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 *uninitialized_var(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;
2217 if (stock->nr_pages) {
2218 page_counter_uncharge(&old->memory, stock->nr_pages);
2219 if (do_memsw_account())
2220 page_counter_uncharge(&old->memsw, stock->nr_pages);
2221 css_put_many(&old->css, stock->nr_pages);
2222 stock->nr_pages = 0;
2224 stock->cached = NULL;
2227 static void drain_local_stock(struct work_struct *dummy)
2229 struct memcg_stock_pcp *stock;
2230 unsigned long flags;
2233 * The only protection from memory hotplug vs. drain_stock races is
2234 * that we always operate on local CPU stock here with IRQ disabled
2236 local_irq_save(flags);
2238 stock = this_cpu_ptr(&memcg_stock);
2240 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2242 local_irq_restore(flags);
2246 * Cache charges(val) to local per_cpu area.
2247 * This will be consumed by consume_stock() function, later.
2249 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2251 struct memcg_stock_pcp *stock;
2252 unsigned long flags;
2254 local_irq_save(flags);
2256 stock = this_cpu_ptr(&memcg_stock);
2257 if (stock->cached != memcg) { /* reset if necessary */
2259 stock->cached = memcg;
2261 stock->nr_pages += nr_pages;
2263 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2266 local_irq_restore(flags);
2270 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2271 * of the hierarchy under it.
2273 static void drain_all_stock(struct mem_cgroup *root_memcg)
2277 /* If someone's already draining, avoid adding running more workers. */
2278 if (!mutex_trylock(&percpu_charge_mutex))
2281 * Notify other cpus that system-wide "drain" is running
2282 * We do not care about races with the cpu hotplug because cpu down
2283 * as well as workers from this path always operate on the local
2284 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2287 for_each_online_cpu(cpu) {
2288 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2289 struct mem_cgroup *memcg;
2293 memcg = stock->cached;
2294 if (memcg && stock->nr_pages &&
2295 mem_cgroup_is_descendant(memcg, root_memcg))
2300 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2302 drain_local_stock(&stock->work);
2304 schedule_work_on(cpu, &stock->work);
2308 mutex_unlock(&percpu_charge_mutex);
2311 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2313 struct memcg_stock_pcp *stock;
2314 struct mem_cgroup *memcg, *mi;
2316 stock = &per_cpu(memcg_stock, cpu);
2319 for_each_mem_cgroup(memcg) {
2322 for (i = 0; i < MEMCG_NR_STAT; i++) {
2326 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2328 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2329 atomic_long_add(x, &memcg->vmstats[i]);
2331 if (i >= NR_VM_NODE_STAT_ITEMS)
2334 for_each_node(nid) {
2335 struct mem_cgroup_per_node *pn;
2337 pn = mem_cgroup_nodeinfo(memcg, nid);
2338 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2341 atomic_long_add(x, &pn->lruvec_stat[i]);
2342 } while ((pn = parent_nodeinfo(pn, nid)));
2346 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2349 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2351 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2352 atomic_long_add(x, &memcg->vmevents[i]);
2359 static void reclaim_high(struct mem_cgroup *memcg,
2360 unsigned int nr_pages,
2364 if (page_counter_read(&memcg->memory) <= memcg->high)
2366 memcg_memory_event(memcg, MEMCG_HIGH);
2367 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2368 } while ((memcg = parent_mem_cgroup(memcg)));
2371 static void high_work_func(struct work_struct *work)
2373 struct mem_cgroup *memcg;
2375 memcg = container_of(work, struct mem_cgroup, high_work);
2376 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2380 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2381 * enough to still cause a significant slowdown in most cases, while still
2382 * allowing diagnostics and tracing to proceed without becoming stuck.
2384 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2387 * When calculating the delay, we use these either side of the exponentiation to
2388 * maintain precision and scale to a reasonable number of jiffies (see the table
2391 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2392 * overage ratio to a delay.
2393 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down down the
2394 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2395 * to produce a reasonable delay curve.
2397 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2398 * reasonable delay curve compared to precision-adjusted overage, not
2399 * penalising heavily at first, but still making sure that growth beyond the
2400 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2401 * example, with a high of 100 megabytes:
2403 * +-------+------------------------+
2404 * | usage | time to allocate in ms |
2405 * +-------+------------------------+
2427 * +-------+------------------------+
2429 #define MEMCG_DELAY_PRECISION_SHIFT 20
2430 #define MEMCG_DELAY_SCALING_SHIFT 14
2433 * Get the number of jiffies that we should penalise a mischievous cgroup which
2434 * is exceeding its memory.high by checking both it and its ancestors.
2436 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2437 unsigned int nr_pages)
2439 unsigned long penalty_jiffies;
2440 u64 max_overage = 0;
2443 unsigned long usage, high;
2446 usage = page_counter_read(&memcg->memory);
2447 high = READ_ONCE(memcg->high);
2453 * Prevent division by 0 in overage calculation by acting as if
2454 * it was a threshold of 1 page
2456 high = max(high, 1UL);
2458 overage = usage - high;
2459 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2460 overage = div64_u64(overage, high);
2462 if (overage > max_overage)
2463 max_overage = overage;
2464 } while ((memcg = parent_mem_cgroup(memcg)) &&
2465 !mem_cgroup_is_root(memcg));
2471 * We use overage compared to memory.high to calculate the number of
2472 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2473 * fairly lenient on small overages, and increasingly harsh when the
2474 * memcg in question makes it clear that it has no intention of stopping
2475 * its crazy behaviour, so we exponentially increase the delay based on
2478 penalty_jiffies = max_overage * max_overage * HZ;
2479 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2480 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2483 * Factor in the task's own contribution to the overage, such that four
2484 * N-sized allocations are throttled approximately the same as one
2485 * 4N-sized allocation.
2487 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2488 * larger the current charge patch is than that.
2490 penalty_jiffies = penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2493 * Clamp the max delay per usermode return so as to still keep the
2494 * application moving forwards and also permit diagnostics, albeit
2497 return min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2501 * Scheduled by try_charge() to be executed from the userland return path
2502 * and reclaims memory over the high limit.
2504 void mem_cgroup_handle_over_high(void)
2506 unsigned long penalty_jiffies;
2507 unsigned long pflags;
2508 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2509 struct mem_cgroup *memcg;
2511 if (likely(!nr_pages))
2514 memcg = get_mem_cgroup_from_mm(current->mm);
2515 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2516 current->memcg_nr_pages_over_high = 0;
2519 * memory.high is breached and reclaim is unable to keep up. Throttle
2520 * allocators proactively to slow down excessive growth.
2522 penalty_jiffies = calculate_high_delay(memcg, nr_pages);
2525 * Don't sleep if the amount of jiffies this memcg owes us is so low
2526 * that it's not even worth doing, in an attempt to be nice to those who
2527 * go only a small amount over their memory.high value and maybe haven't
2528 * been aggressively reclaimed enough yet.
2530 if (penalty_jiffies <= HZ / 100)
2534 * If we exit early, we're guaranteed to die (since
2535 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2536 * need to account for any ill-begotten jiffies to pay them off later.
2538 psi_memstall_enter(&pflags);
2539 schedule_timeout_killable(penalty_jiffies);
2540 psi_memstall_leave(&pflags);
2543 css_put(&memcg->css);
2546 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2547 unsigned int nr_pages)
2549 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2550 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2551 struct mem_cgroup *mem_over_limit;
2552 struct page_counter *counter;
2553 unsigned long nr_reclaimed;
2554 bool may_swap = true;
2555 bool drained = false;
2556 enum oom_status oom_status;
2558 if (mem_cgroup_is_root(memcg))
2561 if (consume_stock(memcg, nr_pages))
2564 if (!do_memsw_account() ||
2565 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2566 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2568 if (do_memsw_account())
2569 page_counter_uncharge(&memcg->memsw, batch);
2570 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2572 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2576 if (batch > nr_pages) {
2582 * Memcg doesn't have a dedicated reserve for atomic
2583 * allocations. But like the global atomic pool, we need to
2584 * put the burden of reclaim on regular allocation requests
2585 * and let these go through as privileged allocations.
2587 if (gfp_mask & __GFP_ATOMIC)
2591 * Unlike in global OOM situations, memcg is not in a physical
2592 * memory shortage. Allow dying and OOM-killed tasks to
2593 * bypass the last charges so that they can exit quickly and
2594 * free their memory.
2596 if (unlikely(should_force_charge()))
2600 * Prevent unbounded recursion when reclaim operations need to
2601 * allocate memory. This might exceed the limits temporarily,
2602 * but we prefer facilitating memory reclaim and getting back
2603 * under the limit over triggering OOM kills in these cases.
2605 if (unlikely(current->flags & PF_MEMALLOC))
2608 if (unlikely(task_in_memcg_oom(current)))
2611 if (!gfpflags_allow_blocking(gfp_mask))
2614 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2616 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2617 gfp_mask, may_swap);
2619 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2623 drain_all_stock(mem_over_limit);
2628 if (gfp_mask & __GFP_NORETRY)
2631 * Even though the limit is exceeded at this point, reclaim
2632 * may have been able to free some pages. Retry the charge
2633 * before killing the task.
2635 * Only for regular pages, though: huge pages are rather
2636 * unlikely to succeed so close to the limit, and we fall back
2637 * to regular pages anyway in case of failure.
2639 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2642 * At task move, charge accounts can be doubly counted. So, it's
2643 * better to wait until the end of task_move if something is going on.
2645 if (mem_cgroup_wait_acct_move(mem_over_limit))
2651 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2654 if (gfp_mask & __GFP_NOFAIL)
2657 if (fatal_signal_pending(current))
2661 * keep retrying as long as the memcg oom killer is able to make
2662 * a forward progress or bypass the charge if the oom killer
2663 * couldn't make any progress.
2665 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2666 get_order(nr_pages * PAGE_SIZE));
2667 switch (oom_status) {
2669 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2677 if (!(gfp_mask & __GFP_NOFAIL))
2681 * The allocation either can't fail or will lead to more memory
2682 * being freed very soon. Allow memory usage go over the limit
2683 * temporarily by force charging it.
2685 page_counter_charge(&memcg->memory, nr_pages);
2686 if (do_memsw_account())
2687 page_counter_charge(&memcg->memsw, nr_pages);
2688 css_get_many(&memcg->css, nr_pages);
2693 css_get_many(&memcg->css, batch);
2694 if (batch > nr_pages)
2695 refill_stock(memcg, batch - nr_pages);
2698 * If the hierarchy is above the normal consumption range, schedule
2699 * reclaim on returning to userland. We can perform reclaim here
2700 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2701 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2702 * not recorded as it most likely matches current's and won't
2703 * change in the meantime. As high limit is checked again before
2704 * reclaim, the cost of mismatch is negligible.
2707 if (page_counter_read(&memcg->memory) > memcg->high) {
2708 /* Don't bother a random interrupted task */
2709 if (in_interrupt()) {
2710 schedule_work(&memcg->high_work);
2713 current->memcg_nr_pages_over_high += batch;
2714 set_notify_resume(current);
2717 } while ((memcg = parent_mem_cgroup(memcg)));
2722 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2724 if (mem_cgroup_is_root(memcg))
2727 page_counter_uncharge(&memcg->memory, nr_pages);
2728 if (do_memsw_account())
2729 page_counter_uncharge(&memcg->memsw, nr_pages);
2731 css_put_many(&memcg->css, nr_pages);
2734 static void lock_page_lru(struct page *page, int *isolated)
2736 pg_data_t *pgdat = page_pgdat(page);
2738 spin_lock_irq(&pgdat->lru_lock);
2739 if (PageLRU(page)) {
2740 struct lruvec *lruvec;
2742 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2744 del_page_from_lru_list(page, lruvec, page_lru(page));
2750 static void unlock_page_lru(struct page *page, int isolated)
2752 pg_data_t *pgdat = page_pgdat(page);
2755 struct lruvec *lruvec;
2757 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2758 VM_BUG_ON_PAGE(PageLRU(page), page);
2760 add_page_to_lru_list(page, lruvec, page_lru(page));
2762 spin_unlock_irq(&pgdat->lru_lock);
2765 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2770 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2773 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2774 * may already be on some other mem_cgroup's LRU. Take care of it.
2777 lock_page_lru(page, &isolated);
2780 * Nobody should be changing or seriously looking at
2781 * page->mem_cgroup at this point:
2783 * - the page is uncharged
2785 * - the page is off-LRU
2787 * - an anonymous fault has exclusive page access, except for
2788 * a locked page table
2790 * - a page cache insertion, a swapin fault, or a migration
2791 * have the page locked
2793 page->mem_cgroup = memcg;
2796 unlock_page_lru(page, isolated);
2799 #ifdef CONFIG_MEMCG_KMEM
2801 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2803 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2804 * cgroup_mutex, etc.
2806 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2810 if (mem_cgroup_disabled())
2813 page = virt_to_head_page(p);
2816 * Slab pages don't have page->mem_cgroup set because corresponding
2817 * kmem caches can be reparented during the lifetime. That's why
2818 * memcg_from_slab_page() should be used instead.
2821 return memcg_from_slab_page(page);
2823 /* All other pages use page->mem_cgroup */
2824 return page->mem_cgroup;
2827 static int memcg_alloc_cache_id(void)
2832 id = ida_simple_get(&memcg_cache_ida,
2833 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2837 if (id < memcg_nr_cache_ids)
2841 * There's no space for the new id in memcg_caches arrays,
2842 * so we have to grow them.
2844 down_write(&memcg_cache_ids_sem);
2846 size = 2 * (id + 1);
2847 if (size < MEMCG_CACHES_MIN_SIZE)
2848 size = MEMCG_CACHES_MIN_SIZE;
2849 else if (size > MEMCG_CACHES_MAX_SIZE)
2850 size = MEMCG_CACHES_MAX_SIZE;
2852 err = memcg_update_all_caches(size);
2854 err = memcg_update_all_list_lrus(size);
2856 memcg_nr_cache_ids = size;
2858 up_write(&memcg_cache_ids_sem);
2861 ida_simple_remove(&memcg_cache_ida, id);
2867 static void memcg_free_cache_id(int id)
2869 ida_simple_remove(&memcg_cache_ida, id);
2872 struct memcg_kmem_cache_create_work {
2873 struct mem_cgroup *memcg;
2874 struct kmem_cache *cachep;
2875 struct work_struct work;
2878 static void memcg_kmem_cache_create_func(struct work_struct *w)
2880 struct memcg_kmem_cache_create_work *cw =
2881 container_of(w, struct memcg_kmem_cache_create_work, work);
2882 struct mem_cgroup *memcg = cw->memcg;
2883 struct kmem_cache *cachep = cw->cachep;
2885 memcg_create_kmem_cache(memcg, cachep);
2887 css_put(&memcg->css);
2892 * Enqueue the creation of a per-memcg kmem_cache.
