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);
422 mem_cgroup_iter_break(NULL, memcg);
428 memcg_shrinker_map_size = size;
429 mutex_unlock(&memcg_shrinker_map_mutex);
433 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
435 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
436 struct memcg_shrinker_map *map;
439 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
440 /* Pairs with smp mb in shrink_slab() */
441 smp_mb__before_atomic();
442 set_bit(shrinker_id, map->map);
448 * mem_cgroup_css_from_page - css of the memcg associated with a page
449 * @page: page of interest
451 * If memcg is bound to the default hierarchy, css of the memcg associated
452 * with @page is returned. The returned css remains associated with @page
453 * until it is released.
455 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
458 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
460 struct mem_cgroup *memcg;
462 memcg = page->mem_cgroup;
464 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
465 memcg = root_mem_cgroup;
471 * page_cgroup_ino - return inode number of the memcg a page is charged to
474 * Look up the closest online ancestor of the memory cgroup @page is charged to
475 * and return its inode number or 0 if @page is not charged to any cgroup. It
476 * is safe to call this function without holding a reference to @page.
478 * Note, this function is inherently racy, because there is nothing to prevent
479 * the cgroup inode from getting torn down and potentially reallocated a moment
480 * after page_cgroup_ino() returns, so it only should be used by callers that
481 * do not care (such as procfs interfaces).
483 ino_t page_cgroup_ino(struct page *page)
485 struct mem_cgroup *memcg;
486 unsigned long ino = 0;
489 if (PageSlab(page) && !PageTail(page))
490 memcg = memcg_from_slab_page(page);
492 memcg = READ_ONCE(page->mem_cgroup);
493 while (memcg && !(memcg->css.flags & CSS_ONLINE))
494 memcg = parent_mem_cgroup(memcg);
496 ino = cgroup_ino(memcg->css.cgroup);
501 static struct mem_cgroup_per_node *
502 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
504 int nid = page_to_nid(page);
506 return memcg->nodeinfo[nid];
509 static struct mem_cgroup_tree_per_node *
510 soft_limit_tree_node(int nid)
512 return soft_limit_tree.rb_tree_per_node[nid];
515 static struct mem_cgroup_tree_per_node *
516 soft_limit_tree_from_page(struct page *page)
518 int nid = page_to_nid(page);
520 return soft_limit_tree.rb_tree_per_node[nid];
523 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
524 struct mem_cgroup_tree_per_node *mctz,
525 unsigned long new_usage_in_excess)
527 struct rb_node **p = &mctz->rb_root.rb_node;
528 struct rb_node *parent = NULL;
529 struct mem_cgroup_per_node *mz_node;
530 bool rightmost = true;
535 mz->usage_in_excess = new_usage_in_excess;
536 if (!mz->usage_in_excess)
540 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
542 if (mz->usage_in_excess < mz_node->usage_in_excess) {
548 * We can't avoid mem cgroups that are over their soft
549 * limit by the same amount
551 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
556 mctz->rb_rightmost = &mz->tree_node;
558 rb_link_node(&mz->tree_node, parent, p);
559 rb_insert_color(&mz->tree_node, &mctz->rb_root);
563 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
564 struct mem_cgroup_tree_per_node *mctz)
569 if (&mz->tree_node == mctz->rb_rightmost)
570 mctz->rb_rightmost = rb_prev(&mz->tree_node);
572 rb_erase(&mz->tree_node, &mctz->rb_root);
576 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
577 struct mem_cgroup_tree_per_node *mctz)
581 spin_lock_irqsave(&mctz->lock, flags);
582 __mem_cgroup_remove_exceeded(mz, mctz);
583 spin_unlock_irqrestore(&mctz->lock, flags);
586 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
588 unsigned long nr_pages = page_counter_read(&memcg->memory);
589 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
590 unsigned long excess = 0;
592 if (nr_pages > soft_limit)
593 excess = nr_pages - soft_limit;
598 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
600 unsigned long excess;
601 struct mem_cgroup_per_node *mz;
602 struct mem_cgroup_tree_per_node *mctz;
604 mctz = soft_limit_tree_from_page(page);
608 * Necessary to update all ancestors when hierarchy is used.
609 * because their event counter is not touched.
611 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
612 mz = mem_cgroup_page_nodeinfo(memcg, page);
613 excess = soft_limit_excess(memcg);
615 * We have to update the tree if mz is on RB-tree or
616 * mem is over its softlimit.
618 if (excess || mz->on_tree) {
621 spin_lock_irqsave(&mctz->lock, flags);
622 /* if on-tree, remove it */
624 __mem_cgroup_remove_exceeded(mz, mctz);
626 * Insert again. mz->usage_in_excess will be updated.
627 * If excess is 0, no tree ops.
629 __mem_cgroup_insert_exceeded(mz, mctz, excess);
630 spin_unlock_irqrestore(&mctz->lock, flags);
635 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
637 struct mem_cgroup_tree_per_node *mctz;
638 struct mem_cgroup_per_node *mz;
642 mz = mem_cgroup_nodeinfo(memcg, nid);
643 mctz = soft_limit_tree_node(nid);
645 mem_cgroup_remove_exceeded(mz, mctz);
649 static struct mem_cgroup_per_node *
650 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
652 struct mem_cgroup_per_node *mz;
656 if (!mctz->rb_rightmost)
657 goto done; /* Nothing to reclaim from */
659 mz = rb_entry(mctz->rb_rightmost,
660 struct mem_cgroup_per_node, tree_node);
662 * Remove the node now but someone else can add it back,
663 * we will to add it back at the end of reclaim to its correct
664 * position in the tree.
666 __mem_cgroup_remove_exceeded(mz, mctz);
667 if (!soft_limit_excess(mz->memcg) ||
668 !css_tryget_online(&mz->memcg->css))
674 static struct mem_cgroup_per_node *
675 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
677 struct mem_cgroup_per_node *mz;
679 spin_lock_irq(&mctz->lock);
680 mz = __mem_cgroup_largest_soft_limit_node(mctz);
681 spin_unlock_irq(&mctz->lock);
686 * __mod_memcg_state - update cgroup memory statistics
687 * @memcg: the memory cgroup
688 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
689 * @val: delta to add to the counter, can be negative
691 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
695 if (mem_cgroup_disabled())
698 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
699 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
700 struct mem_cgroup *mi;
703 * Batch local counters to keep them in sync with
704 * the hierarchical ones.
706 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
707 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
708 atomic_long_add(x, &mi->vmstats[idx]);
711 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
714 static struct mem_cgroup_per_node *
715 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
717 struct mem_cgroup *parent;
719 parent = parent_mem_cgroup(pn->memcg);
722 return mem_cgroup_nodeinfo(parent, nid);
726 * __mod_lruvec_state - update lruvec memory statistics
727 * @lruvec: the lruvec
728 * @idx: the stat item
729 * @val: delta to add to the counter, can be negative
731 * The lruvec is the intersection of the NUMA node and a cgroup. This
732 * function updates the all three counters that are affected by a
733 * change of state at this level: per-node, per-cgroup, per-lruvec.
735 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
738 pg_data_t *pgdat = lruvec_pgdat(lruvec);
739 struct mem_cgroup_per_node *pn;
740 struct mem_cgroup *memcg;
744 __mod_node_page_state(pgdat, idx, val);
746 if (mem_cgroup_disabled())
749 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
753 __mod_memcg_state(memcg, idx, val);
756 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
758 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
759 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
760 struct mem_cgroup_per_node *pi;
762 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
763 atomic_long_add(x, &pi->lruvec_stat[idx]);
766 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
769 void __mod_lruvec_slab_state(void *p, enum node_stat_item idx, int val)
771 struct page *page = virt_to_head_page(p);
772 pg_data_t *pgdat = page_pgdat(page);
773 struct mem_cgroup *memcg;
774 struct lruvec *lruvec;
777 memcg = memcg_from_slab_page(page);
780 * Untracked pages have no memcg, no lruvec. Update only the
781 * node. If we reparent the slab objects to the root memcg,
782 * when we free the slab object, we need to update the per-memcg
783 * vmstats to keep it correct for the root memcg.
786 __mod_node_page_state(pgdat, idx, val);
788 lruvec = mem_cgroup_lruvec(pgdat, memcg);
789 __mod_lruvec_state(lruvec, idx, val);
794 void mod_memcg_obj_state(void *p, int idx, int val)
796 struct mem_cgroup *memcg;
799 memcg = mem_cgroup_from_obj(p);
801 mod_memcg_state(memcg, idx, val);
806 * __count_memcg_events - account VM events in a cgroup
807 * @memcg: the memory cgroup
808 * @idx: the event item
809 * @count: the number of events that occured
811 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
816 if (mem_cgroup_disabled())
819 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
820 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
821 struct mem_cgroup *mi;
824 * Batch local counters to keep them in sync with
825 * the hierarchical ones.
827 __this_cpu_add(memcg->vmstats_local->events[idx], x);
828 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
829 atomic_long_add(x, &mi->vmevents[idx]);
832 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
835 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
837 return atomic_long_read(&memcg->vmevents[event]);
840 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
845 for_each_possible_cpu(cpu)
846 x += per_cpu(memcg->vmstats_local->events[event], cpu);
850 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
852 bool compound, int nr_pages)
855 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
856 * counted as CACHE even if it's on ANON LRU.
859 __mod_memcg_state(memcg, MEMCG_RSS, nr_pages);
861 __mod_memcg_state(memcg, MEMCG_CACHE, nr_pages);
862 if (PageSwapBacked(page))
863 __mod_memcg_state(memcg, NR_SHMEM, nr_pages);
867 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
868 __mod_memcg_state(memcg, MEMCG_RSS_HUGE, nr_pages);
871 /* pagein of a big page is an event. So, ignore page size */
873 __count_memcg_events(memcg, PGPGIN, 1);
875 __count_memcg_events(memcg, PGPGOUT, 1);
876 nr_pages = -nr_pages; /* for event */
879 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
882 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
883 enum mem_cgroup_events_target target)
885 unsigned long val, next;
887 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
888 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
889 /* from time_after() in jiffies.h */
890 if ((long)(next - val) < 0) {
892 case MEM_CGROUP_TARGET_THRESH:
893 next = val + THRESHOLDS_EVENTS_TARGET;
895 case MEM_CGROUP_TARGET_SOFTLIMIT:
896 next = val + SOFTLIMIT_EVENTS_TARGET;
898 case MEM_CGROUP_TARGET_NUMAINFO:
899 next = val + NUMAINFO_EVENTS_TARGET;
904 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
911 * Check events in order.
914 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
916 /* threshold event is triggered in finer grain than soft limit */
917 if (unlikely(mem_cgroup_event_ratelimit(memcg,
918 MEM_CGROUP_TARGET_THRESH))) {
920 bool do_numainfo __maybe_unused;
922 do_softlimit = mem_cgroup_event_ratelimit(memcg,
923 MEM_CGROUP_TARGET_SOFTLIMIT);
925 do_numainfo = mem_cgroup_event_ratelimit(memcg,
926 MEM_CGROUP_TARGET_NUMAINFO);
928 mem_cgroup_threshold(memcg);
929 if (unlikely(do_softlimit))
930 mem_cgroup_update_tree(memcg, page);
932 if (unlikely(do_numainfo))
933 atomic_inc(&memcg->numainfo_events);
938 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
941 * mm_update_next_owner() may clear mm->owner to NULL
942 * if it races with swapoff, page migration, etc.
943 * So this can be called with p == NULL.
948 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
950 EXPORT_SYMBOL(mem_cgroup_from_task);
953 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
954 * @mm: mm from which memcg should be extracted. It can be NULL.
956 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
957 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
960 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
962 struct mem_cgroup *memcg;
964 if (mem_cgroup_disabled())
970 * Page cache insertions can happen withou an
971 * actual mm context, e.g. during disk probing
972 * on boot, loopback IO, acct() writes etc.
975 memcg = root_mem_cgroup;
977 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
978 if (unlikely(!memcg))
979 memcg = root_mem_cgroup;
981 } while (!css_tryget(&memcg->css));
985 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
988 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
989 * @page: page from which memcg should be extracted.
991 * Obtain a reference on page->memcg and returns it if successful. Otherwise
992 * root_mem_cgroup is returned.
994 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
996 struct mem_cgroup *memcg = page->mem_cgroup;
998 if (mem_cgroup_disabled())
1002 if (!memcg || !css_tryget_online(&memcg->css))
1003 memcg = root_mem_cgroup;
1007 EXPORT_SYMBOL(get_mem_cgroup_from_page);
1010 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
1012 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
1014 if (unlikely(current->active_memcg)) {
1015 struct mem_cgroup *memcg = root_mem_cgroup;
1018 if (css_tryget_online(¤t->active_memcg->css))
1019 memcg = current->active_memcg;
1023 return get_mem_cgroup_from_mm(current->mm);
1027 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1028 * @root: hierarchy root
1029 * @prev: previously returned memcg, NULL on first invocation
1030 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1032 * Returns references to children of the hierarchy below @root, or
1033 * @root itself, or %NULL after a full round-trip.
1035 * Caller must pass the return value in @prev on subsequent
1036 * invocations for reference counting, or use mem_cgroup_iter_break()
1037 * to cancel a hierarchy walk before the round-trip is complete.
1039 * Reclaimers can specify a node and a priority level in @reclaim to
1040 * divide up the memcgs in the hierarchy among all concurrent
1041 * reclaimers operating on the same node and priority.
1043 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1044 struct mem_cgroup *prev,
1045 struct mem_cgroup_reclaim_cookie *reclaim)
1047 struct mem_cgroup_reclaim_iter *iter;
1048 struct cgroup_subsys_state *css = NULL;
1049 struct mem_cgroup *memcg = NULL;
1050 struct mem_cgroup *pos = NULL;
1052 if (mem_cgroup_disabled())
1056 root = root_mem_cgroup;
1058 if (prev && !reclaim)
1061 if (!root->use_hierarchy && root != root_mem_cgroup) {
1070 struct mem_cgroup_per_node *mz;
1072 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1073 iter = &mz->iter[reclaim->priority];
1075 if (prev && reclaim->generation != iter->generation)
1079 pos = READ_ONCE(iter->position);
1080 if (!pos || css_tryget(&pos->css))
1083 * css reference reached zero, so iter->position will
1084 * be cleared by ->css_released. However, we should not
1085 * rely on this happening soon, because ->css_released
1086 * is called from a work queue, and by busy-waiting we
1087 * might block it. So we clear iter->position right
1090 (void)cmpxchg(&iter->position, pos, NULL);
1098 css = css_next_descendant_pre(css, &root->css);
1101 * Reclaimers share the hierarchy walk, and a
1102 * new one might jump in right at the end of
1103 * the hierarchy - make sure they see at least
1104 * one group and restart from the beginning.
1112 * Verify the css and acquire a reference. The root
1113 * is provided by the caller, so we know it's alive
1114 * and kicking, and don't take an extra reference.
1116 memcg = mem_cgroup_from_css(css);
1118 if (css == &root->css)
1121 if (css_tryget(css))
1129 * The position could have already been updated by a competing
1130 * thread, so check that the value hasn't changed since we read
1131 * it to avoid reclaiming from the same cgroup twice.
1133 (void)cmpxchg(&iter->position, pos, memcg);
1141 reclaim->generation = iter->generation;
1147 if (prev && prev != root)
1148 css_put(&prev->css);
1154 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1155 * @root: hierarchy root
1156 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1158 void mem_cgroup_iter_break(struct mem_cgroup *root,
1159 struct mem_cgroup *prev)
1162 root = root_mem_cgroup;
1163 if (prev && prev != root)
1164 css_put(&prev->css);
1167 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1168 struct mem_cgroup *dead_memcg)
1170 struct mem_cgroup_reclaim_iter *iter;
1171 struct mem_cgroup_per_node *mz;
1175 for_each_node(nid) {
1176 mz = mem_cgroup_nodeinfo(from, nid);
1177 for (i = 0; i <= DEF_PRIORITY; i++) {
1178 iter = &mz->iter[i];
1179 cmpxchg(&iter->position,
1185 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1187 struct mem_cgroup *memcg = dead_memcg;
1188 struct mem_cgroup *last;
1191 __invalidate_reclaim_iterators(memcg, dead_memcg);
1193 } while ((memcg = parent_mem_cgroup(memcg)));
1196 * When cgruop1 non-hierarchy mode is used,
1197 * parent_mem_cgroup() does not walk all the way up to the
1198 * cgroup root (root_mem_cgroup). So we have to handle
1199 * dead_memcg from cgroup root separately.
1201 if (last != root_mem_cgroup)
1202 __invalidate_reclaim_iterators(root_mem_cgroup,
1207 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1208 * @memcg: hierarchy root
1209 * @fn: function to call for each task
1210 * @arg: argument passed to @fn
1212 * This function iterates over tasks attached to @memcg or to any of its
1213 * descendants and calls @fn for each task. If @fn returns a non-zero
1214 * value, the function breaks the iteration loop and returns the value.
1215 * Otherwise, it will iterate over all tasks and return 0.
1217 * This function must not be called for the root memory cgroup.
1219 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1220 int (*fn)(struct task_struct *, void *), void *arg)
1222 struct mem_cgroup *iter;
1225 BUG_ON(memcg == root_mem_cgroup);
1227 for_each_mem_cgroup_tree(iter, memcg) {
1228 struct css_task_iter it;
1229 struct task_struct *task;
1231 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1232 while (!ret && (task = css_task_iter_next(&it)))
1233 ret = fn(task, arg);
1234 css_task_iter_end(&it);
1236 mem_cgroup_iter_break(memcg, iter);
1244 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1246 * @pgdat: pgdat of the page
1248 * This function is only safe when following the LRU page isolation
1249 * and putback protocol: the LRU lock must be held, and the page must
1250 * either be PageLRU() or the caller must have isolated/allocated it.
1252 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1254 struct mem_cgroup_per_node *mz;
1255 struct mem_cgroup *memcg;
1256 struct lruvec *lruvec;
1258 if (mem_cgroup_disabled()) {
1259 lruvec = &pgdat->lruvec;
1263 memcg = page->mem_cgroup;
1265 * Swapcache readahead pages are added to the LRU - and
1266 * possibly migrated - before they are charged.
1269 memcg = root_mem_cgroup;
1271 mz = mem_cgroup_page_nodeinfo(memcg, page);
1272 lruvec = &mz->lruvec;
1275 * Since a node can be onlined after the mem_cgroup was created,
1276 * we have to be prepared to initialize lruvec->zone here;
1277 * and if offlined then reonlined, we need to reinitialize it.
1279 if (unlikely(lruvec->pgdat != pgdat))
1280 lruvec->pgdat = pgdat;
1285 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1286 * @lruvec: mem_cgroup per zone lru vector
1287 * @lru: index of lru list the page is sitting on
1288 * @zid: zone id of the accounted pages
1289 * @nr_pages: positive when adding or negative when removing
1291 * This function must be called under lru_lock, just before a page is added
1292 * to or just after a page is removed from an lru list (that ordering being
1293 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1295 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1296 int zid, int nr_pages)
1298 struct mem_cgroup_per_node *mz;
1299 unsigned long *lru_size;
1302 if (mem_cgroup_disabled())
1305 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1306 lru_size = &mz->lru_zone_size[zid][lru];
1309 *lru_size += nr_pages;
1312 if (WARN_ONCE(size < 0,
1313 "%s(%p, %d, %d): lru_size %ld\n",
1314 __func__, lruvec, lru, nr_pages, size)) {
1320 *lru_size += nr_pages;
1324 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1325 * @memcg: the memory cgroup
1327 * Returns the maximum amount of memory @mem can be charged with, in
1330 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1332 unsigned long margin = 0;
1333 unsigned long count;
1334 unsigned long limit;
1336 count = page_counter_read(&memcg->memory);
1337 limit = READ_ONCE(memcg->memory.max);
1339 margin = limit - count;
1341 if (do_memsw_account()) {
1342 count = page_counter_read(&memcg->memsw);
1343 limit = READ_ONCE(memcg->memsw.max);
1345 margin = min(margin, limit - count);
1354 * A routine for checking "mem" is under move_account() or not.
