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);
779 /* Untracked pages have no memcg, no lruvec. Update only the node */
780 if (!memcg || memcg == root_mem_cgroup) {
781 __mod_node_page_state(pgdat, idx, val);
783 lruvec = mem_cgroup_lruvec(pgdat, memcg);
784 __mod_lruvec_state(lruvec, idx, val);
789 void mod_memcg_obj_state(void *p, int idx, int val)
791 struct mem_cgroup *memcg;
794 memcg = mem_cgroup_from_obj(p);
796 mod_memcg_state(memcg, idx, val);
801 * __count_memcg_events - account VM events in a cgroup
802 * @memcg: the memory cgroup
803 * @idx: the event item
804 * @count: the number of events that occured
806 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
811 if (mem_cgroup_disabled())
814 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
815 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
816 struct mem_cgroup *mi;
819 * Batch local counters to keep them in sync with
820 * the hierarchical ones.
822 __this_cpu_add(memcg->vmstats_local->events[idx], x);
823 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
824 atomic_long_add(x, &mi->vmevents[idx]);
827 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
830 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
832 return atomic_long_read(&memcg->vmevents[event]);
835 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
840 for_each_possible_cpu(cpu)
841 x += per_cpu(memcg->vmstats_local->events[event], cpu);
845 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
847 bool compound, int nr_pages)
850 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
851 * counted as CACHE even if it's on ANON LRU.
854 __mod_memcg_state(memcg, MEMCG_RSS, nr_pages);
856 __mod_memcg_state(memcg, MEMCG_CACHE, nr_pages);
857 if (PageSwapBacked(page))
858 __mod_memcg_state(memcg, NR_SHMEM, nr_pages);
862 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
863 __mod_memcg_state(memcg, MEMCG_RSS_HUGE, nr_pages);
866 /* pagein of a big page is an event. So, ignore page size */
868 __count_memcg_events(memcg, PGPGIN, 1);
870 __count_memcg_events(memcg, PGPGOUT, 1);
871 nr_pages = -nr_pages; /* for event */
874 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
877 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
878 enum mem_cgroup_events_target target)
880 unsigned long val, next;
882 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
883 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
884 /* from time_after() in jiffies.h */
885 if ((long)(next - val) < 0) {
887 case MEM_CGROUP_TARGET_THRESH:
888 next = val + THRESHOLDS_EVENTS_TARGET;
890 case MEM_CGROUP_TARGET_SOFTLIMIT:
891 next = val + SOFTLIMIT_EVENTS_TARGET;
893 case MEM_CGROUP_TARGET_NUMAINFO:
894 next = val + NUMAINFO_EVENTS_TARGET;
899 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
906 * Check events in order.
909 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
911 /* threshold event is triggered in finer grain than soft limit */
912 if (unlikely(mem_cgroup_event_ratelimit(memcg,
913 MEM_CGROUP_TARGET_THRESH))) {
915 bool do_numainfo __maybe_unused;
917 do_softlimit = mem_cgroup_event_ratelimit(memcg,
918 MEM_CGROUP_TARGET_SOFTLIMIT);
920 do_numainfo = mem_cgroup_event_ratelimit(memcg,
921 MEM_CGROUP_TARGET_NUMAINFO);
923 mem_cgroup_threshold(memcg);
924 if (unlikely(do_softlimit))
925 mem_cgroup_update_tree(memcg, page);
927 if (unlikely(do_numainfo))
928 atomic_inc(&memcg->numainfo_events);
933 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
936 * mm_update_next_owner() may clear mm->owner to NULL
937 * if it races with swapoff, page migration, etc.
938 * So this can be called with p == NULL.
943 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
945 EXPORT_SYMBOL(mem_cgroup_from_task);
948 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
949 * @mm: mm from which memcg should be extracted. It can be NULL.
951 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
952 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
955 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
957 struct mem_cgroup *memcg;
959 if (mem_cgroup_disabled())
965 * Page cache insertions can happen withou an
966 * actual mm context, e.g. during disk probing
967 * on boot, loopback IO, acct() writes etc.
970 memcg = root_mem_cgroup;
972 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
973 if (unlikely(!memcg))
974 memcg = root_mem_cgroup;
976 } while (!css_tryget(&memcg->css));
980 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
983 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
984 * @page: page from which memcg should be extracted.
986 * Obtain a reference on page->memcg and returns it if successful. Otherwise
987 * root_mem_cgroup is returned.
989 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
991 struct mem_cgroup *memcg = page->mem_cgroup;
993 if (mem_cgroup_disabled())
997 if (!memcg || !css_tryget_online(&memcg->css))
998 memcg = root_mem_cgroup;
1002 EXPORT_SYMBOL(get_mem_cgroup_from_page);
1005 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
1007 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
1009 if (unlikely(current->active_memcg)) {
1010 struct mem_cgroup *memcg = root_mem_cgroup;
1013 if (css_tryget_online(¤t->active_memcg->css))
1014 memcg = current->active_memcg;
1018 return get_mem_cgroup_from_mm(current->mm);
1022 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1023 * @root: hierarchy root
1024 * @prev: previously returned memcg, NULL on first invocation
1025 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1027 * Returns references to children of the hierarchy below @root, or
1028 * @root itself, or %NULL after a full round-trip.
1030 * Caller must pass the return value in @prev on subsequent
1031 * invocations for reference counting, or use mem_cgroup_iter_break()
1032 * to cancel a hierarchy walk before the round-trip is complete.
1034 * Reclaimers can specify a node and a priority level in @reclaim to
1035 * divide up the memcgs in the hierarchy among all concurrent
1036 * reclaimers operating on the same node and priority.
1038 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1039 struct mem_cgroup *prev,
1040 struct mem_cgroup_reclaim_cookie *reclaim)
1042 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1043 struct cgroup_subsys_state *css = NULL;
1044 struct mem_cgroup *memcg = NULL;
1045 struct mem_cgroup *pos = NULL;
1047 if (mem_cgroup_disabled())
1051 root = root_mem_cgroup;
1053 if (prev && !reclaim)
1056 if (!root->use_hierarchy && root != root_mem_cgroup) {
1065 struct mem_cgroup_per_node *mz;
1067 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1068 iter = &mz->iter[reclaim->priority];
1070 if (prev && reclaim->generation != iter->generation)
1074 pos = READ_ONCE(iter->position);
1075 if (!pos || css_tryget(&pos->css))
1078 * css reference reached zero, so iter->position will
1079 * be cleared by ->css_released. However, we should not
1080 * rely on this happening soon, because ->css_released
1081 * is called from a work queue, and by busy-waiting we
1082 * might block it. So we clear iter->position right
1085 (void)cmpxchg(&iter->position, pos, NULL);
1093 css = css_next_descendant_pre(css, &root->css);
1096 * Reclaimers share the hierarchy walk, and a
1097 * new one might jump in right at the end of
1098 * the hierarchy - make sure they see at least
1099 * one group and restart from the beginning.
1107 * Verify the css and acquire a reference. The root
1108 * is provided by the caller, so we know it's alive
1109 * and kicking, and don't take an extra reference.
1111 memcg = mem_cgroup_from_css(css);
1113 if (css == &root->css)
1116 if (css_tryget(css))
1124 * The position could have already been updated by a competing
1125 * thread, so check that the value hasn't changed since we read
1126 * it to avoid reclaiming from the same cgroup twice.
1128 (void)cmpxchg(&iter->position, pos, memcg);
1136 reclaim->generation = iter->generation;
1142 if (prev && prev != root)
1143 css_put(&prev->css);
1149 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1150 * @root: hierarchy root
1151 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1153 void mem_cgroup_iter_break(struct mem_cgroup *root,
1154 struct mem_cgroup *prev)
1157 root = root_mem_cgroup;
1158 if (prev && prev != root)
1159 css_put(&prev->css);
1162 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1163 struct mem_cgroup *dead_memcg)
1165 struct mem_cgroup_reclaim_iter *iter;
1166 struct mem_cgroup_per_node *mz;
1170 for_each_node(nid) {
1171 mz = mem_cgroup_nodeinfo(from, nid);
1172 for (i = 0; i <= DEF_PRIORITY; i++) {
1173 iter = &mz->iter[i];
1174 cmpxchg(&iter->position,
1180 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1182 struct mem_cgroup *memcg = dead_memcg;
1183 struct mem_cgroup *last;
1186 __invalidate_reclaim_iterators(memcg, dead_memcg);
1188 } while ((memcg = parent_mem_cgroup(memcg)));
1191 * When cgruop1 non-hierarchy mode is used,
1192 * parent_mem_cgroup() does not walk all the way up to the
1193 * cgroup root (root_mem_cgroup). So we have to handle
1194 * dead_memcg from cgroup root separately.
1196 if (last != root_mem_cgroup)
1197 __invalidate_reclaim_iterators(root_mem_cgroup,
1202 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1203 * @memcg: hierarchy root
1204 * @fn: function to call for each task
1205 * @arg: argument passed to @fn
1207 * This function iterates over tasks attached to @memcg or to any of its
1208 * descendants and calls @fn for each task. If @fn returns a non-zero
1209 * value, the function breaks the iteration loop and returns the value.
1210 * Otherwise, it will iterate over all tasks and return 0.
1212 * This function must not be called for the root memory cgroup.
1214 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1215 int (*fn)(struct task_struct *, void *), void *arg)
1217 struct mem_cgroup *iter;
1220 BUG_ON(memcg == root_mem_cgroup);
1222 for_each_mem_cgroup_tree(iter, memcg) {
1223 struct css_task_iter it;
1224 struct task_struct *task;
1226 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1227 while (!ret && (task = css_task_iter_next(&it)))
1228 ret = fn(task, arg);
1229 css_task_iter_end(&it);
1231 mem_cgroup_iter_break(memcg, iter);
1239 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1241 * @pgdat: pgdat of the page
1243 * This function is only safe when following the LRU page isolation
1244 * and putback protocol: the LRU lock must be held, and the page must
1245 * either be PageLRU() or the caller must have isolated/allocated it.
1247 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1249 struct mem_cgroup_per_node *mz;
1250 struct mem_cgroup *memcg;
1251 struct lruvec *lruvec;
1253 if (mem_cgroup_disabled()) {
1254 lruvec = &pgdat->lruvec;
1258 memcg = page->mem_cgroup;
1260 * Swapcache readahead pages are added to the LRU - and
1261 * possibly migrated - before they are charged.
1264 memcg = root_mem_cgroup;
1266 mz = mem_cgroup_page_nodeinfo(memcg, page);
1267 lruvec = &mz->lruvec;
1270 * Since a node can be onlined after the mem_cgroup was created,
1271 * we have to be prepared to initialize lruvec->zone here;
1272 * and if offlined then reonlined, we need to reinitialize it.
1274 if (unlikely(lruvec->pgdat != pgdat))
1275 lruvec->pgdat = pgdat;
1280 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1281 * @lruvec: mem_cgroup per zone lru vector
1282 * @lru: index of lru list the page is sitting on
1283 * @zid: zone id of the accounted pages
1284 * @nr_pages: positive when adding or negative when removing
1286 * This function must be called under lru_lock, just before a page is added
1287 * to or just after a page is removed from an lru list (that ordering being
1288 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1290 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1291 int zid, int nr_pages)
1293 struct mem_cgroup_per_node *mz;
1294 unsigned long *lru_size;
1297 if (mem_cgroup_disabled())
1300 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1301 lru_size = &mz->lru_zone_size[zid][lru];
1304 *lru_size += nr_pages;
1307 if (WARN_ONCE(size < 0,
1308 "%s(%p, %d, %d): lru_size %ld\n",
1309 __func__, lruvec, lru, nr_pages, size)) {
1315 *lru_size += nr_pages;
1319 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1320 * @memcg: the memory cgroup
1322 * Returns the maximum amount of memory @mem can be charged with, in
1325 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1327 unsigned long margin = 0;
1328 unsigned long count;
1329 unsigned long limit;
1331 count = page_counter_read(&memcg->memory);
1332 limit = READ_ONCE(memcg->memory.max);
1334 margin = limit - count;
1336 if (do_memsw_account()) {
1337 count = page_counter_read(&memcg->memsw);
1338 limit = READ_ONCE(memcg->memsw.max);
1340 margin = min(margin, limit - count);
1349 * A routine for checking "mem" is under move_account() or not.
1351 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1352 * moving cgroups. This is for waiting at high-memory pressure
1355 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1357 struct mem_cgroup *from;
1358 struct mem_cgroup *to;
1361 * Unlike task_move routines, we access mc.to, mc.from not under
1362 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1364 spin_lock(&mc.lock);
1370 ret = mem_cgroup_is_descendant(from, memcg) ||
1371 mem_cgroup_is_descendant(to, memcg);
1373 spin_unlock(&mc.lock);
1377 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1379 if (mc.moving_task && current != mc.moving_task) {
1380 if (mem_cgroup_under_move(memcg)) {
1382 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1383 /* moving charge context might have finished. */
1386 finish_wait(&mc.waitq, &wait);
1393 static char *memory_stat_format(struct mem_cgroup *memcg)
1398 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1403 * Provide statistics on the state of the memory subsystem as
1404 * well as cumulative event counters that show past behavior.
1406 * This list is ordered following a combination of these gradients:
1407 * 1) generic big picture -> specifics and details
1408 * 2) reflecting userspace activity -> reflecting kernel heuristics
1410 * Current memory state:
1413 seq_buf_printf(&s, "anon %llu\n",
1414 (u64)memcg_page_state(memcg, MEMCG_RSS) *
1416 seq_buf_printf(&s, "file %llu\n",
1417 (u64)memcg_page_state(memcg, MEMCG_CACHE) *
1419 seq_buf_printf(&s, "kernel_stack %llu\n",
1420 (u64)memcg_page_state(memcg, MEMCG_KERNEL_STACK_KB) *
1422 seq_buf_printf(&s, "slab %llu\n",
1423 (u64)(memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) +
1424 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE)) *
1426 seq_buf_printf(&s, "sock %llu\n",
1427 (u64)memcg_page_state(memcg, MEMCG_SOCK) *
1430 seq_buf_printf(&s, "shmem %llu\n",
1431 (u64)memcg_page_state(memcg, NR_SHMEM) *
1433 seq_buf_printf(&s, "file_mapped %llu\n",
1434 (u64)memcg_page_state(memcg, NR_FILE_MAPPED) *
1436 seq_buf_printf(&s, "file_dirty %llu\n",
1437 (u64)memcg_page_state(memcg, NR_FILE_DIRTY) *
1439 seq_buf_printf(&s, "file_writeback %llu\n",
1440 (u64)memcg_page_state(memcg, NR_WRITEBACK) *
1444 * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter
1445 * with the NR_ANON_THP vm counter, but right now it's a pain in the
1446 * arse because it requires migrating the work out of rmap to a place
1447 * where the page->mem_cgroup is set up and stable.
1449 seq_buf_printf(&s, "anon_thp %llu\n",
1450 (u64)memcg_page_state(memcg, MEMCG_RSS_HUGE) *
1453 for (i = 0; i < NR_LRU_LISTS; i++)
1454 seq_buf_printf(&s, "%s %llu\n", mem_cgroup_lru_names[i],
1455 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
1458 seq_buf_printf(&s, "slab_reclaimable %llu\n",
1459 (u64)memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) *
1461 seq_buf_printf(&s, "slab_unreclaimable %llu\n",
1462 (u64)memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE) *
1465 /* Accumulated memory events */
1467 seq_buf_printf(&s, "pgfault %lu\n", memcg_events(memcg, PGFAULT));
1468 seq_buf_printf(&s, "pgmajfault %lu\n", memcg_events(memcg, PGMAJFAULT));
1470 seq_buf_printf(&s, "workingset_refault %lu\n",
1471 memcg_page_state(memcg, WORKINGSET_REFAULT));
1472 seq_buf_printf(&s, "workingset_activate %lu\n",
1473 memcg_page_state(memcg, WORKINGSET_ACTIVATE));
1474 seq_buf_printf(&s, "workingset_nodereclaim %lu\n",
1475 memcg_page_state(memcg, WORKINGSET_NODERECLAIM));
1477 seq_buf_printf(&s, "pgrefill %lu\n", memcg_events(memcg, PGREFILL));
1478 seq_buf_printf(&s, "pgscan %lu\n",
1479 memcg_events(memcg, PGSCAN_KSWAPD) +
1480 memcg_events(memcg, PGSCAN_DIRECT));
1481 seq_buf_printf(&s, "pgsteal %lu\n",
1482 memcg_events(memcg, PGSTEAL_KSWAPD) +
1483 memcg_events(memcg, PGSTEAL_DIRECT));
1484 seq_buf_printf(&s, "pgactivate %lu\n", memcg_events(memcg, PGACTIVATE));
1485 seq_buf_printf(&s, "pgdeactivate %lu\n", memcg_events(memcg, PGDEACTIVATE));
1486 seq_buf_printf(&s, "pglazyfree %lu\n", memcg_events(memcg, PGLAZYFREE));
1487 seq_buf_printf(&s, "pglazyfreed %lu\n", memcg_events(memcg, PGLAZYFREED));
1489 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1490 seq_buf_printf(&s, "thp_fault_alloc %lu\n",
1491 memcg_events(memcg, THP_FAULT_ALLOC));
1492 seq_buf_printf(&s, "thp_collapse_alloc %lu\n",
1493 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1494 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1496 /* The above should easily fit into one page */
1497 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1502 #define K(x) ((x) << (PAGE_SHIFT-10))
1504 * mem_cgroup_print_oom_context: Print OOM information relevant to
1505 * memory controller.
1506 * @memcg: The memory cgroup that went over limit
1507 * @p: Task that is going to be killed
1509 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1512 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1517 pr_cont(",oom_memcg=");
1518 pr_cont_cgroup_path(memcg->css.cgroup);
1520 pr_cont(",global_oom");
1522 pr_cont(",task_memcg=");
1523 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1529 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1530 * memory controller.
1531 * @memcg: The memory cgroup that went over limit
1533 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1537 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1538 K((u64)page_counter_read(&memcg->memory)),
1539 K((u64)memcg->memory.max), memcg->memory.failcnt);
1540 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1541 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1542 K((u64)page_counter_read(&memcg->swap)),
1543 K((u64)memcg->swap.max), memcg->swap.failcnt);
1545 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1546 K((u64)page_counter_read(&memcg->memsw)),
1547 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1548 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1549 K((u64)page_counter_read(&memcg->kmem)),
1550 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1553 pr_info("Memory cgroup stats for ");
1554 pr_cont_cgroup_path(memcg->css.cgroup);
1556 buf = memory_stat_format(memcg);
1564 * Return the memory (and swap, if configured) limit for a memcg.