2894 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2895 struct kmem_cache *cachep)
2897 struct memcg_kmem_cache_create_work *cw;
2899 if (!css_tryget_online(&memcg->css))
2902 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2904 css_put(&memcg->css);
2909 cw->cachep = cachep;
2910 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2912 queue_work(memcg_kmem_cache_wq, &cw->work);
2915 static inline bool memcg_kmem_bypass(void)
2917 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2923 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2924 * @cachep: the original global kmem cache
2926 * Return the kmem_cache we're supposed to use for a slab allocation.
2927 * We try to use the current memcg's version of the cache.
2929 * If the cache does not exist yet, if we are the first user of it, we
2930 * create it asynchronously in a workqueue and let the current allocation
2931 * go through with the original cache.
2933 * This function takes a reference to the cache it returns to assure it
2934 * won't get destroyed while we are working with it. Once the caller is
2935 * done with it, memcg_kmem_put_cache() must be called to release the
2938 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2940 struct mem_cgroup *memcg;
2941 struct kmem_cache *memcg_cachep;
2942 struct memcg_cache_array *arr;
2945 VM_BUG_ON(!is_root_cache(cachep));
2947 if (memcg_kmem_bypass())
2952 if (unlikely(current->active_memcg))
2953 memcg = current->active_memcg;
2955 memcg = mem_cgroup_from_task(current);
2957 if (!memcg || memcg == root_mem_cgroup)
2960 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2964 arr = rcu_dereference(cachep->memcg_params.memcg_caches);
2967 * Make sure we will access the up-to-date value. The code updating
2968 * memcg_caches issues a write barrier to match the data dependency
2969 * barrier inside READ_ONCE() (see memcg_create_kmem_cache()).
2971 memcg_cachep = READ_ONCE(arr->entries[kmemcg_id]);
2974 * If we are in a safe context (can wait, and not in interrupt
2975 * context), we could be be predictable and return right away.
2976 * This would guarantee that the allocation being performed
2977 * already belongs in the new cache.
2979 * However, there are some clashes that can arrive from locking.
2980 * For instance, because we acquire the slab_mutex while doing
2981 * memcg_create_kmem_cache, this means no further allocation
2982 * could happen with the slab_mutex held. So it's better to
2985 * If the memcg is dying or memcg_cache is about to be released,
2986 * don't bother creating new kmem_caches. Because memcg_cachep
2987 * is ZEROed as the fist step of kmem offlining, we don't need
2988 * percpu_ref_tryget_live() here. css_tryget_online() check in
2989 * memcg_schedule_kmem_cache_create() will prevent us from
2990 * creation of a new kmem_cache.
2992 if (unlikely(!memcg_cachep))
2993 memcg_schedule_kmem_cache_create(memcg, cachep);
2994 else if (percpu_ref_tryget(&memcg_cachep->memcg_params.refcnt))
2995 cachep = memcg_cachep;
3002 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
3003 * @cachep: the cache returned by memcg_kmem_get_cache
3005 void memcg_kmem_put_cache(struct kmem_cache *cachep)
3007 if (!is_root_cache(cachep))
3008 percpu_ref_put(&cachep->memcg_params.refcnt);
3012 * __memcg_kmem_charge_memcg: charge a kmem page
3013 * @page: page to charge
3014 * @gfp: reclaim mode
3015 * @order: allocation order
3016 * @memcg: memory cgroup to charge
3018 * Returns 0 on success, an error code on failure.
3020 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
3021 struct mem_cgroup *memcg)
3023 unsigned int nr_pages = 1 << order;
3024 struct page_counter *counter;
3027 ret = try_charge(memcg, gfp, nr_pages);
3031 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3032 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3035 * Enforce __GFP_NOFAIL allocation because callers are not
3036 * prepared to see failures and likely do not have any failure
3039 if (gfp & __GFP_NOFAIL) {
3040 page_counter_charge(&memcg->kmem, nr_pages);
3043 cancel_charge(memcg, nr_pages);
3050 * __memcg_kmem_charge: charge a kmem page to the current memory cgroup
3051 * @page: page to charge
3052 * @gfp: reclaim mode
3053 * @order: allocation order
3055 * Returns 0 on success, an error code on failure.
3057 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
3059 struct mem_cgroup *memcg;
3062 if (memcg_kmem_bypass())
3065 memcg = get_mem_cgroup_from_current();
3066 if (!mem_cgroup_is_root(memcg)) {
3067 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
3069 page->mem_cgroup = memcg;
3070 __SetPageKmemcg(page);
3073 css_put(&memcg->css);
3078 * __memcg_kmem_uncharge_memcg: uncharge a kmem page
3079 * @memcg: memcg to uncharge
3080 * @nr_pages: number of pages to uncharge
3082 void __memcg_kmem_uncharge_memcg(struct mem_cgroup *memcg,
3083 unsigned int nr_pages)
3085 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3086 page_counter_uncharge(&memcg->kmem, nr_pages);
3088 page_counter_uncharge(&memcg->memory, nr_pages);
3089 if (do_memsw_account())
3090 page_counter_uncharge(&memcg->memsw, nr_pages);
3093 * __memcg_kmem_uncharge: uncharge a kmem page
3094 * @page: page to uncharge
3095 * @order: allocation order
3097 void __memcg_kmem_uncharge(struct page *page, int order)
3099 struct mem_cgroup *memcg = page->mem_cgroup;
3100 unsigned int nr_pages = 1 << order;
3105 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3106 __memcg_kmem_uncharge_memcg(memcg, nr_pages);
3107 page->mem_cgroup = NULL;
3109 /* slab pages do not have PageKmemcg flag set */
3110 if (PageKmemcg(page))
3111 __ClearPageKmemcg(page);
3113 css_put_many(&memcg->css, nr_pages);
3115 #endif /* CONFIG_MEMCG_KMEM */
3117 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3120 * Because tail pages are not marked as "used", set it. We're under
3121 * pgdat->lru_lock and migration entries setup in all page mappings.
3123 void mem_cgroup_split_huge_fixup(struct page *head)
3127 if (mem_cgroup_disabled())
3130 for (i = 1; i < HPAGE_PMD_NR; i++)
3131 head[i].mem_cgroup = head->mem_cgroup;
3133 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
3135 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3137 #ifdef CONFIG_MEMCG_SWAP
3139 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3140 * @entry: swap entry to be moved
3141 * @from: mem_cgroup which the entry is moved from
3142 * @to: mem_cgroup which the entry is moved to
3144 * It succeeds only when the swap_cgroup's record for this entry is the same
3145 * as the mem_cgroup's id of @from.
3147 * Returns 0 on success, -EINVAL on failure.
3149 * The caller must have charged to @to, IOW, called page_counter_charge() about
3150 * both res and memsw, and called css_get().
3152 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3153 struct mem_cgroup *from, struct mem_cgroup *to)
3155 unsigned short old_id, new_id;
3157 old_id = mem_cgroup_id(from);
3158 new_id = mem_cgroup_id(to);
3160 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3161 mod_memcg_state(from, MEMCG_SWAP, -1);
3162 mod_memcg_state(to, MEMCG_SWAP, 1);
3168 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3169 struct mem_cgroup *from, struct mem_cgroup *to)
3175 static DEFINE_MUTEX(memcg_max_mutex);
3177 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3178 unsigned long max, bool memsw)
3180 bool enlarge = false;
3181 bool drained = false;
3183 bool limits_invariant;
3184 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3187 if (signal_pending(current)) {
3192 mutex_lock(&memcg_max_mutex);
3194 * Make sure that the new limit (memsw or memory limit) doesn't
3195 * break our basic invariant rule memory.max <= memsw.max.
3197 limits_invariant = memsw ? max >= memcg->memory.max :
3198 max <= memcg->memsw.max;
3199 if (!limits_invariant) {
3200 mutex_unlock(&memcg_max_mutex);
3204 if (max > counter->max)
3206 ret = page_counter_set_max(counter, max);
3207 mutex_unlock(&memcg_max_mutex);
3213 drain_all_stock(memcg);
3218 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3219 GFP_KERNEL, !memsw)) {
3225 if (!ret && enlarge)
3226 memcg_oom_recover(memcg);
3231 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3233 unsigned long *total_scanned)
3235 unsigned long nr_reclaimed = 0;
3236 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3237 unsigned long reclaimed;
3239 struct mem_cgroup_tree_per_node *mctz;
3240 unsigned long excess;
3241 unsigned long nr_scanned;
3246 mctz = soft_limit_tree_node(pgdat->node_id);
3249 * Do not even bother to check the largest node if the root
3250 * is empty. Do it lockless to prevent lock bouncing. Races
3251 * are acceptable as soft limit is best effort anyway.
3253 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3257 * This loop can run a while, specially if mem_cgroup's continuously
3258 * keep exceeding their soft limit and putting the system under
3265 mz = mem_cgroup_largest_soft_limit_node(mctz);
3270 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3271 gfp_mask, &nr_scanned);
3272 nr_reclaimed += reclaimed;
3273 *total_scanned += nr_scanned;
3274 spin_lock_irq(&mctz->lock);
3275 __mem_cgroup_remove_exceeded(mz, mctz);
3278 * If we failed to reclaim anything from this memory cgroup
3279 * it is time to move on to the next cgroup
3283 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3285 excess = soft_limit_excess(mz->memcg);
3287 * One school of thought says that we should not add
3288 * back the node to the tree if reclaim returns 0.
3289 * But our reclaim could return 0, simply because due
3290 * to priority we are exposing a smaller subset of
3291 * memory to reclaim from. Consider this as a longer
3294 /* If excess == 0, no tree ops */
3295 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3296 spin_unlock_irq(&mctz->lock);
3297 css_put(&mz->memcg->css);
3300 * Could not reclaim anything and there are no more
3301 * mem cgroups to try or we seem to be looping without
3302 * reclaiming anything.
3304 if (!nr_reclaimed &&
3306 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3308 } while (!nr_reclaimed);
3310 css_put(&next_mz->memcg->css);
3311 return nr_reclaimed;
3315 * Test whether @memcg has children, dead or alive. Note that this
3316 * function doesn't care whether @memcg has use_hierarchy enabled and
3317 * returns %true if there are child csses according to the cgroup
3318 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3320 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3325 ret = css_next_child(NULL, &memcg->css);
3331 * Reclaims as many pages from the given memcg as possible.
3333 * Caller is responsible for holding css reference for memcg.
3335 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3337 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3339 /* we call try-to-free pages for make this cgroup empty */
3340 lru_add_drain_all();
3342 drain_all_stock(memcg);
3344 /* try to free all pages in this cgroup */
3345 while (nr_retries && page_counter_read(&memcg->memory)) {
3348 if (signal_pending(current))
3351 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3355 /* maybe some writeback is necessary */
3356 congestion_wait(BLK_RW_ASYNC, HZ/10);
3364 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3365 char *buf, size_t nbytes,
3368 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3370 if (mem_cgroup_is_root(memcg))
3372 return mem_cgroup_force_empty(memcg) ?: nbytes;
3375 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3378 return mem_cgroup_from_css(css)->use_hierarchy;
3381 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3382 struct cftype *cft, u64 val)
3385 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3386 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3388 if (memcg->use_hierarchy == val)
3392 * If parent's use_hierarchy is set, we can't make any modifications
3393 * in the child subtrees. If it is unset, then the change can
3394 * occur, provided the current cgroup has no children.
3396 * For the root cgroup, parent_mem is NULL, we allow value to be
3397 * set if there are no children.
3399 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3400 (val == 1 || val == 0)) {
3401 if (!memcg_has_children(memcg))
3402 memcg->use_hierarchy = val;
3411 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3415 if (mem_cgroup_is_root(memcg)) {
3416 val = memcg_page_state(memcg, MEMCG_CACHE) +
3417 memcg_page_state(memcg, MEMCG_RSS);
3419 val += memcg_page_state(memcg, MEMCG_SWAP);
3422 val = page_counter_read(&memcg->memory);
3424 val = page_counter_read(&memcg->memsw);
3437 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3440 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3441 struct page_counter *counter;
3443 switch (MEMFILE_TYPE(cft->private)) {
3445 counter = &memcg->memory;
3448 counter = &memcg->memsw;
3451 counter = &memcg->kmem;
3454 counter = &memcg->tcpmem;
3460 switch (MEMFILE_ATTR(cft->private)) {
3462 if (counter == &memcg->memory)
3463 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3464 if (counter == &memcg->memsw)
3465 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3466 return (u64)page_counter_read(counter) * PAGE_SIZE;
3468 return (u64)counter->max * PAGE_SIZE;
3470 return (u64)counter->watermark * PAGE_SIZE;
3472 return counter->failcnt;
3473 case RES_SOFT_LIMIT:
3474 return (u64)memcg->soft_limit * PAGE_SIZE;
3480 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3482 unsigned long stat[MEMCG_NR_STAT] = {0};
3483 struct mem_cgroup *mi;
3486 for_each_online_cpu(cpu)
3487 for (i = 0; i < MEMCG_NR_STAT; i++)
3488 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3490 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3491 for (i = 0; i < MEMCG_NR_STAT; i++)
3492 atomic_long_add(stat[i], &mi->vmstats[i]);
3494 for_each_node(node) {
3495 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3496 struct mem_cgroup_per_node *pi;
3498 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3501 for_each_online_cpu(cpu)
3502 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3504 pn->lruvec_stat_cpu->count[i], cpu);
3506 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3507 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3508 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3512 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3514 unsigned long events[NR_VM_EVENT_ITEMS];
3515 struct mem_cgroup *mi;
3518 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3521 for_each_online_cpu(cpu)
3522 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3523 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3526 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3527 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3528 atomic_long_add(events[i], &mi->vmevents[i]);
3531 #ifdef CONFIG_MEMCG_KMEM
3532 static int memcg_online_kmem(struct mem_cgroup *memcg)
3536 if (cgroup_memory_nokmem)
3539 BUG_ON(memcg->kmemcg_id >= 0);
3540 BUG_ON(memcg->kmem_state);
3542 memcg_id = memcg_alloc_cache_id();
3546 static_branch_inc(&memcg_kmem_enabled_key);
3548 * A memory cgroup is considered kmem-online as soon as it gets
3549 * kmemcg_id. Setting the id after enabling static branching will
3550 * guarantee no one starts accounting before all call sites are
3553 memcg->kmemcg_id = memcg_id;
3554 memcg->kmem_state = KMEM_ONLINE;
3555 INIT_LIST_HEAD(&memcg->kmem_caches);
3560 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3562 struct cgroup_subsys_state *css;
3563 struct mem_cgroup *parent, *child;
3566 if (memcg->kmem_state != KMEM_ONLINE)
3569 * Clear the online state before clearing memcg_caches array
3570 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3571 * guarantees that no cache will be created for this cgroup
3572 * after we are done (see memcg_create_kmem_cache()).
