1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* memcontrol.c - Memory Controller
4 * Copyright IBM Corporation, 2007
5 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
7 * Copyright 2007 OpenVZ SWsoft Inc
8 * Author: Pavel Emelianov <xemul@openvz.org>
11 * Copyright (C) 2009 Nokia Corporation
12 * Author: Kirill A. Shutemov
14 * Kernel Memory Controller
15 * Copyright (C) 2012 Parallels Inc. and Google Inc.
16 * Authors: Glauber Costa and Suleiman Souhlal
19 * Charge lifetime sanitation
20 * Lockless page tracking & accounting
21 * Unified hierarchy configuration model
22 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
25 #include <linux/page_counter.h>
26 #include <linux/memcontrol.h>
27 #include <linux/cgroup.h>
28 #include <linux/pagewalk.h>
29 #include <linux/sched/mm.h>
30 #include <linux/shmem_fs.h>
31 #include <linux/hugetlb.h>
32 #include <linux/pagemap.h>
33 #include <linux/vm_event_item.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/swap_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
59 #include <linux/tracehook.h>
60 #include <linux/psi.h>
61 #include <linux/seq_buf.h>
67 #include <linux/uaccess.h>
69 #include <trace/events/vmscan.h>
71 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
72 EXPORT_SYMBOL(memory_cgrp_subsys);
74 struct mem_cgroup *root_mem_cgroup __read_mostly;
76 #define MEM_CGROUP_RECLAIM_RETRIES 5
78 /* Socket memory accounting disabled? */
79 static bool cgroup_memory_nosocket;
81 /* Kernel memory accounting disabled? */
82 static bool cgroup_memory_nokmem;
84 /* Whether the swap controller is active */
85 #ifdef CONFIG_MEMCG_SWAP
86 int do_swap_account __read_mostly;
88 #define do_swap_account 0
91 #ifdef CONFIG_CGROUP_WRITEBACK
92 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
95 /* Whether legacy memory+swap accounting is active */
96 static bool do_memsw_account(void)
98 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && do_swap_account;
101 static const char *const mem_cgroup_lru_names[] = {
109 #define THRESHOLDS_EVENTS_TARGET 128
110 #define SOFTLIMIT_EVENTS_TARGET 1024
111 #define NUMAINFO_EVENTS_TARGET 1024
114 * Cgroups above their limits are maintained in a RB-Tree, independent of
115 * their hierarchy representation
118 struct mem_cgroup_tree_per_node {
119 struct rb_root rb_root;
120 struct rb_node *rb_rightmost;
124 struct mem_cgroup_tree {
125 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
128 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
131 struct mem_cgroup_eventfd_list {
132 struct list_head list;
133 struct eventfd_ctx *eventfd;
137 * cgroup_event represents events which userspace want to receive.
139 struct mem_cgroup_event {
141 * memcg which the event belongs to.
143 struct mem_cgroup *memcg;
145 * eventfd to signal userspace about the event.
147 struct eventfd_ctx *eventfd;
149 * Each of these stored in a list by the cgroup.
151 struct list_head list;
153 * register_event() callback will be used to add new userspace
154 * waiter for changes related to this event. Use eventfd_signal()
155 * on eventfd to send notification to userspace.
157 int (*register_event)(struct mem_cgroup *memcg,
158 struct eventfd_ctx *eventfd, const char *args);
160 * unregister_event() callback will be called when userspace closes
161 * the eventfd or on cgroup removing. This callback must be set,
162 * if you want provide notification functionality.
164 void (*unregister_event)(struct mem_cgroup *memcg,
165 struct eventfd_ctx *eventfd);
167 * All fields below needed to unregister event when
168 * userspace closes eventfd.
171 wait_queue_head_t *wqh;
172 wait_queue_entry_t wait;
173 struct work_struct remove;
176 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
177 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
179 /* Stuffs for move charges at task migration. */
181 * Types of charges to be moved.
183 #define MOVE_ANON 0x1U
184 #define MOVE_FILE 0x2U
185 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
187 /* "mc" and its members are protected by cgroup_mutex */
188 static struct move_charge_struct {
189 spinlock_t lock; /* for from, to */
190 struct mm_struct *mm;
191 struct mem_cgroup *from;
192 struct mem_cgroup *to;
194 unsigned long precharge;
195 unsigned long moved_charge;
196 unsigned long moved_swap;
197 struct task_struct *moving_task; /* a task moving charges */
198 wait_queue_head_t waitq; /* a waitq for other context */
200 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
201 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
205 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
206 * limit reclaim to prevent infinite loops, if they ever occur.
208 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
209 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
212 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
213 MEM_CGROUP_CHARGE_TYPE_ANON,
214 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
215 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
219 /* for encoding cft->private value on file */
228 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
229 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
230 #define MEMFILE_ATTR(val) ((val) & 0xffff)
231 /* Used for OOM nofiier */
232 #define OOM_CONTROL (0)
235 * Iteration constructs for visiting all cgroups (under a tree). If
236 * loops are exited prematurely (break), mem_cgroup_iter_break() must
237 * be used for reference counting.
239 #define for_each_mem_cgroup_tree(iter, root) \
240 for (iter = mem_cgroup_iter(root, NULL, NULL); \
242 iter = mem_cgroup_iter(root, iter, NULL))
244 #define for_each_mem_cgroup(iter) \
245 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
247 iter = mem_cgroup_iter(NULL, iter, NULL))
249 static inline bool should_force_charge(void)
251 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
252 (current->flags & PF_EXITING);
255 /* Some nice accessors for the vmpressure. */
256 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
259 memcg = root_mem_cgroup;
260 return &memcg->vmpressure;
263 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
265 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
268 #ifdef CONFIG_MEMCG_KMEM
270 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
271 * The main reason for not using cgroup id for this:
272 * this works better in sparse environments, where we have a lot of memcgs,
273 * but only a few kmem-limited. Or also, if we have, for instance, 200
274 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
275 * 200 entry array for that.
277 * The current size of the caches array is stored in memcg_nr_cache_ids. It
278 * will double each time we have to increase it.
280 static DEFINE_IDA(memcg_cache_ida);
281 int memcg_nr_cache_ids;
283 /* Protects memcg_nr_cache_ids */
284 static DECLARE_RWSEM(memcg_cache_ids_sem);
286 void memcg_get_cache_ids(void)
288 down_read(&memcg_cache_ids_sem);
291 void memcg_put_cache_ids(void)
293 up_read(&memcg_cache_ids_sem);
297 * MIN_SIZE is different than 1, because we would like to avoid going through
298 * the alloc/free process all the time. In a small machine, 4 kmem-limited
299 * cgroups is a reasonable guess. In the future, it could be a parameter or
300 * tunable, but that is strictly not necessary.
302 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
303 * this constant directly from cgroup, but it is understandable that this is
304 * better kept as an internal representation in cgroup.c. In any case, the
305 * cgrp_id space is not getting any smaller, and we don't have to necessarily
306 * increase ours as well if it increases.
308 #define MEMCG_CACHES_MIN_SIZE 4
309 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
312 * A lot of the calls to the cache allocation functions are expected to be
313 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
314 * conditional to this static branch, we'll have to allow modules that does
315 * kmem_cache_alloc and the such to see this symbol as well
317 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
318 EXPORT_SYMBOL(memcg_kmem_enabled_key);
320 struct workqueue_struct *memcg_kmem_cache_wq;
323 static int memcg_shrinker_map_size;
324 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
326 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
328 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
331 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
332 int size, int old_size)
334 struct memcg_shrinker_map *new, *old;
337 lockdep_assert_held(&memcg_shrinker_map_mutex);
340 old = rcu_dereference_protected(
341 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
342 /* Not yet online memcg */
346 new = kvmalloc(sizeof(*new) + size, GFP_KERNEL);
350 /* Set all old bits, clear all new bits */
351 memset(new->map, (int)0xff, old_size);
352 memset((void *)new->map + old_size, 0, size - old_size);
354 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
355 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
361 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
363 struct mem_cgroup_per_node *pn;
364 struct memcg_shrinker_map *map;
367 if (mem_cgroup_is_root(memcg))
371 pn = mem_cgroup_nodeinfo(memcg, nid);
372 map = rcu_dereference_protected(pn->shrinker_map, true);
375 rcu_assign_pointer(pn->shrinker_map, NULL);
379 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
381 struct memcg_shrinker_map *map;
382 int nid, size, ret = 0;
384 if (mem_cgroup_is_root(memcg))
387 mutex_lock(&memcg_shrinker_map_mutex);
388 size = memcg_shrinker_map_size;
390 map = kvzalloc(sizeof(*map) + size, GFP_KERNEL);
392 memcg_free_shrinker_maps(memcg);
396 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
398 mutex_unlock(&memcg_shrinker_map_mutex);
403 int memcg_expand_shrinker_maps(int new_id)
405 int size, old_size, ret = 0;
406 struct mem_cgroup *memcg;
408 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
409 old_size = memcg_shrinker_map_size;
410 if (size <= old_size)
413 mutex_lock(&memcg_shrinker_map_mutex);
414 if (!root_mem_cgroup)
417 for_each_mem_cgroup(memcg) {
418 if (mem_cgroup_is_root(memcg))
420 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
426 memcg_shrinker_map_size = size;
427 mutex_unlock(&memcg_shrinker_map_mutex);
431 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
433 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
434 struct memcg_shrinker_map *map;
437 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
438 /* Pairs with smp mb in shrink_slab() */
439 smp_mb__before_atomic();
440 set_bit(shrinker_id, map->map);
446 * mem_cgroup_css_from_page - css of the memcg associated with a page
447 * @page: page of interest
449 * If memcg is bound to the default hierarchy, css of the memcg associated
450 * with @page is returned. The returned css remains associated with @page
451 * until it is released.
453 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
456 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
458 struct mem_cgroup *memcg;
460 memcg = page->mem_cgroup;
462 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
463 memcg = root_mem_cgroup;
469 * page_cgroup_ino - return inode number of the memcg a page is charged to
472 * Look up the closest online ancestor of the memory cgroup @page is charged to
473 * and return its inode number or 0 if @page is not charged to any cgroup. It
474 * is safe to call this function without holding a reference to @page.
476 * Note, this function is inherently racy, because there is nothing to prevent
477 * the cgroup inode from getting torn down and potentially reallocated a moment
478 * after page_cgroup_ino() returns, so it only should be used by callers that
479 * do not care (such as procfs interfaces).
481 ino_t page_cgroup_ino(struct page *page)
483 struct mem_cgroup *memcg;
484 unsigned long ino = 0;
487 if (PageHead(page) && PageSlab(page))
488 memcg = memcg_from_slab_page(page);
490 memcg = READ_ONCE(page->mem_cgroup);
491 while (memcg && !(memcg->css.flags & CSS_ONLINE))
492 memcg = parent_mem_cgroup(memcg);
494 ino = cgroup_ino(memcg->css.cgroup);
499 static struct mem_cgroup_per_node *
500 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
502 int nid = page_to_nid(page);
504 return memcg->nodeinfo[nid];
507 static struct mem_cgroup_tree_per_node *
508 soft_limit_tree_node(int nid)
510 return soft_limit_tree.rb_tree_per_node[nid];
513 static struct mem_cgroup_tree_per_node *
514 soft_limit_tree_from_page(struct page *page)
516 int nid = page_to_nid(page);
518 return soft_limit_tree.rb_tree_per_node[nid];
521 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
522 struct mem_cgroup_tree_per_node *mctz,
523 unsigned long new_usage_in_excess)
525 struct rb_node **p = &mctz->rb_root.rb_node;
526 struct rb_node *parent = NULL;
527 struct mem_cgroup_per_node *mz_node;
528 bool rightmost = true;
533 mz->usage_in_excess = new_usage_in_excess;
534 if (!mz->usage_in_excess)
538 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
540 if (mz->usage_in_excess < mz_node->usage_in_excess) {
546 * We can't avoid mem cgroups that are over their soft
547 * limit by the same amount
549 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
554 mctz->rb_rightmost = &mz->tree_node;
556 rb_link_node(&mz->tree_node, parent, p);
557 rb_insert_color(&mz->tree_node, &mctz->rb_root);
561 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
562 struct mem_cgroup_tree_per_node *mctz)
567 if (&mz->tree_node == mctz->rb_rightmost)
568 mctz->rb_rightmost = rb_prev(&mz->tree_node);
570 rb_erase(&mz->tree_node, &mctz->rb_root);
574 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
575 struct mem_cgroup_tree_per_node *mctz)
579 spin_lock_irqsave(&mctz->lock, flags);
580 __mem_cgroup_remove_exceeded(mz, mctz);
581 spin_unlock_irqrestore(&mctz->lock, flags);
584 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
586 unsigned long nr_pages = page_counter_read(&memcg->memory);
587 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
588 unsigned long excess = 0;
590 if (nr_pages > soft_limit)
591 excess = nr_pages - soft_limit;
596 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
598 unsigned long excess;
599 struct mem_cgroup_per_node *mz;
600 struct mem_cgroup_tree_per_node *mctz;
602 mctz = soft_limit_tree_from_page(page);
606 * Necessary to update all ancestors when hierarchy is used.
607 * because their event counter is not touched.
609 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
610 mz = mem_cgroup_page_nodeinfo(memcg, page);
611 excess = soft_limit_excess(memcg);
613 * We have to update the tree if mz is on RB-tree or
614 * mem is over its softlimit.
616 if (excess || mz->on_tree) {
619 spin_lock_irqsave(&mctz->lock, flags);
620 /* if on-tree, remove it */
622 __mem_cgroup_remove_exceeded(mz, mctz);
624 * Insert again. mz->usage_in_excess will be updated.
625 * If excess is 0, no tree ops.
627 __mem_cgroup_insert_exceeded(mz, mctz, excess);
628 spin_unlock_irqrestore(&mctz->lock, flags);
633 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
635 struct mem_cgroup_tree_per_node *mctz;
636 struct mem_cgroup_per_node *mz;
640 mz = mem_cgroup_nodeinfo(memcg, nid);
641 mctz = soft_limit_tree_node(nid);
643 mem_cgroup_remove_exceeded(mz, mctz);
647 static struct mem_cgroup_per_node *
648 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
650 struct mem_cgroup_per_node *mz;
654 if (!mctz->rb_rightmost)
655 goto done; /* Nothing to reclaim from */
657 mz = rb_entry(mctz->rb_rightmost,
658 struct mem_cgroup_per_node, tree_node);
660 * Remove the node now but someone else can add it back,
661 * we will to add it back at the end of reclaim to its correct
662 * position in the tree.
664 __mem_cgroup_remove_exceeded(mz, mctz);
665 if (!soft_limit_excess(mz->memcg) ||
666 !css_tryget_online(&mz->memcg->css))
672 static struct mem_cgroup_per_node *
673 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
675 struct mem_cgroup_per_node *mz;
677 spin_lock_irq(&mctz->lock);
678 mz = __mem_cgroup_largest_soft_limit_node(mctz);
679 spin_unlock_irq(&mctz->lock);
684 * __mod_memcg_state - update cgroup memory statistics
685 * @memcg: the memory cgroup
686 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
687 * @val: delta to add to the counter, can be negative
689 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
693 if (mem_cgroup_disabled())
696 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
697 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
698 struct mem_cgroup *mi;
701 * Batch local counters to keep them in sync with
702 * the hierarchical ones.
704 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
705 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
706 atomic_long_add(x, &mi->vmstats[idx]);
709 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
712 static struct mem_cgroup_per_node *
713 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
715 struct mem_cgroup *parent;
717 parent = parent_mem_cgroup(pn->memcg);
720 return mem_cgroup_nodeinfo(parent, nid);
724 * __mod_lruvec_state - update lruvec memory statistics
725 * @lruvec: the lruvec
726 * @idx: the stat item
727 * @val: delta to add to the counter, can be negative
729 * The lruvec is the intersection of the NUMA node and a cgroup. This
730 * function updates the all three counters that are affected by a
731 * change of state at this level: per-node, per-cgroup, per-lruvec.
733 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
736 pg_data_t *pgdat = lruvec_pgdat(lruvec);
737 struct mem_cgroup_per_node *pn;
738 struct mem_cgroup *memcg;
742 __mod_node_page_state(pgdat, idx, val);
744 if (mem_cgroup_disabled())
747 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
751 __mod_memcg_state(memcg, idx, val);
754 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
756 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
757 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
758 struct mem_cgroup_per_node *pi;
760 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
761 atomic_long_add(x, &pi->lruvec_stat[idx]);
764 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
767 void __mod_lruvec_slab_state(void *p, enum node_stat_item idx, int val)
769 struct page *page = virt_to_head_page(p);
770 pg_data_t *pgdat = page_pgdat(page);
771 struct mem_cgroup *memcg;
772 struct lruvec *lruvec;
775 memcg = memcg_from_slab_page(page);
777 /* Untracked pages have no memcg, no lruvec. Update only the node */
778 if (!memcg || memcg == root_mem_cgroup) {
779 __mod_node_page_state(pgdat, idx, val);
781 lruvec = mem_cgroup_lruvec(pgdat, memcg);
782 __mod_lruvec_state(lruvec, idx, val);
788 * __count_memcg_events - account VM events in a cgroup
789 * @memcg: the memory cgroup
790 * @idx: the event item
791 * @count: the number of events that occured
793 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
798 if (mem_cgroup_disabled())
801 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
802 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
803 struct mem_cgroup *mi;
806 * Batch local counters to keep them in sync with
807 * the hierarchical ones.
809 __this_cpu_add(memcg->vmstats_local->events[idx], x);
810 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
811 atomic_long_add(x, &mi->vmevents[idx]);
814 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
817 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
819 return atomic_long_read(&memcg->vmevents[event]);
822 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
827 for_each_possible_cpu(cpu)
828 x += per_cpu(memcg->vmstats_local->events[event], cpu);
832 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
834 bool compound, int nr_pages)
837 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
838 * counted as CACHE even if it's on ANON LRU.
841 __mod_memcg_state(memcg, MEMCG_RSS, nr_pages);
843 __mod_memcg_state(memcg, MEMCG_CACHE, nr_pages);
844 if (PageSwapBacked(page))
845 __mod_memcg_state(memcg, NR_SHMEM, nr_pages);
849 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
850 __mod_memcg_state(memcg, MEMCG_RSS_HUGE, nr_pages);
853 /* pagein of a big page is an event. So, ignore page size */
855 __count_memcg_events(memcg, PGPGIN, 1);
857 __count_memcg_events(memcg, PGPGOUT, 1);
858 nr_pages = -nr_pages; /* for event */
861 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
864 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
865 enum mem_cgroup_events_target target)
867 unsigned long val, next;
869 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
870 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
871 /* from time_after() in jiffies.h */
872 if ((long)(next - val) < 0) {
874 case MEM_CGROUP_TARGET_THRESH:
875 next = val + THRESHOLDS_EVENTS_TARGET;
877 case MEM_CGROUP_TARGET_SOFTLIMIT:
878 next = val + SOFTLIMIT_EVENTS_TARGET;
880 case MEM_CGROUP_TARGET_NUMAINFO:
881 next = val + NUMAINFO_EVENTS_TARGET;
886 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
893 * Check events in order.
896 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
898 /* threshold event is triggered in finer grain than soft limit */
899 if (unlikely(mem_cgroup_event_ratelimit(memcg,
900 MEM_CGROUP_TARGET_THRESH))) {
902 bool do_numainfo __maybe_unused;
904 do_softlimit = mem_cgroup_event_ratelimit(memcg,
905 MEM_CGROUP_TARGET_SOFTLIMIT);
907 do_numainfo = mem_cgroup_event_ratelimit(memcg,
908 MEM_CGROUP_TARGET_NUMAINFO);
910 mem_cgroup_threshold(memcg);
911 if (unlikely(do_softlimit))
912 mem_cgroup_update_tree(memcg, page);
914 if (unlikely(do_numainfo))
915 atomic_inc(&memcg->numainfo_events);
920 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
923 * mm_update_next_owner() may clear mm->owner to NULL
924 * if it races with swapoff, page migration, etc.
925 * So this can be called with p == NULL.
930 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
932 EXPORT_SYMBOL(mem_cgroup_from_task);
935 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
936 * @mm: mm from which memcg should be extracted. It can be NULL.