1566 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1570 max = memcg->memory.max;
1571 if (mem_cgroup_swappiness(memcg)) {
1572 unsigned long memsw_max;
1573 unsigned long swap_max;
1575 memsw_max = memcg->memsw.max;
1576 swap_max = memcg->swap.max;
1577 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1578 max = min(max + swap_max, memsw_max);
1583 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1585 return page_counter_read(&memcg->memory);
1588 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1591 struct oom_control oc = {
1595 .gfp_mask = gfp_mask,
1600 if (mutex_lock_killable(&oom_lock))
1603 * A few threads which were not waiting at mutex_lock_killable() can
1604 * fail to bail out. Therefore, check again after holding oom_lock.
1606 ret = should_force_charge() || out_of_memory(&oc);
1607 mutex_unlock(&oom_lock);
1611 #if MAX_NUMNODES > 1
1614 * test_mem_cgroup_node_reclaimable
1615 * @memcg: the target memcg
1616 * @nid: the node ID to be checked.
1617 * @noswap : specify true here if the user wants flle only information.
1619 * This function returns whether the specified memcg contains any
1620 * reclaimable pages on a node. Returns true if there are any reclaimable
1621 * pages in the node.
1623 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1624 int nid, bool noswap)
1626 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
1628 if (lruvec_page_state(lruvec, NR_INACTIVE_FILE) ||
1629 lruvec_page_state(lruvec, NR_ACTIVE_FILE))
1631 if (noswap || !total_swap_pages)
1633 if (lruvec_page_state(lruvec, NR_INACTIVE_ANON) ||
1634 lruvec_page_state(lruvec, NR_ACTIVE_ANON))
1641 * Always updating the nodemask is not very good - even if we have an empty
1642 * list or the wrong list here, we can start from some node and traverse all
1643 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1646 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1650 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1651 * pagein/pageout changes since the last update.
1653 if (!atomic_read(&memcg->numainfo_events))
1655 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1658 /* make a nodemask where this memcg uses memory from */
1659 memcg->scan_nodes = node_states[N_MEMORY];
1661 for_each_node_mask(nid, node_states[N_MEMORY]) {
1663 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1664 node_clear(nid, memcg->scan_nodes);
1667 atomic_set(&memcg->numainfo_events, 0);
1668 atomic_set(&memcg->numainfo_updating, 0);
1672 * Selecting a node where we start reclaim from. Because what we need is just
1673 * reducing usage counter, start from anywhere is O,K. Considering
1674 * memory reclaim from current node, there are pros. and cons.
1676 * Freeing memory from current node means freeing memory from a node which
1677 * we'll use or we've used. So, it may make LRU bad. And if several threads
1678 * hit limits, it will see a contention on a node. But freeing from remote
1679 * node means more costs for memory reclaim because of memory latency.
1681 * Now, we use round-robin. Better algorithm is welcomed.
1683 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1687 mem_cgroup_may_update_nodemask(memcg);
1688 node = memcg->last_scanned_node;
1690 node = next_node_in(node, memcg->scan_nodes);
1692 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1693 * last time it really checked all the LRUs due to rate limiting.
1694 * Fallback to the current node in that case for simplicity.
1696 if (unlikely(node == MAX_NUMNODES))
1697 node = numa_node_id();
1699 memcg->last_scanned_node = node;
1703 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1709 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1712 unsigned long *total_scanned)
1714 struct mem_cgroup *victim = NULL;
1717 unsigned long excess;
1718 unsigned long nr_scanned;
1719 struct mem_cgroup_reclaim_cookie reclaim = {
1724 excess = soft_limit_excess(root_memcg);
1727 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1732 * If we have not been able to reclaim
1733 * anything, it might because there are
1734 * no reclaimable pages under this hierarchy
1739 * We want to do more targeted reclaim.
1740 * excess >> 2 is not to excessive so as to
1741 * reclaim too much, nor too less that we keep
1742 * coming back to reclaim from this cgroup
1744 if (total >= (excess >> 2) ||
1745 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1750 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1751 pgdat, &nr_scanned);
1752 *total_scanned += nr_scanned;
1753 if (!soft_limit_excess(root_memcg))
1756 mem_cgroup_iter_break(root_memcg, victim);
1760 #ifdef CONFIG_LOCKDEP
1761 static struct lockdep_map memcg_oom_lock_dep_map = {
1762 .name = "memcg_oom_lock",
1766 static DEFINE_SPINLOCK(memcg_oom_lock);
1769 * Check OOM-Killer is already running under our hierarchy.
1770 * If someone is running, return false.
1772 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1774 struct mem_cgroup *iter, *failed = NULL;
1776 spin_lock(&memcg_oom_lock);
1778 for_each_mem_cgroup_tree(iter, memcg) {
1779 if (iter->oom_lock) {
1781 * this subtree of our hierarchy is already locked
1782 * so we cannot give a lock.
1785 mem_cgroup_iter_break(memcg, iter);
1788 iter->oom_lock = true;
1793 * OK, we failed to lock the whole subtree so we have
1794 * to clean up what we set up to the failing subtree
1796 for_each_mem_cgroup_tree(iter, memcg) {
1797 if (iter == failed) {
1798 mem_cgroup_iter_break(memcg, iter);
1801 iter->oom_lock = false;
1804 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1806 spin_unlock(&memcg_oom_lock);
1811 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1813 struct mem_cgroup *iter;
1815 spin_lock(&memcg_oom_lock);
1816 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1817 for_each_mem_cgroup_tree(iter, memcg)
1818 iter->oom_lock = false;
1819 spin_unlock(&memcg_oom_lock);
1822 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1824 struct mem_cgroup *iter;
1826 spin_lock(&memcg_oom_lock);
1827 for_each_mem_cgroup_tree(iter, memcg)
1829 spin_unlock(&memcg_oom_lock);
1832 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1834 struct mem_cgroup *iter;
1837 * When a new child is created while the hierarchy is under oom,
1838 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1840 spin_lock(&memcg_oom_lock);
1841 for_each_mem_cgroup_tree(iter, memcg)
1842 if (iter->under_oom > 0)
1844 spin_unlock(&memcg_oom_lock);
1847 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1849 struct oom_wait_info {
1850 struct mem_cgroup *memcg;
1851 wait_queue_entry_t wait;
1854 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1855 unsigned mode, int sync, void *arg)
1857 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1858 struct mem_cgroup *oom_wait_memcg;
1859 struct oom_wait_info *oom_wait_info;
1861 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1862 oom_wait_memcg = oom_wait_info->memcg;
1864 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1865 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1867 return autoremove_wake_function(wait, mode, sync, arg);
1870 static void memcg_oom_recover(struct mem_cgroup *memcg)
1873 * For the following lockless ->under_oom test, the only required
1874 * guarantee is that it must see the state asserted by an OOM when
1875 * this function is called as a result of userland actions
1876 * triggered by the notification of the OOM. This is trivially
1877 * achieved by invoking mem_cgroup_mark_under_oom() before
1878 * triggering notification.
1880 if (memcg && memcg->under_oom)
1881 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1891 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1893 enum oom_status ret;
1896 if (order > PAGE_ALLOC_COSTLY_ORDER)
1899 memcg_memory_event(memcg, MEMCG_OOM);
1902 * We are in the middle of the charge context here, so we
1903 * don't want to block when potentially sitting on a callstack
1904 * that holds all kinds of filesystem and mm locks.
1906 * cgroup1 allows disabling the OOM killer and waiting for outside
1907 * handling until the charge can succeed; remember the context and put
1908 * the task to sleep at the end of the page fault when all locks are
1911 * On the other hand, in-kernel OOM killer allows for an async victim
1912 * memory reclaim (oom_reaper) and that means that we are not solely
1913 * relying on the oom victim to make a forward progress and we can
1914 * invoke the oom killer here.
1916 * Please note that mem_cgroup_out_of_memory might fail to find a
1917 * victim and then we have to bail out from the charge path.
1919 if (memcg->oom_kill_disable) {
1920 if (!current->in_user_fault)
1922 css_get(&memcg->css);
1923 current->memcg_in_oom = memcg;
1924 current->memcg_oom_gfp_mask = mask;
1925 current->memcg_oom_order = order;
1930 mem_cgroup_mark_under_oom(memcg);
1932 locked = mem_cgroup_oom_trylock(memcg);
1935 mem_cgroup_oom_notify(memcg);
1937 mem_cgroup_unmark_under_oom(memcg);
1938 if (mem_cgroup_out_of_memory(memcg, mask, order))
1944 mem_cgroup_oom_unlock(memcg);
1950 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1951 * @handle: actually kill/wait or just clean up the OOM state
1953 * This has to be called at the end of a page fault if the memcg OOM
1954 * handler was enabled.
1956 * Memcg supports userspace OOM handling where failed allocations must
1957 * sleep on a waitqueue until the userspace task resolves the
1958 * situation. Sleeping directly in the charge context with all kinds
1959 * of locks held is not a good idea, instead we remember an OOM state
1960 * in the task and mem_cgroup_oom_synchronize() has to be called at
1961 * the end of the page fault to complete the OOM handling.
1963 * Returns %true if an ongoing memcg OOM situation was detected and
1964 * completed, %false otherwise.
1966 bool mem_cgroup_oom_synchronize(bool handle)
1968 struct mem_cgroup *memcg = current->memcg_in_oom;
1969 struct oom_wait_info owait;
1972 /* OOM is global, do not handle */
1979 owait.memcg = memcg;
1980 owait.wait.flags = 0;
1981 owait.wait.func = memcg_oom_wake_function;
1982 owait.wait.private = current;
1983 INIT_LIST_HEAD(&owait.wait.entry);
1985 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1986 mem_cgroup_mark_under_oom(memcg);
1988 locked = mem_cgroup_oom_trylock(memcg);
1991 mem_cgroup_oom_notify(memcg);
1993 if (locked && !memcg->oom_kill_disable) {
1994 mem_cgroup_unmark_under_oom(memcg);
1995 finish_wait(&memcg_oom_waitq, &owait.wait);
1996 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1997 current->memcg_oom_order);
2000 mem_cgroup_unmark_under_oom(memcg);
2001 finish_wait(&memcg_oom_waitq, &owait.wait);
2005 mem_cgroup_oom_unlock(memcg);
2007 * There is no guarantee that an OOM-lock contender
2008 * sees the wakeups triggered by the OOM kill
2009 * uncharges. Wake any sleepers explicitely.
2011 memcg_oom_recover(memcg);
2014 current->memcg_in_oom = NULL;
2015 css_put(&memcg->css);
2020 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2021 * @victim: task to be killed by the OOM killer
2022 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2024 * Returns a pointer to a memory cgroup, which has to be cleaned up
2025 * by killing all belonging OOM-killable tasks.
2027 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2029 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2030 struct mem_cgroup *oom_domain)
2032 struct mem_cgroup *oom_group = NULL;
2033 struct mem_cgroup *memcg;
2035 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2039 oom_domain = root_mem_cgroup;
2043 memcg = mem_cgroup_from_task(victim);
2044 if (memcg == root_mem_cgroup)
2048 * Traverse the memory cgroup hierarchy from the victim task's
2049 * cgroup up to the OOMing cgroup (or root) to find the
2050 * highest-level memory cgroup with oom.group set.
2052 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2053 if (memcg->oom_group)
2056 if (memcg == oom_domain)
2061 css_get(&oom_group->css);
2068 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2070 pr_info("Tasks in ");
2071 pr_cont_cgroup_path(memcg->css.cgroup);
2072 pr_cont(" are going to be killed due to memory.oom.group set\n");
2076 * lock_page_memcg - lock a page->mem_cgroup binding
2079 * This function protects unlocked LRU pages from being moved to
2082 * It ensures lifetime of the returned memcg. Caller is responsible
2083 * for the lifetime of the page; __unlock_page_memcg() is available
2084 * when @page might get freed inside the locked section.
2086 struct mem_cgroup *lock_page_memcg(struct page *page)
2088 struct mem_cgroup *memcg;
2089 unsigned long flags;
2092 * The RCU lock is held throughout the transaction. The fast
2093 * path can get away without acquiring the memcg->move_lock
2094 * because page moving starts with an RCU grace period.
2096 * The RCU lock also protects the memcg from being freed when
2097 * the page state that is going to change is the only thing
2098 * preventing the page itself from being freed. E.g. writeback
2099 * doesn't hold a page reference and relies on PG_writeback to
2100 * keep off truncation, migration and so forth.
2104 if (mem_cgroup_disabled())
2107 memcg = page->mem_cgroup;
2108 if (unlikely(!memcg))
2111 if (atomic_read(&memcg->moving_account) <= 0)
2114 spin_lock_irqsave(&memcg->move_lock, flags);
2115 if (memcg != page->mem_cgroup) {
2116 spin_unlock_irqrestore(&memcg->move_lock, flags);
2121 * When charge migration first begins, we can have locked and
2122 * unlocked page stat updates happening concurrently. Track
2123 * the task who has the lock for unlock_page_memcg().
2125 memcg->move_lock_task = current;
2126 memcg->move_lock_flags = flags;
2130 EXPORT_SYMBOL(lock_page_memcg);
2133 * __unlock_page_memcg - unlock and unpin a memcg
2136 * Unlock and unpin a memcg returned by lock_page_memcg().
2138 void __unlock_page_memcg(struct mem_cgroup *memcg)
2140 if (memcg && memcg->move_lock_task == current) {
2141 unsigned long flags = memcg->move_lock_flags;
2143 memcg->move_lock_task = NULL;
2144 memcg->move_lock_flags = 0;
2146 spin_unlock_irqrestore(&memcg->move_lock, flags);
2153 * unlock_page_memcg - unlock a page->mem_cgroup binding
2156 void unlock_page_memcg(struct page *page)
2158 __unlock_page_memcg(page->mem_cgroup);
2160 EXPORT_SYMBOL(unlock_page_memcg);
2162 struct memcg_stock_pcp {
2163 struct mem_cgroup *cached; /* this never be root cgroup */
2164 unsigned int nr_pages;
2165 struct work_struct work;
2166 unsigned long flags;
2167 #define FLUSHING_CACHED_CHARGE 0
2169 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2170 static DEFINE_MUTEX(percpu_charge_mutex);
2173 * consume_stock: Try to consume stocked charge on this cpu.
2174 * @memcg: memcg to consume from.
2175 * @nr_pages: how many pages to charge.
2177 * The charges will only happen if @memcg matches the current cpu's memcg
2178 * stock, and at least @nr_pages are available in that stock. Failure to
2179 * service an allocation will refill the stock.
2181 * returns true if successful, false otherwise.
2183 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2185 struct memcg_stock_pcp *stock;
2186 unsigned long flags;
2189 if (nr_pages > MEMCG_CHARGE_BATCH)
2192 local_irq_save(flags);
2194 stock = this_cpu_ptr(&memcg_stock);
2195 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2196 stock->nr_pages -= nr_pages;
2200 local_irq_restore(flags);
2206 * Returns stocks cached in percpu and reset cached information.
2208 static void drain_stock(struct memcg_stock_pcp *stock)
2210 struct mem_cgroup *old = stock->cached;
2212 if (stock->nr_pages) {
2213 page_counter_uncharge(&old->memory, stock->nr_pages);
2214 if (do_memsw_account())
2215 page_counter_uncharge(&old->memsw, stock->nr_pages);
2216 css_put_many(&old->css, stock->nr_pages);
2217 stock->nr_pages = 0;
2219 stock->cached = NULL;
2222 static void drain_local_stock(struct work_struct *dummy)
2224 struct memcg_stock_pcp *stock;
2225 unsigned long flags;
2228 * The only protection from memory hotplug vs. drain_stock races is
2229 * that we always operate on local CPU stock here with IRQ disabled
2231 local_irq_save(flags);
2233 stock = this_cpu_ptr(&memcg_stock);
2235 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2237 local_irq_restore(flags);
2241 * Cache charges(val) to local per_cpu area.
2242 * This will be consumed by consume_stock() function, later.
2244 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2246 struct memcg_stock_pcp *stock;
2247 unsigned long flags;
2249 local_irq_save(flags);
2251 stock = this_cpu_ptr(&memcg_stock);
2252 if (stock->cached != memcg) { /* reset if necessary */
2254 stock->cached = memcg;
2256 stock->nr_pages += nr_pages;
2258 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2261 local_irq_restore(flags);
2265 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2266 * of the hierarchy under it.
2268 static void drain_all_stock(struct mem_cgroup *root_memcg)
2272 /* If someone's already draining, avoid adding running more workers. */
2273 if (!mutex_trylock(&percpu_charge_mutex))
2276 * Notify other cpus that system-wide "drain" is running
2277 * We do not care about races with the cpu hotplug because cpu down
2278 * as well as workers from this path always operate on the local
2279 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2282 for_each_online_cpu(cpu) {
2283 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2284 struct mem_cgroup *memcg;
2288 memcg = stock->cached;
2289 if (memcg && stock->nr_pages &&
2290 mem_cgroup_is_descendant(memcg, root_memcg))
2295 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2297 drain_local_stock(&stock->work);
2299 schedule_work_on(cpu, &stock->work);
2303 mutex_unlock(&percpu_charge_mutex);
2306 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2308 struct memcg_stock_pcp *stock;
2309 struct mem_cgroup *memcg, *mi;
2311 stock = &per_cpu(memcg_stock, cpu);
2314 for_each_mem_cgroup(memcg) {
2317 for (i = 0; i < MEMCG_NR_STAT; i++) {
2321 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2323 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2324 atomic_long_add(x, &memcg->vmstats[i]);
2326 if (i >= NR_VM_NODE_STAT_ITEMS)
2329 for_each_node(nid) {
2330 struct mem_cgroup_per_node *pn;
2332 pn = mem_cgroup_nodeinfo(memcg, nid);
2333 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2336 atomic_long_add(x, &pn->lruvec_stat[i]);
2337 } while ((pn = parent_nodeinfo(pn, nid)));
2341 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2344 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2346 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2347 atomic_long_add(x, &memcg->vmevents[i]);
2354 static void reclaim_high(struct mem_cgroup *memcg,
2355 unsigned int nr_pages,
2359 if (page_counter_read(&memcg->memory) <= memcg->high)
2361 memcg_memory_event(memcg, MEMCG_HIGH);
2362 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2363 } while ((memcg = parent_mem_cgroup(memcg)));
2366 static void high_work_func(struct work_struct *work)
2368 struct mem_cgroup *memcg;
2370 memcg = container_of(work, struct mem_cgroup, high_work);
2371 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2375 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2376 * enough to still cause a significant slowdown in most cases, while still
2377 * allowing diagnostics and tracing to proceed without becoming stuck.