3574 memcg->kmem_state = KMEM_ALLOCATED;
3576 parent = parent_mem_cgroup(memcg);
3578 parent = root_mem_cgroup;
3581 * Deactivate and reparent kmem_caches.
3583 memcg_deactivate_kmem_caches(memcg, parent);
3585 kmemcg_id = memcg->kmemcg_id;
3586 BUG_ON(kmemcg_id < 0);
3589 * Change kmemcg_id of this cgroup and all its descendants to the
3590 * parent's id, and then move all entries from this cgroup's list_lrus
3591 * to ones of the parent. After we have finished, all list_lrus
3592 * corresponding to this cgroup are guaranteed to remain empty. The
3593 * ordering is imposed by list_lru_node->lock taken by
3594 * memcg_drain_all_list_lrus().
3596 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3597 css_for_each_descendant_pre(css, &memcg->css) {
3598 child = mem_cgroup_from_css(css);
3599 BUG_ON(child->kmemcg_id != kmemcg_id);
3600 child->kmemcg_id = parent->kmemcg_id;
3601 if (!memcg->use_hierarchy)
3606 memcg_drain_all_list_lrus(kmemcg_id, parent);
3608 memcg_free_cache_id(kmemcg_id);
3611 static void memcg_free_kmem(struct mem_cgroup *memcg)
3613 /* css_alloc() failed, offlining didn't happen */
3614 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3615 memcg_offline_kmem(memcg);
3617 if (memcg->kmem_state == KMEM_ALLOCATED) {
3618 WARN_ON(!list_empty(&memcg->kmem_caches));
3619 static_branch_dec(&memcg_kmem_enabled_key);
3623 static int memcg_online_kmem(struct mem_cgroup *memcg)
3627 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3630 static void memcg_free_kmem(struct mem_cgroup *memcg)
3633 #endif /* CONFIG_MEMCG_KMEM */
3635 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3640 mutex_lock(&memcg_max_mutex);
3641 ret = page_counter_set_max(&memcg->kmem, max);
3642 mutex_unlock(&memcg_max_mutex);
3646 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3650 mutex_lock(&memcg_max_mutex);
3652 ret = page_counter_set_max(&memcg->tcpmem, max);
3656 if (!memcg->tcpmem_active) {
3658 * The active flag needs to be written after the static_key
3659 * update. This is what guarantees that the socket activation
3660 * function is the last one to run. See mem_cgroup_sk_alloc()
3661 * for details, and note that we don't mark any socket as
3662 * belonging to this memcg until that flag is up.
3664 * We need to do this, because static_keys will span multiple
3665 * sites, but we can't control their order. If we mark a socket
3666 * as accounted, but the accounting functions are not patched in
3667 * yet, we'll lose accounting.
3669 * We never race with the readers in mem_cgroup_sk_alloc(),
3670 * because when this value change, the code to process it is not
3673 static_branch_inc(&memcg_sockets_enabled_key);
3674 memcg->tcpmem_active = true;
3677 mutex_unlock(&memcg_max_mutex);
3682 * The user of this function is...
3685 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3686 char *buf, size_t nbytes, loff_t off)
3688 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3689 unsigned long nr_pages;
3692 buf = strstrip(buf);
3693 ret = page_counter_memparse(buf, "-1", &nr_pages);
3697 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3699 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3703 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3705 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3708 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3711 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3712 "Please report your usecase to linux-mm@kvack.org if you "
3713 "depend on this functionality.\n");
3714 ret = memcg_update_kmem_max(memcg, nr_pages);
3717 ret = memcg_update_tcp_max(memcg, nr_pages);
3721 case RES_SOFT_LIMIT:
3722 memcg->soft_limit = nr_pages;
3726 return ret ?: nbytes;
3729 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3730 size_t nbytes, loff_t off)
3732 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3733 struct page_counter *counter;
3735 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3737 counter = &memcg->memory;
3740 counter = &memcg->memsw;
3743 counter = &memcg->kmem;
3746 counter = &memcg->tcpmem;
3752 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3754 page_counter_reset_watermark(counter);
3757 counter->failcnt = 0;
3766 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3769 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3773 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3774 struct cftype *cft, u64 val)
3776 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3778 if (val & ~MOVE_MASK)
3782 * No kind of locking is needed in here, because ->can_attach() will
3783 * check this value once in the beginning of the process, and then carry
3784 * on with stale data. This means that changes to this value will only
3785 * affect task migrations starting after the change.
3787 memcg->move_charge_at_immigrate = val;
3791 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3792 struct cftype *cft, u64 val)
3800 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3801 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3802 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3804 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3805 int nid, unsigned int lru_mask)
3807 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
3808 unsigned long nr = 0;
3811 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3814 if (!(BIT(lru) & lru_mask))
3816 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3821 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3822 unsigned int lru_mask)
3824 unsigned long nr = 0;
3828 if (!(BIT(lru) & lru_mask))
3830 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3835 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3839 unsigned int lru_mask;
3842 static const struct numa_stat stats[] = {
3843 { "total", LRU_ALL },
3844 { "file", LRU_ALL_FILE },
3845 { "anon", LRU_ALL_ANON },
3846 { "unevictable", BIT(LRU_UNEVICTABLE) },
3848 const struct numa_stat *stat;
3851 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3853 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3854 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3855 seq_printf(m, "%s=%lu", stat->name, nr);
3856 for_each_node_state(nid, N_MEMORY) {
3857 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3859 seq_printf(m, " N%d=%lu", nid, nr);
3864 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3865 struct mem_cgroup *iter;
3868 for_each_mem_cgroup_tree(iter, memcg)
3869 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3870 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3871 for_each_node_state(nid, N_MEMORY) {
3873 for_each_mem_cgroup_tree(iter, memcg)
3874 nr += mem_cgroup_node_nr_lru_pages(
3875 iter, nid, stat->lru_mask);
3876 seq_printf(m, " N%d=%lu", nid, nr);
3883 #endif /* CONFIG_NUMA */
3885 static const unsigned int memcg1_stats[] = {
3896 static const char *const memcg1_stat_names[] = {
3907 /* Universal VM events cgroup1 shows, original sort order */
3908 static const unsigned int memcg1_events[] = {
3915 static const char *const memcg1_event_names[] = {
3922 static int memcg_stat_show(struct seq_file *m, void *v)
3924 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3925 unsigned long memory, memsw;
3926 struct mem_cgroup *mi;
3929 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3930 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3932 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3933 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3935 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3936 memcg_page_state_local(memcg, memcg1_stats[i]) *
3940 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3941 seq_printf(m, "%s %lu\n", memcg1_event_names[i],
3942 memcg_events_local(memcg, memcg1_events[i]));
3944 for (i = 0; i < NR_LRU_LISTS; i++)
3945 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3946 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3949 /* Hierarchical information */
3950 memory = memsw = PAGE_COUNTER_MAX;
3951 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3952 memory = min(memory, mi->memory.max);
3953 memsw = min(memsw, mi->memsw.max);
3955 seq_printf(m, "hierarchical_memory_limit %llu\n",
3956 (u64)memory * PAGE_SIZE);
3957 if (do_memsw_account())
3958 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3959 (u64)memsw * PAGE_SIZE);
3961 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3962 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3964 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3965 (u64)memcg_page_state(memcg, memcg1_stats[i]) *
3969 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3970 seq_printf(m, "total_%s %llu\n", memcg1_event_names[i],
3971 (u64)memcg_events(memcg, memcg1_events[i]));
3973 for (i = 0; i < NR_LRU_LISTS; i++)
3974 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i],
3975 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
3978 #ifdef CONFIG_DEBUG_VM
3981 struct mem_cgroup_per_node *mz;
3982 struct zone_reclaim_stat *rstat;
3983 unsigned long recent_rotated[2] = {0, 0};
3984 unsigned long recent_scanned[2] = {0, 0};
3986 for_each_online_pgdat(pgdat) {
3987 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3988 rstat = &mz->lruvec.reclaim_stat;
3990 recent_rotated[0] += rstat->recent_rotated[0];
3991 recent_rotated[1] += rstat->recent_rotated[1];
3992 recent_scanned[0] += rstat->recent_scanned[0];
3993 recent_scanned[1] += rstat->recent_scanned[1];
3995 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3996 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3997 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3998 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
4005 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4008 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4010 return mem_cgroup_swappiness(memcg);
4013 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4014 struct cftype *cft, u64 val)
4016 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4022 memcg->swappiness = val;
4024 vm_swappiness = val;
4029 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4031 struct mem_cgroup_threshold_ary *t;
4032 unsigned long usage;
4037 t = rcu_dereference(memcg->thresholds.primary);
4039 t = rcu_dereference(memcg->memsw_thresholds.primary);
4044 usage = mem_cgroup_usage(memcg, swap);
4047 * current_threshold points to threshold just below or equal to usage.
4048 * If it's not true, a threshold was crossed after last
4049 * call of __mem_cgroup_threshold().
4051 i = t->current_threshold;
4054 * Iterate backward over array of thresholds starting from
4055 * current_threshold and check if a threshold is crossed.
4056 * If none of thresholds below usage is crossed, we read
4057 * only one element of the array here.
4059 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4060 eventfd_signal(t->entries[i].eventfd, 1);
4062 /* i = current_threshold + 1 */
4066 * Iterate forward over array of thresholds starting from
4067 * current_threshold+1 and check if a threshold is crossed.
4068 * If none of thresholds above usage is crossed, we read
4069 * only one element of the array here.
4071 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4072 eventfd_signal(t->entries[i].eventfd, 1);
4074 /* Update current_threshold */
4075 t->current_threshold = i - 1;
4080 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4083 __mem_cgroup_threshold(memcg, false);
4084 if (do_memsw_account())
4085 __mem_cgroup_threshold(memcg, true);
4087 memcg = parent_mem_cgroup(memcg);
4091 static int compare_thresholds(const void *a, const void *b)
4093 const struct mem_cgroup_threshold *_a = a;
4094 const struct mem_cgroup_threshold *_b = b;
4096 if (_a->threshold > _b->threshold)
4099 if (_a->threshold < _b->threshold)
4105 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4107 struct mem_cgroup_eventfd_list *ev;
4109 spin_lock(&memcg_oom_lock);
4111 list_for_each_entry(ev, &memcg->oom_notify, list)
4112 eventfd_signal(ev->eventfd, 1);
4114 spin_unlock(&memcg_oom_lock);
4118 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4120 struct mem_cgroup *iter;
4122 for_each_mem_cgroup_tree(iter, memcg)
4123 mem_cgroup_oom_notify_cb(iter);
4126 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4127 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4129 struct mem_cgroup_thresholds *thresholds;
4130 struct mem_cgroup_threshold_ary *new;
4131 unsigned long threshold;
4132 unsigned long usage;
4135 ret = page_counter_memparse(args, "-1", &threshold);
4139 mutex_lock(&memcg->thresholds_lock);
4142 thresholds = &memcg->thresholds;
4143 usage = mem_cgroup_usage(memcg, false);
4144 } else if (type == _MEMSWAP) {
4145 thresholds = &memcg->memsw_thresholds;
4146 usage = mem_cgroup_usage(memcg, true);
4150 /* Check if a threshold crossed before adding a new one */
4151 if (thresholds->primary)
4152 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4154 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4156 /* Allocate memory for new array of thresholds */
4157 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4164 /* Copy thresholds (if any) to new array */
4165 if (thresholds->primary) {
4166 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4167 sizeof(struct mem_cgroup_threshold));
4170 /* Add new threshold */
4171 new->entries[size - 1].eventfd = eventfd;
4172 new->entries[size - 1].threshold = threshold;
4174 /* Sort thresholds. Registering of new threshold isn't time-critical */
4175 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4176 compare_thresholds, NULL);
4178 /* Find current threshold */
4179 new->current_threshold = -1;
4180 for (i = 0; i < size; i++) {
4181 if (new->entries[i].threshold <= usage) {
4183 * new->current_threshold will not be used until
4184 * rcu_assign_pointer(), so it's safe to increment
4187 ++new->current_threshold;
4192 /* Free old spare buffer and save old primary buffer as spare */
4193 kfree(thresholds->spare);
4194 thresholds->spare = thresholds->primary;
4196 rcu_assign_pointer(thresholds->primary, new);
4198 /* To be sure that nobody uses thresholds */
4202 mutex_unlock(&memcg->thresholds_lock);
4207 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4208 struct eventfd_ctx *eventfd, const char *args)
4210 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4213 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4214 struct eventfd_ctx *eventfd, const char *args)
4216 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4219 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4220 struct eventfd_ctx *eventfd, enum res_type type)
4222 struct mem_cgroup_thresholds *thresholds;
4223 struct mem_cgroup_threshold_ary *new;
4224 unsigned long usage;
4225 int i, j, size, entries;
4227 mutex_lock(&memcg->thresholds_lock);
4230 thresholds = &memcg->thresholds;
4231 usage = mem_cgroup_usage(memcg, false);
4232 } else if (type == _MEMSWAP) {
4233 thresholds = &memcg->memsw_thresholds;
4234 usage = mem_cgroup_usage(memcg, true);
4238 if (!thresholds->primary)
4241 /* Check if a threshold crossed before removing */
4242 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4244 /* Calculate new number of threshold */
4246 for (i = 0; i < thresholds->primary->size; i++) {
4247 if (thresholds->primary->entries[i].eventfd != eventfd)
4253 new = thresholds->spare;
4255 /* If no items related to eventfd have been cleared, nothing to do */
4259 /* Set thresholds array to NULL if we don't have thresholds */
4268 /* Copy thresholds and find current threshold */
4269 new->current_threshold = -1;
4270 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4271 if (thresholds->primary->entries[i].eventfd == eventfd)
4274 new->entries[j] = thresholds->primary->entries[i];
4275 if (new->entries[j].threshold <= usage) {
4277 * new->current_threshold will not be used
4278 * until rcu_assign_pointer(), so it's safe to increment
4281 ++new->current_threshold;
4287 /* Swap primary and spare array */
4288 thresholds->spare = thresholds->primary;
4290 rcu_assign_pointer(thresholds->primary, new);
4292 /* To be sure that nobody uses thresholds */
4295 /* If all events are unregistered, free the spare array */
4297 kfree(thresholds->spare);
4298 thresholds->spare = NULL;
4301 mutex_unlock(&memcg->thresholds_lock);
4304 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4305 struct eventfd_ctx *eventfd)
4307 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4310 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4311 struct eventfd_ctx *eventfd)
4313 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4316 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4317 struct eventfd_ctx *eventfd, const char *args)
4319 struct mem_cgroup_eventfd_list *event;
4321 event = kmalloc(sizeof(*event), GFP_KERNEL);
4325 spin_lock(&memcg_oom_lock);
4327 event->eventfd = eventfd;
4328 list_add(&event->list, &memcg->oom_notify);
4330 /* already in OOM ? */
4331 if (memcg->under_oom)
4332 eventfd_signal(eventfd, 1);
4333 spin_unlock(&memcg_oom_lock);
4338 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4339 struct eventfd_ctx *eventfd)
4341 struct mem_cgroup_eventfd_list *ev, *tmp;
4343 spin_lock(&memcg_oom_lock);
4345 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4346 if (ev->eventfd == eventfd) {
4347 list_del(&ev->list);
4352 spin_unlock(&memcg_oom_lock);
4355 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4357 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4359 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4360 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4361 seq_printf(sf, "oom_kill %lu\n",
4362 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4366 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4367 struct cftype *cft, u64 val)
4369 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4371 /* cannot set to root cgroup and only 0 and 1 are allowed */
4372 if (!css->parent || !((val == 0) || (val == 1)))
4375 memcg->oom_kill_disable = val;
4377 memcg_oom_recover(memcg);
4382 #ifdef CONFIG_CGROUP_WRITEBACK
4384 #include <trace/events/writeback.h>
4386 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4388 return wb_domain_init(&memcg->cgwb_domain, gfp);
4391 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4393 wb_domain_exit(&memcg->cgwb_domain);
4396 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4398 wb_domain_size_changed(&memcg->cgwb_domain);
4401 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4403 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4405 if (!memcg->css.parent)
4408 return &memcg->cgwb_domain;
4412 * idx can be of type enum memcg_stat_item or node_stat_item.
