current threshold */
4108 new->current_threshold = -1;
4109 for (i = 0; i < size; i++) {
4110 if (new->entries[i].threshold <= usage) {
4112 * new->current_threshold will not be used until
4113 * rcu_assign_pointer(), so it's safe to increment
4116 ++new->current_threshold;
4121 /* Free old spare buffer and save old primary buffer as spare */
4122 kfree(thresholds->spare);
4123 thresholds->spare = thresholds->primary;
4125 rcu_assign_pointer(thresholds->primary, new);
4127 /* To be sure that nobody uses thresholds */
4131 mutex_unlock(&memcg->thresholds_lock);
4136 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4137 struct eventfd_ctx *eventfd, const char *args)
4139 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4142 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4143 struct eventfd_ctx *eventfd, const char *args)
4145 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4148 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4149 struct eventfd_ctx *eventfd, enum res_type type)
4151 struct mem_cgroup_thresholds *thresholds;
4152 struct mem_cgroup_threshold_ary *new;
4153 unsigned long usage;
4156 mutex_lock(&memcg->thresholds_lock);
4159 thresholds = &memcg->thresholds;
4160 usage = mem_cgroup_usage(memcg, false);
4161 } else if (type == _MEMSWAP) {
4162 thresholds = &memcg->memsw_thresholds;
4163 usage = mem_cgroup_usage(memcg, true);
4167 if (!thresholds->primary)
4170 /* Check if a threshold crossed before removing */
4171 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4173 /* Calculate new number of threshold */
4175 for (i = 0; i < thresholds->primary->size; i++) {
4176 if (thresholds->primary->entries[i].eventfd != eventfd)
4180 new = thresholds->spare;
4182 /* Set thresholds array to NULL if we don't have thresholds */
4191 /* Copy thresholds and find current threshold */
4192 new->current_threshold = -1;
4193 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4194 if (thresholds->primary->entries[i].eventfd == eventfd)
4197 new->entries[j] = thresholds->primary->entries[i];
4198 if (new->entries[j].threshold <= usage) {
4200 * new->current_threshold will not be used
4201 * until rcu_assign_pointer(), so it's safe to increment
4204 ++new->current_threshold;
4210 /* Swap primary and spare array */
4211 thresholds->spare = thresholds->primary;
4213 rcu_assign_pointer(thresholds->primary, new);
4215 /* To be sure that nobody uses thresholds */
4218 /* If all events are unregistered, free the spare array */
4220 kfree(thresholds->spare);
4221 thresholds->spare = NULL;
4224 mutex_unlock(&memcg->thresholds_lock);
4227 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4228 struct eventfd_ctx *eventfd)
4230 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4233 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4234 struct eventfd_ctx *eventfd)
4236 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4239 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4240 struct eventfd_ctx *eventfd, const char *args)
4242 struct mem_cgroup_eventfd_list *event;
4244 event = kmalloc(sizeof(*event), GFP_KERNEL);
4248 spin_lock(&memcg_oom_lock);
4250 event->eventfd = eventfd;
4251 list_add(&event->list, &memcg->oom_notify);
4253 /* already in OOM ? */
4254 if (memcg->under_oom)
4255 eventfd_signal(eventfd, 1);
4256 spin_unlock(&memcg_oom_lock);
4261 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4262 struct eventfd_ctx *eventfd)
4264 struct mem_cgroup_eventfd_list *ev, *tmp;
4266 spin_lock(&memcg_oom_lock);
4268 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4269 if (ev->eventfd == eventfd) {
4270 list_del(&ev->list);
4275 spin_unlock(&memcg_oom_lock);
4278 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4280 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4282 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4283 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4284 seq_printf(sf, "oom_kill %lu\n",
4285 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4289 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4290 struct cftype *cft, u64 val)
4292 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4294 /* cannot set to root cgroup and only 0 and 1 are allowed */
4295 if (!css->parent || !((val == 0) || (val == 1)))
4298 memcg->oom_kill_disable = val;
4300 memcg_oom_recover(memcg);
4305 #ifdef CONFIG_CGROUP_WRITEBACK
4307 #include <trace/events/writeback.h>
4309 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4311 return wb_domain_init(&memcg->cgwb_domain, gfp);
4314 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4316 wb_domain_exit(&memcg->cgwb_domain);
4319 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4321 wb_domain_size_changed(&memcg->cgwb_domain);
4324 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4326 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4328 if (!memcg->css.parent)
4331 return &memcg->cgwb_domain;
4335 * idx can be of type enum memcg_stat_item or node_stat_item.
4336 * Keep in sync with memcg_exact_page().
4338 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4340 long x = atomic_long_read(&memcg->vmstats[idx]);
4343 for_each_online_cpu(cpu)
4344 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4351 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4352 * @wb: bdi_writeback in question
4353 * @pfilepages: out parameter for number of file pages
4354 * @pheadroom: out parameter for number of allocatable pages according to memcg
4355 * @pdirty: out parameter for number of dirty pages
4356 * @pwriteback: out parameter for number of pages under writeback
4358 * Determine the numbers of file, headroom, dirty, and writeback pages in
4359 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4360 * is a bit more involved.
4362 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4363 * headroom is calculated as the lowest headroom of itself and the
4364 * ancestors. Note that this doesn't consider the actual amount of
4365 * available memory in the system. The caller should further cap
4366 * *@pheadroom accordingly.
4368 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4369 unsigned long *pheadroom, unsigned long *pdirty,
4370 unsigned long *pwriteback)
4372 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4373 struct mem_cgroup *parent;
4375 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4377 /* this should eventually include NR_UNSTABLE_NFS */
4378 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4379 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4380 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4381 *pheadroom = PAGE_COUNTER_MAX;
4383 while ((parent = parent_mem_cgroup(memcg))) {
4384 unsigned long ceiling = min(memcg->memory.max, memcg->high);
4385 unsigned long used = page_counter_read(&memcg->memory);
4387 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4393 * Foreign dirty flushing
4395 * There's an inherent mismatch between memcg and writeback. The former
4396 * trackes ownership per-page while the latter per-inode. This was a
4397 * deliberate design decision because honoring per-page ownership in the
4398 * writeback path is complicated, may lead to higher CPU and IO overheads
4399 * and deemed unnecessary given that write-sharing an inode across
4400 * different cgroups isn't a common use-case.
4402 * Combined with inode majority-writer ownership switching, this works well
4403 * enough in most cases but there are some pathological cases. For
4404 * example, let's say there are two cgroups A and B which keep writing to
4405 * different but confined parts of the same inode. B owns the inode and
4406 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4407 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4408 * triggering background writeback. A will be slowed down without a way to
4409 * make writeback of the dirty pages happen.
4411 * Conditions like the above can lead to a cgroup getting repatedly and
4412 * severely throttled after making some progress after each
4413 * dirty_expire_interval while the underyling IO device is almost
4416 * Solving this problem completely requires matching the ownership tracking
4417 * granularities between memcg and writeback in either direction. However,
4418 * the more egregious behaviors can be avoided by simply remembering the
4419 * most recent foreign dirtying events and initiating remote flushes on
4420 * them when local writeback isn't enough to keep the memory clean enough.
4422 * The following two functions implement such mechanism. When a foreign
4423 * page - a page whose memcg and writeback ownerships don't match - is
4424 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4425 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4426 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4427 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4428 * foreign bdi_writebacks which haven't expired. Both the numbers of
4429 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4430 * limited to MEMCG_CGWB_FRN_CNT.
4432 * The mechanism only remembers IDs and doesn't hold any object references.
4433 * As being wrong occasionally doesn't matter, updates and accesses to the
4434 * records are lockless and racy.
4436 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4437 struct bdi_writeback *wb)
4439 struct mem_cgroup *memcg = page->mem_cgroup;
4440 struct memcg_cgwb_frn *frn;
4441 u64 now = get_jiffies_64();
4442 u64 oldest_at = now;
4446 trace_track_foreign_dirty(page, wb);
4449 * Pick the slot to use. If there is already a slot for @wb, keep
4450 * using it. If not replace the oldest one which isn't being
4453 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4454 frn = &memcg->cgwb_frn[i];
4455 if (frn->bdi_id == wb->bdi->id &&
4456 frn->memcg_id == wb->memcg_css->id)
4458 if (time_before64(frn->at, oldest_at) &&
4459 atomic_read(&frn->done.cnt) == 1) {
4461 oldest_at = frn->at;
4465 if (i < MEMCG_CGWB_FRN_CNT) {
4467 * Re-using an existing one. Update timestamp lazily to
4468 * avoid making the cacheline hot. We want them to be
4469 * reasonably up-to-date and significantly shorter than
4470 * dirty_expire_interval as that's what expires the record.
4471 * Use the shorter of 1s and dirty_expire_interval / 8.