2379 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2382 * When calculating the delay, we use these either side of the exponentiation to
2383 * maintain precision and scale to a reasonable number of jiffies (see the table
2386 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2387 * overage ratio to a delay.
2388 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down down the
2389 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2390 * to produce a reasonable delay curve.
2392 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2393 * reasonable delay curve compared to precision-adjusted overage, not
2394 * penalising heavily at first, but still making sure that growth beyond the
2395 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2396 * example, with a high of 100 megabytes:
2398 * +-------+------------------------+
2399 * | usage | time to allocate in ms |
2400 * +-------+------------------------+
2422 * +-------+------------------------+
2424 #define MEMCG_DELAY_PRECISION_SHIFT 20
2425 #define MEMCG_DELAY_SCALING_SHIFT 14
2428 * Get the number of jiffies that we should penalise a mischievous cgroup which
2429 * is exceeding its memory.high by checking both it and its ancestors.
2431 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2432 unsigned int nr_pages)
2434 unsigned long penalty_jiffies;
2435 u64 max_overage = 0;
2438 unsigned long usage, high;
2441 usage = page_counter_read(&memcg->memory);
2442 high = READ_ONCE(memcg->high);
2448 * Prevent division by 0 in overage calculation by acting as if
2449 * it was a threshold of 1 page
2451 high = max(high, 1UL);
2453 overage = usage - high;
2454 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2455 overage = div64_u64(overage, high);
2457 if (overage > max_overage)
2458 max_overage = overage;
2459 } while ((memcg = parent_mem_cgroup(memcg)) &&
2460 !mem_cgroup_is_root(memcg));
2466 * We use overage compared to memory.high to calculate the number of
2467 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2468 * fairly lenient on small overages, and increasingly harsh when the
2469 * memcg in question makes it clear that it has no intention of stopping
2470 * its crazy behaviour, so we exponentially increase the delay based on
2473 penalty_jiffies = max_overage * max_overage * HZ;
2474 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2475 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2478 * Factor in the task's own contribution to the overage, such that four
2479 * N-sized allocations are throttled approximately the same as one
2480 * 4N-sized allocation.
2482 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2483 * larger the current charge patch is than that.
2485 penalty_jiffies = penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2488 * Clamp the max delay per usermode return so as to still keep the
2489 * application moving forwards and also permit diagnostics, albeit
2492 return min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2496 * Scheduled by try_charge() to be executed from the userland return path
2497 * and reclaims memory over the high limit.
2499 void mem_cgroup_handle_over_high(void)
2501 unsigned long penalty_jiffies;
2502 unsigned long pflags;
2503 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2504 struct mem_cgroup *memcg;
2506 if (likely(!nr_pages))
2509 memcg = get_mem_cgroup_from_mm(current->mm);
2510 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2511 current->memcg_nr_pages_over_high = 0;
2514 * memory.high is breached and reclaim is unable to keep up. Throttle
2515 * allocators proactively to slow down excessive growth.
2517 penalty_jiffies = calculate_high_delay(memcg, nr_pages);
2520 * Don't sleep if the amount of jiffies this memcg owes us is so low
2521 * that it's not even worth doing, in an attempt to be nice to those who
2522 * go only a small amount over their memory.high value and maybe haven't
2523 * been aggressively reclaimed enough yet.
2525 if (penalty_jiffies <= HZ / 100)
2529 * If we exit early, we're guaranteed to die (since
2530 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2531 * need to account for any ill-begotten jiffies to pay them off later.
2533 psi_memstall_enter(&pflags);
2534 schedule_timeout_killable(penalty_jiffies);
2535 psi_memstall_leave(&pflags);
2538 css_put(&memcg->css);
2541 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2542 unsigned int nr_pages)
2544 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2545 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2546 struct mem_cgroup *mem_over_limit;
2547 struct page_counter *counter;
2548 unsigned long nr_reclaimed;
2549 bool may_swap = true;
2550 bool drained = false;
2551 enum oom_status oom_status;
2553 if (mem_cgroup_is_root(memcg))
2556 if (consume_stock(memcg, nr_pages))
2559 if (!do_memsw_account() ||
2560 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2561 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2563 if (do_memsw_account())
2564 page_counter_uncharge(&memcg->memsw, batch);
2565 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2567 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2571 if (batch > nr_pages) {
2577 * Memcg doesn't have a dedicated reserve for atomic
2578 * allocations. But like the global atomic pool, we need to
2579 * put the burden of reclaim on regular allocation requests
2580 * and let these go through as privileged allocations.
2582 if (gfp_mask & __GFP_ATOMIC)
2586 * Unlike in global OOM situations, memcg is not in a physical
2587 * memory shortage. Allow dying and OOM-killed tasks to
2588 * bypass the last charges so that they can exit quickly and
2589 * free their memory.
2591 if (unlikely(should_force_charge()))
2595 * Prevent unbounded recursion when reclaim operations need to
2596 * allocate memory. This might exceed the limits temporarily,
2597 * but we prefer facilitating memory reclaim and getting back
2598 * under the limit over triggering OOM kills in these cases.
2600 if (unlikely(current->flags & PF_MEMALLOC))
2603 if (unlikely(task_in_memcg_oom(current)))
2606 if (!gfpflags_allow_blocking(gfp_mask))
2609 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2611 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2612 gfp_mask, may_swap);
2614 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2618 drain_all_stock(mem_over_limit);
2623 if (gfp_mask & __GFP_NORETRY)
2626 * Even though the limit is exceeded at this point, reclaim
2627 * may have been able to free some pages. Retry the charge
2628 * before killing the task.
2630 * Only for regular pages, though: huge pages are rather
2631 * unlikely to succeed so close to the limit, and we fall back
2632 * to regular pages anyway in case of failure.
2634 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2637 * At task move, charge accounts can be doubly counted. So, it's
2638 * better to wait until the end of task_move if something is going on.
2640 if (mem_cgroup_wait_acct_move(mem_over_limit))
2646 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2649 if (gfp_mask & __GFP_NOFAIL)
2652 if (fatal_signal_pending(current))
2656 * keep retrying as long as the memcg oom killer is able to make
2657 * a forward progress or bypass the charge if the oom killer
2658 * couldn't make any progress.
2660 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2661 get_order(nr_pages * PAGE_SIZE));
2662 switch (oom_status) {
2664 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2672 if (!(gfp_mask & __GFP_NOFAIL))
2676 * The allocation either can't fail or will lead to more memory
2677 * being freed very soon. Allow memory usage go over the limit
2678 * temporarily by force charging it.
2680 page_counter_charge(&memcg->memory, nr_pages);
2681 if (do_memsw_account())
2682 page_counter_charge(&memcg->memsw, nr_pages);
2683 css_get_many(&memcg->css, nr_pages);
2688 css_get_many(&memcg->css, batch);
2689 if (batch > nr_pages)
2690 refill_stock(memcg, batch - nr_pages);
2693 * If the hierarchy is above the normal consumption range, schedule
2694 * reclaim on returning to userland. We can perform reclaim here
2695 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2696 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2697 * not recorded as it most likely matches current's and won't
2698 * change in the meantime. As high limit is checked again before
2699 * reclaim, the cost of mismatch is negligible.
2702 if (page_counter_read(&memcg->memory) > memcg->high) {
2703 /* Don't bother a random interrupted task */
2704 if (in_interrupt()) {
2705 schedule_work(&memcg->high_work);
2708 current->memcg_nr_pages_over_high += batch;
2709 set_notify_resume(current);
2712 } while ((memcg = parent_mem_cgroup(memcg)));
2717 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2719 if (mem_cgroup_is_root(memcg))
2722 page_counter_uncharge(&memcg->memory, nr_pages);
2723 if (do_memsw_account())
2724 page_counter_uncharge(&memcg->memsw, nr_pages);
2726 css_put_many(&memcg->css, nr_pages);
2729 static void lock_page_lru(struct page *page, int *isolated)
2731 pg_data_t *pgdat = page_pgdat(page);
2733 spin_lock_irq(&pgdat->lru_lock);
2734 if (PageLRU(page)) {
2735 struct lruvec *lruvec;
2737 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2739 del_page_from_lru_list(page, lruvec, page_lru(page));
2745 static void unlock_page_lru(struct page *page, int isolated)
2747 pg_data_t *pgdat = page_pgdat(page);
2750 struct lruvec *lruvec;
2752 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2753 VM_BUG_ON_PAGE(PageLRU(page), page);
2755 add_page_to_lru_list(page, lruvec, page_lru(page));
2757 spin_unlock_irq(&pgdat->lru_lock);
2760 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2765 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2768 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2769 * may already be on some other mem_cgroup's LRU. Take care of it.
2772 lock_page_lru(page, &isolated);
2775 * Nobody should be changing or seriously looking at
2776 * page->mem_cgroup at this point:
2778 * - the page is uncharged
2780 * - the page is off-LRU
2782 * - an anonymous fault has exclusive page access, except for
2783 * a locked page table
2785 * - a page cache insertion, a swapin fault, or a migration
2786 * have the page locked
2788 page->mem_cgroup = memcg;
2791 unlock_page_lru(page, isolated);
2794 #ifdef CONFIG_MEMCG_KMEM
2796 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2798 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2799 * cgroup_mutex, etc.
2801 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2805 if (mem_cgroup_disabled())
2808 page = virt_to_head_page(p);
2811 * Slab pages don't have page->mem_cgroup set because corresponding
2812 * kmem caches can be reparented during the lifetime. That's why
2813 * memcg_from_slab_page() should be used instead.
2816 return memcg_from_slab_page(page);
2818 /* All other pages use page->mem_cgroup */
2819 return page->mem_cgroup;
2822 static int memcg_alloc_cache_id(void)
2827 id = ida_simple_get(&memcg_cache_ida,
2828 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2832 if (id < memcg_nr_cache_ids)
2836 * There's no space for the new id in memcg_caches arrays,
2837 * so we have to grow them.
2839 down_write(&memcg_cache_ids_sem);
2841 size = 2 * (id + 1);
2842 if (size < MEMCG_CACHES_MIN_SIZE)
2843 size = MEMCG_CACHES_MIN_SIZE;
2844 else if (size > MEMCG_CACHES_MAX_SIZE)
2845 size = MEMCG_CACHES_MAX_SIZE;
2847 err = memcg_update_all_caches(size);
2849 err = memcg_update_all_list_lrus(size);
2851 memcg_nr_cache_ids = size;
2853 up_write(&memcg_cache_ids_sem);
2856 ida_simple_remove(&memcg_cache_ida, id);
2862 static void memcg_free_cache_id(int id)
2864 ida_simple_remove(&memcg_cache_ida, id);
2867 struct memcg_kmem_cache_create_work {
2868 struct mem_cgroup *memcg;
2869 struct kmem_cache *cachep;
2870 struct work_struct work;
2873 static void memcg_kmem_cache_create_func(struct work_struct *w)
2875 struct memcg_kmem_cache_create_work *cw =
2876 container_of(w, struct memcg_kmem_cache_create_work, work);
2877 struct mem_cgroup *memcg = cw->memcg;
2878 struct kmem_cache *cachep = cw->cachep;
2880 memcg_create_kmem_cache(memcg, cachep);
2882 css_put(&memcg->css);
2887 * Enqueue the creation of a per-memcg kmem_cache.
2889 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2890 struct kmem_cache *cachep)
2892 struct memcg_kmem_cache_create_work *cw;
2894 if (!css_tryget_online(&memcg->css))
2897 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2899 css_put(&memcg->css);
2904 cw->cachep = cachep;
2905 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2907 queue_work(memcg_kmem_cache_wq, &cw->work);
2910 static inline bool memcg_kmem_bypass(void)
2912 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2918 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2919 * @cachep: the original global kmem cache
2921 * Return the kmem_cache we're supposed to use for a slab allocation.
2922 * We try to use the current memcg's version of the cache.
2924 * If the cache does not exist yet, if we are the first user of it, we
2925 * create it asynchronously in a workqueue and let the current allocation
2926 * go through with the original cache.
2928 * This function takes a reference to the cache it returns to assure it
2929 * won't get destroyed while we are working with it. Once the caller is
2930 * done with it, memcg_kmem_put_cache() must be called to release the
2933 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2935 struct mem_cgroup *memcg;
2936 struct kmem_cache *memcg_cachep;
2937 struct memcg_cache_array *arr;
2940 VM_BUG_ON(!is_root_cache(cachep));
2942 if (memcg_kmem_bypass())
2947 if (unlikely(current->active_memcg))
2948 memcg = current->active_memcg;
2950 memcg = mem_cgroup_from_task(current);
2952 if (!memcg || memcg == root_mem_cgroup)
2955 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2959 arr = rcu_dereference(cachep->memcg_params.memcg_caches);
2962 * Make sure we will access the up-to-date value. The code updating
2963 * memcg_caches issues a write barrier to match the data dependency
2964 * barrier inside READ_ONCE() (see memcg_create_kmem_cache()).
2966 memcg_cachep = READ_ONCE(arr->entries[kmemcg_id]);
2969 * If we are in a safe context (can wait, and not in interrupt
2970 * context), we could be be predictable and return right away.
2971 * This would guarantee that the allocation being performed
2972 * already belongs in the new cache.
2974 * However, there are some clashes that can arrive from locking.
2975 * For instance, because we acquire the slab_mutex while doing
2976 * memcg_create_kmem_cache, this means no further allocation
2977 * could happen with the slab_mutex held. So it's better to
2980 * If the memcg is dying or memcg_cache is about to be released,
2981 * don't bother creating new kmem_caches. Because memcg_cachep
2982 * is ZEROed as the fist step of kmem offlining, we don't need
2983 * percpu_ref_tryget_live() here. css_tryget_online() check in
2984 * memcg_schedule_kmem_cache_create() will prevent us from
2985 * creation of a new kmem_cache.
2987 if (unlikely(!memcg_cachep))
2988 memcg_schedule_kmem_cache_create(memcg, cachep);
2989 else if (percpu_ref_tryget(&memcg_cachep->memcg_params.refcnt))
2990 cachep = memcg_cachep;
2997 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2998 * @cachep: the cache returned by memcg_kmem_get_cache
3000 void memcg_kmem_put_cache(struct kmem_cache *cachep)
3002 if (!is_root_cache(cachep))
3003 percpu_ref_put(&cachep->memcg_params.refcnt);
3007 * __memcg_kmem_charge_memcg: charge a kmem page
3008 * @page: page to charge
3009 * @gfp: reclaim mode
3010 * @order: allocation order
3011 * @memcg: memory cgroup to charge
3013 * Returns 0 on success, an error code on failure.
3015 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
3016 struct mem_cgroup *memcg)
3018 unsigned int nr_pages = 1 << order;
3019 struct page_counter *counter;
3022 ret = try_charge(memcg, gfp, nr_pages);
3026 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3027 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3030 * Enforce __GFP_NOFAIL allocation because callers are not
3031 * prepared to see failures and likely do not have any failure
3034 if (gfp & __GFP_NOFAIL) {
3035 page_counter_charge(&memcg->kmem, nr_pages);
3038 cancel_charge(memcg, nr_pages);
3045 * __memcg_kmem_charge: charge a kmem page to the current memory cgroup
3046 * @page: page to charge
3047 * @gfp: reclaim mode
3048 * @order: allocation order
3050 * Returns 0 on success, an error code on failure.
3052 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
3054 struct mem_cgroup *memcg;
3057 if (memcg_kmem_bypass())
3060 memcg = get_mem_cgroup_from_current();
3061 if (!mem_cgroup_is_root(memcg)) {
3062 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
3064 page->mem_cgroup = memcg;
3065 __SetPageKmemcg(page);
3068 css_put(&memcg->css);
3073 * __memcg_kmem_uncharge_memcg: uncharge a kmem page
3074 * @memcg: memcg to uncharge
3075 * @nr_pages: number of pages to uncharge
3077 void __memcg_kmem_uncharge_memcg(struct mem_cgroup *memcg,
3078 unsigned int nr_pages)
3080 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3081 page_counter_uncharge(&memcg->kmem, nr_pages);
3083 page_counter_uncharge(&memcg->memory, nr_pages);
3084 if (do_memsw_account())
3085 page_counter_uncharge(&memcg->memsw, nr_pages);
3088 * __memcg_kmem_uncharge: uncharge a kmem page
3089 * @page: page to uncharge
3090 * @order: allocation order
3092 void __memcg_kmem_uncharge(struct page *page, int order)
3094 struct mem_cgroup *memcg = page->mem_cgroup;
3095 unsigned int nr_pages = 1 << order;
3100 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3101 __memcg_kmem_uncharge_memcg(memcg, nr_pages);
3102 page->mem_cgroup = NULL;
3104 /* slab pages do not have PageKmemcg flag set */
3105 if (PageKmemcg(page))
3106 __ClearPageKmemcg(page);
3108 css_put_many(&memcg->css, nr_pages);
3110 #endif /* CONFIG_MEMCG_KMEM */
3112 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3115 * Because tail pages are not marked as "used", set it. We're under
3116 * pgdat->lru_lock and migration entries setup in all page mappings.
3118 void mem_cgroup_split_huge_fixup(struct page *head)
3122 if (mem_cgroup_disabled())
3125 for (i = 1; i < HPAGE_PMD_NR; i++)
3126 head[i].mem_cgroup = head->mem_cgroup;
3128 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
3130 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3132 #ifdef CONFIG_MEMCG_SWAP
3134 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3135 * @entry: swap entry to be moved
3136 * @from: mem_cgroup which the entry is moved from
3137 * @to: mem_cgroup which the entry is moved to
3139 * It succeeds only when the swap_cgroup's record for this entry is the same
3140 * as the mem_cgroup's id of @from.
3142 * Returns 0 on success, -EINVAL on failure.
3144 * The caller must have charged to @to, IOW, called page_counter_charge() about
3145 * both res and memsw, and called css_get().
