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 #define THRESHOLDS_EVENTS_TARGET 128
102 #define SOFTLIMIT_EVENTS_TARGET 1024
105 * Cgroups above their limits are maintained in a RB-Tree, independent of
106 * their hierarchy representation
109 struct mem_cgroup_tree_per_node {
110 struct rb_root rb_root;
111 struct rb_node *rb_rightmost;
115 struct mem_cgroup_tree {
116 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
119 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
122 struct mem_cgroup_eventfd_list {
123 struct list_head list;
124 struct eventfd_ctx *eventfd;
128 * cgroup_event represents events which userspace want to receive.
130 struct mem_cgroup_event {
132 * memcg which the event belongs to.
134 struct mem_cgroup *memcg;
136 * eventfd to signal userspace about the event.
138 struct eventfd_ctx *eventfd;
140 * Each of these stored in a list by the cgroup.
142 struct list_head list;
144 * register_event() callback will be used to add new userspace
145 * waiter for changes related to this event. Use eventfd_signal()
146 * on eventfd to send notification to userspace.
148 int (*register_event)(struct mem_cgroup *memcg,
149 struct eventfd_ctx *eventfd, const char *args);
151 * unregister_event() callback will be called when userspace closes
152 * the eventfd or on cgroup removing. This callback must be set,
153 * if you want provide notification functionality.
155 void (*unregister_event)(struct mem_cgroup *memcg,
156 struct eventfd_ctx *eventfd);
158 * All fields below needed to unregister event when
159 * userspace closes eventfd.
162 wait_queue_head_t *wqh;
163 wait_queue_entry_t wait;
164 struct work_struct remove;
167 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
168 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
170 /* Stuffs for move charges at task migration. */
172 * Types of charges to be moved.
174 #define MOVE_ANON 0x1U
175 #define MOVE_FILE 0x2U
176 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
178 /* "mc" and its members are protected by cgroup_mutex */
179 static struct move_charge_struct {
180 spinlock_t lock; /* for from, to */
181 struct mm_struct *mm;
182 struct mem_cgroup *from;
183 struct mem_cgroup *to;
185 unsigned long precharge;
186 unsigned long moved_charge;
187 unsigned long moved_swap;
188 struct task_struct *moving_task; /* a task moving charges */
189 wait_queue_head_t waitq; /* a waitq for other context */
191 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
192 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
196 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
197 * limit reclaim to prevent infinite loops, if they ever occur.
199 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
200 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
203 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
204 MEM_CGROUP_CHARGE_TYPE_ANON,
205 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
206 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
210 /* for encoding cft->private value on file */
219 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
220 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
221 #define MEMFILE_ATTR(val) ((val) & 0xffff)
222 /* Used for OOM nofiier */
223 #define OOM_CONTROL (0)
226 * Iteration constructs for visiting all cgroups (under a tree). If
227 * loops are exited prematurely (break), mem_cgroup_iter_break() must
228 * be used for reference counting.
230 #define for_each_mem_cgroup_tree(iter, root) \
231 for (iter = mem_cgroup_iter(root, NULL, NULL); \
233 iter = mem_cgroup_iter(root, iter, NULL))
235 #define for_each_mem_cgroup(iter) \
236 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
238 iter = mem_cgroup_iter(NULL, iter, NULL))
240 static inline bool should_force_charge(void)
242 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
243 (current->flags & PF_EXITING);
246 /* Some nice accessors for the vmpressure. */
247 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
250 memcg = root_mem_cgroup;
251 return &memcg->vmpressure;
254 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
256 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
259 #ifdef CONFIG_MEMCG_KMEM
261 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
262 * The main reason for not using cgroup id for this:
263 * this works better in sparse environments, where we have a lot of memcgs,
264 * but only a few kmem-limited. Or also, if we have, for instance, 200
265 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
266 * 200 entry array for that.
268 * The current size of the caches array is stored in memcg_nr_cache_ids. It
269 * will double each time we have to increase it.
271 static DEFINE_IDA(memcg_cache_ida);
272 int memcg_nr_cache_ids;
274 /* Protects memcg_nr_cache_ids */
275 static DECLARE_RWSEM(memcg_cache_ids_sem);
277 void memcg_get_cache_ids(void)
279 down_read(&memcg_cache_ids_sem);
282 void memcg_put_cache_ids(void)
284 up_read(&memcg_cache_ids_sem);
288 * MIN_SIZE is different than 1, because we would like to avoid going through
289 * the alloc/free process all the time. In a small machine, 4 kmem-limited
290 * cgroups is a reasonable guess. In the future, it could be a parameter or
291 * tunable, but that is strictly not necessary.
293 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
294 * this constant directly from cgroup, but it is understandable that this is
295 * better kept as an internal representation in cgroup.c. In any case, the
296 * cgrp_id space is not getting any smaller, and we don't have to necessarily
297 * increase ours as well if it increases.
299 #define MEMCG_CACHES_MIN_SIZE 4
300 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
303 * A lot of the calls to the cache allocation functions are expected to be
304 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
305 * conditional to this static branch, we'll have to allow modules that does
306 * kmem_cache_alloc and the such to see this symbol as well
308 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
309 EXPORT_SYMBOL(memcg_kmem_enabled_key);
311 struct workqueue_struct *memcg_kmem_cache_wq;
314 static int memcg_shrinker_map_size;
315 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
317 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
319 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
322 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
323 int size, int old_size)
325 struct memcg_shrinker_map *new, *old;
328 lockdep_assert_held(&memcg_shrinker_map_mutex);
331 old = rcu_dereference_protected(
332 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
333 /* Not yet online memcg */
337 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
341 /* Set all old bits, clear all new bits */
342 memset(new->map, (int)0xff, old_size);
343 memset((void *)new->map + old_size, 0, size - old_size);
345 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
346 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
352 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
354 struct mem_cgroup_per_node *pn;
355 struct memcg_shrinker_map *map;
358 if (mem_cgroup_is_root(memcg))
362 pn = mem_cgroup_nodeinfo(memcg, nid);
363 map = rcu_dereference_protected(pn->shrinker_map, true);
366 rcu_assign_pointer(pn->shrinker_map, NULL);
370 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
372 struct memcg_shrinker_map *map;
373 int nid, size, ret = 0;
375 if (mem_cgroup_is_root(memcg))
378 mutex_lock(&memcg_shrinker_map_mutex);
379 size = memcg_shrinker_map_size;
381 map = kvzalloc_node(sizeof(*map) + size, GFP_KERNEL, nid);
383 memcg_free_shrinker_maps(memcg);
387 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
389 mutex_unlock(&memcg_shrinker_map_mutex);
394 int memcg_expand_shrinker_maps(int new_id)
396 int size, old_size, ret = 0;
397 struct mem_cgroup *memcg;
399 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
400 old_size = memcg_shrinker_map_size;
401 if (size <= old_size)
404 mutex_lock(&memcg_shrinker_map_mutex);
405 if (!root_mem_cgroup)
408 for_each_mem_cgroup(memcg) {
409 if (mem_cgroup_is_root(memcg))
411 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
413 mem_cgroup_iter_break(NULL, memcg);
419 memcg_shrinker_map_size = size;
420 mutex_unlock(&memcg_shrinker_map_mutex);
424 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
426 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
427 struct memcg_shrinker_map *map;
430 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
431 /* Pairs with smp mb in shrink_slab() */
432 smp_mb__before_atomic();
433 set_bit(shrinker_id, map->map);
439 * mem_cgroup_css_from_page - css of the memcg associated with a page
440 * @page: page of interest
442 * If memcg is bound to the default hierarchy, css of the memcg associated
443 * with @page is returned. The returned css remains associated with @page
444 * until it is released.
446 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
449 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
451 struct mem_cgroup *memcg;
453 memcg = page->mem_cgroup;
455 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
456 memcg = root_mem_cgroup;
462 * page_cgroup_ino - return inode number of the memcg a page is charged to
465 * Look up the closest online ancestor of the memory cgroup @page is charged to
466 * and return its inode number or 0 if @page is not charged to any cgroup. It
467 * is safe to call this function without holding a reference to @page.
469 * Note, this function is inherently racy, because there is nothing to prevent
470 * the cgroup inode from getting torn down and potentially reallocated a moment
471 * after page_cgroup_ino() returns, so it only should be used by callers that
472 * do not care (such as procfs interfaces).
474 ino_t page_cgroup_ino(struct page *page)
476 struct mem_cgroup *memcg;
477 unsigned long ino = 0;
480 if (PageSlab(page) && !PageTail(page))
481 memcg = memcg_from_slab_page(page);
483 memcg = READ_ONCE(page->mem_cgroup);
484 while (memcg && !(memcg->css.flags & CSS_ONLINE))
485 memcg = parent_mem_cgroup(memcg);
487 ino = cgroup_ino(memcg->css.cgroup);
492 static struct mem_cgroup_per_node *
493 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
495 int nid = page_to_nid(page);
497 return memcg->nodeinfo[nid];
500 static struct mem_cgroup_tree_per_node *
501 soft_limit_tree_node(int nid)
503 return soft_limit_tree.rb_tree_per_node[nid];
506 static struct mem_cgroup_tree_per_node *
507 soft_limit_tree_from_page(struct page *page)
509 int nid = page_to_nid(page);
511 return soft_limit_tree.rb_tree_per_node[nid];
514 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
515 struct mem_cgroup_tree_per_node *mctz,
516 unsigned long new_usage_in_excess)
518 struct rb_node **p = &mctz->rb_root.rb_node;
519 struct rb_node *parent = NULL;
520 struct mem_cgroup_per_node *mz_node;
521 bool rightmost = true;
526 mz->usage_in_excess = new_usage_in_excess;
527 if (!mz->usage_in_excess)
531 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
533 if (mz->usage_in_excess < mz_node->usage_in_excess) {
539 * We can't avoid mem cgroups that are over their soft
540 * limit by the same amount
542 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
547 mctz->rb_rightmost = &mz->tree_node;
549 rb_link_node(&mz->tree_node, parent, p);
550 rb_insert_color(&mz->tree_node, &mctz->rb_root);
554 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
555 struct mem_cgroup_tree_per_node *mctz)
560 if (&mz->tree_node == mctz->rb_rightmost)
561 mctz->rb_rightmost = rb_prev(&mz->tree_node);
563 rb_erase(&mz->tree_node, &mctz->rb_root);
567 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
568 struct mem_cgroup_tree_per_node *mctz)
572 spin_lock_irqsave(&mctz->lock, flags);
573 __mem_cgroup_remove_exceeded(mz, mctz);
574 spin_unlock_irqrestore(&mctz->lock, flags);
577 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
579 unsigned long nr_pages = page_counter_read(&memcg->memory);
580 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
581 unsigned long excess = 0;
583 if (nr_pages > soft_limit)
584 excess = nr_pages - soft_limit;
589 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
591 unsigned long excess;
592 struct mem_cgroup_per_node *mz;
593 struct mem_cgroup_tree_per_node *mctz;
595 mctz = soft_limit_tree_from_page(page);
599 * Necessary to update all ancestors when hierarchy is used.
600 * because their event counter is not touched.
602 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
603 mz = mem_cgroup_page_nodeinfo(memcg, page);
604 excess = soft_limit_excess(memcg);
606 * We have to update the tree if mz is on RB-tree or
607 * mem is over its softlimit.
609 if (excess || mz->on_tree) {
612 spin_lock_irqsave(&mctz->lock, flags);
613 /* if on-tree, remove it */
615 __mem_cgroup_remove_exceeded(mz, mctz);
617 * Insert again. mz->usage_in_excess will be updated.
618 * If excess is 0, no tree ops.
620 __mem_cgroup_insert_exceeded(mz, mctz, excess);
621 spin_unlock_irqrestore(&mctz->lock, flags);
626 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
628 struct mem_cgroup_tree_per_node *mctz;
629 struct mem_cgroup_per_node *mz;
633 mz = mem_cgroup_nodeinfo(memcg, nid);
634 mctz = soft_limit_tree_node(nid);
636 mem_cgroup_remove_exceeded(mz, mctz);
640 static struct mem_cgroup_per_node *
641 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
643 struct mem_cgroup_per_node *mz;
647 if (!mctz->rb_rightmost)
648 goto done; /* Nothing to reclaim from */
650 mz = rb_entry(mctz->rb_rightmost,
651 struct mem_cgroup_per_node, tree_node);
653 * Remove the node now but someone else can add it back,
654 * we will to add it back at the end of reclaim to its correct
655 * position in the tree.
657 __mem_cgroup_remove_exceeded(mz, mctz);
658 if (!soft_limit_excess(mz->memcg) ||
659 !css_tryget(&mz->memcg->css))
665 static struct mem_cgroup_per_node *
666 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
668 struct mem_cgroup_per_node *mz;
670 spin_lock_irq(&mctz->lock);
671 mz = __mem_cgroup_largest_soft_limit_node(mctz);
672 spin_unlock_irq(&mctz->lock);
677 * __mod_memcg_state - update cgroup memory statistics
678 * @memcg: the memory cgroup
679 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
680 * @val: delta to add to the counter, can be negative
682 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
686 if (mem_cgroup_disabled())
689 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
690 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
691 struct mem_cgroup *mi;
694 * Batch local counters to keep them in sync with
695 * the hierarchical ones.
697 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
698 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
699 atomic_long_add(x, &mi->vmstats[idx]);
702 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
705 static struct mem_cgroup_per_node *
706 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
708 struct mem_cgroup *parent;
710 parent = parent_mem_cgroup(pn->memcg);
713 return mem_cgroup_nodeinfo(parent, nid);
717 * __mod_lruvec_state - update lruvec memory statistics
718 * @lruvec: the lruvec
719 * @idx: the stat item
720 * @val: delta to add to the counter, can be negative
722 * The lruvec is the intersection of the NUMA node and a cgroup. This
723 * function updates the all three counters that are affected by a
724 * change of state at this level: per-node, per-cgroup, per-lruvec.
726 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
729 pg_data_t *pgdat = lruvec_pgdat(lruvec);
730 struct mem_cgroup_per_node *pn;
731 struct mem_cgroup *memcg;
735 __mod_node_page_state(pgdat, idx, val);
737 if (mem_cgroup_disabled())
740 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
744 __mod_memcg_state(memcg, idx, val);
747 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
749 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
750 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
751 struct mem_cgroup_per_node *pi;
753 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
754 atomic_long_add(x, &pi->lruvec_stat[idx]);
757 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
760 void __mod_lruvec_slab_state(void *p, enum node_stat_item idx, int val)
762 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
763 struct mem_cgroup *memcg;
764 struct lruvec *lruvec;
767 memcg = mem_cgroup_from_obj(p);
769 /* Untracked pages have no memcg, no lruvec. Update only the node */
770 if (!memcg || memcg == root_mem_cgroup) {
771 __mod_node_page_state(pgdat, idx, val);
773 lruvec = mem_cgroup_lruvec(memcg, pgdat);
774 __mod_lruvec_state(lruvec, idx, val);
779 void mod_memcg_obj_state(void *p, int idx, int val)
781 struct mem_cgroup *memcg;
784 memcg = mem_cgroup_from_obj(p);
786 mod_memcg_state(memcg, idx, val);
791 * __count_memcg_events - account VM events in a cgroup
792 * @memcg: the memory cgroup
793 * @idx: the event item
794 * @count: the number of events that occured
796 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
801 if (mem_cgroup_disabled())
804 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
805 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
806 struct mem_cgroup *mi;
809 * Batch local counters to keep them in sync with
810 * the hierarchical ones.
812 __this_cpu_add(memcg->vmstats_local->events[idx], x);
813 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
814 atomic_long_add(x, &mi->vmevents[idx]);
817 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
820 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
822 return atomic_long_read(&memcg->vmevents[event]);
825 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
830 for_each_possible_cpu(cpu)
831 x += per_cpu(memcg->vmstats_local->events[event], cpu);
835 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
837 bool compound, int nr_pages)
840 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
841 * counted as CACHE even if it's on ANON LRU.
844 __mod_memcg_state(memcg, MEMCG_RSS, nr_pages);
846 __mod_memcg_state(memcg, MEMCG_CACHE, nr_pages);
847 if (PageSwapBacked(page))
848 __mod_memcg_state(memcg, NR_SHMEM, nr_pages);
852 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
853 __mod_memcg_state(memcg, MEMCG_RSS_HUGE, nr_pages);
856 /* pagein of a big page is an event. So, ignore page size */
858 __count_memcg_events(memcg, PGPGIN, 1);
860 __count_memcg_events(memcg, PGPGOUT, 1);
861 nr_pages = -nr_pages; /* for event */
864 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
867 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
868 enum mem_cgroup_events_target target)
870 unsigned long val, next;
872 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
873 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
874 /* from time_after() in jiffies.h */
875 if ((long)(next - val) < 0) {
877 case MEM_CGROUP_TARGET_THRESH:
878 next = val + THRESHOLDS_EVENTS_TARGET;
880 case MEM_CGROUP_TARGET_SOFTLIMIT:
881 next = val + SOFTLIMIT_EVENTS_TARGET;
886 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
893 * Check events in order.
896 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
898 /* threshold event is triggered in finer grain than soft limit */
899 if (unlikely(mem_cgroup_event_ratelimit(memcg,
900 MEM_CGROUP_TARGET_THRESH))) {
903 do_softlimit = mem_cgroup_event_ratelimit(memcg,
904 MEM_CGROUP_TARGET_SOFTLIMIT);
905 mem_cgroup_threshold(memcg);
906 if (unlikely(do_softlimit))
907 mem_cgroup_update_tree(memcg, page);
911 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
914 * mm_update_next_owner() may clear mm->owner to NULL
915 * if it races with swapoff, page migration, etc.