1356 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1357 * moving cgroups. This is for waiting at high-memory pressure
1360 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1362 struct mem_cgroup *from;
1363 struct mem_cgroup *to;
1366 * Unlike task_move routines, we access mc.to, mc.from not under
1367 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1369 spin_lock(&mc.lock);
1375 ret = mem_cgroup_is_descendant(from, memcg) ||
1376 mem_cgroup_is_descendant(to, memcg);
1378 spin_unlock(&mc.lock);
1382 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1384 if (mc.moving_task && current != mc.moving_task) {
1385 if (mem_cgroup_under_move(memcg)) {
1387 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1388 /* moving charge context might have finished. */
1391 finish_wait(&mc.waitq, &wait);
1398 static char *memory_stat_format(struct mem_cgroup *memcg)
1403 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1408 * Provide statistics on the state of the memory subsystem as
1409 * well as cumulative event counters that show past behavior.
1411 * This list is ordered following a combination of these gradients:
1412 * 1) generic big picture -> specifics and details
1413 * 2) reflecting userspace activity -> reflecting kernel heuristics
1415 * Current memory state:
1418 seq_buf_printf(&s, "anon %llu\n",
1419 (u64)memcg_page_state(memcg, MEMCG_RSS) *
1421 seq_buf_printf(&s, "file %llu\n",
1422 (u64)memcg_page_state(memcg, MEMCG_CACHE) *
1424 seq_buf_printf(&s, "kernel_stack %llu\n",
1425 (u64)memcg_page_state(memcg, MEMCG_KERNEL_STACK_KB) *
1427 seq_buf_printf(&s, "slab %llu\n",
1428 (u64)(memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) +
1429 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE)) *
1431 seq_buf_printf(&s, "sock %llu\n",
1432 (u64)memcg_page_state(memcg, MEMCG_SOCK) *
1435 seq_buf_printf(&s, "shmem %llu\n",
1436 (u64)memcg_page_state(memcg, NR_SHMEM) *
1438 seq_buf_printf(&s, "file_mapped %llu\n",
1439 (u64)memcg_page_state(memcg, NR_FILE_MAPPED) *
1441 seq_buf_printf(&s, "file_dirty %llu\n",
1442 (u64)memcg_page_state(memcg, NR_FILE_DIRTY) *
1444 seq_buf_printf(&s, "file_writeback %llu\n",
1445 (u64)memcg_page_state(memcg, NR_WRITEBACK) *
1449 * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter
1450 * with the NR_ANON_THP vm counter, but right now it's a pain in the
1451 * arse because it requires migrating the work out of rmap to a place
1452 * where the page->mem_cgroup is set up and stable.
1454 seq_buf_printf(&s, "anon_thp %llu\n",
1455 (u64)memcg_page_state(memcg, MEMCG_RSS_HUGE) *
1458 for (i = 0; i < NR_LRU_LISTS; i++)
1459 seq_buf_printf(&s, "%s %llu\n", mem_cgroup_lru_names[i],
1460 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
1463 seq_buf_printf(&s, "slab_reclaimable %llu\n",
1464 (u64)memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) *
1466 seq_buf_printf(&s, "slab_unreclaimable %llu\n",
1467 (u64)memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE) *
1470 /* Accumulated memory events */
1472 seq_buf_printf(&s, "pgfault %lu\n", memcg_events(memcg, PGFAULT));
1473 seq_buf_printf(&s, "pgmajfault %lu\n", memcg_events(memcg, PGMAJFAULT));
1475 seq_buf_printf(&s, "workingset_refault %lu\n",
1476 memcg_page_state(memcg, WORKINGSET_REFAULT));
1477 seq_buf_printf(&s, "workingset_activate %lu\n",
1478 memcg_page_state(memcg, WORKINGSET_ACTIVATE));
1479 seq_buf_printf(&s, "workingset_nodereclaim %lu\n",
1480 memcg_page_state(memcg, WORKINGSET_NODERECLAIM));
1482 seq_buf_printf(&s, "pgrefill %lu\n", memcg_events(memcg, PGREFILL));
1483 seq_buf_printf(&s, "pgscan %lu\n",
1484 memcg_events(memcg, PGSCAN_KSWAPD) +
1485 memcg_events(memcg, PGSCAN_DIRECT));
1486 seq_buf_printf(&s, "pgsteal %lu\n",
1487 memcg_events(memcg, PGSTEAL_KSWAPD) +
1488 memcg_events(memcg, PGSTEAL_DIRECT));
1489 seq_buf_printf(&s, "pgactivate %lu\n", memcg_events(memcg, PGACTIVATE));
1490 seq_buf_printf(&s, "pgdeactivate %lu\n", memcg_events(memcg, PGDEACTIVATE));
1491 seq_buf_printf(&s, "pglazyfree %lu\n", memcg_events(memcg, PGLAZYFREE));
1492 seq_buf_printf(&s, "pglazyfreed %lu\n", memcg_events(memcg, PGLAZYFREED));
1494 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1495 seq_buf_printf(&s, "thp_fault_alloc %lu\n",
1496 memcg_events(memcg, THP_FAULT_ALLOC));
1497 seq_buf_printf(&s, "thp_collapse_alloc %lu\n",
1498 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1499 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1501 /* The above should easily fit into one page */
1502 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1507 #define K(x) ((x) << (PAGE_SHIFT-10))
1509 * mem_cgroup_print_oom_context: Print OOM information relevant to
1510 * memory controller.
1511 * @memcg: The memory cgroup that went over limit
1512 * @p: Task that is going to be killed
1514 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1517 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1522 pr_cont(",oom_memcg=");
1523 pr_cont_cgroup_path(memcg->css.cgroup);
1525 pr_cont(",global_oom");
1527 pr_cont(",task_memcg=");
1528 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1534 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1535 * memory controller.
1536 * @memcg: The memory cgroup that went over limit
1538 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1542 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1543 K((u64)page_counter_read(&memcg->memory)),
1544 K((u64)memcg->memory.max), memcg->memory.failcnt);
1545 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1546 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1547 K((u64)page_counter_read(&memcg->swap)),
1548 K((u64)memcg->swap.max), memcg->swap.failcnt);
1550 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1551 K((u64)page_counter_read(&memcg->memsw)),
1552 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1553 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1554 K((u64)page_counter_read(&memcg->kmem)),
1555 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1558 pr_info("Memory cgroup stats for ");
1559 pr_cont_cgroup_path(memcg->css.cgroup);
1561 buf = memory_stat_format(memcg);
1569 * Return the memory (and swap, if configured) limit for a memcg.
1571 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1575 max = memcg->memory.max;
1576 if (mem_cgroup_swappiness(memcg)) {
1577 unsigned long memsw_max;
1578 unsigned long swap_max;
1580 memsw_max = memcg->memsw.max;
1581 swap_max = memcg->swap.max;
1582 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1583 max = min(max + swap_max, memsw_max);
1588 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1590 return page_counter_read(&memcg->memory);
1593 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1596 struct oom_control oc = {
1600 .gfp_mask = gfp_mask,
1605 if (mutex_lock_killable(&oom_lock))
1608 * A few threads which were not waiting at mutex_lock_killable() can
1609 * fail to bail out. Therefore, check again after holding oom_lock.
1611 ret = should_force_charge() || out_of_memory(&oc);
1612 mutex_unlock(&oom_lock);
1616 #if MAX_NUMNODES > 1
1619 * test_mem_cgroup_node_reclaimable
1620 * @memcg: the target memcg
1621 * @nid: the node ID to be checked.
1622 * @noswap : specify true here if the user wants flle only information.
1624 * This function returns whether the specified memcg contains any
1625 * reclaimable pages on a node. Returns true if there are any reclaimable
1626 * pages in the node.
1628 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1629 int nid, bool noswap)
1631 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
1633 if (lruvec_page_state(lruvec, NR_INACTIVE_FILE) ||
1634 lruvec_page_state(lruvec, NR_ACTIVE_FILE))
1636 if (noswap || !total_swap_pages)
1638 if (lruvec_page_state(lruvec, NR_INACTIVE_ANON) ||
1639 lruvec_page_state(lruvec, NR_ACTIVE_ANON))
1646 * Always updating the nodemask is not very good - even if we have an empty
1647 * list or the wrong list here, we can start from some node and traverse all
1648 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1651 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1655 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1656 * pagein/pageout changes since the last update.
1658 if (!atomic_read(&memcg->numainfo_events))
1660 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1663 /* make a nodemask where this memcg uses memory from */
1664 memcg->scan_nodes = node_states[N_MEMORY];
1666 for_each_node_mask(nid, node_states[N_MEMORY]) {
1668 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1669 node_clear(nid, memcg->scan_nodes);
1672 atomic_set(&memcg->numainfo_events, 0);
1673 atomic_set(&memcg->numainfo_updating, 0);
1677 * Selecting a node where we start reclaim from. Because what we need is just
1678 * reducing usage counter, start from anywhere is O,K. Considering
1679 * memory reclaim from current node, there are pros. and cons.
1681 * Freeing memory from current node means freeing memory from a node which
1682 * we'll use or we've used. So, it may make LRU bad. And if several threads
1683 * hit limits, it will see a contention on a node. But freeing from remote
1684 * node means more costs for memory reclaim because of memory latency.
1686 * Now, we use round-robin. Better algorithm is welcomed.
1688 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1692 mem_cgroup_may_update_nodemask(memcg);
1693 node = memcg->last_scanned_node;
1695 node = next_node_in(node, memcg->scan_nodes);
1697 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1698 * last time it really checked all the LRUs due to rate limiting.
1699 * Fallback to the current node in that case for simplicity.
1701 if (unlikely(node == MAX_NUMNODES))
1702 node = numa_node_id();
1704 memcg->last_scanned_node = node;
1708 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1714 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1717 unsigned long *total_scanned)
1719 struct mem_cgroup *victim = NULL;
1722 unsigned long excess;
1723 unsigned long nr_scanned;
1724 struct mem_cgroup_reclaim_cookie reclaim = {
1729 excess = soft_limit_excess(root_memcg);
1732 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1737 * If we have not been able to reclaim
1738 * anything, it might because there are
1739 * no reclaimable pages under this hierarchy
1744 * We want to do more targeted reclaim.
1745 * excess >> 2 is not to excessive so as to
1746 * reclaim too much, nor too less that we keep
1747 * coming back to reclaim from this cgroup
1749 if (total >= (excess >> 2) ||
1750 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1755 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1756 pgdat, &nr_scanned);
1757 *total_scanned += nr_scanned;
1758 if (!soft_limit_excess(root_memcg))
1761 mem_cgroup_iter_break(root_memcg, victim);
1765 #ifdef CONFIG_LOCKDEP
1766 static struct lockdep_map memcg_oom_lock_dep_map = {
1767 .name = "memcg_oom_lock",
1771 static DEFINE_SPINLOCK(memcg_oom_lock);
1774 * Check OOM-Killer is already running under our hierarchy.
1775 * If someone is running, return false.
1777 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1779 struct mem_cgroup *iter, *failed = NULL;
1781 spin_lock(&memcg_oom_lock);
1783 for_each_mem_cgroup_tree(iter, memcg) {
1784 if (iter->oom_lock) {
1786 * this subtree of our hierarchy is already locked
1787 * so we cannot give a lock.
1790 mem_cgroup_iter_break(memcg, iter);
1793 iter->oom_lock = true;
1798 * OK, we failed to lock the whole subtree so we have
1799 * to clean up what we set up to the failing subtree
1801 for_each_mem_cgroup_tree(iter, memcg) {
1802 if (iter == failed) {
1803 mem_cgroup_iter_break(memcg, iter);
1806 iter->oom_lock = false;
1809 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1811 spin_unlock(&memcg_oom_lock);
1816 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1818 struct mem_cgroup *iter;
1820 spin_lock(&memcg_oom_lock);
1821 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1822 for_each_mem_cgroup_tree(iter, memcg)
1823 iter->oom_lock = false;
1824 spin_unlock(&memcg_oom_lock);
1827 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1829 struct mem_cgroup *iter;
1831 spin_lock(&memcg_oom_lock);
1832 for_each_mem_cgroup_tree(iter, memcg)
1834 spin_unlock(&memcg_oom_lock);
1837 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1839 struct mem_cgroup *iter;
1842 * When a new child is created while the hierarchy is under oom,
1843 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1845 spin_lock(&memcg_oom_lock);
1846 for_each_mem_cgroup_tree(iter, memcg)
1847 if (iter->under_oom > 0)
1849 spin_unlock(&memcg_oom_lock);
1852 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1854 struct oom_wait_info {
1855 struct mem_cgroup *memcg;
1856 wait_queue_entry_t wait;
1859 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1860 unsigned mode, int sync, void *arg)
1862 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1863 struct mem_cgroup *oom_wait_memcg;
1864 struct oom_wait_info *oom_wait_info;
1866 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1867 oom_wait_memcg = oom_wait_info->memcg;
1869 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1870 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1872 return autoremove_wake_function(wait, mode, sync, arg);
1875 static void memcg_oom_recover(struct mem_cgroup *memcg)
1878 * For the following lockless ->under_oom test, the only required
1879 * guarantee is that it must see the state asserted by an OOM when
1880 * this function is called as a result of userland actions
1881 * triggered by the notification of the OOM. This is trivially
1882 * achieved by invoking mem_cgroup_mark_under_oom() before
1883 * triggering notification.
1885 if (memcg && memcg->under_oom)
1886 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1896 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1898 enum oom_status ret;
1901 if (order > PAGE_ALLOC_COSTLY_ORDER)
1904 memcg_memory_event(memcg, MEMCG_OOM);
1907 * We are in the middle of the charge context here, so we
1908 * don't want to block when potentially sitting on a callstack
1909 * that holds all kinds of filesystem and mm locks.
1911 * cgroup1 allows disabling the OOM killer and waiting for outside
1912 * handling until the charge can succeed; remember the context and put
1913 * the task to sleep at the end of the page fault when all locks are
1916 * On the other hand, in-kernel OOM killer allows for an async victim
1917 * memory reclaim (oom_reaper) and that means that we are not solely
1918 * relying on the oom victim to make a forward progress and we can
1919 * invoke the oom killer here.
1921 * Please note that mem_cgroup_out_of_memory might fail to find a
1922 * victim and then we have to bail out from the charge path.
1924 if (memcg->oom_kill_disable) {
1925 if (!current->in_user_fault)
1927 css_get(&memcg->css);
1928 current->memcg_in_oom = memcg;
1929 current->memcg_oom_gfp_mask = mask;
1930 current->memcg_oom_order = order;
1935 mem_cgroup_mark_under_oom(memcg);
1937 locked = mem_cgroup_oom_trylock(memcg);
1940 mem_cgroup_oom_notify(memcg);
1942 mem_cgroup_unmark_under_oom(memcg);
1943 if (mem_cgroup_out_of_memory(memcg, mask, order))
1949 mem_cgroup_oom_unlock(memcg);
1955 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1956 * @handle: actually kill/wait or just clean up the OOM state
1958 * This has to be called at the end of a page fault if the memcg OOM
1959 * handler was enabled.
1961 * Memcg supports userspace OOM handling where failed allocations must
1962 * sleep on a waitqueue until the userspace task resolves the
1963 * situation. Sleeping directly in the charge context with all kinds
1964 * of locks held is not a good idea, instead we remember an OOM state
1965 * in the task and mem_cgroup_oom_synchronize() has to be called at
1966 * the end of the page fault to complete the OOM handling.
1968 * Returns %true if an ongoing memcg OOM situation was detected and
1969 * completed, %false otherwise.
1971 bool mem_cgroup_oom_synchronize(bool handle)
1973 struct mem_cgroup *memcg = current->memcg_in_oom;
1974 struct oom_wait_info owait;
1977 /* OOM is global, do not handle */
1984 owait.memcg = memcg;
1985 owait.wait.flags = 0;
1986 owait.wait.func = memcg_oom_wake_function;
1987 owait.wait.private = current;
1988 INIT_LIST_HEAD(&owait.wait.entry);
1990 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1991 mem_cgroup_mark_under_oom(memcg);
1993 locked = mem_cgroup_oom_trylock(memcg);
1996 mem_cgroup_oom_notify(memcg);
1998 if (locked && !memcg->oom_kill_disable) {
1999 mem_cgroup_unmark_under_oom(memcg);
2000 finish_wait(&memcg_oom_waitq, &owait.wait);
2001 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
2002 current->memcg_oom_order);
2005 mem_cgroup_unmark_under_oom(memcg);
2006 finish_wait(&memcg_oom_waitq, &owait.wait);
2010 mem_cgroup_oom_unlock(memcg);
2012 * There is no guarantee that an OOM-lock contender
2013 * sees the wakeups triggered by the OOM kill
2014 * uncharges. Wake any sleepers explicitely.
2016 memcg_oom_recover(memcg);
2019 current->memcg_in_oom = NULL;
2020 css_put(&memcg->css);
2025 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2026 * @victim: task to be killed by the OOM killer
2027 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2029 * Returns a pointer to a memory cgroup, which has to be cleaned up
2030 * by killing all belonging OOM-killable tasks.
2032 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2034 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2035 struct mem_cgroup *oom_domain)
2037 struct mem_cgroup *oom_group = NULL;
2038 struct mem_cgroup *memcg;
2040 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2044 oom_domain = root_mem_cgroup;
2048 memcg = mem_cgroup_from_task(victim);
2049 if (memcg == root_mem_cgroup)
2053 * Traverse the memory cgroup hierarchy from the victim task's
2054 * cgroup up to the OOMing cgroup (or root) to find the
2055 * highest-level memory cgroup with oom.group set.
2057 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2058 if (memcg->oom_group)
2061 if (memcg == oom_domain)
2066 css_get(&oom_group->css);
2073 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2075 pr_info("Tasks in ");
2076 pr_cont_cgroup_path(memcg->css.cgroup);
2077 pr_cont(" are going to be killed due to memory.oom.group set\n");
2081 * lock_page_memcg - lock a page->mem_cgroup binding
2084 * This function protects unlocked LRU pages from being moved to
2087 * It ensures lifetime of the returned memcg. Caller is responsible
2088 * for the lifetime of the page; __unlock_page_memcg() is available
2089 * when @page might get freed inside the locked section.
2091 struct mem_cgroup *lock_page_memcg(struct page *page)
2093 struct mem_cgroup *memcg;
2094 unsigned long flags;
2097 * The RCU lock is held throughout the transaction. The fast
2098 * path can get away without acquiring the memcg->move_lock
2099 * because page moving starts with an RCU grace period.