3147 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3148 struct mem_cgroup *from, struct mem_cgroup *to)
3150 unsigned short old_id, new_id;
3152 old_id = mem_cgroup_id(from);
3153 new_id = mem_cgroup_id(to);
3155 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3156 mod_memcg_state(from, MEMCG_SWAP, -1);
3157 mod_memcg_state(to, MEMCG_SWAP, 1);
3163 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3164 struct mem_cgroup *from, struct mem_cgroup *to)
3170 static DEFINE_MUTEX(memcg_max_mutex);
3172 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3173 unsigned long max, bool memsw)
3175 bool enlarge = false;
3176 bool drained = false;
3178 bool limits_invariant;
3179 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3182 if (signal_pending(current)) {
3187 mutex_lock(&memcg_max_mutex);
3189 * Make sure that the new limit (memsw or memory limit) doesn't
3190 * break our basic invariant rule memory.max <= memsw.max.
3192 limits_invariant = memsw ? max >= memcg->memory.max :
3193 max <= memcg->memsw.max;
3194 if (!limits_invariant) {
3195 mutex_unlock(&memcg_max_mutex);
3199 if (max > counter->max)
3201 ret = page_counter_set_max(counter, max);
3202 mutex_unlock(&memcg_max_mutex);
3208 drain_all_stock(memcg);
3213 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3214 GFP_KERNEL, !memsw)) {
3220 if (!ret && enlarge)
3221 memcg_oom_recover(memcg);
3226 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3228 unsigned long *total_scanned)
3230 unsigned long nr_reclaimed = 0;
3231 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3232 unsigned long reclaimed;
3234 struct mem_cgroup_tree_per_node *mctz;
3235 unsigned long excess;
3236 unsigned long nr_scanned;
3241 mctz = soft_limit_tree_node(pgdat->node_id);
3244 * Do not even bother to check the largest node if the root
3245 * is empty. Do it lockless to prevent lock bouncing. Races
3246 * are acceptable as soft limit is best effort anyway.
3248 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3252 * This loop can run a while, specially if mem_cgroup's continuously
3253 * keep exceeding their soft limit and putting the system under
3260 mz = mem_cgroup_largest_soft_limit_node(mctz);
3265 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3266 gfp_mask, &nr_scanned);
3267 nr_reclaimed += reclaimed;
3268 *total_scanned += nr_scanned;
3269 spin_lock_irq(&mctz->lock);
3270 __mem_cgroup_remove_exceeded(mz, mctz);
3273 * If we failed to reclaim anything from this memory cgroup
3274 * it is time to move on to the next cgroup
3278 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3280 excess = soft_limit_excess(mz->memcg);
3282 * One school of thought says that we should not add
3283 * back the node to the tree if reclaim returns 0.
3284 * But our reclaim could return 0, simply because due
3285 * to priority we are exposing a smaller subset of
3286 * memory to reclaim from. Consider this as a longer
3289 /* If excess == 0, no tree ops */
3290 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3291 spin_unlock_irq(&mctz->lock);
3292 css_put(&mz->memcg->css);
3295 * Could not reclaim anything and there are no more
3296 * mem cgroups to try or we seem to be looping without
3297 * reclaiming anything.
3299 if (!nr_reclaimed &&
3301 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3303 } while (!nr_reclaimed);
3305 css_put(&next_mz->memcg->css);
3306 return nr_reclaimed;
3310 * Test whether @memcg has children, dead or alive. Note that this
3311 * function doesn't care whether @memcg has use_hierarchy enabled and
3312 * returns %true if there are child csses according to the cgroup
3313 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3315 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3320 ret = css_next_child(NULL, &memcg->css);
3326 * Reclaims as many pages from the given memcg as possible.
3328 * Caller is responsible for holding css reference for memcg.
3330 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3332 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3334 /* we call try-to-free pages for make this cgroup empty */
3335 lru_add_drain_all();
3337 drain_all_stock(memcg);
3339 /* try to free all pages in this cgroup */
3340 while (nr_retries && page_counter_read(&memcg->memory)) {
3343 if (signal_pending(current))
3346 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3350 /* maybe some writeback is necessary */
3351 congestion_wait(BLK_RW_ASYNC, HZ/10);
3359 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3360 char *buf, size_t nbytes,
3363 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3365 if (mem_cgroup_is_root(memcg))
3367 return mem_cgroup_force_empty(memcg) ?: nbytes;
3370 #ifdef CONFIG_MEMCG_SWAP
3371 static int mem_cgroup_force_reclaim(struct cgroup_subsys_state *css,
3372 struct cftype *cft, u64 val)
3374 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3375 unsigned long nr_to_reclaim = val;
3376 unsigned long total = 0;
3379 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
3380 total += try_to_free_mem_cgroup_pages(memcg, nr_to_reclaim,
3384 * If nothing was reclaimed after two attempts, there
3385 * may be no reclaimable pages in this hierarchy.
3386 * If more than nr_to_reclaim pages were already reclaimed,
3387 * finish force reclaim.
3389 if (loop && (!total || total > nr_to_reclaim))
3393 pr_info("%s: [Mem_reclaim] Loop: %d - Total_reclaimed: %lu - nr_to_reclaim: %lu\n",
3394 __func__, loop, total, nr_to_reclaim);
3400 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3403 return mem_cgroup_from_css(css)->use_hierarchy;
3406 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3407 struct cftype *cft, u64 val)
3410 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3411 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3413 if (memcg->use_hierarchy == val)
3417 * If parent's use_hierarchy is set, we can't make any modifications
3418 * in the child subtrees. If it is unset, then the change can
3419 * occur, provided the current cgroup has no children.
3421 * For the root cgroup, parent_mem is NULL, we allow value to be
3422 * set if there are no children.
3424 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3425 (val == 1 || val == 0)) {
3426 if (!memcg_has_children(memcg))
3427 memcg->use_hierarchy = val;
3436 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3440 if (mem_cgroup_is_root(memcg)) {
3441 val = memcg_page_state(memcg, MEMCG_CACHE) +
3442 memcg_page_state(memcg, MEMCG_RSS);
3444 val += memcg_page_state(memcg, MEMCG_SWAP);
3447 val = page_counter_read(&memcg->memory);
3449 val = page_counter_read(&memcg->memsw);
3462 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3465 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3466 struct page_counter *counter;
3468 switch (MEMFILE_TYPE(cft->private)) {
3470 counter = &memcg->memory;
3473 counter = &memcg->memsw;
3476 counter = &memcg->kmem;
3479 counter = &memcg->tcpmem;
3485 switch (MEMFILE_ATTR(cft->private)) {
3487 if (counter == &memcg->memory)
3488 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3489 if (counter == &memcg->memsw)
3490 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3491 return (u64)page_counter_read(counter) * PAGE_SIZE;
3493 return (u64)counter->max * PAGE_SIZE;
3495 return (u64)counter->watermark * PAGE_SIZE;
3497 return counter->failcnt;
3498 case RES_SOFT_LIMIT:
3499 return (u64)memcg->soft_limit * PAGE_SIZE;
3505 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3507 unsigned long stat[MEMCG_NR_STAT] = {0};
3508 struct mem_cgroup *mi;
3511 for_each_online_cpu(cpu)
3512 for (i = 0; i < MEMCG_NR_STAT; i++)
3513 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3515 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3516 for (i = 0; i < MEMCG_NR_STAT; i++)
3517 atomic_long_add(stat[i], &mi->vmstats[i]);
3519 for_each_node(node) {
3520 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3521 struct mem_cgroup_per_node *pi;
3523 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3526 for_each_online_cpu(cpu)
3527 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3529 pn->lruvec_stat_cpu->count[i], cpu);
3531 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3532 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3533 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3537 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3539 unsigned long events[NR_VM_EVENT_ITEMS];
3540 struct mem_cgroup *mi;
3543 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3546 for_each_online_cpu(cpu)
3547 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3548 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3551 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3552 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3553 atomic_long_add(events[i], &mi->vmevents[i]);
3556 #ifdef CONFIG_MEMCG_KMEM
3557 static int memcg_online_kmem(struct mem_cgroup *memcg)
3561 if (cgroup_memory_nokmem)
3564 BUG_ON(memcg->kmemcg_id >= 0);
3565 BUG_ON(memcg->kmem_state);
3567 memcg_id = memcg_alloc_cache_id();
3571 static_branch_inc(&memcg_kmem_enabled_key);
3573 * A memory cgroup is considered kmem-online as soon as it gets
3574 * kmemcg_id. Setting the id after enabling static branching will
3575 * guarantee no one starts accounting before all call sites are
3578 memcg->kmemcg_id = memcg_id;
3579 memcg->kmem_state = KMEM_ONLINE;
3580 INIT_LIST_HEAD(&memcg->kmem_caches);
3585 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3587 struct cgroup_subsys_state *css;
3588 struct mem_cgroup *parent, *child;
3591 if (memcg->kmem_state != KMEM_ONLINE)
3594 * Clear the online state before clearing memcg_caches array
3595 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3596 * guarantees that no cache will be created for this cgroup
3597 * after we are done (see memcg_create_kmem_cache()).
3599 memcg->kmem_state = KMEM_ALLOCATED;
3601 parent = parent_mem_cgroup(memcg);
3603 parent = root_mem_cgroup;
3606 * Deactivate and reparent kmem_caches.
3608 memcg_deactivate_kmem_caches(memcg, parent);
3610 kmemcg_id = memcg->kmemcg_id;
3611 BUG_ON(kmemcg_id < 0);
3614 * Change kmemcg_id of this cgroup and all its descendants to the
3615 * parent's id, and then move all entries from this cgroup's list_lrus
3616 * to ones of the parent. After we have finished, all list_lrus
3617 * corresponding to this cgroup are guaranteed to remain empty. The
3618 * ordering is imposed by list_lru_node->lock taken by
3619 * memcg_drain_all_list_lrus().
3621 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3622 css_for_each_descendant_pre(css, &memcg->css) {
3623 child = mem_cgroup_from_css(css);
3624 BUG_ON(child->kmemcg_id != kmemcg_id);
3625 child->kmemcg_id = parent->kmemcg_id;
3626 if (!memcg->use_hierarchy)
3631 memcg_drain_all_list_lrus(kmemcg_id, parent);
3633 memcg_free_cache_id(kmemcg_id);
3636 static void memcg_free_kmem(struct mem_cgroup *memcg)
3638 /* css_alloc() failed, offlining didn't happen */
3639 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3640 memcg_offline_kmem(memcg);
3642 if (memcg->kmem_state == KMEM_ALLOCATED) {
3643 WARN_ON(!list_empty(&memcg->kmem_caches));
3644 static_branch_dec(&memcg_kmem_enabled_key);
3648 static int memcg_online_kmem(struct mem_cgroup *memcg)
3652 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3655 static void memcg_free_kmem(struct mem_cgroup *memcg)
3658 #endif /* CONFIG_MEMCG_KMEM */
3660 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3665 mutex_lock(&memcg_max_mutex);
3666 ret = page_counter_set_max(&memcg->kmem, max);
3667 mutex_unlock(&memcg_max_mutex);
3671 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3675 mutex_lock(&memcg_max_mutex);
3677 ret = page_counter_set_max(&memcg->tcpmem, max);
3681 if (!memcg->tcpmem_active) {
3683 * The active flag needs to be written after the static_key
3684 * update. This is what guarantees that the socket activation
3685 * function is the last one to run. See mem_cgroup_sk_alloc()
3686 * for details, and note that we don't mark any socket as
3687 * belonging to this memcg until that flag is up.
3689 * We need to do this, because static_keys will span multiple
3690 * sites, but we can't control their order. If we mark a socket
3691 * as accounted, but the accounting functions are not patched in
3692 * yet, we'll lose accounting.
3694 * We never race with the readers in mem_cgroup_sk_alloc(),
3695 * because when this value change, the code to process it is not
3698 static_branch_inc(&memcg_sockets_enabled_key);
3699 memcg->tcpmem_active = true;
3702 mutex_unlock(&memcg_max_mutex);
3707 * The user of this function is...
3710 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3711 char *buf, size_t nbytes, loff_t off)
3713 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3714 unsigned long nr_pages;
3717 buf = strstrip(buf);
3718 ret = page_counter_memparse(buf, "-1", &nr_pages);
3722 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3724 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3728 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3730 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3733 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3736 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3737 "Please report your usecase to linux-mm@kvack.org if you "
3738 "depend on this functionality.\n");
3739 ret = memcg_update_kmem_max(memcg, nr_pages);
3742 ret = memcg_update_tcp_max(memcg, nr_pages);
3746 case RES_SOFT_LIMIT:
3747 memcg->soft_limit = nr_pages;
3751 return ret ?: nbytes;
3754 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3755 size_t nbytes, loff_t off)
3757 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3758 struct page_counter *counter;
3760 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3762 counter = &memcg->memory;
3765 counter = &memcg->memsw;
3768 counter = &memcg->kmem;
3771 counter = &memcg->tcpmem;
3777 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3779 page_counter_reset_watermark(counter);
3782 counter->failcnt = 0;
3791 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3794 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3798 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3799 struct cftype *cft, u64 val)
3801 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3803 if (val & ~MOVE_MASK)
3807 * No kind of locking is needed in here, because ->can_attach() will
3808 * check this value once in the beginning of the process, and then carry
3809 * on with stale data. This means that changes to this value will only
3810 * affect task migrations starting after the change.
3812 memcg->move_charge_at_immigrate = val;
3816 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3817 struct cftype *cft, u64 val)
3825 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3826 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3827 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3829 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3830 int nid, unsigned int lru_mask)
3832 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
3833 unsigned long nr = 0;
3836 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3839 if (!(BIT(lru) & lru_mask))
3841 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3846 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3847 unsigned int lru_mask)
3849 unsigned long nr = 0;
3853 if (!(BIT(lru) & lru_mask))
3855 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3860 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3864 unsigned int lru_mask;
3867 static const struct numa_stat stats[] = {
3868 { "total", LRU_ALL },
3869 { "file", LRU_ALL_FILE },
3870 { "anon", LRU_ALL_ANON },
3871 { "unevictable", BIT(LRU_UNEVICTABLE) },
3873 const struct numa_stat *stat;
3876 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3878 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3879 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3880 seq_printf(m, "%s=%lu", stat->name, nr);
3881 for_each_node_state(nid, N_MEMORY) {
3882 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3884 seq_printf(m, " N%d=%lu", nid, nr);
3889 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3890 struct mem_cgroup *iter;
3893 for_each_mem_cgroup_tree(iter, memcg)
3894 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3895 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3896 for_each_node_state(nid, N_MEMORY) {
3898 for_each_mem_cgroup_tree(iter, memcg)
3899 nr += mem_cgroup_node_nr_lru_pages(
3900 iter, nid, stat->lru_mask);
3901 seq_printf(m, " N%d=%lu", nid, nr);
3908 #endif /* CONFIG_NUMA */
3910 static const unsigned int memcg1_stats[] = {
3921 static const char *const memcg1_stat_names[] = {
3932 /* Universal VM events cgroup1 shows, original sort order */
3933 static const unsigned int memcg1_events[] = {
3940 static const char *const memcg1_event_names[] = {
3947 static int memcg_stat_show(struct seq_file *m, void *v)
3949 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3950 unsigned long memory, memsw;
3951 struct mem_cgroup *mi;
3954 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3955 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3957 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3958 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3960 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3961 memcg_page_state_local(memcg, memcg1_stats[i]) *
3965 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3966 seq_printf(m, "%s %lu\n", memcg1_event_names[i],
3967 memcg_events_local(memcg, memcg1_events[i]));
3969 for (i = 0; i < NR_LRU_LISTS; i++)
3970 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3971 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3974 /* Hierarchical information */
3975 memory = memsw = PAGE_COUNTER_MAX;
3976 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3977 memory = min(memory, mi->memory.max);
3978 memsw = min(memsw, mi->memsw.max);
3980 seq_printf(m, "hierarchical_memory_limit %llu\n",
3981 (u64)memory * PAGE_SIZE);
3982 if (do_memsw_account())
3983 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3984 (u64)memsw * PAGE_SIZE);
3986 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3987 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3989 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3990 (u64)memcg_page_state(memcg, memcg1_stats[i]) *
3994 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3995 seq_printf(m, "total_%s %llu\n", memcg1_event_names[i],
3996 (u64)memcg_events(memcg, memcg1_events[i]));
3998 for (i = 0; i < NR_LRU_LISTS; i++)
3999 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i],
4000 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4003 #ifdef CONFIG_DEBUG_VM
4006 struct mem_cgroup_per_node *mz;
4007 struct zone_reclaim_stat *rstat;
4008 unsigned long recent_rotated[2] = {0, 0};
4009 unsigned long recent_scanned[2] = {0, 0};
4011 for_each_online_pgdat(pgdat) {
4012 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
4013 rstat = &mz->lruvec.reclaim_stat;
4015 recent_rotated[0] += rstat->recent_rotated[0];
4016 recent_rotated[1] += rstat->recent_rotated[1];
4017 recent_scanned[0] += rstat->recent_scanned[0];
4018 recent_scanned[1] += rstat->recent_scanned[1];
4020 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
4021 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
4022 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
4023 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
4030 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4033 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4035 return mem_cgroup_swappiness(memcg);
4038 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4039 struct cftype *cft, u64 val)
4041 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4047 memcg->swappiness = val;
4049 vm_swappiness = val;
4054 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4056 struct mem_cgroup_threshold_ary *t;
4057 unsigned long usage;
4062 t = rcu_dereference(memcg->thresholds.primary);
4064 t = rcu_dereference(memcg->memsw_thresholds.primary);
4069 usage = mem_cgroup_usage(memcg, swap);
4072 * current_threshold points to threshold just below or equal to usage.
4073 * If it's not true, a threshold was crossed after last
4074 * call of __mem_cgroup_threshold().
4076 i = t->current_threshold;
4079 * Iterate backward over array of thresholds starting from
4080 * current_threshold and check if a threshold is crossed.
4081 * If none of thresholds below usage is crossed, we read
4082 * only one element of the array here.
4084 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4085 eventfd_signal(t->entries[i].eventfd, 1);
4087 /* i = current_threshold + 1 */
4091 * Iterate forward over array of thresholds starting from
4092 * current_threshold+1 and check if a threshold is crossed.
4093 * If none of thresholds above usage is crossed, we read
4094 * only one element of the array here.
