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>
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/seq_buf.h>
66 #include <linux/uaccess.h>
68 #include <trace/events/vmscan.h>
70 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
71 EXPORT_SYMBOL(memory_cgrp_subsys);
73 struct mem_cgroup *root_mem_cgroup __read_mostly;
75 #define MEM_CGROUP_RECLAIM_RETRIES 5
77 /* Socket memory accounting disabled? */
78 static bool cgroup_memory_nosocket;
80 /* Kernel memory accounting disabled? */
81 static bool cgroup_memory_nokmem;
83 /* Whether the swap controller is active */
84 #ifdef CONFIG_MEMCG_SWAP
85 int do_swap_account __read_mostly;
87 #define do_swap_account 0
90 /* Whether legacy memory+swap accounting is active */
91 static bool do_memsw_account(void)
93 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && do_swap_account;
96 static const char *const mem_cgroup_lru_names[] = {
104 #define THRESHOLDS_EVENTS_TARGET 128
105 #define SOFTLIMIT_EVENTS_TARGET 1024
106 #define NUMAINFO_EVENTS_TARGET 1024
109 * Cgroups above their limits are maintained in a RB-Tree, independent of
110 * their hierarchy representation
113 struct mem_cgroup_tree_per_node {
114 struct rb_root rb_root;
115 struct rb_node *rb_rightmost;
119 struct mem_cgroup_tree {
120 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
123 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
126 struct mem_cgroup_eventfd_list {
127 struct list_head list;
128 struct eventfd_ctx *eventfd;
132 * cgroup_event represents events which userspace want to receive.
134 struct mem_cgroup_event {
136 * memcg which the event belongs to.
138 struct mem_cgroup *memcg;
140 * eventfd to signal userspace about the event.
142 struct eventfd_ctx *eventfd;
144 * Each of these stored in a list by the cgroup.
146 struct list_head list;
148 * register_event() callback will be used to add new userspace
149 * waiter for changes related to this event. Use eventfd_signal()
150 * on eventfd to send notification to userspace.
152 int (*register_event)(struct mem_cgroup *memcg,
153 struct eventfd_ctx *eventfd, const char *args);
155 * unregister_event() callback will be called when userspace closes
156 * the eventfd or on cgroup removing. This callback must be set,
157 * if you want provide notification functionality.
159 void (*unregister_event)(struct mem_cgroup *memcg,
160 struct eventfd_ctx *eventfd);
162 * All fields below needed to unregister event when
163 * userspace closes eventfd.
166 wait_queue_head_t *wqh;
167 wait_queue_entry_t wait;
168 struct work_struct remove;
171 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
172 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
174 /* Stuffs for move charges at task migration. */
176 * Types of charges to be moved.
178 #define MOVE_ANON 0x1U
179 #define MOVE_FILE 0x2U
180 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
182 /* "mc" and its members are protected by cgroup_mutex */
183 static struct move_charge_struct {
184 spinlock_t lock; /* for from, to */
185 struct mm_struct *mm;
186 struct mem_cgroup *from;
187 struct mem_cgroup *to;
189 unsigned long precharge;
190 unsigned long moved_charge;
191 unsigned long moved_swap;
192 struct task_struct *moving_task; /* a task moving charges */
193 wait_queue_head_t waitq; /* a waitq for other context */
195 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
196 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
200 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
201 * limit reclaim to prevent infinite loops, if they ever occur.
203 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
204 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
207 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
208 MEM_CGROUP_CHARGE_TYPE_ANON,
209 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
210 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
214 /* for encoding cft->private value on file */
223 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
224 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
225 #define MEMFILE_ATTR(val) ((val) & 0xffff)
226 /* Used for OOM nofiier */
227 #define OOM_CONTROL (0)
230 * Iteration constructs for visiting all cgroups (under a tree). If
231 * loops are exited prematurely (break), mem_cgroup_iter_break() must
232 * be used for reference counting.
234 #define for_each_mem_cgroup_tree(iter, root) \
235 for (iter = mem_cgroup_iter(root, NULL, NULL); \
237 iter = mem_cgroup_iter(root, iter, NULL))
239 #define for_each_mem_cgroup(iter) \
240 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
242 iter = mem_cgroup_iter(NULL, iter, NULL))
244 static inline bool should_force_charge(void)
246 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
247 (current->flags & PF_EXITING);
250 /* Some nice accessors for the vmpressure. */
251 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
254 memcg = root_mem_cgroup;
255 return &memcg->vmpressure;
258 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
260 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
263 #ifdef CONFIG_MEMCG_KMEM
265 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
266 * The main reason for not using cgroup id for this:
267 * this works better in sparse environments, where we have a lot of memcgs,
268 * but only a few kmem-limited. Or also, if we have, for instance, 200
269 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
270 * 200 entry array for that.
272 * The current size of the caches array is stored in memcg_nr_cache_ids. It
273 * will double each time we have to increase it.
275 static DEFINE_IDA(memcg_cache_ida);
276 int memcg_nr_cache_ids;
278 /* Protects memcg_nr_cache_ids */
279 static DECLARE_RWSEM(memcg_cache_ids_sem);
281 void memcg_get_cache_ids(void)
283 down_read(&memcg_cache_ids_sem);
286 void memcg_put_cache_ids(void)
288 up_read(&memcg_cache_ids_sem);
292 * MIN_SIZE is different than 1, because we would like to avoid going through
293 * the alloc/free process all the time. In a small machine, 4 kmem-limited
294 * cgroups is a reasonable guess. In the future, it could be a parameter or
295 * tunable, but that is strictly not necessary.
297 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
298 * this constant directly from cgroup, but it is understandable that this is
299 * better kept as an internal representation in cgroup.c. In any case, the
300 * cgrp_id space is not getting any smaller, and we don't have to necessarily
301 * increase ours as well if it increases.
303 #define MEMCG_CACHES_MIN_SIZE 4
304 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
307 * A lot of the calls to the cache allocation functions are expected to be
308 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
309 * conditional to this static branch, we'll have to allow modules that does
310 * kmem_cache_alloc and the such to see this symbol as well
312 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
313 EXPORT_SYMBOL(memcg_kmem_enabled_key);
315 struct workqueue_struct *memcg_kmem_cache_wq;
317 static int memcg_shrinker_map_size;
318 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
320 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
322 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
325 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
326 int size, int old_size)
328 struct memcg_shrinker_map *new, *old;
331 lockdep_assert_held(&memcg_shrinker_map_mutex);
334 old = rcu_dereference_protected(
335 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
336 /* Not yet online memcg */
340 new = kvmalloc(sizeof(*new) + size, GFP_KERNEL);
344 /* Set all old bits, clear all new bits */
345 memset(new->map, (int)0xff, old_size);
346 memset((void *)new->map + old_size, 0, size - old_size);
348 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
349 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
355 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
357 struct mem_cgroup_per_node *pn;
358 struct memcg_shrinker_map *map;
361 if (mem_cgroup_is_root(memcg))
365 pn = mem_cgroup_nodeinfo(memcg, nid);
366 map = rcu_dereference_protected(pn->shrinker_map, true);
369 rcu_assign_pointer(pn->shrinker_map, NULL);
373 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
375 struct memcg_shrinker_map *map;
376 int nid, size, ret = 0;
378 if (mem_cgroup_is_root(memcg))
381 mutex_lock(&memcg_shrinker_map_mutex);
382 size = memcg_shrinker_map_size;
384 map = kvzalloc(sizeof(*map) + size, GFP_KERNEL);
386 memcg_free_shrinker_maps(memcg);
390 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
392 mutex_unlock(&memcg_shrinker_map_mutex);
397 int memcg_expand_shrinker_maps(int new_id)
399 int size, old_size, ret = 0;
400 struct mem_cgroup *memcg;
402 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
403 old_size = memcg_shrinker_map_size;
404 if (size <= old_size)
407 mutex_lock(&memcg_shrinker_map_mutex);
408 if (!root_mem_cgroup)
411 for_each_mem_cgroup(memcg) {
412 if (mem_cgroup_is_root(memcg))
414 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
420 memcg_shrinker_map_size = size;
421 mutex_unlock(&memcg_shrinker_map_mutex);
425 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
427 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
428 struct memcg_shrinker_map *map;
431 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
432 /* Pairs with smp mb in shrink_slab() */
433 smp_mb__before_atomic();
434 set_bit(shrinker_id, map->map);
439 #else /* CONFIG_MEMCG_KMEM */
440 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
444 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg) { }
445 #endif /* CONFIG_MEMCG_KMEM */
448 * mem_cgroup_css_from_page - css of the memcg associated with a page
449 * @page: page of interest
451 * If memcg is bound to the default hierarchy, css of the memcg associated
452 * with @page is returned. The returned css remains associated with @page
453 * until it is released.
455 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
458 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
460 struct mem_cgroup *memcg;
462 memcg = page->mem_cgroup;
464 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
465 memcg = root_mem_cgroup;
471 * page_cgroup_ino - return inode number of the memcg a page is charged to
474 * Look up the closest online ancestor of the memory cgroup @page is charged to
475 * and return its inode number or 0 if @page is not charged to any cgroup. It
476 * is safe to call this function without holding a reference to @page.
478 * Note, this function is inherently racy, because there is nothing to prevent
479 * the cgroup inode from getting torn down and potentially reallocated a moment
480 * after page_cgroup_ino() returns, so it only should be used by callers that
481 * do not care (such as procfs interfaces).
483 ino_t page_cgroup_ino(struct page *page)
485 struct mem_cgroup *memcg;
486 unsigned long ino = 0;
489 if (PageSlab(page) && !PageTail(page))
490 memcg = memcg_from_slab_page(page);
492 memcg = READ_ONCE(page->mem_cgroup);
493 while (memcg && !(memcg->css.flags & CSS_ONLINE))
494 memcg = parent_mem_cgroup(memcg);
496 ino = cgroup_ino(memcg->css.cgroup);
501 static struct mem_cgroup_per_node *
502 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
504 int nid = page_to_nid(page);
506 return memcg->nodeinfo[nid];
509 static struct mem_cgroup_tree_per_node *
510 soft_limit_tree_node(int nid)
512 return soft_limit_tree.rb_tree_per_node[nid];
515 static struct mem_cgroup_tree_per_node *
516 soft_limit_tree_from_page(struct page *page)
518 int nid = page_to_nid(page);
520 return soft_limit_tree.rb_tree_per_node[nid];
523 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
524 struct mem_cgroup_tree_per_node *mctz,
525 unsigned long new_usage_in_excess)
527 struct rb_node **p = &mctz->rb_root.rb_node;
528 struct rb_node *parent = NULL;
529 struct mem_cgroup_per_node *mz_node;
530 bool rightmost = true;
535 mz->usage_in_excess = new_usage_in_excess;
536 if (!mz->usage_in_excess)
540 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
542 if (mz->usage_in_excess < mz_node->usage_in_excess) {
548 * We can't avoid mem cgroups that are over their soft
549 * limit by the same amount
551 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
556 mctz->rb_rightmost = &mz->tree_node;
558 rb_link_node(&mz->tree_node, parent, p);
559 rb_insert_color(&mz->tree_node, &mctz->rb_root);
563 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
564 struct mem_cgroup_tree_per_node *mctz)
569 if (&mz->tree_node == mctz->rb_rightmost)
570 mctz->rb_rightmost = rb_prev(&mz->tree_node);
572 rb_erase(&mz->tree_node, &mctz->rb_root);
576 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
577 struct mem_cgroup_tree_per_node *mctz)
581 spin_lock_irqsave(&mctz->lock, flags);
582 __mem_cgroup_remove_exceeded(mz, mctz);
583 spin_unlock_irqrestore(&mctz->lock, flags);
586 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
588 unsigned long nr_pages = page_counter_read(&memcg->memory);
589 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
590 unsigned long excess = 0;
592 if (nr_pages > soft_limit)
593 excess = nr_pages - soft_limit;
598 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
600 unsigned long excess;
601 struct mem_cgroup_per_node *mz;
602 struct mem_cgroup_tree_per_node *mctz;
604 mctz = soft_limit_tree_from_page(page);
608 * Necessary to update all ancestors when hierarchy is used.
609 * because their event counter is not touched.
611 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
612 mz = mem_cgroup_page_nodeinfo(memcg, page);
613 excess = soft_limit_excess(memcg);
615 * We have to update the tree if mz is on RB-tree or
616 * mem is over its softlimit.
618 if (excess || mz->on_tree) {
621 spin_lock_irqsave(&mctz->lock, flags);
622 /* if on-tree, remove it */
624 __mem_cgroup_remove_exceeded(mz, mctz);
626 * Insert again. mz->usage_in_excess will be updated.
627 * If excess is 0, no tree ops.
629 __mem_cgroup_insert_exceeded(mz, mctz, excess);
630 spin_unlock_irqrestore(&mctz->lock, flags);
635 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
637 struct mem_cgroup_tree_per_node *mctz;
638 struct mem_cgroup_per_node *mz;
642 mz = mem_cgroup_nodeinfo(memcg, nid);
643 mctz = soft_limit_tree_node(nid);
645 mem_cgroup_remove_exceeded(mz, mctz);
649 static struct mem_cgroup_per_node *
650 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
652 struct mem_cgroup_per_node *mz;
656 if (!mctz->rb_rightmost)
657 goto done; /* Nothing to reclaim from */
659 mz = rb_entry(mctz->rb_rightmost,
660 struct mem_cgroup_per_node, tree_node);
662 * Remove the node now but someone else can add it back,
663 * we will to add it back at the end of reclaim to its correct
664 * position in the tree.
666 __mem_cgroup_remove_exceeded(mz, mctz);
667 if (!soft_limit_excess(mz->memcg) ||
668 !css_tryget_online(&mz->memcg->css))
674 static struct mem_cgroup_per_node *
675 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
677 struct mem_cgroup_per_node *mz;
679 spin_lock_irq(&mctz->lock);
680 mz = __mem_cgroup_largest_soft_limit_node(mctz);
681 spin_unlock_irq(&mctz->lock);
686 * __mod_memcg_state - update cgroup memory statistics
687 * @memcg: the memory cgroup
688 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
689 * @val: delta to add to the counter, can be negative
691 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
695 if (mem_cgroup_disabled())
698 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
699 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
700 struct mem_cgroup *mi;
703 * Batch local counters to keep them in sync with
704 * the hierarchical ones.
706 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
707 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
708 atomic_long_add(x, &mi->vmstats[idx]);
711 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
714 static struct mem_cgroup_per_node *
715 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
717 struct mem_cgroup *parent;
719 parent = parent_mem_cgroup(pn->memcg);
722 return mem_cgroup_nodeinfo(parent, nid);
726 * __mod_lruvec_state - update lruvec memory statistics
727 * @lruvec: the lruvec
728 * @idx: the stat item
729 * @val: delta to add to the counter, can be negative
731 * The lruvec is the intersection of the NUMA node and a cgroup. This
732 * function updates the all three counters that are affected by a
733 * change of state at this level: per-node, per-cgroup, per-lruvec.
735 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
738 pg_data_t *pgdat = lruvec_pgdat(lruvec);
739 struct mem_cgroup_per_node *pn;
740 struct mem_cgroup *memcg;
744 __mod_node_page_state(pgdat, idx, val);
746 if (mem_cgroup_disabled())
749 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
753 __mod_memcg_state(memcg, idx, val);
756 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
758 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
759 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
760 struct mem_cgroup_per_node *pi;
762 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
763 atomic_long_add(x, &pi->lruvec_stat[idx]);
766 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
769 void __mod_lruvec_slab_state(void *p, enum node_stat_item idx, int val)
771 struct page *page = virt_to_head_page(p);
772 pg_data_t *pgdat = page_pgdat(page);
773 struct mem_cgroup *memcg;
774 struct lruvec *lruvec;
777 memcg = memcg_from_slab_page(page);
779 /* Untracked pages have no memcg, no lruvec. Update only the node */
780 if (!memcg || memcg == root_mem_cgroup) {
781 __mod_node_page_state(pgdat, idx, val);
783 lruvec = mem_cgroup_lruvec(pgdat, memcg);
784 __mod_lruvec_state(lruvec, idx, val);
790 * __count_memcg_events - account VM events in a cgroup
791 * @memcg: the memory cgroup
792 * @idx: the event item
793 * @count: the number of events that occured
795 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
800 if (mem_cgroup_disabled())
803 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
804 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
805 struct mem_cgroup *mi;
808 * Batch local counters to keep them in sync with
809 * the hierarchical ones.
811 __this_cpu_add(memcg->vmstats_local->events[idx], x);
812 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
813 atomic_long_add(x, &mi->vmevents[idx]);
816 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
819 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
821 return atomic_long_read(&memcg->vmevents[event]);
824 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
829 for_each_possible_cpu(cpu)
830 x += per_cpu(memcg->vmstats_local->events[event], cpu);
834 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
836 bool compound, int nr_pages)
839 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
840 * counted as CACHE even if it's on ANON LRU.
843 __mod_memcg_state(memcg, MEMCG_RSS, nr_pages);
845 __mod_memcg_state(memcg, MEMCG_CACHE, nr_pages);
846 if (PageSwapBacked(page))
847 __mod_memcg_state(memcg, NR_SHMEM, nr_pages);
851 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
852 __mod_memcg_state(memcg, MEMCG_RSS_HUGE, nr_pages);
855 /* pagein of a big page is an event. So, ignore page size */
857 __count_memcg_events(memcg, PGPGIN, 1);
859 __count_memcg_events(memcg, PGPGOUT, 1);
860 nr_pages = -nr_pages; /* for event */
863 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
866 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
867 enum mem_cgroup_events_target target)
869 unsigned long val, next;
871 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
872 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
873 /* from time_after() in jiffies.h */
874 if ((long)(next - val) < 0) {
876 case MEM_CGROUP_TARGET_THRESH:
877 next = val + THRESHOLDS_EVENTS_TARGET;
879 case MEM_CGROUP_TARGET_SOFTLIMIT:
880 next = val + SOFTLIMIT_EVENTS_TARGET;
882 case MEM_CGROUP_TARGET_NUMAINFO:
883 next = val + NUMAINFO_EVENTS_TARGET;
888 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
895 * Check events in order.
898 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
900 /* threshold event is triggered in finer grain than soft limit */
901 if (unlikely(mem_cgroup_event_ratelimit(memcg,
902 MEM_CGROUP_TARGET_THRESH))) {
904 bool do_numainfo __maybe_unused;
906 do_softlimit = mem_cgroup_event_ratelimit(memcg,
907 MEM_CGROUP_TARGET_SOFTLIMIT);
909 do_numainfo = mem_cgroup_event_ratelimit(memcg,
910 MEM_CGROUP_TARGET_NUMAINFO);
912 mem_cgroup_threshold(memcg);
913 if (unlikely(do_softlimit))
914 mem_cgroup_update_tree(memcg, page);
916 if (unlikely(do_numainfo))
917 atomic_inc(&memcg->numainfo_events);
922 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
925 * mm_update_next_owner() may clear mm->owner to NULL
926 * if it races with swapoff, page migration, etc.
927 * So this can be called with p == NULL.
932 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
934 EXPORT_SYMBOL(mem_cgroup_from_task);
937 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
938 * @mm: mm from which memcg should be extracted. It can be NULL.
940 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
941 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
944 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
946 struct mem_cgroup *memcg;
948 if (mem_cgroup_disabled())
954 * Page cache insertions can happen withou an
955 * actual mm context, e.g. during disk probing
956 * on boot, loopback IO, acct() writes etc.