4473 unsigned long update_intv =
4474 min_t(unsigned long, HZ,
4475 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4477 if (time_before64(frn->at, now - update_intv))
4479 } else if (oldest >= 0) {
4480 /* replace the oldest free one */
4481 frn = &memcg->cgwb_frn[oldest];
4482 frn->bdi_id = wb->bdi->id;
4483 frn->memcg_id = wb->memcg_css->id;
4488 /* issue foreign writeback flushes for recorded foreign dirtying events */
4489 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4491 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4492 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4493 u64 now = jiffies_64;
4496 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4497 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4500 * If the record is older than dirty_expire_interval,
4501 * writeback on it has already started. No need to kick it
4502 * off again. Also, don't start a new one if there's
4503 * already one in flight.
4505 if (time_after64(frn->at, now - intv) &&
4506 atomic_read(&frn->done.cnt) == 1) {
4508 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4509 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4510 WB_REASON_FOREIGN_FLUSH,
4516 #else /* CONFIG_CGROUP_WRITEBACK */
4518 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4523 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4527 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4531 #endif /* CONFIG_CGROUP_WRITEBACK */
4534 * DO NOT USE IN NEW FILES.
4536 * "cgroup.event_control" implementation.
4538 * This is way over-engineered. It tries to support fully configurable
4539 * events for each user. Such level of flexibility is completely
4540 * unnecessary especially in the light of the planned unified hierarchy.
4542 * Please deprecate this and replace with something simpler if at all
4547 * Unregister event and free resources.
4549 * Gets called from workqueue.
4551 static void memcg_event_remove(struct work_struct *work)
4553 struct mem_cgroup_event *event =
4554 container_of(work, struct mem_cgroup_event, remove);
4555 struct mem_cgroup *memcg = event->memcg;
4557 remove_wait_queue(event->wqh, &event->wait);
4559 event->unregister_event(memcg, event->eventfd);
4561 /* Notify userspace the event is going away. */
4562 eventfd_signal(event->eventfd, 1);
4564 eventfd_ctx_put(event->eventfd);
4566 css_put(&memcg->css);
4570 * Gets called on EPOLLHUP on eventfd when user closes it.
4572 * Called with wqh->lock held and interrupts disabled.
4574 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4575 int sync, void *key)
4577 struct mem_cgroup_event *event =
4578 container_of(wait, struct mem_cgroup_event, wait);
4579 struct mem_cgroup *memcg = event->memcg;
4580 __poll_t flags = key_to_poll(key);
4582 if (flags & EPOLLHUP) {
4584 * If the event has been detached at cgroup removal, we
4585 * can simply return knowing the other side will cleanup
4588 * We can't race against event freeing since the other
4589 * side will require wqh->lock via remove_wait_queue(),
4592 spin_lock(&memcg->event_list_lock);
4593 if (!list_empty(&event->list)) {
4594 list_del_init(&event->list);
4596 * We are in atomic context, but cgroup_event_remove()
4597 * may sleep, so we have to call it in workqueue.
4599 schedule_work(&event->remove);
4601 spin_unlock(&memcg->event_list_lock);
4607 static void memcg_event_ptable_queue_proc(struct file *file,
4608 wait_queue_head_t *wqh, poll_table *pt)
4610 struct mem_cgroup_event *event =
4611 container_of(pt, struct mem_cgroup_event, pt);
4614 add_wait_queue(wqh, &event->wait);
4618 * DO NOT USE IN NEW FILES.
4620 * Parse input and register new cgroup event handler.
4622 * Input must be in format '<event_fd> <control_fd> <args>'.
4623 * Interpretation of args is defined by control file implementation.
4625 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4626 char *buf, size_t nbytes, loff_t off)
4628 struct cgroup_subsys_state *css = of_css(of);
4629 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4630 struct mem_cgroup_event *event;
4631 struct cgroup_subsys_state *cfile_css;
4632 unsigned int efd, cfd;
4639 buf = strstrip(buf);
4641 efd = simple_strtoul(buf, &endp, 10);
4646 cfd = simple_strtoul(buf, &endp, 10);
4647 if ((*endp != ' ') && (*endp != '\0'))
4651 event = kzalloc(sizeof(*event), GFP_KERNEL);
4655 event->memcg = memcg;
4656 INIT_LIST_HEAD(&event->list);
4657 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4658 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4659 INIT_WORK(&event->remove, memcg_event_remove);
4667 event->eventfd = eventfd_ctx_fileget(efile.file);
4668 if (IS_ERR(event->eventfd)) {
4669 ret = PTR_ERR(event->eventfd);
4676 goto out_put_eventfd;
4679 /* the process need read permission on control file */
4680 /* AV: shouldn't we check that it's been opened for read instead? */
4681 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4686 * Determine the event callbacks and set them in @event. This used
4687 * to be done via struct cftype but cgroup core no longer knows
4688 * about these events. The following is crude but the whole thing
4689 * is for compatibility anyway.
4691 * DO NOT ADD NEW FILES.
4693 name = cfile.file->f_path.dentry->d_name.name;
4695 if (!strcmp(name, "memory.usage_in_bytes")) {
4696 event->register_event = mem_cgroup_usage_register_event;
4697 event->unregister_event = mem_cgroup_usage_unregister_event;
4698 } else if (!strcmp(name, "memory.oom_control")) {
4699 event->register_event = mem_cgroup_oom_register_event;
4700 event->unregister_event = mem_cgroup_oom_unregister_event;
4701 } else if (!strcmp(name, "memory.pressure_level")) {
4702 event->register_event = vmpressure_register_event;
4703 event->unregister_event = vmpressure_unregister_event;
4704 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4705 event->register_event = memsw_cgroup_usage_register_event;
4706 event->unregister_event = memsw_cgroup_usage_unregister_event;
4713 * Verify @cfile should belong to @css. Also, remaining events are
4714 * automatically removed on cgroup destruction but the removal is
4715 * asynchronous, so take an extra ref on @css.
4717 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4718 &memory_cgrp_subsys);
4720 if (IS_ERR(cfile_css))
4722 if (cfile_css != css) {
4727 ret = event->register_event(memcg, event->eventfd, buf);
4731 vfs_poll(efile.file, &event->pt);
4733 spin_lock(&memcg->event_list_lock);
4734 list_add(&event->list, &memcg->event_list);
4735 spin_unlock(&memcg->event_list_lock);
4747 eventfd_ctx_put(event->eventfd);
4756 static struct cftype mem_cgroup_legacy_files[] = {
4758 .name = "usage_in_bytes",
4759 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4760 .read_u64 = mem_cgroup_read_u64,
4763 .name = "max_usage_in_bytes",
4764 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4765 .write = mem_cgroup_reset,
4766 .read_u64 = mem_cgroup_read_u64,
4769 .name = "limit_in_bytes",
4770 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4771 .write = mem_cgroup_write,
4772 .read_u64 = mem_cgroup_read_u64,
4775 .name = "soft_limit_in_bytes",
4776 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4777 .write = mem_cgroup_write,
4778 .read_u64 = mem_cgroup_read_u64,
4782 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4783 .write = mem_cgroup_reset,
4784 .read_u64 = mem_cgroup_read_u64,
4788 .seq_show = memcg_stat_show,
4791 .name = "force_empty",
4792 .write = mem_cgroup_force_empty_write,
4795 .name = "use_hierarchy",
4796 .write_u64 = mem_cgroup_hierarchy_write,
4797 .read_u64 = mem_cgroup_hierarchy_read,
4800 .name = "cgroup.event_control", /* XXX: for compat */
4801 .write = memcg_write_event_control,
4802 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4805 .name = "swappiness",
4806 .read_u64 = mem_cgroup_swappiness_read,
4807 .write_u64 = mem_cgroup_swappiness_write,
4810 .name = "move_charge_at_immigrate",
4811 .read_u64 = mem_cgroup_move_charge_read,
4812 .write_u64 = mem_cgroup_move_charge_write,
4815 .name = "oom_control",
4816 .seq_show = mem_cgroup_oom_control_read,
4817 .write_u64 = mem_cgroup_oom_control_write,
4818 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4821 .name = "pressure_level",
4825 .name = "numa_stat",
4826 .seq_show = memcg_numa_stat_show,
4830 .name = "kmem.limit_in_bytes",
4831 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4832 .write = mem_cgroup_write,
4833 .read_u64 = mem_cgroup_read_u64,
4836 .name = "kmem.usage_in_bytes",
4837 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4838 .read_u64 = mem_cgroup_read_u64,
4841 .name = "kmem.failcnt",
4842 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4843 .write = mem_cgroup_reset,
4844 .read_u64 = mem_cgroup_read_u64,
4847 .name = "kmem.max_usage_in_bytes",
4848 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4849 .write = mem_cgroup_reset,
4850 .read_u64 = mem_cgroup_read_u64,
4852 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4854 .name = "kmem.slabinfo",
4855 .seq_start = memcg_slab_start,
4856 .seq_next = memcg_slab_next,
4857 .seq_stop = memcg_slab_stop,
4858 .seq_show = memcg_slab_show,
4862 .name = "kmem.tcp.limit_in_bytes",
4863 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4864 .write = mem_cgroup_write,
4865 .read_u64 = mem_cgroup_read_u64,
4868 .name = "kmem.tcp.usage_in_bytes",
4869 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4870 .read_u64 = mem_cgroup_read_u64,
4873 .name = "kmem.tcp.failcnt",
4874 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4875 .write = mem_cgroup_reset,
4876 .read_u64 = mem_cgroup_read_u64,
4879 .name = "kmem.tcp.max_usage_in_bytes",
4880 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4881 .write = mem_cgroup_reset,
4882 .read_u64 = mem_cgroup_read_u64,
4884 { }, /* terminate */
4888 * Private memory cgroup IDR
4890 * Swap-out records and page cache shadow entries need to store memcg
4891 * references in constrained space, so we maintain an ID space that is
4892 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4893 * memory-controlled cgroups to 64k.