938 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
939 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
942 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
944 struct mem_cgroup *memcg;
946 if (mem_cgroup_disabled())
952 * Page cache insertions can happen withou an
953 * actual mm context, e.g. during disk probing
954 * on boot, loopback IO, acct() writes etc.
957 memcg = root_mem_cgroup;
959 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
960 if (unlikely(!memcg))
961 memcg = root_mem_cgroup;
963 } while (!css_tryget_online(&memcg->css));
967 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
970 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
971 * @page: page from which memcg should be extracted.
973 * Obtain a reference on page->memcg and returns it if successful. Otherwise
974 * root_mem_cgroup is returned.
976 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
978 struct mem_cgroup *memcg = page->mem_cgroup;
980 if (mem_cgroup_disabled())
984 if (!memcg || !css_tryget_online(&memcg->css))
985 memcg = root_mem_cgroup;
989 EXPORT_SYMBOL(get_mem_cgroup_from_page);
992 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
994 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
996 if (unlikely(current->active_memcg)) {
997 struct mem_cgroup *memcg = root_mem_cgroup;
1000 if (css_tryget_online(¤t->active_memcg->css))
1001 memcg = current->active_memcg;
1005 return get_mem_cgroup_from_mm(current->mm);
1009 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1010 * @root: hierarchy root
1011 * @prev: previously returned memcg, NULL on first invocation
1012 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1014 * Returns references to children of the hierarchy below @root, or
1015 * @root itself, or %NULL after a full round-trip.
1017 * Caller must pass the return value in @prev on subsequent
1018 * invocations for reference counting, or use mem_cgroup_iter_break()
1019 * to cancel a hierarchy walk before the round-trip is complete.
1021 * Reclaimers can specify a node and a priority level in @reclaim to
1022 * divide up the memcgs in the hierarchy among all concurrent
1023 * reclaimers operating on the same node and priority.
1025 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1026 struct mem_cgroup *prev,
1027 struct mem_cgroup_reclaim_cookie *reclaim)
1029 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1030 struct cgroup_subsys_state *css = NULL;
1031 struct mem_cgroup *memcg = NULL;
1032 struct mem_cgroup *pos = NULL;
1034 if (mem_cgroup_disabled())
1038 root = root_mem_cgroup;
1040 if (prev && !reclaim)
1043 if (!root->use_hierarchy && root != root_mem_cgroup) {
1052 struct mem_cgroup_per_node *mz;
1054 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1055 iter = &mz->iter[reclaim->priority];
1057 if (prev && reclaim->generation != iter->generation)
1061 pos = READ_ONCE(iter->position);
1062 if (!pos || css_tryget(&pos->css))
1065 * css reference reached zero, so iter->position will
1066 * be cleared by ->css_released. However, we should not
1067 * rely on this happening soon, because ->css_released
1068 * is called from a work queue, and by busy-waiting we
1069 * might block it. So we clear iter->position right
1072 (void)cmpxchg(&iter->position, pos, NULL);
1080 css = css_next_descendant_pre(css, &root->css);
1083 * Reclaimers share the hierarchy walk, and a
1084 * new one might jump in right at the end of
1085 * the hierarchy - make sure they see at least
1086 * one group and restart from the beginning.
1094 * Verify the css and acquire a reference. The root
1095 * is provided by the caller, so we know it's alive
1096 * and kicking, and don't take an extra reference.
1098 memcg = mem_cgroup_from_css(css);
1100 if (css == &root->css)
1103 if (css_tryget(css))
1111 * The position could have already been updated by a competing
1112 * thread, so check that the value hasn't changed since we read
1113 * it to avoid reclaiming from the same cgroup twice.
1115 (void)cmpxchg(&iter->position, pos, memcg);
1123 reclaim->generation = iter->generation;
1129 if (prev && prev != root)
1130 css_put(&prev->css);
1136 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1137 * @root: hierarchy root
1138 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1140 void mem_cgroup_iter_break(struct mem_cgroup *root,
1141 struct mem_cgroup *prev)
1144 root = root_mem_cgroup;
1145 if (prev && prev != root)
1146 css_put(&prev->css);
1149 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1150 struct mem_cgroup *dead_memcg)
1152 struct mem_cgroup_reclaim_iter *iter;
1153 struct mem_cgroup_per_node *mz;
1157 for_each_node(nid) {
1158 mz = mem_cgroup_nodeinfo(from, nid);
1159 for (i = 0; i <= DEF_PRIORITY; i++) {
1160 iter = &mz->iter[i];
1161 cmpxchg(&iter->position,
1167 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1169 struct mem_cgroup *memcg = dead_memcg;
1170 struct mem_cgroup *last;
1173 __invalidate_reclaim_iterators(memcg, dead_memcg);
1175 } while ((memcg = parent_mem_cgroup(memcg)));
1178 * When cgruop1 non-hierarchy mode is used,
1179 * parent_mem_cgroup() does not walk all the way up to the
1180 * cgroup root (root_mem_cgroup). So we have to handle
1181 * dead_memcg from cgroup root separately.
1183 if (last != root_mem_cgroup)
1184 __invalidate_reclaim_iterators(root_mem_cgroup,
1189 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1190 * @memcg: hierarchy root
1191 * @fn: function to call for each task
1192 * @arg: argument passed to @fn
1194 * This function iterates over tasks attached to @memcg or to any of its
1195 * descendants and calls @fn for each task. If @fn returns a non-zero
1196 * value, the function breaks the iteration loop and returns the value.
1197 * Otherwise, it will iterate over all tasks and return 0.
1199 * This function must not be called for the root memory cgroup.
1201 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1202 int (*fn)(struct task_struct *, void *), void *arg)
1204 struct mem_cgroup *iter;
1207 BUG_ON(memcg == root_mem_cgroup);
1209 for_each_mem_cgroup_tree(iter, memcg) {
1210 struct css_task_iter it;
1211 struct task_struct *task;
1213 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1214 while (!ret && (task = css_task_iter_next(&it)))
1215 ret = fn(task, arg);
1216 css_task_iter_end(&it);
1218 mem_cgroup_iter_break(memcg, iter);
1226 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1228 * @pgdat: pgdat of the page
1230 * This function is only safe when following the LRU page isolation
1231 * and putback protocol: the LRU lock must be held, and the page must
1232 * either be PageLRU() or the caller must have isolated/allocated it.
1234 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1236 struct mem_cgroup_per_node *mz;
1237 struct mem_cgroup *memcg;
1238 struct lruvec *lruvec;
1240 if (mem_cgroup_disabled()) {
1241 lruvec = &pgdat->lruvec;
1245 memcg = page->mem_cgroup;
1247 * Swapcache readahead pages are added to the LRU - and
1248 * possibly migrated - before they are charged.
1251 memcg = root_mem_cgroup;
1253 mz = mem_cgroup_page_nodeinfo(memcg, page);
1254 lruvec = &mz->lruvec;
1257 * Since a node can be onlined after the mem_cgroup was created,
1258 * we have to be prepared to initialize lruvec->zone here;
1259 * and if offlined then reonlined, we need to reinitialize it.
1261 if (unlikely(lruvec->pgdat != pgdat))
1262 lruvec->pgdat = pgdat;
1267 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1268 * @lruvec: mem_cgroup per zone lru vector
1269 * @lru: index of lru list the page is sitting on
1270 * @zid: zone id of the accounted pages
1271 * @nr_pages: positive when adding or negative when removing
1273 * This function must be called under lru_lock, just before a page is added
1274 * to or just after a page is removed from an lru list (that ordering being
1275 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1277 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1278 int zid, int nr_pages)
1280 struct mem_cgroup_per_node *mz;
1281 unsigned long *lru_size;
1284 if (mem_cgroup_disabled())
1287 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1288 lru_size = &mz->lru_zone_size[zid][lru];
1291 *lru_size += nr_pages;
1294 if (WARN_ONCE(size < 0,
1295 "%s(%p, %d, %d): lru_size %ld\n",
1296 __func__, lruvec, lru, nr_pages, size)) {
1302 *lru_size += nr_pages;
1306 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1307 * @memcg: the memory cgroup
1309 * Returns the maximum amount of memory @mem can be charged with, in
1312 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1314 unsigned long margin = 0;
1315 unsigned long count;
1316 unsigned long limit;
1318 count = page_counter_read(&memcg->memory);
1319 limit = READ_ONCE(memcg->memory.max);
1321 margin = limit - count;
1323 if (do_memsw_account()) {
1324 count = page_counter_read(&memcg->memsw);
1325 limit = READ_ONCE(memcg->memsw.max);
1327 margin = min(margin, limit - count);
1336 * A routine for checking "mem" is under move_account() or not.
1338 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1339 * moving cgroups. This is for waiting at high-memory pressure
1342 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1344 struct mem_cgroup *from;
1345 struct mem_cgroup *to;
1348 * Unlike task_move routines, we access mc.to, mc.from not under
1349 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1351 spin_lock(&mc.lock);
1357 ret = mem_cgroup_is_descendant(from, memcg) ||
1358 mem_cgroup_is_descendant(to, memcg);
1360 spin_unlock(&mc.lock);
1364 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1366 if (mc.moving_task && current != mc.moving_task) {
1367 if (mem_cgroup_under_move(memcg)) {
1369 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1370 /* moving charge context might have finished. */
1373 finish_wait(&mc.waitq, &wait);
1380 static char *memory_stat_format(struct mem_cgroup *memcg)
1385 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1390 * Provide statistics on the state of the memory subsystem as
1391 * well as cumulative event counters that show past behavior.
1393 * This list is ordered following a combination of these gradients:
1394 * 1) generic big picture -> specifics and details
1395 * 2) reflecting userspace activity -> reflecting kernel heuristics
1397 * Current memory state:
1400 seq_buf_printf(&s, "anon %llu\n",
1401 (u64)memcg_page_state(memcg, MEMCG_RSS) *
1403 seq_buf_printf(&s, "file %llu\n",
1404 (u64)memcg_page_state(memcg, MEMCG_CACHE) *
1406 seq_buf_printf(&s, "kernel_stack %llu\n",
1407 (u64)memcg_page_state(memcg, MEMCG_KERNEL_STACK_KB) *
1409 seq_buf_printf(&s, "slab %llu\n",
1410 (u64)(memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) +
1411 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE)) *
1413 seq_buf_printf(&s, "sock %llu\n",
1414 (u64)memcg_page_state(memcg, MEMCG_SOCK) *
1417 seq_buf_printf(&s, "shmem %llu\n",
1418 (u64)memcg_page_state(memcg, NR_SHMEM) *
1420 seq_buf_printf(&s, "file_mapped %llu\n",
1421 (u64)memcg_page_state(memcg, NR_FILE_MAPPED) *
1423 seq_buf_printf(&s, "file_dirty %llu\n",
1424 (u64)memcg_page_state(memcg, NR_FILE_DIRTY) *
1426 seq_buf_printf(&s, "file_writeback %llu\n",
1427 (u64)memcg_page_state(memcg, NR_WRITEBACK) *
1431 * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter
1432 * with the NR_ANON_THP vm counter, but right now it's a pain in the
1433 * arse because it requires migrating the work out of rmap to a place
1434 * where the page->mem_cgroup is set up and stable.
1436 seq_buf_printf(&s, "anon_thp %llu\n",
1437 (u64)memcg_page_state(memcg, MEMCG_RSS_HUGE) *
1440 for (i = 0; i < NR_LRU_LISTS; i++)
1441 seq_buf_printf(&s, "%s %llu\n", mem_cgroup_lru_names[i],
1442 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
1445 seq_buf_printf(&s, "slab_reclaimable %llu\n",
1446 (u64)memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) *
1448 seq_buf_printf(&s, "slab_unreclaimable %llu\n",
1449 (u64)memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE) *
1452 /* Accumulated memory events */
1454 seq_buf_printf(&s, "pgfault %lu\n", memcg_events(memcg, PGFAULT));
1455 seq_buf_printf(&s, "pgmajfault %lu\n", memcg_events(memcg, PGMAJFAULT));
1457 seq_buf_printf(&s, "workingset_refault %lu\n",
1458 memcg_page_state(memcg, WORKINGSET_REFAULT));
1459 seq_buf_printf(&s, "workingset_activate %lu\n",
1460 memcg_page_state(memcg, WORKINGSET_ACTIVATE));
1461 seq_buf_printf(&s, "workingset_nodereclaim %lu\n",
1462 memcg_page_state(memcg, WORKINGSET_NODERECLAIM));
1464 seq_buf_printf(&s, "pgrefill %lu\n", memcg_events(memcg, PGREFILL));
1465 seq_buf_printf(&s, "pgscan %lu\n",
1466 memcg_events(memcg, PGSCAN_KSWAPD) +
1467 memcg_events(memcg, PGSCAN_DIRECT));
1468 seq_buf_printf(&s, "pgsteal %lu\n",
1469 memcg_events(memcg, PGSTEAL_KSWAPD) +
1470 memcg_events(memcg, PGSTEAL_DIRECT));
1471 seq_buf_printf(&s, "pgactivate %lu\n", memcg_events(memcg, PGACTIVATE));
1472 seq_buf_printf(&s, "pgdeactivate %lu\n", memcg_events(memcg, PGDEACTIVATE));
1473 seq_buf_printf(&s, "pglazyfree %lu\n", memcg_events(memcg, PGLAZYFREE));
1474 seq_buf_printf(&s, "pglazyfreed %lu\n", memcg_events(memcg, PGLAZYFREED));
1476 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1477 seq_buf_printf(&s, "thp_fault_alloc %lu\n",
1478 memcg_events(memcg, THP_FAULT_ALLOC));
1479 seq_buf_printf(&s, "thp_collapse_alloc %lu\n",
1480 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1481 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1483 /* The above should easily fit into one page */
1484 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1489 #define K(x) ((x) << (PAGE_SHIFT-10))
1491 * mem_cgroup_print_oom_context: Print OOM information relevant to
1492 * memory controller.
1493 * @memcg: The memory cgroup that went over limit
1494 * @p: Task that is going to be killed
1496 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1499 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1504 pr_cont(",oom_memcg=");
1505 pr_cont_cgroup_path(memcg->css.cgroup);
1507 pr_cont(",global_oom");
1509 pr_cont(",task_memcg=");
1510 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1516 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1517 * memory controller.
1518 * @memcg: The memory cgroup that went over limit
1520 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1524 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1525 K((u64)page_counter_read(&memcg->memory)),
1526 K((u64)memcg->memory.max), memcg->memory.failcnt);
1527 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1528 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1529 K((u64)page_counter_read(&memcg->swap)),
1530 K((u64)memcg->swap.max), memcg->swap.failcnt);
1532 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1533 K((u64)page_counter_read(&memcg->memsw)),
1534 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1535 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1536 K((u64)page_counter_read(&memcg->kmem)),
1537 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1540 pr_info("Memory cgroup stats for ");
1541 pr_cont_cgroup_path(memcg->css.cgroup);
1543 buf = memory_stat_format(memcg);
1551 * Return the memory (and swap, if configured) limit for a memcg.
1553 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1557 max = memcg->memory.max;
1558 if (mem_cgroup_swappiness(memcg)) {
1559 unsigned long memsw_max;
1560 unsigned long swap_max;
1562 memsw_max = memcg->memsw.max;
1563 swap_max = memcg->swap.max;
1564 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1565 max = min(max + swap_max, memsw_max);
1570 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1573 struct oom_control oc = {
1577 .gfp_mask = gfp_mask,
1582 if (mutex_lock_killable(&oom_lock))
1585 * A few threads which were not waiting at mutex_lock_killable() can
1586 * fail to bail out. Therefore, check again after holding oom_lock.
1588 ret = should_force_charge() || out_of_memory(&oc);
1589 mutex_unlock(&oom_lock);
1593 #if MAX_NUMNODES > 1
1596 * test_mem_cgroup_node_reclaimable
1597 * @memcg: the target memcg
1598 * @nid: the node ID to be checked.
1599 * @noswap : specify true here if the user wants flle only information.
1601 * This function returns whether the specified memcg contains any
1602 * reclaimable pages on a node. Returns true if there are any reclaimable
1603 * pages in the node.
1605 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1606 int nid, bool noswap)
1608 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
1610 if (lruvec_page_state(lruvec, NR_INACTIVE_FILE) ||
1611 lruvec_page_state(lruvec, NR_ACTIVE_FILE))
1613 if (noswap || !total_swap_pages)
1615 if (lruvec_page_state(lruvec, NR_INACTIVE_ANON) ||
1616 lruvec_page_state(lruvec, NR_ACTIVE_ANON))
1623 * Always updating the nodemask is not very good - even if we have an empty
1624 * list or the wrong list here, we can start from some node and traverse all
1625 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1628 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1632 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1633 * pagein/pageout changes since the last update.
1635 if (!atomic_read(&memcg->numainfo_events))
1637 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1640 /* make a nodemask where this memcg uses memory from */
1641 memcg->scan_nodes = node_states[N_MEMORY];
1643 for_each_node_mask(nid, node_states[N_MEMORY]) {
1645 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1646 node_clear(nid, memcg->scan_nodes);
1649 atomic_set(&memcg->numainfo_events, 0);
1650 atomic_set(&memcg->numainfo_updating, 0);
1654 * Selecting a node where we start reclaim from. Because what we need is just
1655 * reducing usage counter, start from anywhere is O,K. Considering
1656 * memory reclaim from current node, there are pros. and cons.
1658 * Freeing memory from current node means freeing memory from a node which
1659 * we'll use or we've used. So, it may make LRU bad. And if several threads
1660 * hit limits, it will see a contention on a node. But freeing from remote
1661 * node means more costs for memory reclaim because of memory latency.
1663 * Now, we use round-robin. Better algorithm is welcomed.
1665 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1669 mem_cgroup_may_update_nodemask(memcg);
1670 node = memcg->last_scanned_node;
1672 node = next_node_in(node, memcg->scan_nodes);
1674 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1675 * last time it really checked all the LRUs due to rate limiting.
1676 * Fallback to the current node in that case for simplicity.
1678 if (unlikely(node == MAX_NUMNODES))
1679 node = numa_node_id();
1681 memcg->last_scanned_node = node;
1685 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1691 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1694 unsigned long *total_scanned)
1696 struct mem_cgroup *victim = NULL;
1699 unsigned long excess;
1700 unsigned long nr_scanned;
1701 struct mem_cgroup_reclaim_cookie reclaim = {
1706 excess = soft_limit_excess(root_memcg);
1709 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1714 * If we have not been able to reclaim
1715 * anything, it might because there are
1716 * no reclaimable pages under this hierarchy
1721 * We want to do more targeted reclaim.
1722 * excess >> 2 is not to excessive so as to
1723 * reclaim too much, nor too less that we keep
1724 * coming back to reclaim from this cgroup
1726 if (total >= (excess >> 2) ||
1727 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1732 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1733 pgdat, &nr_scanned);
1734 *total_scanned += nr_scanned;
1735 if (!soft_limit_excess(root_memcg))
1738 mem_cgroup_iter_break(root_memcg, victim);
1742 #ifdef CONFIG_LOCKDEP
1743 static struct lockdep_map memcg_oom_lock_dep_map = {
1744 .name = "memcg_oom_lock",
1748 static DEFINE_SPINLOCK(memcg_oom_lock);
1751 * Check OOM-Killer is already running under our hierarchy.
1752 * If someone is running, return false.
1754 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1756 struct mem_cgroup *iter, *failed = NULL;
1758 spin_lock(&memcg_oom_lock);
1760 for_each_mem_cgroup_tree(iter, memcg) {
1761 if (iter->oom_lock) {
1763 * this subtree of our hierarchy is already locked
1764 * so we cannot give a lock.