959 memcg = root_mem_cgroup;
961 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
962 if (unlikely(!memcg))
963 memcg = root_mem_cgroup;
965 } while (!css_tryget(&memcg->css));
969 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
972 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
973 * @page: page from which memcg should be extracted.
975 * Obtain a reference on page->memcg and returns it if successful. Otherwise
976 * root_mem_cgroup is returned.
978 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
980 struct mem_cgroup *memcg = page->mem_cgroup;
982 if (mem_cgroup_disabled())
986 if (!memcg || !css_tryget_online(&memcg->css))
987 memcg = root_mem_cgroup;
991 EXPORT_SYMBOL(get_mem_cgroup_from_page);
994 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
996 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
998 if (unlikely(current->active_memcg)) {
999 struct mem_cgroup *memcg = root_mem_cgroup;
1002 if (css_tryget_online(¤t->active_memcg->css))
1003 memcg = current->active_memcg;
1007 return get_mem_cgroup_from_mm(current->mm);
1011 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1012 * @root: hierarchy root
1013 * @prev: previously returned memcg, NULL on first invocation
1014 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1016 * Returns references to children of the hierarchy below @root, or
1017 * @root itself, or %NULL after a full round-trip.
1019 * Caller must pass the return value in @prev on subsequent
1020 * invocations for reference counting, or use mem_cgroup_iter_break()
1021 * to cancel a hierarchy walk before the round-trip is complete.
1023 * Reclaimers can specify a node and a priority level in @reclaim to
1024 * divide up the memcgs in the hierarchy among all concurrent
1025 * reclaimers operating on the same node and priority.
1027 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1028 struct mem_cgroup *prev,
1029 struct mem_cgroup_reclaim_cookie *reclaim)
1031 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1032 struct cgroup_subsys_state *css = NULL;
1033 struct mem_cgroup *memcg = NULL;
1034 struct mem_cgroup *pos = NULL;
1036 if (mem_cgroup_disabled())
1040 root = root_mem_cgroup;
1042 if (prev && !reclaim)
1045 if (!root->use_hierarchy && root != root_mem_cgroup) {
1054 struct mem_cgroup_per_node *mz;
1056 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1057 iter = &mz->iter[reclaim->priority];
1059 if (prev && reclaim->generation != iter->generation)
1063 pos = READ_ONCE(iter->position);
1064 if (!pos || css_tryget(&pos->css))
1067 * css reference reached zero, so iter->position will
1068 * be cleared by ->css_released. However, we should not
1069 * rely on this happening soon, because ->css_released
1070 * is called from a work queue, and by busy-waiting we
1071 * might block it. So we clear iter->position right
1074 (void)cmpxchg(&iter->position, pos, NULL);
1082 css = css_next_descendant_pre(css, &root->css);
1085 * Reclaimers share the hierarchy walk, and a
1086 * new one might jump in right at the end of
1087 * the hierarchy - make sure they see at least
1088 * one group and restart from the beginning.
1096 * Verify the css and acquire a reference. The root
1097 * is provided by the caller, so we know it's alive
1098 * and kicking, and don't take an extra reference.
1100 memcg = mem_cgroup_from_css(css);
1102 if (css == &root->css)
1105 if (css_tryget(css))
1113 * The position could have already been updated by a competing
1114 * thread, so check that the value hasn't changed since we read
1115 * it to avoid reclaiming from the same cgroup twice.
1117 (void)cmpxchg(&iter->position, pos, memcg);
1125 reclaim->generation = iter->generation;
1131 if (prev && prev != root)
1132 css_put(&prev->css);
1138 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1139 * @root: hierarchy root
1140 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1142 void mem_cgroup_iter_break(struct mem_cgroup *root,
1143 struct mem_cgroup *prev)
1146 root = root_mem_cgroup;
1147 if (prev && prev != root)
1148 css_put(&prev->css);
1151 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1152 struct mem_cgroup *dead_memcg)
1154 struct mem_cgroup_reclaim_iter *iter;
1155 struct mem_cgroup_per_node *mz;
1159 for_each_node(nid) {
1160 mz = mem_cgroup_nodeinfo(from, nid);
1161 for (i = 0; i <= DEF_PRIORITY; i++) {
1162 iter = &mz->iter[i];
1163 cmpxchg(&iter->position,
1169 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1171 struct mem_cgroup *memcg = dead_memcg;
1172 struct mem_cgroup *last;
1175 __invalidate_reclaim_iterators(memcg, dead_memcg);
1177 } while ((memcg = parent_mem_cgroup(memcg)));
1180 * When cgruop1 non-hierarchy mode is used,
1181 * parent_mem_cgroup() does not walk all the way up to the
1182 * cgroup root (root_mem_cgroup). So we have to handle
1183 * dead_memcg from cgroup root separately.
1185 if (last != root_mem_cgroup)
1186 __invalidate_reclaim_iterators(root_mem_cgroup,
1191 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1192 * @memcg: hierarchy root
1193 * @fn: function to call for each task
1194 * @arg: argument passed to @fn
1196 * This function iterates over tasks attached to @memcg or to any of its
1197 * descendants and calls @fn for each task. If @fn returns a non-zero
1198 * value, the function breaks the iteration loop and returns the value.
1199 * Otherwise, it will iterate over all tasks and return 0.
1201 * This function must not be called for the root memory cgroup.
1203 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1204 int (*fn)(struct task_struct *, void *), void *arg)
1206 struct mem_cgroup *iter;
1209 BUG_ON(memcg == root_mem_cgroup);
1211 for_each_mem_cgroup_tree(iter, memcg) {
1212 struct css_task_iter it;
1213 struct task_struct *task;
1215 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1216 while (!ret && (task = css_task_iter_next(&it)))
1217 ret = fn(task, arg);
1218 css_task_iter_end(&it);
1220 mem_cgroup_iter_break(memcg, iter);
1228 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1230 * @pgdat: pgdat of the page
1232 * This function is only safe when following the LRU page isolation
1233 * and putback protocol: the LRU lock must be held, and the page must
1234 * either be PageLRU() or the caller must have isolated/allocated it.
1236 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1238 struct mem_cgroup_per_node *mz;
1239 struct mem_cgroup *memcg;
1240 struct lruvec *lruvec;
1242 if (mem_cgroup_disabled()) {
1243 lruvec = &pgdat->lruvec;
1247 memcg = page->mem_cgroup;
1249 * Swapcache readahead pages are added to the LRU - and
1250 * possibly migrated - before they are charged.
1253 memcg = root_mem_cgroup;
1255 mz = mem_cgroup_page_nodeinfo(memcg, page);
1256 lruvec = &mz->lruvec;
1259 * Since a node can be onlined after the mem_cgroup was created,
1260 * we have to be prepared to initialize lruvec->zone here;
1261 * and if offlined then reonlined, we need to reinitialize it.
1263 if (unlikely(lruvec->pgdat != pgdat))
1264 lruvec->pgdat = pgdat;
1269 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1270 * @lruvec: mem_cgroup per zone lru vector
1271 * @lru: index of lru list the page is sitting on
1272 * @zid: zone id of the accounted pages
1273 * @nr_pages: positive when adding or negative when removing
1275 * This function must be called under lru_lock, just before a page is added
1276 * to or just after a page is removed from an lru list (that ordering being
1277 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1279 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1280 int zid, int nr_pages)
1282 struct mem_cgroup_per_node *mz;
1283 unsigned long *lru_size;
1286 if (mem_cgroup_disabled())
1289 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1290 lru_size = &mz->lru_zone_size[zid][lru];
1293 *lru_size += nr_pages;
1296 if (WARN_ONCE(size < 0,
1297 "%s(%p, %d, %d): lru_size %ld\n",
1298 __func__, lruvec, lru, nr_pages, size)) {
1304 *lru_size += nr_pages;
1308 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1309 * @memcg: the memory cgroup
1311 * Returns the maximum amount of memory @mem can be charged with, in
1314 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1316 unsigned long margin = 0;
1317 unsigned long count;
1318 unsigned long limit;
1320 count = page_counter_read(&memcg->memory);
1321 limit = READ_ONCE(memcg->memory.max);
1323 margin = limit - count;
1325 if (do_memsw_account()) {
1326 count = page_counter_read(&memcg->memsw);
1327 limit = READ_ONCE(memcg->memsw.max);
1329 margin = min(margin, limit - count);
1338 * A routine for checking "mem" is under move_account() or not.
1340 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1341 * moving cgroups. This is for waiting at high-memory pressure
1344 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1346 struct mem_cgroup *from;
1347 struct mem_cgroup *to;
1350 * Unlike task_move routines, we access mc.to, mc.from not under
1351 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1353 spin_lock(&mc.lock);
1359 ret = mem_cgroup_is_descendant(from, memcg) ||
1360 mem_cgroup_is_descendant(to, memcg);
1362 spin_unlock(&mc.lock);
1366 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1368 if (mc.moving_task && current != mc.moving_task) {
1369 if (mem_cgroup_under_move(memcg)) {
1371 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1372 /* moving charge context might have finished. */
1375 finish_wait(&mc.waitq, &wait);
1382 static char *memory_stat_format(struct mem_cgroup *memcg)
1387 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1392 * Provide statistics on the state of the memory subsystem as
1393 * well as cumulative event counters that show past behavior.
1395 * This list is ordered following a combination of these gradients:
1396 * 1) generic big picture -> specifics and details
1397 * 2) reflecting userspace activity -> reflecting kernel heuristics
1399 * Current memory state:
1402 seq_buf_printf(&s, "anon %llu\n",
1403 (u64)memcg_page_state(memcg, MEMCG_RSS) *
1405 seq_buf_printf(&s, "file %llu\n",
1406 (u64)memcg_page_state(memcg, MEMCG_CACHE) *
1408 seq_buf_printf(&s, "kernel_stack %llu\n",
1409 (u64)memcg_page_state(memcg, MEMCG_KERNEL_STACK_KB) *
1411 seq_buf_printf(&s, "slab %llu\n",
1412 (u64)(memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) +
1413 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE)) *
1415 seq_buf_printf(&s, "sock %llu\n",
1416 (u64)memcg_page_state(memcg, MEMCG_SOCK) *
1419 seq_buf_printf(&s, "shmem %llu\n",
1420 (u64)memcg_page_state(memcg, NR_SHMEM) *
1422 seq_buf_printf(&s, "file_mapped %llu\n",
1423 (u64)memcg_page_state(memcg, NR_FILE_MAPPED) *
1425 seq_buf_printf(&s, "file_dirty %llu\n",
1426 (u64)memcg_page_state(memcg, NR_FILE_DIRTY) *
1428 seq_buf_printf(&s, "file_writeback %llu\n",
1429 (u64)memcg_page_state(memcg, NR_WRITEBACK) *
1433 * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter
1434 * with the NR_ANON_THP vm counter, but right now it's a pain in the
1435 * arse because it requires migrating the work out of rmap to a place
1436 * where the page->mem_cgroup is set up and stable.
1438 seq_buf_printf(&s, "anon_thp %llu\n",
1439 (u64)memcg_page_state(memcg, MEMCG_RSS_HUGE) *
1442 for (i = 0; i < NR_LRU_LISTS; i++)
1443 seq_buf_printf(&s, "%s %llu\n", mem_cgroup_lru_names[i],
1444 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
1447 seq_buf_printf(&s, "slab_reclaimable %llu\n",
1448 (u64)memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) *
1450 seq_buf_printf(&s, "slab_unreclaimable %llu\n",
1451 (u64)memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE) *
1454 /* Accumulated memory events */
1456 seq_buf_printf(&s, "pgfault %lu\n", memcg_events(memcg, PGFAULT));
1457 seq_buf_printf(&s, "pgmajfault %lu\n", memcg_events(memcg, PGMAJFAULT));
1459 seq_buf_printf(&s, "workingset_refault %lu\n",
1460 memcg_page_state(memcg, WORKINGSET_REFAULT));
1461 seq_buf_printf(&s, "workingset_activate %lu\n",
1462 memcg_page_state(memcg, WORKINGSET_ACTIVATE));
1463 seq_buf_printf(&s, "workingset_nodereclaim %lu\n",
1464 memcg_page_state(memcg, WORKINGSET_NODERECLAIM));
1466 seq_buf_printf(&s, "pgrefill %lu\n", memcg_events(memcg, PGREFILL));
1467 seq_buf_printf(&s, "pgscan %lu\n",
1468 memcg_events(memcg, PGSCAN_KSWAPD) +
1469 memcg_events(memcg, PGSCAN_DIRECT));
1470 seq_buf_printf(&s, "pgsteal %lu\n",
1471 memcg_events(memcg, PGSTEAL_KSWAPD) +
1472 memcg_events(memcg, PGSTEAL_DIRECT));
1473 seq_buf_printf(&s, "pgactivate %lu\n", memcg_events(memcg, PGACTIVATE));
1474 seq_buf_printf(&s, "pgdeactivate %lu\n", memcg_events(memcg, PGDEACTIVATE));
1475 seq_buf_printf(&s, "pglazyfree %lu\n", memcg_events(memcg, PGLAZYFREE));
1476 seq_buf_printf(&s, "pglazyfreed %lu\n", memcg_events(memcg, PGLAZYFREED));
1478 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1479 seq_buf_printf(&s, "thp_fault_alloc %lu\n",
1480 memcg_events(memcg, THP_FAULT_ALLOC));
1481 seq_buf_printf(&s, "thp_collapse_alloc %lu\n",
1482 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1483 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1485 /* The above should easily fit into one page */
1486 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1491 #define K(x) ((x) << (PAGE_SHIFT-10))
1493 * mem_cgroup_print_oom_context: Print OOM information relevant to
1494 * memory controller.
1495 * @memcg: The memory cgroup that went over limit
1496 * @p: Task that is going to be killed
1498 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1501 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1506 pr_cont(",oom_memcg=");
1507 pr_cont_cgroup_path(memcg->css.cgroup);
1509 pr_cont(",global_oom");
1511 pr_cont(",task_memcg=");
1512 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1518 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1519 * memory controller.
1520 * @memcg: The memory cgroup that went over limit
1522 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1526 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1527 K((u64)page_counter_read(&memcg->memory)),
1528 K((u64)memcg->memory.max), memcg->memory.failcnt);
1529 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1530 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1531 K((u64)page_counter_read(&memcg->swap)),
1532 K((u64)memcg->swap.max), memcg->swap.failcnt);
1534 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1535 K((u64)page_counter_read(&memcg->memsw)),
1536 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1537 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1538 K((u64)page_counter_read(&memcg->kmem)),
1539 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1542 pr_info("Memory cgroup stats for ");
1543 pr_cont_cgroup_path(memcg->css.cgroup);
1545 buf = memory_stat_format(memcg);
1553 * Return the memory (and swap, if configured) limit for a memcg.
1555 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1559 max = memcg->memory.max;
1560 if (mem_cgroup_swappiness(memcg)) {
1561 unsigned long memsw_max;
1562 unsigned long swap_max;
1564 memsw_max = memcg->memsw.max;
1565 swap_max = memcg->swap.max;
1566 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1567 max = min(max + swap_max, memsw_max);
1572 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1575 struct oom_control oc = {
1579 .gfp_mask = gfp_mask,
1584 if (mutex_lock_killable(&oom_lock))
1587 * A few threads which were not waiting at mutex_lock_killable() can
1588 * fail to bail out. Therefore, check again after holding oom_lock.
1590 ret = should_force_charge() || out_of_memory(&oc);
1591 mutex_unlock(&oom_lock);
1595 #if MAX_NUMNODES > 1
1598 * test_mem_cgroup_node_reclaimable
1599 * @memcg: the target memcg
1600 * @nid: the node ID to be checked.
1601 * @noswap : specify true here if the user wants flle only information.
1603 * This function returns whether the specified memcg contains any
1604 * reclaimable pages on a node. Returns true if there are any reclaimable
1605 * pages in the node.
1607 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1608 int nid, bool noswap)
1610 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
1612 if (lruvec_page_state(lruvec, NR_INACTIVE_FILE) ||
1613 lruvec_page_state(lruvec, NR_ACTIVE_FILE))
1615 if (noswap || !total_swap_pages)
1617 if (lruvec_page_state(lruvec, NR_INACTIVE_ANON) ||
1618 lruvec_page_state(lruvec, NR_ACTIVE_ANON))
1625 * Always updating the nodemask is not very good - even if we have an empty
1626 * list or the wrong list here, we can start from some node and traverse all
1627 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1630 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1634 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1635 * pagein/pageout changes since the last update.
1637 if (!atomic_read(&memcg->numainfo_events))
1639 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1642 /* make a nodemask where this memcg uses memory from */
1643 memcg->scan_nodes = node_states[N_MEMORY];
1645 for_each_node_mask(nid, node_states[N_MEMORY]) {
1647 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1648 node_clear(nid, memcg->scan_nodes);
1651 atomic_set(&memcg->numainfo_events, 0);
1652 atomic_set(&memcg->numainfo_updating, 0);
1656 * Selecting a node where we start reclaim from. Because what we need is just
1657 * reducing usage counter, start from anywhere is O,K. Considering
1658 * memory reclaim from current node, there are pros. and cons.
1660 * Freeing memory from current node means freeing memory from a node which
1661 * we'll use or we've used. So, it may make LRU bad. And if several threads
1662 * hit limits, it will see a contention on a node. But freeing from remote
1663 * node means more costs for memory reclaim because of memory latency.
1665 * Now, we use round-robin. Better algorithm is welcomed.
1667 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1671 mem_cgroup_may_update_nodemask(memcg);
1672 node = memcg->last_scanned_node;
1674 node = next_node_in(node, memcg->scan_nodes);
1676 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1677 * last time it really checked all the LRUs due to rate limiting.
1678 * Fallback to the current node in that case for simplicity.
1680 if (unlikely(node == MAX_NUMNODES))
1681 node = numa_node_id();
1683 memcg->last_scanned_node = node;
1687 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1693 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1696 unsigned long *total_scanned)
1698 struct mem_cgroup *victim = NULL;
1701 unsigned long excess;
1702 unsigned long nr_scanned;
1703 struct mem_cgroup_reclaim_cookie reclaim = {
1708 excess = soft_limit_excess(root_memcg);
1711 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1716 * If we have not been able to reclaim
1717 * anything, it might because there are
1718 * no reclaimable pages under this hierarchy
1723 * We want to do more targeted reclaim.
1724 * excess >> 2 is not to excessive so as to
1725 * reclaim too much, nor too less that we keep
1726 * coming back to reclaim from this cgroup
1728 if (total >= (excess >> 2) ||
1729 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1734 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1735 pgdat, &nr_scanned);
1736 *total_scanned += nr_scanned;
1737 if (!soft_limit_excess(root_memcg))
1740 mem_cgroup_iter_break(root_memcg, victim);
1744 #ifdef CONFIG_LOCKDEP
1745 static struct lockdep_map memcg_oom_lock_dep_map = {
1746 .name = "memcg_oom_lock",
1750 static DEFINE_SPINLOCK(memcg_oom_lock);
1753 * Check OOM-Killer is already running under our hierarchy.
1754 * If someone is running, return false.
1756 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1758 struct mem_cgroup *iter, *failed = NULL;
1760 spin_lock(&memcg_oom_lock);
1762 for_each_mem_cgroup_tree(iter, memcg) {
1763 if (iter->oom_lock) {
1765 * this subtree of our hierarchy is already locked
1766 * so we cannot give a lock.