4895 * However, there usually are many references to the oflline CSS after
4896 * the cgroup has been destroyed, such as page cache or reclaimable
4897 * slab objects, that don't need to hang on to the ID. We want to keep
4898 * those dead CSS from occupying IDs, or we might quickly exhaust the
4899 * relatively small ID space and prevent the creation of new cgroups
4900 * even when there are much fewer than 64k cgroups - possibly none.
4902 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4903 * be freed and recycled when it's no longer needed, which is usually
4904 * when the CSS is offlined.
4906 * The only exception to that are records of swapped out tmpfs/shmem
4907 * pages that need to be attributed to live ancestors on swapin. But
4908 * those references are manageable from userspace.
4911 static DEFINE_IDR(mem_cgroup_idr);
4913 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4915 if (memcg->id.id > 0) {
4916 idr_remove(&mem_cgroup_idr, memcg->id.id);
4921 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4923 refcount_add(n, &memcg->id.ref);
4926 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4928 if (refcount_sub_and_test(n, &memcg->id.ref)) {
4929 mem_cgroup_id_remove(memcg);
4931 /* Memcg ID pins CSS */
4932 css_put(&memcg->css);
4936 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4938 mem_cgroup_id_put_many(memcg, 1);
4942 * mem_cgroup_from_id - look up a memcg from a memcg id
4943 * @id: the memcg id to look up
4945 * Caller must hold rcu_read_lock().
4947 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4949 WARN_ON_ONCE(!rcu_read_lock_held());
4950 return idr_find(&mem_cgroup_idr, id);
4953 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4955 struct mem_cgroup_per_node *pn;
4958 * This routine is called against possible nodes.
4959 * But it's BUG to call kmalloc() against offline node.
4961 * TODO: this routine can waste much memory for nodes which will
4962 * never be onlined. It's better to use memory hotplug callback
4965 if (!node_state(node, N_NORMAL_MEMORY))
4967 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4971 pn->lruvec_stat_local = alloc_percpu(struct lruvec_stat);
4972 if (!pn->lruvec_stat_local) {
4977 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
4978 if (!pn->lruvec_stat_cpu) {
4979 free_percpu(pn->lruvec_stat_local);
4984 lruvec_init(&pn->lruvec);
4985 pn->usage_in_excess = 0;
4986 pn->on_tree = false;
4989 memcg->nodeinfo[node] = pn;
4993 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4995 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5000 free_percpu(pn->lruvec_stat_cpu);
5001 free_percpu(pn->lruvec_stat_local);
5005 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5010 * Flush percpu vmstats and vmevents to guarantee the value correctness
5011 * on parent's and all ancestor levels.
5013 memcg_flush_percpu_vmstats(memcg, false);
5014 memcg_flush_percpu_vmevents(memcg);
5016 free_mem_cgroup_per_node_info(memcg, node);
5017 free_percpu(memcg->vmstats_percpu);
5018 free_percpu(memcg->vmstats_local);
5022 static void mem_cgroup_free(struct mem_cgroup *memcg)
5024 memcg_wb_domain_exit(memcg);
5025 __mem_cgroup_free(memcg);
5028 static struct mem_cgroup *mem_cgroup_alloc(void)
5030 struct mem_cgroup *memcg;
5033 int __maybe_unused i;
5035 size = sizeof(struct mem_cgroup);
5036 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5038 memcg = kzalloc(size, GFP_KERNEL);
5042 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5043 1, MEM_CGROUP_ID_MAX,
5045 if (memcg->id.id < 0)
5048 memcg->vmstats_local = alloc_percpu(struct memcg_vmstats_percpu);
5049 if (!memcg->vmstats_local)
5052 memcg->vmstats_percpu = alloc_percpu(struct memcg_vmstats_percpu);
5053 if (!memcg->vmstats_percpu)
5057 if (alloc_mem_cgroup_per_node_info(memcg, node))
5060 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5063 INIT_WORK(&memcg->high_work, high_work_func);
5064 memcg->last_scanned_node = MAX_NUMNODES;
5065 INIT_LIST_HEAD(&memcg->oom_notify);
5066 mutex_init(&memcg->thresholds_lock);
5067 spin_lock_init(&memcg->move_lock);
5068 vmpressure_init(&memcg->vmpressure);
5069 INIT_LIST_HEAD(&memcg->event_list);
5070 spin_lock_init(&memcg->event_list_lock);
5071 memcg->socket_pressure = jiffies;
5072 #ifdef CONFIG_MEMCG_KMEM
5073 memcg->kmemcg_id = -1;
5075 #ifdef CONFIG_CGROUP_WRITEBACK
5076 INIT_LIST_HEAD(&memcg->cgwb_list);
5077 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5078 memcg->cgwb_frn[i].done =
5079 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5081 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5084 mem_cgroup_id_remove(memcg);
5085 __mem_cgroup_free(memcg);
5089 static struct cgroup_subsys_state * __ref
5090 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5092 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5093 struct mem_cgroup *memcg;
5094 long error = -ENOMEM;
5096 memcg = mem_cgroup_alloc();
5098 return ERR_PTR(error);
5100 memcg->high = PAGE_COUNTER_MAX;
5101 memcg->soft_limit = PAGE_COUNTER_MAX;
5103 memcg->swappiness = mem_cgroup_swappiness(parent);
5104 memcg->oom_kill_disable = parent->oom_kill_disable;
5106 if (parent && parent->use_hierarchy) {
5107 memcg->use_hierarchy = true;
5108 page_counter_init(&memcg->memory, &parent->memory);
5109 page_counter_init(&memcg->swap, &parent->swap);
5110 page_counter_init(&memcg->memsw, &parent->memsw);
5111 page_counter_init(&memcg->kmem, &parent->kmem);
5112 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5114 page_counter_init(&memcg->memory, NULL);
5115 page_counter_init(&memcg->swap, NULL);
5116 page_counter_init(&memcg->memsw, NULL);
5117 page_counter_init(&memcg->kmem, NULL);
5118 page_counter_init(&memcg->tcpmem, NULL);
5120 * Deeper hierachy with use_hierarchy == false doesn't make
5121 * much sense so let cgroup subsystem know about this
5122 * unfortunate state in our controller.
5124 if (parent != root_mem_cgroup)
5125 memory_cgrp_subsys.broken_hierarchy = true;
5128 /* The following stuff does not apply to the root */
5130 #ifdef CONFIG_MEMCG_KMEM
5131 INIT_LIST_HEAD(&memcg->kmem_caches);
5133 root_mem_cgroup = memcg;
5137 error = memcg_online_kmem(memcg);
5141 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5142 static_branch_inc(&memcg_sockets_enabled_key);
5146 mem_cgroup_id_remove(memcg);
5147 mem_cgroup_free(memcg);
5148 return ERR_PTR(-ENOMEM);
5151 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5153 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5156 * A memcg must be visible for memcg_expand_shrinker_maps()
5157 * by the time the maps are allocated. So, we allocate maps
5158 * here, when for_each_mem_cgroup() can't skip it.
5160 if (memcg_alloc_shrinker_maps(memcg)) {
5161 mem_cgroup_id_remove(memcg);
5165 /* Online state pins memcg ID, memcg ID pins CSS */
5166 refcount_set(&memcg->id.ref, 1);
5171 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5173 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5174 struct mem_cgroup_event *event, *tmp;
5177 * Unregister events and notify userspace.
5178 * Notify userspace about cgroup removing only after rmdir of cgroup
5179 * directory to avoid race between userspace and kernelspace.
5181 spin_lock(&memcg->event_list_lock);
5182 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5183 list_del_init(&event->list);
5184 schedule_work(&event->remove);
5186 spin_unlock(&memcg->event_list_lock);
5188 page_counter_set_min(&memcg->memory, 0);
5189 page_counter_set_low(&memcg->memory, 0);
5191 memcg_offline_kmem(memcg);
5192 wb_memcg_offline(memcg);
5194 drain_all_stock(memcg);
5196 mem_cgroup_id_put(memcg);
5199 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5201 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5203 invalidate_reclaim_iterators(memcg);
5206 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5208 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5209 int __maybe_unused i;
5211 #ifdef CONFIG_CGROUP_WRITEBACK
5212 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5213 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5215 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5216 static_branch_dec(&memcg_sockets_enabled_key);
5218 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5219 static_branch_dec(&memcg_sockets_enabled_key);
5221 vmpressure_cleanup(&memcg->vmpressure);
5222 cancel_work_sync(&memcg->high_work);
5223 mem_cgroup_remove_from_trees(memcg);
5224 memcg_free_shrinker_maps(memcg);
5225 memcg_free_kmem(memcg);
5226 mem_cgroup_free(memcg);
5230 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5231 * @css: the target css
5233 * Reset the states of the mem_cgroup associated with @css. This is
5234 * invoked when the userland requests disabling on the default hierarchy
5235 * but the memcg is pinned through dependency. The memcg should stop
5236 * applying policies and should revert to the vanilla state as it may be
5237 * made visible again.