916 * So this can be called with p == NULL.
921 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
923 EXPORT_SYMBOL(mem_cgroup_from_task);
926 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
927 * @mm: mm from which memcg should be extracted. It can be NULL.
929 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
930 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
933 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
935 struct mem_cgroup *memcg;
937 if (mem_cgroup_disabled())
943 * Page cache insertions can happen withou an
944 * actual mm context, e.g. during disk probing
945 * on boot, loopback IO, acct() writes etc.
948 memcg = root_mem_cgroup;
950 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
951 if (unlikely(!memcg))
952 memcg = root_mem_cgroup;
954 } while (!css_tryget(&memcg->css));
958 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
961 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
962 * @page: page from which memcg should be extracted.
964 * Obtain a reference on page->memcg and returns it if successful. Otherwise
965 * root_mem_cgroup is returned.
967 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
969 struct mem_cgroup *memcg = page->mem_cgroup;
971 if (mem_cgroup_disabled())
975 /* Page should not get uncharged and freed memcg under us. */
976 if (!memcg || WARN_ON_ONCE(!css_tryget(&memcg->css)))
977 memcg = root_mem_cgroup;
981 EXPORT_SYMBOL(get_mem_cgroup_from_page);
984 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
986 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
988 if (unlikely(current->active_memcg)) {
989 struct mem_cgroup *memcg;
992 /* current->active_memcg must hold a ref. */
993 if (WARN_ON_ONCE(!css_tryget(¤t->active_memcg->css)))
994 memcg = root_mem_cgroup;
996 memcg = current->active_memcg;
1000 return get_mem_cgroup_from_mm(current->mm);
1004 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1005 * @root: hierarchy root
1006 * @prev: previously returned memcg, NULL on first invocation
1007 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1009 * Returns references to children of the hierarchy below @root, or
1010 * @root itself, or %NULL after a full round-trip.
1012 * Caller must pass the return value in @prev on subsequent
1013 * invocations for reference counting, or use mem_cgroup_iter_break()
1014 * to cancel a hierarchy walk before the round-trip is complete.
1016 * Reclaimers can specify a node and a priority level in @reclaim to
1017 * divide up the memcgs in the hierarchy among all concurrent
1018 * reclaimers operating on the same node and priority.
1020 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1021 struct mem_cgroup *prev,
1022 struct mem_cgroup_reclaim_cookie *reclaim)
1024 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1025 struct cgroup_subsys_state *css = NULL;
1026 struct mem_cgroup *memcg = NULL;
1027 struct mem_cgroup *pos = NULL;
1029 if (mem_cgroup_disabled())
1033 root = root_mem_cgroup;
1035 if (prev && !reclaim)
1038 if (!root->use_hierarchy && root != root_mem_cgroup) {
1047 struct mem_cgroup_per_node *mz;
1049 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1052 if (prev && reclaim->generation != iter->generation)
1056 pos = READ_ONCE(iter->position);
1057 if (!pos || css_tryget(&pos->css))
1060 * css reference reached zero, so iter->position will
1061 * be cleared by ->css_released. However, we should not
1062 * rely on this happening soon, because ->css_released
1063 * is called from a work queue, and by busy-waiting we
1064 * might block it. So we clear iter->position right
1067 (void)cmpxchg(&iter->position, pos, NULL);
1075 css = css_next_descendant_pre(css, &root->css);
1078 * Reclaimers share the hierarchy walk, and a
1079 * new one might jump in right at the end of
1080 * the hierarchy - make sure they see at least
1081 * one group and restart from the beginning.
1089 * Verify the css and acquire a reference. The root
1090 * is provided by the caller, so we know it's alive
1091 * and kicking, and don't take an extra reference.
1093 memcg = mem_cgroup_from_css(css);
1095 if (css == &root->css)
1098 if (css_tryget(css))
1106 * The position could have already been updated by a competing
1107 * thread, so check that the value hasn't changed since we read
1108 * it to avoid reclaiming from the same cgroup twice.
1110 (void)cmpxchg(&iter->position, pos, memcg);
1118 reclaim->generation = iter->generation;
1124 if (prev && prev != root)
1125 css_put(&prev->css);
1131 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1132 * @root: hierarchy root
1133 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1135 void mem_cgroup_iter_break(struct mem_cgroup *root,
1136 struct mem_cgroup *prev)
1139 root = root_mem_cgroup;
1140 if (prev && prev != root)
1141 css_put(&prev->css);
1144 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1145 struct mem_cgroup *dead_memcg)
1147 struct mem_cgroup_reclaim_iter *iter;
1148 struct mem_cgroup_per_node *mz;
1151 for_each_node(nid) {
1152 mz = mem_cgroup_nodeinfo(from, nid);
1154 cmpxchg(&iter->position, dead_memcg, NULL);
1158 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1160 struct mem_cgroup *memcg = dead_memcg;
1161 struct mem_cgroup *last;
1164 __invalidate_reclaim_iterators(memcg, dead_memcg);
1166 } while ((memcg = parent_mem_cgroup(memcg)));
1169 * When cgruop1 non-hierarchy mode is used,
1170 * parent_mem_cgroup() does not walk all the way up to the
1171 * cgroup root (root_mem_cgroup). So we have to handle
1172 * dead_memcg from cgroup root separately.
1174 if (last != root_mem_cgroup)
1175 __invalidate_reclaim_iterators(root_mem_cgroup,
1180 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1181 * @memcg: hierarchy root
1182 * @fn: function to call for each task
1183 * @arg: argument passed to @fn
1185 * This function iterates over tasks attached to @memcg or to any of its
1186 * descendants and calls @fn for each task. If @fn returns a non-zero
1187 * value, the function breaks the iteration loop and returns the value.
1188 * Otherwise, it will iterate over all tasks and return 0.
1190 * This function must not be called for the root memory cgroup.
1192 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1193 int (*fn)(struct task_struct *, void *), void *arg)
1195 struct mem_cgroup *iter;
1198 BUG_ON(memcg == root_mem_cgroup);
1200 for_each_mem_cgroup_tree(iter, memcg) {
1201 struct css_task_iter it;
1202 struct task_struct *task;
1204 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1205 while (!ret && (task = css_task_iter_next(&it)))
1206 ret = fn(task, arg);
1207 css_task_iter_end(&it);
1209 mem_cgroup_iter_break(memcg, iter);
1217 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1219 * @pgdat: pgdat of the page
1221 * This function is only safe when following the LRU page isolation
1222 * and putback protocol: the LRU lock must be held, and the page must
1223 * either be PageLRU() or the caller must have isolated/allocated it.
1225 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1227 struct mem_cgroup_per_node *mz;
1228 struct mem_cgroup *memcg;
1229 struct lruvec *lruvec;
1231 if (mem_cgroup_disabled()) {
1232 lruvec = &pgdat->__lruvec;
1236 memcg = page->mem_cgroup;
1238 * Swapcache readahead pages are added to the LRU - and
1239 * possibly migrated - before they are charged.
1242 memcg = root_mem_cgroup;
1244 mz = mem_cgroup_page_nodeinfo(memcg, page);
1245 lruvec = &mz->lruvec;
1248 * Since a node can be onlined after the mem_cgroup was created,
1249 * we have to be prepared to initialize lruvec->zone here;
1250 * and if offlined then reonlined, we need to reinitialize it.
1252 if (unlikely(lruvec->pgdat != pgdat))
1253 lruvec->pgdat = pgdat;
1258 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1259 * @lruvec: mem_cgroup per zone lru vector
1260 * @lru: index of lru list the page is sitting on
1261 * @zid: zone id of the accounted pages
1262 * @nr_pages: positive when adding or negative when removing
1264 * This function must be called under lru_lock, just before a page is added
1265 * to or just after a page is removed from an lru list (that ordering being
1266 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1268 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1269 int zid, int nr_pages)
1271 struct mem_cgroup_per_node *mz;
1272 unsigned long *lru_size;
1275 if (mem_cgroup_disabled())
1278 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1279 lru_size = &mz->lru_zone_size[zid][lru];
1282 *lru_size += nr_pages;
1285 if (WARN_ONCE(size < 0,
1286 "%s(%p, %d, %d): lru_size %ld\n",
1287 __func__, lruvec, lru, nr_pages, size)) {
1293 *lru_size += nr_pages;
1297 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1298 * @memcg: the memory cgroup
1300 * Returns the maximum amount of memory @mem can be charged with, in
1303 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1305 unsigned long margin = 0;
1306 unsigned long count;
1307 unsigned long limit;
1309 count = page_counter_read(&memcg->memory);
1310 limit = READ_ONCE(memcg->memory.max);
1312 margin = limit - count;
1314 if (do_memsw_account()) {
1315 count = page_counter_read(&memcg->memsw);
1316 limit = READ_ONCE(memcg->memsw.max);
1318 margin = min(margin, limit - count);
1327 * A routine for checking "mem" is under move_account() or not.
1329 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1330 * moving cgroups. This is for waiting at high-memory pressure
1333 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1335 struct mem_cgroup *from;
1336 struct mem_cgroup *to;
1339 * Unlike task_move routines, we access mc.to, mc.from not under
1340 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1342 spin_lock(&mc.lock);
1348 ret = mem_cgroup_is_descendant(from, memcg) ||
1349 mem_cgroup_is_descendant(to, memcg);
1351 spin_unlock(&mc.lock);
1355 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1357 if (mc.moving_task && current != mc.moving_task) {
1358 if (mem_cgroup_under_move(memcg)) {
1360 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1361 /* moving charge context might have finished. */
1364 finish_wait(&mc.waitq, &wait);
1371 static char *memory_stat_format(struct mem_cgroup *memcg)
1376 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1381 * Provide statistics on the state of the memory subsystem as
1382 * well as cumulative event counters that show past behavior.
1384 * This list is ordered following a combination of these gradients:
1385 * 1) generic big picture -> specifics and details
1386 * 2) reflecting userspace activity -> reflecting kernel heuristics
1388 * Current memory state:
1391 seq_buf_printf(&s, "anon %llu\n",
1392 (u64)memcg_page_state(memcg, MEMCG_RSS) *
1394 seq_buf_printf(&s, "file %llu\n",
1395 (u64)memcg_page_state(memcg, MEMCG_CACHE) *
1397 seq_buf_printf(&s, "kernel_stack %llu\n",
1398 (u64)memcg_page_state(memcg, MEMCG_KERNEL_STACK_KB) *
1400 seq_buf_printf(&s, "slab %llu\n",
1401 (u64)(memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) +
1402 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE)) *
1404 seq_buf_printf(&s, "sock %llu\n",
1405 (u64)memcg_page_state(memcg, MEMCG_SOCK) *
1408 seq_buf_printf(&s, "shmem %llu\n",
1409 (u64)memcg_page_state(memcg, NR_SHMEM) *
1411 seq_buf_printf(&s, "file_mapped %llu\n",
1412 (u64)memcg_page_state(memcg, NR_FILE_MAPPED) *
1414 seq_buf_printf(&s, "file_dirty %llu\n",
1415 (u64)memcg_page_state(memcg, NR_FILE_DIRTY) *
1417 seq_buf_printf(&s, "file_writeback %llu\n",
1418 (u64)memcg_page_state(memcg, NR_WRITEBACK) *
1422 * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter
1423 * with the NR_ANON_THP vm counter, but right now it's a pain in the
1424 * arse because it requires migrating the work out of rmap to a place
1425 * where the page->mem_cgroup is set up and stable.
1427 seq_buf_printf(&s, "anon_thp %llu\n",
1428 (u64)memcg_page_state(memcg, MEMCG_RSS_HUGE) *
1431 for (i = 0; i < NR_LRU_LISTS; i++)
1432 seq_buf_printf(&s, "%s %llu\n", lru_list_name(i),
1433 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
1436 seq_buf_printf(&s, "slab_reclaimable %llu\n",
1437 (u64)memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) *
1439 seq_buf_printf(&s, "slab_unreclaimable %llu\n",
1440 (u64)memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE) *
1443 /* Accumulated memory events */
1445 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1446 memcg_events(memcg, PGFAULT));
1447 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1448 memcg_events(memcg, PGMAJFAULT));
1450 seq_buf_printf(&s, "workingset_refault %lu\n",
1451 memcg_page_state(memcg, WORKINGSET_REFAULT));
1452 seq_buf_printf(&s, "workingset_activate %lu\n",
1453 memcg_page_state(memcg, WORKINGSET_ACTIVATE));
1454 seq_buf_printf(&s, "workingset_restore %lu\n",
1455 memcg_page_state(memcg, WORKINGSET_RESTORE));
1456 seq_buf_printf(&s, "workingset_nodereclaim %lu\n",
1457 memcg_page_state(memcg, WORKINGSET_NODERECLAIM));
1459 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1460 memcg_events(memcg, PGREFILL));
1461 seq_buf_printf(&s, "pgscan %lu\n",
1462 memcg_events(memcg, PGSCAN_KSWAPD) +
1463 memcg_events(memcg, PGSCAN_DIRECT));
1464 seq_buf_printf(&s, "pgsteal %lu\n",
1465 memcg_events(memcg, PGSTEAL_KSWAPD) +
1466 memcg_events(memcg, PGSTEAL_DIRECT));
1467 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1468 memcg_events(memcg, PGACTIVATE));
1469 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1470 memcg_events(memcg, PGDEACTIVATE));
1471 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1472 memcg_events(memcg, PGLAZYFREE));
1473 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1474 memcg_events(memcg, PGLAZYFREED));
1476 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1477 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1478 memcg_events(memcg, THP_FAULT_ALLOC));
1479 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1480 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1481 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1483 /* The above should easily fit into one page */
1484 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1489 #define K(x) ((x) << (PAGE_SHIFT-10))
1491 * mem_cgroup_print_oom_context: Print OOM information relevant to
1492 * memory controller.
1493 * @memcg: The memory cgroup that went over limit
1494 * @p: Task that is going to be killed
1496 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1499 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1504 pr_cont(",oom_memcg=");
1505 pr_cont_cgroup_path(memcg->css.cgroup);
1507 pr_cont(",global_oom");
1509 pr_cont(",task_memcg=");
1510 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1516 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1517 * memory controller.
1518 * @memcg: The memory cgroup that went over limit
1520 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1524 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1525 K((u64)page_counter_read(&memcg->memory)),
1526 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1527 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1528 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1529 K((u64)page_counter_read(&memcg->swap)),
1530 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1532 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1533 K((u64)page_counter_read(&memcg->memsw)),
1534 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1535 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1536 K((u64)page_counter_read(&memcg->kmem)),
1537 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1540 pr_info("Memory cgroup stats for ");
1541 pr_cont_cgroup_path(memcg->css.cgroup);
1543 buf = memory_stat_format(memcg);
1551 * Return the memory (and swap, if configured) limit for a memcg.
1553 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1557 max = READ_ONCE(memcg->memory.max);
1558 if (mem_cgroup_swappiness(memcg)) {
1559 unsigned long memsw_max;
1560 unsigned long swap_max;
1562 memsw_max = memcg->memsw.max;
1563 swap_max = READ_ONCE(memcg->swap.max);
1564 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1565 max = min(max + swap_max, memsw_max);
1570 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1572 return page_counter_read(&memcg->memory);
1575 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1578 struct oom_control oc = {
1582 .gfp_mask = gfp_mask,
1587 if (mutex_lock_killable(&oom_lock))
1590 * A few threads which were not waiting at mutex_lock_killable() can
1591 * fail to bail out. Therefore, check again after holding oom_lock.
1593 ret = should_force_charge() || out_of_memory(&oc);
1594 mutex_unlock(&oom_lock);
1598 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1601 unsigned long *total_scanned)
1603 struct mem_cgroup *victim = NULL;
1606 unsigned long excess;
1607 unsigned long nr_scanned;
1608 struct mem_cgroup_reclaim_cookie reclaim = {
1612 excess = soft_limit_excess(root_memcg);
1615 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1620 * If we have not been able to reclaim
1621 * anything, it might because there are
1622 * no reclaimable pages under this hierarchy
1627 * We want to do more targeted reclaim.
1628 * excess >> 2 is not to excessive so as to
1629 * reclaim too much, nor too less that we keep
1630 * coming back to reclaim from this cgroup
1632 if (total >= (excess >> 2) ||
1633 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1638 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1639 pgdat, &nr_scanned);
1640 *total_scanned += nr_scanned;
1641 if (!soft_limit_excess(root_memcg))
1644 mem_cgroup_iter_break(root_memcg, victim);
1648 #ifdef CONFIG_LOCKDEP
1649 static struct lockdep_map memcg_oom_lock_dep_map = {
1650 .name = "memcg_oom_lock",
1654 static DEFINE_SPINLOCK(memcg_oom_lock);
1657 * Check OOM-Killer is already running under our hierarchy.
1658 * If someone is running, return false.
1660 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1662 struct mem_cgroup *iter, *failed = NULL;
1664 spin_lock(&memcg_oom_lock);
1666 for_each_mem_cgroup_tree(iter, memcg) {
1667 if (iter->oom_lock) {
1669 * this subtree of our hierarchy is already locked
1670 * so we cannot give a lock.