1769 mem_cgroup_iter_break(memcg, iter);
1772 iter->oom_lock = true;
1777 * OK, we failed to lock the whole subtree so we have
1778 * to clean up what we set up to the failing subtree
1780 for_each_mem_cgroup_tree(iter, memcg) {
1781 if (iter == failed) {
1782 mem_cgroup_iter_break(memcg, iter);
1785 iter->oom_lock = false;
1788 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1790 spin_unlock(&memcg_oom_lock);
1795 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1797 struct mem_cgroup *iter;
1799 spin_lock(&memcg_oom_lock);
1800 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1801 for_each_mem_cgroup_tree(iter, memcg)
1802 iter->oom_lock = false;
1803 spin_unlock(&memcg_oom_lock);
1806 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1808 struct mem_cgroup *iter;
1810 spin_lock(&memcg_oom_lock);
1811 for_each_mem_cgroup_tree(iter, memcg)
1813 spin_unlock(&memcg_oom_lock);
1816 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1818 struct mem_cgroup *iter;
1821 * When a new child is created while the hierarchy is under oom,
1822 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1824 spin_lock(&memcg_oom_lock);
1825 for_each_mem_cgroup_tree(iter, memcg)
1826 if (iter->under_oom > 0)
1828 spin_unlock(&memcg_oom_lock);
1831 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1833 struct oom_wait_info {
1834 struct mem_cgroup *memcg;
1835 wait_queue_entry_t wait;
1838 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1839 unsigned mode, int sync, void *arg)
1841 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1842 struct mem_cgroup *oom_wait_memcg;
1843 struct oom_wait_info *oom_wait_info;
1845 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1846 oom_wait_memcg = oom_wait_info->memcg;
1848 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1849 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1851 return autoremove_wake_function(wait, mode, sync, arg);
1854 static void memcg_oom_recover(struct mem_cgroup *memcg)
1857 * For the following lockless ->under_oom test, the only required
1858 * guarantee is that it must see the state asserted by an OOM when
1859 * this function is called as a result of userland actions
1860 * triggered by the notification of the OOM. This is trivially
1861 * achieved by invoking mem_cgroup_mark_under_oom() before
1862 * triggering notification.
1864 if (memcg && memcg->under_oom)
1865 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1875 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1877 enum oom_status ret;
1880 if (order > PAGE_ALLOC_COSTLY_ORDER)
1883 memcg_memory_event(memcg, MEMCG_OOM);
1886 * We are in the middle of the charge context here, so we
1887 * don't want to block when potentially sitting on a callstack
1888 * that holds all kinds of filesystem and mm locks.
1890 * cgroup1 allows disabling the OOM killer and waiting for outside
1891 * handling until the charge can succeed; remember the context and put
1892 * the task to sleep at the end of the page fault when all locks are
1895 * On the other hand, in-kernel OOM killer allows for an async victim
1896 * memory reclaim (oom_reaper) and that means that we are not solely
1897 * relying on the oom victim to make a forward progress and we can
1898 * invoke the oom killer here.
1900 * Please note that mem_cgroup_out_of_memory might fail to find a
1901 * victim and then we have to bail out from the charge path.
1903 if (memcg->oom_kill_disable) {
1904 if (!current->in_user_fault)
1906 css_get(&memcg->css);
1907 current->memcg_in_oom = memcg;
1908 current->memcg_oom_gfp_mask = mask;
1909 current->memcg_oom_order = order;
1914 mem_cgroup_mark_under_oom(memcg);
1916 locked = mem_cgroup_oom_trylock(memcg);
1919 mem_cgroup_oom_notify(memcg);
1921 mem_cgroup_unmark_under_oom(memcg);
1922 if (mem_cgroup_out_of_memory(memcg, mask, order))
1928 mem_cgroup_oom_unlock(memcg);
1934 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1935 * @handle: actually kill/wait or just clean up the OOM state
1937 * This has to be called at the end of a page fault if the memcg OOM
1938 * handler was enabled.
1940 * Memcg supports userspace OOM handling where failed allocations must
1941 * sleep on a waitqueue until the userspace task resolves the
1942 * situation. Sleeping directly in the charge context with all kinds
1943 * of locks held is not a good idea, instead we remember an OOM state
1944 * in the task and mem_cgroup_oom_synchronize() has to be called at
1945 * the end of the page fault to complete the OOM handling.
1947 * Returns %true if an ongoing memcg OOM situation was detected and
1948 * completed, %false otherwise.
1950 bool mem_cgroup_oom_synchronize(bool handle)
1952 struct mem_cgroup *memcg = current->memcg_in_oom;
1953 struct oom_wait_info owait;
1956 /* OOM is global, do not handle */
1963 owait.memcg = memcg;
1964 owait.wait.flags = 0;
1965 owait.wait.func = memcg_oom_wake_function;
1966 owait.wait.private = current;
1967 INIT_LIST_HEAD(&owait.wait.entry);
1969 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1970 mem_cgroup_mark_under_oom(memcg);
1972 locked = mem_cgroup_oom_trylock(memcg);
1975 mem_cgroup_oom_notify(memcg);
1977 if (locked && !memcg->oom_kill_disable) {
1978 mem_cgroup_unmark_under_oom(memcg);
1979 finish_wait(&memcg_oom_waitq, &owait.wait);
1980 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1981 current->memcg_oom_order);
1984 mem_cgroup_unmark_under_oom(memcg);
1985 finish_wait(&memcg_oom_waitq, &owait.wait);
1989 mem_cgroup_oom_unlock(memcg);
1991 * There is no guarantee that an OOM-lock contender
1992 * sees the wakeups triggered by the OOM kill
1993 * uncharges. Wake any sleepers explicitely.
1995 memcg_oom_recover(memcg);
1998 current->memcg_in_oom = NULL;
1999 css_put(&memcg->css);
2004 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2005 * @victim: task to be killed by the OOM killer
2006 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2008 * Returns a pointer to a memory cgroup, which has to be cleaned up
2009 * by killing all belonging OOM-killable tasks.
2011 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2013 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2014 struct mem_cgroup *oom_domain)
2016 struct mem_cgroup *oom_group = NULL;
2017 struct mem_cgroup *memcg;
2019 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2023 oom_domain = root_mem_cgroup;
2027 memcg = mem_cgroup_from_task(victim);
2028 if (memcg == root_mem_cgroup)
2032 * Traverse the memory cgroup hierarchy from the victim task's
2033 * cgroup up to the OOMing cgroup (or root) to find the
2034 * highest-level memory cgroup with oom.group set.
2036 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2037 if (memcg->oom_group)
2040 if (memcg == oom_domain)
2045 css_get(&oom_group->css);
2052 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2054 pr_info("Tasks in ");
2055 pr_cont_cgroup_path(memcg->css.cgroup);
2056 pr_cont(" are going to be killed due to memory.oom.group set\n");
2060 * lock_page_memcg - lock a page->mem_cgroup binding
2063 * This function protects unlocked LRU pages from being moved to
2066 * It ensures lifetime of the returned memcg. Caller is responsible
2067 * for the lifetime of the page; __unlock_page_memcg() is available
2068 * when @page might get freed inside the locked section.
2070 struct mem_cgroup *lock_page_memcg(struct page *page)
2072 struct mem_cgroup *memcg;
2073 unsigned long flags;
2076 * The RCU lock is held throughout the transaction. The fast
2077 * path can get away without acquiring the memcg->move_lock
2078 * because page moving starts with an RCU grace period.
2080 * The RCU lock also protects the memcg from being freed when
2081 * the page state that is going to change is the only thing
2082 * preventing the page itself from being freed. E.g. writeback
2083 * doesn't hold a page reference and relies on PG_writeback to
2084 * keep off truncation, migration and so forth.
2088 if (mem_cgroup_disabled())
2091 memcg = page->mem_cgroup;
2092 if (unlikely(!memcg))
2095 if (atomic_read(&memcg->moving_account) <= 0)
2098 spin_lock_irqsave(&memcg->move_lock, flags);
2099 if (memcg != page->mem_cgroup) {
2100 spin_unlock_irqrestore(&memcg->move_lock, flags);
2105 * When charge migration first begins, we can have locked and
2106 * unlocked page stat updates happening concurrently. Track
2107 * the task who has the lock for unlock_page_memcg().
2109 memcg->move_lock_task = current;
2110 memcg->move_lock_flags = flags;
2114 EXPORT_SYMBOL(lock_page_memcg);
2117 * __unlock_page_memcg - unlock and unpin a memcg
2120 * Unlock and unpin a memcg returned by lock_page_memcg().
2122 void __unlock_page_memcg(struct mem_cgroup *memcg)
2124 if (memcg && memcg->move_lock_task == current) {
2125 unsigned long flags = memcg->move_lock_flags;
2127 memcg->move_lock_task = NULL;
2128 memcg->move_lock_flags = 0;
2130 spin_unlock_irqrestore(&memcg->move_lock, flags);
2137 * unlock_page_memcg - unlock a page->mem_cgroup binding
2140 void unlock_page_memcg(struct page *page)
2142 __unlock_page_memcg(page->mem_cgroup);
2144 EXPORT_SYMBOL(unlock_page_memcg);
2146 struct memcg_stock_pcp {
2147 struct mem_cgroup *cached; /* this never be root cgroup */
2148 unsigned int nr_pages;
2149 struct work_struct work;
2150 unsigned long flags;
2151 #define FLUSHING_CACHED_CHARGE 0
2153 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2154 static DEFINE_MUTEX(percpu_charge_mutex);
2157 * consume_stock: Try to consume stocked charge on this cpu.
2158 * @memcg: memcg to consume from.
2159 * @nr_pages: how many pages to charge.
2161 * The charges will only happen if @memcg matches the current cpu's memcg
2162 * stock, and at least @nr_pages are available in that stock. Failure to
2163 * service an allocation will refill the stock.
2165 * returns true if successful, false otherwise.
2167 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2169 struct memcg_stock_pcp *stock;
2170 unsigned long flags;
2173 if (nr_pages > MEMCG_CHARGE_BATCH)
2176 local_irq_save(flags);
2178 stock = this_cpu_ptr(&memcg_stock);
2179 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2180 stock->nr_pages -= nr_pages;
2184 local_irq_restore(flags);
2190 * Returns stocks cached in percpu and reset cached information.
2192 static void drain_stock(struct memcg_stock_pcp *stock)
2194 struct mem_cgroup *old = stock->cached;
2196 if (stock->nr_pages) {
2197 page_counter_uncharge(&old->memory, stock->nr_pages);
2198 if (do_memsw_account())
2199 page_counter_uncharge(&old->memsw, stock->nr_pages);
2200 css_put_many(&old->css, stock->nr_pages);
2201 stock->nr_pages = 0;
2203 stock->cached = NULL;
2206 static void drain_local_stock(struct work_struct *dummy)
2208 struct memcg_stock_pcp *stock;
2209 unsigned long flags;
2212 * The only protection from memory hotplug vs. drain_stock races is
2213 * that we always operate on local CPU stock here with IRQ disabled
2215 local_irq_save(flags);
2217 stock = this_cpu_ptr(&memcg_stock);
2219 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2221 local_irq_restore(flags);
2225 * Cache charges(val) to local per_cpu area.
2226 * This will be consumed by consume_stock() function, later.
2228 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2230 struct memcg_stock_pcp *stock;
2231 unsigned long flags;
2233 local_irq_save(flags);
2235 stock = this_cpu_ptr(&memcg_stock);
2236 if (stock->cached != memcg) { /* reset if necessary */
2238 stock->cached = memcg;
2240 stock->nr_pages += nr_pages;
2242 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2245 local_irq_restore(flags);
2249 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2250 * of the hierarchy under it.
2252 static void drain_all_stock(struct mem_cgroup *root_memcg)
2256 /* If someone's already draining, avoid adding running more workers. */
2257 if (!mutex_trylock(&percpu_charge_mutex))
2260 * Notify other cpus that system-wide "drain" is running
2261 * We do not care about races with the cpu hotplug because cpu down
2262 * as well as workers from this path always operate on the local
2263 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2266 for_each_online_cpu(cpu) {
2267 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2268 struct mem_cgroup *memcg;
2270 memcg = stock->cached;
2271 if (!memcg || !stock->nr_pages || !css_tryget(&memcg->css))
2273 if (!mem_cgroup_is_descendant(memcg, root_memcg)) {
2274 css_put(&memcg->css);
2277 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2279 drain_local_stock(&stock->work);
2281 schedule_work_on(cpu, &stock->work);
2283 css_put(&memcg->css);
2286 mutex_unlock(&percpu_charge_mutex);
2289 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2291 struct memcg_stock_pcp *stock;
2292 struct mem_cgroup *memcg, *mi;
2294 stock = &per_cpu(memcg_stock, cpu);
2297 for_each_mem_cgroup(memcg) {
2300 for (i = 0; i < MEMCG_NR_STAT; i++) {
2304 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2306 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2307 atomic_long_add(x, &memcg->vmstats[i]);
2309 if (i >= NR_VM_NODE_STAT_ITEMS)
2312 for_each_node(nid) {
2313 struct mem_cgroup_per_node *pn;
2315 pn = mem_cgroup_nodeinfo(memcg, nid);
2316 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2319 atomic_long_add(x, &pn->lruvec_stat[i]);
2320 } while ((pn = parent_nodeinfo(pn, nid)));
2324 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2327 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2329 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2330 atomic_long_add(x, &memcg->vmevents[i]);
2337 static void reclaim_high(struct mem_cgroup *memcg,
2338 unsigned int nr_pages,
2342 if (page_counter_read(&memcg->memory) <= memcg->high)
2344 memcg_memory_event(memcg, MEMCG_HIGH);
2345 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2346 } while ((memcg = parent_mem_cgroup(memcg)));
2349 static void high_work_func(struct work_struct *work)
2351 struct mem_cgroup *memcg;
2353 memcg = container_of(work, struct mem_cgroup, high_work);
2354 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2358 * Scheduled by try_charge() to be executed from the userland return path
2359 * and reclaims memory over the high limit.
2361 void mem_cgroup_handle_over_high(void)
2363 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2364 struct mem_cgroup *memcg;
2366 if (likely(!nr_pages))
2369 memcg = get_mem_cgroup_from_mm(current->mm);
2370 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2371 css_put(&memcg->css);
2372 current->memcg_nr_pages_over_high = 0;
2375 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2376 unsigned int nr_pages)
2378 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2379 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2380 struct mem_cgroup *mem_over_limit;
2381 struct page_counter *counter;
2382 unsigned long nr_reclaimed;
2383 bool may_swap = true;
2384 bool drained = false;
2385 enum oom_status oom_status;
2387 if (mem_cgroup_is_root(memcg))
2390 if (consume_stock(memcg, nr_pages))
2393 if (!do_memsw_account() ||
2394 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2395 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2397 if (do_memsw_account())
2398 page_counter_uncharge(&memcg->memsw, batch);
2399 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2401 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2405 if (batch > nr_pages) {
2411 * Memcg doesn't have a dedicated reserve for atomic
2412 * allocations. But like the global atomic pool, we need to
2413 * put the burden of reclaim on regular allocation requests
2414 * and let these go through as privileged allocations.
2416 if (gfp_mask & __GFP_ATOMIC)
2420 * Unlike in global OOM situations, memcg is not in a physical
2421 * memory shortage. Allow dying and OOM-killed tasks to
2422 * bypass the last charges so that they can exit quickly and
2423 * free their memory.
2425 if (unlikely(should_force_charge()))
2429 * Prevent unbounded recursion when reclaim operations need to
2430 * allocate memory. This might exceed the limits temporarily,
2431 * but we prefer facilitating memory reclaim and getting back
2432 * under the limit over triggering OOM kills in these cases.
2434 if (unlikely(current->flags & PF_MEMALLOC))
2437 if (unlikely(task_in_memcg_oom(current)))
2440 if (!gfpflags_allow_blocking(gfp_mask))
2443 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2445 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2446 gfp_mask, may_swap);
2448 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2452 drain_all_stock(mem_over_limit);
2457 if (gfp_mask & __GFP_NORETRY)
2460 * Even though the limit is exceeded at this point, reclaim
2461 * may have been able to free some pages. Retry the charge
2462 * before killing the task.
2464 * Only for regular pages, though: huge pages are rather
2465 * unlikely to succeed so close to the limit, and we fall back
2466 * to regular pages anyway in case of failure.
2468 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2471 * At task move, charge accounts can be doubly counted. So, it's
2472 * better to wait until the end of task_move if something is going on.
2474 if (mem_cgroup_wait_acct_move(mem_over_limit))
2480 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2483 if (gfp_mask & __GFP_NOFAIL)
2486 if (fatal_signal_pending(current))
2490 * keep retrying as long as the memcg oom killer is able to make
2491 * a forward progress or bypass the charge if the oom killer
2492 * couldn't make any progress.
2494 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2495 get_order(nr_pages * PAGE_SIZE));
2496 switch (oom_status) {
2498 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2506 if (!(gfp_mask & __GFP_NOFAIL))
2510 * The allocation either can't fail or will lead to more memory
2511 * being freed very soon. Allow memory usage go over the limit
2512 * temporarily by force charging it.
2514 page_counter_charge(&memcg->memory, nr_pages);
2515 if (do_memsw_account())
2516 page_counter_charge(&memcg->memsw, nr_pages);
2517 css_get_many(&memcg->css, nr_pages);
2522 css_get_many(&memcg->css, batch);
2523 if (batch > nr_pages)
2524 refill_stock(memcg, batch - nr_pages);
2527 * If the hierarchy is above the normal consumption range, schedule
2528 * reclaim on returning to userland. We can perform reclaim here
2529 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2530 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2531 * not recorded as it most likely matches current's and won't
2532 * change in the meantime. As high limit is checked again before
2533 * reclaim, the cost of mismatch is negligible.
2536 if (page_counter_read(&memcg->memory) > memcg->high) {
2537 /* Don't bother a random interrupted task */
2538 if (in_interrupt()) {
2539 schedule_work(&memcg->high_work);
2542 current->memcg_nr_pages_over_high += batch;
2543 set_notify_resume(current);
2546 } while ((memcg = parent_mem_cgroup(memcg)));
2551 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2553 if (mem_cgroup_is_root(memcg))
2556 page_counter_uncharge(&memcg->memory, nr_pages);
2557 if (do_memsw_account())
2558 page_counter_uncharge(&memcg->memsw, nr_pages);
2560 css_put_many(&memcg->css, nr_pages);
2563 static void lock_page_lru(struct page *page, int *isolated)
2565 pg_data_t *pgdat = page_pgdat(page);
2567 spin_lock_irq(&pgdat->lru_lock);
2568 if (PageLRU(page)) {
2569 struct lruvec *lruvec;
2571 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2573 del_page_from_lru_list(page, lruvec, page_lru(page));
2579 static void unlock_page_lru(struct page *page, int isolated)
2581 pg_data_t *pgdat = page_pgdat(page);
2584 struct lruvec *lruvec;
2586 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2587 VM_BUG_ON_PAGE(PageLRU(page), page);
2589 add_page_to_lru_list(page, lruvec, page_lru(page));
2591 spin_unlock_irq(&pgdat->lru_lock);
2594 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2599 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2602 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2603 * may already be on some other mem_cgroup's LRU. Take care of it.