1767 mem_cgroup_iter_break(memcg, iter);
1770 iter->oom_lock = true;
1775 * OK, we failed to lock the whole subtree so we have
1776 * to clean up what we set up to the failing subtree
1778 for_each_mem_cgroup_tree(iter, memcg) {
1779 if (iter == failed) {
1780 mem_cgroup_iter_break(memcg, iter);
1783 iter->oom_lock = false;
1786 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1788 spin_unlock(&memcg_oom_lock);
1793 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1795 struct mem_cgroup *iter;
1797 spin_lock(&memcg_oom_lock);
1798 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1799 for_each_mem_cgroup_tree(iter, memcg)
1800 iter->oom_lock = false;
1801 spin_unlock(&memcg_oom_lock);
1804 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1806 struct mem_cgroup *iter;
1808 spin_lock(&memcg_oom_lock);
1809 for_each_mem_cgroup_tree(iter, memcg)
1811 spin_unlock(&memcg_oom_lock);
1814 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1816 struct mem_cgroup *iter;
1819 * When a new child is created while the hierarchy is under oom,
1820 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1822 spin_lock(&memcg_oom_lock);
1823 for_each_mem_cgroup_tree(iter, memcg)
1824 if (iter->under_oom > 0)
1826 spin_unlock(&memcg_oom_lock);
1829 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1831 struct oom_wait_info {
1832 struct mem_cgroup *memcg;
1833 wait_queue_entry_t wait;
1836 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1837 unsigned mode, int sync, void *arg)
1839 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1840 struct mem_cgroup *oom_wait_memcg;
1841 struct oom_wait_info *oom_wait_info;
1843 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1844 oom_wait_memcg = oom_wait_info->memcg;
1846 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1847 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1849 return autoremove_wake_function(wait, mode, sync, arg);
1852 static void memcg_oom_recover(struct mem_cgroup *memcg)
1855 * For the following lockless ->under_oom test, the only required
1856 * guarantee is that it must see the state asserted by an OOM when
1857 * this function is called as a result of userland actions
1858 * triggered by the notification of the OOM. This is trivially
1859 * achieved by invoking mem_cgroup_mark_under_oom() before
1860 * triggering notification.
1862 if (memcg && memcg->under_oom)
1863 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1873 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1875 enum oom_status ret;
1878 if (order > PAGE_ALLOC_COSTLY_ORDER)
1881 memcg_memory_event(memcg, MEMCG_OOM);
1884 * We are in the middle of the charge context here, so we
1885 * don't want to block when potentially sitting on a callstack
1886 * that holds all kinds of filesystem and mm locks.
1888 * cgroup1 allows disabling the OOM killer and waiting for outside
1889 * handling until the charge can succeed; remember the context and put
1890 * the task to sleep at the end of the page fault when all locks are
1893 * On the other hand, in-kernel OOM killer allows for an async victim
1894 * memory reclaim (oom_reaper) and that means that we are not solely
1895 * relying on the oom victim to make a forward progress and we can
1896 * invoke the oom killer here.
1898 * Please note that mem_cgroup_out_of_memory might fail to find a
1899 * victim and then we have to bail out from the charge path.
1901 if (memcg->oom_kill_disable) {
1902 if (!current->in_user_fault)
1904 css_get(&memcg->css);
1905 current->memcg_in_oom = memcg;
1906 current->memcg_oom_gfp_mask = mask;
1907 current->memcg_oom_order = order;
1912 mem_cgroup_mark_under_oom(memcg);
1914 locked = mem_cgroup_oom_trylock(memcg);
1917 mem_cgroup_oom_notify(memcg);
1919 mem_cgroup_unmark_under_oom(memcg);
1920 if (mem_cgroup_out_of_memory(memcg, mask, order))
1926 mem_cgroup_oom_unlock(memcg);
1932 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1933 * @handle: actually kill/wait or just clean up the OOM state
1935 * This has to be called at the end of a page fault if the memcg OOM
1936 * handler was enabled.
1938 * Memcg supports userspace OOM handling where failed allocations must
1939 * sleep on a waitqueue until the userspace task resolves the
1940 * situation. Sleeping directly in the charge context with all kinds
1941 * of locks held is not a good idea, instead we remember an OOM state
1942 * in the task and mem_cgroup_oom_synchronize() has to be called at
1943 * the end of the page fault to complete the OOM handling.
1945 * Returns %true if an ongoing memcg OOM situation was detected and
1946 * completed, %false otherwise.
1948 bool mem_cgroup_oom_synchronize(bool handle)
1950 struct mem_cgroup *memcg = current->memcg_in_oom;
1951 struct oom_wait_info owait;
1954 /* OOM is global, do not handle */
1961 owait.memcg = memcg;
1962 owait.wait.flags = 0;
1963 owait.wait.func = memcg_oom_wake_function;
1964 owait.wait.private = current;
1965 INIT_LIST_HEAD(&owait.wait.entry);
1967 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1968 mem_cgroup_mark_under_oom(memcg);
1970 locked = mem_cgroup_oom_trylock(memcg);
1973 mem_cgroup_oom_notify(memcg);
1975 if (locked && !memcg->oom_kill_disable) {
1976 mem_cgroup_unmark_under_oom(memcg);
1977 finish_wait(&memcg_oom_waitq, &owait.wait);
1978 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1979 current->memcg_oom_order);
1982 mem_cgroup_unmark_under_oom(memcg);
1983 finish_wait(&memcg_oom_waitq, &owait.wait);
1987 mem_cgroup_oom_unlock(memcg);
1989 * There is no guarantee that an OOM-lock contender
1990 * sees the wakeups triggered by the OOM kill
1991 * uncharges. Wake any sleepers explicitely.
1993 memcg_oom_recover(memcg);
1996 current->memcg_in_oom = NULL;
1997 css_put(&memcg->css);
2002 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2003 * @victim: task to be killed by the OOM killer
2004 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2006 * Returns a pointer to a memory cgroup, which has to be cleaned up
2007 * by killing all belonging OOM-killable tasks.
2009 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2011 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2012 struct mem_cgroup *oom_domain)
2014 struct mem_cgroup *oom_group = NULL;
2015 struct mem_cgroup *memcg;
2017 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2021 oom_domain = root_mem_cgroup;
2025 memcg = mem_cgroup_from_task(victim);
2026 if (memcg == root_mem_cgroup)
2030 * Traverse the memory cgroup hierarchy from the victim task's
2031 * cgroup up to the OOMing cgroup (or root) to find the
2032 * highest-level memory cgroup with oom.group set.
2034 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2035 if (memcg->oom_group)
2038 if (memcg == oom_domain)
2043 css_get(&oom_group->css);
2050 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2052 pr_info("Tasks in ");
2053 pr_cont_cgroup_path(memcg->css.cgroup);
2054 pr_cont(" are going to be killed due to memory.oom.group set\n");
2058 * lock_page_memcg - lock a page->mem_cgroup binding
2061 * This function protects unlocked LRU pages from being moved to
2064 * It ensures lifetime of the returned memcg. Caller is responsible
2065 * for the lifetime of the page; __unlock_page_memcg() is available
2066 * when @page might get freed inside the locked section.
2068 struct mem_cgroup *lock_page_memcg(struct page *page)
2070 struct mem_cgroup *memcg;
2071 unsigned long flags;
2074 * The RCU lock is held throughout the transaction. The fast
2075 * path can get away without acquiring the memcg->move_lock
2076 * because page moving starts with an RCU grace period.
2078 * The RCU lock also protects the memcg from being freed when
2079 * the page state that is going to change is the only thing
2080 * preventing the page itself from being freed. E.g. writeback
2081 * doesn't hold a page reference and relies on PG_writeback to
2082 * keep off truncation, migration and so forth.
2086 if (mem_cgroup_disabled())
2089 memcg = page->mem_cgroup;
2090 if (unlikely(!memcg))
2093 if (atomic_read(&memcg->moving_account) <= 0)
2096 spin_lock_irqsave(&memcg->move_lock, flags);
2097 if (memcg != page->mem_cgroup) {
2098 spin_unlock_irqrestore(&memcg->move_lock, flags);
2103 * When charge migration first begins, we can have locked and
2104 * unlocked page stat updates happening concurrently. Track
2105 * the task who has the lock for unlock_page_memcg().
2107 memcg->move_lock_task = current;
2108 memcg->move_lock_flags = flags;
2112 EXPORT_SYMBOL(lock_page_memcg);
2115 * __unlock_page_memcg - unlock and unpin a memcg
2118 * Unlock and unpin a memcg returned by lock_page_memcg().
2120 void __unlock_page_memcg(struct mem_cgroup *memcg)
2122 if (memcg && memcg->move_lock_task == current) {
2123 unsigned long flags = memcg->move_lock_flags;
2125 memcg->move_lock_task = NULL;
2126 memcg->move_lock_flags = 0;
2128 spin_unlock_irqrestore(&memcg->move_lock, flags);
2135 * unlock_page_memcg - unlock a page->mem_cgroup binding
2138 void unlock_page_memcg(struct page *page)
2140 __unlock_page_memcg(page->mem_cgroup);
2142 EXPORT_SYMBOL(unlock_page_memcg);
2144 struct memcg_stock_pcp {
2145 struct mem_cgroup *cached; /* this never be root cgroup */
2146 unsigned int nr_pages;
2147 struct work_struct work;
2148 unsigned long flags;
2149 #define FLUSHING_CACHED_CHARGE 0
2151 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2152 static DEFINE_MUTEX(percpu_charge_mutex);
2155 * consume_stock: Try to consume stocked charge on this cpu.
2156 * @memcg: memcg to consume from.
2157 * @nr_pages: how many pages to charge.
2159 * The charges will only happen if @memcg matches the current cpu's memcg
2160 * stock, and at least @nr_pages are available in that stock. Failure to
2161 * service an allocation will refill the stock.
2163 * returns true if successful, false otherwise.
2165 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2167 struct memcg_stock_pcp *stock;
2168 unsigned long flags;
2171 if (nr_pages > MEMCG_CHARGE_BATCH)
2174 local_irq_save(flags);
2176 stock = this_cpu_ptr(&memcg_stock);
2177 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2178 stock->nr_pages -= nr_pages;
2182 local_irq_restore(flags);
2188 * Returns stocks cached in percpu and reset cached information.
2190 static void drain_stock(struct memcg_stock_pcp *stock)
2192 struct mem_cgroup *old = stock->cached;
2194 if (stock->nr_pages) {
2195 page_counter_uncharge(&old->memory, stock->nr_pages);
2196 if (do_memsw_account())
2197 page_counter_uncharge(&old->memsw, stock->nr_pages);
2198 css_put_many(&old->css, stock->nr_pages);
2199 stock->nr_pages = 0;
2201 stock->cached = NULL;
2204 static void drain_local_stock(struct work_struct *dummy)
2206 struct memcg_stock_pcp *stock;
2207 unsigned long flags;
2210 * The only protection from memory hotplug vs. drain_stock races is
2211 * that we always operate on local CPU stock here with IRQ disabled
2213 local_irq_save(flags);
2215 stock = this_cpu_ptr(&memcg_stock);
2217 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2219 local_irq_restore(flags);
2223 * Cache charges(val) to local per_cpu area.
2224 * This will be consumed by consume_stock() function, later.
2226 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2228 struct memcg_stock_pcp *stock;
2229 unsigned long flags;
2231 local_irq_save(flags);
2233 stock = this_cpu_ptr(&memcg_stock);
2234 if (stock->cached != memcg) { /* reset if necessary */
2236 stock->cached = memcg;
2238 stock->nr_pages += nr_pages;
2240 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2243 local_irq_restore(flags);
2247 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2248 * of the hierarchy under it.
2250 static void drain_all_stock(struct mem_cgroup *root_memcg)
2254 /* If someone's already draining, avoid adding running more workers. */
2255 if (!mutex_trylock(&percpu_charge_mutex))
2258 * Notify other cpus that system-wide "drain" is running
2259 * We do not care about races with the cpu hotplug because cpu down
2260 * as well as workers from this path always operate on the local
2261 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2264 for_each_online_cpu(cpu) {
2265 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2266 struct mem_cgroup *memcg;
2270 memcg = stock->cached;
2271 if (memcg && stock->nr_pages &&
2272 mem_cgroup_is_descendant(memcg, root_memcg))
2277 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2279 drain_local_stock(&stock->work);
2281 schedule_work_on(cpu, &stock->work);
2285 mutex_unlock(&percpu_charge_mutex);
2288 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2290 struct memcg_stock_pcp *stock;
2291 struct mem_cgroup *memcg, *mi;
2293 stock = &per_cpu(memcg_stock, cpu);
2296 for_each_mem_cgroup(memcg) {
2299 for (i = 0; i < MEMCG_NR_STAT; i++) {
2303 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2305 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2306 atomic_long_add(x, &memcg->vmstats[i]);
2308 if (i >= NR_VM_NODE_STAT_ITEMS)
2311 for_each_node(nid) {
2312 struct mem_cgroup_per_node *pn;
2314 pn = mem_cgroup_nodeinfo(memcg, nid);
2315 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2318 atomic_long_add(x, &pn->lruvec_stat[i]);
2319 } while ((pn = parent_nodeinfo(pn, nid)));
2323 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2326 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2328 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2329 atomic_long_add(x, &memcg->vmevents[i]);
2336 static void reclaim_high(struct mem_cgroup *memcg,
2337 unsigned int nr_pages,
2341 if (page_counter_read(&memcg->memory) <= memcg->high)
2343 memcg_memory_event(memcg, MEMCG_HIGH);
2344 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2345 } while ((memcg = parent_mem_cgroup(memcg)));
2348 static void high_work_func(struct work_struct *work)
2350 struct mem_cgroup *memcg;
2352 memcg = container_of(work, struct mem_cgroup, high_work);
2353 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2357 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2358 * enough to still cause a significant slowdown in most cases, while still
2359 * allowing diagnostics and tracing to proceed without becoming stuck.
2361 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2364 * When calculating the delay, we use these either side of the exponentiation to
2365 * maintain precision and scale to a reasonable number of jiffies (see the table
2368 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2369 * overage ratio to a delay.
2370 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down down the
2371 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2372 * to produce a reasonable delay curve.
2374 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2375 * reasonable delay curve compared to precision-adjusted overage, not
2376 * penalising heavily at first, but still making sure that growth beyond the
2377 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2378 * example, with a high of 100 megabytes:
2380 * +-------+------------------------+
2381 * | usage | time to allocate in ms |
2382 * +-------+------------------------+
2404 * +-------+------------------------+
2406 #define MEMCG_DELAY_PRECISION_SHIFT 20
2407 #define MEMCG_DELAY_SCALING_SHIFT 14
2410 * Scheduled by try_charge() to be executed from the userland return path
2411 * and reclaims memory over the high limit.
2413 void mem_cgroup_handle_over_high(void)
2415 unsigned long usage, high, clamped_high;
2416 unsigned long pflags;
2417 unsigned long penalty_jiffies, overage;
2418 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2419 struct mem_cgroup *memcg;
2421 if (likely(!nr_pages))
2424 memcg = get_mem_cgroup_from_mm(current->mm);
2425 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2426 current->memcg_nr_pages_over_high = 0;
2429 * memory.high is breached and reclaim is unable to keep up. Throttle
2430 * allocators proactively to slow down excessive growth.
2432 * We use overage compared to memory.high to calculate the number of
2433 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2434 * fairly lenient on small overages, and increasingly harsh when the
2435 * memcg in question makes it clear that it has no intention of stopping
2436 * its crazy behaviour, so we exponentially increase the delay based on
2440 usage = page_counter_read(&memcg->memory);
2441 high = READ_ONCE(memcg->high);
2447 * Prevent division by 0 in overage calculation by acting as if it was a
2448 * threshold of 1 page
2450 clamped_high = max(high, 1UL);
2452 overage = div_u64((u64)(usage - high) << MEMCG_DELAY_PRECISION_SHIFT,
2455 penalty_jiffies = ((u64)overage * overage * HZ)
2456 >> (MEMCG_DELAY_PRECISION_SHIFT + MEMCG_DELAY_SCALING_SHIFT);
2459 * Factor in the task's own contribution to the overage, such that four
2460 * N-sized allocations are throttled approximately the same as one
2461 * 4N-sized allocation.
2463 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2464 * larger the current charge patch is than that.
2466 penalty_jiffies = penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2469 * Clamp the max delay per usermode return so as to still keep the
2470 * application moving forwards and also permit diagnostics, albeit
2473 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2476 * Don't sleep if the amount of jiffies this memcg owes us is so low
2477 * that it's not even worth doing, in an attempt to be nice to those who
2478 * go only a small amount over their memory.high value and maybe haven't
2479 * been aggressively reclaimed enough yet.
2481 if (penalty_jiffies <= HZ / 100)
2485 * If we exit early, we're guaranteed to die (since
2486 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2487 * need to account for any ill-begotten jiffies to pay them off later.
2489 psi_memstall_enter(&pflags);
2490 schedule_timeout_killable(penalty_jiffies);
2491 psi_memstall_leave(&pflags);
2494 css_put(&memcg->css);
2497 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2498 unsigned int nr_pages)
2500 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2501 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2502 struct mem_cgroup *mem_over_limit;
2503 struct page_counter *counter;
2504 unsigned long nr_reclaimed;
2505 bool may_swap = true;
2506 bool drained = false;
2507 enum oom_status oom_status;
2509 if (mem_cgroup_is_root(memcg))
2512 if (consume_stock(memcg, nr_pages))
2515 if (!do_memsw_account() ||
2516 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2517 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2519 if (do_memsw_account())
2520 page_counter_uncharge(&memcg->memsw, batch);
2521 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2523 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2527 if (batch > nr_pages) {
2533 * Unlike in global OOM situations, memcg is not in a physical
2534 * memory shortage. Allow dying and OOM-killed tasks to
2535 * bypass the last charges so that they can exit quickly and
2536 * free their memory.
2538 if (unlikely(should_force_charge()))
2542 * Prevent unbounded recursion when reclaim operations need to
2543 * allocate memory. This might exceed the limits temporarily,
2544 * but we prefer facilitating memory reclaim and getting back
2545 * under the limit over triggering OOM kills in these cases.
2547 if (unlikely(current->flags & PF_MEMALLOC))
2550 if (unlikely(task_in_memcg_oom(current)))
2553 if (!gfpflags_allow_blocking(gfp_mask))
2556 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2558 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2559 gfp_mask, may_swap);
2561 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2565 drain_all_stock(mem_over_limit);
2570 if (gfp_mask & __GFP_NORETRY)
2573 * Even though the limit is exceeded at this point, reclaim
2574 * may have been able to free some pages. Retry the charge
2575 * before killing the task.
2577 * Only for regular pages, though: huge pages are rather
2578 * unlikely to succeed so close to the limit, and we fall back
2579 * to regular pages anyway in case of failure.
2581 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2584 * At task move, charge accounts can be doubly counted. So, it's
2585 * better to wait until the end of task_move if something is going on.
2587 if (mem_cgroup_wait_acct_move(mem_over_limit))
2593 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2596 if (gfp_mask & __GFP_NOFAIL)
2599 if (fatal_signal_pending(current))
2603 * keep retrying as long as the memcg oom killer is able to make
2604 * a forward progress or bypass the charge if the oom killer
2605 * couldn't make any progress.
2607 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2608 get_order(nr_pages * PAGE_SIZE));
2609 switch (oom_status) {
2611 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2619 if (!(gfp_mask & __GFP_NOFAIL))
2623 * The allocation either can't fail or will lead to more memory
2624 * being freed very soon. Allow memory usage go over the limit
2625 * temporarily by force charging it.
2627 page_counter_charge(&memcg->memory, nr_pages);
2628 if (do_memsw_account())
2629 page_counter_charge(&memcg->memsw, nr_pages);
2630 css_get_many(&memcg->css, nr_pages);
2635 css_get_many(&memcg->css, batch);
2636 if (batch > nr_pages)
2637 refill_stock(memcg, batch - nr_pages);
2640 * If the hierarchy is above the normal consumption range, schedule
2641 * reclaim on returning to userland. We can perform reclaim here
2642 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2643 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2644 * not recorded as it most likely matches current's and won't
2645 * change in the meantime. As high limit is checked again before
2646 * reclaim, the cost of mismatch is negligible.
2649 if (page_counter_read(&memcg->memory) > memcg->high) {
2650 /* Don't bother a random interrupted task */
2651 if (in_interrupt()) {
2652 schedule_work(&memcg->high_work);
2655 current->memcg_nr_pages_over_high += batch;
2656 set_notify_resume(current);
2659 } while ((memcg = parent_mem_cgroup(memcg)));
2664 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2666 if (mem_cgroup_is_root(memcg))
2669 page_counter_uncharge(&memcg->memory, nr_pages);
2670 if (do_memsw_account())
2671 page_counter_uncharge(&memcg->memsw, nr_pages);
2673 css_put_many(&memcg->css, nr_pages);
2676 static void lock_page_lru(struct page *page, int *isolated)
2678 pg_data_t *pgdat = page_pgdat(page);
2680 spin_lock_irq(&pgdat->lru_lock);
2681 if (PageLRU(page)) {
2682 struct lruvec *lruvec;
2684 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2686 del_page_from_lru_list(page, lruvec, page_lru(page));
2692 static void unlock_page_lru(struct page *page, int isolated)
2694 pg_data_t *pgdat = page_pgdat(page);
2697 struct lruvec *lruvec;
2699 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2700 VM_BUG_ON_PAGE(PageLRU(page), page);
2702 add_page_to_lru_list(page, lruvec, page_lru(page));
2704 spin_unlock_irq(&pgdat->lru_lock);
2707 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2712 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2715 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2716 * may already be on some other mem_cgroup's LRU. Take care of it.