5239 * The current implementation only resets the essential configurations.
5240 * This needs to be expanded to cover all the visible parts.
5242 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5244 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5246 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5247 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5248 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
5249 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5250 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5251 page_counter_set_min(&memcg->memory, 0);
5252 page_counter_set_low(&memcg->memory, 0);
5253 memcg->high = PAGE_COUNTER_MAX;
5254 memcg->soft_limit = PAGE_COUNTER_MAX;
5255 memcg_wb_domain_size_changed(memcg);
5259 /* Handlers for move charge at task migration. */
5260 static int mem_cgroup_do_precharge(unsigned long count)
5264 /* Try a single bulk charge without reclaim first, kswapd may wake */
5265 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5267 mc.precharge += count;
5271 /* Try charges one by one with reclaim, but do not retry */
5273 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5287 enum mc_target_type {
5294 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5295 unsigned long addr, pte_t ptent)
5297 struct page *page = vm_normal_page(vma, addr, ptent);
5299 if (!page || !page_mapped(page))
5301 if (PageAnon(page)) {
5302 if (!(mc.flags & MOVE_ANON))
5305 if (!(mc.flags & MOVE_FILE))
5308 if (!get_page_unless_zero(page))
5314 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5315 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5316 pte_t ptent, swp_entry_t *entry)
5318 struct page *page = NULL;
5319 swp_entry_t ent = pte_to_swp_entry(ptent);
5321 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
5325 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5326 * a device and because they are not accessible by CPU they are store
5327 * as special swap entry in the CPU page table.
5329 if (is_device_private_entry(ent)) {
5330 page = device_private_entry_to_page(ent);
5332 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5333 * a refcount of 1 when free (unlike normal page)
5335 if (!page_ref_add_unless(page, 1, 1))
5341 * Because lookup_swap_cache() updates some statistics counter,
5342 * we call find_get_page() with swapper_space directly.
5344 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5345 if (do_memsw_account())
5346 entry->val = ent.val;
5351 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5352 pte_t ptent, swp_entry_t *entry)
5358 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5359 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5361 struct page *page = NULL;
5362 struct address_space *mapping;
5365 if (!vma->vm_file) /* anonymous vma */
5367 if (!(mc.flags & MOVE_FILE))
5370 mapping = vma->vm_file->f_mapping;
5371 pgoff = linear_page_index(vma, addr);
5373 /* page is moved even if it's not RSS of this task(page-faulted). */
5375 /* shmem/tmpfs may report page out on swap: account for that too. */
5376 if (shmem_mapping(mapping)) {
5377 page = find_get_entry(mapping, pgoff);
5378 if (xa_is_value(page)) {
5379 swp_entry_t swp = radix_to_swp_entry(page);
5380 if (do_memsw_account())
5382 page = find_get_page(swap_address_space(swp),
5386 page = find_get_page(mapping, pgoff);
5388 page = find_get_page(mapping, pgoff);
5394 * mem_cgroup_move_account - move account of the page
5396 * @compound: charge the page as compound or small page
5397 * @from: mem_cgroup which the page is moved from.
5398 * @to: mem_cgroup which the page is moved to. @from != @to.
5400 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5402 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5405 static int mem_cgroup_move_account(struct page *page,
5407 struct mem_cgroup *from,
5408 struct mem_cgroup *to)
5410 unsigned long flags;
5411 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5415 VM_BUG_ON(from == to);
5416 VM_BUG_ON_PAGE(PageLRU(page), page);
5417 VM_BUG_ON(compound && !PageTransHuge(page));
5420 * Prevent mem_cgroup_migrate() from looking at
5421 * page->mem_cgroup of its source page while we change it.
5424 if (!trylock_page(page))
5428 if (page->mem_cgroup != from)
5431 anon = PageAnon(page);
5433 spin_lock_irqsave(&from->move_lock, flags);
5435 if (!anon && page_mapped(page)) {
5436 __mod_memcg_state(from, NR_FILE_MAPPED, -nr_pages);
5437 __mod_memcg_state(to, NR_FILE_MAPPED, nr_pages);
5441 * move_lock grabbed above and caller set from->moving_account, so
5442 * mod_memcg_page_state will serialize updates to PageDirty.
5443 * So mapping should be stable for dirty pages.
5445 if (!anon && PageDirty(page)) {
5446 struct address_space *mapping = page_mapping(page);
5448 if (mapping_cap_account_dirty(mapping)) {
5449 __mod_memcg_state(from, NR_FILE_DIRTY, -nr_pages);
5450 __mod_memcg_state(to, NR_FILE_DIRTY, nr_pages);
5454 if (PageWriteback(page)) {
5455 __mod_memcg_state(from, NR_WRITEBACK, -nr_pages);
5456 __mod_memcg_state(to, NR_WRITEBACK, nr_pages);
5460 * It is safe to change page->mem_cgroup here because the page
5461 * is referenced, charged, and isolated - we can't race with
5462 * uncharging, charging, migration, or LRU putback.
5465 /* caller should have done css_get */
5466 page->mem_cgroup = to;
5467 spin_unlock_irqrestore(&from->move_lock, flags);
5471 local_irq_disable();
5472 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
5473 memcg_check_events(to, page);
5474 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
5475 memcg_check_events(from, page);
5484 * get_mctgt_type - get target type of moving charge
5485 * @vma: the vma the pte to be checked belongs
5486 * @addr: the address corresponding to the pte to be checked
5487 * @ptent: the pte to be checked
5488 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5491 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5492 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5493 * move charge. if @target is not NULL, the page is stored in target->page
5494 * with extra refcnt got(Callers should handle it).
5495 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5496 * target for charge migration. if @target is not NULL, the entry is stored
5498 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5499 * (so ZONE_DEVICE page and thus not on the lru).
5500 * For now we such page is charge like a regular page would be as for all
5501 * intent and purposes it is just special memory taking the place of a
5504 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5506 * Called with pte lock held.
5509 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5510 unsigned long addr, pte_t ptent, union mc_target *target)
5512 struct page *page = NULL;
5513 enum mc_target_type ret = MC_TARGET_NONE;
5514 swp_entry_t ent = { .val = 0 };
5516 if (pte_present(ptent))
5517 page = mc_handle_present_pte(vma, addr, ptent);
5518 else if (is_swap_pte(ptent))
5519 page = mc_handle_swap_pte(vma, ptent, &ent);
5520 else if (pte_none(ptent))
5521 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5523 if (!page && !ent.val)
5527 * Do only loose check w/o serialization.
5528 * mem_cgroup_move_account() checks the page is valid or
5529 * not under LRU exclusion.
5531 if (page->mem_cgroup == mc.from) {
5532 ret = MC_TARGET_PAGE;
5533 if (is_device_private_page(page))
5534 ret = MC_TARGET_DEVICE;
5536 target->page = page;
5538 if (!ret || !target)
5542 * There is a swap entry and a page doesn't exist or isn't charged.
5543 * But we cannot move a tail-page in a THP.
5545 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5546 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5547 ret = MC_TARGET_SWAP;
5554 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5556 * We don't consider PMD mapped swapping or file mapped pages because THP does
5557 * not support them for now.
5558 * Caller should make sure that pmd_trans_huge(pmd) is true.
5560 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5561 unsigned long addr, pmd_t pmd, union mc_target *target)
5563 struct page *page = NULL;
5564 enum mc_target_type ret = MC_TARGET_NONE;
5566 if (unlikely(is_swap_pmd(pmd))) {
5567 VM_BUG_ON(thp_migration_supported() &&
5568 !is_pmd_migration_entry(pmd));
5571 page = pmd_page(pmd);
5572 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5573 if (!(mc.flags & MOVE_ANON))
5575 if (page->mem_cgroup == mc.from) {
5576 ret = MC_TARGET_PAGE;
5579 target->page = page;
5585 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5586 unsigned long addr, pmd_t pmd, union mc_target *target)
5588 return MC_TARGET_NONE;
5592 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5593 unsigned long addr, unsigned long end,
5594 struct mm_walk *walk)
5596 struct vm_area_struct *vma = walk->vma;
5600 ptl = pmd_trans_huge_lock(pmd, vma);
5603 * Note their can not be MC_TARGET_DEVICE for now as we do not
5604 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5605 * this might change.