2606 lock_page_lru(page, &isolated);
2609 * Nobody should be changing or seriously looking at
2610 * page->mem_cgroup at this point:
2612 * - the page is uncharged
2614 * - the page is off-LRU
2616 * - an anonymous fault has exclusive page access, except for
2617 * a locked page table
2619 * - a page cache insertion, a swapin fault, or a migration
2620 * have the page locked
2622 page->mem_cgroup = memcg;
2625 unlock_page_lru(page, isolated);
2628 #ifdef CONFIG_MEMCG_KMEM
2629 static int memcg_alloc_cache_id(void)
2634 id = ida_simple_get(&memcg_cache_ida,
2635 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2639 if (id < memcg_nr_cache_ids)
2643 * There's no space for the new id in memcg_caches arrays,
2644 * so we have to grow them.
2646 down_write(&memcg_cache_ids_sem);
2648 size = 2 * (id + 1);
2649 if (size < MEMCG_CACHES_MIN_SIZE)
2650 size = MEMCG_CACHES_MIN_SIZE;
2651 else if (size > MEMCG_CACHES_MAX_SIZE)
2652 size = MEMCG_CACHES_MAX_SIZE;
2654 err = memcg_update_all_caches(size);
2656 err = memcg_update_all_list_lrus(size);
2658 memcg_nr_cache_ids = size;
2660 up_write(&memcg_cache_ids_sem);
2663 ida_simple_remove(&memcg_cache_ida, id);
2669 static void memcg_free_cache_id(int id)
2671 ida_simple_remove(&memcg_cache_ida, id);
2674 struct memcg_kmem_cache_create_work {
2675 struct mem_cgroup *memcg;
2676 struct kmem_cache *cachep;
2677 struct work_struct work;
2680 static void memcg_kmem_cache_create_func(struct work_struct *w)
2682 struct memcg_kmem_cache_create_work *cw =
2683 container_of(w, struct memcg_kmem_cache_create_work, work);
2684 struct mem_cgroup *memcg = cw->memcg;
2685 struct kmem_cache *cachep = cw->cachep;
2687 memcg_create_kmem_cache(memcg, cachep);
2689 css_put(&memcg->css);
2694 * Enqueue the creation of a per-memcg kmem_cache.
2696 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2697 struct kmem_cache *cachep)
2699 struct memcg_kmem_cache_create_work *cw;
2701 if (!css_tryget_online(&memcg->css))
2704 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2709 cw->cachep = cachep;
2710 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2712 queue_work(memcg_kmem_cache_wq, &cw->work);
2715 static inline bool memcg_kmem_bypass(void)
2717 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2723 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2724 * @cachep: the original global kmem cache
2726 * Return the kmem_cache we're supposed to use for a slab allocation.
2727 * We try to use the current memcg's version of the cache.
2729 * If the cache does not exist yet, if we are the first user of it, we
2730 * create it asynchronously in a workqueue and let the current allocation
2731 * go through with the original cache.
2733 * This function takes a reference to the cache it returns to assure it
2734 * won't get destroyed while we are working with it. Once the caller is
2735 * done with it, memcg_kmem_put_cache() must be called to release the
2738 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2740 struct mem_cgroup *memcg;
2741 struct kmem_cache *memcg_cachep;
2742 struct memcg_cache_array *arr;
2745 VM_BUG_ON(!is_root_cache(cachep));
2747 if (memcg_kmem_bypass())
2752 if (unlikely(current->active_memcg))
2753 memcg = current->active_memcg;
2755 memcg = mem_cgroup_from_task(current);
2757 if (!memcg || memcg == root_mem_cgroup)
2760 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2764 arr = rcu_dereference(cachep->memcg_params.memcg_caches);
2767 * Make sure we will access the up-to-date value. The code updating
2768 * memcg_caches issues a write barrier to match the data dependency
2769 * barrier inside READ_ONCE() (see memcg_create_kmem_cache()).
2771 memcg_cachep = READ_ONCE(arr->entries[kmemcg_id]);
2774 * If we are in a safe context (can wait, and not in interrupt
2775 * context), we could be be predictable and return right away.
2776 * This would guarantee that the allocation being performed
2777 * already belongs in the new cache.
2779 * However, there are some clashes that can arrive from locking.
2780 * For instance, because we acquire the slab_mutex while doing
2781 * memcg_create_kmem_cache, this means no further allocation
2782 * could happen with the slab_mutex held. So it's better to
2785 * If the memcg is dying or memcg_cache is about to be released,
2786 * don't bother creating new kmem_caches. Because memcg_cachep
2787 * is ZEROed as the fist step of kmem offlining, we don't need
2788 * percpu_ref_tryget_live() here. css_tryget_online() check in
2789 * memcg_schedule_kmem_cache_create() will prevent us from
2790 * creation of a new kmem_cache.
2792 if (unlikely(!memcg_cachep))
2793 memcg_schedule_kmem_cache_create(memcg, cachep);
2794 else if (percpu_ref_tryget(&memcg_cachep->memcg_params.refcnt))
2795 cachep = memcg_cachep;
2802 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2803 * @cachep: the cache returned by memcg_kmem_get_cache
2805 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2807 if (!is_root_cache(cachep))
2808 percpu_ref_put(&cachep->memcg_params.refcnt);
2812 * __memcg_kmem_charge_memcg: charge a kmem page
2813 * @page: page to charge
2814 * @gfp: reclaim mode
2815 * @order: allocation order
2816 * @memcg: memory cgroup to charge
2818 * Returns 0 on success, an error code on failure.
2820 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2821 struct mem_cgroup *memcg)
2823 unsigned int nr_pages = 1 << order;
2824 struct page_counter *counter;
2827 ret = try_charge(memcg, gfp, nr_pages);
2831 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2832 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2835 * Enforce __GFP_NOFAIL allocation because callers are not
2836 * prepared to see failures and likely do not have any failure
2839 if (gfp & __GFP_NOFAIL) {
2840 page_counter_charge(&memcg->kmem, nr_pages);
2843 cancel_charge(memcg, nr_pages);
2850 * __memcg_kmem_charge: charge a kmem page to the current memory cgroup
2851 * @page: page to charge
2852 * @gfp: reclaim mode
2853 * @order: allocation order
2855 * Returns 0 on success, an error code on failure.
2857 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2859 struct mem_cgroup *memcg;
2862 if (memcg_kmem_bypass())
2865 memcg = get_mem_cgroup_from_current();
2866 if (!mem_cgroup_is_root(memcg)) {
2867 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2869 page->mem_cgroup = memcg;
2870 __SetPageKmemcg(page);
2873 css_put(&memcg->css);
2878 * __memcg_kmem_uncharge_memcg: uncharge a kmem page
2879 * @memcg: memcg to uncharge
2880 * @nr_pages: number of pages to uncharge
2882 void __memcg_kmem_uncharge_memcg(struct mem_cgroup *memcg,
2883 unsigned int nr_pages)
2885 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2886 page_counter_uncharge(&memcg->kmem, nr_pages);
2888 page_counter_uncharge(&memcg->memory, nr_pages);
2889 if (do_memsw_account())
2890 page_counter_uncharge(&memcg->memsw, nr_pages);
2893 * __memcg_kmem_uncharge: uncharge a kmem page
2894 * @page: page to uncharge
2895 * @order: allocation order
2897 void __memcg_kmem_uncharge(struct page *page, int order)
2899 struct mem_cgroup *memcg = page->mem_cgroup;
2900 unsigned int nr_pages = 1 << order;
2905 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2906 __memcg_kmem_uncharge_memcg(memcg, nr_pages);
2907 page->mem_cgroup = NULL;
2909 /* slab pages do not have PageKmemcg flag set */
2910 if (PageKmemcg(page))
2911 __ClearPageKmemcg(page);
2913 css_put_many(&memcg->css, nr_pages);
2915 #endif /* CONFIG_MEMCG_KMEM */
2917 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2920 * Because tail pages are not marked as "used", set it. We're under
2921 * pgdat->lru_lock and migration entries setup in all page mappings.
2923 void mem_cgroup_split_huge_fixup(struct page *head)
2927 if (mem_cgroup_disabled())
2930 for (i = 1; i < HPAGE_PMD_NR; i++)
2931 head[i].mem_cgroup = head->mem_cgroup;
2933 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
2935 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2937 #ifdef CONFIG_MEMCG_SWAP
2939 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2940 * @entry: swap entry to be moved
2941 * @from: mem_cgroup which the entry is moved from
2942 * @to: mem_cgroup which the entry is moved to
2944 * It succeeds only when the swap_cgroup's record for this entry is the same
2945 * as the mem_cgroup's id of @from.
2947 * Returns 0 on success, -EINVAL on failure.
2949 * The caller must have charged to @to, IOW, called page_counter_charge() about
2950 * both res and memsw, and called css_get().
2952 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2953 struct mem_cgroup *from, struct mem_cgroup *to)
2955 unsigned short old_id, new_id;
2957 old_id = mem_cgroup_id(from);
2958 new_id = mem_cgroup_id(to);
2960 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2961 mod_memcg_state(from, MEMCG_SWAP, -1);
2962 mod_memcg_state(to, MEMCG_SWAP, 1);
2968 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2969 struct mem_cgroup *from, struct mem_cgroup *to)
2975 static DEFINE_MUTEX(memcg_max_mutex);
2977 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
2978 unsigned long max, bool memsw)
2980 bool enlarge = false;
2981 bool drained = false;
2983 bool limits_invariant;
2984 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
2987 if (signal_pending(current)) {
2992 mutex_lock(&memcg_max_mutex);
2994 * Make sure that the new limit (memsw or memory limit) doesn't
2995 * break our basic invariant rule memory.max <= memsw.max.
2997 limits_invariant = memsw ? max >= memcg->memory.max :
2998 max <= memcg->memsw.max;
2999 if (!limits_invariant) {
3000 mutex_unlock(&memcg_max_mutex);
3004 if (max > counter->max)
3006 ret = page_counter_set_max(counter, max);
3007 mutex_unlock(&memcg_max_mutex);
3013 drain_all_stock(memcg);
3018 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3019 GFP_KERNEL, !memsw)) {
3025 if (!ret && enlarge)
3026 memcg_oom_recover(memcg);
3031 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3033 unsigned long *total_scanned)
3035 unsigned long nr_reclaimed = 0;
3036 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3037 unsigned long reclaimed;
3039 struct mem_cgroup_tree_per_node *mctz;
3040 unsigned long excess;
3041 unsigned long nr_scanned;
3046 mctz = soft_limit_tree_node(pgdat->node_id);
3049 * Do not even bother to check the largest node if the root
3050 * is empty. Do it lockless to prevent lock bouncing. Races
3051 * are acceptable as soft limit is best effort anyway.
3053 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3057 * This loop can run a while, specially if mem_cgroup's continuously
3058 * keep exceeding their soft limit and putting the system under
3065 mz = mem_cgroup_largest_soft_limit_node(mctz);
3070 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3071 gfp_mask, &nr_scanned);
3072 nr_reclaimed += reclaimed;
3073 *total_scanned += nr_scanned;
3074 spin_lock_irq(&mctz->lock);
3075 __mem_cgroup_remove_exceeded(mz, mctz);
3078 * If we failed to reclaim anything from this memory cgroup
3079 * it is time to move on to the next cgroup
3083 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3085 excess = soft_limit_excess(mz->memcg);
3087 * One school of thought says that we should not add
3088 * back the node to the tree if reclaim returns 0.
3089 * But our reclaim could return 0, simply because due
3090 * to priority we are exposing a smaller subset of
3091 * memory to reclaim from. Consider this as a longer
3094 /* If excess == 0, no tree ops */
3095 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3096 spin_unlock_irq(&mctz->lock);
3097 css_put(&mz->memcg->css);
3100 * Could not reclaim anything and there are no more
3101 * mem cgroups to try or we seem to be looping without
3102 * reclaiming anything.
3104 if (!nr_reclaimed &&
3106 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3108 } while (!nr_reclaimed);
3110 css_put(&next_mz->memcg->css);
3111 return nr_reclaimed;
3115 * Test whether @memcg has children, dead or alive. Note that this
3116 * function doesn't care whether @memcg has use_hierarchy enabled and
3117 * returns %true if there are child csses according to the cgroup
3118 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3120 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3125 ret = css_next_child(NULL, &memcg->css);
3131 * Reclaims as many pages from the given memcg as possible.
3133 * Caller is responsible for holding css reference for memcg.
3135 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3137 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3139 /* we call try-to-free pages for make this cgroup empty */
3140 lru_add_drain_all();
3142 drain_all_stock(memcg);
3144 /* try to free all pages in this cgroup */
3145 while (nr_retries && page_counter_read(&memcg->memory)) {
3148 if (signal_pending(current))
3151 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3155 /* maybe some writeback is necessary */
3156 congestion_wait(BLK_RW_ASYNC, HZ/10);
3164 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3165 char *buf, size_t nbytes,
3168 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3170 if (mem_cgroup_is_root(memcg))
3172 return mem_cgroup_force_empty(memcg) ?: nbytes;
3175 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3178 return mem_cgroup_from_css(css)->use_hierarchy;
3181 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3182 struct cftype *cft, u64 val)
3185 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3186 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3188 if (memcg->use_hierarchy == val)
3192 * If parent's use_hierarchy is set, we can't make any modifications
3193 * in the child subtrees. If it is unset, then the change can
3194 * occur, provided the current cgroup has no children.
3196 * For the root cgroup, parent_mem is NULL, we allow value to be
3197 * set if there are no children.
3199 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3200 (val == 1 || val == 0)) {
3201 if (!memcg_has_children(memcg))
3202 memcg->use_hierarchy = val;
3211 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3215 if (mem_cgroup_is_root(memcg)) {
3216 val = memcg_page_state(memcg, MEMCG_CACHE) +
3217 memcg_page_state(memcg, MEMCG_RSS);
3219 val += memcg_page_state(memcg, MEMCG_SWAP);
3222 val = page_counter_read(&memcg->memory);
3224 val = page_counter_read(&memcg->memsw);
3237 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3240 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3241 struct page_counter *counter;
3243 switch (MEMFILE_TYPE(cft->private)) {
3245 counter = &memcg->memory;
3248 counter = &memcg->memsw;
3251 counter = &memcg->kmem;
3254 counter = &memcg->tcpmem;
3260 switch (MEMFILE_ATTR(cft->private)) {
3262 if (counter == &memcg->memory)
3263 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3264 if (counter == &memcg->memsw)
3265 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3266 return (u64)page_counter_read(counter) * PAGE_SIZE;
3268 return (u64)counter->max * PAGE_SIZE;
3270 return (u64)counter->watermark * PAGE_SIZE;
3272 return counter->failcnt;
3273 case RES_SOFT_LIMIT:
3274 return (u64)memcg->soft_limit * PAGE_SIZE;
3280 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg, bool slab_only)
3282 unsigned long stat[MEMCG_NR_STAT];
3283 struct mem_cgroup *mi;
3285 int min_idx, max_idx;
3288 min_idx = NR_SLAB_RECLAIMABLE;
3289 max_idx = NR_SLAB_UNRECLAIMABLE;
3292 max_idx = MEMCG_NR_STAT;
3295 for (i = min_idx; i < max_idx; i++)
3298 for_each_online_cpu(cpu)
3299 for (i = min_idx; i < max_idx; i++)
3300 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3302 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3303 for (i = min_idx; i < max_idx; i++)
3304 atomic_long_add(stat[i], &mi->vmstats[i]);
3307 max_idx = NR_VM_NODE_STAT_ITEMS;
3309 for_each_node(node) {
3310 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3311 struct mem_cgroup_per_node *pi;
3313 for (i = min_idx; i < max_idx; i++)
3316 for_each_online_cpu(cpu)
3317 for (i = min_idx; i < max_idx; i++)
3319 pn->lruvec_stat_cpu->count[i], cpu);
3321 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3322 for (i = min_idx; i < max_idx; i++)
3323 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3327 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3329 unsigned long events[NR_VM_EVENT_ITEMS];
3330 struct mem_cgroup *mi;
3333 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3336 for_each_online_cpu(cpu)
3337 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3338 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3341 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3342 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3343 atomic_long_add(events[i], &mi->vmevents[i]);
3346 #ifdef CONFIG_MEMCG_KMEM
3347 static int memcg_online_kmem(struct mem_cgroup *memcg)
3351 if (cgroup_memory_nokmem)
3354 BUG_ON(memcg->kmemcg_id >= 0);
3355 BUG_ON(memcg->kmem_state);
3357 memcg_id = memcg_alloc_cache_id();
3361 static_branch_inc(&memcg_kmem_enabled_key);
3363 * A memory cgroup is considered kmem-online as soon as it gets
3364 * kmemcg_id. Setting the id after enabling static branching will
3365 * guarantee no one starts accounting before all call sites are
3368 memcg->kmemcg_id = memcg_id;
3369 memcg->kmem_state = KMEM_ONLINE;
3370 INIT_LIST_HEAD(&memcg->kmem_caches);
3375 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3377 struct cgroup_subsys_state *css;
3378 struct mem_cgroup *parent, *child;
3381 if (memcg->kmem_state != KMEM_ONLINE)
3384 * Clear the online state before clearing memcg_caches array
3385 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3386 * guarantees that no cache will be created for this cgroup
3387 * after we are done (see memcg_create_kmem_cache()).
3389 memcg->kmem_state = KMEM_ALLOCATED;
3391 parent = parent_mem_cgroup(memcg);
3393 parent = root_mem_cgroup;
3396 * Deactivate and reparent kmem_caches. Then flush percpu
3397 * slab statistics to have precise values at the parent and
3398 * all ancestor levels. It's required to keep slab stats
3399 * accurate after the reparenting of kmem_caches.
3401 memcg_deactivate_kmem_caches(memcg, parent);
3402 memcg_flush_percpu_vmstats(memcg, true);
3404 kmemcg_id = memcg->kmemcg_id;
3405 BUG_ON(kmemcg_id < 0);
3408 * Change kmemcg_id of this cgroup and all its descendants to the
3409 * parent's id, and then move all entries from this cgroup's list_lrus
3410 * to ones of the parent. After we have finished, all list_lrus
3411 * corresponding to this cgroup are guaranteed to remain empty. The
3412 * ordering is imposed by list_lru_node->lock taken by
3413 * memcg_drain_all_list_lrus().
3415 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3416 css_for_each_descendant_pre(css, &memcg->css) {
3417 child = mem_cgroup_from_css(css);
3418 BUG_ON(child->kmemcg_id != kmemcg_id);
3419 child->kmemcg_id = parent->kmemcg_id;
3420 if (!memcg->use_hierarchy)
3425 memcg_drain_all_list_lrus(kmemcg_id, parent);
3427 memcg_free_cache_id(kmemcg_id);
3430 static void memcg_free_kmem(struct mem_cgroup *memcg)
3432 /* css_alloc() failed, offlining didn't happen */
3433 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3434 memcg_offline_kmem(memcg);
3436 if (memcg->kmem_state == KMEM_ALLOCATED) {
3437 WARN_ON(!list_empty(&memcg->kmem_caches));
3438 static_branch_dec(&memcg_kmem_enabled_key);
3442 static int memcg_online_kmem(struct mem_cgroup *memcg)
3446 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3449 static void memcg_free_kmem(struct mem_cgroup *memcg)
3452 #endif /* CONFIG_MEMCG_KMEM */
3454 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3459 mutex_lock(&memcg_max_mutex);
3460 ret = page_counter_set_max(&memcg->kmem, max);
3461 mutex_unlock(&memcg_max_mutex);
3465 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3469 mutex_lock(&memcg_max_mutex);
3471 ret = page_counter_set_max(&memcg->tcpmem, max);
3475 if (!memcg->tcpmem_active) {
3477 * The active flag needs to be written after the static_key
3478 * update. This is what guarantees that the socket activation
3479 * function is the last one to run. See mem_cgroup_sk_alloc()
3480 * for details, and note that we don't mark any socket as
3481 * belonging to this memcg until that flag is up.