2101 * The RCU lock also protects the memcg from being freed when
2102 * the page state that is going to change is the only thing
2103 * preventing the page itself from being freed. E.g. writeback
2104 * doesn't hold a page reference and relies on PG_writeback to
2105 * keep off truncation, migration and so forth.
2109 if (mem_cgroup_disabled())
2112 memcg = page->mem_cgroup;
2113 if (unlikely(!memcg))
2116 if (atomic_read(&memcg->moving_account) <= 0)
2119 spin_lock_irqsave(&memcg->move_lock, flags);
2120 if (memcg != page->mem_cgroup) {
2121 spin_unlock_irqrestore(&memcg->move_lock, flags);
2126 * When charge migration first begins, we can have locked and
2127 * unlocked page stat updates happening concurrently. Track
2128 * the task who has the lock for unlock_page_memcg().
2130 memcg->move_lock_task = current;
2131 memcg->move_lock_flags = flags;
2135 EXPORT_SYMBOL(lock_page_memcg);
2138 * __unlock_page_memcg - unlock and unpin a memcg
2141 * Unlock and unpin a memcg returned by lock_page_memcg().
2143 void __unlock_page_memcg(struct mem_cgroup *memcg)
2145 if (memcg && memcg->move_lock_task == current) {
2146 unsigned long flags = memcg->move_lock_flags;
2148 memcg->move_lock_task = NULL;
2149 memcg->move_lock_flags = 0;
2151 spin_unlock_irqrestore(&memcg->move_lock, flags);
2158 * unlock_page_memcg - unlock a page->mem_cgroup binding
2161 void unlock_page_memcg(struct page *page)
2163 __unlock_page_memcg(page->mem_cgroup);
2165 EXPORT_SYMBOL(unlock_page_memcg);
2167 struct memcg_stock_pcp {
2168 struct mem_cgroup *cached; /* this never be root cgroup */
2169 unsigned int nr_pages;
2170 struct work_struct work;
2171 unsigned long flags;
2172 #define FLUSHING_CACHED_CHARGE 0
2174 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2175 static DEFINE_MUTEX(percpu_charge_mutex);
2178 * consume_stock: Try to consume stocked charge on this cpu.
2179 * @memcg: memcg to consume from.
2180 * @nr_pages: how many pages to charge.
2182 * The charges will only happen if @memcg matches the current cpu's memcg
2183 * stock, and at least @nr_pages are available in that stock. Failure to
2184 * service an allocation will refill the stock.
2186 * returns true if successful, false otherwise.
2188 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2190 struct memcg_stock_pcp *stock;
2191 unsigned long flags;
2194 if (nr_pages > MEMCG_CHARGE_BATCH)
2197 local_irq_save(flags);
2199 stock = this_cpu_ptr(&memcg_stock);
2200 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2201 stock->nr_pages -= nr_pages;
2205 local_irq_restore(flags);
2211 * Returns stocks cached in percpu and reset cached information.
2213 static void drain_stock(struct memcg_stock_pcp *stock)
2215 struct mem_cgroup *old = stock->cached;
2220 if (stock->nr_pages) {
2221 page_counter_uncharge(&old->memory, stock->nr_pages);
2222 if (do_memsw_account())
2223 page_counter_uncharge(&old->memsw, stock->nr_pages);
2224 css_put_many(&old->css, stock->nr_pages);
2225 stock->nr_pages = 0;
2229 stock->cached = NULL;
2232 static void drain_local_stock(struct work_struct *dummy)
2234 struct memcg_stock_pcp *stock;
2235 unsigned long flags;
2238 * The only protection from memory hotplug vs. drain_stock races is
2239 * that we always operate on local CPU stock here with IRQ disabled
2241 local_irq_save(flags);
2243 stock = this_cpu_ptr(&memcg_stock);
2245 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2247 local_irq_restore(flags);
2251 * Cache charges(val) to local per_cpu area.
2252 * This will be consumed by consume_stock() function, later.
2254 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2256 struct memcg_stock_pcp *stock;
2257 unsigned long flags;
2259 local_irq_save(flags);
2261 stock = this_cpu_ptr(&memcg_stock);
2262 if (stock->cached != memcg) { /* reset if necessary */
2264 css_get(&memcg->css);
2265 stock->cached = memcg;
2267 stock->nr_pages += nr_pages;
2269 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2272 local_irq_restore(flags);
2276 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2277 * of the hierarchy under it.
2279 static void drain_all_stock(struct mem_cgroup *root_memcg)
2283 /* If someone's already draining, avoid adding running more workers. */
2284 if (!mutex_trylock(&percpu_charge_mutex))
2287 * Notify other cpus that system-wide "drain" is running
2288 * We do not care about races with the cpu hotplug because cpu down
2289 * as well as workers from this path always operate on the local
2290 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2293 for_each_online_cpu(cpu) {
2294 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2295 struct mem_cgroup *memcg;
2299 memcg = stock->cached;
2300 if (memcg && stock->nr_pages &&
2301 mem_cgroup_is_descendant(memcg, root_memcg))
2306 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2308 drain_local_stock(&stock->work);
2310 schedule_work_on(cpu, &stock->work);
2314 mutex_unlock(&percpu_charge_mutex);
2317 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2319 struct memcg_stock_pcp *stock;
2320 struct mem_cgroup *memcg, *mi;
2322 stock = &per_cpu(memcg_stock, cpu);
2325 for_each_mem_cgroup(memcg) {
2328 for (i = 0; i < MEMCG_NR_STAT; i++) {
2332 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2334 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2335 atomic_long_add(x, &memcg->vmstats[i]);
2337 if (i >= NR_VM_NODE_STAT_ITEMS)
2340 for_each_node(nid) {
2341 struct mem_cgroup_per_node *pn;
2343 pn = mem_cgroup_nodeinfo(memcg, nid);
2344 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2347 atomic_long_add(x, &pn->lruvec_stat[i]);
2348 } while ((pn = parent_nodeinfo(pn, nid)));
2352 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2355 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2357 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2358 atomic_long_add(x, &memcg->vmevents[i]);
2365 static void reclaim_high(struct mem_cgroup *memcg,
2366 unsigned int nr_pages,
2370 if (page_counter_read(&memcg->memory) <= memcg->high)
2372 memcg_memory_event(memcg, MEMCG_HIGH);
2373 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2374 } while ((memcg = parent_mem_cgroup(memcg)));
2377 static void high_work_func(struct work_struct *work)
2379 struct mem_cgroup *memcg;
2381 memcg = container_of(work, struct mem_cgroup, high_work);
2382 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2386 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2387 * enough to still cause a significant slowdown in most cases, while still
2388 * allowing diagnostics and tracing to proceed without becoming stuck.
2390 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2393 * When calculating the delay, we use these either side of the exponentiation to
2394 * maintain precision and scale to a reasonable number of jiffies (see the table
2397 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2398 * overage ratio to a delay.
2399 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down down the
2400 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2401 * to produce a reasonable delay curve.
2403 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2404 * reasonable delay curve compared to precision-adjusted overage, not
2405 * penalising heavily at first, but still making sure that growth beyond the
2406 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2407 * example, with a high of 100 megabytes:
2409 * +-------+------------------------+
2410 * | usage | time to allocate in ms |
2411 * +-------+------------------------+
2433 * +-------+------------------------+
2435 #define MEMCG_DELAY_PRECISION_SHIFT 20
2436 #define MEMCG_DELAY_SCALING_SHIFT 14
2439 * Get the number of jiffies that we should penalise a mischievous cgroup which
2440 * is exceeding its memory.high by checking both it and its ancestors.
2442 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2443 unsigned int nr_pages)
2445 unsigned long penalty_jiffies;
2446 u64 max_overage = 0;
2449 unsigned long usage, high;
2452 usage = page_counter_read(&memcg->memory);
2453 high = READ_ONCE(memcg->high);
2459 * Prevent division by 0 in overage calculation by acting as if
2460 * it was a threshold of 1 page
2462 high = max(high, 1UL);
2464 overage = usage - high;
2465 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2466 overage = div64_u64(overage, high);
2468 if (overage > max_overage)
2469 max_overage = overage;
2470 } while ((memcg = parent_mem_cgroup(memcg)) &&
2471 !mem_cgroup_is_root(memcg));
2477 * We use overage compared to memory.high to calculate the number of
2478 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2479 * fairly lenient on small overages, and increasingly harsh when the
2480 * memcg in question makes it clear that it has no intention of stopping
2481 * its crazy behaviour, so we exponentially increase the delay based on
2484 penalty_jiffies = max_overage * max_overage * HZ;
2485 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2486 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2489 * Factor in the task's own contribution to the overage, such that four
2490 * N-sized allocations are throttled approximately the same as one
2491 * 4N-sized allocation.
2493 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2494 * larger the current charge patch is than that.
2496 penalty_jiffies = penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2499 * Clamp the max delay per usermode return so as to still keep the
2500 * application moving forwards and also permit diagnostics, albeit
2503 return min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2507 * Scheduled by try_charge() to be executed from the userland return path
2508 * and reclaims memory over the high limit.
2510 void mem_cgroup_handle_over_high(void)
2512 unsigned long penalty_jiffies;
2513 unsigned long pflags;
2514 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2515 struct mem_cgroup *memcg;
2517 if (likely(!nr_pages))
2520 memcg = get_mem_cgroup_from_mm(current->mm);
2521 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2522 current->memcg_nr_pages_over_high = 0;
2525 * memory.high is breached and reclaim is unable to keep up. Throttle
2526 * allocators proactively to slow down excessive growth.
2528 penalty_jiffies = calculate_high_delay(memcg, nr_pages);
2531 * Don't sleep if the amount of jiffies this memcg owes us is so low
2532 * that it's not even worth doing, in an attempt to be nice to those who
2533 * go only a small amount over their memory.high value and maybe haven't
2534 * been aggressively reclaimed enough yet.
2536 if (penalty_jiffies <= HZ / 100)
2540 * If we exit early, we're guaranteed to die (since
2541 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2542 * need to account for any ill-begotten jiffies to pay them off later.
2544 psi_memstall_enter(&pflags);
2545 schedule_timeout_killable(penalty_jiffies);
2546 psi_memstall_leave(&pflags);
2549 css_put(&memcg->css);
2552 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2553 unsigned int nr_pages)
2555 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2556 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2557 struct mem_cgroup *mem_over_limit;
2558 struct page_counter *counter;
2559 unsigned long nr_reclaimed;
2560 bool may_swap = true;
2561 bool drained = false;
2562 enum oom_status oom_status;
2564 if (mem_cgroup_is_root(memcg))
2567 if (consume_stock(memcg, nr_pages))
2570 if (!do_memsw_account() ||
2571 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2572 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2574 if (do_memsw_account())
2575 page_counter_uncharge(&memcg->memsw, batch);
2576 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2578 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2582 if (batch > nr_pages) {
2588 * Memcg doesn't have a dedicated reserve for atomic
2589 * allocations. But like the global atomic pool, we need to
2590 * put the burden of reclaim on regular allocation requests
2591 * and let these go through as privileged allocations.
2593 if (gfp_mask & __GFP_ATOMIC)
2597 * Unlike in global OOM situations, memcg is not in a physical
2598 * memory shortage. Allow dying and OOM-killed tasks to
2599 * bypass the last charges so that they can exit quickly and
2600 * free their memory.
2602 if (unlikely(should_force_charge()))
2606 * Prevent unbounded recursion when reclaim operations need to
2607 * allocate memory. This might exceed the limits temporarily,
2608 * but we prefer facilitating memory reclaim and getting back
2609 * under the limit over triggering OOM kills in these cases.
2611 if (unlikely(current->flags & PF_MEMALLOC))
2614 if (unlikely(task_in_memcg_oom(current)))
2617 if (!gfpflags_allow_blocking(gfp_mask))
2620 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2622 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2623 gfp_mask, may_swap);
2625 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2629 drain_all_stock(mem_over_limit);
2634 if (gfp_mask & __GFP_NORETRY)
2637 * Even though the limit is exceeded at this point, reclaim
2638 * may have been able to free some pages. Retry the charge
2639 * before killing the task.
2641 * Only for regular pages, though: huge pages are rather
2642 * unlikely to succeed so close to the limit, and we fall back
2643 * to regular pages anyway in case of failure.
2645 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2648 * At task move, charge accounts can be doubly counted. So, it's
2649 * better to wait until the end of task_move if something is going on.
2651 if (mem_cgroup_wait_acct_move(mem_over_limit))
2657 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2660 if (gfp_mask & __GFP_NOFAIL)
2663 if (fatal_signal_pending(current))
2667 * keep retrying as long as the memcg oom killer is able to make
2668 * a forward progress or bypass the charge if the oom killer
2669 * couldn't make any progress.
2671 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2672 get_order(nr_pages * PAGE_SIZE));
2673 switch (oom_status) {
2675 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2683 if (!(gfp_mask & __GFP_NOFAIL))
2687 * The allocation either can't fail or will lead to more memory
2688 * being freed very soon. Allow memory usage go over the limit
2689 * temporarily by force charging it.
2691 page_counter_charge(&memcg->memory, nr_pages);
2692 if (do_memsw_account())
2693 page_counter_charge(&memcg->memsw, nr_pages);
2694 css_get_many(&memcg->css, nr_pages);
2699 css_get_many(&memcg->css, batch);
2700 if (batch > nr_pages)
2701 refill_stock(memcg, batch - nr_pages);
2704 * If the hierarchy is above the normal consumption range, schedule
2705 * reclaim on returning to userland. We can perform reclaim here
2706 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2707 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2708 * not recorded as it most likely matches current's and won't
2709 * change in the meantime. As high limit is checked again before
2710 * reclaim, the cost of mismatch is negligible.
2713 if (page_counter_read(&memcg->memory) > memcg->high) {
2714 /* Don't bother a random interrupted task */
2715 if (in_interrupt()) {
2716 schedule_work(&memcg->high_work);
2719 current->memcg_nr_pages_over_high += batch;
2720 set_notify_resume(current);
2723 } while ((memcg = parent_mem_cgroup(memcg)));
2728 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2730 if (mem_cgroup_is_root(memcg))
2733 page_counter_uncharge(&memcg->memory, nr_pages);
2734 if (do_memsw_account())
2735 page_counter_uncharge(&memcg->memsw, nr_pages);
2737 css_put_many(&memcg->css, nr_pages);
2740 static void lock_page_lru(struct page *page, int *isolated)
2742 pg_data_t *pgdat = page_pgdat(page);
2744 spin_lock_irq(&pgdat->lru_lock);
2745 if (PageLRU(page)) {
2746 struct lruvec *lruvec;
2748 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2750 del_page_from_lru_list(page, lruvec, page_lru(page));
2756 static void unlock_page_lru(struct page *page, int isolated)
2758 pg_data_t *pgdat = page_pgdat(page);
2761 struct lruvec *lruvec;
2763 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2764 VM_BUG_ON_PAGE(PageLRU(page), page);
2766 add_page_to_lru_list(page, lruvec, page_lru(page));
2768 spin_unlock_irq(&pgdat->lru_lock);
2771 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2776 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2779 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2780 * may already be on some other mem_cgroup's LRU. Take care of it.
2783 lock_page_lru(page, &isolated);
2786 * Nobody should be changing or seriously looking at
2787 * page->mem_cgroup at this point:
2789 * - the page is uncharged
2791 * - the page is off-LRU
2793 * - an anonymous fault has exclusive page access, except for
2794 * a locked page table
2796 * - a page cache insertion, a swapin fault, or a migration
2797 * have the page locked
2799 page->mem_cgroup = memcg;
2802 unlock_page_lru(page, isolated);
2805 #ifdef CONFIG_MEMCG_KMEM
2807 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2809 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2810 * cgroup_mutex, etc.
2812 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2816 if (mem_cgroup_disabled())
2819 page = virt_to_head_page(p);
2822 * Slab pages don't have page->mem_cgroup set because corresponding
2823 * kmem caches can be reparented during the lifetime. That's why
2824 * memcg_from_slab_page() should be used instead.
2827 return memcg_from_slab_page(page);
2829 /* All other pages use page->mem_cgroup */
2830 return page->mem_cgroup;
2833 static int memcg_alloc_cache_id(void)
2838 id = ida_simple_get(&memcg_cache_ida,
2839 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2843 if (id < memcg_nr_cache_ids)
2847 * There's no space for the new id in memcg_caches arrays,
2848 * so we have to grow them.
2850 down_write(&memcg_cache_ids_sem);
2852 size = 2 * (id + 1);
2853 if (size < MEMCG_CACHES_MIN_SIZE)
2854 size = MEMCG_CACHES_MIN_SIZE;
2855 else if (size > MEMCG_CACHES_MAX_SIZE)
2856 size = MEMCG_CACHES_MAX_SIZE;
2858 err = memcg_update_all_caches(size);
2860 err = memcg_update_all_list_lrus(size);
2862 memcg_nr_cache_ids = size;
2864 up_write(&memcg_cache_ids_sem);
2867 ida_simple_remove(&memcg_cache_ida, id);
2873 static void memcg_free_cache_id(int id)
2875 ida_simple_remove(&memcg_cache_ida, id);
2878 struct memcg_kmem_cache_create_work {
2879 struct mem_cgroup *memcg;
2880 struct kmem_cache *cachep;
2881 struct work_struct work;
2884 static void memcg_kmem_cache_create_func(struct work_struct *w)
2886 struct memcg_kmem_cache_create_work *cw =
2887 container_of(w, struct memcg_kmem_cache_create_work, work);
2888 struct mem_cgroup *memcg = cw->memcg;
2889 struct kmem_cache *cachep = cw->cachep;
2891 memcg_create_kmem_cache(memcg, cachep);
2893 css_put(&memcg->css);
2898 * Enqueue the creation of a per-memcg kmem_cache.
2900 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2901 struct kmem_cache *cachep)
2903 struct memcg_kmem_cache_create_work *cw;
2905 if (!css_tryget_online(&memcg->css))
2908 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2910 css_put(&memcg->css);
2915 cw->cachep = cachep;
2916 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2918 queue_work(memcg_kmem_cache_wq, &cw->work);
2921 static inline bool memcg_kmem_bypass(void)
2923 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2929 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2930 * @cachep: the original global kmem cache
2932 * Return the kmem_cache we're supposed to use for a slab allocation.
2933 * We try to use the current memcg's version of the cache.
2935 * If the cache does not exist yet, if we are the first user of it, we
2936 * create it asynchronously in a workqueue and let the current allocation
2937 * go through with the original cache.
2939 * This function takes a reference to the cache it returns to assure it
2940 * won't get destroyed while we are working with it. Once the caller is
2941 * done with it, memcg_kmem_put_cache() must be called to release the
2944 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2946 struct mem_cgroup *memcg;
2947 struct kmem_cache *memcg_cachep;
2948 struct memcg_cache_array *arr;
2951 VM_BUG_ON(!is_root_cache(cachep));
2953 if (memcg_kmem_bypass())
2958 if (unlikely(current->active_memcg))
2959 memcg = current->active_memcg;
2961 memcg = mem_cgroup_from_task(current);
2963 if (!memcg || memcg == root_mem_cgroup)
2966 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2970 arr = rcu_dereference(cachep->memcg_params.memcg_caches);
2973 * Make sure we will access the up-to-date value. The code updating
2974 * memcg_caches issues a write barrier to match the data dependency
2975 * barrier inside READ_ONCE() (see memcg_create_kmem_cache()).