2719 lock_page_lru(page, &isolated);
2722 * Nobody should be changing or seriously looking at
2723 * page->mem_cgroup at this point:
2725 * - the page is uncharged
2727 * - the page is off-LRU
2729 * - an anonymous fault has exclusive page access, except for
2730 * a locked page table
2732 * - a page cache insertion, a swapin fault, or a migration
2733 * have the page locked
2735 page->mem_cgroup = memcg;
2738 unlock_page_lru(page, isolated);
2741 #ifdef CONFIG_MEMCG_KMEM
2742 static int memcg_alloc_cache_id(void)
2747 id = ida_simple_get(&memcg_cache_ida,
2748 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2752 if (id < memcg_nr_cache_ids)
2756 * There's no space for the new id in memcg_caches arrays,
2757 * so we have to grow them.
2759 down_write(&memcg_cache_ids_sem);
2761 size = 2 * (id + 1);
2762 if (size < MEMCG_CACHES_MIN_SIZE)
2763 size = MEMCG_CACHES_MIN_SIZE;
2764 else if (size > MEMCG_CACHES_MAX_SIZE)
2765 size = MEMCG_CACHES_MAX_SIZE;
2767 err = memcg_update_all_caches(size);
2769 err = memcg_update_all_list_lrus(size);
2771 memcg_nr_cache_ids = size;
2773 up_write(&memcg_cache_ids_sem);
2776 ida_simple_remove(&memcg_cache_ida, id);
2782 static void memcg_free_cache_id(int id)
2784 ida_simple_remove(&memcg_cache_ida, id);
2787 struct memcg_kmem_cache_create_work {
2788 struct mem_cgroup *memcg;
2789 struct kmem_cache *cachep;
2790 struct work_struct work;
2793 static void memcg_kmem_cache_create_func(struct work_struct *w)
2795 struct memcg_kmem_cache_create_work *cw =
2796 container_of(w, struct memcg_kmem_cache_create_work, work);
2797 struct mem_cgroup *memcg = cw->memcg;
2798 struct kmem_cache *cachep = cw->cachep;
2800 memcg_create_kmem_cache(memcg, cachep);
2802 css_put(&memcg->css);
2807 * Enqueue the creation of a per-memcg kmem_cache.
2809 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2810 struct kmem_cache *cachep)
2812 struct memcg_kmem_cache_create_work *cw;
2814 if (!css_tryget_online(&memcg->css))
2817 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2822 cw->cachep = cachep;
2823 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2825 queue_work(memcg_kmem_cache_wq, &cw->work);
2828 static inline bool memcg_kmem_bypass(void)
2830 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2836 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2837 * @cachep: the original global kmem cache
2839 * Return the kmem_cache we're supposed to use for a slab allocation.
2840 * We try to use the current memcg's version of the cache.
2842 * If the cache does not exist yet, if we are the first user of it, we
2843 * create it asynchronously in a workqueue and let the current allocation
2844 * go through with the original cache.
2846 * This function takes a reference to the cache it returns to assure it
2847 * won't get destroyed while we are working with it. Once the caller is
2848 * done with it, memcg_kmem_put_cache() must be called to release the
2851 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2853 struct mem_cgroup *memcg;
2854 struct kmem_cache *memcg_cachep;
2855 struct memcg_cache_array *arr;
2858 VM_BUG_ON(!is_root_cache(cachep));
2860 if (memcg_kmem_bypass())
2865 if (unlikely(current->active_memcg))
2866 memcg = current->active_memcg;
2868 memcg = mem_cgroup_from_task(current);
2870 if (!memcg || memcg == root_mem_cgroup)
2873 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2877 arr = rcu_dereference(cachep->memcg_params.memcg_caches);
2880 * Make sure we will access the up-to-date value. The code updating
2881 * memcg_caches issues a write barrier to match the data dependency
2882 * barrier inside READ_ONCE() (see memcg_create_kmem_cache()).
2884 memcg_cachep = READ_ONCE(arr->entries[kmemcg_id]);
2887 * If we are in a safe context (can wait, and not in interrupt
2888 * context), we could be be predictable and return right away.
2889 * This would guarantee that the allocation being performed
2890 * already belongs in the new cache.
2892 * However, there are some clashes that can arrive from locking.
2893 * For instance, because we acquire the slab_mutex while doing
2894 * memcg_create_kmem_cache, this means no further allocation
2895 * could happen with the slab_mutex held. So it's better to
2898 * If the memcg is dying or memcg_cache is about to be released,
2899 * don't bother creating new kmem_caches. Because memcg_cachep
2900 * is ZEROed as the fist step of kmem offlining, we don't need
2901 * percpu_ref_tryget_live() here. css_tryget_online() check in
2902 * memcg_schedule_kmem_cache_create() will prevent us from
2903 * creation of a new kmem_cache.
2905 if (unlikely(!memcg_cachep))
2906 memcg_schedule_kmem_cache_create(memcg, cachep);
2907 else if (percpu_ref_tryget(&memcg_cachep->memcg_params.refcnt))
2908 cachep = memcg_cachep;
2915 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2916 * @cachep: the cache returned by memcg_kmem_get_cache
2918 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2920 if (!is_root_cache(cachep))
2921 percpu_ref_put(&cachep->memcg_params.refcnt);
2925 * __memcg_kmem_charge_memcg: charge a kmem page
2926 * @page: page to charge
2927 * @gfp: reclaim mode
2928 * @order: allocation order
2929 * @memcg: memory cgroup to charge
2931 * Returns 0 on success, an error code on failure.
2933 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2934 struct mem_cgroup *memcg)
2936 unsigned int nr_pages = 1 << order;
2937 struct page_counter *counter;
2940 ret = try_charge(memcg, gfp, nr_pages);
2944 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2945 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2946 cancel_charge(memcg, nr_pages);
2953 * __memcg_kmem_charge: charge a kmem page to the current memory cgroup
2954 * @page: page to charge
2955 * @gfp: reclaim mode
2956 * @order: allocation order
2958 * Returns 0 on success, an error code on failure.
2960 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2962 struct mem_cgroup *memcg;
2965 if (memcg_kmem_bypass())
2968 memcg = get_mem_cgroup_from_current();
2969 if (!mem_cgroup_is_root(memcg)) {
2970 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2972 page->mem_cgroup = memcg;
2973 __SetPageKmemcg(page);
2976 css_put(&memcg->css);
2981 * __memcg_kmem_uncharge_memcg: uncharge a kmem page
2982 * @memcg: memcg to uncharge
2983 * @nr_pages: number of pages to uncharge
2985 void __memcg_kmem_uncharge_memcg(struct mem_cgroup *memcg,
2986 unsigned int nr_pages)
2988 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2989 page_counter_uncharge(&memcg->kmem, nr_pages);
2991 page_counter_uncharge(&memcg->memory, nr_pages);
2992 if (do_memsw_account())
2993 page_counter_uncharge(&memcg->memsw, nr_pages);
2996 * __memcg_kmem_uncharge: uncharge a kmem page
2997 * @page: page to uncharge
2998 * @order: allocation order
3000 void __memcg_kmem_uncharge(struct page *page, int order)
3002 struct mem_cgroup *memcg = page->mem_cgroup;
3003 unsigned int nr_pages = 1 << order;
3008 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3009 __memcg_kmem_uncharge_memcg(memcg, nr_pages);
3010 page->mem_cgroup = NULL;
3012 /* slab pages do not have PageKmemcg flag set */
3013 if (PageKmemcg(page))
3014 __ClearPageKmemcg(page);
3016 css_put_many(&memcg->css, nr_pages);
3018 #endif /* CONFIG_MEMCG_KMEM */
3020 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3023 * Because tail pages are not marked as "used", set it. We're under
3024 * pgdat->lru_lock and migration entries setup in all page mappings.
3026 void mem_cgroup_split_huge_fixup(struct page *head)
3030 if (mem_cgroup_disabled())
3033 for (i = 1; i < HPAGE_PMD_NR; i++)
3034 head[i].mem_cgroup = head->mem_cgroup;
3036 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
3038 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3040 #ifdef CONFIG_MEMCG_SWAP
3042 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3043 * @entry: swap entry to be moved
3044 * @from: mem_cgroup which the entry is moved from
3045 * @to: mem_cgroup which the entry is moved to
3047 * It succeeds only when the swap_cgroup's record for this entry is the same
3048 * as the mem_cgroup's id of @from.
3050 * Returns 0 on success, -EINVAL on failure.
3052 * The caller must have charged to @to, IOW, called page_counter_charge() about
3053 * both res and memsw, and called css_get().
3055 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3056 struct mem_cgroup *from, struct mem_cgroup *to)
3058 unsigned short old_id, new_id;
3060 old_id = mem_cgroup_id(from);
3061 new_id = mem_cgroup_id(to);
3063 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3064 mod_memcg_state(from, MEMCG_SWAP, -1);
3065 mod_memcg_state(to, MEMCG_SWAP, 1);
3071 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3072 struct mem_cgroup *from, struct mem_cgroup *to)
3078 static DEFINE_MUTEX(memcg_max_mutex);
3080 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3081 unsigned long max, bool memsw)
3083 bool enlarge = false;
3084 bool drained = false;
3086 bool limits_invariant;
3087 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3090 if (signal_pending(current)) {
3095 mutex_lock(&memcg_max_mutex);
3097 * Make sure that the new limit (memsw or memory limit) doesn't
3098 * break our basic invariant rule memory.max <= memsw.max.
3100 limits_invariant = memsw ? max >= memcg->memory.max :
3101 max <= memcg->memsw.max;
3102 if (!limits_invariant) {
3103 mutex_unlock(&memcg_max_mutex);
3107 if (max > counter->max)
3109 ret = page_counter_set_max(counter, max);
3110 mutex_unlock(&memcg_max_mutex);
3116 drain_all_stock(memcg);
3121 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3122 GFP_KERNEL, !memsw)) {
3128 if (!ret && enlarge)
3129 memcg_oom_recover(memcg);
3134 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3136 unsigned long *total_scanned)
3138 unsigned long nr_reclaimed = 0;
3139 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3140 unsigned long reclaimed;
3142 struct mem_cgroup_tree_per_node *mctz;
3143 unsigned long excess;
3144 unsigned long nr_scanned;
3149 mctz = soft_limit_tree_node(pgdat->node_id);
3152 * Do not even bother to check the largest node if the root
3153 * is empty. Do it lockless to prevent lock bouncing. Races
3154 * are acceptable as soft limit is best effort anyway.
3156 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3160 * This loop can run a while, specially if mem_cgroup's continuously
3161 * keep exceeding their soft limit and putting the system under
3168 mz = mem_cgroup_largest_soft_limit_node(mctz);
3173 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3174 gfp_mask, &nr_scanned);
3175 nr_reclaimed += reclaimed;
3176 *total_scanned += nr_scanned;
3177 spin_lock_irq(&mctz->lock);
3178 __mem_cgroup_remove_exceeded(mz, mctz);
3181 * If we failed to reclaim anything from this memory cgroup
3182 * it is time to move on to the next cgroup
3186 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3188 excess = soft_limit_excess(mz->memcg);
3190 * One school of thought says that we should not add
3191 * back the node to the tree if reclaim returns 0.
3192 * But our reclaim could return 0, simply because due
3193 * to priority we are exposing a smaller subset of
3194 * memory to reclaim from. Consider this as a longer
3197 /* If excess == 0, no tree ops */
3198 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3199 spin_unlock_irq(&mctz->lock);
3200 css_put(&mz->memcg->css);
3203 * Could not reclaim anything and there are no more
3204 * mem cgroups to try or we seem to be looping without
3205 * reclaiming anything.
3207 if (!nr_reclaimed &&
3209 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3211 } while (!nr_reclaimed);
3213 css_put(&next_mz->memcg->css);
3214 return nr_reclaimed;
3218 * Test whether @memcg has children, dead or alive. Note that this
3219 * function doesn't care whether @memcg has use_hierarchy enabled and
3220 * returns %true if there are child csses according to the cgroup
3221 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3223 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3228 ret = css_next_child(NULL, &memcg->css);
3234 * Reclaims as many pages from the given memcg as possible.
3236 * Caller is responsible for holding css reference for memcg.
3238 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3240 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3242 /* we call try-to-free pages for make this cgroup empty */
3243 lru_add_drain_all();
3245 drain_all_stock(memcg);
3247 /* try to free all pages in this cgroup */
3248 while (nr_retries && page_counter_read(&memcg->memory)) {
3251 if (signal_pending(current))
3254 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3258 /* maybe some writeback is necessary */
3259 congestion_wait(BLK_RW_ASYNC, HZ/10);
3267 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3268 char *buf, size_t nbytes,
3271 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3273 if (mem_cgroup_is_root(memcg))
3275 return mem_cgroup_force_empty(memcg) ?: nbytes;
3278 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3281 return mem_cgroup_from_css(css)->use_hierarchy;
3284 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3285 struct cftype *cft, u64 val)
3288 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3289 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3291 if (memcg->use_hierarchy == val)
3295 * If parent's use_hierarchy is set, we can't make any modifications
3296 * in the child subtrees. If it is unset, then the change can
3297 * occur, provided the current cgroup has no children.
3299 * For the root cgroup, parent_mem is NULL, we allow value to be
3300 * set if there are no children.
3302 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3303 (val == 1 || val == 0)) {
3304 if (!memcg_has_children(memcg))
3305 memcg->use_hierarchy = val;
3314 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3318 if (mem_cgroup_is_root(memcg)) {
3319 val = memcg_page_state(memcg, MEMCG_CACHE) +
3320 memcg_page_state(memcg, MEMCG_RSS);
3322 val += memcg_page_state(memcg, MEMCG_SWAP);
3325 val = page_counter_read(&memcg->memory);
3327 val = page_counter_read(&memcg->memsw);
3340 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3343 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3344 struct page_counter *counter;
3346 switch (MEMFILE_TYPE(cft->private)) {
3348 counter = &memcg->memory;
3351 counter = &memcg->memsw;
3354 counter = &memcg->kmem;
3357 counter = &memcg->tcpmem;
3363 switch (MEMFILE_ATTR(cft->private)) {
3365 if (counter == &memcg->memory)
3366 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3367 if (counter == &memcg->memsw)
3368 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3369 return (u64)page_counter_read(counter) * PAGE_SIZE;
3371 return (u64)counter->max * PAGE_SIZE;
3373 return (u64)counter->watermark * PAGE_SIZE;
3375 return counter->failcnt;
3376 case RES_SOFT_LIMIT:
3377 return (u64)memcg->soft_limit * PAGE_SIZE;
3383 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg, bool slab_only)
3385 unsigned long stat[MEMCG_NR_STAT];
3386 struct mem_cgroup *mi;
3388 int min_idx, max_idx;
3391 min_idx = NR_SLAB_RECLAIMABLE;
3392 max_idx = NR_SLAB_UNRECLAIMABLE;
3395 max_idx = MEMCG_NR_STAT;
3398 for (i = min_idx; i < max_idx; i++)
3401 for_each_online_cpu(cpu)
3402 for (i = min_idx; i < max_idx; i++)
3403 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3405 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3406 for (i = min_idx; i < max_idx; i++)
3407 atomic_long_add(stat[i], &mi->vmstats[i]);
3410 max_idx = NR_VM_NODE_STAT_ITEMS;
3412 for_each_node(node) {
3413 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3414 struct mem_cgroup_per_node *pi;
3416 for (i = min_idx; i < max_idx; i++)
3419 for_each_online_cpu(cpu)
3420 for (i = min_idx; i < max_idx; i++)
3422 pn->lruvec_stat_cpu->count[i], cpu);
3424 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3425 for (i = min_idx; i < max_idx; i++)
3426 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3430 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3432 unsigned long events[NR_VM_EVENT_ITEMS];
3433 struct mem_cgroup *mi;
3436 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3439 for_each_online_cpu(cpu)
3440 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3441 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3444 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3445 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3446 atomic_long_add(events[i], &mi->vmevents[i]);
3449 #ifdef CONFIG_MEMCG_KMEM
3450 static int memcg_online_kmem(struct mem_cgroup *memcg)
3454 if (cgroup_memory_nokmem)
3457 BUG_ON(memcg->kmemcg_id >= 0);
3458 BUG_ON(memcg->kmem_state);
3460 memcg_id = memcg_alloc_cache_id();
3464 static_branch_inc(&memcg_kmem_enabled_key);
3466 * A memory cgroup is considered kmem-online as soon as it gets
3467 * kmemcg_id. Setting the id after enabling static branching will
3468 * guarantee no one starts accounting before all call sites are
3471 memcg->kmemcg_id = memcg_id;
3472 memcg->kmem_state = KMEM_ONLINE;
3473 INIT_LIST_HEAD(&memcg->kmem_caches);
3478 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3480 struct cgroup_subsys_state *css;
3481 struct mem_cgroup *parent, *child;
3484 if (memcg->kmem_state != KMEM_ONLINE)
3487 * Clear the online state before clearing memcg_caches array
3488 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3489 * guarantees that no cache will be created for this cgroup
3490 * after we are done (see memcg_create_kmem_cache()).
3492 memcg->kmem_state = KMEM_ALLOCATED;
3494 parent = parent_mem_cgroup(memcg);
3496 parent = root_mem_cgroup;
3499 * Deactivate and reparent kmem_caches. Then flush percpu
3500 * slab statistics to have precise values at the parent and
3501 * all ancestor levels. It's required to keep slab stats
3502 * accurate after the reparenting of kmem_caches.
3504 memcg_deactivate_kmem_caches(memcg, parent);
3505 memcg_flush_percpu_vmstats(memcg, true);
3507 kmemcg_id = memcg->kmemcg_id;
3508 BUG_ON(kmemcg_id < 0);
3511 * Change kmemcg_id of this cgroup and all its descendants to the
3512 * parent's id, and then move all entries from this cgroup's list_lrus
3513 * to ones of the parent. After we have finished, all list_lrus
3514 * corresponding to this cgroup are guaranteed to remain empty. The
3515 * ordering is imposed by list_lru_node->lock taken by
3516 * memcg_drain_all_list_lrus().
3518 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3519 css_for_each_descendant_pre(css, &memcg->css) {
3520 child = mem_cgroup_from_css(css);
3521 BUG_ON(child->kmemcg_id != kmemcg_id);
3522 child->kmemcg_id = parent->kmemcg_id;
3523 if (!memcg->use_hierarchy)
3528 memcg_drain_all_list_lrus(kmemcg_id, parent);
3530 memcg_free_cache_id(kmemcg_id);
3533 static void memcg_free_kmem(struct mem_cgroup *memcg)
3535 /* css_alloc() failed, offlining didn't happen */
3536 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3537 memcg_offline_kmem(memcg);
3539 if (memcg->kmem_state == KMEM_ALLOCATED) {
3540 WARN_ON(!list_empty(&memcg->kmem_caches));
3541 static_branch_dec(&memcg_kmem_enabled_key);
3545 static int memcg_online_kmem(struct mem_cgroup *memcg)
3549 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3552 static void memcg_free_kmem(struct mem_cgroup *memcg)
3555 #endif /* CONFIG_MEMCG_KMEM */
3557 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3562 mutex_lock(&memcg_max_mutex);
3563 ret = page_counter_set_max(&memcg->kmem, max);
3564 mutex_unlock(&memcg_max_mutex);
3568 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3572 mutex_lock(&memcg_max_mutex);
3574 ret = page_counter_set_max(&memcg->tcpmem, max);
3578 if (!memcg->tcpmem_active) {
3580 * The active flag needs to be written after the static_key
3581 * update. This is what guarantees that the socket activation
3582 * function is the last one to run. See mem_cgroup_sk_alloc()
3583 * for details, and note that we don't mark any socket as
3584 * belonging to this memcg until that flag is up.