1673 mem_cgroup_iter_break(memcg, iter);
1676 iter->oom_lock = true;
1681 * OK, we failed to lock the whole subtree so we have
1682 * to clean up what we set up to the failing subtree
1684 for_each_mem_cgroup_tree(iter, memcg) {
1685 if (iter == failed) {
1686 mem_cgroup_iter_break(memcg, iter);
1689 iter->oom_lock = false;
1692 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1694 spin_unlock(&memcg_oom_lock);
1699 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1701 struct mem_cgroup *iter;
1703 spin_lock(&memcg_oom_lock);
1704 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1705 for_each_mem_cgroup_tree(iter, memcg)
1706 iter->oom_lock = false;
1707 spin_unlock(&memcg_oom_lock);
1710 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1712 struct mem_cgroup *iter;
1714 spin_lock(&memcg_oom_lock);
1715 for_each_mem_cgroup_tree(iter, memcg)
1717 spin_unlock(&memcg_oom_lock);
1720 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1722 struct mem_cgroup *iter;
1725 * When a new child is created while the hierarchy is under oom,
1726 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1728 spin_lock(&memcg_oom_lock);
1729 for_each_mem_cgroup_tree(iter, memcg)
1730 if (iter->under_oom > 0)
1732 spin_unlock(&memcg_oom_lock);
1735 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1737 struct oom_wait_info {
1738 struct mem_cgroup *memcg;
1739 wait_queue_entry_t wait;
1742 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1743 unsigned mode, int sync, void *arg)
1745 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1746 struct mem_cgroup *oom_wait_memcg;
1747 struct oom_wait_info *oom_wait_info;
1749 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1750 oom_wait_memcg = oom_wait_info->memcg;
1752 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1753 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1755 return autoremove_wake_function(wait, mode, sync, arg);
1758 static void memcg_oom_recover(struct mem_cgroup *memcg)
1761 * For the following lockless ->under_oom test, the only required
1762 * guarantee is that it must see the state asserted by an OOM when
1763 * this function is called as a result of userland actions
1764 * triggered by the notification of the OOM. This is trivially
1765 * achieved by invoking mem_cgroup_mark_under_oom() before
1766 * triggering notification.
1768 if (memcg && memcg->under_oom)
1769 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1779 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1781 enum oom_status ret;
1784 if (order > PAGE_ALLOC_COSTLY_ORDER)
1787 memcg_memory_event(memcg, MEMCG_OOM);
1790 * We are in the middle of the charge context here, so we
1791 * don't want to block when potentially sitting on a callstack
1792 * that holds all kinds of filesystem and mm locks.
1794 * cgroup1 allows disabling the OOM killer and waiting for outside
1795 * handling until the charge can succeed; remember the context and put
1796 * the task to sleep at the end of the page fault when all locks are
1799 * On the other hand, in-kernel OOM killer allows for an async victim
1800 * memory reclaim (oom_reaper) and that means that we are not solely
1801 * relying on the oom victim to make a forward progress and we can
1802 * invoke the oom killer here.
1804 * Please note that mem_cgroup_out_of_memory might fail to find a
1805 * victim and then we have to bail out from the charge path.
1807 if (memcg->oom_kill_disable) {
1808 if (!current->in_user_fault)
1810 css_get(&memcg->css);
1811 current->memcg_in_oom = memcg;
1812 current->memcg_oom_gfp_mask = mask;
1813 current->memcg_oom_order = order;
1818 mem_cgroup_mark_under_oom(memcg);
1820 locked = mem_cgroup_oom_trylock(memcg);
1823 mem_cgroup_oom_notify(memcg);
1825 mem_cgroup_unmark_under_oom(memcg);
1826 if (mem_cgroup_out_of_memory(memcg, mask, order))
1832 mem_cgroup_oom_unlock(memcg);
1838 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1839 * @handle: actually kill/wait or just clean up the OOM state
1841 * This has to be called at the end of a page fault if the memcg OOM
1842 * handler was enabled.
1844 * Memcg supports userspace OOM handling where failed allocations must
1845 * sleep on a waitqueue until the userspace task resolves the
1846 * situation. Sleeping directly in the charge context with all kinds
1847 * of locks held is not a good idea, instead we remember an OOM state
1848 * in the task and mem_cgroup_oom_synchronize() has to be called at
1849 * the end of the page fault to complete the OOM handling.
1851 * Returns %true if an ongoing memcg OOM situation was detected and
1852 * completed, %false otherwise.
1854 bool mem_cgroup_oom_synchronize(bool handle)
1856 struct mem_cgroup *memcg = current->memcg_in_oom;
1857 struct oom_wait_info owait;
1860 /* OOM is global, do not handle */
1867 owait.memcg = memcg;
1868 owait.wait.flags = 0;
1869 owait.wait.func = memcg_oom_wake_function;
1870 owait.wait.private = current;
1871 INIT_LIST_HEAD(&owait.wait.entry);
1873 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1874 mem_cgroup_mark_under_oom(memcg);
1876 locked = mem_cgroup_oom_trylock(memcg);
1879 mem_cgroup_oom_notify(memcg);
1881 if (locked && !memcg->oom_kill_disable) {
1882 mem_cgroup_unmark_under_oom(memcg);
1883 finish_wait(&memcg_oom_waitq, &owait.wait);
1884 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1885 current->memcg_oom_order);
1888 mem_cgroup_unmark_under_oom(memcg);
1889 finish_wait(&memcg_oom_waitq, &owait.wait);
1893 mem_cgroup_oom_unlock(memcg);
1895 * There is no guarantee that an OOM-lock contender
1896 * sees the wakeups triggered by the OOM kill
1897 * uncharges. Wake any sleepers explicitely.
1899 memcg_oom_recover(memcg);
1902 current->memcg_in_oom = NULL;
1903 css_put(&memcg->css);
1908 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1909 * @victim: task to be killed by the OOM killer
1910 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1912 * Returns a pointer to a memory cgroup, which has to be cleaned up
1913 * by killing all belonging OOM-killable tasks.
1915 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1917 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1918 struct mem_cgroup *oom_domain)
1920 struct mem_cgroup *oom_group = NULL;
1921 struct mem_cgroup *memcg;
1923 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
1927 oom_domain = root_mem_cgroup;
1931 memcg = mem_cgroup_from_task(victim);
1932 if (memcg == root_mem_cgroup)
1936 * If the victim task has been asynchronously moved to a different
1937 * memory cgroup, we might end up killing tasks outside oom_domain.
1938 * In this case it's better to ignore memory.group.oom.
1940 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
1944 * Traverse the memory cgroup hierarchy from the victim task's
1945 * cgroup up to the OOMing cgroup (or root) to find the
1946 * highest-level memory cgroup with oom.group set.
1948 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1949 if (memcg->oom_group)
1952 if (memcg == oom_domain)
1957 css_get(&oom_group->css);
1964 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
1966 pr_info("Tasks in ");
1967 pr_cont_cgroup_path(memcg->css.cgroup);
1968 pr_cont(" are going to be killed due to memory.oom.group set\n");
1972 * lock_page_memcg - lock a page->mem_cgroup binding
1975 * This function protects unlocked LRU pages from being moved to
1978 * It ensures lifetime of the returned memcg. Caller is responsible
1979 * for the lifetime of the page; __unlock_page_memcg() is available
1980 * when @page might get freed inside the locked section.
1982 struct mem_cgroup *lock_page_memcg(struct page *page)
1984 struct mem_cgroup *memcg;
1985 unsigned long flags;
1988 * The RCU lock is held throughout the transaction. The fast
1989 * path can get away without acquiring the memcg->move_lock
1990 * because page moving starts with an RCU grace period.
1992 * The RCU lock also protects the memcg from being freed when
1993 * the page state that is going to change is the only thing
1994 * preventing the page itself from being freed. E.g. writeback
1995 * doesn't hold a page reference and relies on PG_writeback to
1996 * keep off truncation, migration and so forth.
2000 if (mem_cgroup_disabled())
2003 memcg = page->mem_cgroup;
2004 if (unlikely(!memcg))
2007 if (atomic_read(&memcg->moving_account) <= 0)
2010 spin_lock_irqsave(&memcg->move_lock, flags);
2011 if (memcg != page->mem_cgroup) {
2012 spin_unlock_irqrestore(&memcg->move_lock, flags);
2017 * When charge migration first begins, we can have locked and
2018 * unlocked page stat updates happening concurrently. Track
2019 * the task who has the lock for unlock_page_memcg().
2021 memcg->move_lock_task = current;
2022 memcg->move_lock_flags = flags;
2026 EXPORT_SYMBOL(lock_page_memcg);
2029 * __unlock_page_memcg - unlock and unpin a memcg
2032 * Unlock and unpin a memcg returned by lock_page_memcg().
2034 void __unlock_page_memcg(struct mem_cgroup *memcg)
2036 if (memcg && memcg->move_lock_task == current) {
2037 unsigned long flags = memcg->move_lock_flags;
2039 memcg->move_lock_task = NULL;
2040 memcg->move_lock_flags = 0;
2042 spin_unlock_irqrestore(&memcg->move_lock, flags);
2049 * unlock_page_memcg - unlock a page->mem_cgroup binding
2052 void unlock_page_memcg(struct page *page)
2054 __unlock_page_memcg(page->mem_cgroup);
2056 EXPORT_SYMBOL(unlock_page_memcg);
2058 struct memcg_stock_pcp {
2059 struct mem_cgroup *cached; /* this never be root cgroup */
2060 unsigned int nr_pages;
2061 struct work_struct work;
2062 unsigned long flags;
2063 #define FLUSHING_CACHED_CHARGE 0
2065 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2066 static DEFINE_MUTEX(percpu_charge_mutex);
2069 * consume_stock: Try to consume stocked charge on this cpu.
2070 * @memcg: memcg to consume from.
2071 * @nr_pages: how many pages to charge.
2073 * The charges will only happen if @memcg matches the current cpu's memcg
2074 * stock, and at least @nr_pages are available in that stock. Failure to
2075 * service an allocation will refill the stock.
2077 * returns true if successful, false otherwise.
2079 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2081 struct memcg_stock_pcp *stock;
2082 unsigned long flags;
2085 if (nr_pages > MEMCG_CHARGE_BATCH)
2088 local_irq_save(flags);
2090 stock = this_cpu_ptr(&memcg_stock);
2091 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2092 stock->nr_pages -= nr_pages;
2096 local_irq_restore(flags);
2102 * Returns stocks cached in percpu and reset cached information.
2104 static void drain_stock(struct memcg_stock_pcp *stock)
2106 struct mem_cgroup *old = stock->cached;
2108 if (stock->nr_pages) {
2109 page_counter_uncharge(&old->memory, stock->nr_pages);
2110 if (do_memsw_account())
2111 page_counter_uncharge(&old->memsw, stock->nr_pages);
2112 css_put_many(&old->css, stock->nr_pages);
2113 stock->nr_pages = 0;
2115 stock->cached = NULL;
2118 static void drain_local_stock(struct work_struct *dummy)
2120 struct memcg_stock_pcp *stock;
2121 unsigned long flags;
2124 * The only protection from memory hotplug vs. drain_stock races is
2125 * that we always operate on local CPU stock here with IRQ disabled
2127 local_irq_save(flags);
2129 stock = this_cpu_ptr(&memcg_stock);
2131 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2133 local_irq_restore(flags);
2137 * Cache charges(val) to local per_cpu area.
2138 * This will be consumed by consume_stock() function, later.
2140 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2142 struct memcg_stock_pcp *stock;
2143 unsigned long flags;
2145 local_irq_save(flags);
2147 stock = this_cpu_ptr(&memcg_stock);
2148 if (stock->cached != memcg) { /* reset if necessary */
2150 stock->cached = memcg;
2152 stock->nr_pages += nr_pages;
2154 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2157 local_irq_restore(flags);
2161 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2162 * of the hierarchy under it.
2164 static void drain_all_stock(struct mem_cgroup *root_memcg)
2168 /* If someone's already draining, avoid adding running more workers. */
2169 if (!mutex_trylock(&percpu_charge_mutex))
2172 * Notify other cpus that system-wide "drain" is running
2173 * We do not care about races with the cpu hotplug because cpu down
2174 * as well as workers from this path always operate on the local
2175 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2178 for_each_online_cpu(cpu) {
2179 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2180 struct mem_cgroup *memcg;
2184 memcg = stock->cached;
2185 if (memcg && stock->nr_pages &&
2186 mem_cgroup_is_descendant(memcg, root_memcg))
2191 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2193 drain_local_stock(&stock->work);
2195 schedule_work_on(cpu, &stock->work);
2199 mutex_unlock(&percpu_charge_mutex);
2202 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2204 struct memcg_stock_pcp *stock;
2205 struct mem_cgroup *memcg, *mi;
2207 stock = &per_cpu(memcg_stock, cpu);
2210 for_each_mem_cgroup(memcg) {
2213 for (i = 0; i < MEMCG_NR_STAT; i++) {
2217 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2219 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2220 atomic_long_add(x, &memcg->vmstats[i]);
2222 if (i >= NR_VM_NODE_STAT_ITEMS)
2225 for_each_node(nid) {
2226 struct mem_cgroup_per_node *pn;
2228 pn = mem_cgroup_nodeinfo(memcg, nid);
2229 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2232 atomic_long_add(x, &pn->lruvec_stat[i]);
2233 } while ((pn = parent_nodeinfo(pn, nid)));
2237 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2240 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2242 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2243 atomic_long_add(x, &memcg->vmevents[i]);
2250 static void reclaim_high(struct mem_cgroup *memcg,
2251 unsigned int nr_pages,
2255 if (page_counter_read(&memcg->memory) <= READ_ONCE(memcg->high))
2257 memcg_memory_event(memcg, MEMCG_HIGH);
2258 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2259 } while ((memcg = parent_mem_cgroup(memcg)) &&
2260 !mem_cgroup_is_root(memcg));
2263 static void high_work_func(struct work_struct *work)
2265 struct mem_cgroup *memcg;
2267 memcg = container_of(work, struct mem_cgroup, high_work);
2268 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2272 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2273 * enough to still cause a significant slowdown in most cases, while still
2274 * allowing diagnostics and tracing to proceed without becoming stuck.
2276 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2279 * When calculating the delay, we use these either side of the exponentiation to
2280 * maintain precision and scale to a reasonable number of jiffies (see the table
2283 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2284 * overage ratio to a delay.
2285 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down down the
2286 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2287 * to produce a reasonable delay curve.
2289 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2290 * reasonable delay curve compared to precision-adjusted overage, not
2291 * penalising heavily at first, but still making sure that growth beyond the
2292 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2293 * example, with a high of 100 megabytes:
2295 * +-------+------------------------+
2296 * | usage | time to allocate in ms |
2297 * +-------+------------------------+
2319 * +-------+------------------------+
2321 #define MEMCG_DELAY_PRECISION_SHIFT 20
2322 #define MEMCG_DELAY_SCALING_SHIFT 14
2324 static u64 calculate_overage(unsigned long usage, unsigned long high)
2332 * Prevent division by 0 in overage calculation by acting as if
2333 * it was a threshold of 1 page
2335 high = max(high, 1UL);
2337 overage = usage - high;
2338 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2339 return div64_u64(overage, high);
2342 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2344 u64 overage, max_overage = 0;
2347 overage = calculate_overage(page_counter_read(&memcg->memory),
2348 READ_ONCE(memcg->high));
2349 max_overage = max(overage, max_overage);
2350 } while ((memcg = parent_mem_cgroup(memcg)) &&
2351 !mem_cgroup_is_root(memcg));
2357 * Get the number of jiffies that we should penalise a mischievous cgroup which
2358 * is exceeding its memory.high by checking both it and its ancestors.
2360 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2361 unsigned int nr_pages,
2364 unsigned long penalty_jiffies;
2370 * We use overage compared to memory.high to calculate the number of
2371 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2372 * fairly lenient on small overages, and increasingly harsh when the
2373 * memcg in question makes it clear that it has no intention of stopping
2374 * its crazy behaviour, so we exponentially increase the delay based on
2377 penalty_jiffies = max_overage * max_overage * HZ;
2378 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2379 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2382 * Factor in the task's own contribution to the overage, such that four
2383 * N-sized allocations are throttled approximately the same as one
2384 * 4N-sized allocation.
2386 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2387 * larger the current charge patch is than that.
2389 penalty_jiffies = penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2392 * Clamp the max delay per usermode return so as to still keep the
2393 * application moving forwards and also permit diagnostics, albeit
2396 return min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2400 * Scheduled by try_charge() to be executed from the userland return path
2401 * and reclaims memory over the high limit.
2403 void mem_cgroup_handle_over_high(void)
2405 unsigned long penalty_jiffies;
2406 unsigned long pflags;
2407 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2408 struct mem_cgroup *memcg;
2410 if (likely(!nr_pages))
2413 memcg = get_mem_cgroup_from_mm(current->mm);
2414 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2415 current->memcg_nr_pages_over_high = 0;
2418 * memory.high is breached and reclaim is unable to keep up. Throttle
2419 * allocators proactively to slow down excessive growth.
2421 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2422 mem_find_max_overage(memcg));
2425 * Don't sleep if the amount of jiffies this memcg owes us is so low
2426 * that it's not even worth doing, in an attempt to be nice to those who
2427 * go only a small amount over their memory.high value and maybe haven't
2428 * been aggressively reclaimed enough yet.
2430 if (penalty_jiffies <= HZ / 100)
2434 * If we exit early, we're guaranteed to die (since
2435 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2436 * need to account for any ill-begotten jiffies to pay them off later.