5607 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5608 mc.precharge += HPAGE_PMD_NR;
5613 if (pmd_trans_unstable(pmd))
5615 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5616 for (; addr != end; pte++, addr += PAGE_SIZE)
5617 if (get_mctgt_type(vma, addr, *pte, NULL))
5618 mc.precharge++; /* increment precharge temporarily */
5619 pte_unmap_unlock(pte - 1, ptl);
5625 static const struct mm_walk_ops precharge_walk_ops = {
5626 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5629 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5631 unsigned long precharge;
5633 down_read(&mm->mmap_sem);
5634 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5635 up_read(&mm->mmap_sem);
5637 precharge = mc.precharge;
5643 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5645 unsigned long precharge = mem_cgroup_count_precharge(mm);
5647 VM_BUG_ON(mc.moving_task);
5648 mc.moving_task = current;
5649 return mem_cgroup_do_precharge(precharge);
5652 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5653 static void __mem_cgroup_clear_mc(void)
5655 struct mem_cgroup *from = mc.from;
5656 struct mem_cgroup *to = mc.to;
5658 /* we must uncharge all the leftover precharges from mc.to */
5660 cancel_charge(mc.to, mc.precharge);
5664 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5665 * we must uncharge here.
5667 if (mc.moved_charge) {
5668 cancel_charge(mc.from, mc.moved_charge);
5669 mc.moved_charge = 0;
5671 /* we must fixup refcnts and charges */
5672 if (mc.moved_swap) {
5673 /* uncharge swap account from the old cgroup */
5674 if (!mem_cgroup_is_root(mc.from))
5675 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5677 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5680 * we charged both to->memory and to->memsw, so we
5681 * should uncharge to->memory.
5683 if (!mem_cgroup_is_root(mc.to))
5684 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5686 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
5687 css_put_many(&mc.to->css, mc.moved_swap);
5691 memcg_oom_recover(from);
5692 memcg_oom_recover(to);
5693 wake_up_all(&mc.waitq);
5696 static void mem_cgroup_clear_mc(void)
5698 struct mm_struct *mm = mc.mm;
5701 * we must clear moving_task before waking up waiters at the end of
5704 mc.moving_task = NULL;
5705 __mem_cgroup_clear_mc();
5706 spin_lock(&mc.lock);
5710 spin_unlock(&mc.lock);
5715 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5717 struct cgroup_subsys_state *css;
5718 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5719 struct mem_cgroup *from;
5720 struct task_struct *leader, *p;
5721 struct mm_struct *mm;
5722 unsigned long move_flags;
5725 /* charge immigration isn't supported on the default hierarchy */
5726 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5730 * Multi-process migrations only happen on the default hierarchy
5731 * where charge immigration is not used. Perform charge
5732 * immigration if @tset contains a leader and whine if there are
5736 cgroup_taskset_for_each_leader(leader, css, tset) {
5739 memcg = mem_cgroup_from_css(css);
5745 * We are now commited to this value whatever it is. Changes in this
5746 * tunable will only affect upcoming migrations, not the current one.
5747 * So we need to save it, and keep it going.
5749 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5753 from = mem_cgroup_from_task(p);
5755 VM_BUG_ON(from == memcg);
5757 mm = get_task_mm(p);
5760 /* We move charges only when we move a owner of the mm */
5761 if (mm->owner == p) {
5764 VM_BUG_ON(mc.precharge);
5765 VM_BUG_ON(mc.moved_charge);
5766 VM_BUG_ON(mc.moved_swap);
5768 spin_lock(&mc.lock);
5772 mc.flags = move_flags;
5773 spin_unlock(&mc.lock);
5774 /* We set mc.moving_task later */
5776 ret = mem_cgroup_precharge_mc(mm);
5778 mem_cgroup_clear_mc();
5785 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5788 mem_cgroup_clear_mc();
5791 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5792 unsigned long addr, unsigned long end,
5793 struct mm_walk *walk)
5796 struct vm_area_struct *vma = walk->vma;
5799 enum mc_target_type target_type;
5800 union mc_target target;
5803 ptl = pmd_trans_huge_lock(pmd, vma);
5805 if (mc.precharge < HPAGE_PMD_NR) {
5809 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5810 if (target_type == MC_TARGET_PAGE) {
5812 if (!isolate_lru_page(page)) {
5813 if (!mem_cgroup_move_account(page, true,
5815 mc.precharge -= HPAGE_PMD_NR;
5816 mc.moved_charge += HPAGE_PMD_NR;
5818 putback_lru_page(page);
5821 } else if (target_type == MC_TARGET_DEVICE) {
5823 if (!mem_cgroup_move_account(page, true,
5825 mc.precharge -= HPAGE_PMD_NR;
5826 mc.moved_charge += HPAGE_PMD_NR;
5834 if (pmd_trans_unstable(pmd))
5837 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5838 for (; addr != end; addr += PAGE_SIZE) {
5839 pte_t ptent = *(pte++);
5840 bool device = false;
5846 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5847 case MC_TARGET_DEVICE:
5850 case MC_TARGET_PAGE:
5853 * We can have a part of the split pmd here. Moving it
5854 * can be done but it would be too convoluted so simply
5855 * ignore such a partial THP and keep it in original
5856 * memcg. There should be somebody mapping the head.
5858 if (PageTransCompound(page))
5860 if (!device && isolate_lru_page(page))
5862 if (!mem_cgroup_move_account(page, false,
5865 /* we uncharge from mc.from later. */
5869 putback_lru_page(page);
5870 put: /* get_mctgt_type() gets the page */
5873 case MC_TARGET_SWAP:
5875 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5877 /* we fixup refcnts and charges later. */
5885 pte_unmap_unlock(pte - 1, ptl);
5890 * We have consumed all precharges we got in can_attach().
5891 * We try charge one by one, but don't do any additional
5892 * charges to mc.to if we have failed in charge once in attach()
5895 ret = mem_cgroup_do_precharge(1);
5903 static const struct mm_walk_ops charge_walk_ops = {
5904 .pmd_entry = mem_cgroup_move_charge_pte_range,
5907 static void mem_cgroup_move_charge(void)
5909 lru_add_drain_all();
5911 * Signal lock_page_memcg() to take the memcg's move_lock
5912 * while we're moving its pages to another memcg. Then wait
5913 * for already started RCU-only updates to finish.
5915 atomic_inc(&mc.from->moving_account);
5918 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5920 * Someone who are holding the mmap_sem might be waiting in
5921 * waitq. So we cancel all extra charges, wake up all waiters,
5922 * and retry. Because we cancel precharges, we might not be able
5923 * to move enough charges, but moving charge is a best-effort
5924 * feature anyway, so it wouldn't be a big problem.
5926 __mem_cgroup_clear_mc();
5931 * When we have consumed all precharges and failed in doing
5932 * additional charge, the page walk just aborts.
5934 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
5937 up_read(&mc.mm->mmap_sem);
5938 atomic_dec(&mc.from->moving_account);
5941 static void mem_cgroup_move_task(void)
5944 mem_cgroup_move_charge();
5945 mem_cgroup_clear_mc();
5948 #else /* !CONFIG_MMU */
5949 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5953 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5956 static void mem_cgroup_move_task(void)
5962 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5963 * to verify whether we're attached to the default hierarchy on each mount
5966 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5969 * use_hierarchy is forced on the default hierarchy. cgroup core
5970 * guarantees that @root doesn't have any children, so turning it
5971 * on for the root memcg is enough.