2977 memcg_cachep = READ_ONCE(arr->entries[kmemcg_id]);
2980 * If we are in a safe context (can wait, and not in interrupt
2981 * context), we could be be predictable and return right away.
2982 * This would guarantee that the allocation being performed
2983 * already belongs in the new cache.
2985 * However, there are some clashes that can arrive from locking.
2986 * For instance, because we acquire the slab_mutex while doing
2987 * memcg_create_kmem_cache, this means no further allocation
2988 * could happen with the slab_mutex held. So it's better to
2991 * If the memcg is dying or memcg_cache is about to be released,
2992 * don't bother creating new kmem_caches. Because memcg_cachep
2993 * is ZEROed as the fist step of kmem offlining, we don't need
2994 * percpu_ref_tryget_live() here. css_tryget_online() check in
2995 * memcg_schedule_kmem_cache_create() will prevent us from
2996 * creation of a new kmem_cache.
2998 if (unlikely(!memcg_cachep))
2999 memcg_schedule_kmem_cache_create(memcg, cachep);
3000 else if (percpu_ref_tryget(&memcg_cachep->memcg_params.refcnt))
3001 cachep = memcg_cachep;
3008 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
3009 * @cachep: the cache returned by memcg_kmem_get_cache
3011 void memcg_kmem_put_cache(struct kmem_cache *cachep)
3013 if (!is_root_cache(cachep))
3014 percpu_ref_put(&cachep->memcg_params.refcnt);
3018 * __memcg_kmem_charge_memcg: charge a kmem page
3019 * @page: page to charge
3020 * @gfp: reclaim mode
3021 * @order: allocation order
3022 * @memcg: memory cgroup to charge
3024 * Returns 0 on success, an error code on failure.
3026 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
3027 struct mem_cgroup *memcg)
3029 unsigned int nr_pages = 1 << order;
3030 struct page_counter *counter;
3033 ret = try_charge(memcg, gfp, nr_pages);
3037 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3038 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3041 * Enforce __GFP_NOFAIL allocation because callers are not
3042 * prepared to see failures and likely do not have any failure
3045 if (gfp & __GFP_NOFAIL) {
3046 page_counter_charge(&memcg->kmem, nr_pages);
3049 cancel_charge(memcg, nr_pages);
3056 * __memcg_kmem_charge: charge a kmem page to the current memory cgroup
3057 * @page: page to charge
3058 * @gfp: reclaim mode
3059 * @order: allocation order
3061 * Returns 0 on success, an error code on failure.
3063 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
3065 struct mem_cgroup *memcg;
3068 if (memcg_kmem_bypass())
3071 memcg = get_mem_cgroup_from_current();
3072 if (!mem_cgroup_is_root(memcg)) {
3073 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
3075 page->mem_cgroup = memcg;
3076 __SetPageKmemcg(page);
3079 css_put(&memcg->css);
3084 * __memcg_kmem_uncharge_memcg: uncharge a kmem page
3085 * @memcg: memcg to uncharge
3086 * @nr_pages: number of pages to uncharge
3088 void __memcg_kmem_uncharge_memcg(struct mem_cgroup *memcg,
3089 unsigned int nr_pages)
3091 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3092 page_counter_uncharge(&memcg->kmem, nr_pages);
3094 page_counter_uncharge(&memcg->memory, nr_pages);
3095 if (do_memsw_account())
3096 page_counter_uncharge(&memcg->memsw, nr_pages);
3099 * __memcg_kmem_uncharge: uncharge a kmem page
3100 * @page: page to uncharge
3101 * @order: allocation order
3103 void __memcg_kmem_uncharge(struct page *page, int order)
3105 struct mem_cgroup *memcg = page->mem_cgroup;
3106 unsigned int nr_pages = 1 << order;
3111 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3112 __memcg_kmem_uncharge_memcg(memcg, nr_pages);
3113 page->mem_cgroup = NULL;
3115 /* slab pages do not have PageKmemcg flag set */
3116 if (PageKmemcg(page))
3117 __ClearPageKmemcg(page);
3119 css_put_many(&memcg->css, nr_pages);
3121 #endif /* CONFIG_MEMCG_KMEM */
3123 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3126 * Because tail pages are not marked as "used", set it. We're under
3127 * pgdat->lru_lock and migration entries setup in all page mappings.
3129 void mem_cgroup_split_huge_fixup(struct page *head)
3133 if (mem_cgroup_disabled())
3136 for (i = 1; i < HPAGE_PMD_NR; i++)
3137 head[i].mem_cgroup = head->mem_cgroup;
3139 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
3141 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3143 #ifdef CONFIG_MEMCG_SWAP
3145 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3146 * @entry: swap entry to be moved
3147 * @from: mem_cgroup which the entry is moved from
3148 * @to: mem_cgroup which the entry is moved to
3150 * It succeeds only when the swap_cgroup's record for this entry is the same
3151 * as the mem_cgroup's id of @from.
3153 * Returns 0 on success, -EINVAL on failure.
3155 * The caller must have charged to @to, IOW, called page_counter_charge() about
3156 * both res and memsw, and called css_get().
3158 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3159 struct mem_cgroup *from, struct mem_cgroup *to)
3161 unsigned short old_id, new_id;
3163 old_id = mem_cgroup_id(from);
3164 new_id = mem_cgroup_id(to);
3166 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3167 mod_memcg_state(from, MEMCG_SWAP, -1);
3168 mod_memcg_state(to, MEMCG_SWAP, 1);
3174 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3175 struct mem_cgroup *from, struct mem_cgroup *to)
3181 static DEFINE_MUTEX(memcg_max_mutex);
3183 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3184 unsigned long max, bool memsw)
3186 bool enlarge = false;
3187 bool drained = false;
3189 bool limits_invariant;
3190 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3193 if (signal_pending(current)) {
3198 mutex_lock(&memcg_max_mutex);
3200 * Make sure that the new limit (memsw or memory limit) doesn't
3201 * break our basic invariant rule memory.max <= memsw.max.
3203 limits_invariant = memsw ? max >= memcg->memory.max :
3204 max <= memcg->memsw.max;
3205 if (!limits_invariant) {
3206 mutex_unlock(&memcg_max_mutex);
3210 if (max > counter->max)
3212 ret = page_counter_set_max(counter, max);
3213 mutex_unlock(&memcg_max_mutex);
3219 drain_all_stock(memcg);
3224 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3225 GFP_KERNEL, !memsw)) {
3231 if (!ret && enlarge)
3232 memcg_oom_recover(memcg);
3237 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3239 unsigned long *total_scanned)
3241 unsigned long nr_reclaimed = 0;
3242 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3243 unsigned long reclaimed;
3245 struct mem_cgroup_tree_per_node *mctz;
3246 unsigned long excess;
3247 unsigned long nr_scanned;
3252 mctz = soft_limit_tree_node(pgdat->node_id);
3255 * Do not even bother to check the largest node if the root
3256 * is empty. Do it lockless to prevent lock bouncing. Races
3257 * are acceptable as soft limit is best effort anyway.
3259 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3263 * This loop can run a while, specially if mem_cgroup's continuously
3264 * keep exceeding their soft limit and putting the system under
3271 mz = mem_cgroup_largest_soft_limit_node(mctz);
3276 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3277 gfp_mask, &nr_scanned);
3278 nr_reclaimed += reclaimed;
3279 *total_scanned += nr_scanned;
3280 spin_lock_irq(&mctz->lock);
3281 __mem_cgroup_remove_exceeded(mz, mctz);
3284 * If we failed to reclaim anything from this memory cgroup
3285 * it is time to move on to the next cgroup
3289 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3291 excess = soft_limit_excess(mz->memcg);
3293 * One school of thought says that we should not add
3294 * back the node to the tree if reclaim returns 0.
3295 * But our reclaim could return 0, simply because due
3296 * to priority we are exposing a smaller subset of
3297 * memory to reclaim from. Consider this as a longer
3300 /* If excess == 0, no tree ops */
3301 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3302 spin_unlock_irq(&mctz->lock);
3303 css_put(&mz->memcg->css);
3306 * Could not reclaim anything and there are no more
3307 * mem cgroups to try or we seem to be looping without
3308 * reclaiming anything.
3310 if (!nr_reclaimed &&
3312 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3314 } while (!nr_reclaimed);
3316 css_put(&next_mz->memcg->css);
3317 return nr_reclaimed;
3321 * Test whether @memcg has children, dead or alive. Note that this
3322 * function doesn't care whether @memcg has use_hierarchy enabled and
3323 * returns %true if there are child csses according to the cgroup
3324 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3326 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3331 ret = css_next_child(NULL, &memcg->css);
3337 * Reclaims as many pages from the given memcg as possible.
3339 * Caller is responsible for holding css reference for memcg.
3341 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3343 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3345 /* we call try-to-free pages for make this cgroup empty */
3346 lru_add_drain_all();
3348 drain_all_stock(memcg);
3350 /* try to free all pages in this cgroup */
3351 while (nr_retries && page_counter_read(&memcg->memory)) {
3354 if (signal_pending(current))
3357 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3361 /* maybe some writeback is necessary */
3362 congestion_wait(BLK_RW_ASYNC, HZ/10);
3370 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3371 char *buf, size_t nbytes,
3374 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3376 if (mem_cgroup_is_root(memcg))
3378 return mem_cgroup_force_empty(memcg) ?: nbytes;
3381 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3384 return mem_cgroup_from_css(css)->use_hierarchy;
3387 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3388 struct cftype *cft, u64 val)
3391 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3392 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3394 if (memcg->use_hierarchy == val)
3398 * If parent's use_hierarchy is set, we can't make any modifications
3399 * in the child subtrees. If it is unset, then the change can
3400 * occur, provided the current cgroup has no children.
3402 * For the root cgroup, parent_mem is NULL, we allow value to be
3403 * set if there are no children.
3405 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3406 (val == 1 || val == 0)) {
3407 if (!memcg_has_children(memcg))
3408 memcg->use_hierarchy = val;
3417 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3421 if (mem_cgroup_is_root(memcg)) {
3422 val = memcg_page_state(memcg, MEMCG_CACHE) +
3423 memcg_page_state(memcg, MEMCG_RSS);
3425 val += memcg_page_state(memcg, MEMCG_SWAP);
3428 val = page_counter_read(&memcg->memory);
3430 val = page_counter_read(&memcg->memsw);
3443 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3446 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3447 struct page_counter *counter;
3449 switch (MEMFILE_TYPE(cft->private)) {
3451 counter = &memcg->memory;
3454 counter = &memcg->memsw;
3457 counter = &memcg->kmem;
3460 counter = &memcg->tcpmem;
3466 switch (MEMFILE_ATTR(cft->private)) {
3468 if (counter == &memcg->memory)
3469 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3470 if (counter == &memcg->memsw)
3471 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3472 return (u64)page_counter_read(counter) * PAGE_SIZE;
3474 return (u64)counter->max * PAGE_SIZE;
3476 return (u64)counter->watermark * PAGE_SIZE;
3478 return counter->failcnt;
3479 case RES_SOFT_LIMIT:
3480 return (u64)memcg->soft_limit * PAGE_SIZE;
3486 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3488 unsigned long stat[MEMCG_NR_STAT] = {0};
3489 struct mem_cgroup *mi;
3492 for_each_online_cpu(cpu)
3493 for (i = 0; i < MEMCG_NR_STAT; i++)
3494 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3496 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3497 for (i = 0; i < MEMCG_NR_STAT; i++)
3498 atomic_long_add(stat[i], &mi->vmstats[i]);
3500 for_each_node(node) {
3501 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3502 struct mem_cgroup_per_node *pi;
3504 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3507 for_each_online_cpu(cpu)
3508 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3510 pn->lruvec_stat_cpu->count[i], cpu);
3512 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3513 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3514 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3518 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3520 unsigned long events[NR_VM_EVENT_ITEMS];
3521 struct mem_cgroup *mi;
3524 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3527 for_each_online_cpu(cpu)
3528 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3529 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3532 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3533 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3534 atomic_long_add(events[i], &mi->vmevents[i]);
3537 #ifdef CONFIG_MEMCG_KMEM
3538 static int memcg_online_kmem(struct mem_cgroup *memcg)
3542 if (cgroup_memory_nokmem)
3545 BUG_ON(memcg->kmemcg_id >= 0);
3546 BUG_ON(memcg->kmem_state);
3548 memcg_id = memcg_alloc_cache_id();
3552 static_branch_inc(&memcg_kmem_enabled_key);
3554 * A memory cgroup is considered kmem-online as soon as it gets
3555 * kmemcg_id. Setting the id after enabling static branching will
3556 * guarantee no one starts accounting before all call sites are
3559 memcg->kmemcg_id = memcg_id;
3560 memcg->kmem_state = KMEM_ONLINE;
3561 INIT_LIST_HEAD(&memcg->kmem_caches);
3566 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3568 struct cgroup_subsys_state *css;
3569 struct mem_cgroup *parent, *child;
3572 if (memcg->kmem_state != KMEM_ONLINE)
3575 * Clear the online state before clearing memcg_caches array
3576 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3577 * guarantees that no cache will be created for this cgroup
3578 * after we are done (see memcg_create_kmem_cache()).
3580 memcg->kmem_state = KMEM_ALLOCATED;
3582 parent = parent_mem_cgroup(memcg);
3584 parent = root_mem_cgroup;
3587 * Deactivate and reparent kmem_caches.
3589 memcg_deactivate_kmem_caches(memcg, parent);
3591 kmemcg_id = memcg->kmemcg_id;
3592 BUG_ON(kmemcg_id < 0);
3595 * Change kmemcg_id of this cgroup and all its descendants to the
3596 * parent's id, and then move all entries from this cgroup's list_lrus
3597 * to ones of the parent. After we have finished, all list_lrus
3598 * corresponding to this cgroup are guaranteed to remain empty. The
3599 * ordering is imposed by list_lru_node->lock taken by
3600 * memcg_drain_all_list_lrus().
3602 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3603 css_for_each_descendant_pre(css, &memcg->css) {
3604 child = mem_cgroup_from_css(css);
3605 BUG_ON(child->kmemcg_id != kmemcg_id);
3606 child->kmemcg_id = parent->kmemcg_id;
3607 if (!memcg->use_hierarchy)
3612 memcg_drain_all_list_lrus(kmemcg_id, parent);
3614 memcg_free_cache_id(kmemcg_id);
3617 static void memcg_free_kmem(struct mem_cgroup *memcg)
3619 /* css_alloc() failed, offlining didn't happen */
3620 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3621 memcg_offline_kmem(memcg);
3623 if (memcg->kmem_state == KMEM_ALLOCATED) {
3624 WARN_ON(!list_empty(&memcg->kmem_caches));
3625 static_branch_dec(&memcg_kmem_enabled_key);
3629 static int memcg_online_kmem(struct mem_cgroup *memcg)
3633 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3636 static void memcg_free_kmem(struct mem_cgroup *memcg)
3639 #endif /* CONFIG_MEMCG_KMEM */
3641 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3646 mutex_lock(&memcg_max_mutex);
3647 ret = page_counter_set_max(&memcg->kmem, max);
3648 mutex_unlock(&memcg_max_mutex);
3652 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3656 mutex_lock(&memcg_max_mutex);
3658 ret = page_counter_set_max(&memcg->tcpmem, max);
3662 if (!memcg->tcpmem_active) {
3664 * The active flag needs to be written after the static_key
3665 * update. This is what guarantees that the socket activation
3666 * function is the last one to run. See mem_cgroup_sk_alloc()
3667 * for details, and note that we don't mark any socket as
3668 * belonging to this memcg until that flag is up.
3670 * We need to do this, because static_keys will span multiple
3671 * sites, but we can't control their order. If we mark a socket
3672 * as accounted, but the accounting functions are not patched in
3673 * yet, we'll lose accounting.
3675 * We never race with the readers in mem_cgroup_sk_alloc(),
3676 * because when this value change, the code to process it is not
3679 static_branch_inc(&memcg_sockets_enabled_key);
3680 memcg->tcpmem_active = true;
3683 mutex_unlock(&memcg_max_mutex);
3688 * The user of this function is...
3691 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3692 char *buf, size_t nbytes, loff_t off)
3694 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3695 unsigned long nr_pages;
3698 buf = strstrip(buf);
3699 ret = page_counter_memparse(buf, "-1", &nr_pages);
3703 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3705 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3709 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3711 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3714 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3717 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3718 "Please report your usecase to linux-mm@kvack.org if you "
3719 "depend on this functionality.\n");
3720 ret = memcg_update_kmem_max(memcg, nr_pages);
3723 ret = memcg_update_tcp_max(memcg, nr_pages);
3727 case RES_SOFT_LIMIT:
3728 memcg->soft_limit = nr_pages;
3732 return ret ?: nbytes;
3735 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3736 size_t nbytes, loff_t off)
3738 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3739 struct page_counter *counter;
3741 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3743 counter = &memcg->memory;
3746 counter = &memcg->memsw;
3749 counter = &memcg->kmem;
3752 counter = &memcg->tcpmem;
3758 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3760 page_counter_reset_watermark(counter);
3763 counter->failcnt = 0;
3772 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3775 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3779 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3780 struct cftype *cft, u64 val)
3782 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3784 pr_warn_once("Cgroup memory moving (move_charge_at_immigrate) is deprecated. "
3785 "Please report your usecase to linux-mm@kvack.org if you "
3786 "depend on this functionality.\n");
3788 if (val & ~MOVE_MASK)
3792 * No kind of locking is needed in here, because ->can_attach() will
3793 * check this value once in the beginning of the process, and then carry
3794 * on with stale data. This means that changes to this value will only
3795 * affect task migrations starting after the change.