2438 psi_memstall_enter(&pflags);
2439 schedule_timeout_killable(penalty_jiffies);
2440 psi_memstall_leave(&pflags);
2443 css_put(&memcg->css);
2446 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2447 unsigned int nr_pages)
2449 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2450 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2451 struct mem_cgroup *mem_over_limit;
2452 struct page_counter *counter;
2453 unsigned long nr_reclaimed;
2454 bool may_swap = true;
2455 bool drained = false;
2456 enum oom_status oom_status;
2458 if (mem_cgroup_is_root(memcg))
2461 if (consume_stock(memcg, nr_pages))
2464 if (!do_memsw_account() ||
2465 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2466 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2468 if (do_memsw_account())
2469 page_counter_uncharge(&memcg->memsw, batch);
2470 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2472 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2476 if (batch > nr_pages) {
2482 * Memcg doesn't have a dedicated reserve for atomic
2483 * allocations. But like the global atomic pool, we need to
2484 * put the burden of reclaim on regular allocation requests
2485 * and let these go through as privileged allocations.
2487 if (gfp_mask & __GFP_ATOMIC)
2491 * Unlike in global OOM situations, memcg is not in a physical
2492 * memory shortage. Allow dying and OOM-killed tasks to
2493 * bypass the last charges so that they can exit quickly and
2494 * free their memory.
2496 if (unlikely(should_force_charge()))
2500 * Prevent unbounded recursion when reclaim operations need to
2501 * allocate memory. This might exceed the limits temporarily,
2502 * but we prefer facilitating memory reclaim and getting back
2503 * under the limit over triggering OOM kills in these cases.
2505 if (unlikely(current->flags & PF_MEMALLOC))
2508 if (unlikely(task_in_memcg_oom(current)))
2511 if (!gfpflags_allow_blocking(gfp_mask))
2514 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2516 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2517 gfp_mask, may_swap);
2519 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2523 drain_all_stock(mem_over_limit);
2528 if (gfp_mask & __GFP_NORETRY)
2531 * Even though the limit is exceeded at this point, reclaim
2532 * may have been able to free some pages. Retry the charge
2533 * before killing the task.
2535 * Only for regular pages, though: huge pages are rather
2536 * unlikely to succeed so close to the limit, and we fall back
2537 * to regular pages anyway in case of failure.
2539 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2542 * At task move, charge accounts can be doubly counted. So, it's
2543 * better to wait until the end of task_move if something is going on.
2545 if (mem_cgroup_wait_acct_move(mem_over_limit))
2551 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2554 if (gfp_mask & __GFP_NOFAIL)
2557 if (fatal_signal_pending(current))
2561 * keep retrying as long as the memcg oom killer is able to make
2562 * a forward progress or bypass the charge if the oom killer
2563 * couldn't make any progress.
2565 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2566 get_order(nr_pages * PAGE_SIZE));
2567 switch (oom_status) {
2569 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2577 if (!(gfp_mask & __GFP_NOFAIL))
2581 * The allocation either can't fail or will lead to more memory
2582 * being freed very soon. Allow memory usage go over the limit
2583 * temporarily by force charging it.
2585 page_counter_charge(&memcg->memory, nr_pages);
2586 if (do_memsw_account())
2587 page_counter_charge(&memcg->memsw, nr_pages);
2588 css_get_many(&memcg->css, nr_pages);
2593 css_get_many(&memcg->css, batch);
2594 if (batch > nr_pages)
2595 refill_stock(memcg, batch - nr_pages);
2598 * If the hierarchy is above the normal consumption range, schedule
2599 * reclaim on returning to userland. We can perform reclaim here
2600 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2601 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2602 * not recorded as it most likely matches current's and won't
2603 * change in the meantime. As high limit is checked again before
2604 * reclaim, the cost of mismatch is negligible.
2607 if (page_counter_read(&memcg->memory) > READ_ONCE(memcg->high)) {
2608 /* Don't bother a random interrupted task */
2609 if (in_interrupt()) {
2610 schedule_work(&memcg->high_work);
2613 current->memcg_nr_pages_over_high += batch;
2614 set_notify_resume(current);
2617 } while ((memcg = parent_mem_cgroup(memcg)));
2622 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2624 if (mem_cgroup_is_root(memcg))
2627 page_counter_uncharge(&memcg->memory, nr_pages);
2628 if (do_memsw_account())
2629 page_counter_uncharge(&memcg->memsw, nr_pages);
2631 css_put_many(&memcg->css, nr_pages);
2634 static void lock_page_lru(struct page *page, int *isolated)
2636 pg_data_t *pgdat = page_pgdat(page);
2638 spin_lock_irq(&pgdat->lru_lock);
2639 if (PageLRU(page)) {
2640 struct lruvec *lruvec;
2642 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2644 del_page_from_lru_list(page, lruvec, page_lru(page));
2650 static void unlock_page_lru(struct page *page, int isolated)
2652 pg_data_t *pgdat = page_pgdat(page);
2655 struct lruvec *lruvec;
2657 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2658 VM_BUG_ON_PAGE(PageLRU(page), page);
2660 add_page_to_lru_list(page, lruvec, page_lru(page));
2662 spin_unlock_irq(&pgdat->lru_lock);
2665 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2670 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2673 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2674 * may already be on some other mem_cgroup's LRU. Take care of it.
2677 lock_page_lru(page, &isolated);
2680 * Nobody should be changing or seriously looking at
2681 * page->mem_cgroup at this point:
2683 * - the page is uncharged
2685 * - the page is off-LRU
2687 * - an anonymous fault has exclusive page access, except for
2688 * a locked page table
2690 * - a page cache insertion, a swapin fault, or a migration
2691 * have the page locked
2693 page->mem_cgroup = memcg;
2696 unlock_page_lru(page, isolated);
2699 #ifdef CONFIG_MEMCG_KMEM
2701 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2703 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2704 * cgroup_mutex, etc.
2706 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2710 if (mem_cgroup_disabled())
2713 page = virt_to_head_page(p);
2716 * Slab pages don't have page->mem_cgroup set because corresponding
2717 * kmem caches can be reparented during the lifetime. That's why
2718 * memcg_from_slab_page() should be used instead.
2721 return memcg_from_slab_page(page);
2723 /* All other pages use page->mem_cgroup */
2724 return page->mem_cgroup;
2727 static int memcg_alloc_cache_id(void)
2732 id = ida_simple_get(&memcg_cache_ida,
2733 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2737 if (id < memcg_nr_cache_ids)
2741 * There's no space for the new id in memcg_caches arrays,
2742 * so we have to grow them.
2744 down_write(&memcg_cache_ids_sem);
2746 size = 2 * (id + 1);
2747 if (size < MEMCG_CACHES_MIN_SIZE)
2748 size = MEMCG_CACHES_MIN_SIZE;
2749 else if (size > MEMCG_CACHES_MAX_SIZE)
2750 size = MEMCG_CACHES_MAX_SIZE;
2752 err = memcg_update_all_caches(size);
2754 err = memcg_update_all_list_lrus(size);
2756 memcg_nr_cache_ids = size;
2758 up_write(&memcg_cache_ids_sem);
2761 ida_simple_remove(&memcg_cache_ida, id);
2767 static void memcg_free_cache_id(int id)
2769 ida_simple_remove(&memcg_cache_ida, id);
2772 struct memcg_kmem_cache_create_work {
2773 struct mem_cgroup *memcg;
2774 struct kmem_cache *cachep;
2775 struct work_struct work;
2778 static void memcg_kmem_cache_create_func(struct work_struct *w)
2780 struct memcg_kmem_cache_create_work *cw =
2781 container_of(w, struct memcg_kmem_cache_create_work, work);
2782 struct mem_cgroup *memcg = cw->memcg;
2783 struct kmem_cache *cachep = cw->cachep;
2785 memcg_create_kmem_cache(memcg, cachep);
2787 css_put(&memcg->css);
2792 * Enqueue the creation of a per-memcg kmem_cache.
2794 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2795 struct kmem_cache *cachep)
2797 struct memcg_kmem_cache_create_work *cw;
2799 if (!css_tryget_online(&memcg->css))
2802 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2807 cw->cachep = cachep;
2808 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2810 queue_work(memcg_kmem_cache_wq, &cw->work);
2813 static inline bool memcg_kmem_bypass(void)
2815 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2821 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2822 * @cachep: the original global kmem cache
2824 * Return the kmem_cache we're supposed to use for a slab allocation.
2825 * We try to use the current memcg's version of the cache.
2827 * If the cache does not exist yet, if we are the first user of it, we
2828 * create it asynchronously in a workqueue and let the current allocation
2829 * go through with the original cache.
2831 * This function takes a reference to the cache it returns to assure it
2832 * won't get destroyed while we are working with it. Once the caller is
2833 * done with it, memcg_kmem_put_cache() must be called to release the
2836 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2838 struct mem_cgroup *memcg;
2839 struct kmem_cache *memcg_cachep;
2840 struct memcg_cache_array *arr;
2843 VM_BUG_ON(!is_root_cache(cachep));
2845 if (memcg_kmem_bypass())
2850 if (unlikely(current->active_memcg))
2851 memcg = current->active_memcg;
2853 memcg = mem_cgroup_from_task(current);
2855 if (!memcg || memcg == root_mem_cgroup)
2858 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2862 arr = rcu_dereference(cachep->memcg_params.memcg_caches);
2865 * Make sure we will access the up-to-date value. The code updating
2866 * memcg_caches issues a write barrier to match the data dependency
2867 * barrier inside READ_ONCE() (see memcg_create_kmem_cache()).
2869 memcg_cachep = READ_ONCE(arr->entries[kmemcg_id]);
2872 * If we are in a safe context (can wait, and not in interrupt
2873 * context), we could be be predictable and return right away.
2874 * This would guarantee that the allocation being performed
2875 * already belongs in the new cache.
2877 * However, there are some clashes that can arrive from locking.
2878 * For instance, because we acquire the slab_mutex while doing
2879 * memcg_create_kmem_cache, this means no further allocation
2880 * could happen with the slab_mutex held. So it's better to
2883 * If the memcg is dying or memcg_cache is about to be released,
2884 * don't bother creating new kmem_caches. Because memcg_cachep
2885 * is ZEROed as the fist step of kmem offlining, we don't need
2886 * percpu_ref_tryget_live() here. css_tryget_online() check in
2887 * memcg_schedule_kmem_cache_create() will prevent us from
2888 * creation of a new kmem_cache.
2890 if (unlikely(!memcg_cachep))
2891 memcg_schedule_kmem_cache_create(memcg, cachep);
2892 else if (percpu_ref_tryget(&memcg_cachep->memcg_params.refcnt))
2893 cachep = memcg_cachep;
2900 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2901 * @cachep: the cache returned by memcg_kmem_get_cache
2903 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2905 if (!is_root_cache(cachep))
2906 percpu_ref_put(&cachep->memcg_params.refcnt);
2910 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
2911 * @memcg: memory cgroup to charge
2912 * @gfp: reclaim mode
2913 * @nr_pages: number of pages to charge
2915 * Returns 0 on success, an error code on failure.
2917 int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
2918 unsigned int nr_pages)
2920 struct page_counter *counter;
2923 ret = try_charge(memcg, gfp, nr_pages);
2927 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2928 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2931 * Enforce __GFP_NOFAIL allocation because callers are not
2932 * prepared to see failures and likely do not have any failure
2935 if (gfp & __GFP_NOFAIL) {
2936 page_counter_charge(&memcg->kmem, nr_pages);
2939 cancel_charge(memcg, nr_pages);
2946 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
2947 * @memcg: memcg to uncharge
2948 * @nr_pages: number of pages to uncharge
2950 void __memcg_kmem_uncharge(struct mem_cgroup *memcg, unsigned int nr_pages)
2952 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2953 page_counter_uncharge(&memcg->kmem, nr_pages);
2955 page_counter_uncharge(&memcg->memory, nr_pages);
2956 if (do_memsw_account())
2957 page_counter_uncharge(&memcg->memsw, nr_pages);
2961 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
2962 * @page: page to charge
2963 * @gfp: reclaim mode
2964 * @order: allocation order
2966 * Returns 0 on success, an error code on failure.
2968 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
2970 struct mem_cgroup *memcg;
2973 if (memcg_kmem_bypass())
2976 memcg = get_mem_cgroup_from_current();
2977 if (!mem_cgroup_is_root(memcg)) {
2978 ret = __memcg_kmem_charge(memcg, gfp, 1 << order);
2980 page->mem_cgroup = memcg;
2981 __SetPageKmemcg(page);
2984 css_put(&memcg->css);
2989 * __memcg_kmem_uncharge_page: uncharge a kmem page
2990 * @page: page to uncharge
2991 * @order: allocation order
2993 void __memcg_kmem_uncharge_page(struct page *page, int order)
2995 struct mem_cgroup *memcg = page->mem_cgroup;
2996 unsigned int nr_pages = 1 << order;
3001 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3002 __memcg_kmem_uncharge(memcg, nr_pages);
3003 page->mem_cgroup = NULL;
3005 /* slab pages do not have PageKmemcg flag set */
3006 if (PageKmemcg(page))
3007 __ClearPageKmemcg(page);
3009 css_put_many(&memcg->css, nr_pages);
3011 #endif /* CONFIG_MEMCG_KMEM */
3013 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3016 * Because tail pages are not marked as "used", set it. We're under
3017 * pgdat->lru_lock and migration entries setup in all page mappings.
3019 void mem_cgroup_split_huge_fixup(struct page *head)
3023 if (mem_cgroup_disabled())
3026 for (i = 1; i < HPAGE_PMD_NR; i++)
3027 head[i].mem_cgroup = head->mem_cgroup;
3029 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
3031 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3033 #ifdef CONFIG_MEMCG_SWAP
3035 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3036 * @entry: swap entry to be moved
3037 * @from: mem_cgroup which the entry is moved from
3038 * @to: mem_cgroup which the entry is moved to
3040 * It succeeds only when the swap_cgroup's record for this entry is the same
3041 * as the mem_cgroup's id of @from.
3043 * Returns 0 on success, -EINVAL on failure.
3045 * The caller must have charged to @to, IOW, called page_counter_charge() about
3046 * both res and memsw, and called css_get().
3048 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3049 struct mem_cgroup *from, struct mem_cgroup *to)
3051 unsigned short old_id, new_id;
3053 old_id = mem_cgroup_id(from);
3054 new_id = mem_cgroup_id(to);
3056 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3057 mod_memcg_state(from, MEMCG_SWAP, -1);
3058 mod_memcg_state(to, MEMCG_SWAP, 1);
3064 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3065 struct mem_cgroup *from, struct mem_cgroup *to)
3071 static DEFINE_MUTEX(memcg_max_mutex);
3073 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3074 unsigned long max, bool memsw)
3076 bool enlarge = false;
3077 bool drained = false;
3079 bool limits_invariant;
3080 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3083 if (signal_pending(current)) {
3088 mutex_lock(&memcg_max_mutex);
3090 * Make sure that the new limit (memsw or memory limit) doesn't
3091 * break our basic invariant rule memory.max <= memsw.max.
3093 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3094 max <= memcg->memsw.max;
3095 if (!limits_invariant) {
3096 mutex_unlock(&memcg_max_mutex);
3100 if (max > counter->max)
3102 ret = page_counter_set_max(counter, max);
3103 mutex_unlock(&memcg_max_mutex);
3109 drain_all_stock(memcg);
3114 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3115 GFP_KERNEL, !memsw)) {
3121 if (!ret && enlarge)
3122 memcg_oom_recover(memcg);
3127 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3129 unsigned long *total_scanned)
3131 unsigned long nr_reclaimed = 0;
3132 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3133 unsigned long reclaimed;
3135 struct mem_cgroup_tree_per_node *mctz;
3136 unsigned long excess;
3137 unsigned long nr_scanned;
3142 mctz = soft_limit_tree_node(pgdat->node_id);
3145 * Do not even bother to check the largest node if the root
3146 * is empty. Do it lockless to prevent lock bouncing. Races
3147 * are acceptable as soft limit is best effort anyway.
3149 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3153 * This loop can run a while, specially if mem_cgroup's continuously
3154 * keep exceeding their soft limit and putting the system under
3161 mz = mem_cgroup_largest_soft_limit_node(mctz);
3166 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3167 gfp_mask, &nr_scanned);
3168 nr_reclaimed += reclaimed;
3169 *total_scanned += nr_scanned;
3170 spin_lock_irq(&mctz->lock);
3171 __mem_cgroup_remove_exceeded(mz, mctz);
3174 * If we failed to reclaim anything from this memory cgroup
3175 * it is time to move on to the next cgroup
3179 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3181 excess = soft_limit_excess(mz->memcg);
3183 * One school of thought says that we should not add
3184 * back the node to the tree if reclaim returns 0.
3185 * But our reclaim could return 0, simply because due
3186 * to priority we are exposing a smaller subset of
3187 * memory to reclaim from. Consider this as a longer
3190 /* If excess == 0, no tree ops */
3191 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3192 spin_unlock_irq(&mctz->lock);
3193 css_put(&mz->memcg->css);
3196 * Could not reclaim anything and there are no more
3197 * mem cgroups to try or we seem to be looping without
3198 * reclaiming anything.
3200 if (!nr_reclaimed &&
3202 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3204 } while (!nr_reclaimed);
3206 css_put(&next_mz->memcg->css);
3207 return nr_reclaimed;
3211 * Test whether @memcg has children, dead or alive. Note that this
3212 * function doesn't care whether @memcg has use_hierarchy enabled and
3213 * returns %true if there are child csses according to the cgroup
3214 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3216 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3221 ret = css_next_child(NULL, &memcg->css);
3227 * Reclaims as many pages from the given memcg as possible.