3586 * We need to do this, because static_keys will span multiple
3587 * sites, but we can't control their order. If we mark a socket
3588 * as accounted, but the accounting functions are not patched in
3589 * yet, we'll lose accounting.
3591 * We never race with the readers in mem_cgroup_sk_alloc(),
3592 * because when this value change, the code to process it is not
3595 static_branch_inc(&memcg_sockets_enabled_key);
3596 memcg->tcpmem_active = true;
3599 mutex_unlock(&memcg_max_mutex);
3604 * The user of this function is...
3607 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3608 char *buf, size_t nbytes, loff_t off)
3610 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3611 unsigned long nr_pages;
3614 buf = strstrip(buf);
3615 ret = page_counter_memparse(buf, "-1", &nr_pages);
3619 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3621 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3625 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3627 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3630 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3633 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3634 "Please report your usecase to linux-mm@kvack.org if you "
3635 "depend on this functionality.\n");
3636 ret = memcg_update_kmem_max(memcg, nr_pages);
3639 ret = memcg_update_tcp_max(memcg, nr_pages);
3643 case RES_SOFT_LIMIT:
3644 memcg->soft_limit = nr_pages;
3648 return ret ?: nbytes;
3651 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3652 size_t nbytes, loff_t off)
3654 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3655 struct page_counter *counter;
3657 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3659 counter = &memcg->memory;
3662 counter = &memcg->memsw;
3665 counter = &memcg->kmem;
3668 counter = &memcg->tcpmem;
3674 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3676 page_counter_reset_watermark(counter);
3679 counter->failcnt = 0;
3688 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3691 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3695 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3696 struct cftype *cft, u64 val)
3698 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3700 if (val & ~MOVE_MASK)
3704 * No kind of locking is needed in here, because ->can_attach() will
3705 * check this value once in the beginning of the process, and then carry
3706 * on with stale data. This means that changes to this value will only
3707 * affect task migrations starting after the change.
3709 memcg->move_charge_at_immigrate = val;
3713 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3714 struct cftype *cft, u64 val)
3722 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3723 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3724 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3726 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3727 int nid, unsigned int lru_mask)
3729 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
3730 unsigned long nr = 0;
3733 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3736 if (!(BIT(lru) & lru_mask))
3738 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3743 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3744 unsigned int lru_mask)
3746 unsigned long nr = 0;
3750 if (!(BIT(lru) & lru_mask))
3752 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3757 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3761 unsigned int lru_mask;
3764 static const struct numa_stat stats[] = {
3765 { "total", LRU_ALL },
3766 { "file", LRU_ALL_FILE },
3767 { "anon", LRU_ALL_ANON },
3768 { "unevictable", BIT(LRU_UNEVICTABLE) },
3770 const struct numa_stat *stat;
3773 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3775 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3776 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3777 seq_printf(m, "%s=%lu", stat->name, nr);
3778 for_each_node_state(nid, N_MEMORY) {
3779 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3781 seq_printf(m, " N%d=%lu", nid, nr);
3786 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3787 struct mem_cgroup *iter;
3790 for_each_mem_cgroup_tree(iter, memcg)
3791 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3792 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3793 for_each_node_state(nid, N_MEMORY) {
3795 for_each_mem_cgroup_tree(iter, memcg)
3796 nr += mem_cgroup_node_nr_lru_pages(
3797 iter, nid, stat->lru_mask);
3798 seq_printf(m, " N%d=%lu", nid, nr);
3805 #endif /* CONFIG_NUMA */
3807 static const unsigned int memcg1_stats[] = {
3818 static const char *const memcg1_stat_names[] = {
3829 /* Universal VM events cgroup1 shows, original sort order */
3830 static const unsigned int memcg1_events[] = {
3837 static const char *const memcg1_event_names[] = {
3844 static int memcg_stat_show(struct seq_file *m, void *v)
3846 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3847 unsigned long memory, memsw;
3848 struct mem_cgroup *mi;
3851 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3852 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3854 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3855 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3857 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3858 memcg_page_state_local(memcg, memcg1_stats[i]) *
3862 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3863 seq_printf(m, "%s %lu\n", memcg1_event_names[i],
3864 memcg_events_local(memcg, memcg1_events[i]));
3866 for (i = 0; i < NR_LRU_LISTS; i++)
3867 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3868 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3871 /* Hierarchical information */
3872 memory = memsw = PAGE_COUNTER_MAX;
3873 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3874 memory = min(memory, mi->memory.max);
3875 memsw = min(memsw, mi->memsw.max);
3877 seq_printf(m, "hierarchical_memory_limit %llu\n",
3878 (u64)memory * PAGE_SIZE);
3879 if (do_memsw_account())
3880 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3881 (u64)memsw * PAGE_SIZE);
3883 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3884 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3886 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3887 (u64)memcg_page_state(memcg, memcg1_stats[i]) *
3891 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3892 seq_printf(m, "total_%s %llu\n", memcg1_event_names[i],
3893 (u64)memcg_events(memcg, memcg1_events[i]));
3895 for (i = 0; i < NR_LRU_LISTS; i++)
3896 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i],
3897 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
3900 #ifdef CONFIG_DEBUG_VM
3903 struct mem_cgroup_per_node *mz;
3904 struct zone_reclaim_stat *rstat;
3905 unsigned long recent_rotated[2] = {0, 0};
3906 unsigned long recent_scanned[2] = {0, 0};
3908 for_each_online_pgdat(pgdat) {
3909 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3910 rstat = &mz->lruvec.reclaim_stat;
3912 recent_rotated[0] += rstat->recent_rotated[0];
3913 recent_rotated[1] += rstat->recent_rotated[1];
3914 recent_scanned[0] += rstat->recent_scanned[0];
3915 recent_scanned[1] += rstat->recent_scanned[1];
3917 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3918 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3919 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3920 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3927 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3930 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3932 return mem_cgroup_swappiness(memcg);
3935 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3936 struct cftype *cft, u64 val)
3938 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3944 memcg->swappiness = val;
3946 vm_swappiness = val;
3951 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3953 struct mem_cgroup_threshold_ary *t;
3954 unsigned long usage;
3959 t = rcu_dereference(memcg->thresholds.primary);
3961 t = rcu_dereference(memcg->memsw_thresholds.primary);
3966 usage = mem_cgroup_usage(memcg, swap);
3969 * current_threshold points to threshold just below or equal to usage.
3970 * If it's not true, a threshold was crossed after last
3971 * call of __mem_cgroup_threshold().
3973 i = t->current_threshold;
3976 * Iterate backward over array of thresholds starting from
3977 * current_threshold and check if a threshold is crossed.
3978 * If none of thresholds below usage is crossed, we read
3979 * only one element of the array here.
3981 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3982 eventfd_signal(t->entries[i].eventfd, 1);
3984 /* i = current_threshold + 1 */
3988 * Iterate forward over array of thresholds starting from
3989 * current_threshold+1 and check if a threshold is crossed.
3990 * If none of thresholds above usage is crossed, we read
3991 * only one element of the array here.
3993 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3994 eventfd_signal(t->entries[i].eventfd, 1);
3996 /* Update current_threshold */
3997 t->current_threshold = i - 1;
4002 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4005 __mem_cgroup_threshold(memcg, false);
4006 if (do_memsw_account())
4007 __mem_cgroup_threshold(memcg, true);
4009 memcg = parent_mem_cgroup(memcg);
4013 static int compare_thresholds(const void *a, const void *b)
4015 const struct mem_cgroup_threshold *_a = a;
4016 const struct mem_cgroup_threshold *_b = b;
4018 if (_a->threshold > _b->threshold)
4021 if (_a->threshold < _b->threshold)
4027 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4029 struct mem_cgroup_eventfd_list *ev;
4031 spin_lock(&memcg_oom_lock);
4033 list_for_each_entry(ev, &memcg->oom_notify, list)
4034 eventfd_signal(ev->eventfd, 1);
4036 spin_unlock(&memcg_oom_lock);
4040 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4042 struct mem_cgroup *iter;
4044 for_each_mem_cgroup_tree(iter, memcg)
4045 mem_cgroup_oom_notify_cb(iter);
4048 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4049 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4051 struct mem_cgroup_thresholds *thresholds;
4052 struct mem_cgroup_threshold_ary *new;
4053 unsigned long threshold;
4054 unsigned long usage;
4057 ret = page_counter_memparse(args, "-1", &threshold);
4061 mutex_lock(&memcg->thresholds_lock);
4064 thresholds = &memcg->thresholds;
4065 usage = mem_cgroup_usage(memcg, false);
4066 } else if (type == _MEMSWAP) {
4067 thresholds = &memcg->memsw_thresholds;
4068 usage = mem_cgroup_usage(memcg, true);
4072 /* Check if a threshold crossed before adding a new one */
4073 if (thresholds->primary)
4074 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4076 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4078 /* Allocate memory for new array of thresholds */
4079 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4086 /* Copy thresholds (if any) to new array */
4087 if (thresholds->primary) {
4088 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4089 sizeof(struct mem_cgroup_threshold));
4092 /* Add new threshold */
4093 new->entries[size - 1].eventfd = eventfd;
4094 new->entries[size - 1].threshold = threshold;
4096 /* Sort thresholds. Registering of new threshold isn't time-critical */
4097 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4098 compare_thresholds, NULL);
4100 /* Find current threshold */
4101 new->current_threshold = -1;
4102 for (i = 0; i < size; i++) {
4103 if (new->entries[i].threshold <= usage) {
4105 * new->current_threshold will not be used until
4106 * rcu_assign_pointer(), so it's safe to increment
4109 ++new->current_threshold;
4114 /* Free old spare buffer and save old primary buffer as spare */
4115 kfree(thresholds->spare);
4116 thresholds->spare = thresholds->primary;
4118 rcu_assign_pointer(thresholds->primary, new);
4120 /* To be sure that nobody uses thresholds */
4124 mutex_unlock(&memcg->thresholds_lock);
4129 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4130 struct eventfd_ctx *eventfd, const char *args)
4132 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4135 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4136 struct eventfd_ctx *eventfd, const char *args)
4138 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4141 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4142 struct eventfd_ctx *eventfd, enum res_type type)
4144 struct mem_cgroup_thresholds *thresholds;
4145 struct mem_cgroup_threshold_ary *new;
4146 unsigned long usage;
4149 mutex_lock(&memcg->thresholds_lock);
4152 thresholds = &memcg->thresholds;
4153 usage = mem_cgroup_usage(memcg, false);
4154 } else if (type == _MEMSWAP) {
4155 thresholds = &memcg->memsw_thresholds;
4156 usage = mem_cgroup_usage(memcg, true);
4160 if (!thresholds->primary)
4163 /* Check if a threshold crossed before removing */
4164 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4166 /* Calculate new number of threshold */
4168 for (i = 0; i < thresholds->primary->size; i++) {
4169 if (thresholds->primary->entries[i].eventfd != eventfd)
4173 new = thresholds->spare;
4175 /* Set thresholds array to NULL if we don't have thresholds */
4184 /* Copy thresholds and find current threshold */
4185 new->current_threshold = -1;
4186 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4187 if (thresholds->primary->entries[i].eventfd == eventfd)
4190 new->entries[j] = thresholds->primary->entries[i];
4191 if (new->entries[j].threshold <= usage) {
4193 * new->current_threshold will not be used
4194 * until rcu_assign_pointer(), so it's safe to increment
4197 ++new->current_threshold;
4203 /* Swap primary and spare array */
4204 thresholds->spare = thresholds->primary;
4206 rcu_assign_pointer(thresholds->primary, new);
4208 /* To be sure that nobody uses thresholds */
4211 /* If all events are unregistered, free the spare array */
4213 kfree(thresholds->spare);
4214 thresholds->spare = NULL;
4217 mutex_unlock(&memcg->thresholds_lock);
4220 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4221 struct eventfd_ctx *eventfd)
4223 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4226 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4227 struct eventfd_ctx *eventfd)
4229 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4232 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4233 struct eventfd_ctx *eventfd, const char *args)
4235 struct mem_cgroup_eventfd_list *event;
4237 event = kmalloc(sizeof(*event), GFP_KERNEL);
4241 spin_lock(&memcg_oom_lock);
4243 event->eventfd = eventfd;
4244 list_add(&event->list, &memcg->oom_notify);
4246 /* already in OOM ? */
4247 if (memcg->under_oom)
4248 eventfd_signal(eventfd, 1);
4249 spin_unlock(&memcg_oom_lock);
4254 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4255 struct eventfd_ctx *eventfd)
4257 struct mem_cgroup_eventfd_list *ev, *tmp;
4259 spin_lock(&memcg_oom_lock);
4261 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4262 if (ev->eventfd == eventfd) {
4263 list_del(&ev->list);
4268 spin_unlock(&memcg_oom_lock);
4271 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4273 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4275 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4276 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4277 seq_printf(sf, "oom_kill %lu\n",
4278 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4282 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4283 struct cftype *cft, u64 val)
4285 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4287 /* cannot set to root cgroup and only 0 and 1 are allowed */
4288 if (!css->parent || !((val == 0) || (val == 1)))
4291 memcg->oom_kill_disable = val;
4293 memcg_oom_recover(memcg);
4298 #ifdef CONFIG_CGROUP_WRITEBACK
4300 #include <trace/events/writeback.h>
4302 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4304 return wb_domain_init(&memcg->cgwb_domain, gfp);
4307 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4309 wb_domain_exit(&memcg->cgwb_domain);
4312 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4314 wb_domain_size_changed(&memcg->cgwb_domain);
4317 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4319 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4321 if (!memcg->css.parent)
4324 return &memcg->cgwb_domain;
4328 * idx can be of type enum memcg_stat_item or node_stat_item.
4329 * Keep in sync with memcg_exact_page().
4331 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4333 long x = atomic_long_read(&memcg->vmstats[idx]);
4336 for_each_online_cpu(cpu)
4337 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4344 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4345 * @wb: bdi_writeback in question
4346 * @pfilepages: out parameter for number of file pages
4347 * @pheadroom: out parameter for number of allocatable pages according to memcg
4348 * @pdirty: out parameter for number of dirty pages
4349 * @pwriteback: out parameter for number of pages under writeback
4351 * Determine the numbers of file, headroom, dirty, and writeback pages in
4352 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4353 * is a bit more involved.
4355 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4356 * headroom is calculated as the lowest headroom of itself and the
4357 * ancestors. Note that this doesn't consider the actual amount of
4358 * available memory in the system. The caller should further cap
4359 * *@pheadroom accordingly.
4361 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4362 unsigned long *pheadroom, unsigned long *pdirty,
4363 unsigned long *pwriteback)
4365 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4366 struct mem_cgroup *parent;
4368 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4370 /* this should eventually include NR_UNSTABLE_NFS */
4371 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4372 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4373 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4374 *pheadroom = PAGE_COUNTER_MAX;
4376 while ((parent = parent_mem_cgroup(memcg))) {
4377 unsigned long ceiling = min(memcg->memory.max, memcg->high);
4378 unsigned long used = page_counter_read(&memcg->memory);
4380 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4386 * Foreign dirty flushing
4388 * There's an inherent mismatch between memcg and writeback. The former
4389 * trackes ownership per-page while the latter per-inode. This was a
4390 * deliberate design decision because honoring per-page ownership in the
4391 * writeback path is complicated, may lead to higher CPU and IO overheads
4392 * and deemed unnecessary given that write-sharing an inode across
4393 * different cgroups isn't a common use-case.
4395 * Combined with inode majority-writer ownership switching, this works well
4396 * enough in most cases but there are some pathological cases. For
4397 * example, let's say there are two cgroups A and B which keep writing to
4398 * different but confined parts of the same inode. B owns the inode and
4399 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4400 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4401 * triggering background writeback. A will be slowed down without a way to
4402 * make writeback of the dirty pages happen.
4404 * Conditions like the above can lead to a cgroup getting repatedly and
4405 * severely throttled after making some progress after each
4406 * dirty_expire_interval while the underyling IO device is almost
4409 * Solving this problem completely requires matching the ownership tracking
4410 * granularities between memcg and writeback in either direction. However,
4411 * the more egregious behaviors can be avoided by simply remembering the
4412 * most recent foreign dirtying events and initiating remote flushes on
4413 * them when local writeback isn't enough to keep the memory clean enough.
4415 * The following two functions implement such mechanism. When a foreign
4416 * page - a page whose memcg and writeback ownerships don't match - is
4417 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4418 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4419 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4420 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4421 * foreign bdi_writebacks which haven't expired. Both the numbers of
4422 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4423 * limited to MEMCG_CGWB_FRN_CNT.
4425 * The mechanism only remembers IDs and doesn't hold any object references.
4426 * As being wrong occasionally doesn't matter, updates and accesses to the
4427 * records are lockless and racy.
4429 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4430 struct bdi_writeback *wb)
4432 struct mem_cgroup *memcg = page->mem_cgroup;
4433 struct memcg_cgwb_frn *frn;
4434 u64 now = get_jiffies_64();
4435 u64 oldest_at = now;
4439 trace_track_foreign_dirty(page, wb);
4442 * Pick the slot to use. If there is already a slot for @wb, keep
4443 * using it. If not replace the oldest one which isn't being
4446 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4447 frn = &memcg->cgwb_frn[i];
4448 if (frn->bdi_id == wb->bdi->id &&
4449 frn->memcg_id == wb->memcg_css->id)
4451 if (time_before64(frn->at, oldest_at) &&
4452 atomic_read(&frn->done.cnt) == 1) {
4454 oldest_at = frn->at;
4458 if (i < MEMCG_CGWB_FRN_CNT) {
4460 * Re-using an existing one. Update timestamp lazily to
4461 * avoid making the cacheline hot. We want them to be
4462 * reasonably up-to-date and significantly shorter than
4463 * dirty_expire_interval as that's what expires the record.
4464 * Use the shorter of 1s and dirty_expire_interval / 8.
4466 unsigned long update_intv =
4467 min_t(unsigned long, HZ,
4468 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4470 if (time_before64(frn->at, now - update_intv))
4472 } else if (oldest >= 0) {
4473 /* replace the oldest free one */
4474 frn = &memcg->cgwb_frn[oldest];
4475 frn->bdi_id = wb->bdi->id;
4476 frn->memcg_id = wb->memcg_css->id;
4481 /* issue foreign writeback flushes for recorded foreign dirtying events */
4482 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4484 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4485 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4486 u64 now = jiffies_64;
4489 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4490 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4493 * If the record is older than dirty_expire_interval,
4494 * writeback on it has already started. No need to kick it
4495 * off again. Also, don't start a new one if there's
4496 * already one in flight.
4498 if (time_after64(frn->at, now - intv) &&
4499 atomic_read(&frn->done.cnt) == 1) {
4501 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4502 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4503 WB_REASON_FOREIGN_FLUSH,
4509 #else /* CONFIG_CGROUP_WRITEBACK */
4511 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4516 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4520 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4524 #endif /* CONFIG_CGROUP_WRITEBACK */
4527 * DO NOT USE IN NEW FILES.
4529 * "cgroup.event_control" implementation.
4531 * This is way over-engineered. It tries to support fully configurable
4532 * events for each user. Such level of flexibility is completely
4533 * unnecessary especially in the light of the planned unified hierarchy.
4535 * Please deprecate this and replace with something simpler if at all
4540 * Unregister event and free resources.
4542 * Gets called from workqueue.
4544 static void memcg_event_remove(struct work_struct *work)
4546 struct mem_cgroup_event *event =
4547 container_of(work, struct mem_cgroup_event, remove);
4548 struct mem_cgroup *memcg = event->memcg;
4550 remove_wait_queue(event->wqh, &event->wait);
4552 event->unregister_event(memcg, event->eventfd);
4554 /* Notify userspace the event is going away. */
4555 eventfd_signal(event->eventfd, 1);
4557 eventfd_ctx_put(event->eventfd);
4559 css_put(&memcg->css);
4563 * Gets called on EPOLLHUP on eventfd when user closes it.
4565 * Called with wqh->lock held and interrupts disabled.
4567 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4568 int sync, void *key)
4570 struct mem_cgroup_event *event =
4571 container_of(wait, struct mem_cgroup_event, wait);
4572 struct mem_cgroup *memcg = event->memcg;
4573 __poll_t flags = key_to_poll(key);
4575 if (flags & EPOLLHUP) {
4577 * If the event has been detached at cgroup removal, we
4578 * can simply return knowing the other side will cleanup
4581 * We can't race against event freeing since the other
4582 * side will require wqh->lock via remove_wait_queue(),
4585 spin_lock(&memcg->event_list_lock);
4586 if (!list_empty(&event->list)) {
4587 list_del_init(&event->list);
4589 * We are in atomic context, but cgroup_event_remove()
4590 * may sleep, so we have to call it in workqueue.