3797 memcg->move_charge_at_immigrate = val;
3801 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3802 struct cftype *cft, u64 val)
3810 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3811 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3812 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3814 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3815 int nid, unsigned int lru_mask)
3817 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
3818 unsigned long nr = 0;
3821 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3824 if (!(BIT(lru) & lru_mask))
3826 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3831 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3832 unsigned int lru_mask)
3834 unsigned long nr = 0;
3838 if (!(BIT(lru) & lru_mask))
3840 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3845 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3849 unsigned int lru_mask;
3852 static const struct numa_stat stats[] = {
3853 { "total", LRU_ALL },
3854 { "file", LRU_ALL_FILE },
3855 { "anon", LRU_ALL_ANON },
3856 { "unevictable", BIT(LRU_UNEVICTABLE) },
3858 const struct numa_stat *stat;
3861 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3863 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3864 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3865 seq_printf(m, "%s=%lu", stat->name, nr);
3866 for_each_node_state(nid, N_MEMORY) {
3867 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3869 seq_printf(m, " N%d=%lu", nid, nr);
3874 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3875 struct mem_cgroup *iter;
3878 for_each_mem_cgroup_tree(iter, memcg)
3879 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3880 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3881 for_each_node_state(nid, N_MEMORY) {
3883 for_each_mem_cgroup_tree(iter, memcg)
3884 nr += mem_cgroup_node_nr_lru_pages(
3885 iter, nid, stat->lru_mask);
3886 seq_printf(m, " N%d=%lu", nid, nr);
3893 #endif /* CONFIG_NUMA */
3895 static const unsigned int memcg1_stats[] = {
3906 static const char *const memcg1_stat_names[] = {
3917 /* Universal VM events cgroup1 shows, original sort order */
3918 static const unsigned int memcg1_events[] = {
3925 static const char *const memcg1_event_names[] = {
3932 static int memcg_stat_show(struct seq_file *m, void *v)
3934 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3935 unsigned long memory, memsw;
3936 struct mem_cgroup *mi;
3939 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3940 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3942 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3943 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3945 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3946 memcg_page_state_local(memcg, memcg1_stats[i]) *
3950 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3951 seq_printf(m, "%s %lu\n", memcg1_event_names[i],
3952 memcg_events_local(memcg, memcg1_events[i]));
3954 for (i = 0; i < NR_LRU_LISTS; i++)
3955 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3956 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3959 /* Hierarchical information */
3960 memory = memsw = PAGE_COUNTER_MAX;
3961 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3962 memory = min(memory, mi->memory.max);
3963 memsw = min(memsw, mi->memsw.max);
3965 seq_printf(m, "hierarchical_memory_limit %llu\n",
3966 (u64)memory * PAGE_SIZE);
3967 if (do_memsw_account())
3968 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3969 (u64)memsw * PAGE_SIZE);
3971 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3972 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3974 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3975 (u64)memcg_page_state(memcg, memcg1_stats[i]) *
3979 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3980 seq_printf(m, "total_%s %llu\n", memcg1_event_names[i],
3981 (u64)memcg_events(memcg, memcg1_events[i]));
3983 for (i = 0; i < NR_LRU_LISTS; i++)
3984 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i],
3985 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
3988 #ifdef CONFIG_DEBUG_VM
3991 struct mem_cgroup_per_node *mz;
3992 struct zone_reclaim_stat *rstat;
3993 unsigned long recent_rotated[2] = {0, 0};
3994 unsigned long recent_scanned[2] = {0, 0};
3996 for_each_online_pgdat(pgdat) {
3997 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3998 rstat = &mz->lruvec.reclaim_stat;
4000 recent_rotated[0] += rstat->recent_rotated[0];
4001 recent_rotated[1] += rstat->recent_rotated[1];
4002 recent_scanned[0] += rstat->recent_scanned[0];
4003 recent_scanned[1] += rstat->recent_scanned[1];
4005 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
4006 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
4007 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
4008 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
4015 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4018 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4020 return mem_cgroup_swappiness(memcg);
4023 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4024 struct cftype *cft, u64 val)
4026 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4032 memcg->swappiness = val;
4034 vm_swappiness = val;
4039 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4041 struct mem_cgroup_threshold_ary *t;
4042 unsigned long usage;
4047 t = rcu_dereference(memcg->thresholds.primary);
4049 t = rcu_dereference(memcg->memsw_thresholds.primary);
4054 usage = mem_cgroup_usage(memcg, swap);
4057 * current_threshold points to threshold just below or equal to usage.
4058 * If it's not true, a threshold was crossed after last
4059 * call of __mem_cgroup_threshold().
4061 i = t->current_threshold;
4064 * Iterate backward over array of thresholds starting from
4065 * current_threshold and check if a threshold is crossed.
4066 * If none of thresholds below usage is crossed, we read
4067 * only one element of the array here.
4069 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4070 eventfd_signal(t->entries[i].eventfd, 1);
4072 /* i = current_threshold + 1 */
4076 * Iterate forward over array of thresholds starting from
4077 * current_threshold+1 and check if a threshold is crossed.
4078 * If none of thresholds above usage is crossed, we read
4079 * only one element of the array here.
4081 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4082 eventfd_signal(t->entries[i].eventfd, 1);
4084 /* Update current_threshold */
4085 t->current_threshold = i - 1;
4090 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4093 __mem_cgroup_threshold(memcg, false);
4094 if (do_memsw_account())
4095 __mem_cgroup_threshold(memcg, true);
4097 memcg = parent_mem_cgroup(memcg);
4101 static int compare_thresholds(const void *a, const void *b)
4103 const struct mem_cgroup_threshold *_a = a;
4104 const struct mem_cgroup_threshold *_b = b;
4106 if (_a->threshold > _b->threshold)
4109 if (_a->threshold < _b->threshold)
4115 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4117 struct mem_cgroup_eventfd_list *ev;
4119 spin_lock(&memcg_oom_lock);
4121 list_for_each_entry(ev, &memcg->oom_notify, list)
4122 eventfd_signal(ev->eventfd, 1);
4124 spin_unlock(&memcg_oom_lock);
4128 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4130 struct mem_cgroup *iter;
4132 for_each_mem_cgroup_tree(iter, memcg)
4133 mem_cgroup_oom_notify_cb(iter);
4136 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4137 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4139 struct mem_cgroup_thresholds *thresholds;
4140 struct mem_cgroup_threshold_ary *new;
4141 unsigned long threshold;
4142 unsigned long usage;
4145 ret = page_counter_memparse(args, "-1", &threshold);
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 /* Check if a threshold crossed before adding a new one */
4161 if (thresholds->primary)
4162 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4164 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4166 /* Allocate memory for new array of thresholds */
4167 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4174 /* Copy thresholds (if any) to new array */
4175 if (thresholds->primary) {
4176 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4177 sizeof(struct mem_cgroup_threshold));
4180 /* Add new threshold */
4181 new->entries[size - 1].eventfd = eventfd;
4182 new->entries[size - 1].threshold = threshold;
4184 /* Sort thresholds. Registering of new threshold isn't time-critical */
4185 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4186 compare_thresholds, NULL);
4188 /* Find current threshold */
4189 new->current_threshold = -1;
4190 for (i = 0; i < size; i++) {
4191 if (new->entries[i].threshold <= usage) {
4193 * new->current_threshold will not be used until
4194 * rcu_assign_pointer(), so it's safe to increment
4197 ++new->current_threshold;
4202 /* Free old spare buffer and save old primary buffer as spare */
4203 kfree(thresholds->spare);
4204 thresholds->spare = thresholds->primary;
4206 rcu_assign_pointer(thresholds->primary, new);
4208 /* To be sure that nobody uses thresholds */
4212 mutex_unlock(&memcg->thresholds_lock);
4217 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4218 struct eventfd_ctx *eventfd, const char *args)
4220 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4223 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4224 struct eventfd_ctx *eventfd, const char *args)
4226 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4229 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4230 struct eventfd_ctx *eventfd, enum res_type type)
4232 struct mem_cgroup_thresholds *thresholds;
4233 struct mem_cgroup_threshold_ary *new;
4234 unsigned long usage;
4235 int i, j, size, entries;
4237 mutex_lock(&memcg->thresholds_lock);
4240 thresholds = &memcg->thresholds;
4241 usage = mem_cgroup_usage(memcg, false);
4242 } else if (type == _MEMSWAP) {
4243 thresholds = &memcg->memsw_thresholds;
4244 usage = mem_cgroup_usage(memcg, true);
4248 if (!thresholds->primary)
4251 /* Check if a threshold crossed before removing */
4252 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4254 /* Calculate new number of threshold */
4256 for (i = 0; i < thresholds->primary->size; i++) {
4257 if (thresholds->primary->entries[i].eventfd != eventfd)
4263 new = thresholds->spare;
4265 /* If no items related to eventfd have been cleared, nothing to do */
4269 /* Set thresholds array to NULL if we don't have thresholds */
4278 /* Copy thresholds and find current threshold */
4279 new->current_threshold = -1;
4280 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4281 if (thresholds->primary->entries[i].eventfd == eventfd)
4284 new->entries[j] = thresholds->primary->entries[i];
4285 if (new->entries[j].threshold <= usage) {
4287 * new->current_threshold will not be used
4288 * until rcu_assign_pointer(), so it's safe to increment
4291 ++new->current_threshold;
4297 /* Swap primary and spare array */
4298 thresholds->spare = thresholds->primary;
4300 rcu_assign_pointer(thresholds->primary, new);
4302 /* To be sure that nobody uses thresholds */
4305 /* If all events are unregistered, free the spare array */
4307 kfree(thresholds->spare);
4308 thresholds->spare = NULL;
4311 mutex_unlock(&memcg->thresholds_lock);
4314 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4315 struct eventfd_ctx *eventfd)
4317 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4320 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4321 struct eventfd_ctx *eventfd)
4323 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4326 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4327 struct eventfd_ctx *eventfd, const char *args)
4329 struct mem_cgroup_eventfd_list *event;
4331 event = kmalloc(sizeof(*event), GFP_KERNEL);
4335 spin_lock(&memcg_oom_lock);
4337 event->eventfd = eventfd;
4338 list_add(&event->list, &memcg->oom_notify);
4340 /* already in OOM ? */
4341 if (memcg->under_oom)
4342 eventfd_signal(eventfd, 1);
4343 spin_unlock(&memcg_oom_lock);
4348 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4349 struct eventfd_ctx *eventfd)
4351 struct mem_cgroup_eventfd_list *ev, *tmp;
4353 spin_lock(&memcg_oom_lock);
4355 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4356 if (ev->eventfd == eventfd) {
4357 list_del(&ev->list);
4362 spin_unlock(&memcg_oom_lock);
4365 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4367 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4369 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4370 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4371 seq_printf(sf, "oom_kill %lu\n",
4372 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4376 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4377 struct cftype *cft, u64 val)
4379 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4381 /* cannot set to root cgroup and only 0 and 1 are allowed */
4382 if (!css->parent || !((val == 0) || (val == 1)))
4385 memcg->oom_kill_disable = val;
4387 memcg_oom_recover(memcg);
4392 #ifdef CONFIG_CGROUP_WRITEBACK
4394 #include <trace/events/writeback.h>
4396 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4398 return wb_domain_init(&memcg->cgwb_domain, gfp);
4401 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4403 wb_domain_exit(&memcg->cgwb_domain);
4406 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4408 wb_domain_size_changed(&memcg->cgwb_domain);
4411 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4413 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4415 if (!memcg->css.parent)
4418 return &memcg->cgwb_domain;
4422 * idx can be of type enum memcg_stat_item or node_stat_item.
4423 * Keep in sync with memcg_exact_page().
4425 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4427 long x = atomic_long_read(&memcg->vmstats[idx]);
4430 for_each_online_cpu(cpu)
4431 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4438 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4439 * @wb: bdi_writeback in question
4440 * @pfilepages: out parameter for number of file pages
4441 * @pheadroom: out parameter for number of allocatable pages according to memcg
4442 * @pdirty: out parameter for number of dirty pages
4443 * @pwriteback: out parameter for number of pages under writeback
4445 * Determine the numbers of file, headroom, dirty, and writeback pages in
4446 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4447 * is a bit more involved.
4449 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4450 * headroom is calculated as the lowest headroom of itself and the
4451 * ancestors. Note that this doesn't consider the actual amount of
4452 * available memory in the system. The caller should further cap
4453 * *@pheadroom accordingly.
4455 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4456 unsigned long *pheadroom, unsigned long *pdirty,
4457 unsigned long *pwriteback)
4459 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4460 struct mem_cgroup *parent;
4462 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4464 /* this should eventually include NR_UNSTABLE_NFS */
4465 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4466 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4467 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4468 *pheadroom = PAGE_COUNTER_MAX;
4470 while ((parent = parent_mem_cgroup(memcg))) {
4471 unsigned long ceiling = min(memcg->memory.max, memcg->high);
4472 unsigned long used = page_counter_read(&memcg->memory);
4474 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4480 * Foreign dirty flushing
4482 * There's an inherent mismatch between memcg and writeback. The former
4483 * trackes ownership per-page while the latter per-inode. This was a
4484 * deliberate design decision because honoring per-page ownership in the
4485 * writeback path is complicated, may lead to higher CPU and IO overheads
4486 * and deemed unnecessary given that write-sharing an inode across
4487 * different cgroups isn't a common use-case.
4489 * Combined with inode majority-writer ownership switching, this works well
4490 * enough in most cases but there are some pathological cases. For
4491 * example, let's say there are two cgroups A and B which keep writing to
4492 * different but confined parts of the same inode. B owns the inode and
4493 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4494 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4495 * triggering background writeback. A will be slowed down without a way to
4496 * make writeback of the dirty pages happen.
4498 * Conditions like the above can lead to a cgroup getting repatedly and
4499 * severely throttled after making some progress after each
4500 * dirty_expire_interval while the underyling IO device is almost
4503 * Solving this problem completely requires matching the ownership tracking
4504 * granularities between memcg and writeback in either direction. However,
4505 * the more egregious behaviors can be avoided by simply remembering the
4506 * most recent foreign dirtying events and initiating remote flushes on
4507 * them when local writeback isn't enough to keep the memory clean enough.
4509 * The following two functions implement such mechanism. When a foreign
4510 * page - a page whose memcg and writeback ownerships don't match - is
4511 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4512 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4513 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4514 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4515 * foreign bdi_writebacks which haven't expired. Both the numbers of
4516 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4517 * limited to MEMCG_CGWB_FRN_CNT.
4519 * The mechanism only remembers IDs and doesn't hold any object references.
4520 * As being wrong occasionally doesn't matter, updates and accesses to the
4521 * records are lockless and racy.
4523 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4524 struct bdi_writeback *wb)
4526 struct mem_cgroup *memcg = page->mem_cgroup;
4527 struct memcg_cgwb_frn *frn;
4528 u64 now = get_jiffies_64();
4529 u64 oldest_at = now;
4533 trace_track_foreign_dirty(page, wb);
4536 * Pick the slot to use. If there is already a slot for @wb, keep
4537 * using it. If not replace the oldest one which isn't being
4540 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4541 frn = &memcg->cgwb_frn[i];
4542 if (frn->bdi_id == wb->bdi->id &&
4543 frn->memcg_id == wb->memcg_css->id)
4545 if (time_before64(frn->at, oldest_at) &&
4546 atomic_read(&frn->done.cnt) == 1) {
4548 oldest_at = frn->at;
4552 if (i < MEMCG_CGWB_FRN_CNT) {
4554 * Re-using an existing one. Update timestamp lazily to
4555 * avoid making the cacheline hot. We want them to be
4556 * reasonably up-to-date and significantly shorter than
4557 * dirty_expire_interval as that's what expires the record.
4558 * Use the shorter of 1s and dirty_expire_interval / 8.
4560 unsigned long update_intv =
4561 min_t(unsigned long, HZ,
4562 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4564 if (time_before64(frn->at, now - update_intv))
4566 } else if (oldest >= 0) {
4567 /* replace the oldest free one */
4568 frn = &memcg->cgwb_frn[oldest];
4569 frn->bdi_id = wb->bdi->id;
4570 frn->memcg_id = wb->memcg_css->id;
4575 /* issue foreign writeback flushes for recorded foreign dirtying events */
4576 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4578 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4579 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4580 u64 now = jiffies_64;
4583 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4584 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4587 * If the record is older than dirty_expire_interval,
4588 * writeback on it has already started. No need to kick it
4589 * off again. Also, don't start a new one if there's
4590 * already one in flight.
4592 if (time_after64(frn->at, now - intv) &&
4593 atomic_read(&frn->done.cnt) == 1) {
4595 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4596 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4597 WB_REASON_FOREIGN_FLUSH,
4603 #else /* CONFIG_CGROUP_WRITEBACK */
4605 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4610 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4614 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4618 #endif /* CONFIG_CGROUP_WRITEBACK */
4621 * DO NOT USE IN NEW FILES.
4623 * "cgroup.event_control" implementation.
4625 * This is way over-engineered. It tries to support fully configurable
4626 * events for each user. Such level of flexibility is completely
4627 * unnecessary especially in the light of the planned unified hierarchy.
4629 * Please deprecate this and replace with something simpler if at all
4634 * Unregister event and free resources.
4636 * Gets called from workqueue.
4638 static void memcg_event_remove(struct work_struct *work)
4640 struct mem_cgroup_event *event =
4641 container_of(work, struct mem_cgroup_event, remove);
4642 struct mem_cgroup *memcg = event->memcg;
4644 remove_wait_queue(event->wqh, &event->wait);
4646 event->unregister_event(memcg, event->eventfd);
4648 /* Notify userspace the event is going away. */
4649 eventfd_signal(event->eventfd, 1);
4651 eventfd_ctx_put(event->eventfd);
4653 css_put(&memcg->css);
4657 * Gets called on EPOLLHUP on eventfd when user closes it.
4659 * Called with wqh->lock held and interrupts disabled.
4661 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4662 int sync, void *key)
4664 struct mem_cgroup_event *event =
4665 container_of(wait, struct mem_cgroup_event, wait);
4666 struct mem_cgroup *memcg = event->memcg;
4667 __poll_t flags = key_to_poll(key);
4669 if (flags & EPOLLHUP) {
4671 * If the event has been detached at cgroup removal, we
4672 * can simply return knowing the other side will cleanup
4675 * We can't race against event freeing since the other
4676 * side will require wqh->lock via remove_wait_queue(),
4679 spin_lock(&memcg->event_list_lock);
4680 if (!list_empty(&event->list)) {
4681 list_del_init(&event->list);
4683 * We are in atomic context, but cgroup_event_remove()
4684 * may sleep, so we have to call it in workqueue.
4686 schedule_work(&event->remove);
4688 spin_unlock(&memcg->event_list_lock);
4694 static void memcg_event_ptable_queue_proc(struct file *file,
4695 wait_queue_head_t *wqh, poll_table *pt)
4697 struct mem_cgroup_event *event =
4698 container_of(pt, struct mem_cgroup_event, pt);
4701 add_wait_queue(wqh, &event->wait);
4705 * DO NOT USE IN NEW FILES.
4707 * Parse input and register new cgroup event handler.
4709 * Input must be in format '<event_fd> <control_fd> <args>'.
4710 * Interpretation of args is defined by control file implementation.
4712 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4713 char *buf, size_t nbytes, loff_t off)
4715 struct cgroup_subsys_state *css = of_css(of);
4716 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4717 struct mem_cgroup_event *event;
4718 struct cgroup_subsys_state *cfile_css;
4719 unsigned int efd, cfd;
4722 struct dentry *cdentry;
4727 buf = strstrip(buf);
4729 efd = simple_strtoul(buf, &endp, 10);
4734 cfd = simple_strtoul(buf, &endp, 10);
4735 if ((*endp != ' ') && (*endp != '\0'))
4739 event = kzalloc(sizeof(*event), GFP_KERNEL);
4743 event->memcg = memcg;
4744 INIT_LIST_HEAD(&event->list);
4745 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4746 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4747 INIT_WORK(&event->remove, memcg_event_remove);
4755 event->eventfd = eventfd_ctx_fileget(efile.file);
4756 if (IS_ERR(event->eventfd)) {
4757 ret = PTR_ERR(event->eventfd);
4764 goto out_put_eventfd;
4767 /* the process need read permission on control file */
4768 /* AV: shouldn't we check that it's been opened for read instead? */
4769 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4774 * The control file must be a regular cgroup1 file. As a regular cgroup
4775 * file can't be renamed, it's safe to access its name afterwards.
4777 cdentry = cfile.file->f_path.dentry;
4778 if (cdentry->d_sb->s_type != &cgroup_fs_type || !d_is_reg(cdentry)) {
4784 * Determine the event callbacks and set them in @event. This used
4785 * to be done via struct cftype but cgroup core no longer knows
4786 * about these events. The following is crude but the whole thing
4787 * is for compatibility anyway.
4789 * DO NOT ADD NEW FILES.