3229 * Caller is responsible for holding css reference for memcg.
3231 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3233 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3235 /* we call try-to-free pages for make this cgroup empty */
3236 lru_add_drain_all();
3238 drain_all_stock(memcg);
3240 /* try to free all pages in this cgroup */
3241 while (nr_retries && page_counter_read(&memcg->memory)) {
3244 if (signal_pending(current))
3247 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3251 /* maybe some writeback is necessary */
3252 congestion_wait(BLK_RW_ASYNC, HZ/10);
3260 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3261 char *buf, size_t nbytes,
3264 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3266 if (mem_cgroup_is_root(memcg))
3268 return mem_cgroup_force_empty(memcg) ?: nbytes;
3271 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3274 return mem_cgroup_from_css(css)->use_hierarchy;
3277 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3278 struct cftype *cft, u64 val)
3281 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3282 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3284 if (memcg->use_hierarchy == val)
3288 * If parent's use_hierarchy is set, we can't make any modifications
3289 * in the child subtrees. If it is unset, then the change can
3290 * occur, provided the current cgroup has no children.
3292 * For the root cgroup, parent_mem is NULL, we allow value to be
3293 * set if there are no children.
3295 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3296 (val == 1 || val == 0)) {
3297 if (!memcg_has_children(memcg))
3298 memcg->use_hierarchy = val;
3307 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3311 if (mem_cgroup_is_root(memcg)) {
3312 val = memcg_page_state(memcg, MEMCG_CACHE) +
3313 memcg_page_state(memcg, MEMCG_RSS);
3315 val += memcg_page_state(memcg, MEMCG_SWAP);
3318 val = page_counter_read(&memcg->memory);
3320 val = page_counter_read(&memcg->memsw);
3333 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3336 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3337 struct page_counter *counter;
3339 switch (MEMFILE_TYPE(cft->private)) {
3341 counter = &memcg->memory;
3344 counter = &memcg->memsw;
3347 counter = &memcg->kmem;
3350 counter = &memcg->tcpmem;
3356 switch (MEMFILE_ATTR(cft->private)) {
3358 if (counter == &memcg->memory)
3359 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3360 if (counter == &memcg->memsw)
3361 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3362 return (u64)page_counter_read(counter) * PAGE_SIZE;
3364 return (u64)counter->max * PAGE_SIZE;
3366 return (u64)counter->watermark * PAGE_SIZE;
3368 return counter->failcnt;
3369 case RES_SOFT_LIMIT:
3370 return (u64)memcg->soft_limit * PAGE_SIZE;
3376 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3378 unsigned long stat[MEMCG_NR_STAT] = {0};
3379 struct mem_cgroup *mi;
3382 for_each_online_cpu(cpu)
3383 for (i = 0; i < MEMCG_NR_STAT; i++)
3384 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3386 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3387 for (i = 0; i < MEMCG_NR_STAT; i++)
3388 atomic_long_add(stat[i], &mi->vmstats[i]);
3390 for_each_node(node) {
3391 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3392 struct mem_cgroup_per_node *pi;
3394 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3397 for_each_online_cpu(cpu)
3398 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3400 pn->lruvec_stat_cpu->count[i], cpu);
3402 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3403 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3404 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3408 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3410 unsigned long events[NR_VM_EVENT_ITEMS];
3411 struct mem_cgroup *mi;
3414 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3417 for_each_online_cpu(cpu)
3418 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3419 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3422 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3423 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3424 atomic_long_add(events[i], &mi->vmevents[i]);
3427 #ifdef CONFIG_MEMCG_KMEM
3428 static int memcg_online_kmem(struct mem_cgroup *memcg)
3432 if (cgroup_memory_nokmem)
3435 BUG_ON(memcg->kmemcg_id >= 0);
3436 BUG_ON(memcg->kmem_state);
3438 memcg_id = memcg_alloc_cache_id();
3442 static_branch_inc(&memcg_kmem_enabled_key);
3444 * A memory cgroup is considered kmem-online as soon as it gets
3445 * kmemcg_id. Setting the id after enabling static branching will
3446 * guarantee no one starts accounting before all call sites are
3449 memcg->kmemcg_id = memcg_id;
3450 memcg->kmem_state = KMEM_ONLINE;
3451 INIT_LIST_HEAD(&memcg->kmem_caches);
3456 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3458 struct cgroup_subsys_state *css;
3459 struct mem_cgroup *parent, *child;
3462 if (memcg->kmem_state != KMEM_ONLINE)
3465 * Clear the online state before clearing memcg_caches array
3466 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3467 * guarantees that no cache will be created for this cgroup
3468 * after we are done (see memcg_create_kmem_cache()).
3470 memcg->kmem_state = KMEM_ALLOCATED;
3472 parent = parent_mem_cgroup(memcg);
3474 parent = root_mem_cgroup;
3477 * Deactivate and reparent kmem_caches.
3479 memcg_deactivate_kmem_caches(memcg, parent);
3481 kmemcg_id = memcg->kmemcg_id;
3482 BUG_ON(kmemcg_id < 0);
3485 * Change kmemcg_id of this cgroup and all its descendants to the
3486 * parent's id, and then move all entries from this cgroup's list_lrus
3487 * to ones of the parent. After we have finished, all list_lrus
3488 * corresponding to this cgroup are guaranteed to remain empty. The
3489 * ordering is imposed by list_lru_node->lock taken by
3490 * memcg_drain_all_list_lrus().
3492 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3493 css_for_each_descendant_pre(css, &memcg->css) {
3494 child = mem_cgroup_from_css(css);
3495 BUG_ON(child->kmemcg_id != kmemcg_id);
3496 child->kmemcg_id = parent->kmemcg_id;
3497 if (!memcg->use_hierarchy)
3502 memcg_drain_all_list_lrus(kmemcg_id, parent);
3504 memcg_free_cache_id(kmemcg_id);
3507 static void memcg_free_kmem(struct mem_cgroup *memcg)
3509 /* css_alloc() failed, offlining didn't happen */
3510 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3511 memcg_offline_kmem(memcg);
3513 if (memcg->kmem_state == KMEM_ALLOCATED) {
3514 WARN_ON(!list_empty(&memcg->kmem_caches));
3515 static_branch_dec(&memcg_kmem_enabled_key);
3519 static int memcg_online_kmem(struct mem_cgroup *memcg)
3523 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3526 static void memcg_free_kmem(struct mem_cgroup *memcg)
3529 #endif /* CONFIG_MEMCG_KMEM */
3531 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3536 mutex_lock(&memcg_max_mutex);
3537 ret = page_counter_set_max(&memcg->kmem, max);
3538 mutex_unlock(&memcg_max_mutex);
3542 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3546 mutex_lock(&memcg_max_mutex);
3548 ret = page_counter_set_max(&memcg->tcpmem, max);
3552 if (!memcg->tcpmem_active) {
3554 * The active flag needs to be written after the static_key
3555 * update. This is what guarantees that the socket activation
3556 * function is the last one to run. See mem_cgroup_sk_alloc()
3557 * for details, and note that we don't mark any socket as
3558 * belonging to this memcg until that flag is up.
3560 * We need to do this, because static_keys will span multiple
3561 * sites, but we can't control their order. If we mark a socket
3562 * as accounted, but the accounting functions are not patched in
3563 * yet, we'll lose accounting.
3565 * We never race with the readers in mem_cgroup_sk_alloc(),
3566 * because when this value change, the code to process it is not
3569 static_branch_inc(&memcg_sockets_enabled_key);
3570 memcg->tcpmem_active = true;
3573 mutex_unlock(&memcg_max_mutex);
3578 * The user of this function is...
3581 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3582 char *buf, size_t nbytes, loff_t off)
3584 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3585 unsigned long nr_pages;
3588 buf = strstrip(buf);
3589 ret = page_counter_memparse(buf, "-1", &nr_pages);
3593 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3595 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3599 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3601 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3604 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3607 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3608 "Please report your usecase to linux-mm@kvack.org if you "
3609 "depend on this functionality.\n");
3610 ret = memcg_update_kmem_max(memcg, nr_pages);
3613 ret = memcg_update_tcp_max(memcg, nr_pages);
3617 case RES_SOFT_LIMIT:
3618 memcg->soft_limit = nr_pages;
3622 return ret ?: nbytes;
3625 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3626 size_t nbytes, loff_t off)
3628 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3629 struct page_counter *counter;
3631 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3633 counter = &memcg->memory;
3636 counter = &memcg->memsw;
3639 counter = &memcg->kmem;
3642 counter = &memcg->tcpmem;
3648 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3650 page_counter_reset_watermark(counter);
3653 counter->failcnt = 0;
3662 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3665 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3669 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3670 struct cftype *cft, u64 val)
3672 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3674 if (val & ~MOVE_MASK)
3678 * No kind of locking is needed in here, because ->can_attach() will
3679 * check this value once in the beginning of the process, and then carry
3680 * on with stale data. This means that changes to this value will only
3681 * affect task migrations starting after the change.
3683 memcg->move_charge_at_immigrate = val;
3687 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3688 struct cftype *cft, u64 val)
3696 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3697 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3698 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3700 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3701 int nid, unsigned int lru_mask)
3703 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3704 unsigned long nr = 0;
3707 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3710 if (!(BIT(lru) & lru_mask))
3712 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3717 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3718 unsigned int lru_mask)
3720 unsigned long nr = 0;
3724 if (!(BIT(lru) & lru_mask))
3726 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3731 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3735 unsigned int lru_mask;
3738 static const struct numa_stat stats[] = {
3739 { "total", LRU_ALL },
3740 { "file", LRU_ALL_FILE },
3741 { "anon", LRU_ALL_ANON },
3742 { "unevictable", BIT(LRU_UNEVICTABLE) },
3744 const struct numa_stat *stat;
3747 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3749 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3750 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3751 seq_printf(m, "%s=%lu", stat->name, nr);
3752 for_each_node_state(nid, N_MEMORY) {
3753 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3755 seq_printf(m, " N%d=%lu", nid, nr);
3760 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3761 struct mem_cgroup *iter;
3764 for_each_mem_cgroup_tree(iter, memcg)
3765 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3766 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3767 for_each_node_state(nid, N_MEMORY) {
3769 for_each_mem_cgroup_tree(iter, memcg)
3770 nr += mem_cgroup_node_nr_lru_pages(
3771 iter, nid, stat->lru_mask);
3772 seq_printf(m, " N%d=%lu", nid, nr);
3779 #endif /* CONFIG_NUMA */
3781 static const unsigned int memcg1_stats[] = {
3792 static const char *const memcg1_stat_names[] = {
3803 /* Universal VM events cgroup1 shows, original sort order */
3804 static const unsigned int memcg1_events[] = {
3811 static int memcg_stat_show(struct seq_file *m, void *v)
3813 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3814 unsigned long memory, memsw;
3815 struct mem_cgroup *mi;
3818 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3820 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3821 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3823 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3824 memcg_page_state_local(memcg, memcg1_stats[i]) *
3828 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3829 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
3830 memcg_events_local(memcg, memcg1_events[i]));
3832 for (i = 0; i < NR_LRU_LISTS; i++)
3833 seq_printf(m, "%s %lu\n", lru_list_name(i),
3834 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3837 /* Hierarchical information */
3838 memory = memsw = PAGE_COUNTER_MAX;
3839 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3840 memory = min(memory, READ_ONCE(mi->memory.max));
3841 memsw = min(memsw, READ_ONCE(mi->memsw.max));
3843 seq_printf(m, "hierarchical_memory_limit %llu\n",
3844 (u64)memory * PAGE_SIZE);
3845 if (do_memsw_account())
3846 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3847 (u64)memsw * PAGE_SIZE);
3849 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3850 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3852 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3853 (u64)memcg_page_state(memcg, memcg1_stats[i]) *
3857 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3858 seq_printf(m, "total_%s %llu\n",
3859 vm_event_name(memcg1_events[i]),
3860 (u64)memcg_events(memcg, memcg1_events[i]));
3862 for (i = 0; i < NR_LRU_LISTS; i++)
3863 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
3864 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
3867 #ifdef CONFIG_DEBUG_VM
3870 struct mem_cgroup_per_node *mz;
3871 struct zone_reclaim_stat *rstat;
3872 unsigned long recent_rotated[2] = {0, 0};
3873 unsigned long recent_scanned[2] = {0, 0};
3875 for_each_online_pgdat(pgdat) {
3876 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3877 rstat = &mz->lruvec.reclaim_stat;
3879 recent_rotated[0] += rstat->recent_rotated[0];
3880 recent_rotated[1] += rstat->recent_rotated[1];
3881 recent_scanned[0] += rstat->recent_scanned[0];
3882 recent_scanned[1] += rstat->recent_scanned[1];
3884 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3885 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3886 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3887 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3894 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3897 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3899 return mem_cgroup_swappiness(memcg);
3902 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3903 struct cftype *cft, u64 val)
3905 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3911 memcg->swappiness = val;
3913 vm_swappiness = val;
3918 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3920 struct mem_cgroup_threshold_ary *t;
3921 unsigned long usage;
3926 t = rcu_dereference(memcg->thresholds.primary);
3928 t = rcu_dereference(memcg->memsw_thresholds.primary);
3933 usage = mem_cgroup_usage(memcg, swap);
3936 * current_threshold points to threshold just below or equal to usage.
3937 * If it's not true, a threshold was crossed after last
3938 * call of __mem_cgroup_threshold().
3940 i = t->current_threshold;
3943 * Iterate backward over array of thresholds starting from
3944 * current_threshold and check if a threshold is crossed.
3945 * If none of thresholds below usage is crossed, we read
3946 * only one element of the array here.
3948 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3949 eventfd_signal(t->entries[i].eventfd, 1);
3951 /* i = current_threshold + 1 */
3955 * Iterate forward over array of thresholds starting from
3956 * current_threshold+1 and check if a threshold is crossed.
3957 * If none of thresholds above usage is crossed, we read
3958 * only one element of the array here.