4592 schedule_work(&event->remove);
4594 spin_unlock(&memcg->event_list_lock);
4600 static void memcg_event_ptable_queue_proc(struct file *file,
4601 wait_queue_head_t *wqh, poll_table *pt)
4603 struct mem_cgroup_event *event =
4604 container_of(pt, struct mem_cgroup_event, pt);
4607 add_wait_queue(wqh, &event->wait);
4611 * DO NOT USE IN NEW FILES.
4613 * Parse input and register new cgroup event handler.
4615 * Input must be in format '<event_fd> <control_fd> <args>'.
4616 * Interpretation of args is defined by control file implementation.
4618 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4619 char *buf, size_t nbytes, loff_t off)
4621 struct cgroup_subsys_state *css = of_css(of);
4622 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4623 struct mem_cgroup_event *event;
4624 struct cgroup_subsys_state *cfile_css;
4625 unsigned int efd, cfd;
4632 buf = strstrip(buf);
4634 efd = simple_strtoul(buf, &endp, 10);
4639 cfd = simple_strtoul(buf, &endp, 10);
4640 if ((*endp != ' ') && (*endp != '\0'))
4644 event = kzalloc(sizeof(*event), GFP_KERNEL);
4648 event->memcg = memcg;
4649 INIT_LIST_HEAD(&event->list);
4650 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4651 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4652 INIT_WORK(&event->remove, memcg_event_remove);
4660 event->eventfd = eventfd_ctx_fileget(efile.file);
4661 if (IS_ERR(event->eventfd)) {
4662 ret = PTR_ERR(event->eventfd);
4669 goto out_put_eventfd;
4672 /* the process need read permission on control file */
4673 /* AV: shouldn't we check that it's been opened for read instead? */
4674 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4679 * Determine the event callbacks and set them in @event. This used
4680 * to be done via struct cftype but cgroup core no longer knows
4681 * about these events. The following is crude but the whole thing
4682 * is for compatibility anyway.
4684 * DO NOT ADD NEW FILES.
4686 name = cfile.file->f_path.dentry->d_name.name;
4688 if (!strcmp(name, "memory.usage_in_bytes")) {
4689 event->register_event = mem_cgroup_usage_register_event;
4690 event->unregister_event = mem_cgroup_usage_unregister_event;
4691 } else if (!strcmp(name, "memory.oom_control")) {
4692 event->register_event = mem_cgroup_oom_register_event;
4693 event->unregister_event = mem_cgroup_oom_unregister_event;
4694 } else if (!strcmp(name, "memory.pressure_level")) {
4695 event->register_event = vmpressure_register_event;
4696 event->unregister_event = vmpressure_unregister_event;
4697 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4698 event->register_event = memsw_cgroup_usage_register_event;
4699 event->unregister_event = memsw_cgroup_usage_unregister_event;
4706 * Verify @cfile should belong to @css. Also, remaining events are
4707 * automatically removed on cgroup destruction but the removal is
4708 * asynchronous, so take an extra ref on @css.
4710 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4711 &memory_cgrp_subsys);
4713 if (IS_ERR(cfile_css))
4715 if (cfile_css != css) {
4720 ret = event->register_event(memcg, event->eventfd, buf);
4724 vfs_poll(efile.file, &event->pt);
4726 spin_lock(&memcg->event_list_lock);
4727 list_add(&event->list, &memcg->event_list);
4728 spin_unlock(&memcg->event_list_lock);
4740 eventfd_ctx_put(event->eventfd);
4749 static struct cftype mem_cgroup_legacy_files[] = {
4751 .name = "usage_in_bytes",
4752 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4753 .read_u64 = mem_cgroup_read_u64,
4756 .name = "max_usage_in_bytes",
4757 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4758 .write = mem_cgroup_reset,
4759 .read_u64 = mem_cgroup_read_u64,
4762 .name = "limit_in_bytes",
4763 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4764 .write = mem_cgroup_write,
4765 .read_u64 = mem_cgroup_read_u64,
4768 .name = "soft_limit_in_bytes",
4769 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4770 .write = mem_cgroup_write,
4771 .read_u64 = mem_cgroup_read_u64,
4775 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4776 .write = mem_cgroup_reset,
4777 .read_u64 = mem_cgroup_read_u64,
4781 .seq_show = memcg_stat_show,
4784 .name = "force_empty",
4785 .write = mem_cgroup_force_empty_write,
4788 .name = "use_hierarchy",
4789 .write_u64 = mem_cgroup_hierarchy_write,
4790 .read_u64 = mem_cgroup_hierarchy_read,
4793 .name = "cgroup.event_control", /* XXX: for compat */
4794 .write = memcg_write_event_control,
4795 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4798 .name = "swappiness",
4799 .read_u64 = mem_cgroup_swappiness_read,
4800 .write_u64 = mem_cgroup_swappiness_write,
4803 .name = "move_charge_at_immigrate",
4804 .read_u64 = mem_cgroup_move_charge_read,
4805 .write_u64 = mem_cgroup_move_charge_write,
4808 .name = "oom_control",
4809 .seq_show = mem_cgroup_oom_control_read,
4810 .write_u64 = mem_cgroup_oom_control_write,
4811 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4814 .name = "pressure_level",
4818 .name = "numa_stat",
4819 .seq_show = memcg_numa_stat_show,
4823 .name = "kmem.limit_in_bytes",
4824 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4825 .write = mem_cgroup_write,
4826 .read_u64 = mem_cgroup_read_u64,
4829 .name = "kmem.usage_in_bytes",
4830 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4831 .read_u64 = mem_cgroup_read_u64,
4834 .name = "kmem.failcnt",
4835 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4836 .write = mem_cgroup_reset,
4837 .read_u64 = mem_cgroup_read_u64,
4840 .name = "kmem.max_usage_in_bytes",
4841 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4842 .write = mem_cgroup_reset,
4843 .read_u64 = mem_cgroup_read_u64,
4845 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4847 .name = "kmem.slabinfo",
4848 .seq_start = memcg_slab_start,
4849 .seq_next = memcg_slab_next,
4850 .seq_stop = memcg_slab_stop,
4851 .seq_show = memcg_slab_show,
4855 .name = "kmem.tcp.limit_in_bytes",
4856 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4857 .write = mem_cgroup_write,
4858 .read_u64 = mem_cgroup_read_u64,
4861 .name = "kmem.tcp.usage_in_bytes",
4862 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4863 .read_u64 = mem_cgroup_read_u64,
4866 .name = "kmem.tcp.failcnt",
4867 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4868 .write = mem_cgroup_reset,
4869 .read_u64 = mem_cgroup_read_u64,
4872 .name = "kmem.tcp.max_usage_in_bytes",
4873 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4874 .write = mem_cgroup_reset,
4875 .read_u64 = mem_cgroup_read_u64,
4877 { }, /* terminate */
4881 * Private memory cgroup IDR
4883 * Swap-out records and page cache shadow entries need to store memcg
4884 * references in constrained space, so we maintain an ID space that is
4885 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4886 * memory-controlled cgroups to 64k.
4888 * However, there usually are many references to the oflline CSS after
4889 * the cgroup has been destroyed, such as page cache or reclaimable
4890 * slab objects, that don't need to hang on to the ID. We want to keep
4891 * those dead CSS from occupying IDs, or we might quickly exhaust the
4892 * relatively small ID space and prevent the creation of new cgroups
4893 * even when there are much fewer than 64k cgroups - possibly none.
4895 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4896 * be freed and recycled when it's no longer needed, which is usually
4897 * when the CSS is offlined.
4899 * The only exception to that are records of swapped out tmpfs/shmem
4900 * pages that need to be attributed to live ancestors on swapin. But
4901 * those references are manageable from userspace.
4904 static DEFINE_IDR(mem_cgroup_idr);
4906 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4908 if (memcg->id.id > 0) {
4909 idr_remove(&mem_cgroup_idr, memcg->id.id);
4914 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4916 refcount_add(n, &memcg->id.ref);
4919 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4921 if (refcount_sub_and_test(n, &memcg->id.ref)) {
4922 mem_cgroup_id_remove(memcg);
4924 /* Memcg ID pins CSS */
4925 css_put(&memcg->css);
4929 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4931 mem_cgroup_id_put_many(memcg, 1);
4935 * mem_cgroup_from_id - look up a memcg from a memcg id
4936 * @id: the memcg id to look up
4938 * Caller must hold rcu_read_lock().
4940 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4942 WARN_ON_ONCE(!rcu_read_lock_held());
4943 return idr_find(&mem_cgroup_idr, id);
4946 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4948 struct mem_cgroup_per_node *pn;
4951 * This routine is called against possible nodes.
4952 * But it's BUG to call kmalloc() against offline node.
4954 * TODO: this routine can waste much memory for nodes which will
4955 * never be onlined. It's better to use memory hotplug callback
4958 if (!node_state(node, N_NORMAL_MEMORY))
4960 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4964 pn->lruvec_stat_local = alloc_percpu(struct lruvec_stat);
4965 if (!pn->lruvec_stat_local) {
4970 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
4971 if (!pn->lruvec_stat_cpu) {
4972 free_percpu(pn->lruvec_stat_local);
4977 lruvec_init(&pn->lruvec);
4978 pn->usage_in_excess = 0;
4979 pn->on_tree = false;
4982 memcg->nodeinfo[node] = pn;
4986 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4988 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
4993 free_percpu(pn->lruvec_stat_cpu);
4994 free_percpu(pn->lruvec_stat_local);
4998 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5003 * Flush percpu vmstats and vmevents to guarantee the value correctness
5004 * on parent's and all ancestor levels.
5006 memcg_flush_percpu_vmstats(memcg, false);
5007 memcg_flush_percpu_vmevents(memcg);
5009 free_mem_cgroup_per_node_info(memcg, node);
5010 free_percpu(memcg->vmstats_percpu);
5011 free_percpu(memcg->vmstats_local);
5015 static void mem_cgroup_free(struct mem_cgroup *memcg)
5017 memcg_wb_domain_exit(memcg);
5018 __mem_cgroup_free(memcg);
5021 static struct mem_cgroup *mem_cgroup_alloc(void)
5023 struct mem_cgroup *memcg;
5026 int __maybe_unused i;
5028 size = sizeof(struct mem_cgroup);
5029 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5031 memcg = kzalloc(size, GFP_KERNEL);
5035 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5036 1, MEM_CGROUP_ID_MAX,
5038 if (memcg->id.id < 0)
5041 memcg->vmstats_local = alloc_percpu(struct memcg_vmstats_percpu);
5042 if (!memcg->vmstats_local)
5045 memcg->vmstats_percpu = alloc_percpu(struct memcg_vmstats_percpu);
5046 if (!memcg->vmstats_percpu)
5050 if (alloc_mem_cgroup_per_node_info(memcg, node))
5053 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5056 INIT_WORK(&memcg->high_work, high_work_func);
5057 memcg->last_scanned_node = MAX_NUMNODES;
5058 INIT_LIST_HEAD(&memcg->oom_notify);
5059 mutex_init(&memcg->thresholds_lock);
5060 spin_lock_init(&memcg->move_lock);
5061 vmpressure_init(&memcg->vmpressure);
5062 INIT_LIST_HEAD(&memcg->event_list);
5063 spin_lock_init(&memcg->event_list_lock);
5064 memcg->socket_pressure = jiffies;
5065 #ifdef CONFIG_MEMCG_KMEM
5066 memcg->kmemcg_id = -1;
5068 #ifdef CONFIG_CGROUP_WRITEBACK
5069 INIT_LIST_HEAD(&memcg->cgwb_list);
5070 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5071 memcg->cgwb_frn[i].done =
5072 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5074 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5075 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5076 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5077 memcg->deferred_split_queue.split_queue_len = 0;
5079 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5082 mem_cgroup_id_remove(memcg);
5083 __mem_cgroup_free(memcg);
5087 static struct cgroup_subsys_state * __ref
5088 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5090 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5091 struct mem_cgroup *memcg;
5092 long error = -ENOMEM;
5094 memcg = mem_cgroup_alloc();
5096 return ERR_PTR(error);
5098 memcg->high = PAGE_COUNTER_MAX;
5099 memcg->soft_limit = PAGE_COUNTER_MAX;
5101 memcg->swappiness = mem_cgroup_swappiness(parent);
5102 memcg->oom_kill_disable = parent->oom_kill_disable;
5104 if (parent && parent->use_hierarchy) {
5105 memcg->use_hierarchy = true;
5106 page_counter_init(&memcg->memory, &parent->memory);
5107 page_counter_init(&memcg->swap, &parent->swap);
5108 page_counter_init(&memcg->memsw, &parent->memsw);
5109 page_counter_init(&memcg->kmem, &parent->kmem);
5110 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5112 page_counter_init(&memcg->memory, NULL);
5113 page_counter_init(&memcg->swap, NULL);
5114 page_counter_init(&memcg->memsw, NULL);
5115 page_counter_init(&memcg->kmem, NULL);
5116 page_counter_init(&memcg->tcpmem, NULL);
5118 * Deeper hierachy with use_hierarchy == false doesn't make
5119 * much sense so let cgroup subsystem know about this
5120 * unfortunate state in our controller.
5122 if (parent != root_mem_cgroup)
5123 memory_cgrp_subsys.broken_hierarchy = true;
5126 /* The following stuff does not apply to the root */
5128 #ifdef CONFIG_MEMCG_KMEM
5129 INIT_LIST_HEAD(&memcg->kmem_caches);
5131 root_mem_cgroup = memcg;
5135 error = memcg_online_kmem(memcg);
5139 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5140 static_branch_inc(&memcg_sockets_enabled_key);
5144 mem_cgroup_id_remove(memcg);
5145 mem_cgroup_free(memcg);
5146 return ERR_PTR(-ENOMEM);
5149 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5151 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5154 * A memcg must be visible for memcg_expand_shrinker_maps()
5155 * by the time the maps are allocated. So, we allocate maps
5156 * here, when for_each_mem_cgroup() can't skip it.
5158 if (memcg_alloc_shrinker_maps(memcg)) {
5159 mem_cgroup_id_remove(memcg);
5163 /* Online state pins memcg ID, memcg ID pins CSS */
5164 refcount_set(&memcg->id.ref, 1);
5169 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5171 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5172 struct mem_cgroup_event *event, *tmp;
5175 * Unregister events and notify userspace.
5176 * Notify userspace about cgroup removing only after rmdir of cgroup
5177 * directory to avoid race between userspace and kernelspace.
5179 spin_lock(&memcg->event_list_lock);
5180 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5181 list_del_init(&event->list);
5182 schedule_work(&event->remove);
5184 spin_unlock(&memcg->event_list_lock);
5186 page_counter_set_min(&memcg->memory, 0);
5187 page_counter_set_low(&memcg->memory, 0);
5189 memcg_offline_kmem(memcg);
5190 wb_memcg_offline(memcg);
5192 drain_all_stock(memcg);
5194 mem_cgroup_id_put(memcg);
5197 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5199 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5201 invalidate_reclaim_iterators(memcg);
5204 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5206 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5207 int __maybe_unused i;
5209 #ifdef CONFIG_CGROUP_WRITEBACK
5210 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5211 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5213 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5214 static_branch_dec(&memcg_sockets_enabled_key);
5216 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5217 static_branch_dec(&memcg_sockets_enabled_key);
5219 vmpressure_cleanup(&memcg->vmpressure);
5220 cancel_work_sync(&memcg->high_work);
5221 mem_cgroup_remove_from_trees(memcg);
5222 memcg_free_shrinker_maps(memcg);
5223 memcg_free_kmem(memcg);
5224 mem_cgroup_free(memcg);
5228 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5229 * @css: the target css
5231 * Reset the states of the mem_cgroup associated with @css. This is
5232 * invoked when the userland requests disabling on the default hierarchy
5233 * but the memcg is pinned through dependency. The memcg should stop
5234 * applying policies and should revert to the vanilla state as it may be
5235 * made visible again.
5237 * The current implementation only resets the essential configurations.
5238 * This needs to be expanded to cover all the visible parts.
5240 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5242 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5244 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5245 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5246 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
5247 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5248 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5249 page_counter_set_min(&memcg->memory, 0);
5250 page_counter_set_low(&memcg->memory, 0);
5251 memcg->high = PAGE_COUNTER_MAX;
5252 memcg->soft_limit = PAGE_COUNTER_MAX;
5253 memcg_wb_domain_size_changed(memcg);
5257 /* Handlers for move charge at task migration. */
5258 static int mem_cgroup_do_precharge(unsigned long count)
5262 /* Try a single bulk charge without reclaim first, kswapd may wake */
5263 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5265 mc.precharge += count;
5269 /* Try charges one by one with reclaim, but do not retry */
5271 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5285 enum mc_target_type {
5292 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5293 unsigned long addr, pte_t ptent)
5295 struct page *page = vm_normal_page(vma, addr, ptent);
5297 if (!page || !page_mapped(page))
5299 if (PageAnon(page)) {
5300 if (!(mc.flags & MOVE_ANON))
5303 if (!(mc.flags & MOVE_FILE))
5306 if (!get_page_unless_zero(page))
5312 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5313 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5314 pte_t ptent, swp_entry_t *entry)
5316 struct page *page = NULL;
5317 swp_entry_t ent = pte_to_swp_entry(ptent);
5319 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
5323 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5324 * a device and because they are not accessible by CPU they are store
5325 * as special swap entry in the CPU page table.
5327 if (is_device_private_entry(ent)) {
5328 page = device_private_entry_to_page(ent);
5330 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5331 * a refcount of 1 when free (unlike normal page)
5333 if (!page_ref_add_unless(page, 1, 1))
5339 * Because lookup_swap_cache() updates some statistics counter,
5340 * we call find_get_page() with swapper_space directly.
5342 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5343 if (do_memsw_account())
5344 entry->val = ent.val;
5349 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5350 pte_t ptent, swp_entry_t *entry)
5356 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5357 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5359 struct page *page = NULL;
5360 struct address_space *mapping;
5363 if (!vma->vm_file) /* anonymous vma */
5365 if (!(mc.flags & MOVE_FILE))
5368 mapping = vma->vm_file->f_mapping;
5369 pgoff = linear_page_index(vma, addr);
5371 /* page is moved even if it's not RSS of this task(page-faulted). */
5373 /* shmem/tmpfs may report page out on swap: account for that too. */
5374 if (shmem_mapping(mapping)) {
5375 page = find_get_entry(mapping, pgoff);
5376 if (xa_is_value(page)) {
5377 swp_entry_t swp = radix_to_swp_entry(page);
5378 if (do_memsw_account())
5380 page = find_get_page(swap_address_space(swp),
5384 page = find_get_page(mapping, pgoff);
5386 page = find_get_page(mapping, pgoff);
5392 * mem_cgroup_move_account - move account of the page
5394 * @compound: charge the page as compound or small page
5395 * @from: mem_cgroup which the page is moved from.
5396 * @to: mem_cgroup which the page is moved to. @from != @to.
5398 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5400 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5403 static int mem_cgroup_move_account(struct page *page,
5405 struct mem_cgroup *from,
5406 struct mem_cgroup *to)
5408 unsigned long flags;
5409 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5413 VM_BUG_ON(from == to);
5414 VM_BUG_ON_PAGE(PageLRU(page), page);
5415 VM_BUG_ON(compound && !PageTransHuge(page));
5418 * Prevent mem_cgroup_migrate() from looking at
5419 * page->mem_cgroup of its source page while we change it.
5422 if (!trylock_page(page))
5426 if (page->mem_cgroup != from)
5429 anon = PageAnon(page);
5431 spin_lock_irqsave(&from->move_lock, flags);
5433 if (!anon && page_mapped(page)) {
5434 __mod_memcg_state(from, NR_FILE_MAPPED, -nr_pages);
5435 __mod_memcg_state(to, NR_FILE_MAPPED, nr_pages);
5439 * move_lock grabbed above and caller set from->moving_account, so
5440 * mod_memcg_page_state will serialize updates to PageDirty.
5441 * So mapping should be stable for dirty pages.