3483 * We need to do this, because static_keys will span multiple
3484 * sites, but we can't control their order. If we mark a socket
3485 * as accounted, but the accounting functions are not patched in
3486 * yet, we'll lose accounting.
3488 * We never race with the readers in mem_cgroup_sk_alloc(),
3489 * because when this value change, the code to process it is not
3492 static_branch_inc(&memcg_sockets_enabled_key);
3493 memcg->tcpmem_active = true;
3496 mutex_unlock(&memcg_max_mutex);
3501 * The user of this function is...
3504 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3505 char *buf, size_t nbytes, loff_t off)
3507 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3508 unsigned long nr_pages;
3511 buf = strstrip(buf);
3512 ret = page_counter_memparse(buf, "-1", &nr_pages);
3516 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3518 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3522 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3524 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3527 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3530 ret = memcg_update_kmem_max(memcg, nr_pages);
3533 ret = memcg_update_tcp_max(memcg, nr_pages);
3537 case RES_SOFT_LIMIT:
3538 memcg->soft_limit = nr_pages;
3542 return ret ?: nbytes;
3545 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3546 size_t nbytes, loff_t off)
3548 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3549 struct page_counter *counter;
3551 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3553 counter = &memcg->memory;
3556 counter = &memcg->memsw;
3559 counter = &memcg->kmem;
3562 counter = &memcg->tcpmem;
3568 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3570 page_counter_reset_watermark(counter);
3573 counter->failcnt = 0;
3582 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3585 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3589 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3590 struct cftype *cft, u64 val)
3592 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3594 if (val & ~MOVE_MASK)
3598 * No kind of locking is needed in here, because ->can_attach() will
3599 * check this value once in the beginning of the process, and then carry
3600 * on with stale data. This means that changes to this value will only
3601 * affect task migrations starting after the change.
3603 memcg->move_charge_at_immigrate = val;
3607 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3608 struct cftype *cft, u64 val)
3616 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3617 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3618 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3620 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3621 int nid, unsigned int lru_mask)
3623 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
3624 unsigned long nr = 0;
3627 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3630 if (!(BIT(lru) & lru_mask))
3632 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3637 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3638 unsigned int lru_mask)
3640 unsigned long nr = 0;
3644 if (!(BIT(lru) & lru_mask))
3646 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3651 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3655 unsigned int lru_mask;
3658 static const struct numa_stat stats[] = {
3659 { "total", LRU_ALL },
3660 { "file", LRU_ALL_FILE },
3661 { "anon", LRU_ALL_ANON },
3662 { "unevictable", BIT(LRU_UNEVICTABLE) },
3664 const struct numa_stat *stat;
3667 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3669 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3670 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3671 seq_printf(m, "%s=%lu", stat->name, nr);
3672 for_each_node_state(nid, N_MEMORY) {
3673 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3675 seq_printf(m, " N%d=%lu", nid, nr);
3680 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3681 struct mem_cgroup *iter;
3684 for_each_mem_cgroup_tree(iter, memcg)
3685 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3686 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3687 for_each_node_state(nid, N_MEMORY) {
3689 for_each_mem_cgroup_tree(iter, memcg)
3690 nr += mem_cgroup_node_nr_lru_pages(
3691 iter, nid, stat->lru_mask);
3692 seq_printf(m, " N%d=%lu", nid, nr);
3699 #endif /* CONFIG_NUMA */
3701 static const unsigned int memcg1_stats[] = {
3712 static const char *const memcg1_stat_names[] = {
3723 /* Universal VM events cgroup1 shows, original sort order */
3724 static const unsigned int memcg1_events[] = {
3731 static const char *const memcg1_event_names[] = {
3738 static int memcg_stat_show(struct seq_file *m, void *v)
3740 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3741 unsigned long memory, memsw;
3742 struct mem_cgroup *mi;
3745 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3746 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3748 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3749 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3751 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3752 memcg_page_state_local(memcg, memcg1_stats[i]) *
3756 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3757 seq_printf(m, "%s %lu\n", memcg1_event_names[i],
3758 memcg_events_local(memcg, memcg1_events[i]));
3760 for (i = 0; i < NR_LRU_LISTS; i++)
3761 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3762 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3765 /* Hierarchical information */
3766 memory = memsw = PAGE_COUNTER_MAX;
3767 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3768 memory = min(memory, mi->memory.max);
3769 memsw = min(memsw, mi->memsw.max);
3771 seq_printf(m, "hierarchical_memory_limit %llu\n",
3772 (u64)memory * PAGE_SIZE);
3773 if (do_memsw_account())
3774 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3775 (u64)memsw * PAGE_SIZE);
3777 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3778 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3780 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3781 (u64)memcg_page_state(memcg, memcg1_stats[i]) *
3785 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3786 seq_printf(m, "total_%s %llu\n", memcg1_event_names[i],
3787 (u64)memcg_events(memcg, memcg1_events[i]));
3789 for (i = 0; i < NR_LRU_LISTS; i++)
3790 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i],
3791 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
3794 #ifdef CONFIG_DEBUG_VM
3797 struct mem_cgroup_per_node *mz;
3798 struct zone_reclaim_stat *rstat;
3799 unsigned long recent_rotated[2] = {0, 0};
3800 unsigned long recent_scanned[2] = {0, 0};
3802 for_each_online_pgdat(pgdat) {
3803 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3804 rstat = &mz->lruvec.reclaim_stat;
3806 recent_rotated[0] += rstat->recent_rotated[0];
3807 recent_rotated[1] += rstat->recent_rotated[1];
3808 recent_scanned[0] += rstat->recent_scanned[0];
3809 recent_scanned[1] += rstat->recent_scanned[1];
3811 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3812 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3813 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3814 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3821 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3824 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3826 return mem_cgroup_swappiness(memcg);
3829 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3830 struct cftype *cft, u64 val)
3832 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3838 memcg->swappiness = val;
3840 vm_swappiness = val;
3845 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3847 struct mem_cgroup_threshold_ary *t;
3848 unsigned long usage;
3853 t = rcu_dereference(memcg->thresholds.primary);
3855 t = rcu_dereference(memcg->memsw_thresholds.primary);
3860 usage = mem_cgroup_usage(memcg, swap);
3863 * current_threshold points to threshold just below or equal to usage.
3864 * If it's not true, a threshold was crossed after last
3865 * call of __mem_cgroup_threshold().
3867 i = t->current_threshold;
3870 * Iterate backward over array of thresholds starting from
3871 * current_threshold and check if a threshold is crossed.
3872 * If none of thresholds below usage is crossed, we read
3873 * only one element of the array here.
3875 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3876 eventfd_signal(t->entries[i].eventfd, 1);
3878 /* i = current_threshold + 1 */
3882 * Iterate forward over array of thresholds starting from
3883 * current_threshold+1 and check if a threshold is crossed.
3884 * If none of thresholds above usage is crossed, we read
3885 * only one element of the array here.
3887 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3888 eventfd_signal(t->entries[i].eventfd, 1);
3890 /* Update current_threshold */
3891 t->current_threshold = i - 1;
3896 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3899 __mem_cgroup_threshold(memcg, false);
3900 if (do_memsw_account())
3901 __mem_cgroup_threshold(memcg, true);
3903 memcg = parent_mem_cgroup(memcg);
3907 static int compare_thresholds(const void *a, const void *b)
3909 const struct mem_cgroup_threshold *_a = a;
3910 const struct mem_cgroup_threshold *_b = b;
3912 if (_a->threshold > _b->threshold)
3915 if (_a->threshold < _b->threshold)
3921 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3923 struct mem_cgroup_eventfd_list *ev;
3925 spin_lock(&memcg_oom_lock);
3927 list_for_each_entry(ev, &memcg->oom_notify, list)
3928 eventfd_signal(ev->eventfd, 1);
3930 spin_unlock(&memcg_oom_lock);
3934 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3936 struct mem_cgroup *iter;
3938 for_each_mem_cgroup_tree(iter, memcg)
3939 mem_cgroup_oom_notify_cb(iter);
3942 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3943 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3945 struct mem_cgroup_thresholds *thresholds;
3946 struct mem_cgroup_threshold_ary *new;
3947 unsigned long threshold;
3948 unsigned long usage;
3951 ret = page_counter_memparse(args, "-1", &threshold);
3955 mutex_lock(&memcg->thresholds_lock);
3958 thresholds = &memcg->thresholds;
3959 usage = mem_cgroup_usage(memcg, false);
3960 } else if (type == _MEMSWAP) {
3961 thresholds = &memcg->memsw_thresholds;
3962 usage = mem_cgroup_usage(memcg, true);
3966 /* Check if a threshold crossed before adding a new one */
3967 if (thresholds->primary)
3968 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3970 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3972 /* Allocate memory for new array of thresholds */
3973 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
3980 /* Copy thresholds (if any) to new array */
3981 if (thresholds->primary) {
3982 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3983 sizeof(struct mem_cgroup_threshold));
3986 /* Add new threshold */
3987 new->entries[size - 1].eventfd = eventfd;
3988 new->entries[size - 1].threshold = threshold;
3990 /* Sort thresholds. Registering of new threshold isn't time-critical */
3991 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3992 compare_thresholds, NULL);
3994 /* Find current threshold */
3995 new->current_threshold = -1;
3996 for (i = 0; i < size; i++) {
3997 if (new->entries[i].threshold <= usage) {
3999 * new->current_threshold will not be used until
4000 * rcu_assign_pointer(), so it's safe to increment
4003 ++new->current_threshold;
4008 /* Free old spare buffer and save old primary buffer as spare */
4009 kfree(thresholds->spare);
4010 thresholds->spare = thresholds->primary;
4012 rcu_assign_pointer(thresholds->primary, new);
4014 /* To be sure that nobody uses thresholds */
4018 mutex_unlock(&memcg->thresholds_lock);
4023 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4024 struct eventfd_ctx *eventfd, const char *args)
4026 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4029 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4030 struct eventfd_ctx *eventfd, const char *args)
4032 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4035 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4036 struct eventfd_ctx *eventfd, enum res_type type)
4038 struct mem_cgroup_thresholds *thresholds;
4039 struct mem_cgroup_threshold_ary *new;
4040 unsigned long usage;
4043 mutex_lock(&memcg->thresholds_lock);
4046 thresholds = &memcg->thresholds;
4047 usage = mem_cgroup_usage(memcg, false);
4048 } else if (type == _MEMSWAP) {
4049 thresholds = &memcg->memsw_thresholds;
4050 usage = mem_cgroup_usage(memcg, true);
4054 if (!thresholds->primary)
4057 /* Check if a threshold crossed before removing */
4058 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4060 /* Calculate new number of threshold */
4062 for (i = 0; i < thresholds->primary->size; i++) {
4063 if (thresholds->primary->entries[i].eventfd != eventfd)
4067 new = thresholds->spare;
4069 /* Set thresholds array to NULL if we don't have thresholds */
4078 /* Copy thresholds and find current threshold */
4079 new->current_threshold = -1;
4080 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4081 if (thresholds->primary->entries[i].eventfd == eventfd)
4084 new->entries[j] = thresholds->primary->entries[i];
4085 if (new->entries[j].threshold <= usage) {
4087 * new->current_threshold will not be used
4088 * until rcu_assign_pointer(), so it's safe to increment
4091 ++new->current_threshold;
4097 /* Swap primary and spare array */
4098 thresholds->spare = thresholds->primary;
4100 rcu_assign_pointer(thresholds->primary, new);
4102 /* To be sure that nobody uses thresholds */
4105 /* If all events are unregistered, free the spare array */
4107 kfree(thresholds->spare);
4108 thresholds->spare = NULL;
4111 mutex_unlock(&memcg->thresholds_lock);
4114 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4115 struct eventfd_ctx *eventfd)
4117 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4120 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4121 struct eventfd_ctx *eventfd)
4123 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4126 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4127 struct eventfd_ctx *eventfd, const char *args)
4129 struct mem_cgroup_eventfd_list *event;
4131 event = kmalloc(sizeof(*event), GFP_KERNEL);
4135 spin_lock(&memcg_oom_lock);
4137 event->eventfd = eventfd;
4138 list_add(&event->list, &memcg->oom_notify);
4140 /* already in OOM ? */
4141 if (memcg->under_oom)
4142 eventfd_signal(eventfd, 1);
4143 spin_unlock(&memcg_oom_lock);
4148 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4149 struct eventfd_ctx *eventfd)
4151 struct mem_cgroup_eventfd_list *ev, *tmp;
4153 spin_lock(&memcg_oom_lock);
4155 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4156 if (ev->eventfd == eventfd) {
4157 list_del(&ev->list);
4162 spin_unlock(&memcg_oom_lock);
4165 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4167 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4169 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4170 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4171 seq_printf(sf, "oom_kill %lu\n",
4172 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4176 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4177 struct cftype *cft, u64 val)
4179 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4181 /* cannot set to root cgroup and only 0 and 1 are allowed */
4182 if (!css->parent || !((val == 0) || (val == 1)))
4185 memcg->oom_kill_disable = val;
4187 memcg_oom_recover(memcg);
4192 #ifdef CONFIG_CGROUP_WRITEBACK
4194 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4196 return wb_domain_init(&memcg->cgwb_domain, gfp);
4199 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4201 wb_domain_exit(&memcg->cgwb_domain);
4204 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4206 wb_domain_size_changed(&memcg->cgwb_domain);
4209 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4211 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4213 if (!memcg->css.parent)
4216 return &memcg->cgwb_domain;
4220 * idx can be of type enum memcg_stat_item or node_stat_item.
4221 * Keep in sync with memcg_exact_page().
4223 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4225 long x = atomic_long_read(&memcg->vmstats[idx]);
4228 for_each_online_cpu(cpu)
4229 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4236 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4237 * @wb: bdi_writeback in question
4238 * @pfilepages: out parameter for number of file pages
4239 * @pheadroom: out parameter for number of allocatable pages according to memcg
4240 * @pdirty: out parameter for number of dirty pages
4241 * @pwriteback: out parameter for number of pages under writeback
4243 * Determine the numbers of file, headroom, dirty, and writeback pages in
4244 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4245 * is a bit more involved.
4247 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4248 * headroom is calculated as the lowest headroom of itself and the
4249 * ancestors. Note that this doesn't consider the actual amount of
4250 * available memory in the system. The caller should further cap
4251 * *@pheadroom accordingly.
4253 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4254 unsigned long *pheadroom, unsigned long *pdirty,
4255 unsigned long *pwriteback)
4257 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4258 struct mem_cgroup *parent;
4260 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4262 /* this should eventually include NR_UNSTABLE_NFS */
4263 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4264 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4265 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4266 *pheadroom = PAGE_COUNTER_MAX;
4268 while ((parent = parent_mem_cgroup(memcg))) {
4269 unsigned long ceiling = min(memcg->memory.max, memcg->high);
4270 unsigned long used = page_counter_read(&memcg->memory);
4272 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4277 #else /* CONFIG_CGROUP_WRITEBACK */
4279 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4284 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4288 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4292 #endif /* CONFIG_CGROUP_WRITEBACK */
4295 * DO NOT USE IN NEW FILES.
4297 * "cgroup.event_control" implementation.
4299 * This is way over-engineered. It tries to support fully configurable
4300 * events for each user. Such level of flexibility is completely
4301 * unnecessary especially in the light of the planned unified hierarchy.
4303 * Please deprecate this and replace with something simpler if at all
4308 * Unregister event and free resources.
4310 * Gets called from workqueue.
4312 static void memcg_event_remove(struct work_struct *work)
4314 struct mem_cgroup_event *event =
4315 container_of(work, struct mem_cgroup_event, remove);
4316 struct mem_cgroup *memcg = event->memcg;
4318 remove_wait_queue(event->wqh, &event->wait);
4320 event->unregister_event(memcg, event->eventfd);
4322 /* Notify userspace the event is going away. */
4323 eventfd_signal(event->eventfd, 1);
4325 eventfd_ctx_put(event->eventfd);
4327 css_put(&memcg->css);
4331 * Gets called on EPOLLHUP on eventfd when user closes it.
4333 * Called with wqh->lock held and interrupts disabled.
4335 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4336 int sync, void *key)
4338 struct mem_cgroup_event *event =
4339 container_of(wait, struct mem_cgroup_event, wait);
4340 struct mem_cgroup *memcg = event->memcg;
4341 __poll_t flags = key_to_poll(key);
4343 if (flags & EPOLLHUP) {
4345 * If the event has been detached at cgroup removal, we
4346 * can simply return knowing the other side will cleanup
4349 * We can't race against event freeing since the other
4350 * side will require wqh->lock via remove_wait_queue(),
4353 spin_lock(&memcg->event_list_lock);
4354 if (!list_empty(&event->list)) {
4355 list_del_init(&event->list);
4357 * We are in atomic context, but cgroup_event_remove()
4358 * may sleep, so we have to call it in workqueue.
4360 schedule_work(&event->remove);
4362 spin_unlock(&memcg->event_list_lock);
4368 static void memcg_event_ptable_queue_proc(struct file *file,
4369 wait_queue_head_t *wqh, poll_table *pt)
4371 struct mem_cgroup_event *event =
4372 container_of(pt, struct mem_cgroup_event, pt);
4375 add_wait_queue(wqh, &event->wait);
4379 * DO NOT USE IN NEW FILES.
4381 * Parse input and register new cgroup event handler.
4383 * Input must be in format '<event_fd> <control_fd> <args>'.
4384 * Interpretation of args is defined by control file implementation.
4386 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4387 char *buf, size_t nbytes, loff_t off)
4389 struct cgroup_subsys_state *css = of_css(of);
4390 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4391 struct mem_cgroup_event *event;
4392 struct cgroup_subsys_state *cfile_css;
4393 unsigned int efd, cfd;
4400 buf = strstrip(buf);
4402 efd = simple_strtoul(buf, &endp, 10);
4407 cfd = simple_strtoul(buf, &endp, 10);
4408 if ((*endp != ' ') && (*endp != '\0'))
4412 event = kzalloc(sizeof(*event), GFP_KERNEL);
4416 event->memcg = memcg;
4417 INIT_LIST_HEAD(&event->list);
4418 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4419 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4420 INIT_WORK(&event->remove, memcg_event_remove);
4428 event->eventfd = eventfd_ctx_fileget(efile.file);
4429 if (IS_ERR(event->eventfd)) {
4430 ret = PTR_ERR(event->eventfd);
4437 goto out_put_eventfd;
4440 /* the process need read permission on control file */
4441 /* AV: shouldn't we check that it's been opened for read instead? */
4442 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4447 * Determine the event callbacks and set them in @event. This used
4448 * to be done via struct cftype but cgroup core no longer knows
4449 * about these events. The following is crude but the whole thing
4450 * is for compatibility anyway.
4452 * DO NOT ADD NEW FILES.