3960 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3961 eventfd_signal(t->entries[i].eventfd, 1);
3963 /* Update current_threshold */
3964 t->current_threshold = i - 1;
3969 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3972 __mem_cgroup_threshold(memcg, false);
3973 if (do_memsw_account())
3974 __mem_cgroup_threshold(memcg, true);
3976 memcg = parent_mem_cgroup(memcg);
3980 static int compare_thresholds(const void *a, const void *b)
3982 const struct mem_cgroup_threshold *_a = a;
3983 const struct mem_cgroup_threshold *_b = b;
3985 if (_a->threshold > _b->threshold)
3988 if (_a->threshold < _b->threshold)
3994 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3996 struct mem_cgroup_eventfd_list *ev;
3998 spin_lock(&memcg_oom_lock);
4000 list_for_each_entry(ev, &memcg->oom_notify, list)
4001 eventfd_signal(ev->eventfd, 1);
4003 spin_unlock(&memcg_oom_lock);
4007 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4009 struct mem_cgroup *iter;
4011 for_each_mem_cgroup_tree(iter, memcg)
4012 mem_cgroup_oom_notify_cb(iter);
4015 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4016 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4018 struct mem_cgroup_thresholds *thresholds;
4019 struct mem_cgroup_threshold_ary *new;
4020 unsigned long threshold;
4021 unsigned long usage;
4024 ret = page_counter_memparse(args, "-1", &threshold);
4028 mutex_lock(&memcg->thresholds_lock);
4031 thresholds = &memcg->thresholds;
4032 usage = mem_cgroup_usage(memcg, false);
4033 } else if (type == _MEMSWAP) {
4034 thresholds = &memcg->memsw_thresholds;
4035 usage = mem_cgroup_usage(memcg, true);
4039 /* Check if a threshold crossed before adding a new one */
4040 if (thresholds->primary)
4041 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4043 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4045 /* Allocate memory for new array of thresholds */
4046 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4053 /* Copy thresholds (if any) to new array */
4054 if (thresholds->primary) {
4055 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4056 sizeof(struct mem_cgroup_threshold));
4059 /* Add new threshold */
4060 new->entries[size - 1].eventfd = eventfd;
4061 new->entries[size - 1].threshold = threshold;
4063 /* Sort thresholds. Registering of new threshold isn't time-critical */
4064 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4065 compare_thresholds, NULL);
4067 /* Find current threshold */
4068 new->current_threshold = -1;
4069 for (i = 0; i < size; i++) {
4070 if (new->entries[i].threshold <= usage) {
4072 * new->current_threshold will not be used until
4073 * rcu_assign_pointer(), so it's safe to increment
4076 ++new->current_threshold;
4081 /* Free old spare buffer and save old primary buffer as spare */
4082 kfree(thresholds->spare);
4083 thresholds->spare = thresholds->primary;
4085 rcu_assign_pointer(thresholds->primary, new);
4087 /* To be sure that nobody uses thresholds */
4091 mutex_unlock(&memcg->thresholds_lock);
4096 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4097 struct eventfd_ctx *eventfd, const char *args)
4099 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4102 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4103 struct eventfd_ctx *eventfd, const char *args)
4105 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4108 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4109 struct eventfd_ctx *eventfd, enum res_type type)
4111 struct mem_cgroup_thresholds *thresholds;
4112 struct mem_cgroup_threshold_ary *new;
4113 unsigned long usage;
4114 int i, j, size, entries;
4116 mutex_lock(&memcg->thresholds_lock);
4119 thresholds = &memcg->thresholds;
4120 usage = mem_cgroup_usage(memcg, false);
4121 } else if (type == _MEMSWAP) {
4122 thresholds = &memcg->memsw_thresholds;
4123 usage = mem_cgroup_usage(memcg, true);
4127 if (!thresholds->primary)
4130 /* Check if a threshold crossed before removing */
4131 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4133 /* Calculate new number of threshold */
4135 for (i = 0; i < thresholds->primary->size; i++) {
4136 if (thresholds->primary->entries[i].eventfd != eventfd)
4142 new = thresholds->spare;
4144 /* If no items related to eventfd have been cleared, nothing to do */
4148 /* Set thresholds array to NULL if we don't have thresholds */
4157 /* Copy thresholds and find current threshold */
4158 new->current_threshold = -1;
4159 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4160 if (thresholds->primary->entries[i].eventfd == eventfd)
4163 new->entries[j] = thresholds->primary->entries[i];
4164 if (new->entries[j].threshold <= usage) {
4166 * new->current_threshold will not be used
4167 * until rcu_assign_pointer(), so it's safe to increment
4170 ++new->current_threshold;
4176 /* Swap primary and spare array */
4177 thresholds->spare = thresholds->primary;
4179 rcu_assign_pointer(thresholds->primary, new);
4181 /* To be sure that nobody uses thresholds */
4184 /* If all events are unregistered, free the spare array */
4186 kfree(thresholds->spare);
4187 thresholds->spare = NULL;
4190 mutex_unlock(&memcg->thresholds_lock);
4193 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4194 struct eventfd_ctx *eventfd)
4196 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4199 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4200 struct eventfd_ctx *eventfd)
4202 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4205 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4206 struct eventfd_ctx *eventfd, const char *args)
4208 struct mem_cgroup_eventfd_list *event;
4210 event = kmalloc(sizeof(*event), GFP_KERNEL);
4214 spin_lock(&memcg_oom_lock);
4216 event->eventfd = eventfd;
4217 list_add(&event->list, &memcg->oom_notify);
4219 /* already in OOM ? */
4220 if (memcg->under_oom)
4221 eventfd_signal(eventfd, 1);
4222 spin_unlock(&memcg_oom_lock);
4227 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4228 struct eventfd_ctx *eventfd)
4230 struct mem_cgroup_eventfd_list *ev, *tmp;
4232 spin_lock(&memcg_oom_lock);
4234 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4235 if (ev->eventfd == eventfd) {
4236 list_del(&ev->list);
4241 spin_unlock(&memcg_oom_lock);
4244 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4246 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4248 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4249 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4250 seq_printf(sf, "oom_kill %lu\n",
4251 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4255 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4256 struct cftype *cft, u64 val)
4258 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4260 /* cannot set to root cgroup and only 0 and 1 are allowed */
4261 if (!css->parent || !((val == 0) || (val == 1)))
4264 memcg->oom_kill_disable = val;
4266 memcg_oom_recover(memcg);
4271 #ifdef CONFIG_CGROUP_WRITEBACK
4273 #include <trace/events/writeback.h>
4275 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4277 return wb_domain_init(&memcg->cgwb_domain, gfp);
4280 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4282 wb_domain_exit(&memcg->cgwb_domain);
4285 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4287 wb_domain_size_changed(&memcg->cgwb_domain);
4290 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4292 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4294 if (!memcg->css.parent)
4297 return &memcg->cgwb_domain;
4301 * idx can be of type enum memcg_stat_item or node_stat_item.
4302 * Keep in sync with memcg_exact_page().
4304 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4306 long x = atomic_long_read(&memcg->vmstats[idx]);
4309 for_each_online_cpu(cpu)
4310 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4317 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4318 * @wb: bdi_writeback in question
4319 * @pfilepages: out parameter for number of file pages
4320 * @pheadroom: out parameter for number of allocatable pages according to memcg
4321 * @pdirty: out parameter for number of dirty pages
4322 * @pwriteback: out parameter for number of pages under writeback
4324 * Determine the numbers of file, headroom, dirty, and writeback pages in
4325 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4326 * is a bit more involved.
4328 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4329 * headroom is calculated as the lowest headroom of itself and the
4330 * ancestors. Note that this doesn't consider the actual amount of
4331 * available memory in the system. The caller should further cap
4332 * *@pheadroom accordingly.
4334 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4335 unsigned long *pheadroom, unsigned long *pdirty,
4336 unsigned long *pwriteback)
4338 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4339 struct mem_cgroup *parent;
4341 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4343 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4344 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4345 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4346 *pheadroom = PAGE_COUNTER_MAX;
4348 while ((parent = parent_mem_cgroup(memcg))) {
4349 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4350 READ_ONCE(memcg->high));
4351 unsigned long used = page_counter_read(&memcg->memory);
4353 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4359 * Foreign dirty flushing
4361 * There's an inherent mismatch between memcg and writeback. The former
4362 * trackes ownership per-page while the latter per-inode. This was a
4363 * deliberate design decision because honoring per-page ownership in the
4364 * writeback path is complicated, may lead to higher CPU and IO overheads
4365 * and deemed unnecessary given that write-sharing an inode across
4366 * different cgroups isn't a common use-case.
4368 * Combined with inode majority-writer ownership switching, this works well
4369 * enough in most cases but there are some pathological cases. For
4370 * example, let's say there are two cgroups A and B which keep writing to
4371 * different but confined parts of the same inode. B owns the inode and
4372 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4373 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4374 * triggering background writeback. A will be slowed down without a way to
4375 * make writeback of the dirty pages happen.
4377 * Conditions like the above can lead to a cgroup getting repatedly and
4378 * severely throttled after making some progress after each
4379 * dirty_expire_interval while the underyling IO device is almost
4382 * Solving this problem completely requires matching the ownership tracking
4383 * granularities between memcg and writeback in either direction. However,
4384 * the more egregious behaviors can be avoided by simply remembering the
4385 * most recent foreign dirtying events and initiating remote flushes on
4386 * them when local writeback isn't enough to keep the memory clean enough.
4388 * The following two functions implement such mechanism. When a foreign
4389 * page - a page whose memcg and writeback ownerships don't match - is
4390 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4391 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4392 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4393 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4394 * foreign bdi_writebacks which haven't expired. Both the numbers of
4395 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4396 * limited to MEMCG_CGWB_FRN_CNT.
4398 * The mechanism only remembers IDs and doesn't hold any object references.
4399 * As being wrong occasionally doesn't matter, updates and accesses to the
4400 * records are lockless and racy.
4402 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4403 struct bdi_writeback *wb)
4405 struct mem_cgroup *memcg = page->mem_cgroup;
4406 struct memcg_cgwb_frn *frn;
4407 u64 now = get_jiffies_64();
4408 u64 oldest_at = now;
4412 trace_track_foreign_dirty(page, wb);
4415 * Pick the slot to use. If there is already a slot for @wb, keep
4416 * using it. If not replace the oldest one which isn't being
4419 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4420 frn = &memcg->cgwb_frn[i];
4421 if (frn->bdi_id == wb->bdi->id &&
4422 frn->memcg_id == wb->memcg_css->id)
4424 if (time_before64(frn->at, oldest_at) &&
4425 atomic_read(&frn->done.cnt) == 1) {
4427 oldest_at = frn->at;
4431 if (i < MEMCG_CGWB_FRN_CNT) {
4433 * Re-using an existing one. Update timestamp lazily to
4434 * avoid making the cacheline hot. We want them to be
4435 * reasonably up-to-date and significantly shorter than
4436 * dirty_expire_interval as that's what expires the record.
4437 * Use the shorter of 1s and dirty_expire_interval / 8.
4439 unsigned long update_intv =
4440 min_t(unsigned long, HZ,
4441 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4443 if (time_before64(frn->at, now - update_intv))
4445 } else if (oldest >= 0) {
4446 /* replace the oldest free one */
4447 frn = &memcg->cgwb_frn[oldest];
4448 frn->bdi_id = wb->bdi->id;
4449 frn->memcg_id = wb->memcg_css->id;
4454 /* issue foreign writeback flushes for recorded foreign dirtying events */
4455 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4457 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4458 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4459 u64 now = jiffies_64;
4462 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4463 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4466 * If the record is older than dirty_expire_interval,
4467 * writeback on it has already started. No need to kick it
4468 * off again. Also, don't start a new one if there's
4469 * already one in flight.
4471 if (time_after64(frn->at, now - intv) &&
4472 atomic_read(&frn->done.cnt) == 1) {
4474 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4475 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4476 WB_REASON_FOREIGN_FLUSH,
4482 #else /* CONFIG_CGROUP_WRITEBACK */
4484 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4489 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4493 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4497 #endif /* CONFIG_CGROUP_WRITEBACK */
4500 * DO NOT USE IN NEW FILES.
4502 * "cgroup.event_control" implementation.
4504 * This is way over-engineered. It tries to support fully configurable
4505 * events for each user. Such level of flexibility is completely
4506 * unnecessary especially in the light of the planned unified hierarchy.
4508 * Please deprecate this and replace with something simpler if at all
4513 * Unregister event and free resources.
4515 * Gets called from workqueue.
4517 static void memcg_event_remove(struct work_struct *work)
4519 struct mem_cgroup_event *event =
4520 container_of(work, struct mem_cgroup_event, remove);
4521 struct mem_cgroup *memcg = event->memcg;
4523 remove_wait_queue(event->wqh, &event->wait);
4525 event->unregister_event(memcg, event->eventfd);
4527 /* Notify userspace the event is going away. */
4528 eventfd_signal(event->eventfd, 1);
4530 eventfd_ctx_put(event->eventfd);
4532 css_put(&memcg->css);
4536 * Gets called on EPOLLHUP on eventfd when user closes it.
4538 * Called with wqh->lock held and interrupts disabled.
4540 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4541 int sync, void *key)
4543 struct mem_cgroup_event *event =
4544 container_of(wait, struct mem_cgroup_event, wait);
4545 struct mem_cgroup *memcg = event->memcg;
4546 __poll_t flags = key_to_poll(key);
4548 if (flags & EPOLLHUP) {
4550 * If the event has been detached at cgroup removal, we
4551 * can simply return knowing the other side will cleanup
4554 * We can't race against event freeing since the other
4555 * side will require wqh->lock via remove_wait_queue(),
4558 spin_lock(&memcg->event_list_lock);
4559 if (!list_empty(&event->list)) {
4560 list_del_init(&event->list);
4562 * We are in atomic context, but cgroup_event_remove()
4563 * may sleep, so we have to call it in workqueue.
4565 schedule_work(&event->remove);
4567 spin_unlock(&memcg->event_list_lock);
4573 static void memcg_event_ptable_queue_proc(struct file *file,
4574 wait_queue_head_t *wqh, poll_table *pt)
4576 struct mem_cgroup_event *event =
4577 container_of(pt, struct mem_cgroup_event, pt);
4580 add_wait_queue(wqh, &event->wait);
4584 * DO NOT USE IN NEW FILES.
4586 * Parse input and register new cgroup event handler.
4588 * Input must be in format '<event_fd> <control_fd> <args>'.
4589 * Interpretation of args is defined by control file implementation.
4591 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4592 char *buf, size_t nbytes, loff_t off)
4594 struct cgroup_subsys_state *css = of_css(of);
4595 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4596 struct mem_cgroup_event *event;
4597 struct cgroup_subsys_state *cfile_css;
4598 unsigned int efd, cfd;
4605 buf = strstrip(buf);
4607 efd = simple_strtoul(buf, &endp, 10);
4612 cfd = simple_strtoul(buf, &endp, 10);
4613 if ((*endp != ' ') && (*endp != '\0'))
4617 event = kzalloc(sizeof(*event), GFP_KERNEL);
4621 event->memcg = memcg;
4622 INIT_LIST_HEAD(&event->list);
4623 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4624 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4625 INIT_WORK(&event->remove, memcg_event_remove);
4633 event->eventfd = eventfd_ctx_fileget(efile.file);
4634 if (IS_ERR(event->eventfd)) {
4635 ret = PTR_ERR(event->eventfd);
4642 goto out_put_eventfd;
4645 /* the process need read permission on control file */
4646 /* AV: shouldn't we check that it's been opened for read instead? */
4647 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4652 * Determine the event callbacks and set them in @event. This used
4653 * to be done via struct cftype but cgroup core no longer knows
4654 * about these events. The following is crude but the whole thing
4655 * is for compatibility anyway.
4657 * DO NOT ADD NEW FILES.
4659 name = cfile.file->f_path.dentry->d_name.name;
4661 if (!strcmp(name, "memory.usage_in_bytes")) {
4662 event->register_event = mem_cgroup_usage_register_event;
4663 event->unregister_event = mem_cgroup_usage_unregister_event;
4664 } else if (!strcmp(name, "memory.oom_control")) {
4665 event->register_event = mem_cgroup_oom_register_event;
4666 event->unregister_event = mem_cgroup_oom_unregister_event;
4667 } else if (!strcmp(name, "memory.pressure_level")) {
4668 event->register_event = vmpressure_register_event;
4669 event->unregister_event = vmpressure_unregister_event;
4670 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4671 event->register_event = memsw_cgroup_usage_register_event;
4672 event->unregister_event = memsw_cgroup_usage_unregister_event;
4679 * Verify @cfile should belong to @css. Also, remaining events are
4680 * automatically removed on cgroup destruction but the removal is
4681 * asynchronous, so take an extra ref on @css.
4683 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4684 &memory_cgrp_subsys);
4686 if (IS_ERR(cfile_css))
4688 if (cfile_css != css) {
4693 ret = event->register_event(memcg, event->eventfd, buf);
4697 vfs_poll(efile.file, &event->pt);
4699 spin_lock(&memcg->event_list_lock);
4700 list_add(&event->list, &memcg->event_list);
4701 spin_unlock(&memcg->event_list_lock);
4713 eventfd_ctx_put(event->eventfd);
4722 static struct cftype mem_cgroup_legacy_files[] = {
4724 .name = "usage_in_bytes",
4725 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4726 .read_u64 = mem_cgroup_read_u64,
4729 .name = "max_usage_in_bytes",
4730 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4731 .write = mem_cgroup_reset,
4732 .read_u64 = mem_cgroup_read_u64,
4735 .name = "limit_in_bytes",
4736 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4737 .write = mem_cgroup_write,
4738 .read_u64 = mem_cgroup_read_u64,
4741 .name = "soft_limit_in_bytes",
4742 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4743 .write = mem_cgroup_write,
4744 .read_u64 = mem_cgroup_read_u64,
4748 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4749 .write = mem_cgroup_reset,
4750 .read_u64 = mem_cgroup_read_u64,
4754 .seq_show = memcg_stat_show,
4757 .name = "force_empty",
4758 .write = mem_cgroup_force_empty_write,
4761 .name = "use_hierarchy",
4762 .write_u64 = mem_cgroup_hierarchy_write,
4763 .read_u64 = mem_cgroup_hierarchy_read,
4766 .name = "cgroup.event_control", /* XXX: for compat */
4767 .write = memcg_write_event_control,
4768 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4771 .name = "swappiness",
4772 .read_u64 = mem_cgroup_swappiness_read,
4773 .write_u64 = mem_cgroup_swappiness_write,
4776 .name = "move_charge_at_immigrate",
4777 .read_u64 = mem_cgroup_move_charge_read,
4778 .write_u64 = mem_cgroup_move_charge_write,
4781 .name = "oom_control",
4782 .seq_show = mem_cgroup_oom_control_read,
4783 .write_u64 = mem_cgroup_oom_control_write,
4784 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4787 .name = "pressure_level",
4791 .name = "numa_stat",
4792 .seq_show = memcg_numa_stat_show,
4796 .name = "kmem.limit_in_bytes",
4797 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4798 .write = mem_cgroup_write,
4799 .read_u64 = mem_cgroup_read_u64,
4802 .name = "kmem.usage_in_bytes",
4803 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4804 .read_u64 = mem_cgroup_read_u64,
4807 .name = "kmem.failcnt",
4808 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4809 .write = mem_cgroup_reset,
4810 .read_u64 = mem_cgroup_read_u64,
4813 .name = "kmem.max_usage_in_bytes",
4814 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4815 .write = mem_cgroup_reset,
4816 .read_u64 = mem_cgroup_read_u64,
4818 #if defined(CONFIG_MEMCG_KMEM) && \
4819 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
4821 .name = "kmem.slabinfo",
4822 .seq_start = memcg_slab_start,
4823 .seq_next = memcg_slab_next,
4824 .seq_stop = memcg_slab_stop,
4825 .seq_show = memcg_slab_show,
4829 .name = "kmem.tcp.limit_in_bytes",
4830 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4831 .write = mem_cgroup_write,
4832 .read_u64 = mem_cgroup_read_u64,
4835 .name = "kmem.tcp.usage_in_bytes",
4836 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4837 .read_u64 = mem_cgroup_read_u64,
4840 .name = "kmem.tcp.failcnt",
4841 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4842 .write = mem_cgroup_reset,
4843 .read_u64 = mem_cgroup_read_u64,
4846 .name = "kmem.tcp.max_usage_in_bytes",
4847 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4848 .write = mem_cgroup_reset,
4849 .read_u64 = mem_cgroup_read_u64,
4851 { }, /* terminate */
4855 * Private memory cgroup IDR
4857 * Swap-out records and page cache shadow entries need to store memcg
4858 * references in constrained space, so we maintain an ID space that is
4859 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4860 * memory-controlled cgroups to 64k.