5443 if (!anon && PageDirty(page)) {
5444 struct address_space *mapping = page_mapping(page);
5446 if (mapping_cap_account_dirty(mapping)) {
5447 __mod_memcg_state(from, NR_FILE_DIRTY, -nr_pages);
5448 __mod_memcg_state(to, NR_FILE_DIRTY, nr_pages);
5452 if (PageWriteback(page)) {
5453 __mod_memcg_state(from, NR_WRITEBACK, -nr_pages);
5454 __mod_memcg_state(to, NR_WRITEBACK, nr_pages);
5457 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5458 if (compound && !list_empty(page_deferred_list(page))) {
5459 spin_lock(&from->deferred_split_queue.split_queue_lock);
5460 list_del_init(page_deferred_list(page));
5461 from->deferred_split_queue.split_queue_len--;
5462 spin_unlock(&from->deferred_split_queue.split_queue_lock);
5466 * It is safe to change page->mem_cgroup here because the page
5467 * is referenced, charged, and isolated - we can't race with
5468 * uncharging, charging, migration, or LRU putback.
5471 /* caller should have done css_get */
5472 page->mem_cgroup = to;
5474 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5475 if (compound && list_empty(page_deferred_list(page))) {
5476 spin_lock(&to->deferred_split_queue.split_queue_lock);
5477 list_add_tail(page_deferred_list(page),
5478 &to->deferred_split_queue.split_queue);
5479 to->deferred_split_queue.split_queue_len++;
5480 spin_unlock(&to->deferred_split_queue.split_queue_lock);
5484 spin_unlock_irqrestore(&from->move_lock, flags);
5488 local_irq_disable();
5489 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
5490 memcg_check_events(to, page);
5491 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
5492 memcg_check_events(from, page);
5501 * get_mctgt_type - get target type of moving charge
5502 * @vma: the vma the pte to be checked belongs
5503 * @addr: the address corresponding to the pte to be checked
5504 * @ptent: the pte to be checked
5505 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5508 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5509 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5510 * move charge. if @target is not NULL, the page is stored in target->page
5511 * with extra refcnt got(Callers should handle it).
5512 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5513 * target for charge migration. if @target is not NULL, the entry is stored
5515 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5516 * (so ZONE_DEVICE page and thus not on the lru).
5517 * For now we such page is charge like a regular page would be as for all
5518 * intent and purposes it is just special memory taking the place of a
5521 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5523 * Called with pte lock held.
5526 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5527 unsigned long addr, pte_t ptent, union mc_target *target)
5529 struct page *page = NULL;
5530 enum mc_target_type ret = MC_TARGET_NONE;
5531 swp_entry_t ent = { .val = 0 };
5533 if (pte_present(ptent))
5534 page = mc_handle_present_pte(vma, addr, ptent);
5535 else if (is_swap_pte(ptent))
5536 page = mc_handle_swap_pte(vma, ptent, &ent);
5537 else if (pte_none(ptent))
5538 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5540 if (!page && !ent.val)
5544 * Do only loose check w/o serialization.
5545 * mem_cgroup_move_account() checks the page is valid or
5546 * not under LRU exclusion.
5548 if (page->mem_cgroup == mc.from) {
5549 ret = MC_TARGET_PAGE;
5550 if (is_device_private_page(page))
5551 ret = MC_TARGET_DEVICE;
5553 target->page = page;
5555 if (!ret || !target)
5559 * There is a swap entry and a page doesn't exist or isn't charged.
5560 * But we cannot move a tail-page in a THP.
5562 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5563 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5564 ret = MC_TARGET_SWAP;
5571 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5573 * We don't consider PMD mapped swapping or file mapped pages because THP does
5574 * not support them for now.
5575 * Caller should make sure that pmd_trans_huge(pmd) is true.
5577 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5578 unsigned long addr, pmd_t pmd, union mc_target *target)
5580 struct page *page = NULL;
5581 enum mc_target_type ret = MC_TARGET_NONE;
5583 if (unlikely(is_swap_pmd(pmd))) {
5584 VM_BUG_ON(thp_migration_supported() &&
5585 !is_pmd_migration_entry(pmd));
5588 page = pmd_page(pmd);
5589 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5590 if (!(mc.flags & MOVE_ANON))
5592 if (page->mem_cgroup == mc.from) {
5593 ret = MC_TARGET_PAGE;
5596 target->page = page;
5602 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5603 unsigned long addr, pmd_t pmd, union mc_target *target)
5605 return MC_TARGET_NONE;
5609 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5610 unsigned long addr, unsigned long end,
5611 struct mm_walk *walk)
5613 struct vm_area_struct *vma = walk->vma;
5617 ptl = pmd_trans_huge_lock(pmd, vma);
5620 * Note their can not be MC_TARGET_DEVICE for now as we do not
5621 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5622 * this might change.
5624 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5625 mc.precharge += HPAGE_PMD_NR;
5630 if (pmd_trans_unstable(pmd))
5632 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5633 for (; addr != end; pte++, addr += PAGE_SIZE)
5634 if (get_mctgt_type(vma, addr, *pte, NULL))
5635 mc.precharge++; /* increment precharge temporarily */
5636 pte_unmap_unlock(pte - 1, ptl);
5642 static const struct mm_walk_ops precharge_walk_ops = {
5643 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5646 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5648 unsigned long precharge;
5650 down_read(&mm->mmap_sem);
5651 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5652 up_read(&mm->mmap_sem);
5654 precharge = mc.precharge;
5660 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5662 unsigned long precharge = mem_cgroup_count_precharge(mm);
5664 VM_BUG_ON(mc.moving_task);
5665 mc.moving_task = current;
5666 return mem_cgroup_do_precharge(precharge);
5669 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5670 static void __mem_cgroup_clear_mc(void)
5672 struct mem_cgroup *from = mc.from;
5673 struct mem_cgroup *to = mc.to;
5675 /* we must uncharge all the leftover precharges from mc.to */
5677 cancel_charge(mc.to, mc.precharge);
5681 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5682 * we must uncharge here.
5684 if (mc.moved_charge) {
5685 cancel_charge(mc.from, mc.moved_charge);
5686 mc.moved_charge = 0;
5688 /* we must fixup refcnts and charges */
5689 if (mc.moved_swap) {
5690 /* uncharge swap account from the old cgroup */
5691 if (!mem_cgroup_is_root(mc.from))
5692 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5694 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5697 * we charged both to->memory and to->memsw, so we
5698 * should uncharge to->memory.
5700 if (!mem_cgroup_is_root(mc.to))
5701 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5703 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
5704 css_put_many(&mc.to->css, mc.moved_swap);
5708 memcg_oom_recover(from);
5709 memcg_oom_recover(to);
5710 wake_up_all(&mc.waitq);
5713 static void mem_cgroup_clear_mc(void)
5715 struct mm_struct *mm = mc.mm;
5718 * we must clear moving_task before waking up waiters at the end of
5721 mc.moving_task = NULL;
5722 __mem_cgroup_clear_mc();
5723 spin_lock(&mc.lock);
5727 spin_unlock(&mc.lock);
5732 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5734 struct cgroup_subsys_state *css;
5735 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5736 struct mem_cgroup *from;
5737 struct task_struct *leader, *p;
5738 struct mm_struct *mm;
5739 unsigned long move_flags;
5742 /* charge immigration isn't supported on the default hierarchy */
5743 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5747 * Multi-process migrations only happen on the default hierarchy
5748 * where charge immigration is not used. Perform charge
5749 * immigration if @tset contains a leader and whine if there are
5753 cgroup_taskset_for_each_leader(leader, css, tset) {
5756 memcg = mem_cgroup_from_css(css);
5762 * We are now commited to this value whatever it is. Changes in this
5763 * tunable will only affect upcoming migrations, not the current one.
5764 * So we need to save it, and keep it going.
5766 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5770 from = mem_cgroup_from_task(p);
5772 VM_BUG_ON(from == memcg);
5774 mm = get_task_mm(p);
5777 /* We move charges only when we move a owner of the mm */
5778 if (mm->owner == p) {
5781 VM_BUG_ON(mc.precharge);
5782 VM_BUG_ON(mc.moved_charge);
5783 VM_BUG_ON(mc.moved_swap);
5785 spin_lock(&mc.lock);
5789 mc.flags = move_flags;
5790 spin_unlock(&mc.lock);
5791 /* We set mc.moving_task later */
5793 ret = mem_cgroup_precharge_mc(mm);
5795 mem_cgroup_clear_mc();
5802 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5805 mem_cgroup_clear_mc();
5808 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5809 unsigned long addr, unsigned long end,
5810 struct mm_walk *walk)
5813 struct vm_area_struct *vma = walk->vma;
5816 enum mc_target_type target_type;
5817 union mc_target target;
5820 ptl = pmd_trans_huge_lock(pmd, vma);
5822 if (mc.precharge < HPAGE_PMD_NR) {
5826 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5827 if (target_type == MC_TARGET_PAGE) {
5829 if (!isolate_lru_page(page)) {
5830 if (!mem_cgroup_move_account(page, true,
5832 mc.precharge -= HPAGE_PMD_NR;
5833 mc.moved_charge += HPAGE_PMD_NR;
5835 putback_lru_page(page);
5838 } else if (target_type == MC_TARGET_DEVICE) {
5840 if (!mem_cgroup_move_account(page, true,
5842 mc.precharge -= HPAGE_PMD_NR;
5843 mc.moved_charge += HPAGE_PMD_NR;
5851 if (pmd_trans_unstable(pmd))
5854 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5855 for (; addr != end; addr += PAGE_SIZE) {
5856 pte_t ptent = *(pte++);
5857 bool device = false;
5863 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5864 case MC_TARGET_DEVICE:
5867 case MC_TARGET_PAGE:
5870 * We can have a part of the split pmd here. Moving it
5871 * can be done but it would be too convoluted so simply
5872 * ignore such a partial THP and keep it in original
5873 * memcg. There should be somebody mapping the head.
5875 if (PageTransCompound(page))
5877 if (!device && isolate_lru_page(page))
5879 if (!mem_cgroup_move_account(page, false,
5882 /* we uncharge from mc.from later. */
5886 putback_lru_page(page);
5887 put: /* get_mctgt_type() gets the page */
5890 case MC_TARGET_SWAP:
5892 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5894 /* we fixup refcnts and charges later. */
5902 pte_unmap_unlock(pte - 1, ptl);
5907 * We have consumed all precharges we got in can_attach().
5908 * We try charge one by one, but don't do any additional
5909 * charges to mc.to if we have failed in charge once in attach()
5912 ret = mem_cgroup_do_precharge(1);
5920 static const struct mm_walk_ops charge_walk_ops = {
5921 .pmd_entry = mem_cgroup_move_charge_pte_range,
5924 static void mem_cgroup_move_charge(void)
5926 lru_add_drain_all();
5928 * Signal lock_page_memcg() to take the memcg's move_lock
5929 * while we're moving its pages to another memcg. Then wait
5930 * for already started RCU-only updates to finish.
5932 atomic_inc(&mc.from->moving_account);
5935 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5937 * Someone who are holding the mmap_sem might be waiting in
5938 * waitq. So we cancel all extra charges, wake up all waiters,
5939 * and retry. Because we cancel precharges, we might not be able
5940 * to move enough charges, but moving charge is a best-effort
5941 * feature anyway, so it wouldn't be a big problem.
5943 __mem_cgroup_clear_mc();
5948 * When we have consumed all precharges and failed in doing
5949 * additional charge, the page walk just aborts.
5951 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
5954 up_read(&mc.mm->mmap_sem);
5955 atomic_dec(&mc.from->moving_account);
5958 static void mem_cgroup_move_task(void)
5961 mem_cgroup_move_charge();
5962 mem_cgroup_clear_mc();
5965 #else /* !CONFIG_MMU */
5966 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5970 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5973 static void mem_cgroup_move_task(void)
5979 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5980 * to verify whether we're attached to the default hierarchy on each mount
5983 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5986 * use_hierarchy is forced on the default hierarchy. cgroup core
5987 * guarantees that @root doesn't have any children, so turning it
5988 * on for the root memcg is enough.
5990 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5991 root_mem_cgroup->use_hierarchy = true;
5993 root_mem_cgroup->use_hierarchy = false;
5996 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
5998 if (value == PAGE_COUNTER_MAX)
5999 seq_puts(m, "max\n");
6001 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6006 static u64 memory_current_read(struct cgroup_subsys_state *css,
6009 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6011 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6014 static int memory_min_show(struct seq_file *m, void *v)
6016 return seq_puts_memcg_tunable(m,
6017 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6020 static ssize_t memory_min_write(struct kernfs_open_file *of,
6021 char *buf, size_t nbytes, loff_t off)
6023 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6027 buf = strstrip(buf);
6028 err = page_counter_memparse(buf, "max", &min);
6032 page_counter_set_min(&memcg->memory, min);
6037 static int memory_low_show(struct seq_file *m, void *v)
6039 return seq_puts_memcg_tunable(m,
6040 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6043 static ssize_t memory_low_write(struct kernfs_open_file *of,
6044 char *buf, size_t nbytes, loff_t off)
6046 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6050 buf = strstrip(buf);
6051 err = page_counter_memparse(buf, "max", &low);
6055 page_counter_set_low(&memcg->memory, low);
6060 static int memory_high_show(struct seq_file *m, void *v)
6062 return seq_puts_memcg_tunable(m, READ_ONCE(mem_cgroup_from_seq(m)->high));
6065 static ssize_t memory_high_write(struct kernfs_open_file *of,
6066 char *buf, size_t nbytes, loff_t off)
6068 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6069 unsigned long nr_pages;
6073 buf = strstrip(buf);
6074 err = page_counter_memparse(buf, "max", &high);
6080 nr_pages = page_counter_read(&memcg->memory);
6081 if (nr_pages > high)
6082 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6085 memcg_wb_domain_size_changed(memcg);
6089 static int memory_max_show(struct seq_file *m, void *v)
6091 return seq_puts_memcg_tunable(m,
6092 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6095 static ssize_t memory_max_write(struct kernfs_open_file *of,
6096 char *buf, size_t nbytes, loff_t off)
6098 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6099 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
6100 bool drained = false;
6104 buf = strstrip(buf);
6105 err = page_counter_memparse(buf, "max", &max);
6109 xchg(&memcg->memory.max, max);
6112 unsigned long nr_pages = page_counter_read(&memcg->memory);
6114 if (nr_pages <= max)
6117 if (signal_pending(current)) {
6123 drain_all_stock(memcg);
6129 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6135 memcg_memory_event(memcg, MEMCG_OOM);
6136 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6140 memcg_wb_domain_size_changed(memcg);
6144 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6146 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6147 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6148 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6149 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6150 seq_printf(m, "oom_kill %lu\n",
6151 atomic_long_read(&events[MEMCG_OOM_KILL]));
6154 static int memory_events_show(struct seq_file *m, void *v)
6156 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6158 __memory_events_show(m, memcg->memory_events);
6162 static int memory_events_local_show(struct seq_file *m, void *v)
6164 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6166 __memory_events_show(m, memcg->memory_events_local);
6170 static int memory_stat_show(struct seq_file *m, void *v)
6172 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6175 buf = memory_stat_format(memcg);
6183 static int memory_oom_group_show(struct seq_file *m, void *v)
6185 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6187 seq_printf(m, "%d\n", memcg->oom_group);
6192 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6193 char *buf, size_t nbytes, loff_t off)
6195 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6198 buf = strstrip(buf);
6202 ret = kstrtoint(buf, 0, &oom_group);
6206 if (oom_group != 0 && oom_group != 1)
6209 memcg->oom_group = oom_group;
6214 static struct cftype memory_files[] = {
6217 .flags = CFTYPE_NOT_ON_ROOT,
6218 .read_u64 = memory_current_read,
6222 .flags = CFTYPE_NOT_ON_ROOT,
6223 .seq_show = memory_min_show,
6224 .write = memory_min_write,
6228 .flags = CFTYPE_NOT_ON_ROOT,
6229 .seq_show = memory_low_show,
6230 .write = memory_low_write,
6234 .flags = CFTYPE_NOT_ON_ROOT,
6235 .seq_show = memory_high_show,
6236 .write = memory_high_write,
6240 .flags = CFTYPE_NOT_ON_ROOT,
6241 .seq_show = memory_max_show,
6242 .write = memory_max_write,
6246 .flags = CFTYPE_NOT_ON_ROOT,
6247 .file_offset = offsetof(struct mem_cgroup, events_file),
6248 .seq_show = memory_events_show,
6251 .name = "events.local",
6252 .flags = CFTYPE_NOT_ON_ROOT,
6253 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6254 .seq_show = memory_events_local_show,
6258 .flags = CFTYPE_NOT_ON_ROOT,
6259 .seq_show = memory_stat_show,
6262 .name = "oom.group",
6263 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6264 .seq_show = memory_oom_group_show,
6265 .write = memory_oom_group_write,
6270 struct cgroup_subsys memory_cgrp_subsys = {
6271 .css_alloc = mem_cgroup_css_alloc,
6272 .css_online = mem_cgroup_css_online,
6273 .css_offline = mem_cgroup_css_offline,
6274 .css_released = mem_cgroup_css_released,
6275 .css_free = mem_cgroup_css_free,
6276 .css_reset = mem_cgroup_css_reset,
6277 .can_attach = mem_cgroup_can_attach,
6278 .cancel_attach = mem_cgroup_cancel_attach,
6279 .post_attach = mem_cgroup_move_task,
6280 .bind = mem_cgroup_bind,
6281 .dfl_cftypes = memory_files,
6282 .legacy_cftypes = mem_cgroup_legacy_files,
6287 * mem_cgroup_protected - check if memory consumption is in the normal range
6288 * @root: the top ancestor of the sub-tree being checked
6289 * @memcg: the memory cgroup to check
6291 * WARNING: This function is not stateless! It can only be used as part
6292 * of a top-down tree iteration, not for isolated queries.
6294 * Returns one of the following:
6295 * MEMCG_PROT_NONE: cgroup memory is not protected
6296 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
6297 * an unprotected supply of reclaimable memory from other cgroups.
6298 * MEMCG_PROT_MIN: cgroup memory is protected
6300 * @root is exclusive; it is never protected when looked at directly
6302 * To provide a proper hierarchical behavior, effective memory.min/low values
6303 * are used. Below is the description of how effective memory.low is calculated.
6304 * Effective memory.min values is calculated in the same way.
6306 * Effective memory.low is always equal or less than the original memory.low.
6307 * If there is no memory.low overcommittment (which is always true for
6308 * top-level memory cgroups), these two values are equal.
6309 * Otherwise, it's a part of parent's effective memory.low,
6310 * calculated as a cgroup's memory.low usage divided by sum of sibling's
6311 * memory.low usages, where memory.low usage is the size of actually
6315 * elow = min( memory.low, parent->elow * ------------------ ),
6316 * siblings_low_usage
6318 * | memory.current, if memory.current < memory.low
6323 * Such definition of the effective memory.low provides the expected
6324 * hierarchical behavior: parent's memory.low value is limiting
6325 * children, unprotected memory is reclaimed first and cgroups,
6326 * which are not using their guarantee do not affect actual memory
6329 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
6331 * A A/memory.low = 2G, A/memory.current = 6G
6333 * BC DE B/memory.low = 3G B/memory.current = 2G
6334 * C/memory.low = 1G C/memory.current = 2G
6335 * D/memory.low = 0 D/memory.current = 2G
6336 * E/memory.low = 10G E/memory.current = 0
6338 * and the memory pressure is applied, the following memory distribution
6339 * is expected (approximately):
6341 * A/memory.current = 2G
6343 * B/memory.current = 1.3G
6344 * C/memory.current = 0.6G
6345 * D/memory.current = 0
6346 * E/memory.current = 0
6348 * These calculations require constant tracking of the actual low usages
6349 * (see propagate_protected_usage()), as well as recursive calculation of
6350 * effective memory.low values. But as we do call mem_cgroup_protected()
6351 * path for each memory cgroup top-down from the reclaim,
6352 * it's possible to optimize this part, and save calculated elow
6353 * for next usage. This part is intentionally racy, but it's ok,
6354 * as memory.low is a best-effort mechanism.