4454 name = cfile.file->f_path.dentry->d_name.name;
4456 if (!strcmp(name, "memory.usage_in_bytes")) {
4457 event->register_event = mem_cgroup_usage_register_event;
4458 event->unregister_event = mem_cgroup_usage_unregister_event;
4459 } else if (!strcmp(name, "memory.oom_control")) {
4460 event->register_event = mem_cgroup_oom_register_event;
4461 event->unregister_event = mem_cgroup_oom_unregister_event;
4462 } else if (!strcmp(name, "memory.pressure_level")) {
4463 event->register_event = vmpressure_register_event;
4464 event->unregister_event = vmpressure_unregister_event;
4465 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4466 event->register_event = memsw_cgroup_usage_register_event;
4467 event->unregister_event = memsw_cgroup_usage_unregister_event;
4474 * Verify @cfile should belong to @css. Also, remaining events are
4475 * automatically removed on cgroup destruction but the removal is
4476 * asynchronous, so take an extra ref on @css.
4478 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4479 &memory_cgrp_subsys);
4481 if (IS_ERR(cfile_css))
4483 if (cfile_css != css) {
4488 ret = event->register_event(memcg, event->eventfd, buf);
4492 vfs_poll(efile.file, &event->pt);
4494 spin_lock(&memcg->event_list_lock);
4495 list_add(&event->list, &memcg->event_list);
4496 spin_unlock(&memcg->event_list_lock);
4508 eventfd_ctx_put(event->eventfd);
4517 static struct cftype mem_cgroup_legacy_files[] = {
4519 .name = "usage_in_bytes",
4520 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4521 .read_u64 = mem_cgroup_read_u64,
4524 .name = "max_usage_in_bytes",
4525 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4526 .write = mem_cgroup_reset,
4527 .read_u64 = mem_cgroup_read_u64,
4530 .name = "limit_in_bytes",
4531 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4532 .write = mem_cgroup_write,
4533 .read_u64 = mem_cgroup_read_u64,
4536 .name = "soft_limit_in_bytes",
4537 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4538 .write = mem_cgroup_write,
4539 .read_u64 = mem_cgroup_read_u64,
4543 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4544 .write = mem_cgroup_reset,
4545 .read_u64 = mem_cgroup_read_u64,
4549 .seq_show = memcg_stat_show,
4552 .name = "force_empty",
4553 .write = mem_cgroup_force_empty_write,
4556 .name = "use_hierarchy",
4557 .write_u64 = mem_cgroup_hierarchy_write,
4558 .read_u64 = mem_cgroup_hierarchy_read,
4561 .name = "cgroup.event_control", /* XXX: for compat */
4562 .write = memcg_write_event_control,
4563 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4566 .name = "swappiness",
4567 .read_u64 = mem_cgroup_swappiness_read,
4568 .write_u64 = mem_cgroup_swappiness_write,
4571 .name = "move_charge_at_immigrate",
4572 .read_u64 = mem_cgroup_move_charge_read,
4573 .write_u64 = mem_cgroup_move_charge_write,
4576 .name = "oom_control",
4577 .seq_show = mem_cgroup_oom_control_read,
4578 .write_u64 = mem_cgroup_oom_control_write,
4579 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4582 .name = "pressure_level",
4586 .name = "numa_stat",
4587 .seq_show = memcg_numa_stat_show,
4591 .name = "kmem.limit_in_bytes",
4592 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4593 .write = mem_cgroup_write,
4594 .read_u64 = mem_cgroup_read_u64,
4597 .name = "kmem.usage_in_bytes",
4598 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4599 .read_u64 = mem_cgroup_read_u64,
4602 .name = "kmem.failcnt",
4603 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4604 .write = mem_cgroup_reset,
4605 .read_u64 = mem_cgroup_read_u64,
4608 .name = "kmem.max_usage_in_bytes",
4609 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4610 .write = mem_cgroup_reset,
4611 .read_u64 = mem_cgroup_read_u64,
4613 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4615 .name = "kmem.slabinfo",
4616 .seq_start = memcg_slab_start,
4617 .seq_next = memcg_slab_next,
4618 .seq_stop = memcg_slab_stop,
4619 .seq_show = memcg_slab_show,
4623 .name = "kmem.tcp.limit_in_bytes",
4624 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4625 .write = mem_cgroup_write,
4626 .read_u64 = mem_cgroup_read_u64,
4629 .name = "kmem.tcp.usage_in_bytes",
4630 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4631 .read_u64 = mem_cgroup_read_u64,
4634 .name = "kmem.tcp.failcnt",
4635 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4636 .write = mem_cgroup_reset,
4637 .read_u64 = mem_cgroup_read_u64,
4640 .name = "kmem.tcp.max_usage_in_bytes",
4641 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4642 .write = mem_cgroup_reset,
4643 .read_u64 = mem_cgroup_read_u64,
4645 { }, /* terminate */
4649 * Private memory cgroup IDR
4651 * Swap-out records and page cache shadow entries need to store memcg
4652 * references in constrained space, so we maintain an ID space that is
4653 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4654 * memory-controlled cgroups to 64k.
4656 * However, there usually are many references to the oflline CSS after
4657 * the cgroup has been destroyed, such as page cache or reclaimable
4658 * slab objects, that don't need to hang on to the ID. We want to keep
4659 * those dead CSS from occupying IDs, or we might quickly exhaust the
4660 * relatively small ID space and prevent the creation of new cgroups
4661 * even when there are much fewer than 64k cgroups - possibly none.
4663 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4664 * be freed and recycled when it's no longer needed, which is usually
4665 * when the CSS is offlined.
4667 * The only exception to that are records of swapped out tmpfs/shmem
4668 * pages that need to be attributed to live ancestors on swapin. But
4669 * those references are manageable from userspace.
4672 static DEFINE_IDR(mem_cgroup_idr);
4674 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4676 if (memcg->id.id > 0) {
4677 idr_remove(&mem_cgroup_idr, memcg->id.id);
4682 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4684 refcount_add(n, &memcg->id.ref);
4687 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4689 if (refcount_sub_and_test(n, &memcg->id.ref)) {
4690 mem_cgroup_id_remove(memcg);
4692 /* Memcg ID pins CSS */
4693 css_put(&memcg->css);
4697 static inline void mem_cgroup_id_get(struct mem_cgroup *memcg)
4699 mem_cgroup_id_get_many(memcg, 1);
4702 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4704 mem_cgroup_id_put_many(memcg, 1);
4708 * mem_cgroup_from_id - look up a memcg from a memcg id
4709 * @id: the memcg id to look up
4711 * Caller must hold rcu_read_lock().
4713 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4715 WARN_ON_ONCE(!rcu_read_lock_held());
4716 return idr_find(&mem_cgroup_idr, id);
4719 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4721 struct mem_cgroup_per_node *pn;
4724 * This routine is called against possible nodes.
4725 * But it's BUG to call kmalloc() against offline node.
4727 * TODO: this routine can waste much memory for nodes which will
4728 * never be onlined. It's better to use memory hotplug callback
4731 if (!node_state(node, N_NORMAL_MEMORY))
4733 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4737 pn->lruvec_stat_local = alloc_percpu(struct lruvec_stat);
4738 if (!pn->lruvec_stat_local) {
4743 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
4744 if (!pn->lruvec_stat_cpu) {
4745 free_percpu(pn->lruvec_stat_local);
4750 lruvec_init(&pn->lruvec);
4751 pn->usage_in_excess = 0;
4752 pn->on_tree = false;
4755 memcg->nodeinfo[node] = pn;
4759 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4761 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
4766 free_percpu(pn->lruvec_stat_cpu);
4767 free_percpu(pn->lruvec_stat_local);
4771 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4776 free_mem_cgroup_per_node_info(memcg, node);
4777 free_percpu(memcg->vmstats_percpu);
4778 free_percpu(memcg->vmstats_local);
4782 static void mem_cgroup_free(struct mem_cgroup *memcg)
4784 memcg_wb_domain_exit(memcg);
4786 * Flush percpu vmstats and vmevents to guarantee the value correctness
4787 * on parent's and all ancestor levels.
4789 memcg_flush_percpu_vmstats(memcg, false);
4790 memcg_flush_percpu_vmevents(memcg);
4791 __mem_cgroup_free(memcg);
4794 static struct mem_cgroup *mem_cgroup_alloc(void)
4796 struct mem_cgroup *memcg;
4800 size = sizeof(struct mem_cgroup);
4801 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4803 memcg = kzalloc(size, GFP_KERNEL);
4807 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4808 1, MEM_CGROUP_ID_MAX,
4810 if (memcg->id.id < 0)
4813 memcg->vmstats_local = alloc_percpu(struct memcg_vmstats_percpu);
4814 if (!memcg->vmstats_local)
4817 memcg->vmstats_percpu = alloc_percpu(struct memcg_vmstats_percpu);
4818 if (!memcg->vmstats_percpu)
4822 if (alloc_mem_cgroup_per_node_info(memcg, node))
4825 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4828 INIT_WORK(&memcg->high_work, high_work_func);
4829 memcg->last_scanned_node = MAX_NUMNODES;
4830 INIT_LIST_HEAD(&memcg->oom_notify);
4831 mutex_init(&memcg->thresholds_lock);
4832 spin_lock_init(&memcg->move_lock);
4833 vmpressure_init(&memcg->vmpressure);
4834 INIT_LIST_HEAD(&memcg->event_list);
4835 spin_lock_init(&memcg->event_list_lock);
4836 memcg->socket_pressure = jiffies;
4837 #ifdef CONFIG_MEMCG_KMEM
4838 memcg->kmemcg_id = -1;
4840 #ifdef CONFIG_CGROUP_WRITEBACK
4841 INIT_LIST_HEAD(&memcg->cgwb_list);
4843 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4846 mem_cgroup_id_remove(memcg);
4847 __mem_cgroup_free(memcg);
4851 static struct cgroup_subsys_state * __ref
4852 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4854 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4855 struct mem_cgroup *memcg;
4856 long error = -ENOMEM;
4858 memcg = mem_cgroup_alloc();
4860 return ERR_PTR(error);
4862 memcg->high = PAGE_COUNTER_MAX;
4863 memcg->soft_limit = PAGE_COUNTER_MAX;
4865 memcg->swappiness = mem_cgroup_swappiness(parent);
4866 memcg->oom_kill_disable = parent->oom_kill_disable;
4868 if (parent && parent->use_hierarchy) {
4869 memcg->use_hierarchy = true;
4870 page_counter_init(&memcg->memory, &parent->memory);
4871 page_counter_init(&memcg->swap, &parent->swap);
4872 page_counter_init(&memcg->memsw, &parent->memsw);
4873 page_counter_init(&memcg->kmem, &parent->kmem);
4874 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4876 page_counter_init(&memcg->memory, NULL);
4877 page_counter_init(&memcg->swap, NULL);
4878 page_counter_init(&memcg->memsw, NULL);
4879 page_counter_init(&memcg->kmem, NULL);
4880 page_counter_init(&memcg->tcpmem, NULL);
4882 * Deeper hierachy with use_hierarchy == false doesn't make
4883 * much sense so let cgroup subsystem know about this
4884 * unfortunate state in our controller.
4886 if (parent != root_mem_cgroup)
4887 memory_cgrp_subsys.broken_hierarchy = true;
4890 /* The following stuff does not apply to the root */
4892 #ifdef CONFIG_MEMCG_KMEM
4893 INIT_LIST_HEAD(&memcg->kmem_caches);
4895 root_mem_cgroup = memcg;
4899 error = memcg_online_kmem(memcg);
4903 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4904 static_branch_inc(&memcg_sockets_enabled_key);
4908 mem_cgroup_id_remove(memcg);
4909 mem_cgroup_free(memcg);
4910 return ERR_PTR(-ENOMEM);
4913 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
4915 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4918 * A memcg must be visible for memcg_expand_shrinker_maps()
4919 * by the time the maps are allocated. So, we allocate maps
4920 * here, when for_each_mem_cgroup() can't skip it.
4922 if (memcg_alloc_shrinker_maps(memcg)) {
4923 mem_cgroup_id_remove(memcg);
4927 /* Online state pins memcg ID, memcg ID pins CSS */
4928 refcount_set(&memcg->id.ref, 1);
4933 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4935 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4936 struct mem_cgroup_event *event, *tmp;
4939 * Unregister events and notify userspace.
4940 * Notify userspace about cgroup removing only after rmdir of cgroup
4941 * directory to avoid race between userspace and kernelspace.
4943 spin_lock(&memcg->event_list_lock);
4944 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4945 list_del_init(&event->list);
4946 schedule_work(&event->remove);
4948 spin_unlock(&memcg->event_list_lock);
4950 page_counter_set_min(&memcg->memory, 0);
4951 page_counter_set_low(&memcg->memory, 0);
4953 memcg_offline_kmem(memcg);
4954 wb_memcg_offline(memcg);
4956 drain_all_stock(memcg);
4958 mem_cgroup_id_put(memcg);
4961 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4963 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4965 invalidate_reclaim_iterators(memcg);
4968 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4970 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4972 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4973 static_branch_dec(&memcg_sockets_enabled_key);
4975 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4976 static_branch_dec(&memcg_sockets_enabled_key);
4978 vmpressure_cleanup(&memcg->vmpressure);
4979 cancel_work_sync(&memcg->high_work);
4980 mem_cgroup_remove_from_trees(memcg);
4981 memcg_free_shrinker_maps(memcg);
4982 memcg_free_kmem(memcg);
4983 mem_cgroup_free(memcg);
4987 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4988 * @css: the target css
4990 * Reset the states of the mem_cgroup associated with @css. This is
4991 * invoked when the userland requests disabling on the default hierarchy
4992 * but the memcg is pinned through dependency. The memcg should stop
4993 * applying policies and should revert to the vanilla state as it may be
4994 * made visible again.
4996 * The current implementation only resets the essential configurations.
4997 * This needs to be expanded to cover all the visible parts.
4999 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5001 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5003 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5004 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5005 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
5006 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5007 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5008 page_counter_set_min(&memcg->memory, 0);
5009 page_counter_set_low(&memcg->memory, 0);
5010 memcg->high = PAGE_COUNTER_MAX;
5011 memcg->soft_limit = PAGE_COUNTER_MAX;
5012 memcg_wb_domain_size_changed(memcg);
5016 /* Handlers for move charge at task migration. */
5017 static int mem_cgroup_do_precharge(unsigned long count)
5021 /* Try a single bulk charge without reclaim first, kswapd may wake */
5022 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5024 mc.precharge += count;
5028 /* Try charges one by one with reclaim, but do not retry */
5030 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5044 enum mc_target_type {
5051 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5052 unsigned long addr, pte_t ptent)
5054 struct page *page = vm_normal_page(vma, addr, ptent);
5056 if (!page || !page_mapped(page))
5058 if (PageAnon(page)) {
5059 if (!(mc.flags & MOVE_ANON))
5062 if (!(mc.flags & MOVE_FILE))
5065 if (!get_page_unless_zero(page))
5071 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5072 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5073 pte_t ptent, swp_entry_t *entry)
5075 struct page *page = NULL;
5076 swp_entry_t ent = pte_to_swp_entry(ptent);
5078 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
5082 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5083 * a device and because they are not accessible by CPU they are store
5084 * as special swap entry in the CPU page table.
5086 if (is_device_private_entry(ent)) {
5087 page = device_private_entry_to_page(ent);
5089 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5090 * a refcount of 1 when free (unlike normal page)
5092 if (!page_ref_add_unless(page, 1, 1))
5098 * Because lookup_swap_cache() updates some statistics counter,
5099 * we call find_get_page() with swapper_space directly.
5101 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5102 if (do_memsw_account())
5103 entry->val = ent.val;
5108 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5109 pte_t ptent, swp_entry_t *entry)
5115 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5116 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5118 struct page *page = NULL;
5119 struct address_space *mapping;
5122 if (!vma->vm_file) /* anonymous vma */
5124 if (!(mc.flags & MOVE_FILE))
5127 mapping = vma->vm_file->f_mapping;
5128 pgoff = linear_page_index(vma, addr);
5130 /* page is moved even if it's not RSS of this task(page-faulted). */
5132 /* shmem/tmpfs may report page out on swap: account for that too. */
5133 if (shmem_mapping(mapping)) {
5134 page = find_get_entry(mapping, pgoff);
5135 if (xa_is_value(page)) {
5136 swp_entry_t swp = radix_to_swp_entry(page);
5137 if (do_memsw_account())
5139 page = find_get_page(swap_address_space(swp),
5143 page = find_get_page(mapping, pgoff);
5145 page = find_get_page(mapping, pgoff);
5151 * mem_cgroup_move_account - move account of the page
5153 * @compound: charge the page as compound or small page
5154 * @from: mem_cgroup which the page is moved from.
5155 * @to: mem_cgroup which the page is moved to. @from != @to.
5157 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5159 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5162 static int mem_cgroup_move_account(struct page *page,
5164 struct mem_cgroup *from,
5165 struct mem_cgroup *to)
5167 unsigned long flags;
5168 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5172 VM_BUG_ON(from == to);
5173 VM_BUG_ON_PAGE(PageLRU(page), page);
5174 VM_BUG_ON(compound && !PageTransHuge(page));
5177 * Prevent mem_cgroup_migrate() from looking at
5178 * page->mem_cgroup of its source page while we change it.
5181 if (!trylock_page(page))
5185 if (page->mem_cgroup != from)
5188 anon = PageAnon(page);
5190 spin_lock_irqsave(&from->move_lock, flags);
5192 if (!anon && page_mapped(page)) {
5193 __mod_memcg_state(from, NR_FILE_MAPPED, -nr_pages);
5194 __mod_memcg_state(to, NR_FILE_MAPPED, nr_pages);
5198 * move_lock grabbed above and caller set from->moving_account, so
5199 * mod_memcg_page_state will serialize updates to PageDirty.
5200 * So mapping should be stable for dirty pages.
5202 if (!anon && PageDirty(page)) {
5203 struct address_space *mapping = page_mapping(page);
5205 if (mapping_cap_account_dirty(mapping)) {
5206 __mod_memcg_state(from, NR_FILE_DIRTY, -nr_pages);
5207 __mod_memcg_state(to, NR_FILE_DIRTY, nr_pages);
5211 if (PageWriteback(page)) {
5212 __mod_memcg_state(from, NR_WRITEBACK, -nr_pages);
5213 __mod_memcg_state(to, NR_WRITEBACK, nr_pages);
5217 * It is safe to change page->mem_cgroup here because the page
5218 * is referenced, charged, and isolated - we can't race with
5219 * uncharging, charging, migration, or LRU putback.
5222 /* caller should have done css_get */
5223 page->mem_cgroup = to;
5224 spin_unlock_irqrestore(&from->move_lock, flags);
5228 local_irq_disable();
5229 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
5230 memcg_check_events(to, page);
5231 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
5232 memcg_check_events(from, page);
5241 * get_mctgt_type - get target type of moving charge
5242 * @vma: the vma the pte to be checked belongs
5243 * @addr: the address corresponding to the pte to be checked
5244 * @ptent: the pte to be checked
5245 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5248 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5249 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5250 * move charge. if @target is not NULL, the page is stored in target->page
5251 * with extra refcnt got(Callers should handle it).
5252 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5253 * target for charge migration. if @target is not NULL, the entry is stored
5255 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5256 * (so ZONE_DEVICE page and thus not on the lru).
5257 * For now we such page is charge like a regular page would be as for all
5258 * intent and purposes it is just special memory taking the place of a
5261 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5263 * Called with pte lock held.