4791 name = cdentry->d_name.name;
4793 if (!strcmp(name, "memory.usage_in_bytes")) {
4794 event->register_event = mem_cgroup_usage_register_event;
4795 event->unregister_event = mem_cgroup_usage_unregister_event;
4796 } else if (!strcmp(name, "memory.oom_control")) {
4797 event->register_event = mem_cgroup_oom_register_event;
4798 event->unregister_event = mem_cgroup_oom_unregister_event;
4799 } else if (!strcmp(name, "memory.pressure_level")) {
4800 event->register_event = vmpressure_register_event;
4801 event->unregister_event = vmpressure_unregister_event;
4802 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4803 event->register_event = memsw_cgroup_usage_register_event;
4804 event->unregister_event = memsw_cgroup_usage_unregister_event;
4811 * Verify @cfile should belong to @css. Also, remaining events are
4812 * automatically removed on cgroup destruction but the removal is
4813 * asynchronous, so take an extra ref on @css.
4815 cfile_css = css_tryget_online_from_dir(cdentry->d_parent,
4816 &memory_cgrp_subsys);
4818 if (IS_ERR(cfile_css))
4820 if (cfile_css != css) {
4825 ret = event->register_event(memcg, event->eventfd, buf);
4829 vfs_poll(efile.file, &event->pt);
4831 spin_lock(&memcg->event_list_lock);
4832 list_add(&event->list, &memcg->event_list);
4833 spin_unlock(&memcg->event_list_lock);
4845 eventfd_ctx_put(event->eventfd);
4854 static struct cftype mem_cgroup_legacy_files[] = {
4856 .name = "usage_in_bytes",
4857 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4858 .read_u64 = mem_cgroup_read_u64,
4861 .name = "max_usage_in_bytes",
4862 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4863 .write = mem_cgroup_reset,
4864 .read_u64 = mem_cgroup_read_u64,
4867 .name = "limit_in_bytes",
4868 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4869 .write = mem_cgroup_write,
4870 .read_u64 = mem_cgroup_read_u64,
4873 .name = "soft_limit_in_bytes",
4874 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4875 .write = mem_cgroup_write,
4876 .read_u64 = mem_cgroup_read_u64,
4880 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4881 .write = mem_cgroup_reset,
4882 .read_u64 = mem_cgroup_read_u64,
4886 .seq_show = memcg_stat_show,
4889 .name = "force_empty",
4890 .write = mem_cgroup_force_empty_write,
4893 .name = "use_hierarchy",
4894 .write_u64 = mem_cgroup_hierarchy_write,
4895 .read_u64 = mem_cgroup_hierarchy_read,
4898 .name = "cgroup.event_control", /* XXX: for compat */
4899 .write = memcg_write_event_control,
4900 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4903 .name = "swappiness",
4904 .read_u64 = mem_cgroup_swappiness_read,
4905 .write_u64 = mem_cgroup_swappiness_write,
4908 .name = "move_charge_at_immigrate",
4909 .read_u64 = mem_cgroup_move_charge_read,
4910 .write_u64 = mem_cgroup_move_charge_write,
4913 .name = "oom_control",
4914 .seq_show = mem_cgroup_oom_control_read,
4915 .write_u64 = mem_cgroup_oom_control_write,
4916 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4919 .name = "pressure_level",
4923 .name = "numa_stat",
4924 .seq_show = memcg_numa_stat_show,
4928 .name = "kmem.limit_in_bytes",
4929 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4930 .write = mem_cgroup_write,
4931 .read_u64 = mem_cgroup_read_u64,
4934 .name = "kmem.usage_in_bytes",
4935 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4936 .read_u64 = mem_cgroup_read_u64,
4939 .name = "kmem.failcnt",
4940 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4941 .write = mem_cgroup_reset,
4942 .read_u64 = mem_cgroup_read_u64,
4945 .name = "kmem.max_usage_in_bytes",
4946 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4947 .write = mem_cgroup_reset,
4948 .read_u64 = mem_cgroup_read_u64,
4950 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4952 .name = "kmem.slabinfo",
4953 .seq_start = memcg_slab_start,
4954 .seq_next = memcg_slab_next,
4955 .seq_stop = memcg_slab_stop,
4956 .seq_show = memcg_slab_show,
4960 .name = "kmem.tcp.limit_in_bytes",
4961 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4962 .write = mem_cgroup_write,
4963 .read_u64 = mem_cgroup_read_u64,
4966 .name = "kmem.tcp.usage_in_bytes",
4967 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4968 .read_u64 = mem_cgroup_read_u64,
4971 .name = "kmem.tcp.failcnt",
4972 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4973 .write = mem_cgroup_reset,
4974 .read_u64 = mem_cgroup_read_u64,
4977 .name = "kmem.tcp.max_usage_in_bytes",
4978 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4979 .write = mem_cgroup_reset,
4980 .read_u64 = mem_cgroup_read_u64,
4982 { }, /* terminate */
4986 * Private memory cgroup IDR
4988 * Swap-out records and page cache shadow entries need to store memcg
4989 * references in constrained space, so we maintain an ID space that is
4990 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4991 * memory-controlled cgroups to 64k.
4993 * However, there usually are many references to the oflline CSS after
4994 * the cgroup has been destroyed, such as page cache or reclaimable
4995 * slab objects, that don't need to hang on to the ID. We want to keep
4996 * those dead CSS from occupying IDs, or we might quickly exhaust the
4997 * relatively small ID space and prevent the creation of new cgroups
4998 * even when there are much fewer than 64k cgroups - possibly none.
5000 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5001 * be freed and recycled when it's no longer needed, which is usually
5002 * when the CSS is offlined.
5004 * The only exception to that are records of swapped out tmpfs/shmem
5005 * pages that need to be attributed to live ancestors on swapin. But
5006 * those references are manageable from userspace.
5009 static DEFINE_IDR(mem_cgroup_idr);
5011 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5013 if (memcg->id.id > 0) {
5014 idr_remove(&mem_cgroup_idr, memcg->id.id);
5019 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
5021 refcount_add(n, &memcg->id.ref);
5024 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5026 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5027 mem_cgroup_id_remove(memcg);
5029 /* Memcg ID pins CSS */
5030 css_put(&memcg->css);
5034 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5036 mem_cgroup_id_put_many(memcg, 1);
5040 * mem_cgroup_from_id - look up a memcg from a memcg id
5041 * @id: the memcg id to look up
5043 * Caller must hold rcu_read_lock().
5045 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5047 WARN_ON_ONCE(!rcu_read_lock_held());
5048 return idr_find(&mem_cgroup_idr, id);
5051 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5053 struct mem_cgroup_per_node *pn;
5056 * This routine is called against possible nodes.
5057 * But it's BUG to call kmalloc() against offline node.
5059 * TODO: this routine can waste much memory for nodes which will
5060 * never be onlined. It's better to use memory hotplug callback
5063 if (!node_state(node, N_NORMAL_MEMORY))
5065 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5069 pn->lruvec_stat_local = alloc_percpu(struct lruvec_stat);
5070 if (!pn->lruvec_stat_local) {
5075 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
5076 if (!pn->lruvec_stat_cpu) {
5077 free_percpu(pn->lruvec_stat_local);
5082 lruvec_init(&pn->lruvec);
5083 pn->usage_in_excess = 0;
5084 pn->on_tree = false;
5087 memcg->nodeinfo[node] = pn;
5091 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5093 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5098 free_percpu(pn->lruvec_stat_cpu);
5099 free_percpu(pn->lruvec_stat_local);
5103 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5108 free_mem_cgroup_per_node_info(memcg, node);
5109 free_percpu(memcg->vmstats_percpu);
5110 free_percpu(memcg->vmstats_local);
5114 static void mem_cgroup_free(struct mem_cgroup *memcg)
5116 memcg_wb_domain_exit(memcg);
5118 * Flush percpu vmstats and vmevents to guarantee the value correctness
5119 * on parent's and all ancestor levels.
5121 memcg_flush_percpu_vmstats(memcg);
5122 memcg_flush_percpu_vmevents(memcg);
5123 __mem_cgroup_free(memcg);
5126 static struct mem_cgroup *mem_cgroup_alloc(void)
5128 struct mem_cgroup *memcg;
5131 int __maybe_unused i;
5132 long error = -ENOMEM;
5134 size = sizeof(struct mem_cgroup);
5135 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5137 memcg = kzalloc(size, GFP_KERNEL);
5139 return ERR_PTR(error);
5141 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5142 1, MEM_CGROUP_ID_MAX,
5144 if (memcg->id.id < 0) {
5145 error = memcg->id.id;
5149 memcg->vmstats_local = alloc_percpu(struct memcg_vmstats_percpu);
5150 if (!memcg->vmstats_local)
5153 memcg->vmstats_percpu = alloc_percpu(struct memcg_vmstats_percpu);
5154 if (!memcg->vmstats_percpu)
5158 if (alloc_mem_cgroup_per_node_info(memcg, node))
5161 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5164 INIT_WORK(&memcg->high_work, high_work_func);
5165 memcg->last_scanned_node = MAX_NUMNODES;
5166 INIT_LIST_HEAD(&memcg->oom_notify);
5167 mutex_init(&memcg->thresholds_lock);
5168 spin_lock_init(&memcg->move_lock);
5169 vmpressure_init(&memcg->vmpressure);
5170 INIT_LIST_HEAD(&memcg->event_list);
5171 spin_lock_init(&memcg->event_list_lock);
5172 memcg->socket_pressure = jiffies;
5173 #ifdef CONFIG_MEMCG_KMEM
5174 memcg->kmemcg_id = -1;
5176 #ifdef CONFIG_CGROUP_WRITEBACK
5177 INIT_LIST_HEAD(&memcg->cgwb_list);
5178 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5179 memcg->cgwb_frn[i].done =
5180 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5182 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5183 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5184 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5185 memcg->deferred_split_queue.split_queue_len = 0;
5187 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5190 mem_cgroup_id_remove(memcg);
5191 __mem_cgroup_free(memcg);
5192 return ERR_PTR(error);
5195 static struct cgroup_subsys_state * __ref
5196 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5198 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5199 struct mem_cgroup *memcg;
5200 long error = -ENOMEM;
5202 memcg = mem_cgroup_alloc();
5204 return ERR_CAST(memcg);
5206 memcg->high = PAGE_COUNTER_MAX;
5207 memcg->soft_limit = PAGE_COUNTER_MAX;
5209 memcg->swappiness = mem_cgroup_swappiness(parent);
5210 memcg->oom_kill_disable = parent->oom_kill_disable;
5212 if (parent && parent->use_hierarchy) {
5213 memcg->use_hierarchy = true;
5214 page_counter_init(&memcg->memory, &parent->memory);
5215 page_counter_init(&memcg->swap, &parent->swap);
5216 page_counter_init(&memcg->memsw, &parent->memsw);
5217 page_counter_init(&memcg->kmem, &parent->kmem);
5218 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5220 page_counter_init(&memcg->memory, NULL);
5221 page_counter_init(&memcg->swap, NULL);
5222 page_counter_init(&memcg->memsw, NULL);
5223 page_counter_init(&memcg->kmem, NULL);
5224 page_counter_init(&memcg->tcpmem, NULL);
5226 * Deeper hierachy with use_hierarchy == false doesn't make
5227 * much sense so let cgroup subsystem know about this
5228 * unfortunate state in our controller.
5230 if (parent != root_mem_cgroup)
5231 memory_cgrp_subsys.broken_hierarchy = true;
5234 /* The following stuff does not apply to the root */
5236 #ifdef CONFIG_MEMCG_KMEM
5237 INIT_LIST_HEAD(&memcg->kmem_caches);
5239 root_mem_cgroup = memcg;
5243 error = memcg_online_kmem(memcg);
5247 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5248 static_branch_inc(&memcg_sockets_enabled_key);
5252 mem_cgroup_id_remove(memcg);
5253 mem_cgroup_free(memcg);
5254 return ERR_PTR(error);
5257 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5259 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5262 * A memcg must be visible for memcg_expand_shrinker_maps()
5263 * by the time the maps are allocated. So, we allocate maps
5264 * here, when for_each_mem_cgroup() can't skip it.
5266 if (memcg_alloc_shrinker_maps(memcg)) {
5267 mem_cgroup_id_remove(memcg);
5271 /* Online state pins memcg ID, memcg ID pins CSS */
5272 refcount_set(&memcg->id.ref, 1);
5277 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5279 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5280 struct mem_cgroup_event *event, *tmp;
5283 * Unregister events and notify userspace.
5284 * Notify userspace about cgroup removing only after rmdir of cgroup
5285 * directory to avoid race between userspace and kernelspace.
5287 spin_lock(&memcg->event_list_lock);
5288 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5289 list_del_init(&event->list);
5290 schedule_work(&event->remove);
5292 spin_unlock(&memcg->event_list_lock);
5294 page_counter_set_min(&memcg->memory, 0);
5295 page_counter_set_low(&memcg->memory, 0);
5297 memcg_offline_kmem(memcg);
5298 wb_memcg_offline(memcg);
5300 drain_all_stock(memcg);
5302 mem_cgroup_id_put(memcg);
5305 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5307 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5309 invalidate_reclaim_iterators(memcg);
5312 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5314 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5315 int __maybe_unused i;
5317 #ifdef CONFIG_CGROUP_WRITEBACK
5318 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5319 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5321 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5322 static_branch_dec(&memcg_sockets_enabled_key);
5324 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5325 static_branch_dec(&memcg_sockets_enabled_key);
5327 vmpressure_cleanup(&memcg->vmpressure);
5328 cancel_work_sync(&memcg->high_work);
5329 mem_cgroup_remove_from_trees(memcg);
5330 memcg_free_shrinker_maps(memcg);
5331 memcg_free_kmem(memcg);
5332 mem_cgroup_free(memcg);
5336 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5337 * @css: the target css
5339 * Reset the states of the mem_cgroup associated with @css. This is
5340 * invoked when the userland requests disabling on the default hierarchy
5341 * but the memcg is pinned through dependency. The memcg should stop
5342 * applying policies and should revert to the vanilla state as it may be
5343 * made visible again.
5345 * The current implementation only resets the essential configurations.
5346 * This needs to be expanded to cover all the visible parts.
5348 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5350 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5352 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5353 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5354 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
5355 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5356 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5357 page_counter_set_min(&memcg->memory, 0);
5358 page_counter_set_low(&memcg->memory, 0);
5359 memcg->high = PAGE_COUNTER_MAX;
5360 memcg->soft_limit = PAGE_COUNTER_MAX;
5361 memcg_wb_domain_size_changed(memcg);
5365 /* Handlers for move charge at task migration. */
5366 static int mem_cgroup_do_precharge(unsigned long count)
5370 /* Try a single bulk charge without reclaim first, kswapd may wake */
5371 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5373 mc.precharge += count;
5377 /* Try charges one by one with reclaim, but do not retry */
5379 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5393 enum mc_target_type {
5400 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5401 unsigned long addr, pte_t ptent)
5403 struct page *page = vm_normal_page(vma, addr, ptent);
5405 if (!page || !page_mapped(page))
5407 if (PageAnon(page)) {
5408 if (!(mc.flags & MOVE_ANON))
5411 if (!(mc.flags & MOVE_FILE))
5414 if (!get_page_unless_zero(page))
5420 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5421 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5422 pte_t ptent, swp_entry_t *entry)
5424 struct page *page = NULL;
5425 swp_entry_t ent = pte_to_swp_entry(ptent);
5427 if (!(mc.flags & MOVE_ANON))
5431 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5432 * a device and because they are not accessible by CPU they are store
5433 * as special swap entry in the CPU page table.
5435 if (is_device_private_entry(ent)) {
5436 page = device_private_entry_to_page(ent);
5438 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5439 * a refcount of 1 when free (unlike normal page)
5441 if (!page_ref_add_unless(page, 1, 1))
5446 if (non_swap_entry(ent))
5450 * Because lookup_swap_cache() updates some statistics counter,
5451 * we call find_get_page() with swapper_space directly.
5453 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5454 if (do_memsw_account())
5455 entry->val = ent.val;
5460 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5461 pte_t ptent, swp_entry_t *entry)
5467 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5468 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5470 struct page *page = NULL;
5471 struct address_space *mapping;
5474 if (!vma->vm_file) /* anonymous vma */
5476 if (!(mc.flags & MOVE_FILE))
5479 mapping = vma->vm_file->f_mapping;
5480 pgoff = linear_page_index(vma, addr);
5482 /* page is moved even if it's not RSS of this task(page-faulted). */
5484 /* shmem/tmpfs may report page out on swap: account for that too. */
5485 if (shmem_mapping(mapping)) {
5486 page = find_get_entry(mapping, pgoff);
5487 if (xa_is_value(page)) {
5488 swp_entry_t swp = radix_to_swp_entry(page);
5489 if (do_memsw_account())
5491 page = find_get_page(swap_address_space(swp),
5495 page = find_get_page(mapping, pgoff);
5497 page = find_get_page(mapping, pgoff);
5503 * mem_cgroup_move_account - move account of the page
5505 * @compound: charge the page as compound or small page
5506 * @from: mem_cgroup which the page is moved from.
5507 * @to: mem_cgroup which the page is moved to. @from != @to.
5509 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5511 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5514 static int mem_cgroup_move_account(struct page *page,
5516 struct mem_cgroup *from,
5517 struct mem_cgroup *to)
5519 struct lruvec *from_vec, *to_vec;
5520 struct pglist_data *pgdat;
5521 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5525 VM_BUG_ON(from == to);
5526 VM_BUG_ON_PAGE(PageLRU(page), page);
5527 VM_BUG_ON(compound && !PageTransHuge(page));
5530 * Prevent mem_cgroup_migrate() from looking at
5531 * page->mem_cgroup of its source page while we change it.
5534 if (!trylock_page(page))
5538 if (page->mem_cgroup != from)
5541 anon = PageAnon(page);
5543 pgdat = page_pgdat(page);
5544 from_vec = mem_cgroup_lruvec(pgdat, from);
5545 to_vec = mem_cgroup_lruvec(pgdat, to);
5547 lock_page_memcg(page);
5549 if (!anon && page_mapped(page)) {
5550 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5551 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5554 if (!anon && PageDirty(page)) {
5555 struct address_space *mapping = page_mapping(page);
5557 if (mapping_cap_account_dirty(mapping)) {
5558 __mod_lruvec_state(from_vec, NR_FILE_DIRTY, -nr_pages);
5559 __mod_lruvec_state(to_vec, NR_FILE_DIRTY, nr_pages);
5563 if (PageWriteback(page)) {
5564 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5565 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5569 * All state has been migrated, let's switch to the new memcg.
5571 * It is safe to change page->mem_cgroup here because the page
5572 * is referenced, charged, isolated, and locked: we can't race
5573 * with (un)charging, migration, LRU putback, or anything else
5574 * that would rely on a stable page->mem_cgroup.
5576 * Note that lock_page_memcg is a memcg lock, not a page lock,
5577 * to save space. As soon as we switch page->mem_cgroup to a
5578 * new memcg that isn't locked, the above state can change
5579 * concurrently again. Make sure we're truly done with it.
5583 page->mem_cgroup = to; /* caller should have done css_get */
5585 __unlock_page_memcg(from);
5589 local_irq_disable();
5590 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
5591 memcg_check_events(to, page);
5592 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
5593 memcg_check_events(from, page);
5602 * get_mctgt_type - get target type of moving charge
5603 * @vma: the vma the pte to be checked belongs
5604 * @addr: the address corresponding to the pte to be checked
5605 * @ptent: the pte to be checked
5606 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5609 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5610 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5611 * move charge. if @target is not NULL, the page is stored in target->page
5612 * with extra refcnt got(Callers should handle it).