5973 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5974 root_mem_cgroup->use_hierarchy = true;
5976 root_mem_cgroup->use_hierarchy = false;
5979 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
5981 if (value == PAGE_COUNTER_MAX)
5982 seq_puts(m, "max\n");
5984 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
5989 static u64 memory_current_read(struct cgroup_subsys_state *css,
5992 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5994 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5997 static int memory_min_show(struct seq_file *m, void *v)
5999 return seq_puts_memcg_tunable(m,
6000 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6003 static ssize_t memory_min_write(struct kernfs_open_file *of,
6004 char *buf, size_t nbytes, loff_t off)
6006 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6010 buf = strstrip(buf);
6011 err = page_counter_memparse(buf, "max", &min);
6015 page_counter_set_min(&memcg->memory, min);
6020 static int memory_low_show(struct seq_file *m, void *v)
6022 return seq_puts_memcg_tunable(m,
6023 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6026 static ssize_t memory_low_write(struct kernfs_open_file *of,
6027 char *buf, size_t nbytes, loff_t off)
6029 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6033 buf = strstrip(buf);
6034 err = page_counter_memparse(buf, "max", &low);
6038 page_counter_set_low(&memcg->memory, low);
6043 static int memory_high_show(struct seq_file *m, void *v)
6045 return seq_puts_memcg_tunable(m, READ_ONCE(mem_cgroup_from_seq(m)->high));
6048 static ssize_t memory_high_write(struct kernfs_open_file *of,
6049 char *buf, size_t nbytes, loff_t off)
6051 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6052 unsigned long nr_pages;
6056 buf = strstrip(buf);
6057 err = page_counter_memparse(buf, "max", &high);
6063 nr_pages = page_counter_read(&memcg->memory);
6064 if (nr_pages > high)
6065 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6068 memcg_wb_domain_size_changed(memcg);
6072 static int memory_max_show(struct seq_file *m, void *v)
6074 return seq_puts_memcg_tunable(m,
6075 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6078 static ssize_t memory_max_write(struct kernfs_open_file *of,
6079 char *buf, size_t nbytes, loff_t off)
6081 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6082 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
6083 bool drained = false;
6087 buf = strstrip(buf);
6088 err = page_counter_memparse(buf, "max", &max);
6092 xchg(&memcg->memory.max, max);
6095 unsigned long nr_pages = page_counter_read(&memcg->memory);
6097 if (nr_pages <= max)
6100 if (signal_pending(current)) {
6106 drain_all_stock(memcg);
6112 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6118 memcg_memory_event(memcg, MEMCG_OOM);
6119 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6123 memcg_wb_domain_size_changed(memcg);
6127 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6129 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6130 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6131 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6132 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6133 seq_printf(m, "oom_kill %lu\n",
6134 atomic_long_read(&events[MEMCG_OOM_KILL]));
6137 static int memory_events_show(struct seq_file *m, void *v)
6139 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6141 __memory_events_show(m, memcg->memory_events);
6145 static int memory_events_local_show(struct seq_file *m, void *v)
6147 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6149 __memory_events_show(m, memcg->memory_events_local);
6153 static int memory_stat_show(struct seq_file *m, void *v)
6155 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6158 buf = memory_stat_format(memcg);
6166 static int memory_oom_group_show(struct seq_file *m, void *v)
6168 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6170 seq_printf(m, "%d\n", memcg->oom_group);
6175 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6176 char *buf, size_t nbytes, loff_t off)
6178 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6181 buf = strstrip(buf);
6185 ret = kstrtoint(buf, 0, &oom_group);
6189 if (oom_group != 0 && oom_group != 1)
6192 memcg->oom_group = oom_group;
6197 static struct cftype memory_files[] = {
6200 .flags = CFTYPE_NOT_ON_ROOT,
6201 .read_u64 = memory_current_read,
6205 .flags = CFTYPE_NOT_ON_ROOT,
6206 .seq_show = memory_min_show,
6207 .write = memory_min_write,
6211 .flags = CFTYPE_NOT_ON_ROOT,
6212 .seq_show = memory_low_show,
6213 .write = memory_low_write,
6217 .flags = CFTYPE_NOT_ON_ROOT,
6218 .seq_show = memory_high_show,
6219 .write = memory_high_write,
6223 .flags = CFTYPE_NOT_ON_ROOT,
6224 .seq_show = memory_max_show,
6225 .write = memory_max_write,
6229 .flags = CFTYPE_NOT_ON_ROOT,
6230 .file_offset = offsetof(struct mem_cgroup, events_file),
6231 .seq_show = memory_events_show,
6234 .name = "events.local",
6235 .flags = CFTYPE_NOT_ON_ROOT,
6236 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6237 .seq_show = memory_events_local_show,
6241 .flags = CFTYPE_NOT_ON_ROOT,
6242 .seq_show = memory_stat_show,
6245 .name = "oom.group",
6246 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6247 .seq_show = memory_oom_group_show,
6248 .write = memory_oom_group_write,
6253 struct cgroup_subsys memory_cgrp_subsys = {
6254 .css_alloc = mem_cgroup_css_alloc,
6255 .css_online = mem_cgroup_css_online,
6256 .css_offline = mem_cgroup_css_offline,
6257 .css_released = mem_cgroup_css_released,
6258 .css_free = mem_cgroup_css_free,
6259 .css_reset = mem_cgroup_css_reset,
6260 .can_attach = mem_cgroup_can_attach,
6261 .cancel_attach = mem_cgroup_cancel_attach,
6262 .post_attach = mem_cgroup_move_task,
6263 .bind = mem_cgroup_bind,
6264 .dfl_cftypes = memory_files,
6265 .legacy_cftypes = mem_cgroup_legacy_files,
6270 * mem_cgroup_protected - check if memory consumption is in the normal range
6271 * @root: the top ancestor of the sub-tree being checked
6272 * @memcg: the memory cgroup to check
6274 * WARNING: This function is not stateless! It can only be used as part
6275 * of a top-down tree iteration, not for isolated queries.
6277 * Returns one of the following:
6278 * MEMCG_PROT_NONE: cgroup memory is not protected
6279 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
6280 * an unprotected supply of reclaimable memory from other cgroups.
6281 * MEMCG_PROT_MIN: cgroup memory is protected
6283 * @root is exclusive; it is never protected when looked at directly
6285 * To provide a proper hierarchical behavior, effective memory.min/low values
6286 * are used. Below is the description of how effective memory.low is calculated.
6287 * Effective memory.min values is calculated in the same way.
6289 * Effective memory.low is always equal or less than the original memory.low.
6290 * If there is no memory.low overcommittment (which is always true for
6291 * top-level memory cgroups), these two values are equal.
6292 * Otherwise, it's a part of parent's effective memory.low,
6293 * calculated as a cgroup's memory.low usage divided by sum of sibling's
6294 * memory.low usages, where memory.low usage is the size of actually
6298 * elow = min( memory.low, parent->elow * ------------------ ),
6299 * siblings_low_usage
6301 * | memory.current, if memory.current < memory.low
6306 * Such definition of the effective memory.low provides the expected
6307 * hierarchical behavior: parent's memory.low value is limiting
6308 * children, unprotected memory is reclaimed first and cgroups,
6309 * which are not using their guarantee do not affect actual memory
6312 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
6314 * A A/memory.low = 2G, A/memory.current = 6G
6316 * BC DE B/memory.low = 3G B/memory.current = 2G
6317 * C/memory.low = 1G C/memory.current = 2G
6318 * D/memory.low = 0 D/memory.current = 2G
6319 * E/memory.low = 10G E/memory.current = 0
6321 * and the memory pressure is applied, the following memory distribution
6322 * is expected (approximately):
6324 * A/memory.current = 2G
6326 * B/memory.current = 1.3G
6327 * C/memory.current = 0.6G
6328 * D/memory.current = 0
6329 * E/memory.current = 0
6331 * These calculations require constant tracking of the actual low usages
6332 * (see propagate_protected_usage()), as well as recursive calculation of
6333 * effective memory.low values. But as we do call mem_cgroup_protected()
6334 * path for each memory cgroup top-down from the reclaim,
6335 * it's possible to optimize this part, and save calculated elow
6336 * for next usage. This part is intentionally racy, but it's ok,
6337 * as memory.low is a best-effort mechanism.
6339 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
6340 struct mem_cgroup *memcg)
6342 struct mem_cgroup *parent;
6343 unsigned long emin, parent_emin;
6344 unsigned long elow, parent_elow;
6345 unsigned long usage;
6347 if (mem_cgroup_disabled())
6348 return MEMCG_PROT_NONE;
6351 root = root_mem_cgroup;
6353 return MEMCG_PROT_NONE;
6355 usage = page_counter_read(&memcg->memory);
6357 return MEMCG_PROT_NONE;
6359 emin = memcg->memory.min;
6360 elow = memcg->memory.low;
6362 parent = parent_mem_cgroup(memcg);
6363 /* No parent means a non-hierarchical mode on v1 memcg */
6365 return MEMCG_PROT_NONE;
6370 parent_emin = READ_ONCE(parent->memory.emin);
6371 emin = min(emin, parent_emin);
6372 if (emin && parent_emin) {
6373 unsigned long min_usage, siblings_min_usage;
6375 min_usage = min(usage, memcg->memory.min);
6376 siblings_min_usage = atomic_long_read(
6377 &parent->memory.children_min_usage);
6379 if (min_usage && siblings_min_usage)
6380 emin = min(emin, parent_emin * min_usage /
6381 siblings_min_usage);
6384 parent_elow = READ_ONCE(parent->memory.elow);
6385 elow = min(elow, parent_elow);
6386 if (elow && parent_elow) {
6387 unsigned long low_usage, siblings_low_usage;
6389 low_usage = min(usage, memcg->memory.low);
6390 siblings_low_usage = atomic_long_read(
6391 &parent->memory.children_low_usage);
6393 if (low_usage && siblings_low_usage)
6394 elow = min(elow, parent_elow * low_usage /
6395 siblings_low_usage);
6399 memcg->memory.emin = emin;
6400 memcg->memory.elow = elow;
6403 return MEMCG_PROT_MIN;
6404 else if (usage <= elow)
6405 return MEMCG_PROT_LOW;
6407 return MEMCG_PROT_NONE;
6411 * mem_cgroup_try_charge - try charging a page
6412 * @page: page to charge
6413 * @mm: mm context of the victim
6414 * @gfp_mask: reclaim mode
6415 * @memcgp: charged memcg return
6416 * @compound: charge the page as compound or small page
6418 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6419 * pages according to @gfp_mask if necessary.
6421 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6422 * Otherwise, an error code is returned.
6424 * After page->mapping has been set up, the caller must finalize the
6425 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6426 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6428 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
6429 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6432 struct mem_cgroup *memcg = NULL;
6433 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6436 if (mem_cgroup_disabled())
6439 if (PageSwapCache(page)) {
6441 * Every swap fault against a single page tries to charge the
6442 * page, bail as early as possible. shmem_unuse() encounters
6443 * already charged pages, too. The USED bit is protected by
6444 * the page lock, which serializes swap cache removal, which
6445 * in turn serializes uncharging.