5266 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5267 unsigned long addr, pte_t ptent, union mc_target *target)
5269 struct page *page = NULL;
5270 enum mc_target_type ret = MC_TARGET_NONE;
5271 swp_entry_t ent = { .val = 0 };
5273 if (pte_present(ptent))
5274 page = mc_handle_present_pte(vma, addr, ptent);
5275 else if (is_swap_pte(ptent))
5276 page = mc_handle_swap_pte(vma, ptent, &ent);
5277 else if (pte_none(ptent))
5278 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5280 if (!page && !ent.val)
5284 * Do only loose check w/o serialization.
5285 * mem_cgroup_move_account() checks the page is valid or
5286 * not under LRU exclusion.
5288 if (page->mem_cgroup == mc.from) {
5289 ret = MC_TARGET_PAGE;
5290 if (is_device_private_page(page))
5291 ret = MC_TARGET_DEVICE;
5293 target->page = page;
5295 if (!ret || !target)
5299 * There is a swap entry and a page doesn't exist or isn't charged.
5300 * But we cannot move a tail-page in a THP.
5302 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5303 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5304 ret = MC_TARGET_SWAP;
5311 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5313 * We don't consider PMD mapped swapping or file mapped pages because THP does
5314 * not support them for now.
5315 * Caller should make sure that pmd_trans_huge(pmd) is true.
5317 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5318 unsigned long addr, pmd_t pmd, union mc_target *target)
5320 struct page *page = NULL;
5321 enum mc_target_type ret = MC_TARGET_NONE;
5323 if (unlikely(is_swap_pmd(pmd))) {
5324 VM_BUG_ON(thp_migration_supported() &&
5325 !is_pmd_migration_entry(pmd));
5328 page = pmd_page(pmd);
5329 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5330 if (!(mc.flags & MOVE_ANON))
5332 if (page->mem_cgroup == mc.from) {
5333 ret = MC_TARGET_PAGE;
5336 target->page = page;
5342 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5343 unsigned long addr, pmd_t pmd, union mc_target *target)
5345 return MC_TARGET_NONE;
5349 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5350 unsigned long addr, unsigned long end,
5351 struct mm_walk *walk)
5353 struct vm_area_struct *vma = walk->vma;
5357 ptl = pmd_trans_huge_lock(pmd, vma);
5360 * Note their can not be MC_TARGET_DEVICE for now as we do not
5361 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5362 * this might change.
5364 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5365 mc.precharge += HPAGE_PMD_NR;
5370 if (pmd_trans_unstable(pmd))
5372 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5373 for (; addr != end; pte++, addr += PAGE_SIZE)
5374 if (get_mctgt_type(vma, addr, *pte, NULL))
5375 mc.precharge++; /* increment precharge temporarily */
5376 pte_unmap_unlock(pte - 1, ptl);
5382 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5384 unsigned long precharge;
5386 struct mm_walk mem_cgroup_count_precharge_walk = {
5387 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5390 down_read(&mm->mmap_sem);
5391 walk_page_range(0, mm->highest_vm_end,
5392 &mem_cgroup_count_precharge_walk);
5393 up_read(&mm->mmap_sem);
5395 precharge = mc.precharge;
5401 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5403 unsigned long precharge = mem_cgroup_count_precharge(mm);
5405 VM_BUG_ON(mc.moving_task);
5406 mc.moving_task = current;
5407 return mem_cgroup_do_precharge(precharge);
5410 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5411 static void __mem_cgroup_clear_mc(void)
5413 struct mem_cgroup *from = mc.from;
5414 struct mem_cgroup *to = mc.to;
5416 /* we must uncharge all the leftover precharges from mc.to */
5418 cancel_charge(mc.to, mc.precharge);
5422 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5423 * we must uncharge here.
5425 if (mc.moved_charge) {
5426 cancel_charge(mc.from, mc.moved_charge);
5427 mc.moved_charge = 0;
5429 /* we must fixup refcnts and charges */
5430 if (mc.moved_swap) {
5431 /* uncharge swap account from the old cgroup */
5432 if (!mem_cgroup_is_root(mc.from))
5433 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5435 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5438 * we charged both to->memory and to->memsw, so we
5439 * should uncharge to->memory.
5441 if (!mem_cgroup_is_root(mc.to))
5442 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5444 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
5445 css_put_many(&mc.to->css, mc.moved_swap);
5449 memcg_oom_recover(from);
5450 memcg_oom_recover(to);
5451 wake_up_all(&mc.waitq);
5454 static void mem_cgroup_clear_mc(void)
5456 struct mm_struct *mm = mc.mm;
5459 * we must clear moving_task before waking up waiters at the end of
5462 mc.moving_task = NULL;
5463 __mem_cgroup_clear_mc();
5464 spin_lock(&mc.lock);
5468 spin_unlock(&mc.lock);
5473 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5475 struct cgroup_subsys_state *css;
5476 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5477 struct mem_cgroup *from;
5478 struct task_struct *leader, *p;
5479 struct mm_struct *mm;
5480 unsigned long move_flags;
5483 /* charge immigration isn't supported on the default hierarchy */
5484 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5488 * Multi-process migrations only happen on the default hierarchy
5489 * where charge immigration is not used. Perform charge
5490 * immigration if @tset contains a leader and whine if there are
5494 cgroup_taskset_for_each_leader(leader, css, tset) {
5497 memcg = mem_cgroup_from_css(css);
5503 * We are now commited to this value whatever it is. Changes in this
5504 * tunable will only affect upcoming migrations, not the current one.
5505 * So we need to save it, and keep it going.
5507 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5511 from = mem_cgroup_from_task(p);
5513 VM_BUG_ON(from == memcg);
5515 mm = get_task_mm(p);
5518 /* We move charges only when we move a owner of the mm */
5519 if (mm->owner == p) {
5522 VM_BUG_ON(mc.precharge);
5523 VM_BUG_ON(mc.moved_charge);
5524 VM_BUG_ON(mc.moved_swap);
5526 spin_lock(&mc.lock);
5530 mc.flags = move_flags;
5531 spin_unlock(&mc.lock);
5532 /* We set mc.moving_task later */
5534 ret = mem_cgroup_precharge_mc(mm);
5536 mem_cgroup_clear_mc();
5543 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5546 mem_cgroup_clear_mc();
5549 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5550 unsigned long addr, unsigned long end,
5551 struct mm_walk *walk)
5554 struct vm_area_struct *vma = walk->vma;
5557 enum mc_target_type target_type;
5558 union mc_target target;
5561 ptl = pmd_trans_huge_lock(pmd, vma);
5563 if (mc.precharge < HPAGE_PMD_NR) {
5567 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5568 if (target_type == MC_TARGET_PAGE) {
5570 if (!isolate_lru_page(page)) {
5571 if (!mem_cgroup_move_account(page, true,
5573 mc.precharge -= HPAGE_PMD_NR;
5574 mc.moved_charge += HPAGE_PMD_NR;
5576 putback_lru_page(page);
5579 } else if (target_type == MC_TARGET_DEVICE) {
5581 if (!mem_cgroup_move_account(page, true,
5583 mc.precharge -= HPAGE_PMD_NR;
5584 mc.moved_charge += HPAGE_PMD_NR;
5592 if (pmd_trans_unstable(pmd))
5595 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5596 for (; addr != end; addr += PAGE_SIZE) {
5597 pte_t ptent = *(pte++);
5598 bool device = false;
5604 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5605 case MC_TARGET_DEVICE:
5608 case MC_TARGET_PAGE:
5611 * We can have a part of the split pmd here. Moving it
5612 * can be done but it would be too convoluted so simply
5613 * ignore such a partial THP and keep it in original
5614 * memcg. There should be somebody mapping the head.
5616 if (PageTransCompound(page))
5618 if (!device && isolate_lru_page(page))
5620 if (!mem_cgroup_move_account(page, false,
5623 /* we uncharge from mc.from later. */
5627 putback_lru_page(page);
5628 put: /* get_mctgt_type() gets the page */
5631 case MC_TARGET_SWAP:
5633 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5635 /* we fixup refcnts and charges later. */
5643 pte_unmap_unlock(pte - 1, ptl);
5648 * We have consumed all precharges we got in can_attach().
5649 * We try charge one by one, but don't do any additional
5650 * charges to mc.to if we have failed in charge once in attach()
5653 ret = mem_cgroup_do_precharge(1);
5661 static void mem_cgroup_move_charge(void)
5663 struct mm_walk mem_cgroup_move_charge_walk = {
5664 .pmd_entry = mem_cgroup_move_charge_pte_range,
5668 lru_add_drain_all();
5670 * Signal lock_page_memcg() to take the memcg's move_lock
5671 * while we're moving its pages to another memcg. Then wait
5672 * for already started RCU-only updates to finish.
5674 atomic_inc(&mc.from->moving_account);
5677 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5679 * Someone who are holding the mmap_sem might be waiting in
5680 * waitq. So we cancel all extra charges, wake up all waiters,
5681 * and retry. Because we cancel precharges, we might not be able
5682 * to move enough charges, but moving charge is a best-effort
5683 * feature anyway, so it wouldn't be a big problem.
5685 __mem_cgroup_clear_mc();
5690 * When we have consumed all precharges and failed in doing
5691 * additional charge, the page walk just aborts.
5693 walk_page_range(0, mc.mm->highest_vm_end, &mem_cgroup_move_charge_walk);
5695 up_read(&mc.mm->mmap_sem);
5696 atomic_dec(&mc.from->moving_account);
5699 static void mem_cgroup_move_task(void)
5702 mem_cgroup_move_charge();
5703 mem_cgroup_clear_mc();
5706 #else /* !CONFIG_MMU */
5707 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5711 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5714 static void mem_cgroup_move_task(void)
5720 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5721 * to verify whether we're attached to the default hierarchy on each mount
5724 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5727 * use_hierarchy is forced on the default hierarchy. cgroup core
5728 * guarantees that @root doesn't have any children, so turning it
5729 * on for the root memcg is enough.
5731 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5732 root_mem_cgroup->use_hierarchy = true;
5734 root_mem_cgroup->use_hierarchy = false;
5737 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
5739 if (value == PAGE_COUNTER_MAX)
5740 seq_puts(m, "max\n");
5742 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
5747 static u64 memory_current_read(struct cgroup_subsys_state *css,
5750 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5752 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5755 static int memory_min_show(struct seq_file *m, void *v)
5757 return seq_puts_memcg_tunable(m,
5758 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
5761 static ssize_t memory_min_write(struct kernfs_open_file *of,
5762 char *buf, size_t nbytes, loff_t off)
5764 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5768 buf = strstrip(buf);
5769 err = page_counter_memparse(buf, "max", &min);
5773 page_counter_set_min(&memcg->memory, min);
5778 static int memory_low_show(struct seq_file *m, void *v)
5780 return seq_puts_memcg_tunable(m,
5781 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
5784 static ssize_t memory_low_write(struct kernfs_open_file *of,
5785 char *buf, size_t nbytes, loff_t off)
5787 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5791 buf = strstrip(buf);
5792 err = page_counter_memparse(buf, "max", &low);
5796 page_counter_set_low(&memcg->memory, low);
5801 static int memory_high_show(struct seq_file *m, void *v)
5803 return seq_puts_memcg_tunable(m, READ_ONCE(mem_cgroup_from_seq(m)->high));
5806 static ssize_t memory_high_write(struct kernfs_open_file *of,
5807 char *buf, size_t nbytes, loff_t off)
5809 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5810 unsigned long nr_pages;
5814 buf = strstrip(buf);
5815 err = page_counter_memparse(buf, "max", &high);
5821 nr_pages = page_counter_read(&memcg->memory);
5822 if (nr_pages > high)
5823 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5826 memcg_wb_domain_size_changed(memcg);
5830 static int memory_max_show(struct seq_file *m, void *v)
5832 return seq_puts_memcg_tunable(m,
5833 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
5836 static ssize_t memory_max_write(struct kernfs_open_file *of,
5837 char *buf, size_t nbytes, loff_t off)
5839 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5840 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5841 bool drained = false;
5845 buf = strstrip(buf);
5846 err = page_counter_memparse(buf, "max", &max);
5850 xchg(&memcg->memory.max, max);
5853 unsigned long nr_pages = page_counter_read(&memcg->memory);
5855 if (nr_pages <= max)
5858 if (signal_pending(current)) {
5864 drain_all_stock(memcg);
5870 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5876 memcg_memory_event(memcg, MEMCG_OOM);
5877 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5881 memcg_wb_domain_size_changed(memcg);
5885 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
5887 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
5888 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
5889 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
5890 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
5891 seq_printf(m, "oom_kill %lu\n",
5892 atomic_long_read(&events[MEMCG_OOM_KILL]));
5895 static int memory_events_show(struct seq_file *m, void *v)
5897 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5899 __memory_events_show(m, memcg->memory_events);
5903 static int memory_events_local_show(struct seq_file *m, void *v)
5905 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5907 __memory_events_show(m, memcg->memory_events_local);
5911 static int memory_stat_show(struct seq_file *m, void *v)
5913 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5916 buf = memory_stat_format(memcg);
5924 static int memory_oom_group_show(struct seq_file *m, void *v)
5926 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5928 seq_printf(m, "%d\n", memcg->oom_group);
5933 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
5934 char *buf, size_t nbytes, loff_t off)
5936 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5939 buf = strstrip(buf);
5943 ret = kstrtoint(buf, 0, &oom_group);
5947 if (oom_group != 0 && oom_group != 1)
5950 memcg->oom_group = oom_group;
5955 static struct cftype memory_files[] = {
5958 .flags = CFTYPE_NOT_ON_ROOT,
5959 .read_u64 = memory_current_read,
5963 .flags = CFTYPE_NOT_ON_ROOT,
5964 .seq_show = memory_min_show,
5965 .write = memory_min_write,
5969 .flags = CFTYPE_NOT_ON_ROOT,
5970 .seq_show = memory_low_show,
5971 .write = memory_low_write,
5975 .flags = CFTYPE_NOT_ON_ROOT,
5976 .seq_show = memory_high_show,
5977 .write = memory_high_write,
5981 .flags = CFTYPE_NOT_ON_ROOT,
5982 .seq_show = memory_max_show,
5983 .write = memory_max_write,
5987 .flags = CFTYPE_NOT_ON_ROOT,
5988 .file_offset = offsetof(struct mem_cgroup, events_file),
5989 .seq_show = memory_events_show,
5992 .name = "events.local",
5993 .flags = CFTYPE_NOT_ON_ROOT,
5994 .file_offset = offsetof(struct mem_cgroup, events_local_file),
5995 .seq_show = memory_events_local_show,
5999 .flags = CFTYPE_NOT_ON_ROOT,
6000 .seq_show = memory_stat_show,
6003 .name = "oom.group",
6004 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6005 .seq_show = memory_oom_group_show,
6006 .write = memory_oom_group_write,
6011 struct cgroup_subsys memory_cgrp_subsys = {
6012 .css_alloc = mem_cgroup_css_alloc,
6013 .css_online = mem_cgroup_css_online,
6014 .css_offline = mem_cgroup_css_offline,
6015 .css_released = mem_cgroup_css_released,
6016 .css_free = mem_cgroup_css_free,
6017 .css_reset = mem_cgroup_css_reset,
6018 .can_attach = mem_cgroup_can_attach,
6019 .cancel_attach = mem_cgroup_cancel_attach,
6020 .post_attach = mem_cgroup_move_task,
6021 .bind = mem_cgroup_bind,
6022 .dfl_cftypes = memory_files,
6023 .legacy_cftypes = mem_cgroup_legacy_files,
6028 * mem_cgroup_protected - check if memory consumption is in the normal range
6029 * @root: the top ancestor of the sub-tree being checked
6030 * @memcg: the memory cgroup to check
6032 * WARNING: This function is not stateless! It can only be used as part
6033 * of a top-down tree iteration, not for isolated queries.
6035 * Returns one of the following:
6036 * MEMCG_PROT_NONE: cgroup memory is not protected
6037 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
6038 * an unprotected supply of reclaimable memory from other cgroups.
6039 * MEMCG_PROT_MIN: cgroup memory is protected
6041 * @root is exclusive; it is never protected when looked at directly
6043 * To provide a proper hierarchical behavior, effective memory.min/low values
6044 * are used. Below is the description of how effective memory.low is calculated.
6045 * Effective memory.min values is calculated in the same way.
6047 * Effective memory.low is always equal or less than the original memory.low.
6048 * If there is no memory.low overcommittment (which is always true for
6049 * top-level memory cgroups), these two values are equal.
6050 * Otherwise, it's a part of parent's effective memory.low,
6051 * calculated as a cgroup's memory.low usage divided by sum of sibling's
6052 * memory.low usages, where memory.low usage is the size of actually
6056 * elow = min( memory.low, parent->elow * ------------------ ),
6057 * siblings_low_usage
6059 * | memory.current, if memory.current < memory.low
6064 * Such definition of the effective memory.low provides the expected
6065 * hierarchical behavior: parent's memory.low value is limiting
6066 * children, unprotected memory is reclaimed first and cgroups,
6067 * which are not using their guarantee do not affect actual memory
6070 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
6072 * A A/memory.low = 2G, A/memory.current = 6G
6074 * BC DE B/memory.low = 3G B/memory.current = 2G
6075 * C/memory.low = 1G C/memory.current = 2G
6076 * D/memory.low = 0 D/memory.current = 2G
6077 * E/memory.low = 10G E/memory.current = 0
6079 * and the memory pressure is applied, the following memory distribution
6080 * is expected (approximately):
6082 * A/memory.current = 2G
6084 * B/memory.current = 1.3G
6085 * C/memory.current = 0.6G
6086 * D/memory.current = 0
6087 * E/memory.current = 0
6089 * These calculations require constant tracking of the actual low usages
6090 * (see propagate_protected_usage()), as well as recursive calculation of
6091 * effective memory.low values. But as we do call mem_cgroup_protected()
6092 * path for each memory cgroup top-down from the reclaim,
6093 * it's possible to optimize this part, and save calculated elow
6094 * for next usage. This part is intentionally racy, but it's ok,
6095 * as memory.low is a best-effort mechanism.
6097 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
6098 struct mem_cgroup *memcg)
6100 struct mem_cgroup *parent;
6101 unsigned long emin, parent_emin;
6102 unsigned long elow, parent_elow;
6103 unsigned long usage;
6105 if (mem_cgroup_disabled())
6106 return MEMCG_PROT_NONE;
6109 root = root_mem_cgroup;
6111 return MEMCG_PROT_NONE;
6113 usage = page_counter_read(&memcg->memory);
6115 return MEMCG_PROT_NONE;
6117 emin = memcg->memory.min;
6118 elow = memcg->memory.low;
6120 parent = parent_mem_cgroup(memcg);
6121 /* No parent means a non-hierarchical mode on v1 memcg */
6123 return MEMCG_PROT_NONE;
6128 parent_emin = READ_ONCE(parent->memory.emin);
6129 emin = min(emin, parent_emin);
6130 if (emin && parent_emin) {
6131 unsigned long min_usage, siblings_min_usage;
6133 min_usage = min(usage, memcg->memory.min);
6134 siblings_min_usage = atomic_long_read(
6135 &parent->memory.children_min_usage);
6137 if (min_usage && siblings_min_usage)
6138 emin = min(emin, parent_emin * min_usage /
6139 siblings_min_usage);
6142 parent_elow = READ_ONCE(parent->memory.elow);
6143 elow = min(elow, parent_elow);
6144 if (elow && parent_elow) {
6145 unsigned long low_usage, siblings_low_usage;
6147 low_usage = min(usage, memcg->memory.low);
6148 siblings_low_usage = atomic_long_read(
6149 &parent->memory.children_low_usage);
6151 if (low_usage && siblings_low_usage)
6152 elow = min(elow, parent_elow * low_usage /
6153 siblings_low_usage);
6157 memcg->memory.emin = emin;
6158 memcg->memory.elow = elow;
6161 return MEMCG_PROT_MIN;
6162 else if (usage <= elow)
6163 return MEMCG_PROT_LOW;
6165 return MEMCG_PROT_NONE;
6169 * mem_cgroup_try_charge - try charging a page
6170 * @page: page to charge
6171 * @mm: mm context of the victim
6172 * @gfp_mask: reclaim mode
6173 * @memcgp: charged memcg return
6174 * @compound: charge the page as compound or small page
6176 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6177 * pages according to @gfp_mask if necessary.