6356 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
6357 struct mem_cgroup *memcg)
6359 struct mem_cgroup *parent;
6360 unsigned long emin, parent_emin;
6361 unsigned long elow, parent_elow;
6362 unsigned long usage;
6364 if (mem_cgroup_disabled())
6365 return MEMCG_PROT_NONE;
6368 root = root_mem_cgroup;
6370 return MEMCG_PROT_NONE;
6372 usage = page_counter_read(&memcg->memory);
6374 return MEMCG_PROT_NONE;
6376 emin = memcg->memory.min;
6377 elow = memcg->memory.low;
6379 parent = parent_mem_cgroup(memcg);
6380 /* No parent means a non-hierarchical mode on v1 memcg */
6382 return MEMCG_PROT_NONE;
6387 parent_emin = READ_ONCE(parent->memory.emin);
6388 emin = min(emin, parent_emin);
6389 if (emin && parent_emin) {
6390 unsigned long min_usage, siblings_min_usage;
6392 min_usage = min(usage, memcg->memory.min);
6393 siblings_min_usage = atomic_long_read(
6394 &parent->memory.children_min_usage);
6396 if (min_usage && siblings_min_usage)
6397 emin = min(emin, parent_emin * min_usage /
6398 siblings_min_usage);
6401 parent_elow = READ_ONCE(parent->memory.elow);
6402 elow = min(elow, parent_elow);
6403 if (elow && parent_elow) {
6404 unsigned long low_usage, siblings_low_usage;
6406 low_usage = min(usage, memcg->memory.low);
6407 siblings_low_usage = atomic_long_read(
6408 &parent->memory.children_low_usage);
6410 if (low_usage && siblings_low_usage)
6411 elow = min(elow, parent_elow * low_usage /
6412 siblings_low_usage);
6416 memcg->memory.emin = emin;
6417 memcg->memory.elow = elow;
6420 return MEMCG_PROT_MIN;
6421 else if (usage <= elow)
6422 return MEMCG_PROT_LOW;
6424 return MEMCG_PROT_NONE;
6428 * mem_cgroup_try_charge - try charging a page
6429 * @page: page to charge
6430 * @mm: mm context of the victim
6431 * @gfp_mask: reclaim mode
6432 * @memcgp: charged memcg return
6433 * @compound: charge the page as compound or small page
6435 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6436 * pages according to @gfp_mask if necessary.
6438 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6439 * Otherwise, an error code is returned.
6441 * After page->mapping has been set up, the caller must finalize the
6442 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6443 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6445 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
6446 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6449 struct mem_cgroup *memcg = NULL;
6450 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6453 if (mem_cgroup_disabled())
6456 if (PageSwapCache(page)) {
6458 * Every swap fault against a single page tries to charge the
6459 * page, bail as early as possible. shmem_unuse() encounters
6460 * already charged pages, too. The USED bit is protected by
6461 * the page lock, which serializes swap cache removal, which
6462 * in turn serializes uncharging.
6464 VM_BUG_ON_PAGE(!PageLocked(page), page);
6465 if (compound_head(page)->mem_cgroup)
6468 if (do_swap_account) {
6469 swp_entry_t ent = { .val = page_private(page), };
6470 unsigned short id = lookup_swap_cgroup_id(ent);
6473 memcg = mem_cgroup_from_id(id);
6474 if (memcg && !css_tryget_online(&memcg->css))
6481 memcg = get_mem_cgroup_from_mm(mm);
6483 ret = try_charge(memcg, gfp_mask, nr_pages);
6485 css_put(&memcg->css);
6491 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
6492 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6495 struct mem_cgroup *memcg;
6498 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound);
6500 mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask);
6505 * mem_cgroup_commit_charge - commit a page charge
6506 * @page: page to charge
6507 * @memcg: memcg to charge the page to
6508 * @lrucare: page might be on LRU already
6509 * @compound: charge the page as compound or small page
6511 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6512 * after page->mapping has been set up. This must happen atomically
6513 * as part of the page instantiation, i.e. under the page table lock
6514 * for anonymous pages, under the page lock for page and swap cache.
6516 * In addition, the page must not be on the LRU during the commit, to
6517 * prevent racing with task migration. If it might be, use @lrucare.
6519 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6521 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
6522 bool lrucare, bool compound)
6524 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6526 VM_BUG_ON_PAGE(!page->mapping, page);
6527 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
6529 if (mem_cgroup_disabled())
6532 * Swap faults will attempt to charge the same page multiple
6533 * times. But reuse_swap_page() might have removed the page
6534 * from swapcache already, so we can't check PageSwapCache().
6539 commit_charge(page, memcg, lrucare);
6541 local_irq_disable();
6542 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
6543 memcg_check_events(memcg, page);
6546 if (do_memsw_account() && PageSwapCache(page)) {
6547 swp_entry_t entry = { .val = page_private(page) };
6549 * The swap entry might not get freed for a long time,
6550 * let's not wait for it. The page already received a
6551 * memory+swap charge, drop the swap entry duplicate.
6553 mem_cgroup_uncharge_swap(entry, nr_pages);
6558 * mem_cgroup_cancel_charge - cancel a page charge
6559 * @page: page to charge
6560 * @memcg: memcg to charge the page to
6561 * @compound: charge the page as compound or small page
6563 * Cancel a charge transaction started by mem_cgroup_try_charge().
6565 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
6568 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6570 if (mem_cgroup_disabled())
6573 * Swap faults will attempt to charge the same page multiple
6574 * times. But reuse_swap_page() might have removed the page
6575 * from swapcache already, so we can't check PageSwapCache().
6580 cancel_charge(memcg, nr_pages);
6583 struct uncharge_gather {
6584 struct mem_cgroup *memcg;
6585 unsigned long pgpgout;
6586 unsigned long nr_anon;
6587 unsigned long nr_file;
6588 unsigned long nr_kmem;
6589 unsigned long nr_huge;
6590 unsigned long nr_shmem;
6591 struct page *dummy_page;
6594 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6596 memset(ug, 0, sizeof(*ug));
6599 static void uncharge_batch(const struct uncharge_gather *ug)
6601 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
6602 unsigned long flags;
6604 if (!mem_cgroup_is_root(ug->memcg)) {
6605 page_counter_uncharge(&ug->memcg->memory, nr_pages);
6606 if (do_memsw_account())
6607 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
6608 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6609 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6610 memcg_oom_recover(ug->memcg);
6613 local_irq_save(flags);
6614 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
6615 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
6616 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
6617 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
6618 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6619 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, nr_pages);
6620 memcg_check_events(ug->memcg, ug->dummy_page);
6621 local_irq_restore(flags);
6623 if (!mem_cgroup_is_root(ug->memcg))
6624 css_put_many(&ug->memcg->css, nr_pages);
6627 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6629 VM_BUG_ON_PAGE(PageLRU(page), page);
6630 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
6631 !PageHWPoison(page) , page);
6633 if (!page->mem_cgroup)
6637 * Nobody should be changing or seriously looking at
6638 * page->mem_cgroup at this point, we have fully
6639 * exclusive access to the page.
6642 if (ug->memcg != page->mem_cgroup) {
6645 uncharge_gather_clear(ug);
6647 ug->memcg = page->mem_cgroup;
6650 if (!PageKmemcg(page)) {
6651 unsigned int nr_pages = 1;
6653 if (PageTransHuge(page)) {
6654 nr_pages = compound_nr(page);
6655 ug->nr_huge += nr_pages;
6658 ug->nr_anon += nr_pages;
6660 ug->nr_file += nr_pages;
6661 if (PageSwapBacked(page))
6662 ug->nr_shmem += nr_pages;
6666 ug->nr_kmem += compound_nr(page);
6667 __ClearPageKmemcg(page);
6670 ug->dummy_page = page;
6671 page->mem_cgroup = NULL;
6674 static void uncharge_list(struct list_head *page_list)
6676 struct uncharge_gather ug;
6677 struct list_head *next;
6679 uncharge_gather_clear(&ug);
6682 * Note that the list can be a single page->lru; hence the
6683 * do-while loop instead of a simple list_for_each_entry().
6685 next = page_list->next;
6689 page = list_entry(next, struct page, lru);
6690 next = page->lru.next;
6692 uncharge_page(page, &ug);
6693 } while (next != page_list);
6696 uncharge_batch(&ug);
6700 * mem_cgroup_uncharge - uncharge a page
6701 * @page: page to uncharge
6703 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6704 * mem_cgroup_commit_charge().
6706 void mem_cgroup_uncharge(struct page *page)
6708 struct uncharge_gather ug;
6710 if (mem_cgroup_disabled())
6713 /* Don't touch page->lru of any random page, pre-check: */
6714 if (!page->mem_cgroup)
6717 uncharge_gather_clear(&ug);
6718 uncharge_page(page, &ug);
6719 uncharge_batch(&ug);
6723 * mem_cgroup_uncharge_list - uncharge a list of page
6724 * @page_list: list of pages to uncharge
6726 * Uncharge a list of pages previously charged with
6727 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6729 void mem_cgroup_uncharge_list(struct list_head *page_list)
6731 if (mem_cgroup_disabled())
6734 if (!list_empty(page_list))
6735 uncharge_list(page_list);
6739 * mem_cgroup_migrate - charge a page's replacement
6740 * @oldpage: currently circulating page
6741 * @newpage: replacement page
6743 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6744 * be uncharged upon free.
6746 * Both pages must be locked, @newpage->mapping must be set up.
6748 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6750 struct mem_cgroup *memcg;
6751 unsigned int nr_pages;
6753 unsigned long flags;
6755 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6756 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6757 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6758 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6761 if (mem_cgroup_disabled())
6764 /* Page cache replacement: new page already charged? */
6765 if (newpage->mem_cgroup)
6768 /* Swapcache readahead pages can get replaced before being charged */
6769 memcg = oldpage->mem_cgroup;
6773 /* Force-charge the new page. The old one will be freed soon */
6774 compound = PageTransHuge(newpage);
6775 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
6777 page_counter_charge(&memcg->memory, nr_pages);
6778 if (do_memsw_account())
6779 page_counter_charge(&memcg->memsw, nr_pages);
6780 css_get_many(&memcg->css, nr_pages);
6782 commit_charge(newpage, memcg, false);
6784 local_irq_save(flags);
6785 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
6786 memcg_check_events(memcg, newpage);
6787 local_irq_restore(flags);
6790 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6791 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6793 void mem_cgroup_sk_alloc(struct sock *sk)
6795 struct mem_cgroup *memcg;
6797 if (!mem_cgroup_sockets_enabled)
6801 * Socket cloning can throw us here with sk_memcg already
6802 * filled. It won't however, necessarily happen from
6803 * process context. So the test for root memcg given
6804 * the current task's memcg won't help us in this case.
6806 * Respecting the original socket's memcg is a better
6807 * decision in this case.
6810 css_get(&sk->sk_memcg->css);
6815 memcg = mem_cgroup_from_task(current);
6816 if (memcg == root_mem_cgroup)
6818 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6820 if (css_tryget_online(&memcg->css))
6821 sk->sk_memcg = memcg;
6826 void mem_cgroup_sk_free(struct sock *sk)
6829 css_put(&sk->sk_memcg->css);
6833 * mem_cgroup_charge_skmem - charge socket memory
6834 * @memcg: memcg to charge
6835 * @nr_pages: number of pages to charge
6837 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6838 * @memcg's configured limit, %false if the charge had to be forced.
6840 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6842 gfp_t gfp_mask = GFP_KERNEL;
6844 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6845 struct page_counter *fail;
6847 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6848 memcg->tcpmem_pressure = 0;
6851 page_counter_charge(&memcg->tcpmem, nr_pages);
6852 memcg->tcpmem_pressure = 1;
6856 /* Don't block in the packet receive path */
6858 gfp_mask = GFP_NOWAIT;
6860 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6862 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6865 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6870 * mem_cgroup_uncharge_skmem - uncharge socket memory
6871 * @memcg: memcg to uncharge
6872 * @nr_pages: number of pages to uncharge
6874 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6876 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6877 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6881 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6883 refill_stock(memcg, nr_pages);
6886 static int __init cgroup_memory(char *s)
6890 while ((token = strsep(&s, ",")) != NULL) {
6893 if (!strcmp(token, "nosocket"))
6894 cgroup_memory_nosocket = true;
6895 if (!strcmp(token, "nokmem"))
6896 cgroup_memory_nokmem = true;
6900 __setup("cgroup.memory=", cgroup_memory);
6903 * subsys_initcall() for memory controller.
6905 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6906 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6907 * basically everything that doesn't depend on a specific mem_cgroup structure
6908 * should be initialized from here.
6910 static int __init mem_cgroup_init(void)
6914 #ifdef CONFIG_MEMCG_KMEM
6916 * Kmem cache creation is mostly done with the slab_mutex held,
6917 * so use a workqueue with limited concurrency to avoid stalling
6918 * all worker threads in case lots of cgroups are created and
6919 * destroyed simultaneously.
6921 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
6922 BUG_ON(!memcg_kmem_cache_wq);
6925 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6926 memcg_hotplug_cpu_dead);
6928 for_each_possible_cpu(cpu)
6929 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
6932 for_each_node(node) {
6933 struct mem_cgroup_tree_per_node *rtpn;
6935 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
6936 node_online(node) ? node : NUMA_NO_NODE);
6938 rtpn->rb_root = RB_ROOT;
6939 rtpn->rb_rightmost = NULL;
6940 spin_lock_init(&rtpn->lock);
6941 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6946 subsys_initcall(mem_cgroup_init);
6948 #ifdef CONFIG_MEMCG_SWAP
6949 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
6951 while (!refcount_inc_not_zero(&memcg->id.ref)) {
6953 * The root cgroup cannot be destroyed, so it's refcount must
6956 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
6960 memcg = parent_mem_cgroup(memcg);
6962 memcg = root_mem_cgroup;
6968 * mem_cgroup_swapout - transfer a memsw charge to swap
6969 * @page: page whose memsw charge to transfer
6970 * @entry: swap entry to move the charge to
6972 * Transfer the memsw charge of @page to @entry.
6974 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6976 struct mem_cgroup *memcg, *swap_memcg;
6977 unsigned int nr_entries;
6978 unsigned short oldid;
6980 VM_BUG_ON_PAGE(PageLRU(page), page);
6981 VM_BUG_ON_PAGE(page_count(page), page);
6983 if (!do_memsw_account())
6986 memcg = page->mem_cgroup;
6988 /* Readahead page, never charged */
6993 * In case the memcg owning these pages has been offlined and doesn't
6994 * have an ID allocated to it anymore, charge the closest online
6995 * ancestor for the swap instead and transfer the memory+swap charge.
6997 swap_memcg = mem_cgroup_id_get_online(memcg);
6998 nr_entries = hpage_nr_pages(page);
6999 /* Get references for the tail pages, too */
7001 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7002 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7004 VM_BUG_ON_PAGE(oldid, page);
7005 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7007 page->mem_cgroup = NULL;
7009 if (!mem_cgroup_is_root(memcg))
7010 page_counter_uncharge(&memcg->memory, nr_entries);
7012 if (memcg != swap_memcg) {
7013 if (!mem_cgroup_is_root(swap_memcg))
7014 page_counter_charge(&swap_memcg->memsw, nr_entries);
7015 page_counter_uncharge(&memcg->memsw, nr_entries);
7019 * Interrupts should be disabled here because the caller holds the
7020 * i_pages lock which is taken with interrupts-off. It is
7021 * important here to have the interrupts disabled because it is the
7022 * only synchronisation we have for updating the per-CPU variables.
7024 VM_BUG_ON(!irqs_disabled());
7025 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
7027 memcg_check_events(memcg, page);
7029 if (!mem_cgroup_is_root(memcg))
7030 css_put_many(&memcg->css, nr_entries);
7034 * mem_cgroup_try_charge_swap - try charging swap space for a page
7035 * @page: page being added to swap
7036 * @entry: swap entry to charge
7038 * Try to charge @page's memcg for the swap space at @entry.
7040 * Returns 0 on success, -ENOMEM on failure.
7042 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7044 unsigned int nr_pages = hpage_nr_pages(page);
7045 struct page_counter *counter;
7046 struct mem_cgroup *memcg;
7047 unsigned short oldid;
7049 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
7052 memcg = page->mem_cgroup;
7054 /* Readahead page, never charged */
7059 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7063 memcg = mem_cgroup_id_get_online(memcg);
7065 if (!mem_cgroup_is_root(memcg) &&
7066 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7067 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7068 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7069 mem_cgroup_id_put(memcg);
7073 /* Get references for the tail pages, too */
7075 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7076 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7077 VM_BUG_ON_PAGE(oldid, page);
7078 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7084 * mem_cgroup_uncharge_swap - uncharge swap space
7085 * @entry: swap entry to uncharge
7086 * @nr_pages: the amount of swap space to uncharge
7088 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7090 struct mem_cgroup *memcg;
7093 if (!do_swap_account)
7096 id = swap_cgroup_record(entry, 0, nr_pages);
7098 memcg = mem_cgroup_from_id(id);
7100 if (!mem_cgroup_is_root(memcg)) {
7101 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7102 page_counter_uncharge(&memcg->swap, nr_pages);
7104 page_counter_uncharge(&memcg->memsw, nr_pages);
7106 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7107 mem_cgroup_id_put_many(memcg, nr_pages);
7112 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7114 long nr_swap_pages = get_nr_swap_pages();
7116 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7117 return nr_swap_pages;
7118 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7119 nr_swap_pages = min_t(long, nr_swap_pages,
7120 READ_ONCE(memcg->swap.max) -
7121 page_counter_read(&memcg->swap));
7122 return nr_swap_pages;
7125 bool mem_cgroup_swap_full(struct page *page)
7127 struct mem_cgroup *memcg;
7129 VM_BUG_ON_PAGE(!PageLocked(page), page);
7133 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7136 memcg = page->mem_cgroup;
7140 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7141 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.max)
7147 /* for remember boot option*/
7148 #ifdef CONFIG_MEMCG_SWAP_ENABLED
7149 static int really_do_swap_account __initdata = 1;
7151 static int really_do_swap_account __initdata;
7154 static int __init enable_swap_account(char *s)
7156 if (!strcmp(s, "1"))
7157 really_do_swap_account = 1;
7158 else if (!strcmp(s, "0"))
7159 really_do_swap_account = 0;
7162 __setup("swapaccount=", enable_swap_account);
7164 static u64 swap_current_read(struct cgroup_subsys_state *css,
7167 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7169 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7172 static int swap_max_show(struct seq_file *m, void *v)
7174 return seq_puts_memcg_tunable(m,
7175 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7178 static ssize_t swap_max_write(struct kernfs_open_file *of,
7179 char *buf, size_t nbytes, loff_t off)
7181 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7185 buf = strstrip(buf);
7186 err = page_counter_memparse(buf, "max", &max);
7190 xchg(&memcg->swap.max, max);
7195 static int swap_events_show(struct seq_file *m, void *v)
7197 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7199 seq_printf(m, "max %lu\n",
7200 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7201 seq_printf(m, "fail %lu\n",
7202 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7207 static struct cftype swap_files[] = {
7209 .name = "swap.current",
7210 .flags = CFTYPE_NOT_ON_ROOT,
7211 .read_u64 = swap_current_read,
7215 .flags = CFTYPE_NOT_ON_ROOT,
7216 .seq_show = swap_max_show,
7217 .write = swap_max_write,
7220 .name = "swap.events",
7221 .flags = CFTYPE_NOT_ON_ROOT,
7222 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7223 .seq_show = swap_events_show,
7228 static struct cftype memsw_cgroup_files[] = {
7230 .name = "memsw.usage_in_bytes",
7231 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7232 .read_u64 = mem_cgroup_read_u64,
7235 .name = "memsw.max_usage_in_bytes",
7236 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7237 .write = mem_cgroup_reset,
7238 .read_u64 = mem_cgroup_read_u64,
7241 .name = "memsw.limit_in_bytes",
7242 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7243 .write = mem_cgroup_write,
7244 .read_u64 = mem_cgroup_read_u64,
7247 .name = "memsw.failcnt",
7248 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7249 .write = mem_cgroup_reset,
7250 .read_u64 = mem_cgroup_read_u64,
7252 { }, /* terminate */
7255 static int __init mem_cgroup_swap_init(void)
7257 if (!mem_cgroup_disabled() && really_do_swap_account) {
7258 do_swap_account = 1;
7259 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
7261 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
7262 memsw_cgroup_files));
7266 subsys_initcall(mem_cgroup_swap_init);
7268 #endif /* CONFIG_MEMCG_SWAP */