4413 * Keep in sync with memcg_exact_page().
4415 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4417 long x = atomic_long_read(&memcg->vmstats[idx]);
4420 for_each_online_cpu(cpu)
4421 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4428 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4429 * @wb: bdi_writeback in question
4430 * @pfilepages: out parameter for number of file pages
4431 * @pheadroom: out parameter for number of allocatable pages according to memcg
4432 * @pdirty: out parameter for number of dirty pages
4433 * @pwriteback: out parameter for number of pages under writeback
4435 * Determine the numbers of file, headroom, dirty, and writeback pages in
4436 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4437 * is a bit more involved.
4439 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4440 * headroom is calculated as the lowest headroom of itself and the
4441 * ancestors. Note that this doesn't consider the actual amount of
4442 * available memory in the system. The caller should further cap
4443 * *@pheadroom accordingly.
4445 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4446 unsigned long *pheadroom, unsigned long *pdirty,
4447 unsigned long *pwriteback)
4449 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4450 struct mem_cgroup *parent;
4452 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4454 /* this should eventually include NR_UNSTABLE_NFS */
4455 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4456 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4457 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4458 *pheadroom = PAGE_COUNTER_MAX;
4460 while ((parent = parent_mem_cgroup(memcg))) {
4461 unsigned long ceiling = min(memcg->memory.max, memcg->high);
4462 unsigned long used = page_counter_read(&memcg->memory);
4464 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4470 * Foreign dirty flushing
4472 * There's an inherent mismatch between memcg and writeback. The former
4473 * trackes ownership per-page while the latter per-inode. This was a
4474 * deliberate design decision because honoring per-page ownership in the
4475 * writeback path is complicated, may lead to higher CPU and IO overheads
4476 * and deemed unnecessary given that write-sharing an inode across
4477 * different cgroups isn't a common use-case.
4479 * Combined with inode majority-writer ownership switching, this works well
4480 * enough in most cases but there are some pathological cases. For
4481 * example, let's say there are two cgroups A and B which keep writing to
4482 * different but confined parts of the same inode. B owns the inode and
4483 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4484 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4485 * triggering background writeback. A will be slowed down without a way to
4486 * make writeback of the dirty pages happen.
4488 * Conditions like the above can lead to a cgroup getting repatedly and
4489 * severely throttled after making some progress after each
4490 * dirty_expire_interval while the underyling IO device is almost
4493 * Solving this problem completely requires matching the ownership tracking
4494 * granularities between memcg and writeback in either direction. However,
4495 * the more egregious behaviors can be avoided by simply remembering the
4496 * most recent foreign dirtying events and initiating remote flushes on
4497 * them when local writeback isn't enough to keep the memory clean enough.
4499 * The following two functions implement such mechanism. When a foreign
4500 * page - a page whose memcg and writeback ownerships don't match - is
4501 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4502 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4503 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4504 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4505 * foreign bdi_writebacks which haven't expired. Both the numbers of
4506 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4507 * limited to MEMCG_CGWB_FRN_CNT.
4509 * The mechanism only remembers IDs and doesn't hold any object references.
4510 * As being wrong occasionally doesn't matter, updates and accesses to the
4511 * records are lockless and racy.
4513 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4514 struct bdi_writeback *wb)
4516 struct mem_cgroup *memcg = page->mem_cgroup;
4517 struct memcg_cgwb_frn *frn;
4518 u64 now = get_jiffies_64();
4519 u64 oldest_at = now;
4523 trace_track_foreign_dirty(page, wb);
4526 * Pick the slot to use. If there is already a slot for @wb, keep
4527 * using it. If not replace the oldest one which isn't being
4530 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4531 frn = &memcg->cgwb_frn[i];
4532 if (frn->bdi_id == wb->bdi->id &&
4533 frn->memcg_id == wb->memcg_css->id)
4535 if (time_before64(frn->at, oldest_at) &&
4536 atomic_read(&frn->done.cnt) == 1) {
4538 oldest_at = frn->at;
4542 if (i < MEMCG_CGWB_FRN_CNT) {
4544 * Re-using an existing one. Update timestamp lazily to
4545 * avoid making the cacheline hot. We want them to be
4546 * reasonably up-to-date and significantly shorter than
4547 * dirty_expire_interval as that's what expires the record.
4548 * Use the shorter of 1s and dirty_expire_interval / 8.
4550 unsigned long update_intv =
4551 min_t(unsigned long, HZ,
4552 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4554 if (time_before64(frn->at, now - update_intv))
4556 } else if (oldest >= 0) {
4557 /* replace the oldest free one */
4558 frn = &memcg->cgwb_frn[oldest];
4559 frn->bdi_id = wb->bdi->id;
4560 frn->memcg_id = wb->memcg_css->id;
4565 /* issue foreign writeback flushes for recorded foreign dirtying events */
4566 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4568 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4569 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4570 u64 now = jiffies_64;
4573 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4574 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4577 * If the record is older than dirty_expire_interval,
4578 * writeback on it has already started. No need to kick it
4579 * off again. Also, don't start a new one if there's
4580 * already one in flight.
4582 if (time_after64(frn->at, now - intv) &&
4583 atomic_read(&frn->done.cnt) == 1) {
4585 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4586 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4587 WB_REASON_FOREIGN_FLUSH,
4593 #else /* CONFIG_CGROUP_WRITEBACK */
4595 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4600 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4604 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4608 #endif /* CONFIG_CGROUP_WRITEBACK */
4611 * DO NOT USE IN NEW FILES.
4613 * "cgroup.event_control" implementation.
4615 * This is way over-engineered. It tries to support fully configurable
4616 * events for each user. Such level of flexibility is completely
4617 * unnecessary especially in the light of the planned unified hierarchy.
4619 * Please deprecate this and replace with something simpler if at all
4624 * Unregister event and free resources.
4626 * Gets called from workqueue.
4628 static void memcg_event_remove(struct work_struct *work)
4630 struct mem_cgroup_event *event =
4631 container_of(work, struct mem_cgroup_event, remove);
4632 struct mem_cgroup *memcg = event->memcg;
4634 remove_wait_queue(event->wqh, &event->wait);
4636 event->unregister_event(memcg, event->eventfd);
4638 /* Notify userspace the event is going away. */
4639 eventfd_signal(event->eventfd, 1);
4641 eventfd_ctx_put(event->eventfd);
4643 css_put(&memcg->css);
4647 * Gets called on EPOLLHUP on eventfd when user closes it.
4649 * Called with wqh->lock held and interrupts disabled.
4651 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4652 int sync, void *key)
4654 struct mem_cgroup_event *event =
4655 container_of(wait, struct mem_cgroup_event, wait);
4656 struct mem_cgroup *memcg = event->memcg;
4657 __poll_t flags = key_to_poll(key);
4659 if (flags & EPOLLHUP) {
4661 * If the event has been detached at cgroup removal, we
4662 * can simply return knowing the other side will cleanup
4665 * We can't race against event freeing since the other
4666 * side will require wqh->lock via remove_wait_queue(),
4669 spin_lock(&memcg->event_list_lock);
4670 if (!list_empty(&event->list)) {
4671 list_del_init(&event->list);
4673 * We are in atomic context, but cgroup_event_remove()
4674 * may sleep, so we have to call it in workqueue.
4676 schedule_work(&event->remove);
4678 spin_unlock(&memcg->event_list_lock);
4684 static void memcg_event_ptable_queue_proc(struct file *file,
4685 wait_queue_head_t *wqh, poll_table *pt)
4687 struct mem_cgroup_event *event =
4688 container_of(pt, struct mem_cgroup_event, pt);
4691 add_wait_queue(wqh, &event->wait);
4695 * DO NOT USE IN NEW FILES.
4697 * Parse input and register new cgroup event handler.
4699 * Input must be in format '<event_fd> <control_fd> <args>'.
4700 * Interpretation of args is defined by control file implementation.
4702 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4703 char *buf, size_t nbytes, loff_t off)
4705 struct cgroup_subsys_state *css = of_css(of);
4706 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4707 struct mem_cgroup_event *event;
4708 struct cgroup_subsys_state *cfile_css;
4709 unsigned int efd, cfd;
4716 buf = strstrip(buf);
4718 efd = simple_strtoul(buf, &endp, 10);
4723 cfd = simple_strtoul(buf, &endp, 10);
4724 if ((*endp != ' ') && (*endp != '\0'))
4728 event = kzalloc(sizeof(*event), GFP_KERNEL);
4732 event->memcg = memcg;
4733 INIT_LIST_HEAD(&event->list);
4734 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4735 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4736 INIT_WORK(&event->remove, memcg_event_remove);
4744 event->eventfd = eventfd_ctx_fileget(efile.file);
4745 if (IS_ERR(event->eventfd)) {
4746 ret = PTR_ERR(event->eventfd);
4753 goto out_put_eventfd;
4756 /* the process need read permission on control file */
4757 /* AV: shouldn't we check that it's been opened for read instead? */
4758 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4763 * Determine the event callbacks and set them in @event. This used
4764 * to be done via struct cftype but cgroup core no longer knows
4765 * about these events. The following is crude but the whole thing
4766 * is for compatibility anyway.
4768 * DO NOT ADD NEW FILES.
4770 name = cfile.file->f_path.dentry->d_name.name;
4772 if (!strcmp(name, "memory.usage_in_bytes")) {
4773 event->register_event = mem_cgroup_usage_register_event;
4774 event->unregister_event = mem_cgroup_usage_unregister_event;
4775 } else if (!strcmp(name, "memory.oom_control")) {
4776 event->register_event = mem_cgroup_oom_register_event;
4777 event->unregister_event = mem_cgroup_oom_unregister_event;
4778 } else if (!strcmp(name, "memory.pressure_level")) {
4779 event->register_event = vmpressure_register_event;
4780 event->unregister_event = vmpressure_unregister_event;
4781 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4782 event->register_event = memsw_cgroup_usage_register_event;
4783 event->unregister_event = memsw_cgroup_usage_unregister_event;
4790 * Verify @cfile should belong to @css. Also, remaining events are
4791 * automatically removed on cgroup destruction but the removal is
4792 * asynchronous, so take an extra ref on @css.