6179 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6180 * Otherwise, an error code is returned.
6182 * After page->mapping has been set up, the caller must finalize the
6183 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6184 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6186 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
6187 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6190 struct mem_cgroup *memcg = NULL;
6191 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6194 if (mem_cgroup_disabled())
6197 if (PageSwapCache(page)) {
6199 * Every swap fault against a single page tries to charge the
6200 * page, bail as early as possible. shmem_unuse() encounters
6201 * already charged pages, too. The USED bit is protected by
6202 * the page lock, which serializes swap cache removal, which
6203 * in turn serializes uncharging.
6205 VM_BUG_ON_PAGE(!PageLocked(page), page);
6206 if (compound_head(page)->mem_cgroup)
6209 if (do_swap_account) {
6210 swp_entry_t ent = { .val = page_private(page), };
6211 unsigned short id = lookup_swap_cgroup_id(ent);
6214 memcg = mem_cgroup_from_id(id);
6215 if (memcg && !css_tryget_online(&memcg->css))
6222 memcg = get_mem_cgroup_from_mm(mm);
6224 ret = try_charge(memcg, gfp_mask, nr_pages);
6226 css_put(&memcg->css);
6232 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
6233 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6236 struct mem_cgroup *memcg;
6239 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound);
6241 mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask);
6246 * mem_cgroup_commit_charge - commit a page charge
6247 * @page: page to charge
6248 * @memcg: memcg to charge the page to
6249 * @lrucare: page might be on LRU already
6250 * @compound: charge the page as compound or small page
6252 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6253 * after page->mapping has been set up. This must happen atomically
6254 * as part of the page instantiation, i.e. under the page table lock
6255 * for anonymous pages, under the page lock for page and swap cache.
6257 * In addition, the page must not be on the LRU during the commit, to
6258 * prevent racing with task migration. If it might be, use @lrucare.
6260 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6262 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
6263 bool lrucare, bool compound)
6265 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6267 VM_BUG_ON_PAGE(!page->mapping, page);
6268 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
6270 if (mem_cgroup_disabled())
6273 * Swap faults will attempt to charge the same page multiple
6274 * times. But reuse_swap_page() might have removed the page
6275 * from swapcache already, so we can't check PageSwapCache().
6280 commit_charge(page, memcg, lrucare);
6282 local_irq_disable();
6283 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
6284 memcg_check_events(memcg, page);
6287 if (do_memsw_account() && PageSwapCache(page)) {
6288 swp_entry_t entry = { .val = page_private(page) };
6290 * The swap entry might not get freed for a long time,
6291 * let's not wait for it. The page already received a
6292 * memory+swap charge, drop the swap entry duplicate.
6294 mem_cgroup_uncharge_swap(entry, nr_pages);
6299 * mem_cgroup_cancel_charge - cancel a page charge
6300 * @page: page to charge
6301 * @memcg: memcg to charge the page to
6302 * @compound: charge the page as compound or small page
6304 * Cancel a charge transaction started by mem_cgroup_try_charge().
6306 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
6309 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6311 if (mem_cgroup_disabled())
6314 * Swap faults will attempt to charge the same page multiple
6315 * times. But reuse_swap_page() might have removed the page
6316 * from swapcache already, so we can't check PageSwapCache().
6321 cancel_charge(memcg, nr_pages);
6324 struct uncharge_gather {
6325 struct mem_cgroup *memcg;
6326 unsigned long pgpgout;
6327 unsigned long nr_anon;
6328 unsigned long nr_file;
6329 unsigned long nr_kmem;
6330 unsigned long nr_huge;
6331 unsigned long nr_shmem;
6332 struct page *dummy_page;
6335 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6337 memset(ug, 0, sizeof(*ug));
6340 static void uncharge_batch(const struct uncharge_gather *ug)
6342 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
6343 unsigned long flags;
6345 if (!mem_cgroup_is_root(ug->memcg)) {
6346 page_counter_uncharge(&ug->memcg->memory, nr_pages);
6347 if (do_memsw_account())
6348 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
6349 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6350 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6351 memcg_oom_recover(ug->memcg);
6354 local_irq_save(flags);
6355 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
6356 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
6357 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
6358 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
6359 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6360 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, nr_pages);
6361 memcg_check_events(ug->memcg, ug->dummy_page);
6362 local_irq_restore(flags);
6364 if (!mem_cgroup_is_root(ug->memcg))
6365 css_put_many(&ug->memcg->css, nr_pages);
6368 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6370 VM_BUG_ON_PAGE(PageLRU(page), page);
6371 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
6372 !PageHWPoison(page) , page);
6374 if (!page->mem_cgroup)
6378 * Nobody should be changing or seriously looking at
6379 * page->mem_cgroup at this point, we have fully
6380 * exclusive access to the page.
6383 if (ug->memcg != page->mem_cgroup) {
6386 uncharge_gather_clear(ug);
6388 ug->memcg = page->mem_cgroup;
6391 if (!PageKmemcg(page)) {
6392 unsigned int nr_pages = 1;
6394 if (PageTransHuge(page)) {
6395 nr_pages <<= compound_order(page);
6396 ug->nr_huge += nr_pages;
6399 ug->nr_anon += nr_pages;
6401 ug->nr_file += nr_pages;
6402 if (PageSwapBacked(page))
6403 ug->nr_shmem += nr_pages;
6407 ug->nr_kmem += 1 << compound_order(page);
6408 __ClearPageKmemcg(page);
6411 ug->dummy_page = page;
6412 page->mem_cgroup = NULL;
6415 static void uncharge_list(struct list_head *page_list)
6417 struct uncharge_gather ug;
6418 struct list_head *next;
6420 uncharge_gather_clear(&ug);
6423 * Note that the list can be a single page->lru; hence the
6424 * do-while loop instead of a simple list_for_each_entry().
6426 next = page_list->next;
6430 page = list_entry(next, struct page, lru);
6431 next = page->lru.next;
6433 uncharge_page(page, &ug);
6434 } while (next != page_list);
6437 uncharge_batch(&ug);
6441 * mem_cgroup_uncharge - uncharge a page
6442 * @page: page to uncharge
6444 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6445 * mem_cgroup_commit_charge().
6447 void mem_cgroup_uncharge(struct page *page)
6449 struct uncharge_gather ug;
6451 if (mem_cgroup_disabled())
6454 /* Don't touch page->lru of any random page, pre-check: */
6455 if (!page->mem_cgroup)
6458 uncharge_gather_clear(&ug);
6459 uncharge_page(page, &ug);
6460 uncharge_batch(&ug);
6464 * mem_cgroup_uncharge_list - uncharge a list of page
6465 * @page_list: list of pages to uncharge
6467 * Uncharge a list of pages previously charged with
6468 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6470 void mem_cgroup_uncharge_list(struct list_head *page_list)
6472 if (mem_cgroup_disabled())
6475 if (!list_empty(page_list))
6476 uncharge_list(page_list);
6480 * mem_cgroup_migrate - charge a page's replacement
6481 * @oldpage: currently circulating page
6482 * @newpage: replacement page
6484 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6485 * be uncharged upon free.
6487 * Both pages must be locked, @newpage->mapping must be set up.
6489 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6491 struct mem_cgroup *memcg;
6492 unsigned int nr_pages;
6494 unsigned long flags;
6496 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6497 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6498 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6499 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6502 if (mem_cgroup_disabled())
6505 /* Page cache replacement: new page already charged? */
6506 if (newpage->mem_cgroup)
6509 /* Swapcache readahead pages can get replaced before being charged */
6510 memcg = oldpage->mem_cgroup;
6514 /* Force-charge the new page. The old one will be freed soon */
6515 compound = PageTransHuge(newpage);
6516 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
6518 page_counter_charge(&memcg->memory, nr_pages);
6519 if (do_memsw_account())
6520 page_counter_charge(&memcg->memsw, nr_pages);
6521 css_get_many(&memcg->css, nr_pages);
6523 commit_charge(newpage, memcg, false);
6525 local_irq_save(flags);
6526 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
6527 memcg_check_events(memcg, newpage);
6528 local_irq_restore(flags);
6531 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6532 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6534 void mem_cgroup_sk_alloc(struct sock *sk)
6536 struct mem_cgroup *memcg;
6538 if (!mem_cgroup_sockets_enabled)
6542 * Socket cloning can throw us here with sk_memcg already
6543 * filled. It won't however, necessarily happen from
6544 * process context. So the test for root memcg given
6545 * the current task's memcg won't help us in this case.
6547 * Respecting the original socket's memcg is a better
6548 * decision in this case.
6551 css_get(&sk->sk_memcg->css);
6556 memcg = mem_cgroup_from_task(current);
6557 if (memcg == root_mem_cgroup)
6559 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6561 if (css_tryget_online(&memcg->css))
6562 sk->sk_memcg = memcg;
6567 void mem_cgroup_sk_free(struct sock *sk)
6570 css_put(&sk->sk_memcg->css);
6574 * mem_cgroup_charge_skmem - charge socket memory
6575 * @memcg: memcg to charge
6576 * @nr_pages: number of pages to charge
6578 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6579 * @memcg's configured limit, %false if the charge had to be forced.
6581 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6583 gfp_t gfp_mask = GFP_KERNEL;
6585 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6586 struct page_counter *fail;
6588 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6589 memcg->tcpmem_pressure = 0;
6592 page_counter_charge(&memcg->tcpmem, nr_pages);
6593 memcg->tcpmem_pressure = 1;
6597 /* Don't block in the packet receive path */
6599 gfp_mask = GFP_NOWAIT;
6601 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6603 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6606 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6611 * mem_cgroup_uncharge_skmem - uncharge socket memory
6612 * @memcg: memcg to uncharge
6613 * @nr_pages: number of pages to uncharge
6615 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6617 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6618 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6622 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6624 refill_stock(memcg, nr_pages);
6627 static int __init cgroup_memory(char *s)
6631 while ((token = strsep(&s, ",")) != NULL) {
6634 if (!strcmp(token, "nosocket"))
6635 cgroup_memory_nosocket = true;
6636 if (!strcmp(token, "nokmem"))
6637 cgroup_memory_nokmem = true;
6641 __setup("cgroup.memory=", cgroup_memory);
6644 * subsys_initcall() for memory controller.
6646 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6647 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6648 * basically everything that doesn't depend on a specific mem_cgroup structure
6649 * should be initialized from here.
6651 static int __init mem_cgroup_init(void)
6655 #ifdef CONFIG_MEMCG_KMEM
6657 * Kmem cache creation is mostly done with the slab_mutex held,
6658 * so use a workqueue with limited concurrency to avoid stalling
6659 * all worker threads in case lots of cgroups are created and
6660 * destroyed simultaneously.
6662 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
6663 BUG_ON(!memcg_kmem_cache_wq);
6666 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6667 memcg_hotplug_cpu_dead);
6669 for_each_possible_cpu(cpu)
6670 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
6673 for_each_node(node) {
6674 struct mem_cgroup_tree_per_node *rtpn;
6676 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
6677 node_online(node) ? node : NUMA_NO_NODE);
6679 rtpn->rb_root = RB_ROOT;
6680 rtpn->rb_rightmost = NULL;
6681 spin_lock_init(&rtpn->lock);
6682 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6687 subsys_initcall(mem_cgroup_init);
6689 #ifdef CONFIG_MEMCG_SWAP
6690 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
6692 while (!refcount_inc_not_zero(&memcg->id.ref)) {
6694 * The root cgroup cannot be destroyed, so it's refcount must
6697 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
6701 memcg = parent_mem_cgroup(memcg);
6703 memcg = root_mem_cgroup;
6709 * mem_cgroup_swapout - transfer a memsw charge to swap
6710 * @page: page whose memsw charge to transfer
6711 * @entry: swap entry to move the charge to
6713 * Transfer the memsw charge of @page to @entry.
6715 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6717 struct mem_cgroup *memcg, *swap_memcg;
6718 unsigned int nr_entries;
6719 unsigned short oldid;
6721 VM_BUG_ON_PAGE(PageLRU(page), page);
6722 VM_BUG_ON_PAGE(page_count(page), page);
6724 if (!do_memsw_account())
6727 memcg = page->mem_cgroup;
6729 /* Readahead page, never charged */
6734 * In case the memcg owning these pages has been offlined and doesn't
6735 * have an ID allocated to it anymore, charge the closest online
6736 * ancestor for the swap instead and transfer the memory+swap charge.
6738 swap_memcg = mem_cgroup_id_get_online(memcg);
6739 nr_entries = hpage_nr_pages(page);
6740 /* Get references for the tail pages, too */
6742 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
6743 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
6745 VM_BUG_ON_PAGE(oldid, page);
6746 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
6748 page->mem_cgroup = NULL;
6750 if (!mem_cgroup_is_root(memcg))
6751 page_counter_uncharge(&memcg->memory, nr_entries);
6753 if (memcg != swap_memcg) {
6754 if (!mem_cgroup_is_root(swap_memcg))
6755 page_counter_charge(&swap_memcg->memsw, nr_entries);
6756 page_counter_uncharge(&memcg->memsw, nr_entries);
6760 * Interrupts should be disabled here because the caller holds the
6761 * i_pages lock which is taken with interrupts-off. It is
6762 * important here to have the interrupts disabled because it is the
6763 * only synchronisation we have for updating the per-CPU variables.
6765 VM_BUG_ON(!irqs_disabled());
6766 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
6768 memcg_check_events(memcg, page);
6770 if (!mem_cgroup_is_root(memcg))
6771 css_put_many(&memcg->css, nr_entries);
6775 * mem_cgroup_try_charge_swap - try charging swap space for a page
6776 * @page: page being added to swap
6777 * @entry: swap entry to charge
6779 * Try to charge @page's memcg for the swap space at @entry.
6781 * Returns 0 on success, -ENOMEM on failure.
6783 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
6785 unsigned int nr_pages = hpage_nr_pages(page);
6786 struct page_counter *counter;
6787 struct mem_cgroup *memcg;
6788 unsigned short oldid;
6790 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
6793 memcg = page->mem_cgroup;
6795 /* Readahead page, never charged */
6800 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6804 memcg = mem_cgroup_id_get_online(memcg);
6806 if (!mem_cgroup_is_root(memcg) &&
6807 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
6808 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
6809 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6810 mem_cgroup_id_put(memcg);
6814 /* Get references for the tail pages, too */
6816 mem_cgroup_id_get_many(memcg, nr_pages - 1);
6817 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
6818 VM_BUG_ON_PAGE(oldid, page);
6819 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
6825 * mem_cgroup_uncharge_swap - uncharge swap space
6826 * @entry: swap entry to uncharge
6827 * @nr_pages: the amount of swap space to uncharge
6829 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
6831 struct mem_cgroup *memcg;
6834 if (!do_swap_account)
6837 id = swap_cgroup_record(entry, 0, nr_pages);
6839 memcg = mem_cgroup_from_id(id);
6841 if (!mem_cgroup_is_root(memcg)) {
6842 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6843 page_counter_uncharge(&memcg->swap, nr_pages);
6845 page_counter_uncharge(&memcg->memsw, nr_pages);
6847 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
6848 mem_cgroup_id_put_many(memcg, nr_pages);
6853 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
6855 long nr_swap_pages = get_nr_swap_pages();
6857 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6858 return nr_swap_pages;
6859 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6860 nr_swap_pages = min_t(long, nr_swap_pages,
6861 READ_ONCE(memcg->swap.max) -
6862 page_counter_read(&memcg->swap));
6863 return nr_swap_pages;
6866 bool mem_cgroup_swap_full(struct page *page)
6868 struct mem_cgroup *memcg;
6870 VM_BUG_ON_PAGE(!PageLocked(page), page);
6874 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6877 memcg = page->mem_cgroup;
6881 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6882 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.max)
6888 /* for remember boot option*/
6889 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6890 static int really_do_swap_account __initdata = 1;
6892 static int really_do_swap_account __initdata;
6895 static int __init enable_swap_account(char *s)
6897 if (!strcmp(s, "1"))
6898 really_do_swap_account = 1;
6899 else if (!strcmp(s, "0"))
6900 really_do_swap_account = 0;
6903 __setup("swapaccount=", enable_swap_account);
6905 static u64 swap_current_read(struct cgroup_subsys_state *css,
6908 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6910 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
6913 static int swap_max_show(struct seq_file *m, void *v)
6915 return seq_puts_memcg_tunable(m,
6916 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
6919 static ssize_t swap_max_write(struct kernfs_open_file *of,
6920 char *buf, size_t nbytes, loff_t off)
6922 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6926 buf = strstrip(buf);
6927 err = page_counter_memparse(buf, "max", &max);
6931 xchg(&memcg->swap.max, max);
6936 static int swap_events_show(struct seq_file *m, void *v)
6938 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6940 seq_printf(m, "max %lu\n",
6941 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
6942 seq_printf(m, "fail %lu\n",
6943 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
6948 static struct cftype swap_files[] = {
6950 .name = "swap.current",
6951 .flags = CFTYPE_NOT_ON_ROOT,
6952 .read_u64 = swap_current_read,
6956 .flags = CFTYPE_NOT_ON_ROOT,
6957 .seq_show = swap_max_show,
6958 .write = swap_max_write,
6961 .name = "swap.events",
6962 .flags = CFTYPE_NOT_ON_ROOT,
6963 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
6964 .seq_show = swap_events_show,
6969 static struct cftype memsw_cgroup_files[] = {
6971 .name = "memsw.usage_in_bytes",
6972 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6973 .read_u64 = mem_cgroup_read_u64,
6976 .name = "memsw.max_usage_in_bytes",
6977 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6978 .write = mem_cgroup_reset,
6979 .read_u64 = mem_cgroup_read_u64,
6982 .name = "memsw.limit_in_bytes",
6983 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6984 .write = mem_cgroup_write,
6985 .read_u64 = mem_cgroup_read_u64,
6988 .name = "memsw.failcnt",
6989 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6990 .write = mem_cgroup_reset,
6991 .read_u64 = mem_cgroup_read_u64,
6993 { }, /* terminate */
6996 static int __init mem_cgroup_swap_init(void)
6998 if (!mem_cgroup_disabled() && really_do_swap_account) {
6999 do_swap_account = 1;
7000 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
7002 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
7003 memsw_cgroup_files));
7007 subsys_initcall(mem_cgroup_swap_init);
7009 #endif /* CONFIG_MEMCG_SWAP */