4096 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4097 eventfd_signal(t->entries[i].eventfd, 1);
4099 /* Update current_threshold */
4100 t->current_threshold = i - 1;
4105 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4108 __mem_cgroup_threshold(memcg, false);
4109 if (do_memsw_account())
4110 __mem_cgroup_threshold(memcg, true);
4112 memcg = parent_mem_cgroup(memcg);
4116 static int compare_thresholds(const void *a, const void *b)
4118 const struct mem_cgroup_threshold *_a = a;
4119 const struct mem_cgroup_threshold *_b = b;
4121 if (_a->threshold > _b->threshold)
4124 if (_a->threshold < _b->threshold)
4130 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4132 struct mem_cgroup_eventfd_list *ev;
4134 spin_lock(&memcg_oom_lock);
4136 list_for_each_entry(ev, &memcg->oom_notify, list)
4137 eventfd_signal(ev->eventfd, 1);
4139 spin_unlock(&memcg_oom_lock);
4143 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4145 struct mem_cgroup *iter;
4147 for_each_mem_cgroup_tree(iter, memcg)
4148 mem_cgroup_oom_notify_cb(iter);
4151 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4152 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4154 struct mem_cgroup_thresholds *thresholds;
4155 struct mem_cgroup_threshold_ary *new;
4156 unsigned long threshold;
4157 unsigned long usage;
4160 ret = page_counter_memparse(args, "-1", &threshold);
4164 mutex_lock(&memcg->thresholds_lock);
4167 thresholds = &memcg->thresholds;
4168 usage = mem_cgroup_usage(memcg, false);
4169 } else if (type == _MEMSWAP) {
4170 thresholds = &memcg->memsw_thresholds;
4171 usage = mem_cgroup_usage(memcg, true);
4175 /* Check if a threshold crossed before adding a new one */
4176 if (thresholds->primary)
4177 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4179 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4181 /* Allocate memory for new array of thresholds */
4182 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4189 /* Copy thresholds (if any) to new array */
4190 if (thresholds->primary) {
4191 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4192 sizeof(struct mem_cgroup_threshold));
4195 /* Add new threshold */
4196 new->entries[size - 1].eventfd = eventfd;
4197 new->entries[size - 1].threshold = threshold;
4199 /* Sort thresholds. Registering of new threshold isn't time-critical */
4200 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4201 compare_thresholds, NULL);
4203 /* Find current threshold */
4204 new->current_threshold = -1;
4205 for (i = 0; i < size; i++) {
4206 if (new->entries[i].threshold <= usage) {
4208 * new->current_threshold will not be used until
4209 * rcu_assign_pointer(), so it's safe to increment
4212 ++new->current_threshold;
4217 /* Free old spare buffer and save old primary buffer as spare */
4218 kfree(thresholds->spare);
4219 thresholds->spare = thresholds->primary;
4221 rcu_assign_pointer(thresholds->primary, new);
4223 /* To be sure that nobody uses thresholds */
4227 mutex_unlock(&memcg->thresholds_lock);
4232 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4233 struct eventfd_ctx *eventfd, const char *args)
4235 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4238 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4239 struct eventfd_ctx *eventfd, const char *args)
4241 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4244 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4245 struct eventfd_ctx *eventfd, enum res_type type)
4247 struct mem_cgroup_thresholds *thresholds;
4248 struct mem_cgroup_threshold_ary *new;
4249 unsigned long usage;
4250 int i, j, size, entries;
4252 mutex_lock(&memcg->thresholds_lock);
4255 thresholds = &memcg->thresholds;
4256 usage = mem_cgroup_usage(memcg, false);
4257 } else if (type == _MEMSWAP) {
4258 thresholds = &memcg->memsw_thresholds;
4259 usage = mem_cgroup_usage(memcg, true);
4263 if (!thresholds->primary)
4266 /* Check if a threshold crossed before removing */
4267 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4269 /* Calculate new number of threshold */
4271 for (i = 0; i < thresholds->primary->size; i++) {
4272 if (thresholds->primary->entries[i].eventfd != eventfd)
4278 new = thresholds->spare;
4280 /* If no items related to eventfd have been cleared, nothing to do */
4284 /* Set thresholds array to NULL if we don't have thresholds */
4293 /* Copy thresholds and find current threshold */
4294 new->current_threshold = -1;
4295 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4296 if (thresholds->primary->entries[i].eventfd == eventfd)
4299 new->entries[j] = thresholds->primary->entries[i];
4300 if (new->entries[j].threshold <= usage) {
4302 * new->current_threshold will not be used
4303 * until rcu_assign_pointer(), so it's safe to increment
4306 ++new->current_threshold;
4312 /* Swap primary and spare array */
4313 thresholds->spare = thresholds->primary;
4315 rcu_assign_pointer(thresholds->primary, new);
4317 /* To be sure that nobody uses thresholds */
4320 /* If all events are unregistered, free the spare array */
4322 kfree(thresholds->spare);
4323 thresholds->spare = NULL;
4326 mutex_unlock(&memcg->thresholds_lock);
4329 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4330 struct eventfd_ctx *eventfd)
4332 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4335 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4336 struct eventfd_ctx *eventfd)
4338 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4341 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4342 struct eventfd_ctx *eventfd, const char *args)
4344 struct mem_cgroup_eventfd_list *event;
4346 event = kmalloc(sizeof(*event), GFP_KERNEL);
4350 spin_lock(&memcg_oom_lock);
4352 event->eventfd = eventfd;
4353 list_add(&event->list, &memcg->oom_notify);
4355 /* already in OOM ? */
4356 if (memcg->under_oom)
4357 eventfd_signal(eventfd, 1);
4358 spin_unlock(&memcg_oom_lock);
4363 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4364 struct eventfd_ctx *eventfd)
4366 struct mem_cgroup_eventfd_list *ev, *tmp;
4368 spin_lock(&memcg_oom_lock);
4370 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4371 if (ev->eventfd == eventfd) {
4372 list_del(&ev->list);
4377 spin_unlock(&memcg_oom_lock);
4380 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4382 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4384 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4385 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4386 seq_printf(sf, "oom_kill %lu\n",
4387 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4391 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4392 struct cftype *cft, u64 val)
4394 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4396 /* cannot set to root cgroup and only 0 and 1 are allowed */
4397 if (!css->parent || !((val == 0) || (val == 1)))
4400 memcg->oom_kill_disable = val;
4402 memcg_oom_recover(memcg);
4407 #ifdef CONFIG_CGROUP_WRITEBACK
4409 #include <trace/events/writeback.h>
4411 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4413 return wb_domain_init(&memcg->cgwb_domain, gfp);
4416 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4418 wb_domain_exit(&memcg->cgwb_domain);
4421 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4423 wb_domain_size_changed(&memcg->cgwb_domain);
4426 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4428 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4430 if (!memcg->css.parent)
4433 return &memcg->cgwb_domain;
4437 * idx can be of type enum memcg_stat_item or node_stat_item.
4438 * Keep in sync with memcg_exact_page().
4440 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4442 long x = atomic_long_read(&memcg->vmstats[idx]);
4445 for_each_online_cpu(cpu)
4446 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4453 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4454 * @wb: bdi_writeback in question
4455 * @pfilepages: out parameter for number of file pages
4456 * @pheadroom: out parameter for number of allocatable pages according to memcg
4457 * @pdirty: out parameter for number of dirty pages
4458 * @pwriteback: out parameter for number of pages under writeback
4460 * Determine the numbers of file, headroom, dirty, and writeback pages in
4461 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4462 * is a bit more involved.
4464 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4465 * headroom is calculated as the lowest headroom of itself and the
4466 * ancestors. Note that this doesn't consider the actual amount of
4467 * available memory in the system. The caller should further cap
4468 * *@pheadroom accordingly.
4470 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4471 unsigned long *pheadroom, unsigned long *pdirty,
4472 unsigned long *pwriteback)
4474 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4475 struct mem_cgroup *parent;
4477 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4479 /* this should eventually include NR_UNSTABLE_NFS */
4480 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4481 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4482 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4483 *pheadroom = PAGE_COUNTER_MAX;
4485 while ((parent = parent_mem_cgroup(memcg))) {
4486 unsigned long ceiling = min(memcg->memory.max, memcg->high);
4487 unsigned long used = page_counter_read(&memcg->memory);
4489 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4495 * Foreign dirty flushing
4497 * There's an inherent mismatch between memcg and writeback. The former
4498 * trackes ownership per-page while the latter per-inode. This was a
4499 * deliberate design decision because honoring per-page ownership in the
4500 * writeback path is complicated, may lead to higher CPU and IO overheads
4501 * and deemed unnecessary given that write-sharing an inode across
4502 * different cgroups isn't a common use-case.
4504 * Combined with inode majority-writer ownership switching, this works well
4505 * enough in most cases but there are some pathological cases. For
4506 * example, let's say there are two cgroups A and B which keep writing to
4507 * different but confined parts of the same inode. B owns the inode and
4508 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4509 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4510 * triggering background writeback. A will be slowed down without a way to
4511 * make writeback of the dirty pages happen.
4513 * Conditions like the above can lead to a cgroup getting repatedly and
4514 * severely throttled after making some progress after each
4515 * dirty_expire_interval while the underyling IO device is almost
4518 * Solving this problem completely requires matching the ownership tracking
4519 * granularities between memcg and writeback in either direction. However,
4520 * the more egregious behaviors can be avoided by simply remembering the
4521 * most recent foreign dirtying events and initiating remote flushes on
4522 * them when local writeback isn't enough to keep the memory clean enough.
4524 * The following two functions implement such mechanism. When a foreign
4525 * page - a page whose memcg and writeback ownerships don't match - is
4526 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4527 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4528 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4529 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4530 * foreign bdi_writebacks which haven't expired. Both the numbers of
4531 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4532 * limited to MEMCG_CGWB_FRN_CNT.
4534 * The mechanism only remembers IDs and doesn't hold any object references.
4535 * As being wrong occasionally doesn't matter, updates and accesses to the
4536 * records are lockless and racy.
4538 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4539 struct bdi_writeback *wb)
4541 struct mem_cgroup *memcg = page->mem_cgroup;
4542 struct memcg_cgwb_frn *frn;
4543 u64 now = get_jiffies_64();
4544 u64 oldest_at = now;
4548 trace_track_foreign_dirty(page, wb);
4551 * Pick the slot to use. If there is already a slot for @wb, keep
4552 * using it. If not replace the oldest one which isn't being
4555 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4556 frn = &memcg->cgwb_frn[i];
4557 if (frn->bdi_id == wb->bdi->id &&
4558 frn->memcg_id == wb->memcg_css->id)
4560 if (time_before64(frn->at, oldest_at) &&
4561 atomic_read(&frn->done.cnt) == 1) {
4563 oldest_at = frn->at;
4567 if (i < MEMCG_CGWB_FRN_CNT) {
4569 * Re-using an existing one. Update timestamp lazily to
4570 * avoid making the cacheline hot. We want them to be
4571 * reasonably up-to-date and significantly shorter than
4572 * dirty_expire_interval as that's what expires the record.
4573 * Use the shorter of 1s and dirty_expire_interval / 8.
4575 unsigned long update_intv =
4576 min_t(unsigned long, HZ,
4577 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4579 if (time_before64(frn->at, now - update_intv))
4581 } else if (oldest >= 0) {
4582 /* replace the oldest free one */
4583 frn = &memcg->cgwb_frn[oldest];
4584 frn->bdi_id = wb->bdi->id;
4585 frn->memcg_id = wb->memcg_css->id;
4590 /* issue foreign writeback flushes for recorded foreign dirtying events */
4591 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4593 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4594 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4595 u64 now = jiffies_64;
4598 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4599 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4602 * If the record is older than dirty_expire_interval,
4603 * writeback on it has already started. No need to kick it
4604 * off again. Also, don't start a new one if there's
4605 * already one in flight.
4607 if (time_after64(frn->at, now - intv) &&
4608 atomic_read(&frn->done.cnt) == 1) {
4610 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4611 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4612 WB_REASON_FOREIGN_FLUSH,
4618 #else /* CONFIG_CGROUP_WRITEBACK */
4620 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4625 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4629 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4633 #endif /* CONFIG_CGROUP_WRITEBACK */
4636 * DO NOT USE IN NEW FILES.
4638 * "cgroup.event_control" implementation.
4640 * This is way over-engineered. It tries to support fully configurable
4641 * events for each user. Such level of flexibility is completely
4642 * unnecessary especially in the light of the planned unified hierarchy.
4644 * Please deprecate this and replace with something simpler if at all
4649 * Unregister event and free resources.
4651 * Gets called from workqueue.
4653 static void memcg_event_remove(struct work_struct *work)
4655 struct mem_cgroup_event *event =
4656 container_of(work, struct mem_cgroup_event, remove);
4657 struct mem_cgroup *memcg = event->memcg;
4659 remove_wait_queue(event->wqh, &event->wait);
4661 event->unregister_event(memcg, event->eventfd);
4663 /* Notify userspace the event is going away. */
4664 eventfd_signal(event->eventfd, 1);
4666 eventfd_ctx_put(event->eventfd);
4668 css_put(&memcg->css);
4672 * Gets called on EPOLLHUP on eventfd when user closes it.
4674 * Called with wqh->lock held and interrupts disabled.
4676 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4677 int sync, void *key)
4679 struct mem_cgroup_event *event =
4680 container_of(wait, struct mem_cgroup_event, wait);
4681 struct mem_cgroup *memcg = event->memcg;
4682 __poll_t flags = key_to_poll(key);
4684 if (flags & EPOLLHUP) {
4686 * If the event has been detached at cgroup removal, we
4687 * can simply return knowing the other side will cleanup
4690 * We can't race against event freeing since the other
4691 * side will require wqh->lock via remove_wait_queue(),
4694 spin_lock(&memcg->event_list_lock);
4695 if (!list_empty(&event->list)) {
4696 list_del_init(&event->list);
4698 * We are in atomic context, but cgroup_event_remove()
4699 * may sleep, so we have to call it in workqueue.
4701 schedule_work(&event->remove);
4703 spin_unlock(&memcg->event_list_lock);
4709 static void memcg_event_ptable_queue_proc(struct file *file,
4710 wait_queue_head_t *wqh, poll_table *pt)
4712 struct mem_cgroup_event *event =
4713 container_of(pt, struct mem_cgroup_event, pt);
4716 add_wait_queue(wqh, &event->wait);
4720 * DO NOT USE IN NEW FILES.
4722 * Parse input and register new cgroup event handler.
4724 * Input must be in format '<event_fd> <control_fd> <args>'.
4725 * Interpretation of args is defined by control file implementation.
4727 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4728 char *buf, size_t nbytes, loff_t off)
4730 struct cgroup_subsys_state *css = of_css(of);
4731 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4732 struct mem_cgroup_event *event;
4733 struct cgroup_subsys_state *cfile_css;
4734 unsigned int efd, cfd;
4741 buf = strstrip(buf);
4743 efd = simple_strtoul(buf, &endp, 10);
4748 cfd = simple_strtoul(buf, &endp, 10);
4749 if ((*endp != ' ') && (*endp != '\0'))
4753 event = kzalloc(sizeof(*event), GFP_KERNEL);
4757 event->memcg = memcg;
4758 INIT_LIST_HEAD(&event->list);
4759 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4760 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4761 INIT_WORK(&event->remove, memcg_event_remove);
4769 event->eventfd = eventfd_ctx_fileget(efile.file);
4770 if (IS_ERR(event->eventfd)) {
4771 ret = PTR_ERR(event->eventfd);
4778 goto out_put_eventfd;
4781 /* the process need read permission on control file */
4782 /* AV: shouldn't we check that it's been opened for read instead? */
4783 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4788 * Determine the event callbacks and set them in @event. This used
4789 * to be done via struct cftype but cgroup core no longer knows
4790 * about these events. The following is crude but the whole thing
4791 * is for compatibility anyway.
4793 * DO NOT ADD NEW FILES.
4795 name = cfile.file->f_path.dentry->d_name.name;
4797 if (!strcmp(name, "memory.usage_in_bytes")) {
4798 event->register_event = mem_cgroup_usage_register_event;
4799 event->unregister_event = mem_cgroup_usage_unregister_event;
4800 } else if (!strcmp(name, "memory.oom_control")) {
4801 event->register_event = mem_cgroup_oom_register_event;
4802 event->unregister_event = mem_cgroup_oom_unregister_event;
4803 } else if (!strcmp(name, "memory.pressure_level")) {
4804 event->register_event = vmpressure_register_event;
4805 event->unregister_event = vmpressure_unregister_event;
4806 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4807 event->register_event = memsw_cgroup_usage_register_event;
4808 event->unregister_event = memsw_cgroup_usage_unregister_event;
4815 * Verify @cfile should belong to @css. Also, remaining events are
4816 * automatically removed on cgroup destruction but the removal is
4817 * asynchronous, so take an extra ref on @css.