4794 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4795 &memory_cgrp_subsys);
4797 if (IS_ERR(cfile_css))
4799 if (cfile_css != css) {
4804 ret = event->register_event(memcg, event->eventfd, buf);
4808 vfs_poll(efile.file, &event->pt);
4810 spin_lock(&memcg->event_list_lock);
4811 list_add(&event->list, &memcg->event_list);
4812 spin_unlock(&memcg->event_list_lock);
4824 eventfd_ctx_put(event->eventfd);
4833 static struct cftype mem_cgroup_legacy_files[] = {
4835 .name = "usage_in_bytes",
4836 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4837 .read_u64 = mem_cgroup_read_u64,
4840 .name = "max_usage_in_bytes",
4841 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4842 .write = mem_cgroup_reset,
4843 .read_u64 = mem_cgroup_read_u64,
4846 .name = "limit_in_bytes",
4847 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4848 .write = mem_cgroup_write,
4849 .read_u64 = mem_cgroup_read_u64,
4852 .name = "soft_limit_in_bytes",
4853 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4854 .write = mem_cgroup_write,
4855 .read_u64 = mem_cgroup_read_u64,
4859 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4860 .write = mem_cgroup_reset,
4861 .read_u64 = mem_cgroup_read_u64,
4865 .seq_show = memcg_stat_show,
4868 .name = "force_empty",
4869 .write = mem_cgroup_force_empty_write,
4872 .name = "use_hierarchy",
4873 .write_u64 = mem_cgroup_hierarchy_write,
4874 .read_u64 = mem_cgroup_hierarchy_read,
4877 .name = "cgroup.event_control", /* XXX: for compat */
4878 .write = memcg_write_event_control,
4879 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4882 .name = "swappiness",
4883 .read_u64 = mem_cgroup_swappiness_read,
4884 .write_u64 = mem_cgroup_swappiness_write,
4887 .name = "move_charge_at_immigrate",
4888 .read_u64 = mem_cgroup_move_charge_read,
4889 .write_u64 = mem_cgroup_move_charge_write,
4892 .name = "oom_control",
4893 .seq_show = mem_cgroup_oom_control_read,
4894 .write_u64 = mem_cgroup_oom_control_write,
4895 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4898 .name = "pressure_level",
4902 .name = "numa_stat",
4903 .seq_show = memcg_numa_stat_show,
4907 .name = "kmem.limit_in_bytes",
4908 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4909 .write = mem_cgroup_write,
4910 .read_u64 = mem_cgroup_read_u64,
4913 .name = "kmem.usage_in_bytes",
4914 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4915 .read_u64 = mem_cgroup_read_u64,
4918 .name = "kmem.failcnt",
4919 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4920 .write = mem_cgroup_reset,
4921 .read_u64 = mem_cgroup_read_u64,
4924 .name = "kmem.max_usage_in_bytes",
4925 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4926 .write = mem_cgroup_reset,
4927 .read_u64 = mem_cgroup_read_u64,
4929 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4931 .name = "kmem.slabinfo",
4932 .seq_start = memcg_slab_start,
4933 .seq_next = memcg_slab_next,
4934 .seq_stop = memcg_slab_stop,
4935 .seq_show = memcg_slab_show,
4939 .name = "kmem.tcp.limit_in_bytes",
4940 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4941 .write = mem_cgroup_write,
4942 .read_u64 = mem_cgroup_read_u64,
4945 .name = "kmem.tcp.usage_in_bytes",
4946 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4947 .read_u64 = mem_cgroup_read_u64,
4950 .name = "kmem.tcp.failcnt",
4951 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4952 .write = mem_cgroup_reset,
4953 .read_u64 = mem_cgroup_read_u64,
4956 .name = "kmem.tcp.max_usage_in_bytes",
4957 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4958 .write = mem_cgroup_reset,
4959 .read_u64 = mem_cgroup_read_u64,
4961 { }, /* terminate */
4965 * Private memory cgroup IDR
4967 * Swap-out records and page cache shadow entries need to store memcg
4968 * references in constrained space, so we maintain an ID space that is
4969 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4970 * memory-controlled cgroups to 64k.
4972 * However, there usually are many references to the oflline CSS after
4973 * the cgroup has been destroyed, such as page cache or reclaimable
4974 * slab objects, that don't need to hang on to the ID. We want to keep
4975 * those dead CSS from occupying IDs, or we might quickly exhaust the
4976 * relatively small ID space and prevent the creation of new cgroups
4977 * even when there are much fewer than 64k cgroups - possibly none.
4979 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4980 * be freed and recycled when it's no longer needed, which is usually
4981 * when the CSS is offlined.
4983 * The only exception to that are records of swapped out tmpfs/shmem
4984 * pages that need to be attributed to live ancestors on swapin. But
4985 * those references are manageable from userspace.
4988 static DEFINE_IDR(mem_cgroup_idr);
4990 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4992 if (memcg->id.id > 0) {
4993 idr_remove(&mem_cgroup_idr, memcg->id.id);
4998 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
5000 refcount_add(n, &memcg->id.ref);
5003 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5005 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5006 mem_cgroup_id_remove(memcg);
5008 /* Memcg ID pins CSS */
5009 css_put(&memcg->css);
5013 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5015 mem_cgroup_id_put_many(memcg, 1);
5019 * mem_cgroup_from_id - look up a memcg from a memcg id
5020 * @id: the memcg id to look up
5022 * Caller must hold rcu_read_lock().
5024 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5026 WARN_ON_ONCE(!rcu_read_lock_held());
5027 return idr_find(&mem_cgroup_idr, id);
5030 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5032 struct mem_cgroup_per_node *pn;
5035 * This routine is called against possible nodes.
5036 * But it's BUG to call kmalloc() against offline node.
5038 * TODO: this routine can waste much memory for nodes which will
5039 * never be onlined. It's better to use memory hotplug callback
5042 if (!node_state(node, N_NORMAL_MEMORY))
5044 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5048 pn->lruvec_stat_local = alloc_percpu(struct lruvec_stat);
5049 if (!pn->lruvec_stat_local) {
5054 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
5055 if (!pn->lruvec_stat_cpu) {
5056 free_percpu(pn->lruvec_stat_local);
5061 lruvec_init(&pn->lruvec);
5062 pn->usage_in_excess = 0;
5063 pn->on_tree = false;
5066 memcg->nodeinfo[node] = pn;
5070 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5072 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5077 free_percpu(pn->lruvec_stat_cpu);
5078 free_percpu(pn->lruvec_stat_local);
5082 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5087 free_mem_cgroup_per_node_info(memcg, node);
5088 free_percpu(memcg->vmstats_percpu);
5089 free_percpu(memcg->vmstats_local);
5093 static void mem_cgroup_free(struct mem_cgroup *memcg)
5095 memcg_wb_domain_exit(memcg);
5097 * Flush percpu vmstats and vmevents to guarantee the value correctness
5098 * on parent's and all ancestor levels.
5100 memcg_flush_percpu_vmstats(memcg);
5101 memcg_flush_percpu_vmevents(memcg);
5102 __mem_cgroup_free(memcg);
5105 static struct mem_cgroup *mem_cgroup_alloc(void)
5107 struct mem_cgroup *memcg;
5110 int __maybe_unused i;
5111 long error = -ENOMEM;
5113 size = sizeof(struct mem_cgroup);
5114 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5116 memcg = kzalloc(size, GFP_KERNEL);
5118 return ERR_PTR(error);
5120 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5121 1, MEM_CGROUP_ID_MAX,
5123 if (memcg->id.id < 0) {
5124 error = memcg->id.id;
5128 memcg->vmstats_local = alloc_percpu(struct memcg_vmstats_percpu);
5129 if (!memcg->vmstats_local)
5132 memcg->vmstats_percpu = alloc_percpu(struct memcg_vmstats_percpu);
5133 if (!memcg->vmstats_percpu)
5137 if (alloc_mem_cgroup_per_node_info(memcg, node))
5140 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5143 INIT_WORK(&memcg->high_work, high_work_func);
5144 memcg->last_scanned_node = MAX_NUMNODES;
5145 INIT_LIST_HEAD(&memcg->oom_notify);
5146 mutex_init(&memcg->thresholds_lock);
5147 spin_lock_init(&memcg->move_lock);
5148 vmpressure_init(&memcg->vmpressure);
5149 INIT_LIST_HEAD(&memcg->event_list);
5150 spin_lock_init(&memcg->event_list_lock);
5151 memcg->socket_pressure = jiffies;
5152 #ifdef CONFIG_MEMCG_KMEM
5153 memcg->kmemcg_id = -1;
5155 #ifdef CONFIG_CGROUP_WRITEBACK
5156 INIT_LIST_HEAD(&memcg->cgwb_list);
5157 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5158 memcg->cgwb_frn[i].done =
5159 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5161 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5162 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5163 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5164 memcg->deferred_split_queue.split_queue_len = 0;
5166 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5169 mem_cgroup_id_remove(memcg);
5170 __mem_cgroup_free(memcg);
5171 return ERR_PTR(error);
5174 static struct cgroup_subsys_state * __ref
5175 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5177 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5178 struct mem_cgroup *memcg;
5179 long error = -ENOMEM;
5181 memcg = mem_cgroup_alloc();
5183 return ERR_CAST(memcg);
5185 memcg->high = PAGE_COUNTER_MAX;
5186 memcg->soft_limit = PAGE_COUNTER_MAX;
5188 memcg->swappiness = mem_cgroup_swappiness(parent);
5189 memcg->oom_kill_disable = parent->oom_kill_disable;
5191 if (parent && parent->use_hierarchy) {
5192 memcg->use_hierarchy = true;
5193 page_counter_init(&memcg->memory, &parent->memory);
5194 page_counter_init(&memcg->swap, &parent->swap);
5195 page_counter_init(&memcg->memsw, &parent->memsw);
5196 page_counter_init(&memcg->kmem, &parent->kmem);
5197 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5199 page_counter_init(&memcg->memory, NULL);
5200 page_counter_init(&memcg->swap, NULL);
5201 page_counter_init(&memcg->memsw, NULL);
5202 page_counter_init(&memcg->kmem, NULL);
5203 page_counter_init(&memcg->tcpmem, NULL);
5205 * Deeper hierachy with use_hierarchy == false doesn't make
5206 * much sense so let cgroup subsystem know about this
5207 * unfortunate state in our controller.
5209 if (parent != root_mem_cgroup)
5210 memory_cgrp_subsys.broken_hierarchy = true;
5213 /* The following stuff does not apply to the root */
5215 #ifdef CONFIG_MEMCG_KMEM
5216 INIT_LIST_HEAD(&memcg->kmem_caches);
5218 root_mem_cgroup = memcg;
5222 error = memcg_online_kmem(memcg);
5226 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5227 static_branch_inc(&memcg_sockets_enabled_key);
5231 mem_cgroup_id_remove(memcg);
5232 mem_cgroup_free(memcg);
5233 return ERR_PTR(error);
5236 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5238 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5241 * A memcg must be visible for memcg_expand_shrinker_maps()
5242 * by the time the maps are allocated. So, we allocate maps
5243 * here, when for_each_mem_cgroup() can't skip it.
5245 if (memcg_alloc_shrinker_maps(memcg)) {
5246 mem_cgroup_id_remove(memcg);
5250 /* Online state pins memcg ID, memcg ID pins CSS */
5251 refcount_set(&memcg->id.ref, 1);
5256 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5258 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5259 struct mem_cgroup_event *event, *tmp;
5262 * Unregister events and notify userspace.
5263 * Notify userspace about cgroup removing only after rmdir of cgroup
5264 * directory to avoid race between userspace and kernelspace.
5266 spin_lock(&memcg->event_list_lock);
5267 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5268 list_del_init(&event->list);
5269 schedule_work(&event->remove);
5271 spin_unlock(&memcg->event_list_lock);
5273 page_counter_set_min(&memcg->memory, 0);
5274 page_counter_set_low(&memcg->memory, 0);
5276 memcg_offline_kmem(memcg);
5277 wb_memcg_offline(memcg);
5279 drain_all_stock(memcg);
5281 mem_cgroup_id_put(memcg);
5284 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5286 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5288 invalidate_reclaim_iterators(memcg);
5291 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5293 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5294 int __maybe_unused i;
5296 #ifdef CONFIG_CGROUP_WRITEBACK
5297 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5298 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5300 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5301 static_branch_dec(&memcg_sockets_enabled_key);
5303 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5304 static_branch_dec(&memcg_sockets_enabled_key);
5306 vmpressure_cleanup(&memcg->vmpressure);
5307 cancel_work_sync(&memcg->high_work);
5308 mem_cgroup_remove_from_trees(memcg);
5309 memcg_free_shrinker_maps(memcg);
5310 memcg_free_kmem(memcg);
5311 mem_cgroup_free(memcg);
5315 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5316 * @css: the target css
5318 * Reset the states of the mem_cgroup associated with @css. This is
5319 * invoked when the userland requests disabling on the default hierarchy
5320 * but the memcg is pinned through dependency. The memcg should stop
5321 * applying policies and should revert to the vanilla state as it may be
5322 * made visible again.
5324 * The current implementation only resets the essential configurations.
5325 * This needs to be expanded to cover all the visible parts.
5327 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5329 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5331 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5332 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5333 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
5334 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5335 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5336 page_counter_set_min(&memcg->memory, 0);
5337 page_counter_set_low(&memcg->memory, 0);
5338 memcg->high = PAGE_COUNTER_MAX;
5339 memcg->soft_limit = PAGE_COUNTER_MAX;
5340 memcg_wb_domain_size_changed(memcg);
5344 /* Handlers for move charge at task migration. */
5345 static int mem_cgroup_do_precharge(unsigned long count)
5349 /* Try a single bulk charge without reclaim first, kswapd may wake */
5350 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5352 mc.precharge += count;
5356 /* Try charges one by one with reclaim, but do not retry */
5358 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5372 enum mc_target_type {
5379 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5380 unsigned long addr, pte_t ptent)
5382 struct page *page = vm_normal_page(vma, addr, ptent);
5384 if (!page || !page_mapped(page))
5386 if (PageAnon(page)) {
5387 if (!(mc.flags & MOVE_ANON))
5390 if (!(mc.flags & MOVE_FILE))
5393 if (!get_page_unless_zero(page))
5399 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5400 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5401 pte_t ptent, swp_entry_t *entry)
5403 struct page *page = NULL;
5404 swp_entry_t ent = pte_to_swp_entry(ptent);
5406 if (!(mc.flags & MOVE_ANON))
5410 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5411 * a device and because they are not accessible by CPU they are store
5412 * as special swap entry in the CPU page table.
5414 if (is_device_private_entry(ent)) {
5415 page = device_private_entry_to_page(ent);
5417 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5418 * a refcount of 1 when free (unlike normal page)
5420 if (!page_ref_add_unless(page, 1, 1))
5425 if (non_swap_entry(ent))
5429 * Because lookup_swap_cache() updates some statistics counter,
5430 * we call find_get_page() with swapper_space directly.