4862 * However, there usually are many references to the oflline CSS after
4863 * the cgroup has been destroyed, such as page cache or reclaimable
4864 * slab objects, that don't need to hang on to the ID. We want to keep
4865 * those dead CSS from occupying IDs, or we might quickly exhaust the
4866 * relatively small ID space and prevent the creation of new cgroups
4867 * even when there are much fewer than 64k cgroups - possibly none.
4869 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4870 * be freed and recycled when it's no longer needed, which is usually
4871 * when the CSS is offlined.
4873 * The only exception to that are records of swapped out tmpfs/shmem
4874 * pages that need to be attributed to live ancestors on swapin. But
4875 * those references are manageable from userspace.
4878 static DEFINE_IDR(mem_cgroup_idr);
4880 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4882 if (memcg->id.id > 0) {
4883 idr_remove(&mem_cgroup_idr, memcg->id.id);
4888 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
4891 refcount_add(n, &memcg->id.ref);
4894 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4896 if (refcount_sub_and_test(n, &memcg->id.ref)) {
4897 mem_cgroup_id_remove(memcg);
4899 /* Memcg ID pins CSS */
4900 css_put(&memcg->css);
4904 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4906 mem_cgroup_id_put_many(memcg, 1);
4910 * mem_cgroup_from_id - look up a memcg from a memcg id
4911 * @id: the memcg id to look up
4913 * Caller must hold rcu_read_lock().
4915 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4917 WARN_ON_ONCE(!rcu_read_lock_held());
4918 return idr_find(&mem_cgroup_idr, id);
4921 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4923 struct mem_cgroup_per_node *pn;
4926 * This routine is called against possible nodes.
4927 * But it's BUG to call kmalloc() against offline node.
4929 * TODO: this routine can waste much memory for nodes which will
4930 * never be onlined. It's better to use memory hotplug callback
4933 if (!node_state(node, N_NORMAL_MEMORY))
4935 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4939 pn->lruvec_stat_local = alloc_percpu(struct lruvec_stat);
4940 if (!pn->lruvec_stat_local) {
4945 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
4946 if (!pn->lruvec_stat_cpu) {
4947 free_percpu(pn->lruvec_stat_local);
4952 lruvec_init(&pn->lruvec);
4953 pn->usage_in_excess = 0;
4954 pn->on_tree = false;
4957 memcg->nodeinfo[node] = pn;
4961 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4963 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
4968 free_percpu(pn->lruvec_stat_cpu);
4969 free_percpu(pn->lruvec_stat_local);
4973 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4978 free_mem_cgroup_per_node_info(memcg, node);
4979 free_percpu(memcg->vmstats_percpu);
4980 free_percpu(memcg->vmstats_local);
4984 static void mem_cgroup_free(struct mem_cgroup *memcg)
4986 memcg_wb_domain_exit(memcg);
4988 * Flush percpu vmstats and vmevents to guarantee the value correctness
4989 * on parent's and all ancestor levels.
4991 memcg_flush_percpu_vmstats(memcg);
4992 memcg_flush_percpu_vmevents(memcg);
4993 __mem_cgroup_free(memcg);
4996 static struct mem_cgroup *mem_cgroup_alloc(void)
4998 struct mem_cgroup *memcg;
5001 int __maybe_unused i;
5002 long error = -ENOMEM;
5004 size = sizeof(struct mem_cgroup);
5005 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5007 memcg = kzalloc(size, GFP_KERNEL);
5009 return ERR_PTR(error);
5011 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5012 1, MEM_CGROUP_ID_MAX,
5014 if (memcg->id.id < 0) {
5015 error = memcg->id.id;
5019 memcg->vmstats_local = alloc_percpu(struct memcg_vmstats_percpu);
5020 if (!memcg->vmstats_local)
5023 memcg->vmstats_percpu = alloc_percpu(struct memcg_vmstats_percpu);
5024 if (!memcg->vmstats_percpu)
5028 if (alloc_mem_cgroup_per_node_info(memcg, node))
5031 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5034 INIT_WORK(&memcg->high_work, high_work_func);
5035 INIT_LIST_HEAD(&memcg->oom_notify);
5036 mutex_init(&memcg->thresholds_lock);
5037 spin_lock_init(&memcg->move_lock);
5038 vmpressure_init(&memcg->vmpressure);
5039 INIT_LIST_HEAD(&memcg->event_list);
5040 spin_lock_init(&memcg->event_list_lock);
5041 memcg->socket_pressure = jiffies;
5042 #ifdef CONFIG_MEMCG_KMEM
5043 memcg->kmemcg_id = -1;
5045 #ifdef CONFIG_CGROUP_WRITEBACK
5046 INIT_LIST_HEAD(&memcg->cgwb_list);
5047 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5048 memcg->cgwb_frn[i].done =
5049 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5051 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5052 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5053 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5054 memcg->deferred_split_queue.split_queue_len = 0;
5056 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5059 mem_cgroup_id_remove(memcg);
5060 __mem_cgroup_free(memcg);
5061 return ERR_PTR(error);
5064 static struct cgroup_subsys_state * __ref
5065 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5067 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5068 struct mem_cgroup *memcg;
5069 long error = -ENOMEM;
5071 memcg = mem_cgroup_alloc();
5073 return ERR_CAST(memcg);
5075 WRITE_ONCE(memcg->high, PAGE_COUNTER_MAX);
5076 memcg->soft_limit = PAGE_COUNTER_MAX;
5078 memcg->swappiness = mem_cgroup_swappiness(parent);
5079 memcg->oom_kill_disable = parent->oom_kill_disable;
5081 if (parent && parent->use_hierarchy) {
5082 memcg->use_hierarchy = true;
5083 page_counter_init(&memcg->memory, &parent->memory);
5084 page_counter_init(&memcg->swap, &parent->swap);
5085 page_counter_init(&memcg->memsw, &parent->memsw);
5086 page_counter_init(&memcg->kmem, &parent->kmem);
5087 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5089 page_counter_init(&memcg->memory, NULL);
5090 page_counter_init(&memcg->swap, NULL);
5091 page_counter_init(&memcg->memsw, NULL);
5092 page_counter_init(&memcg->kmem, NULL);
5093 page_counter_init(&memcg->tcpmem, NULL);
5095 * Deeper hierachy with use_hierarchy == false doesn't make
5096 * much sense so let cgroup subsystem know about this
5097 * unfortunate state in our controller.
5099 if (parent != root_mem_cgroup)
5100 memory_cgrp_subsys.broken_hierarchy = true;
5103 /* The following stuff does not apply to the root */
5105 #ifdef CONFIG_MEMCG_KMEM
5106 INIT_LIST_HEAD(&memcg->kmem_caches);
5108 root_mem_cgroup = memcg;
5112 error = memcg_online_kmem(memcg);
5116 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5117 static_branch_inc(&memcg_sockets_enabled_key);
5121 mem_cgroup_id_remove(memcg);
5122 mem_cgroup_free(memcg);
5123 return ERR_PTR(error);
5126 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5128 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5131 * A memcg must be visible for memcg_expand_shrinker_maps()
5132 * by the time the maps are allocated. So, we allocate maps
5133 * here, when for_each_mem_cgroup() can't skip it.
5135 if (memcg_alloc_shrinker_maps(memcg)) {
5136 mem_cgroup_id_remove(memcg);
5140 /* Online state pins memcg ID, memcg ID pins CSS */
5141 refcount_set(&memcg->id.ref, 1);
5146 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5148 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5149 struct mem_cgroup_event *event, *tmp;
5152 * Unregister events and notify userspace.
5153 * Notify userspace about cgroup removing only after rmdir of cgroup
5154 * directory to avoid race between userspace and kernelspace.
5156 spin_lock(&memcg->event_list_lock);
5157 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5158 list_del_init(&event->list);
5159 schedule_work(&event->remove);
5161 spin_unlock(&memcg->event_list_lock);
5163 page_counter_set_min(&memcg->memory, 0);
5164 page_counter_set_low(&memcg->memory, 0);
5166 memcg_offline_kmem(memcg);
5167 wb_memcg_offline(memcg);
5169 drain_all_stock(memcg);
5171 mem_cgroup_id_put(memcg);
5174 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5176 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5178 invalidate_reclaim_iterators(memcg);
5181 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5183 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5184 int __maybe_unused i;
5186 #ifdef CONFIG_CGROUP_WRITEBACK
5187 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5188 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5190 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5191 static_branch_dec(&memcg_sockets_enabled_key);
5193 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5194 static_branch_dec(&memcg_sockets_enabled_key);
5196 vmpressure_cleanup(&memcg->vmpressure);
5197 cancel_work_sync(&memcg->high_work);
5198 mem_cgroup_remove_from_trees(memcg);
5199 memcg_free_shrinker_maps(memcg);
5200 memcg_free_kmem(memcg);
5201 mem_cgroup_free(memcg);
5205 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5206 * @css: the target css
5208 * Reset the states of the mem_cgroup associated with @css. This is
5209 * invoked when the userland requests disabling on the default hierarchy
5210 * but the memcg is pinned through dependency. The memcg should stop
5211 * applying policies and should revert to the vanilla state as it may be
5212 * made visible again.
5214 * The current implementation only resets the essential configurations.
5215 * This needs to be expanded to cover all the visible parts.
5217 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5219 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5221 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5222 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5223 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
5224 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5225 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5226 page_counter_set_min(&memcg->memory, 0);
5227 page_counter_set_low(&memcg->memory, 0);
5228 WRITE_ONCE(memcg->high, PAGE_COUNTER_MAX);
5229 memcg->soft_limit = PAGE_COUNTER_MAX;
5230 memcg_wb_domain_size_changed(memcg);
5234 /* Handlers for move charge at task migration. */
5235 static int mem_cgroup_do_precharge(unsigned long count)
5239 /* Try a single bulk charge without reclaim first, kswapd may wake */
5240 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5242 mc.precharge += count;
5246 /* Try charges one by one with reclaim, but do not retry */
5248 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5262 enum mc_target_type {
5269 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5270 unsigned long addr, pte_t ptent)
5272 struct page *page = vm_normal_page(vma, addr, ptent);
5274 if (!page || !page_mapped(page))
5276 if (PageAnon(page)) {
5277 if (!(mc.flags & MOVE_ANON))
5280 if (!(mc.flags & MOVE_FILE))
5283 if (!get_page_unless_zero(page))
5289 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5290 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5291 pte_t ptent, swp_entry_t *entry)
5293 struct page *page = NULL;
5294 swp_entry_t ent = pte_to_swp_entry(ptent);
5296 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
5300 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5301 * a device and because they are not accessible by CPU they are store
5302 * as special swap entry in the CPU page table.
5304 if (is_device_private_entry(ent)) {
5305 page = device_private_entry_to_page(ent);
5307 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5308 * a refcount of 1 when free (unlike normal page)
5310 if (!page_ref_add_unless(page, 1, 1))
5316 * Because lookup_swap_cache() updates some statistics counter,
5317 * we call find_get_page() with swapper_space directly.
5319 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5320 if (do_memsw_account())
5321 entry->val = ent.val;
5326 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5327 pte_t ptent, swp_entry_t *entry)
5333 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5334 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5336 struct page *page = NULL;
5337 struct address_space *mapping;
5340 if (!vma->vm_file) /* anonymous vma */
5342 if (!(mc.flags & MOVE_FILE))
5345 mapping = vma->vm_file->f_mapping;
5346 pgoff = linear_page_index(vma, addr);
5348 /* page is moved even if it's not RSS of this task(page-faulted). */
5350 /* shmem/tmpfs may report page out on swap: account for that too. */
5351 if (shmem_mapping(mapping)) {
5352 page = find_get_entry(mapping, pgoff);
5353 if (xa_is_value(page)) {
5354 swp_entry_t swp = radix_to_swp_entry(page);
5355 if (do_memsw_account())
5357 page = find_get_page(swap_address_space(swp),
5361 page = find_get_page(mapping, pgoff);
5363 page = find_get_page(mapping, pgoff);
5369 * mem_cgroup_move_account - move account of the page
5371 * @compound: charge the page as compound or small page
5372 * @from: mem_cgroup which the page is moved from.
5373 * @to: mem_cgroup which the page is moved to. @from != @to.
5375 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5377 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5380 static int mem_cgroup_move_account(struct page *page,
5382 struct mem_cgroup *from,
5383 struct mem_cgroup *to)
5385 struct lruvec *from_vec, *to_vec;
5386 struct pglist_data *pgdat;
5387 unsigned long flags;
5388 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5392 VM_BUG_ON(from == to);
5393 VM_BUG_ON_PAGE(PageLRU(page), page);
5394 VM_BUG_ON(compound && !PageTransHuge(page));
5397 * Prevent mem_cgroup_migrate() from looking at
5398 * page->mem_cgroup of its source page while we change it.
5401 if (!trylock_page(page))
5405 if (page->mem_cgroup != from)
5408 anon = PageAnon(page);
5410 pgdat = page_pgdat(page);
5411 from_vec = mem_cgroup_lruvec(from, pgdat);
5412 to_vec = mem_cgroup_lruvec(to, pgdat);
5414 spin_lock_irqsave(&from->move_lock, flags);
5416 if (!anon && page_mapped(page)) {
5417 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5418 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5422 * move_lock grabbed above and caller set from->moving_account, so
5423 * mod_memcg_page_state will serialize updates to PageDirty.
5424 * So mapping should be stable for dirty pages.
5426 if (!anon && PageDirty(page)) {
5427 struct address_space *mapping = page_mapping(page);
5429 if (mapping_cap_account_dirty(mapping)) {
5430 __mod_lruvec_state(from_vec, NR_FILE_DIRTY, -nr_pages);
5431 __mod_lruvec_state(to_vec, NR_FILE_DIRTY, nr_pages);
5435 if (PageWriteback(page)) {
5436 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5437 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5441 * It is safe to change page->mem_cgroup here because the page
5442 * is referenced, charged, and isolated - we can't race with
5443 * uncharging, charging, migration, or LRU putback.
5446 /* caller should have done css_get */
5447 page->mem_cgroup = to;
5449 spin_unlock_irqrestore(&from->move_lock, flags);
5453 local_irq_disable();
5454 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
5455 memcg_check_events(to, page);
5456 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
5457 memcg_check_events(from, page);
5466 * get_mctgt_type - get target type of moving charge
5467 * @vma: the vma the pte to be checked belongs
5468 * @addr: the address corresponding to the pte to be checked
5469 * @ptent: the pte to be checked
5470 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5473 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5474 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5475 * move charge. if @target is not NULL, the page is stored in target->page
5476 * with extra refcnt got(Callers should handle it).
5477 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5478 * target for charge migration. if @target is not NULL, the entry is stored
5480 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5481 * (so ZONE_DEVICE page and thus not on the lru).
5482 * For now we such page is charge like a regular page would be as for all
5483 * intent and purposes it is just special memory taking the place of a
5486 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5488 * Called with pte lock held.
5491 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5492 unsigned long addr, pte_t ptent, union mc_target *target)
5494 struct page *page = NULL;
5495 enum mc_target_type ret = MC_TARGET_NONE;
5496 swp_entry_t ent = { .val = 0 };
5498 if (pte_present(ptent))
5499 page = mc_handle_present_pte(vma, addr, ptent);
5500 else if (is_swap_pte(ptent))
5501 page = mc_handle_swap_pte(vma, ptent, &ent);
5502 else if (pte_none(ptent))
5503 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5505 if (!page && !ent.val)
5509 * Do only loose check w/o serialization.
5510 * mem_cgroup_move_account() checks the page is valid or
5511 * not under LRU exclusion.
5513 if (page->mem_cgroup == mc.from) {
5514 ret = MC_TARGET_PAGE;
5515 if (is_device_private_page(page))
5516 ret = MC_TARGET_DEVICE;
5518 target->page = page;
5520 if (!ret || !target)
5524 * There is a swap entry and a page doesn't exist or isn't charged.
5525 * But we cannot move a tail-page in a THP.
5527 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5528 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5529 ret = MC_TARGET_SWAP;
5536 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5538 * We don't consider PMD mapped swapping or file mapped pages because THP does
5539 * not support them for now.
5540 * Caller should make sure that pmd_trans_huge(pmd) is true.
5542 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5543 unsigned long addr, pmd_t pmd, union mc_target *target)
5545 struct page *page = NULL;
5546 enum mc_target_type ret = MC_TARGET_NONE;
5548 if (unlikely(is_swap_pmd(pmd))) {
5549 VM_BUG_ON(thp_migration_supported() &&
5550 !is_pmd_migration_entry(pmd));
5553 page = pmd_page(pmd);
5554 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5555 if (!(mc.flags & MOVE_ANON))
5557 if (page->mem_cgroup == mc.from) {
5558 ret = MC_TARGET_PAGE;
5561 target->page = page;
5567 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5568 unsigned long addr, pmd_t pmd, union mc_target *target)
5570 return MC_TARGET_NONE;
5574 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5575 unsigned long addr, unsigned long end,
5576 struct mm_walk *walk)
5578 struct vm_area_struct *vma = walk->vma;
5582 ptl = pmd_trans_huge_lock(pmd, vma);
5585 * Note their can not be MC_TARGET_DEVICE for now as we do not
5586 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5587 * this might change.