4819 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4820 &memory_cgrp_subsys);
4822 if (IS_ERR(cfile_css))
4824 if (cfile_css != css) {
4829 ret = event->register_event(memcg, event->eventfd, buf);
4833 vfs_poll(efile.file, &event->pt);
4835 spin_lock(&memcg->event_list_lock);
4836 list_add(&event->list, &memcg->event_list);
4837 spin_unlock(&memcg->event_list_lock);
4849 eventfd_ctx_put(event->eventfd);
4858 static struct cftype mem_cgroup_legacy_files[] = {
4860 .name = "usage_in_bytes",
4861 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4862 .read_u64 = mem_cgroup_read_u64,
4865 .name = "max_usage_in_bytes",
4866 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4867 .write = mem_cgroup_reset,
4868 .read_u64 = mem_cgroup_read_u64,
4871 .name = "limit_in_bytes",
4872 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4873 .write = mem_cgroup_write,
4874 .read_u64 = mem_cgroup_read_u64,
4877 .name = "soft_limit_in_bytes",
4878 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4879 .write = mem_cgroup_write,
4880 .read_u64 = mem_cgroup_read_u64,
4884 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4885 .write = mem_cgroup_reset,
4886 .read_u64 = mem_cgroup_read_u64,
4890 .seq_show = memcg_stat_show,
4893 .name = "force_empty",
4894 .write = mem_cgroup_force_empty_write,
4897 .name = "use_hierarchy",
4898 .write_u64 = mem_cgroup_hierarchy_write,
4899 .read_u64 = mem_cgroup_hierarchy_read,
4902 .name = "cgroup.event_control", /* XXX: for compat */
4903 .write = memcg_write_event_control,
4904 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4907 .name = "swappiness",
4908 .read_u64 = mem_cgroup_swappiness_read,
4909 .write_u64 = mem_cgroup_swappiness_write,
4912 .name = "move_charge_at_immigrate",
4913 .read_u64 = mem_cgroup_move_charge_read,
4914 .write_u64 = mem_cgroup_move_charge_write,
4917 .name = "oom_control",
4918 .seq_show = mem_cgroup_oom_control_read,
4919 .write_u64 = mem_cgroup_oom_control_write,
4920 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4923 .name = "pressure_level",
4927 .name = "numa_stat",
4928 .seq_show = memcg_numa_stat_show,
4932 .name = "kmem.limit_in_bytes",
4933 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4934 .write = mem_cgroup_write,
4935 .read_u64 = mem_cgroup_read_u64,
4938 .name = "kmem.usage_in_bytes",
4939 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4940 .read_u64 = mem_cgroup_read_u64,
4943 .name = "kmem.failcnt",
4944 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4945 .write = mem_cgroup_reset,
4946 .read_u64 = mem_cgroup_read_u64,
4949 .name = "kmem.max_usage_in_bytes",
4950 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4951 .write = mem_cgroup_reset,
4952 .read_u64 = mem_cgroup_read_u64,
4954 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4956 .name = "kmem.slabinfo",
4957 .seq_start = memcg_slab_start,
4958 .seq_next = memcg_slab_next,
4959 .seq_stop = memcg_slab_stop,
4960 .seq_show = memcg_slab_show,
4964 .name = "kmem.tcp.limit_in_bytes",
4965 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4966 .write = mem_cgroup_write,
4967 .read_u64 = mem_cgroup_read_u64,
4970 .name = "kmem.tcp.usage_in_bytes",
4971 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4972 .read_u64 = mem_cgroup_read_u64,
4975 .name = "kmem.tcp.failcnt",
4976 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4977 .write = mem_cgroup_reset,
4978 .read_u64 = mem_cgroup_read_u64,
4981 .name = "kmem.tcp.max_usage_in_bytes",
4982 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4983 .write = mem_cgroup_reset,
4984 .read_u64 = mem_cgroup_read_u64,
4986 { }, /* terminate */
4990 * Private memory cgroup IDR
4992 * Swap-out records and page cache shadow entries need to store memcg
4993 * references in constrained space, so we maintain an ID space that is
4994 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4995 * memory-controlled cgroups to 64k.
4997 * However, there usually are many references to the oflline CSS after
4998 * the cgroup has been destroyed, such as page cache or reclaimable
4999 * slab objects, that don't need to hang on to the ID. We want to keep
5000 * those dead CSS from occupying IDs, or we might quickly exhaust the
5001 * relatively small ID space and prevent the creation of new cgroups
5002 * even when there are much fewer than 64k cgroups - possibly none.
5004 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5005 * be freed and recycled when it's no longer needed, which is usually
5006 * when the CSS is offlined.
5008 * The only exception to that are records of swapped out tmpfs/shmem
5009 * pages that need to be attributed to live ancestors on swapin. But
5010 * those references are manageable from userspace.
5013 static DEFINE_IDR(mem_cgroup_idr);
5015 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5017 if (memcg->id.id > 0) {
5018 idr_remove(&mem_cgroup_idr, memcg->id.id);
5023 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
5025 refcount_add(n, &memcg->id.ref);
5028 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5030 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5031 mem_cgroup_id_remove(memcg);
5033 /* Memcg ID pins CSS */
5034 css_put(&memcg->css);
5038 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5040 mem_cgroup_id_put_many(memcg, 1);
5044 * mem_cgroup_from_id - look up a memcg from a memcg id
5045 * @id: the memcg id to look up
5047 * Caller must hold rcu_read_lock().
5049 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5051 WARN_ON_ONCE(!rcu_read_lock_held());
5052 return idr_find(&mem_cgroup_idr, id);
5055 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5057 struct mem_cgroup_per_node *pn;
5060 * This routine is called against possible nodes.
5061 * But it's BUG to call kmalloc() against offline node.
5063 * TODO: this routine can waste much memory for nodes which will
5064 * never be onlined. It's better to use memory hotplug callback
5067 if (!node_state(node, N_NORMAL_MEMORY))
5069 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5073 pn->lruvec_stat_local = alloc_percpu(struct lruvec_stat);
5074 if (!pn->lruvec_stat_local) {
5079 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
5080 if (!pn->lruvec_stat_cpu) {
5081 free_percpu(pn->lruvec_stat_local);
5086 lruvec_init(&pn->lruvec);
5087 pn->usage_in_excess = 0;
5088 pn->on_tree = false;
5091 memcg->nodeinfo[node] = pn;
5095 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5097 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5102 free_percpu(pn->lruvec_stat_cpu);
5103 free_percpu(pn->lruvec_stat_local);
5107 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5112 free_mem_cgroup_per_node_info(memcg, node);
5113 free_percpu(memcg->vmstats_percpu);
5114 free_percpu(memcg->vmstats_local);
5118 static void mem_cgroup_free(struct mem_cgroup *memcg)
5120 memcg_wb_domain_exit(memcg);
5122 * Flush percpu vmstats and vmevents to guarantee the value correctness
5123 * on parent's and all ancestor levels.
5125 memcg_flush_percpu_vmstats(memcg);
5126 memcg_flush_percpu_vmevents(memcg);
5127 __mem_cgroup_free(memcg);
5130 static struct mem_cgroup *mem_cgroup_alloc(void)
5132 struct mem_cgroup *memcg;
5135 int __maybe_unused i;
5136 long error = -ENOMEM;
5138 size = sizeof(struct mem_cgroup);
5139 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5141 memcg = kzalloc(size, GFP_KERNEL);
5143 return ERR_PTR(error);
5145 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5146 1, MEM_CGROUP_ID_MAX,
5148 if (memcg->id.id < 0) {
5149 error = memcg->id.id;
5153 memcg->vmstats_local = alloc_percpu(struct memcg_vmstats_percpu);
5154 if (!memcg->vmstats_local)
5157 memcg->vmstats_percpu = alloc_percpu(struct memcg_vmstats_percpu);
5158 if (!memcg->vmstats_percpu)
5162 if (alloc_mem_cgroup_per_node_info(memcg, node))
5165 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5168 INIT_WORK(&memcg->high_work, high_work_func);
5169 memcg->last_scanned_node = MAX_NUMNODES;
5170 INIT_LIST_HEAD(&memcg->oom_notify);
5171 mutex_init(&memcg->thresholds_lock);
5172 spin_lock_init(&memcg->move_lock);
5173 vmpressure_init(&memcg->vmpressure);
5174 INIT_LIST_HEAD(&memcg->event_list);
5175 spin_lock_init(&memcg->event_list_lock);
5176 memcg->socket_pressure = jiffies;
5177 #ifdef CONFIG_MEMCG_KMEM
5178 memcg->kmemcg_id = -1;
5180 #ifdef CONFIG_CGROUP_WRITEBACK
5181 INIT_LIST_HEAD(&memcg->cgwb_list);
5182 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5183 memcg->cgwb_frn[i].done =
5184 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5186 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5187 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5188 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5189 memcg->deferred_split_queue.split_queue_len = 0;
5191 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5194 mem_cgroup_id_remove(memcg);
5195 __mem_cgroup_free(memcg);
5196 return ERR_PTR(error);
5199 static struct cgroup_subsys_state * __ref
5200 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5202 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5203 struct mem_cgroup *memcg;
5204 long error = -ENOMEM;
5206 memcg = mem_cgroup_alloc();
5208 return ERR_CAST(memcg);
5210 memcg->high = PAGE_COUNTER_MAX;
5211 memcg->soft_limit = PAGE_COUNTER_MAX;
5213 memcg->swappiness = mem_cgroup_swappiness(parent);
5214 memcg->oom_kill_disable = parent->oom_kill_disable;
5216 if (parent && parent->use_hierarchy) {
5217 memcg->use_hierarchy = true;
5218 page_counter_init(&memcg->memory, &parent->memory);
5219 page_counter_init(&memcg->swap, &parent->swap);
5220 page_counter_init(&memcg->memsw, &parent->memsw);
5221 page_counter_init(&memcg->kmem, &parent->kmem);
5222 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5224 page_counter_init(&memcg->memory, NULL);
5225 page_counter_init(&memcg->swap, NULL);
5226 page_counter_init(&memcg->memsw, NULL);
5227 page_counter_init(&memcg->kmem, NULL);
5228 page_counter_init(&memcg->tcpmem, NULL);
5230 * Deeper hierachy with use_hierarchy == false doesn't make
5231 * much sense so let cgroup subsystem know about this
5232 * unfortunate state in our controller.
5234 if (parent != root_mem_cgroup)
5235 memory_cgrp_subsys.broken_hierarchy = true;
5238 /* The following stuff does not apply to the root */
5240 #ifdef CONFIG_MEMCG_KMEM
5241 INIT_LIST_HEAD(&memcg->kmem_caches);
5243 root_mem_cgroup = memcg;
5247 error = memcg_online_kmem(memcg);
5251 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5252 static_branch_inc(&memcg_sockets_enabled_key);
5256 mem_cgroup_id_remove(memcg);
5257 mem_cgroup_free(memcg);
5258 return ERR_PTR(error);
5261 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5263 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5266 * A memcg must be visible for memcg_expand_shrinker_maps()
5267 * by the time the maps are allocated. So, we allocate maps
5268 * here, when for_each_mem_cgroup() can't skip it.
5270 if (memcg_alloc_shrinker_maps(memcg)) {
5271 mem_cgroup_id_remove(memcg);
5275 /* Online state pins memcg ID, memcg ID pins CSS */
5276 refcount_set(&memcg->id.ref, 1);
5281 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5283 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5284 struct mem_cgroup_event *event, *tmp;
5287 * Unregister events and notify userspace.
5288 * Notify userspace about cgroup removing only after rmdir of cgroup
5289 * directory to avoid race between userspace and kernelspace.
5291 spin_lock(&memcg->event_list_lock);
5292 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5293 list_del_init(&event->list);
5294 schedule_work(&event->remove);
5296 spin_unlock(&memcg->event_list_lock);
5298 page_counter_set_min(&memcg->memory, 0);
5299 page_counter_set_low(&memcg->memory, 0);
5301 memcg_offline_kmem(memcg);
5302 wb_memcg_offline(memcg);
5304 drain_all_stock(memcg);
5306 mem_cgroup_id_put(memcg);
5309 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5311 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5313 invalidate_reclaim_iterators(memcg);
5316 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5318 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5319 int __maybe_unused i;
5321 #ifdef CONFIG_CGROUP_WRITEBACK
5322 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5323 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5325 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5326 static_branch_dec(&memcg_sockets_enabled_key);
5328 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5329 static_branch_dec(&memcg_sockets_enabled_key);
5331 vmpressure_cleanup(&memcg->vmpressure);
5332 cancel_work_sync(&memcg->high_work);
5333 mem_cgroup_remove_from_trees(memcg);
5334 memcg_free_shrinker_maps(memcg);
5335 memcg_free_kmem(memcg);
5336 mem_cgroup_free(memcg);
5340 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5341 * @css: the target css
5343 * Reset the states of the mem_cgroup associated with @css. This is
5344 * invoked when the userland requests disabling on the default hierarchy
5345 * but the memcg is pinned through dependency. The memcg should stop
5346 * applying policies and should revert to the vanilla state as it may be
5347 * made visible again.
5349 * The current implementation only resets the essential configurations.
5350 * This needs to be expanded to cover all the visible parts.
5352 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5354 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5356 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5357 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5358 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
5359 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5360 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5361 page_counter_set_min(&memcg->memory, 0);
5362 page_counter_set_low(&memcg->memory, 0);
5363 memcg->high = PAGE_COUNTER_MAX;
5364 memcg->soft_limit = PAGE_COUNTER_MAX;
5365 memcg_wb_domain_size_changed(memcg);
5369 /* Handlers for move charge at task migration. */
5370 static int mem_cgroup_do_precharge(unsigned long count)
5374 /* Try a single bulk charge without reclaim first, kswapd may wake */
5375 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5377 mc.precharge += count;
5381 /* Try charges one by one with reclaim, but do not retry */
5383 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5397 enum mc_target_type {
5404 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5405 unsigned long addr, pte_t ptent)
5407 struct page *page = vm_normal_page(vma, addr, ptent);
5409 if (!page || !page_mapped(page))
5411 if (PageAnon(page)) {
5412 if (!(mc.flags & MOVE_ANON))
5415 if (!(mc.flags & MOVE_FILE))
5418 if (!get_page_unless_zero(page))
5424 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5425 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5426 pte_t ptent, swp_entry_t *entry)
5428 struct page *page = NULL;
5429 swp_entry_t ent = pte_to_swp_entry(ptent);
5431 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
5435 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5436 * a device and because they are not accessible by CPU they are store
5437 * as special swap entry in the CPU page table.
5439 if (is_device_private_entry(ent)) {
5440 page = device_private_entry_to_page(ent);
5442 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5443 * a refcount of 1 when free (unlike normal page)
5445 if (!page_ref_add_unless(page, 1, 1))
5451 * Because lookup_swap_cache() updates some statistics counter,
5452 * we call find_get_page() with swapper_space directly.
5454 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5455 if (do_memsw_account())
5456 entry->val = ent.val;
5461 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5462 pte_t ptent, swp_entry_t *entry)
5468 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5469 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5471 struct page *page = NULL;
5472 struct address_space *mapping;
5475 if (!vma->vm_file) /* anonymous vma */
5477 if (!(mc.flags & MOVE_FILE))
5480 mapping = vma->vm_file->f_mapping;
5481 pgoff = linear_page_index(vma, addr);
5483 /* page is moved even if it's not RSS of this task(page-faulted). */
5485 /* shmem/tmpfs may report page out on swap: account for that too. */
5486 if (shmem_mapping(mapping)) {
5487 page = find_get_entry(mapping, pgoff);
5488 if (xa_is_value(page)) {
5489 swp_entry_t swp = radix_to_swp_entry(page);
5490 if (do_memsw_account())
5492 page = find_get_page(swap_address_space(swp),
5496 page = find_get_page(mapping, pgoff);
5498 page = find_get_page(mapping, pgoff);
5504 * mem_cgroup_move_account - move account of the page
5506 * @compound: charge the page as compound or small page
5507 * @from: mem_cgroup which the page is moved from.
5508 * @to: mem_cgroup which the page is moved to. @from != @to.
5510 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5512 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5515 static int mem_cgroup_move_account(struct page *page,
5517 struct mem_cgroup *from,
5518 struct mem_cgroup *to)
5520 struct lruvec *from_vec, *to_vec;
5521 struct pglist_data *pgdat;
5522 unsigned long flags;
5523 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5527 VM_BUG_ON(from == to);
5528 VM_BUG_ON_PAGE(PageLRU(page), page);
5529 VM_BUG_ON(compound && !PageTransHuge(page));
5532 * Prevent mem_cgroup_migrate() from looking at
5533 * page->mem_cgroup of its source page while we change it.
5536 if (!trylock_page(page))
5540 if (page->mem_cgroup != from)
5543 anon = PageAnon(page);
5545 pgdat = page_pgdat(page);
5546 from_vec = mem_cgroup_lruvec(pgdat, from);
5547 to_vec = mem_cgroup_lruvec(pgdat, to);
5549 spin_lock_irqsave(&from->move_lock, flags);
5551 if (!anon && page_mapped(page)) {
5552 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5553 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5557 * move_lock grabbed above and caller set from->moving_account, so
5558 * mod_memcg_page_state will serialize updates to PageDirty.
5559 * So mapping should be stable for dirty pages.
5561 if (!anon && PageDirty(page)) {
5562 struct address_space *mapping = page_mapping(page);
5564 if (mapping_cap_account_dirty(mapping)) {
5565 __mod_lruvec_state(from_vec, NR_FILE_DIRTY, -nr_pages);
5566 __mod_lruvec_state(to_vec, NR_FILE_DIRTY, nr_pages);
5570 if (PageWriteback(page)) {
5571 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5572 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5576 * It is safe to change page->mem_cgroup here because the page
5577 * is referenced, charged, and isolated - we can't race with
5578 * uncharging, charging, migration, or LRU putback.
5581 /* caller should have done css_get */
5582 page->mem_cgroup = to;
5584 spin_unlock_irqrestore(&from->move_lock, flags);
5588 local_irq_disable();
5589 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
5590 memcg_check_events(to, page);
5591 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
5592 memcg_check_events(from, page);
5601 * get_mctgt_type - get target type of moving charge
5602 * @vma: the vma the pte to be checked belongs
5603 * @addr: the address corresponding to the pte to be checked
5604 * @ptent: the pte to be checked
5605 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5608 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5609 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5610 * move charge. if @target is not NULL, the page is stored in target->page
5611 * with extra refcnt got(Callers should handle it).
5612 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5613 * target for charge migration. if @target is not NULL, the entry is stored
5615 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5616 * (so ZONE_DEVICE page and thus not on the lru).
5617 * For now we such page is charge like a regular page would be as for all
5618 * intent and purposes it is just special memory taking the place of a
5621 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5623 * Called with pte lock held.
5626 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5627 unsigned long addr, pte_t ptent, union mc_target *target)
5629 struct page *page = NULL;
5630 enum mc_target_type ret = MC_TARGET_NONE;
5631 swp_entry_t ent = { .val = 0 };
5633 if (pte_present(ptent))
5634 page = mc_handle_present_pte(vma, addr, ptent);
5635 else if (is_swap_pte(ptent))
5636 page = mc_handle_swap_pte(vma, ptent, &ent);
5637 else if (pte_none(ptent))
5638 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5640 if (!page && !ent.val)
5644 * Do only loose check w/o serialization.
5645 * mem_cgroup_move_account() checks the page is valid or
5646 * not under LRU exclusion.
5648 if (page->mem_cgroup == mc.from) {
5649 ret = MC_TARGET_PAGE;
5650 if (is_device_private_page(page))
5651 ret = MC_TARGET_DEVICE;
5653 target->page = page;
5655 if (!ret || !target)
5659 * There is a swap entry and a page doesn't exist or isn't charged.
5660 * But we cannot move a tail-page in a THP.
5662 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5663 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5664 ret = MC_TARGET_SWAP;
5671 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5673 * We don't consider PMD mapped swapping or file mapped pages because THP does
5674 * not support them for now.
5675 * Caller should make sure that pmd_trans_huge(pmd) is true.