5432 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5433 if (do_memsw_account())
5434 entry->val = ent.val;
5439 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5440 pte_t ptent, swp_entry_t *entry)
5446 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5447 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5449 struct page *page = NULL;
5450 struct address_space *mapping;
5453 if (!vma->vm_file) /* anonymous vma */
5455 if (!(mc.flags & MOVE_FILE))
5458 mapping = vma->vm_file->f_mapping;
5459 pgoff = linear_page_index(vma, addr);
5461 /* page is moved even if it's not RSS of this task(page-faulted). */
5463 /* shmem/tmpfs may report page out on swap: account for that too. */
5464 if (shmem_mapping(mapping)) {
5465 page = find_get_entry(mapping, pgoff);
5466 if (xa_is_value(page)) {
5467 swp_entry_t swp = radix_to_swp_entry(page);
5468 if (do_memsw_account())
5470 page = find_get_page(swap_address_space(swp),
5474 page = find_get_page(mapping, pgoff);
5476 page = find_get_page(mapping, pgoff);
5482 * mem_cgroup_move_account - move account of the page
5484 * @compound: charge the page as compound or small page
5485 * @from: mem_cgroup which the page is moved from.
5486 * @to: mem_cgroup which the page is moved to. @from != @to.
5488 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5490 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5493 static int mem_cgroup_move_account(struct page *page,
5495 struct mem_cgroup *from,
5496 struct mem_cgroup *to)
5498 struct lruvec *from_vec, *to_vec;
5499 struct pglist_data *pgdat;
5500 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5504 VM_BUG_ON(from == to);
5505 VM_BUG_ON_PAGE(PageLRU(page), page);
5506 VM_BUG_ON(compound && !PageTransHuge(page));
5509 * Prevent mem_cgroup_migrate() from looking at
5510 * page->mem_cgroup of its source page while we change it.
5513 if (!trylock_page(page))
5517 if (page->mem_cgroup != from)
5520 anon = PageAnon(page);
5522 pgdat = page_pgdat(page);
5523 from_vec = mem_cgroup_lruvec(pgdat, from);
5524 to_vec = mem_cgroup_lruvec(pgdat, to);
5526 lock_page_memcg(page);
5528 if (!anon && page_mapped(page)) {
5529 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5530 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5533 if (!anon && PageDirty(page)) {
5534 struct address_space *mapping = page_mapping(page);
5536 if (mapping_cap_account_dirty(mapping)) {
5537 __mod_lruvec_state(from_vec, NR_FILE_DIRTY, -nr_pages);
5538 __mod_lruvec_state(to_vec, NR_FILE_DIRTY, nr_pages);
5542 if (PageWriteback(page)) {
5543 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5544 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5548 * All state has been migrated, let's switch to the new memcg.
5550 * It is safe to change page->mem_cgroup here because the page
5551 * is referenced, charged, isolated, and locked: we can't race
5552 * with (un)charging, migration, LRU putback, or anything else
5553 * that would rely on a stable page->mem_cgroup.
5555 * Note that lock_page_memcg is a memcg lock, not a page lock,
5556 * to save space. As soon as we switch page->mem_cgroup to a
5557 * new memcg that isn't locked, the above state can change
5558 * concurrently again. Make sure we're truly done with it.
5562 page->mem_cgroup = to; /* caller should have done css_get */
5564 __unlock_page_memcg(from);
5568 local_irq_disable();
5569 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
5570 memcg_check_events(to, page);
5571 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
5572 memcg_check_events(from, page);
5581 * get_mctgt_type - get target type of moving charge
5582 * @vma: the vma the pte to be checked belongs
5583 * @addr: the address corresponding to the pte to be checked
5584 * @ptent: the pte to be checked
5585 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5588 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5589 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5590 * move charge. if @target is not NULL, the page is stored in target->page
5591 * with extra refcnt got(Callers should handle it).
5592 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5593 * target for charge migration. if @target is not NULL, the entry is stored
5595 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5596 * (so ZONE_DEVICE page and thus not on the lru).
5597 * For now we such page is charge like a regular page would be as for all
5598 * intent and purposes it is just special memory taking the place of a
5601 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5603 * Called with pte lock held.
5606 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5607 unsigned long addr, pte_t ptent, union mc_target *target)
5609 struct page *page = NULL;
5610 enum mc_target_type ret = MC_TARGET_NONE;
5611 swp_entry_t ent = { .val = 0 };
5613 if (pte_present(ptent))
5614 page = mc_handle_present_pte(vma, addr, ptent);
5615 else if (is_swap_pte(ptent))
5616 page = mc_handle_swap_pte(vma, ptent, &ent);
5617 else if (pte_none(ptent))
5618 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5620 if (!page && !ent.val)
5624 * Do only loose check w/o serialization.
5625 * mem_cgroup_move_account() checks the page is valid or
5626 * not under LRU exclusion.
5628 if (page->mem_cgroup == mc.from) {
5629 ret = MC_TARGET_PAGE;
5630 if (is_device_private_page(page))
5631 ret = MC_TARGET_DEVICE;
5633 target->page = page;
5635 if (!ret || !target)
5639 * There is a swap entry and a page doesn't exist or isn't charged.
5640 * But we cannot move a tail-page in a THP.
5642 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5643 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5644 ret = MC_TARGET_SWAP;
5651 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5653 * We don't consider PMD mapped swapping or file mapped pages because THP does
5654 * not support them for now.
5655 * Caller should make sure that pmd_trans_huge(pmd) is true.
5657 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5658 unsigned long addr, pmd_t pmd, union mc_target *target)
5660 struct page *page = NULL;
5661 enum mc_target_type ret = MC_TARGET_NONE;
5663 if (unlikely(is_swap_pmd(pmd))) {
5664 VM_BUG_ON(thp_migration_supported() &&
5665 !is_pmd_migration_entry(pmd));
5668 page = pmd_page(pmd);
5669 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5670 if (!(mc.flags & MOVE_ANON))
5672 if (page->mem_cgroup == mc.from) {
5673 ret = MC_TARGET_PAGE;
5676 target->page = page;
5682 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5683 unsigned long addr, pmd_t pmd, union mc_target *target)
5685 return MC_TARGET_NONE;
5689 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5690 unsigned long addr, unsigned long end,
5691 struct mm_walk *walk)
5693 struct vm_area_struct *vma = walk->vma;
5697 ptl = pmd_trans_huge_lock(pmd, vma);
5700 * Note their can not be MC_TARGET_DEVICE for now as we do not
5701 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5702 * this might change.
5704 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5705 mc.precharge += HPAGE_PMD_NR;
5710 if (pmd_trans_unstable(pmd))
5712 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5713 for (; addr != end; pte++, addr += PAGE_SIZE)
5714 if (get_mctgt_type(vma, addr, *pte, NULL))
5715 mc.precharge++; /* increment precharge temporarily */
5716 pte_unmap_unlock(pte - 1, ptl);
5722 static const struct mm_walk_ops precharge_walk_ops = {
5723 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5726 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5728 unsigned long precharge;
5730 down_read(&mm->mmap_sem);
5731 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5732 up_read(&mm->mmap_sem);
5734 precharge = mc.precharge;
5740 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5742 unsigned long precharge = mem_cgroup_count_precharge(mm);
5744 VM_BUG_ON(mc.moving_task);
5745 mc.moving_task = current;
5746 return mem_cgroup_do_precharge(precharge);
5749 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5750 static void __mem_cgroup_clear_mc(void)
5752 struct mem_cgroup *from = mc.from;
5753 struct mem_cgroup *to = mc.to;
5755 /* we must uncharge all the leftover precharges from mc.to */
5757 cancel_charge(mc.to, mc.precharge);
5761 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5762 * we must uncharge here.
5764 if (mc.moved_charge) {
5765 cancel_charge(mc.from, mc.moved_charge);
5766 mc.moved_charge = 0;
5768 /* we must fixup refcnts and charges */
5769 if (mc.moved_swap) {
5770 /* uncharge swap account from the old cgroup */
5771 if (!mem_cgroup_is_root(mc.from))
5772 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5774 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5777 * we charged both to->memory and to->memsw, so we
5778 * should uncharge to->memory.
5780 if (!mem_cgroup_is_root(mc.to))
5781 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5783 css_put_many(&mc.to->css, mc.moved_swap);
5787 memcg_oom_recover(from);
5788 memcg_oom_recover(to);
5789 wake_up_all(&mc.waitq);
5792 static void mem_cgroup_clear_mc(void)
5794 struct mm_struct *mm = mc.mm;
5797 * we must clear moving_task before waking up waiters at the end of
5800 mc.moving_task = NULL;
5801 __mem_cgroup_clear_mc();
5802 spin_lock(&mc.lock);
5806 spin_unlock(&mc.lock);
5811 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5813 struct cgroup_subsys_state *css;
5814 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5815 struct mem_cgroup *from;
5816 struct task_struct *leader, *p;
5817 struct mm_struct *mm;
5818 unsigned long move_flags;
5821 /* charge immigration isn't supported on the default hierarchy */
5822 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5826 * Multi-process migrations only happen on the default hierarchy
5827 * where charge immigration is not used. Perform charge
5828 * immigration if @tset contains a leader and whine if there are
5832 cgroup_taskset_for_each_leader(leader, css, tset) {
5835 memcg = mem_cgroup_from_css(css);
5841 * We are now commited to this value whatever it is. Changes in this
5842 * tunable will only affect upcoming migrations, not the current one.
5843 * So we need to save it, and keep it going.
5845 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5849 from = mem_cgroup_from_task(p);
5851 VM_BUG_ON(from == memcg);
5853 mm = get_task_mm(p);
5856 /* We move charges only when we move a owner of the mm */
5857 if (mm->owner == p) {
5860 VM_BUG_ON(mc.precharge);
5861 VM_BUG_ON(mc.moved_charge);
5862 VM_BUG_ON(mc.moved_swap);
5864 spin_lock(&mc.lock);
5868 mc.flags = move_flags;
5869 spin_unlock(&mc.lock);
5870 /* We set mc.moving_task later */
5872 ret = mem_cgroup_precharge_mc(mm);
5874 mem_cgroup_clear_mc();
5881 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5884 mem_cgroup_clear_mc();
5887 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5888 unsigned long addr, unsigned long end,
5889 struct mm_walk *walk)
5892 struct vm_area_struct *vma = walk->vma;
5895 enum mc_target_type target_type;
5896 union mc_target target;
5899 ptl = pmd_trans_huge_lock(pmd, vma);
5901 if (mc.precharge < HPAGE_PMD_NR) {
5905 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5906 if (target_type == MC_TARGET_PAGE) {
5908 if (!isolate_lru_page(page)) {
5909 if (!mem_cgroup_move_account(page, true,
5911 mc.precharge -= HPAGE_PMD_NR;
5912 mc.moved_charge += HPAGE_PMD_NR;
5914 putback_lru_page(page);
5917 } else if (target_type == MC_TARGET_DEVICE) {
5919 if (!mem_cgroup_move_account(page, true,
5921 mc.precharge -= HPAGE_PMD_NR;
5922 mc.moved_charge += HPAGE_PMD_NR;
5930 if (pmd_trans_unstable(pmd))
5933 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5934 for (; addr != end; addr += PAGE_SIZE) {
5935 pte_t ptent = *(pte++);
5936 bool device = false;
5942 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5943 case MC_TARGET_DEVICE:
5946 case MC_TARGET_PAGE:
5949 * We can have a part of the split pmd here. Moving it
5950 * can be done but it would be too convoluted so simply
5951 * ignore such a partial THP and keep it in original
5952 * memcg. There should be somebody mapping the head.
5954 if (PageTransCompound(page))
5956 if (!device && isolate_lru_page(page))
5958 if (!mem_cgroup_move_account(page, false,
5961 /* we uncharge from mc.from later. */
5965 putback_lru_page(page);
5966 put: /* get_mctgt_type() gets the page */
5969 case MC_TARGET_SWAP:
5971 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5973 mem_cgroup_id_get_many(mc.to, 1);
5974 /* we fixup other refcnts and charges later. */
5982 pte_unmap_unlock(pte - 1, ptl);
5987 * We have consumed all precharges we got in can_attach().
5988 * We try charge one by one, but don't do any additional
5989 * charges to mc.to if we have failed in charge once in attach()
5992 ret = mem_cgroup_do_precharge(1);
6000 static const struct mm_walk_ops charge_walk_ops = {
6001 .pmd_entry = mem_cgroup_move_charge_pte_range,
6004 static void mem_cgroup_move_charge(void)
6006 lru_add_drain_all();
6008 * Signal lock_page_memcg() to take the memcg's move_lock
6009 * while we're moving its pages to another memcg. Then wait
6010 * for already started RCU-only updates to finish.
6012 atomic_inc(&mc.from->moving_account);
6015 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
6017 * Someone who are holding the mmap_sem might be waiting in
6018 * waitq. So we cancel all extra charges, wake up all waiters,
6019 * and retry. Because we cancel precharges, we might not be able
6020 * to move enough charges, but moving charge is a best-effort
6021 * feature anyway, so it wouldn't be a big problem.
6023 __mem_cgroup_clear_mc();
6028 * When we have consumed all precharges and failed in doing
6029 * additional charge, the page walk just aborts.
6031 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6034 up_read(&mc.mm->mmap_sem);
6035 atomic_dec(&mc.from->moving_account);
6038 static void mem_cgroup_move_task(void)
6041 mem_cgroup_move_charge();
6042 mem_cgroup_clear_mc();
6045 #else /* !CONFIG_MMU */
6046 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6050 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6053 static void mem_cgroup_move_task(void)
6059 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6060 * to verify whether we're attached to the default hierarchy on each mount
6063 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6066 * use_hierarchy is forced on the default hierarchy. cgroup core
6067 * guarantees that @root doesn't have any children, so turning it
6068 * on for the root memcg is enough.