5613 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5614 * target for charge migration. if @target is not NULL, the entry is stored
5616 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5617 * (so ZONE_DEVICE page and thus not on the lru).
5618 * For now we such page is charge like a regular page would be as for all
5619 * intent and purposes it is just special memory taking the place of a
5622 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5624 * Called with pte lock held.
5627 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5628 unsigned long addr, pte_t ptent, union mc_target *target)
5630 struct page *page = NULL;
5631 enum mc_target_type ret = MC_TARGET_NONE;
5632 swp_entry_t ent = { .val = 0 };
5634 if (pte_present(ptent))
5635 page = mc_handle_present_pte(vma, addr, ptent);
5636 else if (is_swap_pte(ptent))
5637 page = mc_handle_swap_pte(vma, ptent, &ent);
5638 else if (pte_none(ptent))
5639 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5641 if (!page && !ent.val)
5645 * Do only loose check w/o serialization.
5646 * mem_cgroup_move_account() checks the page is valid or
5647 * not under LRU exclusion.
5649 if (page->mem_cgroup == mc.from) {
5650 ret = MC_TARGET_PAGE;
5651 if (is_device_private_page(page))
5652 ret = MC_TARGET_DEVICE;
5654 target->page = page;
5656 if (!ret || !target)
5660 * There is a swap entry and a page doesn't exist or isn't charged.
5661 * But we cannot move a tail-page in a THP.
5663 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5664 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5665 ret = MC_TARGET_SWAP;
5672 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5674 * We don't consider PMD mapped swapping or file mapped pages because THP does
5675 * not support them for now.
5676 * Caller should make sure that pmd_trans_huge(pmd) is true.
5678 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5679 unsigned long addr, pmd_t pmd, union mc_target *target)
5681 struct page *page = NULL;
5682 enum mc_target_type ret = MC_TARGET_NONE;
5684 if (unlikely(is_swap_pmd(pmd))) {
5685 VM_BUG_ON(thp_migration_supported() &&
5686 !is_pmd_migration_entry(pmd));
5689 page = pmd_page(pmd);
5690 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5691 if (!(mc.flags & MOVE_ANON))
5693 if (page->mem_cgroup == mc.from) {
5694 ret = MC_TARGET_PAGE;
5697 target->page = page;
5703 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5704 unsigned long addr, pmd_t pmd, union mc_target *target)
5706 return MC_TARGET_NONE;
5710 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5711 unsigned long addr, unsigned long end,
5712 struct mm_walk *walk)
5714 struct vm_area_struct *vma = walk->vma;
5718 ptl = pmd_trans_huge_lock(pmd, vma);
5721 * Note their can not be MC_TARGET_DEVICE for now as we do not
5722 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5723 * this might change.
5725 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5726 mc.precharge += HPAGE_PMD_NR;
5731 if (pmd_trans_unstable(pmd))
5733 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5734 for (; addr != end; pte++, addr += PAGE_SIZE)
5735 if (get_mctgt_type(vma, addr, *pte, NULL))
5736 mc.precharge++; /* increment precharge temporarily */
5737 pte_unmap_unlock(pte - 1, ptl);
5743 static const struct mm_walk_ops precharge_walk_ops = {
5744 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5747 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5749 unsigned long precharge;
5751 down_read(&mm->mmap_sem);
5752 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5753 up_read(&mm->mmap_sem);
5755 precharge = mc.precharge;
5761 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5763 unsigned long precharge = mem_cgroup_count_precharge(mm);
5765 VM_BUG_ON(mc.moving_task);
5766 mc.moving_task = current;
5767 return mem_cgroup_do_precharge(precharge);
5770 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5771 static void __mem_cgroup_clear_mc(void)
5773 struct mem_cgroup *from = mc.from;
5774 struct mem_cgroup *to = mc.to;
5776 /* we must uncharge all the leftover precharges from mc.to */
5778 cancel_charge(mc.to, mc.precharge);
5782 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5783 * we must uncharge here.
5785 if (mc.moved_charge) {
5786 cancel_charge(mc.from, mc.moved_charge);
5787 mc.moved_charge = 0;
5789 /* we must fixup refcnts and charges */
5790 if (mc.moved_swap) {
5791 /* uncharge swap account from the old cgroup */
5792 if (!mem_cgroup_is_root(mc.from))
5793 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5795 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5798 * we charged both to->memory and to->memsw, so we
5799 * should uncharge to->memory.
5801 if (!mem_cgroup_is_root(mc.to))
5802 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5804 css_put_many(&mc.to->css, mc.moved_swap);
5808 memcg_oom_recover(from);
5809 memcg_oom_recover(to);
5810 wake_up_all(&mc.waitq);
5813 static void mem_cgroup_clear_mc(void)
5815 struct mm_struct *mm = mc.mm;
5818 * we must clear moving_task before waking up waiters at the end of
5821 mc.moving_task = NULL;
5822 __mem_cgroup_clear_mc();
5823 spin_lock(&mc.lock);
5827 spin_unlock(&mc.lock);
5832 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5834 struct cgroup_subsys_state *css;
5835 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5836 struct mem_cgroup *from;
5837 struct task_struct *leader, *p;
5838 struct mm_struct *mm;
5839 unsigned long move_flags;
5842 /* charge immigration isn't supported on the default hierarchy */
5843 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5847 * Multi-process migrations only happen on the default hierarchy
5848 * where charge immigration is not used. Perform charge
5849 * immigration if @tset contains a leader and whine if there are
5853 cgroup_taskset_for_each_leader(leader, css, tset) {
5856 memcg = mem_cgroup_from_css(css);
5862 * We are now commited to this value whatever it is. Changes in this
5863 * tunable will only affect upcoming migrations, not the current one.
5864 * So we need to save it, and keep it going.
5866 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5870 from = mem_cgroup_from_task(p);
5872 VM_BUG_ON(from == memcg);
5874 mm = get_task_mm(p);
5877 /* We move charges only when we move a owner of the mm */
5878 if (mm->owner == p) {
5881 VM_BUG_ON(mc.precharge);
5882 VM_BUG_ON(mc.moved_charge);
5883 VM_BUG_ON(mc.moved_swap);
5885 spin_lock(&mc.lock);
5889 mc.flags = move_flags;
5890 spin_unlock(&mc.lock);
5891 /* We set mc.moving_task later */
5893 ret = mem_cgroup_precharge_mc(mm);
5895 mem_cgroup_clear_mc();
5902 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5905 mem_cgroup_clear_mc();
5908 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5909 unsigned long addr, unsigned long end,
5910 struct mm_walk *walk)
5913 struct vm_area_struct *vma = walk->vma;
5916 enum mc_target_type target_type;
5917 union mc_target target;
5920 ptl = pmd_trans_huge_lock(pmd, vma);
5922 if (mc.precharge < HPAGE_PMD_NR) {
5926 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5927 if (target_type == MC_TARGET_PAGE) {
5929 if (!isolate_lru_page(page)) {
5930 if (!mem_cgroup_move_account(page, true,
5932 mc.precharge -= HPAGE_PMD_NR;
5933 mc.moved_charge += HPAGE_PMD_NR;
5935 putback_lru_page(page);
5938 } else if (target_type == MC_TARGET_DEVICE) {
5940 if (!mem_cgroup_move_account(page, true,
5942 mc.precharge -= HPAGE_PMD_NR;
5943 mc.moved_charge += HPAGE_PMD_NR;
5951 if (pmd_trans_unstable(pmd))
5954 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5955 for (; addr != end; addr += PAGE_SIZE) {
5956 pte_t ptent = *(pte++);
5957 bool device = false;
5963 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5964 case MC_TARGET_DEVICE:
5967 case MC_TARGET_PAGE:
5970 * We can have a part of the split pmd here. Moving it
5971 * can be done but it would be too convoluted so simply
5972 * ignore such a partial THP and keep it in original
5973 * memcg. There should be somebody mapping the head.
5975 if (PageTransCompound(page))
5977 if (!device && isolate_lru_page(page))
5979 if (!mem_cgroup_move_account(page, false,
5982 /* we uncharge from mc.from later. */
5986 putback_lru_page(page);
5987 put: /* get_mctgt_type() gets the page */
5990 case MC_TARGET_SWAP:
5992 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5994 mem_cgroup_id_get_many(mc.to, 1);
5995 /* we fixup other refcnts and charges later. */
6003 pte_unmap_unlock(pte - 1, ptl);
6008 * We have consumed all precharges we got in can_attach().
6009 * We try charge one by one, but don't do any additional
6010 * charges to mc.to if we have failed in charge once in attach()
6013 ret = mem_cgroup_do_precharge(1);
6021 static const struct mm_walk_ops charge_walk_ops = {
6022 .pmd_entry = mem_cgroup_move_charge_pte_range,
6025 static void mem_cgroup_move_charge(void)
6027 lru_add_drain_all();
6029 * Signal lock_page_memcg() to take the memcg's move_lock
6030 * while we're moving its pages to another memcg. Then wait
6031 * for already started RCU-only updates to finish.
6033 atomic_inc(&mc.from->moving_account);
6036 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
6038 * Someone who are holding the mmap_sem might be waiting in
6039 * waitq. So we cancel all extra charges, wake up all waiters,
6040 * and retry. Because we cancel precharges, we might not be able
6041 * to move enough charges, but moving charge is a best-effort
6042 * feature anyway, so it wouldn't be a big problem.
6044 __mem_cgroup_clear_mc();
6049 * When we have consumed all precharges and failed in doing
6050 * additional charge, the page walk just aborts.
6052 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6055 up_read(&mc.mm->mmap_sem);
6056 atomic_dec(&mc.from->moving_account);
6059 static void mem_cgroup_move_task(void)
6062 mem_cgroup_move_charge();
6063 mem_cgroup_clear_mc();
6066 #else /* !CONFIG_MMU */
6067 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6071 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6074 static void mem_cgroup_move_task(void)
6080 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6081 * to verify whether we're attached to the default hierarchy on each mount
6084 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6087 * use_hierarchy is forced on the default hierarchy. cgroup core
6088 * guarantees that @root doesn't have any children, so turning it
6089 * on for the root memcg is enough.
6091 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6092 root_mem_cgroup->use_hierarchy = true;
6094 root_mem_cgroup->use_hierarchy = false;
6097 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6099 if (value == PAGE_COUNTER_MAX)
6100 seq_puts(m, "max\n");
6102 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6107 static u64 memory_current_read(struct cgroup_subsys_state *css,
6110 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6112 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6115 static int memory_min_show(struct seq_file *m, void *v)
6117 return seq_puts_memcg_tunable(m,
6118 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6121 static ssize_t memory_min_write(struct kernfs_open_file *of,
6122 char *buf, size_t nbytes, loff_t off)
6124 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6128 buf = strstrip(buf);
6129 err = page_counter_memparse(buf, "max", &min);
6133 page_counter_set_min(&memcg->memory, min);
6138 static int memory_low_show(struct seq_file *m, void *v)
6140 return seq_puts_memcg_tunable(m,
6141 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6144 static ssize_t memory_low_write(struct kernfs_open_file *of,
6145 char *buf, size_t nbytes, loff_t off)
6147 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6151 buf = strstrip(buf);
6152 err = page_counter_memparse(buf, "max", &low);
6156 page_counter_set_low(&memcg->memory, low);
6161 static int memory_high_show(struct seq_file *m, void *v)
6163 return seq_puts_memcg_tunable(m, READ_ONCE(mem_cgroup_from_seq(m)->high));
6166 static ssize_t memory_high_write(struct kernfs_open_file *of,
6167 char *buf, size_t nbytes, loff_t off)
6169 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6170 unsigned long nr_pages;
6174 buf = strstrip(buf);
6175 err = page_counter_memparse(buf, "max", &high);
6181 nr_pages = page_counter_read(&memcg->memory);
6182 if (nr_pages > high)
6183 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6186 memcg_wb_domain_size_changed(memcg);
6190 static int memory_max_show(struct seq_file *m, void *v)
6192 return seq_puts_memcg_tunable(m,
6193 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6196 static ssize_t memory_max_write(struct kernfs_open_file *of,
6197 char *buf, size_t nbytes, loff_t off)
6199 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6200 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
6201 bool drained = false;
6205 buf = strstrip(buf);
6206 err = page_counter_memparse(buf, "max", &max);
6210 xchg(&memcg->memory.max, max);
6213 unsigned long nr_pages = page_counter_read(&memcg->memory);
6215 if (nr_pages <= max)
6218 if (signal_pending(current)) {
6224 drain_all_stock(memcg);
6230 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6236 memcg_memory_event(memcg, MEMCG_OOM);
6237 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6241 memcg_wb_domain_size_changed(memcg);
6245 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6247 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6248 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6249 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6250 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6251 seq_printf(m, "oom_kill %lu\n",
6252 atomic_long_read(&events[MEMCG_OOM_KILL]));
6255 static int memory_events_show(struct seq_file *m, void *v)
6257 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6259 __memory_events_show(m, memcg->memory_events);
6263 static int memory_events_local_show(struct seq_file *m, void *v)
6265 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6267 __memory_events_show(m, memcg->memory_events_local);
6271 static int memory_stat_show(struct seq_file *m, void *v)
6273 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6276 buf = memory_stat_format(memcg);
6284 static int memory_oom_group_show(struct seq_file *m, void *v)
6286 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6288 seq_printf(m, "%d\n", memcg->oom_group);
6293 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6294 char *buf, size_t nbytes, loff_t off)
6296 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6299 buf = strstrip(buf);
6303 ret = kstrtoint(buf, 0, &oom_group);
6307 if (oom_group != 0 && oom_group != 1)
6310 memcg->oom_group = oom_group;
6315 static struct cftype memory_files[] = {
6318 .flags = CFTYPE_NOT_ON_ROOT,
6319 .read_u64 = memory_current_read,
6323 .flags = CFTYPE_NOT_ON_ROOT,
6324 .seq_show = memory_min_show,
6325 .write = memory_min_write,
6329 .flags = CFTYPE_NOT_ON_ROOT,
6330 .seq_show = memory_low_show,
6331 .write = memory_low_write,
6335 .flags = CFTYPE_NOT_ON_ROOT,
6336 .seq_show = memory_high_show,
6337 .write = memory_high_write,
6341 .flags = CFTYPE_NOT_ON_ROOT,
6342 .seq_show = memory_max_show,
6343 .write = memory_max_write,
6347 .flags = CFTYPE_NOT_ON_ROOT,
6348 .file_offset = offsetof(struct mem_cgroup, events_file),
6349 .seq_show = memory_events_show,
6352 .name = "events.local",
6353 .flags = CFTYPE_NOT_ON_ROOT,
6354 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6355 .seq_show = memory_events_local_show,
6359 .flags = CFTYPE_NOT_ON_ROOT,
6360 .seq_show = memory_stat_show,
6363 .name = "oom.group",
6364 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6365 .seq_show = memory_oom_group_show,
6366 .write = memory_oom_group_write,
6371 struct cgroup_subsys memory_cgrp_subsys = {
6372 .css_alloc = mem_cgroup_css_alloc,
6373 .css_online = mem_cgroup_css_online,
6374 .css_offline = mem_cgroup_css_offline,
6375 .css_released = mem_cgroup_css_released,
6376 .css_free = mem_cgroup_css_free,
6377 .css_reset = mem_cgroup_css_reset,
6378 .can_attach = mem_cgroup_can_attach,
6379 .cancel_attach = mem_cgroup_cancel_attach,
6380 .post_attach = mem_cgroup_move_task,
6381 .bind = mem_cgroup_bind,
6382 .dfl_cftypes = memory_files,
6383 .legacy_cftypes = mem_cgroup_legacy_files,
6388 * mem_cgroup_protected - check if memory consumption is in the normal range
6389 * @root: the top ancestor of the sub-tree being checked
6390 * @memcg: the memory cgroup to check
6392 * WARNING: This function is not stateless! It can only be used as part
6393 * of a top-down tree iteration, not for isolated queries.
6395 * Returns one of the following:
6396 * MEMCG_PROT_NONE: cgroup memory is not protected
6397 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
6398 * an unprotected supply of reclaimable memory from other cgroups.
6399 * MEMCG_PROT_MIN: cgroup memory is protected
6401 * @root is exclusive; it is never protected when looked at directly
6403 * To provide a proper hierarchical behavior, effective memory.min/low values
6404 * are used. Below is the description of how effective memory.low is calculated.
6405 * Effective memory.min values is calculated in the same way.
6407 * Effective memory.low is always equal or less than the original memory.low.
6408 * If there is no memory.low overcommittment (which is always true for
6409 * top-level memory cgroups), these two values are equal.
6410 * Otherwise, it's a part of parent's effective memory.low,
6411 * calculated as a cgroup's memory.low usage divided by sum of sibling's
6412 * memory.low usages, where memory.low usage is the size of actually
6416 * elow = min( memory.low, parent->elow * ------------------ ),
6417 * siblings_low_usage
6419 * | memory.current, if memory.current < memory.low
6424 * Such definition of the effective memory.low provides the expected
6425 * hierarchical behavior: parent's memory.low value is limiting
6426 * children, unprotected memory is reclaimed first and cgroups,
6427 * which are not using their guarantee do not affect actual memory
6430 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
6432 * A A/memory.low = 2G, A/memory.current = 6G
6434 * BC DE B/memory.low = 3G B/memory.current = 2G
6435 * C/memory.low = 1G C/memory.current = 2G
6436 * D/memory.low = 0 D/memory.current = 2G
6437 * E/memory.low = 10G E/memory.current = 0
6439 * and the memory pressure is applied, the following memory distribution
6440 * is expected (approximately):
6442 * A/memory.current = 2G
6444 * B/memory.current = 1.3G
6445 * C/memory.current = 0.6G
6446 * D/memory.current = 0
6447 * E/memory.current = 0
6449 * These calculations require constant tracking of the actual low usages
6450 * (see propagate_protected_usage()), as well as recursive calculation of
6451 * effective memory.low values. But as we do call mem_cgroup_protected()
6452 * path for each memory cgroup top-down from the reclaim,
6453 * it's possible to optimize this part, and save calculated elow
6454 * for next usage. This part is intentionally racy, but it's ok,
6455 * as memory.low is a best-effort mechanism.
6457 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
6458 struct mem_cgroup *memcg)
6460 struct mem_cgroup *parent;
6461 unsigned long emin, parent_emin;
6462 unsigned long elow, parent_elow;
6463 unsigned long usage;
6465 if (mem_cgroup_disabled())
6466 return MEMCG_PROT_NONE;
6469 root = root_mem_cgroup;
6472 * Effective values of the reclaim targets are ignored so they
6473 * can be stale. Have a look at mem_cgroup_protection for more
6475 * TODO: calculation should be more robust so that we do not need
6476 * that special casing.