5677 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5678 unsigned long addr, pmd_t pmd, union mc_target *target)
5680 struct page *page = NULL;
5681 enum mc_target_type ret = MC_TARGET_NONE;
5683 if (unlikely(is_swap_pmd(pmd))) {
5684 VM_BUG_ON(thp_migration_supported() &&
5685 !is_pmd_migration_entry(pmd));
5688 page = pmd_page(pmd);
5689 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5690 if (!(mc.flags & MOVE_ANON))
5692 if (page->mem_cgroup == mc.from) {
5693 ret = MC_TARGET_PAGE;
5696 target->page = page;
5702 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5703 unsigned long addr, pmd_t pmd, union mc_target *target)
5705 return MC_TARGET_NONE;
5709 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5710 unsigned long addr, unsigned long end,
5711 struct mm_walk *walk)
5713 struct vm_area_struct *vma = walk->vma;
5717 ptl = pmd_trans_huge_lock(pmd, vma);
5720 * Note their can not be MC_TARGET_DEVICE for now as we do not
5721 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5722 * this might change.
5724 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5725 mc.precharge += HPAGE_PMD_NR;
5730 if (pmd_trans_unstable(pmd))
5732 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5733 for (; addr != end; pte++, addr += PAGE_SIZE)
5734 if (get_mctgt_type(vma, addr, *pte, NULL))
5735 mc.precharge++; /* increment precharge temporarily */
5736 pte_unmap_unlock(pte - 1, ptl);
5742 static const struct mm_walk_ops precharge_walk_ops = {
5743 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5746 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5748 unsigned long precharge;
5750 down_read(&mm->mmap_sem);
5751 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5752 up_read(&mm->mmap_sem);
5754 precharge = mc.precharge;
5760 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5762 unsigned long precharge = mem_cgroup_count_precharge(mm);
5764 VM_BUG_ON(mc.moving_task);
5765 mc.moving_task = current;
5766 return mem_cgroup_do_precharge(precharge);
5769 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5770 static void __mem_cgroup_clear_mc(void)
5772 struct mem_cgroup *from = mc.from;
5773 struct mem_cgroup *to = mc.to;
5775 /* we must uncharge all the leftover precharges from mc.to */
5777 cancel_charge(mc.to, mc.precharge);
5781 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5782 * we must uncharge here.
5784 if (mc.moved_charge) {
5785 cancel_charge(mc.from, mc.moved_charge);
5786 mc.moved_charge = 0;
5788 /* we must fixup refcnts and charges */
5789 if (mc.moved_swap) {
5790 /* uncharge swap account from the old cgroup */
5791 if (!mem_cgroup_is_root(mc.from))
5792 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5794 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5797 * we charged both to->memory and to->memsw, so we
5798 * should uncharge to->memory.
5800 if (!mem_cgroup_is_root(mc.to))
5801 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5803 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
5804 css_put_many(&mc.to->css, mc.moved_swap);
5808 memcg_oom_recover(from);
5809 memcg_oom_recover(to);
5810 wake_up_all(&mc.waitq);
5813 static void mem_cgroup_clear_mc(void)
5815 struct mm_struct *mm = mc.mm;
5818 * we must clear moving_task before waking up waiters at the end of
5821 mc.moving_task = NULL;
5822 __mem_cgroup_clear_mc();
5823 spin_lock(&mc.lock);
5827 spin_unlock(&mc.lock);
5832 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5834 struct cgroup_subsys_state *css;
5835 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5836 struct mem_cgroup *from;
5837 struct task_struct *leader, *p;
5838 struct mm_struct *mm;
5839 unsigned long move_flags;
5842 /* charge immigration isn't supported on the default hierarchy */
5843 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5847 * Multi-process migrations only happen on the default hierarchy
5848 * where charge immigration is not used. Perform charge
5849 * immigration if @tset contains a leader and whine if there are
5853 cgroup_taskset_for_each_leader(leader, css, tset) {
5856 memcg = mem_cgroup_from_css(css);
5862 * We are now commited to this value whatever it is. Changes in this
5863 * tunable will only affect upcoming migrations, not the current one.
5864 * So we need to save it, and keep it going.
5866 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5870 from = mem_cgroup_from_task(p);
5872 VM_BUG_ON(from == memcg);
5874 mm = get_task_mm(p);
5877 /* We move charges only when we move a owner of the mm */
5878 if (mm->owner == p) {
5881 VM_BUG_ON(mc.precharge);
5882 VM_BUG_ON(mc.moved_charge);
5883 VM_BUG_ON(mc.moved_swap);
5885 spin_lock(&mc.lock);
5889 mc.flags = move_flags;
5890 spin_unlock(&mc.lock);
5891 /* We set mc.moving_task later */
5893 ret = mem_cgroup_precharge_mc(mm);
5895 mem_cgroup_clear_mc();
5902 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5905 mem_cgroup_clear_mc();
5908 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5909 unsigned long addr, unsigned long end,
5910 struct mm_walk *walk)
5913 struct vm_area_struct *vma = walk->vma;
5916 enum mc_target_type target_type;
5917 union mc_target target;
5920 ptl = pmd_trans_huge_lock(pmd, vma);
5922 if (mc.precharge < HPAGE_PMD_NR) {
5926 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5927 if (target_type == MC_TARGET_PAGE) {
5929 if (!isolate_lru_page(page)) {
5930 if (!mem_cgroup_move_account(page, true,
5932 mc.precharge -= HPAGE_PMD_NR;
5933 mc.moved_charge += HPAGE_PMD_NR;
5935 putback_lru_page(page);
5938 } else if (target_type == MC_TARGET_DEVICE) {
5940 if (!mem_cgroup_move_account(page, true,
5942 mc.precharge -= HPAGE_PMD_NR;
5943 mc.moved_charge += HPAGE_PMD_NR;
5951 if (pmd_trans_unstable(pmd))
5954 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5955 for (; addr != end; addr += PAGE_SIZE) {
5956 pte_t ptent = *(pte++);
5957 bool device = false;
5963 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5964 case MC_TARGET_DEVICE:
5967 case MC_TARGET_PAGE:
5970 * We can have a part of the split pmd here. Moving it
5971 * can be done but it would be too convoluted so simply
5972 * ignore such a partial THP and keep it in original
5973 * memcg. There should be somebody mapping the head.
5975 if (PageTransCompound(page))
5977 if (!device && isolate_lru_page(page))
5979 if (!mem_cgroup_move_account(page, false,
5982 /* we uncharge from mc.from later. */
5986 putback_lru_page(page);
5987 put: /* get_mctgt_type() gets the page */
5990 case MC_TARGET_SWAP:
5992 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5994 /* we fixup refcnts and charges later. */
6002 pte_unmap_unlock(pte - 1, ptl);
6007 * We have consumed all precharges we got in can_attach().
6008 * We try charge one by one, but don't do any additional
6009 * charges to mc.to if we have failed in charge once in attach()
6012 ret = mem_cgroup_do_precharge(1);
6020 static const struct mm_walk_ops charge_walk_ops = {
6021 .pmd_entry = mem_cgroup_move_charge_pte_range,
6024 static void mem_cgroup_move_charge(void)
6026 lru_add_drain_all();
6028 * Signal lock_page_memcg() to take the memcg's move_lock
6029 * while we're moving its pages to another memcg. Then wait
6030 * for already started RCU-only updates to finish.
6032 atomic_inc(&mc.from->moving_account);
6035 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
6037 * Someone who are holding the mmap_sem might be waiting in
6038 * waitq. So we cancel all extra charges, wake up all waiters,
6039 * and retry. Because we cancel precharges, we might not be able
6040 * to move enough charges, but moving charge is a best-effort
6041 * feature anyway, so it wouldn't be a big problem.
6043 __mem_cgroup_clear_mc();
6048 * When we have consumed all precharges and failed in doing
6049 * additional charge, the page walk just aborts.
6051 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6054 up_read(&mc.mm->mmap_sem);
6055 atomic_dec(&mc.from->moving_account);
6058 static void mem_cgroup_move_task(void)
6061 mem_cgroup_move_charge();
6062 mem_cgroup_clear_mc();
6065 #else /* !CONFIG_MMU */
6066 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6070 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6073 static void mem_cgroup_move_task(void)
6079 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6080 * to verify whether we're attached to the default hierarchy on each mount
6083 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6086 * use_hierarchy is forced on the default hierarchy. cgroup core
6087 * guarantees that @root doesn't have any children, so turning it
6088 * on for the root memcg is enough.
6090 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6091 root_mem_cgroup->use_hierarchy = true;
6093 root_mem_cgroup->use_hierarchy = false;
6096 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6098 if (value == PAGE_COUNTER_MAX)
6099 seq_puts(m, "max\n");
6101 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6106 static u64 memory_current_read(struct cgroup_subsys_state *css,
6109 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6111 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6114 static int memory_min_show(struct seq_file *m, void *v)
6116 return seq_puts_memcg_tunable(m,
6117 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6120 static ssize_t memory_min_write(struct kernfs_open_file *of,
6121 char *buf, size_t nbytes, loff_t off)
6123 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6127 buf = strstrip(buf);
6128 err = page_counter_memparse(buf, "max", &min);
6132 page_counter_set_min(&memcg->memory, min);
6137 static int memory_low_show(struct seq_file *m, void *v)
6139 return seq_puts_memcg_tunable(m,
6140 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6143 static ssize_t memory_low_write(struct kernfs_open_file *of,
6144 char *buf, size_t nbytes, loff_t off)
6146 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6150 buf = strstrip(buf);
6151 err = page_counter_memparse(buf, "max", &low);
6155 page_counter_set_low(&memcg->memory, low);
6160 static int memory_high_show(struct seq_file *m, void *v)
6162 return seq_puts_memcg_tunable(m, READ_ONCE(mem_cgroup_from_seq(m)->high));
6165 static ssize_t memory_high_write(struct kernfs_open_file *of,
6166 char *buf, size_t nbytes, loff_t off)
6168 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6169 unsigned long nr_pages;
6173 buf = strstrip(buf);
6174 err = page_counter_memparse(buf, "max", &high);
6180 nr_pages = page_counter_read(&memcg->memory);
6181 if (nr_pages > high)
6182 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6185 memcg_wb_domain_size_changed(memcg);
6189 static int memory_max_show(struct seq_file *m, void *v)
6191 return seq_puts_memcg_tunable(m,
6192 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6195 static ssize_t memory_max_write(struct kernfs_open_file *of,
6196 char *buf, size_t nbytes, loff_t off)
6198 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6199 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
6200 bool drained = false;
6204 buf = strstrip(buf);
6205 err = page_counter_memparse(buf, "max", &max);
6209 xchg(&memcg->memory.max, max);
6212 unsigned long nr_pages = page_counter_read(&memcg->memory);
6214 if (nr_pages <= max)
6217 if (signal_pending(current)) {
6223 drain_all_stock(memcg);
6229 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6235 memcg_memory_event(memcg, MEMCG_OOM);
6236 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6240 memcg_wb_domain_size_changed(memcg);
6244 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6246 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6247 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6248 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6249 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6250 seq_printf(m, "oom_kill %lu\n",
6251 atomic_long_read(&events[MEMCG_OOM_KILL]));
6254 static int memory_events_show(struct seq_file *m, void *v)
6256 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6258 __memory_events_show(m, memcg->memory_events);
6262 static int memory_events_local_show(struct seq_file *m, void *v)
6264 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6266 __memory_events_show(m, memcg->memory_events_local);
6270 static int memory_stat_show(struct seq_file *m, void *v)
6272 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6275 buf = memory_stat_format(memcg);
6283 static int memory_oom_group_show(struct seq_file *m, void *v)
6285 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6287 seq_printf(m, "%d\n", memcg->oom_group);
6292 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6293 char *buf, size_t nbytes, loff_t off)
6295 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6298 buf = strstrip(buf);
6302 ret = kstrtoint(buf, 0, &oom_group);
6306 if (oom_group != 0 && oom_group != 1)
6309 memcg->oom_group = oom_group;
6314 static struct cftype memory_files[] = {
6317 .flags = CFTYPE_NOT_ON_ROOT,
6318 .read_u64 = memory_current_read,
6322 .flags = CFTYPE_NOT_ON_ROOT,
6323 .seq_show = memory_min_show,
6324 .write = memory_min_write,
6328 .flags = CFTYPE_NOT_ON_ROOT,
6329 .seq_show = memory_low_show,
6330 .write = memory_low_write,
6334 .flags = CFTYPE_NOT_ON_ROOT,
6335 .seq_show = memory_high_show,
6336 .write = memory_high_write,
6340 .flags = CFTYPE_NOT_ON_ROOT,
6341 .seq_show = memory_max_show,
6342 .write = memory_max_write,
6346 .flags = CFTYPE_NOT_ON_ROOT,
6347 .file_offset = offsetof(struct mem_cgroup, events_file),
6348 .seq_show = memory_events_show,
6351 .name = "events.local",
6352 .flags = CFTYPE_NOT_ON_ROOT,
6353 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6354 .seq_show = memory_events_local_show,
6358 .flags = CFTYPE_NOT_ON_ROOT,
6359 .seq_show = memory_stat_show,
6362 .name = "oom.group",
6363 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6364 .seq_show = memory_oom_group_show,
6365 .write = memory_oom_group_write,
6370 struct cgroup_subsys memory_cgrp_subsys = {
6371 .css_alloc = mem_cgroup_css_alloc,
6372 .css_online = mem_cgroup_css_online,
6373 .css_offline = mem_cgroup_css_offline,
6374 .css_released = mem_cgroup_css_released,
6375 .css_free = mem_cgroup_css_free,
6376 .css_reset = mem_cgroup_css_reset,
6377 .can_attach = mem_cgroup_can_attach,
6378 .cancel_attach = mem_cgroup_cancel_attach,
6379 .post_attach = mem_cgroup_move_task,
6380 .bind = mem_cgroup_bind,
6381 .dfl_cftypes = memory_files,
6382 .legacy_cftypes = mem_cgroup_legacy_files,
6387 * mem_cgroup_protected - check if memory consumption is in the normal range
6388 * @root: the top ancestor of the sub-tree being checked
6389 * @memcg: the memory cgroup to check
6391 * WARNING: This function is not stateless! It can only be used as part
6392 * of a top-down tree iteration, not for isolated queries.
6394 * Returns one of the following:
6395 * MEMCG_PROT_NONE: cgroup memory is not protected
6396 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
6397 * an unprotected supply of reclaimable memory from other cgroups.
6398 * MEMCG_PROT_MIN: cgroup memory is protected
6400 * @root is exclusive; it is never protected when looked at directly
6402 * To provide a proper hierarchical behavior, effective memory.min/low values
6403 * are used. Below is the description of how effective memory.low is calculated.
6404 * Effective memory.min values is calculated in the same way.
6406 * Effective memory.low is always equal or less than the original memory.low.
6407 * If there is no memory.low overcommittment (which is always true for
6408 * top-level memory cgroups), these two values are equal.
6409 * Otherwise, it's a part of parent's effective memory.low,
6410 * calculated as a cgroup's memory.low usage divided by sum of sibling's
6411 * memory.low usages, where memory.low usage is the size of actually
6415 * elow = min( memory.low, parent->elow * ------------------ ),
6416 * siblings_low_usage
6418 * | memory.current, if memory.current < memory.low
6423 * Such definition of the effective memory.low provides the expected
6424 * hierarchical behavior: parent's memory.low value is limiting
6425 * children, unprotected memory is reclaimed first and cgroups,
6426 * which are not using their guarantee do not affect actual memory
6429 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
6431 * A A/memory.low = 2G, A/memory.current = 6G
6433 * BC DE B/memory.low = 3G B/memory.current = 2G
6434 * C/memory.low = 1G C/memory.current = 2G
6435 * D/memory.low = 0 D/memory.current = 2G
6436 * E/memory.low = 10G E/memory.current = 0
6438 * and the memory pressure is applied, the following memory distribution
6439 * is expected (approximately):
6441 * A/memory.current = 2G
6443 * B/memory.current = 1.3G
6444 * C/memory.current = 0.6G
6445 * D/memory.current = 0
6446 * E/memory.current = 0
6448 * These calculations require constant tracking of the actual low usages
6449 * (see propagate_protected_usage()), as well as recursive calculation of
6450 * effective memory.low values. But as we do call mem_cgroup_protected()
6451 * path for each memory cgroup top-down from the reclaim,
6452 * it's possible to optimize this part, and save calculated elow
6453 * for next usage. This part is intentionally racy, but it's ok,
6454 * as memory.low is a best-effort mechanism.
6456 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
6457 struct mem_cgroup *memcg)
6459 struct mem_cgroup *parent;
6460 unsigned long emin, parent_emin;
6461 unsigned long elow, parent_elow;
6462 unsigned long usage;
6464 if (mem_cgroup_disabled())
6465 return MEMCG_PROT_NONE;
6468 root = root_mem_cgroup;
6470 return MEMCG_PROT_NONE;
6472 usage = page_counter_read(&memcg->memory);
6474 return MEMCG_PROT_NONE;
6476 emin = memcg->memory.min;
6477 elow = memcg->memory.low;
6479 parent = parent_mem_cgroup(memcg);
6480 /* No parent means a non-hierarchical mode on v1 memcg */
6482 return MEMCG_PROT_NONE;
6487 parent_emin = READ_ONCE(parent->memory.emin);
6488 emin = min(emin, parent_emin);
6489 if (emin && parent_emin) {
6490 unsigned long min_usage, siblings_min_usage;
6492 min_usage = min(usage, memcg->memory.min);
6493 siblings_min_usage = atomic_long_read(
6494 &parent->memory.children_min_usage);
6496 if (min_usage && siblings_min_usage)
6497 emin = min(emin, parent_emin * min_usage /
6498 siblings_min_usage);
6501 parent_elow = READ_ONCE(parent->memory.elow);
6502 elow = min(elow, parent_elow);
6503 if (elow && parent_elow) {
6504 unsigned long low_usage, siblings_low_usage;
6506 low_usage = min(usage, memcg->memory.low);
6507 siblings_low_usage = atomic_long_read(
6508 &parent->memory.children_low_usage);
6510 if (low_usage && siblings_low_usage)
6511 elow = min(elow, parent_elow * low_usage /
6512 siblings_low_usage);
6516 memcg->memory.emin = emin;
6517 memcg->memory.elow = elow;
6520 return MEMCG_PROT_MIN;
6521 else if (usage <= elow)
6522 return MEMCG_PROT_LOW;
6524 return MEMCG_PROT_NONE;
6528 * mem_cgroup_try_charge - try charging a page
6529 * @page: page to charge
6530 * @mm: mm context of the victim
6531 * @gfp_mask: reclaim mode
6532 * @memcgp: charged memcg return
6533 * @compound: charge the page as compound or small page
6535 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6536 * pages according to @gfp_mask if necessary.