6070 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6071 root_mem_cgroup->use_hierarchy = true;
6073 root_mem_cgroup->use_hierarchy = false;
6076 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6078 if (value == PAGE_COUNTER_MAX)
6079 seq_puts(m, "max\n");
6081 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6086 static u64 memory_current_read(struct cgroup_subsys_state *css,
6089 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6091 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6094 static int memory_min_show(struct seq_file *m, void *v)
6096 return seq_puts_memcg_tunable(m,
6097 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6100 static ssize_t memory_min_write(struct kernfs_open_file *of,
6101 char *buf, size_t nbytes, loff_t off)
6103 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6107 buf = strstrip(buf);
6108 err = page_counter_memparse(buf, "max", &min);
6112 page_counter_set_min(&memcg->memory, min);
6117 static int memory_low_show(struct seq_file *m, void *v)
6119 return seq_puts_memcg_tunable(m,
6120 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6123 static ssize_t memory_low_write(struct kernfs_open_file *of,
6124 char *buf, size_t nbytes, loff_t off)
6126 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6130 buf = strstrip(buf);
6131 err = page_counter_memparse(buf, "max", &low);
6135 page_counter_set_low(&memcg->memory, low);
6140 static int memory_high_show(struct seq_file *m, void *v)
6142 return seq_puts_memcg_tunable(m, READ_ONCE(mem_cgroup_from_seq(m)->high));
6145 static ssize_t memory_high_write(struct kernfs_open_file *of,
6146 char *buf, size_t nbytes, loff_t off)
6148 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6149 unsigned long nr_pages;
6153 buf = strstrip(buf);
6154 err = page_counter_memparse(buf, "max", &high);
6160 nr_pages = page_counter_read(&memcg->memory);
6161 if (nr_pages > high)
6162 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6165 memcg_wb_domain_size_changed(memcg);
6169 static int memory_max_show(struct seq_file *m, void *v)
6171 return seq_puts_memcg_tunable(m,
6172 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6175 static ssize_t memory_max_write(struct kernfs_open_file *of,
6176 char *buf, size_t nbytes, loff_t off)
6178 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6179 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
6180 bool drained = false;
6184 buf = strstrip(buf);
6185 err = page_counter_memparse(buf, "max", &max);
6189 xchg(&memcg->memory.max, max);
6192 unsigned long nr_pages = page_counter_read(&memcg->memory);
6194 if (nr_pages <= max)
6197 if (signal_pending(current)) {
6203 drain_all_stock(memcg);
6209 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6215 memcg_memory_event(memcg, MEMCG_OOM);
6216 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6220 memcg_wb_domain_size_changed(memcg);
6224 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6226 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6227 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6228 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6229 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6230 seq_printf(m, "oom_kill %lu\n",
6231 atomic_long_read(&events[MEMCG_OOM_KILL]));
6234 static int memory_events_show(struct seq_file *m, void *v)
6236 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6238 __memory_events_show(m, memcg->memory_events);
6242 static int memory_events_local_show(struct seq_file *m, void *v)
6244 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6246 __memory_events_show(m, memcg->memory_events_local);
6250 static int memory_stat_show(struct seq_file *m, void *v)
6252 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6255 buf = memory_stat_format(memcg);
6263 static int memory_oom_group_show(struct seq_file *m, void *v)
6265 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6267 seq_printf(m, "%d\n", memcg->oom_group);
6272 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6273 char *buf, size_t nbytes, loff_t off)
6275 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6278 buf = strstrip(buf);
6282 ret = kstrtoint(buf, 0, &oom_group);
6286 if (oom_group != 0 && oom_group != 1)
6289 memcg->oom_group = oom_group;
6294 static struct cftype memory_files[] = {
6297 .flags = CFTYPE_NOT_ON_ROOT,
6298 .read_u64 = memory_current_read,
6302 .flags = CFTYPE_NOT_ON_ROOT,
6303 .seq_show = memory_min_show,
6304 .write = memory_min_write,
6308 .flags = CFTYPE_NOT_ON_ROOT,
6309 .seq_show = memory_low_show,
6310 .write = memory_low_write,
6314 .flags = CFTYPE_NOT_ON_ROOT,
6315 .seq_show = memory_high_show,
6316 .write = memory_high_write,
6320 .flags = CFTYPE_NOT_ON_ROOT,
6321 .seq_show = memory_max_show,
6322 .write = memory_max_write,
6326 .flags = CFTYPE_NOT_ON_ROOT,
6327 .file_offset = offsetof(struct mem_cgroup, events_file),
6328 .seq_show = memory_events_show,
6331 .name = "events.local",
6332 .flags = CFTYPE_NOT_ON_ROOT,
6333 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6334 .seq_show = memory_events_local_show,
6338 .flags = CFTYPE_NOT_ON_ROOT,
6339 .seq_show = memory_stat_show,
6342 .name = "oom.group",
6343 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6344 .seq_show = memory_oom_group_show,
6345 .write = memory_oom_group_write,
6350 struct cgroup_subsys memory_cgrp_subsys = {
6351 .css_alloc = mem_cgroup_css_alloc,
6352 .css_online = mem_cgroup_css_online,
6353 .css_offline = mem_cgroup_css_offline,
6354 .css_released = mem_cgroup_css_released,
6355 .css_free = mem_cgroup_css_free,
6356 .css_reset = mem_cgroup_css_reset,
6357 .can_attach = mem_cgroup_can_attach,
6358 .cancel_attach = mem_cgroup_cancel_attach,
6359 .post_attach = mem_cgroup_move_task,
6360 .bind = mem_cgroup_bind,
6361 .dfl_cftypes = memory_files,
6362 .legacy_cftypes = mem_cgroup_legacy_files,
6367 * mem_cgroup_protected - check if memory consumption is in the normal range
6368 * @root: the top ancestor of the sub-tree being checked
6369 * @memcg: the memory cgroup to check
6371 * WARNING: This function is not stateless! It can only be used as part
6372 * of a top-down tree iteration, not for isolated queries.
6374 * Returns one of the following:
6375 * MEMCG_PROT_NONE: cgroup memory is not protected
6376 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
6377 * an unprotected supply of reclaimable memory from other cgroups.
6378 * MEMCG_PROT_MIN: cgroup memory is protected
6380 * @root is exclusive; it is never protected when looked at directly
6382 * To provide a proper hierarchical behavior, effective memory.min/low values
6383 * are used. Below is the description of how effective memory.low is calculated.
6384 * Effective memory.min values is calculated in the same way.
6386 * Effective memory.low is always equal or less than the original memory.low.
6387 * If there is no memory.low overcommittment (which is always true for
6388 * top-level memory cgroups), these two values are equal.
6389 * Otherwise, it's a part of parent's effective memory.low,
6390 * calculated as a cgroup's memory.low usage divided by sum of sibling's
6391 * memory.low usages, where memory.low usage is the size of actually
6395 * elow = min( memory.low, parent->elow * ------------------ ),
6396 * siblings_low_usage
6398 * | memory.current, if memory.current < memory.low
6403 * Such definition of the effective memory.low provides the expected
6404 * hierarchical behavior: parent's memory.low value is limiting
6405 * children, unprotected memory is reclaimed first and cgroups,
6406 * which are not using their guarantee do not affect actual memory
6409 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
6411 * A A/memory.low = 2G, A/memory.current = 6G
6413 * BC DE B/memory.low = 3G B/memory.current = 2G
6414 * C/memory.low = 1G C/memory.current = 2G
6415 * D/memory.low = 0 D/memory.current = 2G
6416 * E/memory.low = 10G E/memory.current = 0
6418 * and the memory pressure is applied, the following memory distribution
6419 * is expected (approximately):
6421 * A/memory.current = 2G
6423 * B/memory.current = 1.3G
6424 * C/memory.current = 0.6G
6425 * D/memory.current = 0
6426 * E/memory.current = 0
6428 * These calculations require constant tracking of the actual low usages
6429 * (see propagate_protected_usage()), as well as recursive calculation of
6430 * effective memory.low values. But as we do call mem_cgroup_protected()
6431 * path for each memory cgroup top-down from the reclaim,
6432 * it's possible to optimize this part, and save calculated elow
6433 * for next usage. This part is intentionally racy, but it's ok,
6434 * as memory.low is a best-effort mechanism.
6436 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
6437 struct mem_cgroup *memcg)
6439 struct mem_cgroup *parent;
6440 unsigned long emin, parent_emin;
6441 unsigned long elow, parent_elow;
6442 unsigned long usage;
6444 if (mem_cgroup_disabled())
6445 return MEMCG_PROT_NONE;
6448 root = root_mem_cgroup;
6451 * Effective values of the reclaim targets are ignored so they
6452 * can be stale. Have a look at mem_cgroup_protection for more
6454 * TODO: calculation should be more robust so that we do not need
6455 * that special casing.
6458 return MEMCG_PROT_NONE;
6460 usage = page_counter_read(&memcg->memory);
6462 return MEMCG_PROT_NONE;
6464 emin = memcg->memory.min;
6465 elow = memcg->memory.low;
6467 parent = parent_mem_cgroup(memcg);
6468 /* No parent means a non-hierarchical mode on v1 memcg */
6470 return MEMCG_PROT_NONE;
6475 parent_emin = READ_ONCE(parent->memory.emin);
6476 emin = min(emin, parent_emin);
6477 if (emin && parent_emin) {
6478 unsigned long min_usage, siblings_min_usage;
6480 min_usage = min(usage, memcg->memory.min);
6481 siblings_min_usage = atomic_long_read(
6482 &parent->memory.children_min_usage);
6484 if (min_usage && siblings_min_usage)
6485 emin = min(emin, parent_emin * min_usage /
6486 siblings_min_usage);
6489 parent_elow = READ_ONCE(parent->memory.elow);
6490 elow = min(elow, parent_elow);
6491 if (elow && parent_elow) {
6492 unsigned long low_usage, siblings_low_usage;
6494 low_usage = min(usage, memcg->memory.low);
6495 siblings_low_usage = atomic_long_read(
6496 &parent->memory.children_low_usage);
6498 if (low_usage && siblings_low_usage)
6499 elow = min(elow, parent_elow * low_usage /
6500 siblings_low_usage);
6504 memcg->memory.emin = emin;
6505 memcg->memory.elow = elow;
6508 return MEMCG_PROT_MIN;
6509 else if (usage <= elow)
6510 return MEMCG_PROT_LOW;
6512 return MEMCG_PROT_NONE;
6516 * mem_cgroup_try_charge - try charging a page
6517 * @page: page to charge
6518 * @mm: mm context of the victim
6519 * @gfp_mask: reclaim mode
6520 * @memcgp: charged memcg return
6521 * @compound: charge the page as compound or small page
6523 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6524 * pages according to @gfp_mask if necessary.
6526 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6527 * Otherwise, an error code is returned.
6529 * After page->mapping has been set up, the caller must finalize the
6530 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6531 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6533 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
6534 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6537 struct mem_cgroup *memcg = NULL;
6538 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6541 if (mem_cgroup_disabled())
6544 if (PageSwapCache(page)) {
6546 * Every swap fault against a single page tries to charge the
6547 * page, bail as early as possible. shmem_unuse() encounters
6548 * already charged pages, too. The USED bit is protected by
6549 * the page lock, which serializes swap cache removal, which
6550 * in turn serializes uncharging.
6552 VM_BUG_ON_PAGE(!PageLocked(page), page);
6553 if (compound_head(page)->mem_cgroup)
6556 if (do_swap_account) {
6557 swp_entry_t ent = { .val = page_private(page), };
6558 unsigned short id = lookup_swap_cgroup_id(ent);
6561 memcg = mem_cgroup_from_id(id);
6562 if (memcg && !css_tryget_online(&memcg->css))
6569 memcg = get_mem_cgroup_from_mm(mm);
6571 ret = try_charge(memcg, gfp_mask, nr_pages);
6573 css_put(&memcg->css);
6579 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
6580 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6583 struct mem_cgroup *memcg;
6586 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound);
6588 mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask);
6593 * mem_cgroup_commit_charge - commit a page charge
6594 * @page: page to charge
6595 * @memcg: memcg to charge the page to
6596 * @lrucare: page might be on LRU already
6597 * @compound: charge the page as compound or small page
6599 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6600 * after page->mapping has been set up. This must happen atomically
6601 * as part of the page instantiation, i.e. under the page table lock
6602 * for anonymous pages, under the page lock for page and swap cache.
6604 * In addition, the page must not be on the LRU during the commit, to
6605 * prevent racing with task migration. If it might be, use @lrucare.
6607 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6609 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
6610 bool lrucare, bool compound)
6612 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6614 VM_BUG_ON_PAGE(!page->mapping, page);
6615 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
6617 if (mem_cgroup_disabled())
6620 * Swap faults will attempt to charge the same page multiple
6621 * times. But reuse_swap_page() might have removed the page
6622 * from swapcache already, so we can't check PageSwapCache().
6627 commit_charge(page, memcg, lrucare);
6629 local_irq_disable();
6630 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
6631 memcg_check_events(memcg, page);
6634 if (do_memsw_account() && PageSwapCache(page)) {
6635 swp_entry_t entry = { .val = page_private(page) };
6637 * The swap entry might not get freed for a long time,
6638 * let's not wait for it. The page already received a
6639 * memory+swap charge, drop the swap entry duplicate.
6641 mem_cgroup_uncharge_swap(entry, nr_pages);
6646 * mem_cgroup_cancel_charge - cancel a page charge
6647 * @page: page to charge
6648 * @memcg: memcg to charge the page to
6649 * @compound: charge the page as compound or small page
6651 * Cancel a charge transaction started by mem_cgroup_try_charge().
6653 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
6656 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6658 if (mem_cgroup_disabled())
6661 * Swap faults will attempt to charge the same page multiple
6662 * times. But reuse_swap_page() might have removed the page
6663 * from swapcache already, so we can't check PageSwapCache().