6447 VM_BUG_ON_PAGE(!PageLocked(page), page);
6448 if (compound_head(page)->mem_cgroup)
6451 if (do_swap_account) {
6452 swp_entry_t ent = { .val = page_private(page), };
6453 unsigned short id = lookup_swap_cgroup_id(ent);
6456 memcg = mem_cgroup_from_id(id);
6457 if (memcg && !css_tryget_online(&memcg->css))
6464 memcg = get_mem_cgroup_from_mm(mm);
6466 ret = try_charge(memcg, gfp_mask, nr_pages);
6468 css_put(&memcg->css);
6474 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
6475 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6478 struct mem_cgroup *memcg;
6481 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound);
6483 mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask);
6488 * mem_cgroup_commit_charge - commit a page charge
6489 * @page: page to charge
6490 * @memcg: memcg to charge the page to
6491 * @lrucare: page might be on LRU already
6492 * @compound: charge the page as compound or small page
6494 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6495 * after page->mapping has been set up. This must happen atomically
6496 * as part of the page instantiation, i.e. under the page table lock
6497 * for anonymous pages, under the page lock for page and swap cache.
6499 * In addition, the page must not be on the LRU during the commit, to
6500 * prevent racing with task migration. If it might be, use @lrucare.
6502 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6504 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
6505 bool lrucare, bool compound)
6507 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6509 VM_BUG_ON_PAGE(!page->mapping, page);
6510 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
6512 if (mem_cgroup_disabled())
6515 * Swap faults will attempt to charge the same page multiple
6516 * times. But reuse_swap_page() might have removed the page
6517 * from swapcache already, so we can't check PageSwapCache().
6522 commit_charge(page, memcg, lrucare);
6524 local_irq_disable();
6525 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
6526 memcg_check_events(memcg, page);
6529 if (do_memsw_account() && PageSwapCache(page)) {
6530 swp_entry_t entry = { .val = page_private(page) };
6532 * The swap entry might not get freed for a long time,
6533 * let's not wait for it. The page already received a
6534 * memory+swap charge, drop the swap entry duplicate.
6536 mem_cgroup_uncharge_swap(entry, nr_pages);
6541 * mem_cgroup_cancel_charge - cancel a page charge
6542 * @page: page to charge
6543 * @memcg: memcg to charge the page to
6544 * @compound: charge the page as compound or small page
6546 * Cancel a charge transaction started by mem_cgroup_try_charge().
6548 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
6551 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6553 if (mem_cgroup_disabled())
6556 * Swap faults will attempt to charge the same page multiple
6557 * times. But reuse_swap_page() might have removed the page
6558 * from swapcache already, so we can't check PageSwapCache().
6563 cancel_charge(memcg, nr_pages);
6566 struct uncharge_gather {
6567 struct mem_cgroup *memcg;
6568 unsigned long pgpgout;
6569 unsigned long nr_anon;
6570 unsigned long nr_file;
6571 unsigned long nr_kmem;
6572 unsigned long nr_huge;
6573 unsigned long nr_shmem;
6574 struct page *dummy_page;
6577 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6579 memset(ug, 0, sizeof(*ug));
6582 static void uncharge_batch(const struct uncharge_gather *ug)
6584 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
6585 unsigned long flags;
6587 if (!mem_cgroup_is_root(ug->memcg)) {
6588 page_counter_uncharge(&ug->memcg->memory, nr_pages);
6589 if (do_memsw_account())
6590 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
6591 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6592 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6593 memcg_oom_recover(ug->memcg);
6596 local_irq_save(flags);
6597 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
6598 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
6599 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
6600 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
6601 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6602 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, nr_pages);
6603 memcg_check_events(ug->memcg, ug->dummy_page);
6604 local_irq_restore(flags);
6606 if (!mem_cgroup_is_root(ug->memcg))
6607 css_put_many(&ug->memcg->css, nr_pages);
6610 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6612 VM_BUG_ON_PAGE(PageLRU(page), page);
6613 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
6614 !PageHWPoison(page) , page);
6616 if (!page->mem_cgroup)
6620 * Nobody should be changing or seriously looking at
6621 * page->mem_cgroup at this point, we have fully
6622 * exclusive access to the page.
6625 if (ug->memcg != page->mem_cgroup) {
6628 uncharge_gather_clear(ug);
6630 ug->memcg = page->mem_cgroup;
6633 if (!PageKmemcg(page)) {
6634 unsigned int nr_pages = 1;
6636 if (PageTransHuge(page)) {
6637 nr_pages = compound_nr(page);
6638 ug->nr_huge += nr_pages;
6641 ug->nr_anon += nr_pages;
6643 ug->nr_file += nr_pages;
6644 if (PageSwapBacked(page))
6645 ug->nr_shmem += nr_pages;
6649 ug->nr_kmem += compound_nr(page);
6650 __ClearPageKmemcg(page);
6653 ug->dummy_page = page;
6654 page->mem_cgroup = NULL;
6657 static void uncharge_list(struct list_head *page_list)
6659 struct uncharge_gather ug;
6660 struct list_head *next;
6662 uncharge_gather_clear(&ug);
6665 * Note that the list can be a single page->lru; hence the
6666 * do-while loop instead of a simple list_for_each_entry().
6668 next = page_list->next;
6672 page = list_entry(next, struct page, lru);
6673 next = page->lru.next;
6675 uncharge_page(page, &ug);
6676 } while (next != page_list);
6679 uncharge_batch(&ug);
6683 * mem_cgroup_uncharge - uncharge a page
6684 * @page: page to uncharge
6686 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6687 * mem_cgroup_commit_charge().
6689 void mem_cgroup_uncharge(struct page *page)
6691 struct uncharge_gather ug;
6693 if (mem_cgroup_disabled())
6696 /* Don't touch page->lru of any random page, pre-check: */
6697 if (!page->mem_cgroup)
6700 uncharge_gather_clear(&ug);
6701 uncharge_page(page, &ug);
6702 uncharge_batch(&ug);
6706 * mem_cgroup_uncharge_list - uncharge a list of page
6707 * @page_list: list of pages to uncharge
6709 * Uncharge a list of pages previously charged with
6710 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6712 void mem_cgroup_uncharge_list(struct list_head *page_list)
6714 if (mem_cgroup_disabled())
6717 if (!list_empty(page_list))
6718 uncharge_list(page_list);
6722 * mem_cgroup_migrate - charge a page's replacement
6723 * @oldpage: currently circulating page
6724 * @newpage: replacement page
6726 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6727 * be uncharged upon free.
6729 * Both pages must be locked, @newpage->mapping must be set up.
6731 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6733 struct mem_cgroup *memcg;
6734 unsigned int nr_pages;
6736 unsigned long flags;
6738 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6739 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6740 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6741 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6744 if (mem_cgroup_disabled())
6747 /* Page cache replacement: new page already charged? */
6748 if (newpage->mem_cgroup)
6751 /* Swapcache readahead pages can get replaced before being charged */
6752 memcg = oldpage->mem_cgroup;
6756 /* Force-charge the new page. The old one will be freed soon */
6757 compound = PageTransHuge(newpage);
6758 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
6760 page_counter_charge(&memcg->memory, nr_pages);
6761 if (do_memsw_account())
6762 page_counter_charge(&memcg->memsw, nr_pages);
6763 css_get_many(&memcg->css, nr_pages);
6765 commit_charge(newpage, memcg, false);
6767 local_irq_save(flags);
6768 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
6769 memcg_check_events(memcg, newpage);
6770 local_irq_restore(flags);
6773 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6774 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6776 void mem_cgroup_sk_alloc(struct sock *sk)
6778 struct mem_cgroup *memcg;
6780 if (!mem_cgroup_sockets_enabled)
6784 * Socket cloning can throw us here with sk_memcg already
6785 * filled. It won't however, necessarily happen from
6786 * process context. So the test for root memcg given
6787 * the current task's memcg won't help us in this case.
6789 * Respecting the original socket's memcg is a better
6790 * decision in this case.
6793 css_get(&sk->sk_memcg->css);
6798 memcg = mem_cgroup_from_task(current);
6799 if (memcg == root_mem_cgroup)
6801 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6803 if (css_tryget_online(&memcg->css))
6804 sk->sk_memcg = memcg;
6809 void mem_cgroup_sk_free(struct sock *sk)
6812 css_put(&sk->sk_memcg->css);
6816 * mem_cgroup_charge_skmem - charge socket memory
6817 * @memcg: memcg to charge
6818 * @nr_pages: number of pages to charge
6820 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6821 * @memcg's configured limit, %false if the charge had to be forced.