6538 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6539 * Otherwise, an error code is returned.
6541 * After page->mapping has been set up, the caller must finalize the
6542 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6543 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6545 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
6546 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6549 struct mem_cgroup *memcg = NULL;
6550 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6553 if (mem_cgroup_disabled())
6556 if (PageSwapCache(page)) {
6558 * Every swap fault against a single page tries to charge the
6559 * page, bail as early as possible. shmem_unuse() encounters
6560 * already charged pages, too. The USED bit is protected by
6561 * the page lock, which serializes swap cache removal, which
6562 * in turn serializes uncharging.
6564 VM_BUG_ON_PAGE(!PageLocked(page), page);
6565 if (compound_head(page)->mem_cgroup)
6568 if (do_swap_account) {
6569 swp_entry_t ent = { .val = page_private(page), };
6570 unsigned short id = lookup_swap_cgroup_id(ent);
6573 memcg = mem_cgroup_from_id(id);
6574 if (memcg && !css_tryget_online(&memcg->css))
6581 memcg = get_mem_cgroup_from_mm(mm);
6583 ret = try_charge(memcg, gfp_mask, nr_pages);
6585 css_put(&memcg->css);
6591 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
6592 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6595 struct mem_cgroup *memcg;
6598 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound);
6600 mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask);
6605 * mem_cgroup_commit_charge - commit a page charge
6606 * @page: page to charge
6607 * @memcg: memcg to charge the page to
6608 * @lrucare: page might be on LRU already
6609 * @compound: charge the page as compound or small page
6611 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6612 * after page->mapping has been set up. This must happen atomically
6613 * as part of the page instantiation, i.e. under the page table lock
6614 * for anonymous pages, under the page lock for page and swap cache.
6616 * In addition, the page must not be on the LRU during the commit, to
6617 * prevent racing with task migration. If it might be, use @lrucare.
6619 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6621 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
6622 bool lrucare, bool compound)
6624 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6626 VM_BUG_ON_PAGE(!page->mapping, page);
6627 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
6629 if (mem_cgroup_disabled())
6632 * Swap faults will attempt to charge the same page multiple
6633 * times. But reuse_swap_page() might have removed the page
6634 * from swapcache already, so we can't check PageSwapCache().
6639 commit_charge(page, memcg, lrucare);
6641 local_irq_disable();
6642 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
6643 memcg_check_events(memcg, page);
6646 if (do_memsw_account() && PageSwapCache(page)) {
6647 swp_entry_t entry = { .val = page_private(page) };
6649 * The swap entry might not get freed for a long time,
6650 * let's not wait for it. The page already received a
6651 * memory+swap charge, drop the swap entry duplicate.
6653 mem_cgroup_uncharge_swap(entry, nr_pages);
6658 * mem_cgroup_cancel_charge - cancel a page charge
6659 * @page: page to charge
6660 * @memcg: memcg to charge the page to
6661 * @compound: charge the page as compound or small page
6663 * Cancel a charge transaction started by mem_cgroup_try_charge().
6665 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
6668 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6670 if (mem_cgroup_disabled())
6673 * Swap faults will attempt to charge the same page multiple
6674 * times. But reuse_swap_page() might have removed the page
6675 * from swapcache already, so we can't check PageSwapCache().
6680 cancel_charge(memcg, nr_pages);
6683 struct uncharge_gather {
6684 struct mem_cgroup *memcg;
6685 unsigned long pgpgout;
6686 unsigned long nr_anon;
6687 unsigned long nr_file;
6688 unsigned long nr_kmem;
6689 unsigned long nr_huge;
6690 unsigned long nr_shmem;
6691 struct page *dummy_page;
6694 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6696 memset(ug, 0, sizeof(*ug));
6699 static void uncharge_batch(const struct uncharge_gather *ug)
6701 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
6702 unsigned long flags;
6704 if (!mem_cgroup_is_root(ug->memcg)) {
6705 page_counter_uncharge(&ug->memcg->memory, nr_pages);
6706 if (do_memsw_account())
6707 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
6708 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6709 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6710 memcg_oom_recover(ug->memcg);
6713 local_irq_save(flags);
6714 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
6715 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
6716 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
6717 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
6718 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6719 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, nr_pages);
6720 memcg_check_events(ug->memcg, ug->dummy_page);
6721 local_irq_restore(flags);
6723 if (!mem_cgroup_is_root(ug->memcg))
6724 css_put_many(&ug->memcg->css, nr_pages);
6727 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6729 VM_BUG_ON_PAGE(PageLRU(page), page);
6730 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
6731 !PageHWPoison(page) , page);
6733 if (!page->mem_cgroup)
6737 * Nobody should be changing or seriously looking at
6738 * page->mem_cgroup at this point, we have fully
6739 * exclusive access to the page.
6742 if (ug->memcg != page->mem_cgroup) {
6745 uncharge_gather_clear(ug);
6747 ug->memcg = page->mem_cgroup;
6750 if (!PageKmemcg(page)) {
6751 unsigned int nr_pages = 1;
6753 if (PageTransHuge(page)) {
6754 nr_pages = compound_nr(page);
6755 ug->nr_huge += nr_pages;
6758 ug->nr_anon += nr_pages;
6760 ug->nr_file += nr_pages;
6761 if (PageSwapBacked(page))
6762 ug->nr_shmem += nr_pages;
6766 ug->nr_kmem += compound_nr(page);
6767 __ClearPageKmemcg(page);
6770 ug->dummy_page = page;
6771 page->mem_cgroup = NULL;
6774 static void uncharge_list(struct list_head *page_list)
6776 struct uncharge_gather ug;
6777 struct list_head *next;
6779 uncharge_gather_clear(&ug);
6782 * Note that the list can be a single page->lru; hence the
6783 * do-while loop instead of a simple list_for_each_entry().
6785 next = page_list->next;
6789 page = list_entry(next, struct page, lru);
6790 next = page->lru.next;
6792 uncharge_page(page, &ug);
6793 } while (next != page_list);
6796 uncharge_batch(&ug);
6800 * mem_cgroup_uncharge - uncharge a page
6801 * @page: page to uncharge
6803 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6804 * mem_cgroup_commit_charge().
6806 void mem_cgroup_uncharge(struct page *page)
6808 struct uncharge_gather ug;
6810 if (mem_cgroup_disabled())
6813 /* Don't touch page->lru of any random page, pre-check: */
6814 if (!page->mem_cgroup)
6817 uncharge_gather_clear(&ug);
6818 uncharge_page(page, &ug);
6819 uncharge_batch(&ug);
6823 * mem_cgroup_uncharge_list - uncharge a list of page
6824 * @page_list: list of pages to uncharge
6826 * Uncharge a list of pages previously charged with
6827 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6829 void mem_cgroup_uncharge_list(struct list_head *page_list)
6831 if (mem_cgroup_disabled())
6834 if (!list_empty(page_list))
6835 uncharge_list(page_list);
6839 * mem_cgroup_migrate - charge a page's replacement
6840 * @oldpage: currently circulating page
6841 * @newpage: replacement page
6843 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6844 * be uncharged upon free.
6846 * Both pages must be locked, @newpage->mapping must be set up.
6848 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6850 struct mem_cgroup *memcg;
6851 unsigned int nr_pages;
6853 unsigned long flags;
6855 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6856 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6857 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6858 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6861 if (mem_cgroup_disabled())
6864 /* Page cache replacement: new page already charged? */
6865 if (newpage->mem_cgroup)
6868 /* Swapcache readahead pages can get replaced before being charged */
6869 memcg = oldpage->mem_cgroup;
6873 /* Force-charge the new page. The old one will be freed soon */
6874 compound = PageTransHuge(newpage);
6875 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
6877 page_counter_charge(&memcg->memory, nr_pages);
6878 if (do_memsw_account())
6879 page_counter_charge(&memcg->memsw, nr_pages);
6880 css_get_many(&memcg->css, nr_pages);
6882 commit_charge(newpage, memcg, false);
6884 local_irq_save(flags);
6885 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
6886 memcg_check_events(memcg, newpage);
6887 local_irq_restore(flags);
6890 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6891 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6893 void mem_cgroup_sk_alloc(struct sock *sk)
6895 struct mem_cgroup *memcg;
6897 if (!mem_cgroup_sockets_enabled)
6900 /* Do not associate the sock with unrelated interrupted task's memcg. */
6905 memcg = mem_cgroup_from_task(current);
6906 if (memcg == root_mem_cgroup)
6908 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6910 if (css_tryget_online(&memcg->css))
6911 sk->sk_memcg = memcg;
6916 void mem_cgroup_sk_free(struct sock *sk)
6919 css_put(&sk->sk_memcg->css);
6923 * mem_cgroup_charge_skmem - charge socket memory
6924 * @memcg: memcg to charge
6925 * @nr_pages: number of pages to charge
6927 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6928 * @memcg's configured limit, %false if the charge had to be forced.
6930 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6932 gfp_t gfp_mask = GFP_KERNEL;
6934 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6935 struct page_counter *fail;
6937 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6938 memcg->tcpmem_pressure = 0;
6941 page_counter_charge(&memcg->tcpmem, nr_pages);
6942 memcg->tcpmem_pressure = 1;
6946 /* Don't block in the packet receive path */
6948 gfp_mask = GFP_NOWAIT;
6950 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6952 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6955 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6960 * mem_cgroup_uncharge_skmem - uncharge socket memory
6961 * @memcg: memcg to uncharge
6962 * @nr_pages: number of pages to uncharge
6964 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6966 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6967 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6971 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6973 refill_stock(memcg, nr_pages);
6976 static int __init cgroup_memory(char *s)
6980 while ((token = strsep(&s, ",")) != NULL) {
6983 if (!strcmp(token, "nosocket"))
6984 cgroup_memory_nosocket = true;
6985 if (!strcmp(token, "nokmem"))
6986 cgroup_memory_nokmem = true;
6990 __setup("cgroup.memory=", cgroup_memory);
6993 * subsys_initcall() for memory controller.
6995 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6996 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6997 * basically everything that doesn't depend on a specific mem_cgroup structure
6998 * should be initialized from here.
7000 static int __init mem_cgroup_init(void)
7004 #ifdef CONFIG_MEMCG_KMEM
7006 * Kmem cache creation is mostly done with the slab_mutex held,
7007 * so use a workqueue with limited concurrency to avoid stalling
7008 * all worker threads in case lots of cgroups are created and
7009 * destroyed simultaneously.
7011 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
7012 BUG_ON(!memcg_kmem_cache_wq);
7015 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7016 memcg_hotplug_cpu_dead);
7018 for_each_possible_cpu(cpu)
7019 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7022 for_each_node(node) {
7023 struct mem_cgroup_tree_per_node *rtpn;
7025 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7026 node_online(node) ? node : NUMA_NO_NODE);
7028 rtpn->rb_root = RB_ROOT;
7029 rtpn->rb_rightmost = NULL;
7030 spin_lock_init(&rtpn->lock);
7031 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7036 subsys_initcall(mem_cgroup_init);
7038 #ifdef CONFIG_MEMCG_SWAP
7039 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7041 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7043 * The root cgroup cannot be destroyed, so it's refcount must
7046 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7050 memcg = parent_mem_cgroup(memcg);
7052 memcg = root_mem_cgroup;
7058 * mem_cgroup_swapout - transfer a memsw charge to swap
7059 * @page: page whose memsw charge to transfer
7060 * @entry: swap entry to move the charge to
7062 * Transfer the memsw charge of @page to @entry.
7064 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7066 struct mem_cgroup *memcg, *swap_memcg;
7067 unsigned int nr_entries;
7068 unsigned short oldid;
7070 VM_BUG_ON_PAGE(PageLRU(page), page);
7071 VM_BUG_ON_PAGE(page_count(page), page);
7073 if (!do_memsw_account())
7076 memcg = page->mem_cgroup;
7078 /* Readahead page, never charged */
7083 * In case the memcg owning these pages has been offlined and doesn't
7084 * have an ID allocated to it anymore, charge the closest online
7085 * ancestor for the swap instead and transfer the memory+swap charge.
7087 swap_memcg = mem_cgroup_id_get_online(memcg);
7088 nr_entries = hpage_nr_pages(page);
7089 /* Get references for the tail pages, too */
7091 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7092 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7094 VM_BUG_ON_PAGE(oldid, page);
7095 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7097 page->mem_cgroup = NULL;
7099 if (!mem_cgroup_is_root(memcg))
7100 page_counter_uncharge(&memcg->memory, nr_entries);
7102 if (memcg != swap_memcg) {
7103 if (!mem_cgroup_is_root(swap_memcg))
7104 page_counter_charge(&swap_memcg->memsw, nr_entries);
7105 page_counter_uncharge(&memcg->memsw, nr_entries);
7109 * Interrupts should be disabled here because the caller holds the
7110 * i_pages lock which is taken with interrupts-off. It is
7111 * important here to have the interrupts disabled because it is the
7112 * only synchronisation we have for updating the per-CPU variables.
7114 VM_BUG_ON(!irqs_disabled());
7115 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
7117 memcg_check_events(memcg, page);
7119 if (!mem_cgroup_is_root(memcg))
7120 css_put_many(&memcg->css, nr_entries);
7124 * mem_cgroup_try_charge_swap - try charging swap space for a page
7125 * @page: page being added to swap
7126 * @entry: swap entry to charge
7128 * Try to charge @page's memcg for the swap space at @entry.
7130 * Returns 0 on success, -ENOMEM on failure.
7132 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7134 unsigned int nr_pages = hpage_nr_pages(page);
7135 struct page_counter *counter;
7136 struct mem_cgroup *memcg;
7137 unsigned short oldid;
7139 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
7142 memcg = page->mem_cgroup;
7144 /* Readahead page, never charged */
7149 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7153 memcg = mem_cgroup_id_get_online(memcg);
7155 if (!mem_cgroup_is_root(memcg) &&
7156 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7157 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7158 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7159 mem_cgroup_id_put(memcg);
7163 /* Get references for the tail pages, too */
7165 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7166 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7167 VM_BUG_ON_PAGE(oldid, page);
7168 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7174 * mem_cgroup_uncharge_swap - uncharge swap space
7175 * @entry: swap entry to uncharge
7176 * @nr_pages: the amount of swap space to uncharge
7178 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7180 struct mem_cgroup *memcg;
7183 if (!do_swap_account)
7186 id = swap_cgroup_record(entry, 0, nr_pages);
7188 memcg = mem_cgroup_from_id(id);
7190 if (!mem_cgroup_is_root(memcg)) {
7191 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7192 page_counter_uncharge(&memcg->swap, nr_pages);
7194 page_counter_uncharge(&memcg->memsw, nr_pages);
7196 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7197 mem_cgroup_id_put_many(memcg, nr_pages);
7202 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7204 long nr_swap_pages = get_nr_swap_pages();
7206 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7207 return nr_swap_pages;
7208 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7209 nr_swap_pages = min_t(long, nr_swap_pages,
7210 READ_ONCE(memcg->swap.max) -
7211 page_counter_read(&memcg->swap));
7212 return nr_swap_pages;
7215 bool mem_cgroup_swap_full(struct page *page)
7217 struct mem_cgroup *memcg;
7219 VM_BUG_ON_PAGE(!PageLocked(page), page);
7223 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7226 memcg = page->mem_cgroup;
7230 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7231 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.max)
7237 /* for remember boot option*/
7238 #ifdef CONFIG_MEMCG_SWAP_ENABLED
7239 static int really_do_swap_account __initdata = 1;
7241 static int really_do_swap_account __initdata;
7244 static int __init enable_swap_account(char *s)
7246 if (!strcmp(s, "1"))
7247 really_do_swap_account = 1;
7248 else if (!strcmp(s, "0"))
7249 really_do_swap_account = 0;
7252 __setup("swapaccount=", enable_swap_account);
7254 static u64 swap_current_read(struct cgroup_subsys_state *css,
7257 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7259 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7262 static int swap_max_show(struct seq_file *m, void *v)
7264 return seq_puts_memcg_tunable(m,
7265 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7268 static ssize_t swap_max_write(struct kernfs_open_file *of,
7269 char *buf, size_t nbytes, loff_t off)
7271 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7275 buf = strstrip(buf);
7276 err = page_counter_memparse(buf, "max", &max);
7280 xchg(&memcg->swap.max, max);
7285 static int swap_events_show(struct seq_file *m, void *v)
7287 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7289 seq_printf(m, "max %lu\n",
7290 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7291 seq_printf(m, "fail %lu\n",
7292 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7297 static struct cftype swap_files[] = {
7299 .name = "swap.current",
7300 .flags = CFTYPE_NOT_ON_ROOT,
7301 .read_u64 = swap_current_read,
7305 .flags = CFTYPE_NOT_ON_ROOT,
7306 .seq_show = swap_max_show,
7307 .write = swap_max_write,
7310 .name = "swap.events",
7311 .flags = CFTYPE_NOT_ON_ROOT,
7312 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7313 .seq_show = swap_events_show,
7318 static struct cftype memsw_cgroup_files[] = {
7320 .name = "memsw.usage_in_bytes",
7321 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7322 .read_u64 = mem_cgroup_read_u64,
7325 .name = "memsw.max_usage_in_bytes",
7326 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7327 .write = mem_cgroup_reset,
7328 .read_u64 = mem_cgroup_read_u64,
7331 .name = "memsw.limit_in_bytes",
7332 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7333 .write = mem_cgroup_write,
7334 .read_u64 = mem_cgroup_read_u64,
7337 .name = "memsw.failcnt",
7338 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7339 .write = mem_cgroup_reset,
7340 .read_u64 = mem_cgroup_read_u64,
7343 .name = "force_reclaim",
7344 .write_u64 = mem_cgroup_force_reclaim,
7346 { }, /* terminate */
7349 static int __init mem_cgroup_swap_init(void)
7351 if (!mem_cgroup_disabled() && really_do_swap_account) {
7352 do_swap_account = 1;
7353 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
7355 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
7356 memsw_cgroup_files));
7360 subsys_initcall(mem_cgroup_swap_init);
7362 #endif /* CONFIG_MEMCG_SWAP */