6668 cancel_charge(memcg, nr_pages);
6671 struct uncharge_gather {
6672 struct mem_cgroup *memcg;
6673 unsigned long pgpgout;
6674 unsigned long nr_anon;
6675 unsigned long nr_file;
6676 unsigned long nr_kmem;
6677 unsigned long nr_huge;
6678 unsigned long nr_shmem;
6679 struct page *dummy_page;
6682 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6684 memset(ug, 0, sizeof(*ug));
6687 static void uncharge_batch(const struct uncharge_gather *ug)
6689 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
6690 unsigned long flags;
6692 if (!mem_cgroup_is_root(ug->memcg)) {
6693 page_counter_uncharge(&ug->memcg->memory, nr_pages);
6694 if (do_memsw_account())
6695 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
6696 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6697 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6698 memcg_oom_recover(ug->memcg);
6701 local_irq_save(flags);
6702 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
6703 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
6704 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
6705 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
6706 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6707 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, nr_pages);
6708 memcg_check_events(ug->memcg, ug->dummy_page);
6709 local_irq_restore(flags);
6711 if (!mem_cgroup_is_root(ug->memcg))
6712 css_put_many(&ug->memcg->css, nr_pages);
6715 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6717 VM_BUG_ON_PAGE(PageLRU(page), page);
6718 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
6719 !PageHWPoison(page) , page);
6721 if (!page->mem_cgroup)
6725 * Nobody should be changing or seriously looking at
6726 * page->mem_cgroup at this point, we have fully
6727 * exclusive access to the page.
6730 if (ug->memcg != page->mem_cgroup) {
6733 uncharge_gather_clear(ug);
6735 ug->memcg = page->mem_cgroup;
6738 if (!PageKmemcg(page)) {
6739 unsigned int nr_pages = 1;
6741 if (PageTransHuge(page)) {
6742 nr_pages = compound_nr(page);
6743 ug->nr_huge += nr_pages;
6746 ug->nr_anon += nr_pages;
6748 ug->nr_file += nr_pages;
6749 if (PageSwapBacked(page))
6750 ug->nr_shmem += nr_pages;
6754 ug->nr_kmem += compound_nr(page);
6755 __ClearPageKmemcg(page);
6758 ug->dummy_page = page;
6759 page->mem_cgroup = NULL;
6762 static void uncharge_list(struct list_head *page_list)
6764 struct uncharge_gather ug;
6765 struct list_head *next;
6767 uncharge_gather_clear(&ug);
6770 * Note that the list can be a single page->lru; hence the
6771 * do-while loop instead of a simple list_for_each_entry().
6773 next = page_list->next;
6777 page = list_entry(next, struct page, lru);
6778 next = page->lru.next;
6780 uncharge_page(page, &ug);
6781 } while (next != page_list);
6784 uncharge_batch(&ug);
6788 * mem_cgroup_uncharge - uncharge a page
6789 * @page: page to uncharge
6791 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6792 * mem_cgroup_commit_charge().
6794 void mem_cgroup_uncharge(struct page *page)
6796 struct uncharge_gather ug;
6798 if (mem_cgroup_disabled())
6801 /* Don't touch page->lru of any random page, pre-check: */
6802 if (!page->mem_cgroup)
6805 uncharge_gather_clear(&ug);
6806 uncharge_page(page, &ug);
6807 uncharge_batch(&ug);
6811 * mem_cgroup_uncharge_list - uncharge a list of page
6812 * @page_list: list of pages to uncharge
6814 * Uncharge a list of pages previously charged with
6815 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6817 void mem_cgroup_uncharge_list(struct list_head *page_list)
6819 if (mem_cgroup_disabled())
6822 if (!list_empty(page_list))
6823 uncharge_list(page_list);
6827 * mem_cgroup_migrate - charge a page's replacement
6828 * @oldpage: currently circulating page
6829 * @newpage: replacement page
6831 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6832 * be uncharged upon free.
6834 * Both pages must be locked, @newpage->mapping must be set up.
6836 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6838 struct mem_cgroup *memcg;
6839 unsigned int nr_pages;
6841 unsigned long flags;
6843 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6844 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6845 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6846 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6849 if (mem_cgroup_disabled())
6852 /* Page cache replacement: new page already charged? */
6853 if (newpage->mem_cgroup)
6856 /* Swapcache readahead pages can get replaced before being charged */
6857 memcg = oldpage->mem_cgroup;
6861 /* Force-charge the new page. The old one will be freed soon */
6862 compound = PageTransHuge(newpage);
6863 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
6865 page_counter_charge(&memcg->memory, nr_pages);
6866 if (do_memsw_account())
6867 page_counter_charge(&memcg->memsw, nr_pages);
6868 css_get_many(&memcg->css, nr_pages);
6870 commit_charge(newpage, memcg, false);
6872 local_irq_save(flags);
6873 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
6874 memcg_check_events(memcg, newpage);
6875 local_irq_restore(flags);
6878 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6879 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6881 void mem_cgroup_sk_alloc(struct sock *sk)
6883 struct mem_cgroup *memcg;
6885 if (!mem_cgroup_sockets_enabled)
6888 /* Do not associate the sock with unrelated interrupted task's memcg. */
6893 memcg = mem_cgroup_from_task(current);
6894 if (memcg == root_mem_cgroup)
6896 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6898 if (css_tryget_online(&memcg->css))
6899 sk->sk_memcg = memcg;
6904 void mem_cgroup_sk_free(struct sock *sk)
6907 css_put(&sk->sk_memcg->css);
6911 * mem_cgroup_charge_skmem - charge socket memory
6912 * @memcg: memcg to charge
6913 * @nr_pages: number of pages to charge
6915 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6916 * @memcg's configured limit, %false if the charge had to be forced.
6918 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6920 gfp_t gfp_mask = GFP_KERNEL;
6922 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6923 struct page_counter *fail;
6925 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6926 memcg->tcpmem_pressure = 0;
6929 page_counter_charge(&memcg->tcpmem, nr_pages);
6930 memcg->tcpmem_pressure = 1;
6934 /* Don't block in the packet receive path */
6936 gfp_mask = GFP_NOWAIT;
6938 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6940 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6943 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6948 * mem_cgroup_uncharge_skmem - uncharge socket memory
6949 * @memcg: memcg to uncharge
6950 * @nr_pages: number of pages to uncharge
6952 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6954 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6955 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6959 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6961 refill_stock(memcg, nr_pages);
6964 static int __init cgroup_memory(char *s)
6968 while ((token = strsep(&s, ",")) != NULL) {
6971 if (!strcmp(token, "nosocket"))
6972 cgroup_memory_nosocket = true;
6973 if (!strcmp(token, "nokmem"))
6974 cgroup_memory_nokmem = true;
6978 __setup("cgroup.memory=", cgroup_memory);
6981 * subsys_initcall() for memory controller.
6983 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6984 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6985 * basically everything that doesn't depend on a specific mem_cgroup structure
6986 * should be initialized from here.
6988 static int __init mem_cgroup_init(void)
6992 #ifdef CONFIG_MEMCG_KMEM
6994 * Kmem cache creation is mostly done with the slab_mutex held,
6995 * so use a workqueue with limited concurrency to avoid stalling
6996 * all worker threads in case lots of cgroups are created and
6997 * destroyed simultaneously.
6999 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
7000 BUG_ON(!memcg_kmem_cache_wq);
7003 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7004 memcg_hotplug_cpu_dead);
7006 for_each_possible_cpu(cpu)
7007 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7010 for_each_node(node) {
7011 struct mem_cgroup_tree_per_node *rtpn;
7013 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7014 node_online(node) ? node : NUMA_NO_NODE);
7016 rtpn->rb_root = RB_ROOT;
7017 rtpn->rb_rightmost = NULL;
7018 spin_lock_init(&rtpn->lock);
7019 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7024 subsys_initcall(mem_cgroup_init);
7026 #ifdef CONFIG_MEMCG_SWAP
7027 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7029 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7031 * The root cgroup cannot be destroyed, so it's refcount must
7034 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7038 memcg = parent_mem_cgroup(memcg);
7040 memcg = root_mem_cgroup;
7046 * mem_cgroup_swapout - transfer a memsw charge to swap
7047 * @page: page whose memsw charge to transfer
7048 * @entry: swap entry to move the charge to
7050 * Transfer the memsw charge of @page to @entry.
7052 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7054 struct mem_cgroup *memcg, *swap_memcg;
7055 unsigned int nr_entries;
7056 unsigned short oldid;
7058 VM_BUG_ON_PAGE(PageLRU(page), page);
7059 VM_BUG_ON_PAGE(page_count(page), page);
7061 if (!do_memsw_account())
7064 memcg = page->mem_cgroup;
7066 /* Readahead page, never charged */
7071 * In case the memcg owning these pages has been offlined and doesn't
7072 * have an ID allocated to it anymore, charge the closest online
7073 * ancestor for the swap instead and transfer the memory+swap charge.
7075 swap_memcg = mem_cgroup_id_get_online(memcg);
7076 nr_entries = hpage_nr_pages(page);
7077 /* Get references for the tail pages, too */
7079 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7080 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7082 VM_BUG_ON_PAGE(oldid, page);
7083 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7085 page->mem_cgroup = NULL;
7087 if (!mem_cgroup_is_root(memcg))
7088 page_counter_uncharge(&memcg->memory, nr_entries);
7090 if (memcg != swap_memcg) {
7091 if (!mem_cgroup_is_root(swap_memcg))
7092 page_counter_charge(&swap_memcg->memsw, nr_entries);
7093 page_counter_uncharge(&memcg->memsw, nr_entries);
7097 * Interrupts should be disabled here because the caller holds the
7098 * i_pages lock which is taken with interrupts-off. It is
7099 * important here to have the interrupts disabled because it is the
7100 * only synchronisation we have for updating the per-CPU variables.
7102 VM_BUG_ON(!irqs_disabled());
7103 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
7105 memcg_check_events(memcg, page);
7107 if (!mem_cgroup_is_root(memcg))
7108 css_put_many(&memcg->css, nr_entries);
7112 * mem_cgroup_try_charge_swap - try charging swap space for a page
7113 * @page: page being added to swap
7114 * @entry: swap entry to charge
7116 * Try to charge @page's memcg for the swap space at @entry.
7118 * Returns 0 on success, -ENOMEM on failure.
7120 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7122 unsigned int nr_pages = hpage_nr_pages(page);
7123 struct page_counter *counter;
7124 struct mem_cgroup *memcg;
7125 unsigned short oldid;
7127 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
7130 memcg = page->mem_cgroup;
7132 /* Readahead page, never charged */
7137 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7141 memcg = mem_cgroup_id_get_online(memcg);
7143 if (!mem_cgroup_is_root(memcg) &&
7144 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7145 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7146 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7147 mem_cgroup_id_put(memcg);
7151 /* Get references for the tail pages, too */
7153 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7154 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7155 VM_BUG_ON_PAGE(oldid, page);
7156 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7162 * mem_cgroup_uncharge_swap - uncharge swap space
7163 * @entry: swap entry to uncharge
7164 * @nr_pages: the amount of swap space to uncharge
7166 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7168 struct mem_cgroup *memcg;
7171 if (!do_swap_account)
7174 id = swap_cgroup_record(entry, 0, nr_pages);
7176 memcg = mem_cgroup_from_id(id);
7178 if (!mem_cgroup_is_root(memcg)) {
7179 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7180 page_counter_uncharge(&memcg->swap, nr_pages);
7182 page_counter_uncharge(&memcg->memsw, nr_pages);
7184 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7185 mem_cgroup_id_put_many(memcg, nr_pages);
7190 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7192 long nr_swap_pages = get_nr_swap_pages();
7194 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7195 return nr_swap_pages;
7196 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7197 nr_swap_pages = min_t(long, nr_swap_pages,
7198 READ_ONCE(memcg->swap.max) -
7199 page_counter_read(&memcg->swap));
7200 return nr_swap_pages;
7203 bool mem_cgroup_swap_full(struct page *page)
7205 struct mem_cgroup *memcg;
7207 VM_BUG_ON_PAGE(!PageLocked(page), page);
7211 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7214 memcg = page->mem_cgroup;
7218 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7219 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.max)
7225 /* for remember boot option*/
7226 #ifdef CONFIG_MEMCG_SWAP_ENABLED
7227 static int really_do_swap_account __initdata = 1;
7229 static int really_do_swap_account __initdata;
7232 static int __init enable_swap_account(char *s)
7234 if (!strcmp(s, "1"))
7235 really_do_swap_account = 1;
7236 else if (!strcmp(s, "0"))
7237 really_do_swap_account = 0;
7240 __setup("swapaccount=", enable_swap_account);
7242 static u64 swap_current_read(struct cgroup_subsys_state *css,
7245 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7247 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7250 static int swap_max_show(struct seq_file *m, void *v)
7252 return seq_puts_memcg_tunable(m,
7253 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7256 static ssize_t swap_max_write(struct kernfs_open_file *of,
7257 char *buf, size_t nbytes, loff_t off)
7259 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7263 buf = strstrip(buf);
7264 err = page_counter_memparse(buf, "max", &max);
7268 xchg(&memcg->swap.max, max);
7273 static int swap_events_show(struct seq_file *m, void *v)
7275 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7277 seq_printf(m, "max %lu\n",
7278 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7279 seq_printf(m, "fail %lu\n",
7280 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7285 static struct cftype swap_files[] = {
7287 .name = "swap.current",
7288 .flags = CFTYPE_NOT_ON_ROOT,
7289 .read_u64 = swap_current_read,
7293 .flags = CFTYPE_NOT_ON_ROOT,
7294 .seq_show = swap_max_show,
7295 .write = swap_max_write,
7298 .name = "swap.events",
7299 .flags = CFTYPE_NOT_ON_ROOT,
7300 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7301 .seq_show = swap_events_show,
7306 static struct cftype memsw_cgroup_files[] = {
7308 .name = "memsw.usage_in_bytes",
7309 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7310 .read_u64 = mem_cgroup_read_u64,
7313 .name = "memsw.max_usage_in_bytes",
7314 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7315 .write = mem_cgroup_reset,
7316 .read_u64 = mem_cgroup_read_u64,
7319 .name = "memsw.limit_in_bytes",
7320 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7321 .write = mem_cgroup_write,
7322 .read_u64 = mem_cgroup_read_u64,
7325 .name = "memsw.failcnt",
7326 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7327 .write = mem_cgroup_reset,
7328 .read_u64 = mem_cgroup_read_u64,
7330 { }, /* terminate */
7333 static int __init mem_cgroup_swap_init(void)
7335 if (!mem_cgroup_disabled() && really_do_swap_account) {
7336 do_swap_account = 1;
7337 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
7339 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
7340 memsw_cgroup_files));
7344 subsys_initcall(mem_cgroup_swap_init);
7346 #endif /* CONFIG_MEMCG_SWAP */