6479 return MEMCG_PROT_NONE;
6481 usage = page_counter_read(&memcg->memory);
6483 return MEMCG_PROT_NONE;
6485 emin = memcg->memory.min;
6486 elow = memcg->memory.low;
6488 parent = parent_mem_cgroup(memcg);
6489 /* No parent means a non-hierarchical mode on v1 memcg */
6491 return MEMCG_PROT_NONE;
6496 parent_emin = READ_ONCE(parent->memory.emin);
6497 emin = min(emin, parent_emin);
6498 if (emin && parent_emin) {
6499 unsigned long min_usage, siblings_min_usage;
6501 min_usage = min(usage, memcg->memory.min);
6502 siblings_min_usage = atomic_long_read(
6503 &parent->memory.children_min_usage);
6505 if (min_usage && siblings_min_usage)
6506 emin = min(emin, parent_emin * min_usage /
6507 siblings_min_usage);
6510 parent_elow = READ_ONCE(parent->memory.elow);
6511 elow = min(elow, parent_elow);
6512 if (elow && parent_elow) {
6513 unsigned long low_usage, siblings_low_usage;
6515 low_usage = min(usage, memcg->memory.low);
6516 siblings_low_usage = atomic_long_read(
6517 &parent->memory.children_low_usage);
6519 if (low_usage && siblings_low_usage)
6520 elow = min(elow, parent_elow * low_usage /
6521 siblings_low_usage);
6525 memcg->memory.emin = emin;
6526 memcg->memory.elow = elow;
6529 return MEMCG_PROT_MIN;
6530 else if (usage <= elow)
6531 return MEMCG_PROT_LOW;
6533 return MEMCG_PROT_NONE;
6537 * mem_cgroup_try_charge - try charging a page
6538 * @page: page to charge
6539 * @mm: mm context of the victim
6540 * @gfp_mask: reclaim mode
6541 * @memcgp: charged memcg return
6542 * @compound: charge the page as compound or small page
6544 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6545 * pages according to @gfp_mask if necessary.
6547 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6548 * Otherwise, an error code is returned.
6550 * After page->mapping has been set up, the caller must finalize the
6551 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6552 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6554 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
6555 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6558 struct mem_cgroup *memcg = NULL;
6559 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6562 if (mem_cgroup_disabled())
6565 if (PageSwapCache(page)) {
6567 * Every swap fault against a single page tries to charge the
6568 * page, bail as early as possible. shmem_unuse() encounters
6569 * already charged pages, too. The USED bit is protected by
6570 * the page lock, which serializes swap cache removal, which
6571 * in turn serializes uncharging.
6573 VM_BUG_ON_PAGE(!PageLocked(page), page);
6574 if (compound_head(page)->mem_cgroup)
6577 if (do_swap_account) {
6578 swp_entry_t ent = { .val = page_private(page), };
6579 unsigned short id = lookup_swap_cgroup_id(ent);
6582 memcg = mem_cgroup_from_id(id);
6583 if (memcg && !css_tryget_online(&memcg->css))
6590 memcg = get_mem_cgroup_from_mm(mm);
6592 ret = try_charge(memcg, gfp_mask, nr_pages);
6594 css_put(&memcg->css);
6600 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
6601 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6604 struct mem_cgroup *memcg;
6607 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound);
6609 mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask);
6614 * mem_cgroup_commit_charge - commit a page charge
6615 * @page: page to charge
6616 * @memcg: memcg to charge the page to
6617 * @lrucare: page might be on LRU already
6618 * @compound: charge the page as compound or small page
6620 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6621 * after page->mapping has been set up. This must happen atomically
6622 * as part of the page instantiation, i.e. under the page table lock
6623 * for anonymous pages, under the page lock for page and swap cache.
6625 * In addition, the page must not be on the LRU during the commit, to
6626 * prevent racing with task migration. If it might be, use @lrucare.
6628 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6630 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
6631 bool lrucare, bool compound)
6633 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6635 VM_BUG_ON_PAGE(!page->mapping, page);
6636 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
6638 if (mem_cgroup_disabled())
6641 * Swap faults will attempt to charge the same page multiple
6642 * times. But reuse_swap_page() might have removed the page
6643 * from swapcache already, so we can't check PageSwapCache().
6648 commit_charge(page, memcg, lrucare);
6650 local_irq_disable();
6651 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
6652 memcg_check_events(memcg, page);
6655 if (do_memsw_account() && PageSwapCache(page)) {
6656 swp_entry_t entry = { .val = page_private(page) };
6658 * The swap entry might not get freed for a long time,
6659 * let's not wait for it. The page already received a
6660 * memory+swap charge, drop the swap entry duplicate.
6662 mem_cgroup_uncharge_swap(entry, nr_pages);
6667 * mem_cgroup_cancel_charge - cancel a page charge
6668 * @page: page to charge
6669 * @memcg: memcg to charge the page to
6670 * @compound: charge the page as compound or small page
6672 * Cancel a charge transaction started by mem_cgroup_try_charge().
6674 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
6677 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6679 if (mem_cgroup_disabled())
6682 * Swap faults will attempt to charge the same page multiple
6683 * times. But reuse_swap_page() might have removed the page
6684 * from swapcache already, so we can't check PageSwapCache().
6689 cancel_charge(memcg, nr_pages);
6692 struct uncharge_gather {
6693 struct mem_cgroup *memcg;
6694 unsigned long pgpgout;
6695 unsigned long nr_anon;
6696 unsigned long nr_file;
6697 unsigned long nr_kmem;
6698 unsigned long nr_huge;
6699 unsigned long nr_shmem;
6700 struct page *dummy_page;
6703 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6705 memset(ug, 0, sizeof(*ug));
6708 static void uncharge_batch(const struct uncharge_gather *ug)
6710 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
6711 unsigned long flags;
6713 if (!mem_cgroup_is_root(ug->memcg)) {
6714 page_counter_uncharge(&ug->memcg->memory, nr_pages);
6715 if (do_memsw_account())
6716 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
6717 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6718 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6719 memcg_oom_recover(ug->memcg);
6722 local_irq_save(flags);
6723 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
6724 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
6725 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
6726 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
6727 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6728 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, nr_pages);
6729 memcg_check_events(ug->memcg, ug->dummy_page);
6730 local_irq_restore(flags);
6732 if (!mem_cgroup_is_root(ug->memcg))
6733 css_put_many(&ug->memcg->css, nr_pages);
6736 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6738 VM_BUG_ON_PAGE(PageLRU(page), page);
6739 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
6740 !PageHWPoison(page) , page);
6742 if (!page->mem_cgroup)
6746 * Nobody should be changing or seriously looking at
6747 * page->mem_cgroup at this point, we have fully
6748 * exclusive access to the page.
6751 if (ug->memcg != page->mem_cgroup) {
6754 uncharge_gather_clear(ug);
6756 ug->memcg = page->mem_cgroup;
6759 if (!PageKmemcg(page)) {
6760 unsigned int nr_pages = 1;
6762 if (PageTransHuge(page)) {
6763 nr_pages = compound_nr(page);
6764 ug->nr_huge += nr_pages;
6767 ug->nr_anon += nr_pages;
6769 ug->nr_file += nr_pages;
6770 if (PageSwapBacked(page))
6771 ug->nr_shmem += nr_pages;
6775 ug->nr_kmem += compound_nr(page);
6776 __ClearPageKmemcg(page);
6779 ug->dummy_page = page;
6780 page->mem_cgroup = NULL;
6783 static void uncharge_list(struct list_head *page_list)
6785 struct uncharge_gather ug;
6786 struct list_head *next;
6788 uncharge_gather_clear(&ug);
6791 * Note that the list can be a single page->lru; hence the
6792 * do-while loop instead of a simple list_for_each_entry().
6794 next = page_list->next;
6798 page = list_entry(next, struct page, lru);
6799 next = page->lru.next;
6801 uncharge_page(page, &ug);
6802 } while (next != page_list);
6805 uncharge_batch(&ug);
6809 * mem_cgroup_uncharge - uncharge a page
6810 * @page: page to uncharge
6812 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6813 * mem_cgroup_commit_charge().
6815 void mem_cgroup_uncharge(struct page *page)
6817 struct uncharge_gather ug;
6819 if (mem_cgroup_disabled())
6822 /* Don't touch page->lru of any random page, pre-check: */
6823 if (!page->mem_cgroup)
6826 uncharge_gather_clear(&ug);
6827 uncharge_page(page, &ug);
6828 uncharge_batch(&ug);
6832 * mem_cgroup_uncharge_list - uncharge a list of page
6833 * @page_list: list of pages to uncharge
6835 * Uncharge a list of pages previously charged with
6836 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6838 void mem_cgroup_uncharge_list(struct list_head *page_list)
6840 if (mem_cgroup_disabled())
6843 if (!list_empty(page_list))
6844 uncharge_list(page_list);
6848 * mem_cgroup_migrate - charge a page's replacement
6849 * @oldpage: currently circulating page
6850 * @newpage: replacement page
6852 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6853 * be uncharged upon free.
6855 * Both pages must be locked, @newpage->mapping must be set up.
6857 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6859 struct mem_cgroup *memcg;
6860 unsigned int nr_pages;
6862 unsigned long flags;
6864 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6865 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6866 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6867 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6870 if (mem_cgroup_disabled())
6873 /* Page cache replacement: new page already charged? */
6874 if (newpage->mem_cgroup)
6877 /* Swapcache readahead pages can get replaced before being charged */
6878 memcg = oldpage->mem_cgroup;
6882 /* Force-charge the new page. The old one will be freed soon */
6883 compound = PageTransHuge(newpage);
6884 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
6886 page_counter_charge(&memcg->memory, nr_pages);
6887 if (do_memsw_account())
6888 page_counter_charge(&memcg->memsw, nr_pages);
6889 css_get_many(&memcg->css, nr_pages);
6891 commit_charge(newpage, memcg, false);
6893 local_irq_save(flags);
6894 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
6895 memcg_check_events(memcg, newpage);
6896 local_irq_restore(flags);
6899 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6900 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6902 void mem_cgroup_sk_alloc(struct sock *sk)
6904 struct mem_cgroup *memcg;
6906 if (!mem_cgroup_sockets_enabled)
6909 /* Do not associate the sock with unrelated interrupted task's memcg. */
6914 memcg = mem_cgroup_from_task(current);
6915 if (memcg == root_mem_cgroup)
6917 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6919 if (css_tryget_online(&memcg->css))
6920 sk->sk_memcg = memcg;
6925 void mem_cgroup_sk_free(struct sock *sk)
6928 css_put(&sk->sk_memcg->css);
6932 * mem_cgroup_charge_skmem - charge socket memory
6933 * @memcg: memcg to charge
6934 * @nr_pages: number of pages to charge
6936 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6937 * @memcg's configured limit, %false if the charge had to be forced.
6939 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6941 gfp_t gfp_mask = GFP_KERNEL;
6943 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6944 struct page_counter *fail;
6946 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6947 memcg->tcpmem_pressure = 0;
6950 page_counter_charge(&memcg->tcpmem, nr_pages);
6951 memcg->tcpmem_pressure = 1;
6955 /* Don't block in the packet receive path */
6957 gfp_mask = GFP_NOWAIT;
6959 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6961 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6964 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6969 * mem_cgroup_uncharge_skmem - uncharge socket memory
6970 * @memcg: memcg to uncharge
6971 * @nr_pages: number of pages to uncharge
6973 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6975 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6976 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6980 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6982 refill_stock(memcg, nr_pages);
6985 static int __init cgroup_memory(char *s)
6989 while ((token = strsep(&s, ",")) != NULL) {
6992 if (!strcmp(token, "nosocket"))
6993 cgroup_memory_nosocket = true;
6994 if (!strcmp(token, "nokmem"))
6995 cgroup_memory_nokmem = true;
6999 __setup("cgroup.memory=", cgroup_memory);
7002 * subsys_initcall() for memory controller.
7004 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7005 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7006 * basically everything that doesn't depend on a specific mem_cgroup structure
7007 * should be initialized from here.
7009 static int __init mem_cgroup_init(void)
7013 #ifdef CONFIG_MEMCG_KMEM
7015 * Kmem cache creation is mostly done with the slab_mutex held,
7016 * so use a workqueue with limited concurrency to avoid stalling
7017 * all worker threads in case lots of cgroups are created and
7018 * destroyed simultaneously.
7020 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
7021 BUG_ON(!memcg_kmem_cache_wq);
7024 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7025 memcg_hotplug_cpu_dead);
7027 for_each_possible_cpu(cpu)
7028 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7031 for_each_node(node) {
7032 struct mem_cgroup_tree_per_node *rtpn;
7034 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7035 node_online(node) ? node : NUMA_NO_NODE);
7037 rtpn->rb_root = RB_ROOT;
7038 rtpn->rb_rightmost = NULL;
7039 spin_lock_init(&rtpn->lock);
7040 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7045 subsys_initcall(mem_cgroup_init);
7047 #ifdef CONFIG_MEMCG_SWAP
7048 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7050 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7052 * The root cgroup cannot be destroyed, so it's refcount must
7055 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7059 memcg = parent_mem_cgroup(memcg);
7061 memcg = root_mem_cgroup;
7067 * mem_cgroup_swapout - transfer a memsw charge to swap
7068 * @page: page whose memsw charge to transfer
7069 * @entry: swap entry to move the charge to
7071 * Transfer the memsw charge of @page to @entry.
7073 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7075 struct mem_cgroup *memcg, *swap_memcg;
7076 unsigned int nr_entries;
7077 unsigned short oldid;
7079 VM_BUG_ON_PAGE(PageLRU(page), page);
7080 VM_BUG_ON_PAGE(page_count(page), page);
7082 if (!do_memsw_account())
7085 memcg = page->mem_cgroup;
7087 /* Readahead page, never charged */
7092 * In case the memcg owning these pages has been offlined and doesn't
7093 * have an ID allocated to it anymore, charge the closest online
7094 * ancestor for the swap instead and transfer the memory+swap charge.
7096 swap_memcg = mem_cgroup_id_get_online(memcg);
7097 nr_entries = hpage_nr_pages(page);
7098 /* Get references for the tail pages, too */
7100 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7101 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7103 VM_BUG_ON_PAGE(oldid, page);
7104 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7106 page->mem_cgroup = NULL;
7108 if (!mem_cgroup_is_root(memcg))
7109 page_counter_uncharge(&memcg->memory, nr_entries);
7111 if (memcg != swap_memcg) {
7112 if (!mem_cgroup_is_root(swap_memcg))
7113 page_counter_charge(&swap_memcg->memsw, nr_entries);
7114 page_counter_uncharge(&memcg->memsw, nr_entries);
7118 * Interrupts should be disabled here because the caller holds the
7119 * i_pages lock which is taken with interrupts-off. It is
7120 * important here to have the interrupts disabled because it is the
7121 * only synchronisation we have for updating the per-CPU variables.
7123 VM_BUG_ON(!irqs_disabled());
7124 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
7126 memcg_check_events(memcg, page);
7128 if (!mem_cgroup_is_root(memcg))
7129 css_put_many(&memcg->css, nr_entries);
7133 * mem_cgroup_try_charge_swap - try charging swap space for a page
7134 * @page: page being added to swap
7135 * @entry: swap entry to charge
7137 * Try to charge @page's memcg for the swap space at @entry.
7139 * Returns 0 on success, -ENOMEM on failure.
7141 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7143 unsigned int nr_pages = hpage_nr_pages(page);
7144 struct page_counter *counter;
7145 struct mem_cgroup *memcg;
7146 unsigned short oldid;
7148 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
7151 memcg = page->mem_cgroup;
7153 /* Readahead page, never charged */
7158 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7162 memcg = mem_cgroup_id_get_online(memcg);
7164 if (!mem_cgroup_is_root(memcg) &&
7165 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7166 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7167 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7168 mem_cgroup_id_put(memcg);
7172 /* Get references for the tail pages, too */
7174 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7175 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7176 VM_BUG_ON_PAGE(oldid, page);
7177 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7183 * mem_cgroup_uncharge_swap - uncharge swap space
7184 * @entry: swap entry to uncharge
7185 * @nr_pages: the amount of swap space to uncharge
7187 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7189 struct mem_cgroup *memcg;
7192 if (!do_swap_account)
7195 id = swap_cgroup_record(entry, 0, nr_pages);
7197 memcg = mem_cgroup_from_id(id);
7199 if (!mem_cgroup_is_root(memcg)) {
7200 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7201 page_counter_uncharge(&memcg->swap, nr_pages);
7203 page_counter_uncharge(&memcg->memsw, nr_pages);
7205 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7206 mem_cgroup_id_put_many(memcg, nr_pages);
7211 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7213 long nr_swap_pages = get_nr_swap_pages();
7215 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7216 return nr_swap_pages;
7217 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7218 nr_swap_pages = min_t(long, nr_swap_pages,
7219 READ_ONCE(memcg->swap.max) -
7220 page_counter_read(&memcg->swap));
7221 return nr_swap_pages;
7224 bool mem_cgroup_swap_full(struct page *page)
7226 struct mem_cgroup *memcg;
7228 VM_BUG_ON_PAGE(!PageLocked(page), page);
7232 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7235 memcg = page->mem_cgroup;
7239 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7240 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.max)
7246 /* for remember boot option*/
7247 #ifdef CONFIG_MEMCG_SWAP_ENABLED
7248 static int really_do_swap_account __initdata = 1;
7250 static int really_do_swap_account __initdata;
7253 static int __init enable_swap_account(char *s)
7255 if (!strcmp(s, "1"))
7256 really_do_swap_account = 1;
7257 else if (!strcmp(s, "0"))
7258 really_do_swap_account = 0;
7261 __setup("swapaccount=", enable_swap_account);
7263 static u64 swap_current_read(struct cgroup_subsys_state *css,
7266 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7268 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7271 static int swap_max_show(struct seq_file *m, void *v)
7273 return seq_puts_memcg_tunable(m,
7274 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7277 static ssize_t swap_max_write(struct kernfs_open_file *of,
7278 char *buf, size_t nbytes, loff_t off)
7280 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7284 buf = strstrip(buf);
7285 err = page_counter_memparse(buf, "max", &max);
7289 xchg(&memcg->swap.max, max);
7294 static int swap_events_show(struct seq_file *m, void *v)
7296 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7298 seq_printf(m, "max %lu\n",
7299 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7300 seq_printf(m, "fail %lu\n",
7301 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7306 static struct cftype swap_files[] = {
7308 .name = "swap.current",
7309 .flags = CFTYPE_NOT_ON_ROOT,
7310 .read_u64 = swap_current_read,
7314 .flags = CFTYPE_NOT_ON_ROOT,
7315 .seq_show = swap_max_show,
7316 .write = swap_max_write,
7319 .name = "swap.events",
7320 .flags = CFTYPE_NOT_ON_ROOT,
7321 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7322 .seq_show = swap_events_show,
7327 static struct cftype memsw_cgroup_files[] = {
7329 .name = "memsw.usage_in_bytes",
7330 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7331 .read_u64 = mem_cgroup_read_u64,
7334 .name = "memsw.max_usage_in_bytes",
7335 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7336 .write = mem_cgroup_reset,
7337 .read_u64 = mem_cgroup_read_u64,
7340 .name = "memsw.limit_in_bytes",
7341 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7342 .write = mem_cgroup_write,
7343 .read_u64 = mem_cgroup_read_u64,
7346 .name = "memsw.failcnt",
7347 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7348 .write = mem_cgroup_reset,
7349 .read_u64 = mem_cgroup_read_u64,
7351 { }, /* terminate */
7354 static int __init mem_cgroup_swap_init(void)
7356 if (!mem_cgroup_disabled() && really_do_swap_account) {
7357 do_swap_account = 1;
7358 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
7360 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
7361 memsw_cgroup_files));
7365 subsys_initcall(mem_cgroup_swap_init);
7367 #endif /* CONFIG_MEMCG_SWAP */