6823 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6825 gfp_t gfp_mask = GFP_KERNEL;
6827 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6828 struct page_counter *fail;
6830 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6831 memcg->tcpmem_pressure = 0;
6834 page_counter_charge(&memcg->tcpmem, nr_pages);
6835 memcg->tcpmem_pressure = 1;
6839 /* Don't block in the packet receive path */
6841 gfp_mask = GFP_NOWAIT;
6843 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6845 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6848 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6853 * mem_cgroup_uncharge_skmem - uncharge socket memory
6854 * @memcg: memcg to uncharge
6855 * @nr_pages: number of pages to uncharge
6857 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6859 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6860 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6864 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6866 refill_stock(memcg, nr_pages);
6869 static int __init cgroup_memory(char *s)
6873 while ((token = strsep(&s, ",")) != NULL) {
6876 if (!strcmp(token, "nosocket"))
6877 cgroup_memory_nosocket = true;
6878 if (!strcmp(token, "nokmem"))
6879 cgroup_memory_nokmem = true;
6883 __setup("cgroup.memory=", cgroup_memory);
6886 * subsys_initcall() for memory controller.
6888 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6889 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6890 * basically everything that doesn't depend on a specific mem_cgroup structure
6891 * should be initialized from here.
6893 static int __init mem_cgroup_init(void)
6897 #ifdef CONFIG_MEMCG_KMEM
6899 * Kmem cache creation is mostly done with the slab_mutex held,
6900 * so use a workqueue with limited concurrency to avoid stalling
6901 * all worker threads in case lots of cgroups are created and
6902 * destroyed simultaneously.
6904 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
6905 BUG_ON(!memcg_kmem_cache_wq);
6908 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6909 memcg_hotplug_cpu_dead);
6911 for_each_possible_cpu(cpu)
6912 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
6915 for_each_node(node) {
6916 struct mem_cgroup_tree_per_node *rtpn;
6918 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
6919 node_online(node) ? node : NUMA_NO_NODE);
6921 rtpn->rb_root = RB_ROOT;
6922 rtpn->rb_rightmost = NULL;
6923 spin_lock_init(&rtpn->lock);
6924 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6929 subsys_initcall(mem_cgroup_init);
6931 #ifdef CONFIG_MEMCG_SWAP
6932 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
6934 while (!refcount_inc_not_zero(&memcg->id.ref)) {
6936 * The root cgroup cannot be destroyed, so it's refcount must
6939 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
6943 memcg = parent_mem_cgroup(memcg);
6945 memcg = root_mem_cgroup;
6951 * mem_cgroup_swapout - transfer a memsw charge to swap
6952 * @page: page whose memsw charge to transfer
6953 * @entry: swap entry to move the charge to
6955 * Transfer the memsw charge of @page to @entry.
6957 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6959 struct mem_cgroup *memcg, *swap_memcg;
6960 unsigned int nr_entries;
6961 unsigned short oldid;
6963 VM_BUG_ON_PAGE(PageLRU(page), page);
6964 VM_BUG_ON_PAGE(page_count(page), page);
6966 if (!do_memsw_account())
6969 memcg = page->mem_cgroup;
6971 /* Readahead page, never charged */
6976 * In case the memcg owning these pages has been offlined and doesn't
6977 * have an ID allocated to it anymore, charge the closest online
6978 * ancestor for the swap instead and transfer the memory+swap charge.
6980 swap_memcg = mem_cgroup_id_get_online(memcg);
6981 nr_entries = hpage_nr_pages(page);
6982 /* Get references for the tail pages, too */
6984 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
6985 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
6987 VM_BUG_ON_PAGE(oldid, page);
6988 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
6990 page->mem_cgroup = NULL;
6992 if (!mem_cgroup_is_root(memcg))
6993 page_counter_uncharge(&memcg->memory, nr_entries);
6995 if (memcg != swap_memcg) {
6996 if (!mem_cgroup_is_root(swap_memcg))
6997 page_counter_charge(&swap_memcg->memsw, nr_entries);
6998 page_counter_uncharge(&memcg->memsw, nr_entries);
7002 * Interrupts should be disabled here because the caller holds the
7003 * i_pages lock which is taken with interrupts-off. It is
7004 * important here to have the interrupts disabled because it is the
7005 * only synchronisation we have for updating the per-CPU variables.
7007 VM_BUG_ON(!irqs_disabled());
7008 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
7010 memcg_check_events(memcg, page);
7012 if (!mem_cgroup_is_root(memcg))
7013 css_put_many(&memcg->css, nr_entries);
7017 * mem_cgroup_try_charge_swap - try charging swap space for a page
7018 * @page: page being added to swap
7019 * @entry: swap entry to charge
7021 * Try to charge @page's memcg for the swap space at @entry.
7023 * Returns 0 on success, -ENOMEM on failure.
7025 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7027 unsigned int nr_pages = hpage_nr_pages(page);
7028 struct page_counter *counter;
7029 struct mem_cgroup *memcg;
7030 unsigned short oldid;
7032 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
7035 memcg = page->mem_cgroup;
7037 /* Readahead page, never charged */
7042 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7046 memcg = mem_cgroup_id_get_online(memcg);
7048 if (!mem_cgroup_is_root(memcg) &&
7049 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7050 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7051 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7052 mem_cgroup_id_put(memcg);
7056 /* Get references for the tail pages, too */
7058 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7059 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7060 VM_BUG_ON_PAGE(oldid, page);
7061 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7067 * mem_cgroup_uncharge_swap - uncharge swap space
7068 * @entry: swap entry to uncharge
7069 * @nr_pages: the amount of swap space to uncharge
7071 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7073 struct mem_cgroup *memcg;
7076 if (!do_swap_account)
7079 id = swap_cgroup_record(entry, 0, nr_pages);
7081 memcg = mem_cgroup_from_id(id);
7083 if (!mem_cgroup_is_root(memcg)) {
7084 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7085 page_counter_uncharge(&memcg->swap, nr_pages);
7087 page_counter_uncharge(&memcg->memsw, nr_pages);
7089 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7090 mem_cgroup_id_put_many(memcg, nr_pages);
7095 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7097 long nr_swap_pages = get_nr_swap_pages();
7099 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7100 return nr_swap_pages;
7101 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7102 nr_swap_pages = min_t(long, nr_swap_pages,
7103 READ_ONCE(memcg->swap.max) -
7104 page_counter_read(&memcg->swap));
7105 return nr_swap_pages;
7108 bool mem_cgroup_swap_full(struct page *page)
7110 struct mem_cgroup *memcg;
7112 VM_BUG_ON_PAGE(!PageLocked(page), page);
7116 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7119 memcg = page->mem_cgroup;
7123 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7124 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.max)
7130 /* for remember boot option*/
7131 #ifdef CONFIG_MEMCG_SWAP_ENABLED
7132 static int really_do_swap_account __initdata = 1;
7134 static int really_do_swap_account __initdata;
7137 static int __init enable_swap_account(char *s)
7139 if (!strcmp(s, "1"))
7140 really_do_swap_account = 1;
7141 else if (!strcmp(s, "0"))
7142 really_do_swap_account = 0;
7145 __setup("swapaccount=", enable_swap_account);
7147 static u64 swap_current_read(struct cgroup_subsys_state *css,
7150 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7152 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7155 static int swap_max_show(struct seq_file *m, void *v)
7157 return seq_puts_memcg_tunable(m,
7158 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7161 static ssize_t swap_max_write(struct kernfs_open_file *of,
7162 char *buf, size_t nbytes, loff_t off)
7164 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7168 buf = strstrip(buf);
7169 err = page_counter_memparse(buf, "max", &max);
7173 xchg(&memcg->swap.max, max);
7178 static int swap_events_show(struct seq_file *m, void *v)
7180 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7182 seq_printf(m, "max %lu\n",
7183 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7184 seq_printf(m, "fail %lu\n",
7185 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7190 static struct cftype swap_files[] = {
7192 .name = "swap.current",
7193 .flags = CFTYPE_NOT_ON_ROOT,
7194 .read_u64 = swap_current_read,
7198 .flags = CFTYPE_NOT_ON_ROOT,
7199 .seq_show = swap_max_show,
7200 .write = swap_max_write,
7203 .name = "swap.events",
7204 .flags = CFTYPE_NOT_ON_ROOT,
7205 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7206 .seq_show = swap_events_show,
7211 static struct cftype memsw_cgroup_files[] = {
7213 .name = "memsw.usage_in_bytes",
7214 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7215 .read_u64 = mem_cgroup_read_u64,
7218 .name = "memsw.max_usage_in_bytes",
7219 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7220 .write = mem_cgroup_reset,
7221 .read_u64 = mem_cgroup_read_u64,
7224 .name = "memsw.limit_in_bytes",
7225 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7226 .write = mem_cgroup_write,
7227 .read_u64 = mem_cgroup_read_u64,
7230 .name = "memsw.failcnt",
7231 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7232 .write = mem_cgroup_reset,
7233 .read_u64 = mem_cgroup_read_u64,
7235 { }, /* terminate */
7238 static int __init mem_cgroup_swap_init(void)
7240 if (!mem_cgroup_disabled() && really_do_swap_account) {
7241 do_swap_account = 1;
7242 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
7244 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
7245 memsw_cgroup_files));
7249 subsys_initcall(mem_cgroup_swap_init);
7251 #endif /* CONFIG_MEMCG_SWAP */