5589 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5590 mc.precharge += HPAGE_PMD_NR;
5595 if (pmd_trans_unstable(pmd))
5597 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5598 for (; addr != end; pte++, addr += PAGE_SIZE)
5599 if (get_mctgt_type(vma, addr, *pte, NULL))
5600 mc.precharge++; /* increment precharge temporarily */
5601 pte_unmap_unlock(pte - 1, ptl);
5607 static const struct mm_walk_ops precharge_walk_ops = {
5608 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5611 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5613 unsigned long precharge;
5615 down_read(&mm->mmap_sem);
5616 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5617 up_read(&mm->mmap_sem);
5619 precharge = mc.precharge;
5625 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5627 unsigned long precharge = mem_cgroup_count_precharge(mm);
5629 VM_BUG_ON(mc.moving_task);
5630 mc.moving_task = current;
5631 return mem_cgroup_do_precharge(precharge);
5634 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5635 static void __mem_cgroup_clear_mc(void)
5637 struct mem_cgroup *from = mc.from;
5638 struct mem_cgroup *to = mc.to;
5640 /* we must uncharge all the leftover precharges from mc.to */
5642 cancel_charge(mc.to, mc.precharge);
5646 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5647 * we must uncharge here.
5649 if (mc.moved_charge) {
5650 cancel_charge(mc.from, mc.moved_charge);
5651 mc.moved_charge = 0;
5653 /* we must fixup refcnts and charges */
5654 if (mc.moved_swap) {
5655 /* uncharge swap account from the old cgroup */
5656 if (!mem_cgroup_is_root(mc.from))
5657 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5659 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5662 * we charged both to->memory and to->memsw, so we
5663 * should uncharge to->memory.
5665 if (!mem_cgroup_is_root(mc.to))
5666 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5668 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
5669 css_put_many(&mc.to->css, mc.moved_swap);
5673 memcg_oom_recover(from);
5674 memcg_oom_recover(to);
5675 wake_up_all(&mc.waitq);
5678 static void mem_cgroup_clear_mc(void)
5680 struct mm_struct *mm = mc.mm;
5683 * we must clear moving_task before waking up waiters at the end of
5686 mc.moving_task = NULL;
5687 __mem_cgroup_clear_mc();
5688 spin_lock(&mc.lock);
5692 spin_unlock(&mc.lock);
5697 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5699 struct cgroup_subsys_state *css;
5700 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5701 struct mem_cgroup *from;
5702 struct task_struct *leader, *p;
5703 struct mm_struct *mm;
5704 unsigned long move_flags;
5707 /* charge immigration isn't supported on the default hierarchy */
5708 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5712 * Multi-process migrations only happen on the default hierarchy
5713 * where charge immigration is not used. Perform charge
5714 * immigration if @tset contains a leader and whine if there are
5718 cgroup_taskset_for_each_leader(leader, css, tset) {
5721 memcg = mem_cgroup_from_css(css);
5727 * We are now commited to this value whatever it is. Changes in this
5728 * tunable will only affect upcoming migrations, not the current one.
5729 * So we need to save it, and keep it going.
5731 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5735 from = mem_cgroup_from_task(p);
5737 VM_BUG_ON(from == memcg);
5739 mm = get_task_mm(p);
5742 /* We move charges only when we move a owner of the mm */
5743 if (mm->owner == p) {
5746 VM_BUG_ON(mc.precharge);
5747 VM_BUG_ON(mc.moved_charge);
5748 VM_BUG_ON(mc.moved_swap);
5750 spin_lock(&mc.lock);
5754 mc.flags = move_flags;
5755 spin_unlock(&mc.lock);
5756 /* We set mc.moving_task later */
5758 ret = mem_cgroup_precharge_mc(mm);
5760 mem_cgroup_clear_mc();
5767 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5770 mem_cgroup_clear_mc();
5773 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5774 unsigned long addr, unsigned long end,
5775 struct mm_walk *walk)
5778 struct vm_area_struct *vma = walk->vma;
5781 enum mc_target_type target_type;
5782 union mc_target target;
5785 ptl = pmd_trans_huge_lock(pmd, vma);
5787 if (mc.precharge < HPAGE_PMD_NR) {
5791 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5792 if (target_type == MC_TARGET_PAGE) {
5794 if (!isolate_lru_page(page)) {
5795 if (!mem_cgroup_move_account(page, true,
5797 mc.precharge -= HPAGE_PMD_NR;
5798 mc.moved_charge += HPAGE_PMD_NR;
5800 putback_lru_page(page);
5803 } else if (target_type == MC_TARGET_DEVICE) {
5805 if (!mem_cgroup_move_account(page, true,
5807 mc.precharge -= HPAGE_PMD_NR;
5808 mc.moved_charge += HPAGE_PMD_NR;
5816 if (pmd_trans_unstable(pmd))
5819 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5820 for (; addr != end; addr += PAGE_SIZE) {
5821 pte_t ptent = *(pte++);
5822 bool device = false;
5828 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5829 case MC_TARGET_DEVICE:
5832 case MC_TARGET_PAGE:
5835 * We can have a part of the split pmd here. Moving it
5836 * can be done but it would be too convoluted so simply
5837 * ignore such a partial THP and keep it in original
5838 * memcg. There should be somebody mapping the head.
5840 if (PageTransCompound(page))
5842 if (!device && isolate_lru_page(page))
5844 if (!mem_cgroup_move_account(page, false,
5847 /* we uncharge from mc.from later. */
5851 putback_lru_page(page);
5852 put: /* get_mctgt_type() gets the page */
5855 case MC_TARGET_SWAP:
5857 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5859 /* we fixup refcnts and charges later. */
5867 pte_unmap_unlock(pte - 1, ptl);
5872 * We have consumed all precharges we got in can_attach().
5873 * We try charge one by one, but don't do any additional
5874 * charges to mc.to if we have failed in charge once in attach()
5877 ret = mem_cgroup_do_precharge(1);
5885 static const struct mm_walk_ops charge_walk_ops = {
5886 .pmd_entry = mem_cgroup_move_charge_pte_range,
5889 static void mem_cgroup_move_charge(void)
5891 lru_add_drain_all();
5893 * Signal lock_page_memcg() to take the memcg's move_lock
5894 * while we're moving its pages to another memcg. Then wait
5895 * for already started RCU-only updates to finish.
5897 atomic_inc(&mc.from->moving_account);
5900 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5902 * Someone who are holding the mmap_sem might be waiting in
5903 * waitq. So we cancel all extra charges, wake up all waiters,
5904 * and retry. Because we cancel precharges, we might not be able
5905 * to move enough charges, but moving charge is a best-effort
5906 * feature anyway, so it wouldn't be a big problem.
5908 __mem_cgroup_clear_mc();
5913 * When we have consumed all precharges and failed in doing
5914 * additional charge, the page walk just aborts.
5916 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
5919 up_read(&mc.mm->mmap_sem);
5920 atomic_dec(&mc.from->moving_account);
5923 static void mem_cgroup_move_task(void)
5926 mem_cgroup_move_charge();
5927 mem_cgroup_clear_mc();
5930 #else /* !CONFIG_MMU */
5931 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5935 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5938 static void mem_cgroup_move_task(void)
5944 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5945 * to verify whether we're attached to the default hierarchy on each mount
5948 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5951 * use_hierarchy is forced on the default hierarchy. cgroup core
5952 * guarantees that @root doesn't have any children, so turning it
5953 * on for the root memcg is enough.
5955 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5956 root_mem_cgroup->use_hierarchy = true;
5958 root_mem_cgroup->use_hierarchy = false;
5961 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
5963 if (value == PAGE_COUNTER_MAX)
5964 seq_puts(m, "max\n");
5966 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
5971 static u64 memory_current_read(struct cgroup_subsys_state *css,
5974 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5976 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5979 static int memory_min_show(struct seq_file *m, void *v)
5981 return seq_puts_memcg_tunable(m,
5982 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
5985 static ssize_t memory_min_write(struct kernfs_open_file *of,
5986 char *buf, size_t nbytes, loff_t off)
5988 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5992 buf = strstrip(buf);
5993 err = page_counter_memparse(buf, "max", &min);
5997 page_counter_set_min(&memcg->memory, min);
6002 static int memory_low_show(struct seq_file *m, void *v)
6004 return seq_puts_memcg_tunable(m,
6005 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6008 static ssize_t memory_low_write(struct kernfs_open_file *of,
6009 char *buf, size_t nbytes, loff_t off)
6011 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6015 buf = strstrip(buf);
6016 err = page_counter_memparse(buf, "max", &low);
6020 page_counter_set_low(&memcg->memory, low);
6025 static int memory_high_show(struct seq_file *m, void *v)
6027 return seq_puts_memcg_tunable(m, READ_ONCE(mem_cgroup_from_seq(m)->high));
6030 static ssize_t memory_high_write(struct kernfs_open_file *of,
6031 char *buf, size_t nbytes, loff_t off)
6033 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6034 unsigned int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
6035 bool drained = false;
6039 buf = strstrip(buf);
6040 err = page_counter_memparse(buf, "max", &high);
6044 WRITE_ONCE(memcg->high, high);
6047 unsigned long nr_pages = page_counter_read(&memcg->memory);
6048 unsigned long reclaimed;
6050 if (nr_pages <= high)
6053 if (signal_pending(current))
6057 drain_all_stock(memcg);
6062 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6065 if (!reclaimed && !nr_retries--)
6072 static int memory_max_show(struct seq_file *m, void *v)
6074 return seq_puts_memcg_tunable(m,
6075 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6078 static ssize_t memory_max_write(struct kernfs_open_file *of,
6079 char *buf, size_t nbytes, loff_t off)
6081 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6082 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
6083 bool drained = false;
6087 buf = strstrip(buf);
6088 err = page_counter_memparse(buf, "max", &max);
6092 xchg(&memcg->memory.max, max);
6095 unsigned long nr_pages = page_counter_read(&memcg->memory);
6097 if (nr_pages <= max)
6100 if (signal_pending(current))
6104 drain_all_stock(memcg);
6110 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6116 memcg_memory_event(memcg, MEMCG_OOM);
6117 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6121 memcg_wb_domain_size_changed(memcg);
6125 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6127 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6128 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6129 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6130 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6131 seq_printf(m, "oom_kill %lu\n",
6132 atomic_long_read(&events[MEMCG_OOM_KILL]));
6135 static int memory_events_show(struct seq_file *m, void *v)
6137 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6139 __memory_events_show(m, memcg->memory_events);
6143 static int memory_events_local_show(struct seq_file *m, void *v)
6145 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6147 __memory_events_show(m, memcg->memory_events_local);
6151 static int memory_stat_show(struct seq_file *m, void *v)
6153 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6156 buf = memory_stat_format(memcg);
6164 static int memory_oom_group_show(struct seq_file *m, void *v)
6166 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6168 seq_printf(m, "%d\n", memcg->oom_group);
6173 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6174 char *buf, size_t nbytes, loff_t off)
6176 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6179 buf = strstrip(buf);
6183 ret = kstrtoint(buf, 0, &oom_group);
6187 if (oom_group != 0 && oom_group != 1)
6190 memcg->oom_group = oom_group;
6195 static struct cftype memory_files[] = {
6198 .flags = CFTYPE_NOT_ON_ROOT,
6199 .read_u64 = memory_current_read,
6203 .flags = CFTYPE_NOT_ON_ROOT,
6204 .seq_show = memory_min_show,
6205 .write = memory_min_write,
6209 .flags = CFTYPE_NOT_ON_ROOT,
6210 .seq_show = memory_low_show,
6211 .write = memory_low_write,
6215 .flags = CFTYPE_NOT_ON_ROOT,
6216 .seq_show = memory_high_show,
6217 .write = memory_high_write,
6221 .flags = CFTYPE_NOT_ON_ROOT,
6222 .seq_show = memory_max_show,
6223 .write = memory_max_write,
6227 .flags = CFTYPE_NOT_ON_ROOT,
6228 .file_offset = offsetof(struct mem_cgroup, events_file),
6229 .seq_show = memory_events_show,
6232 .name = "events.local",
6233 .flags = CFTYPE_NOT_ON_ROOT,
6234 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6235 .seq_show = memory_events_local_show,
6239 .seq_show = memory_stat_show,
6242 .name = "oom.group",
6243 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6244 .seq_show = memory_oom_group_show,
6245 .write = memory_oom_group_write,
6250 struct cgroup_subsys memory_cgrp_subsys = {
6251 .css_alloc = mem_cgroup_css_alloc,
6252 .css_online = mem_cgroup_css_online,
6253 .css_offline = mem_cgroup_css_offline,
6254 .css_released = mem_cgroup_css_released,
6255 .css_free = mem_cgroup_css_free,
6256 .css_reset = mem_cgroup_css_reset,
6257 .can_attach = mem_cgroup_can_attach,
6258 .cancel_attach = mem_cgroup_cancel_attach,
6259 .post_attach = mem_cgroup_move_task,
6260 .bind = mem_cgroup_bind,
6261 .dfl_cftypes = memory_files,
6262 .legacy_cftypes = mem_cgroup_legacy_files,
6267 * This function calculates an individual cgroup's effective
6268 * protection which is derived from its own memory.min/low, its
6269 * parent's and siblings' settings, as well as the actual memory
6270 * distribution in the tree.
6272 * The following rules apply to the effective protection values:
6274 * 1. At the first level of reclaim, effective protection is equal to
6275 * the declared protection in memory.min and memory.low.
6277 * 2. To enable safe delegation of the protection configuration, at
6278 * subsequent levels the effective protection is capped to the
6279 * parent's effective protection.
6281 * 3. To make complex and dynamic subtrees easier to configure, the
6282 * user is allowed to overcommit the declared protection at a given
6283 * level. If that is the case, the parent's effective protection is
6284 * distributed to the children in proportion to how much protection
6285 * they have declared and how much of it they are utilizing.
6287 * This makes distribution proportional, but also work-conserving:
6288 * if one cgroup claims much more protection than it uses memory,
6289 * the unused remainder is available to its siblings.
6291 * 4. Conversely, when the declared protection is undercommitted at a
6292 * given level, the distribution of the larger parental protection
6293 * budget is NOT proportional. A cgroup's protection from a sibling
6294 * is capped to its own memory.min/low setting.
6296 * 5. However, to allow protecting recursive subtrees from each other
6297 * without having to declare each individual cgroup's fixed share
6298 * of the ancestor's claim to protection, any unutilized -
6299 * "floating" - protection from up the tree is distributed in
6300 * proportion to each cgroup's *usage*. This makes the protection
6301 * neutral wrt sibling cgroups and lets them compete freely over
6302 * the shared parental protection budget, but it protects the
6303 * subtree as a whole from neighboring subtrees.
6305 * Note that 4. and 5. are not in conflict: 4. is about protecting
6306 * against immediate siblings whereas 5. is about protecting against
6307 * neighboring subtrees.
6309 static unsigned long effective_protection(unsigned long usage,
6310 unsigned long parent_usage,
6311 unsigned long setting,
6312 unsigned long parent_effective,
6313 unsigned long siblings_protected)
6315 unsigned long protected;
6318 protected = min(usage, setting);
6320 * If all cgroups at this level combined claim and use more
6321 * protection then what the parent affords them, distribute
6322 * shares in proportion to utilization.
6324 * We are using actual utilization rather than the statically
6325 * claimed protection in order to be work-conserving: claimed
6326 * but unused protection is available to siblings that would
6327 * otherwise get a smaller chunk than what they claimed.
6329 if (siblings_protected > parent_effective)
6330 return protected * parent_effective / siblings_protected;
6333 * Ok, utilized protection of all children is within what the
6334 * parent affords them, so we know whatever this child claims
6335 * and utilizes is effectively protected.
6337 * If there is unprotected usage beyond this value, reclaim
6338 * will apply pressure in proportion to that amount.
6340 * If there is unutilized protection, the cgroup will be fully
6341 * shielded from reclaim, but we do return a smaller value for
6342 * protection than what the group could enjoy in theory. This
6343 * is okay. With the overcommit distribution above, effective
6344 * protection is always dependent on how memory is actually
6345 * consumed among the siblings anyway.
6350 * If the children aren't claiming (all of) the protection
6351 * afforded to them by the parent, distribute the remainder in
6352 * proportion to the (unprotected) memory of each cgroup. That
6353 * way, cgroups that aren't explicitly prioritized wrt each
6354 * other compete freely over the allowance, but they are
6355 * collectively protected from neighboring trees.
6357 * We're using unprotected memory for the weight so that if
6358 * some cgroups DO claim explicit protection, we don't protect
6359 * the same bytes twice.
6361 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6364 if (parent_effective > siblings_protected && usage > protected) {
6365 unsigned long unclaimed;
6367 unclaimed = parent_effective - siblings_protected;
6368 unclaimed *= usage - protected;
6369 unclaimed /= parent_usage - siblings_protected;
6378 * mem_cgroup_protected - check if memory consumption is in the normal range
6379 * @root: the top ancestor of the sub-tree being checked
6380 * @memcg: the memory cgroup to check
6382 * WARNING: This function is not stateless! It can only be used as part
6383 * of a top-down tree iteration, not for isolated queries.
6385 * Returns one of the following:
6386 * MEMCG_PROT_NONE: cgroup memory is not protected
6387 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
6388 * an unprotected supply of reclaimable memory from other cgroups.
6389 * MEMCG_PROT_MIN: cgroup memory is protected
6391 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
6392 struct mem_cgroup *memcg)
6394 unsigned long usage, parent_usage;
6395 struct mem_cgroup *parent;
6397 if (mem_cgroup_disabled())
6398 return MEMCG_PROT_NONE;
6401 root = root_mem_cgroup;
6403 return MEMCG_PROT_NONE;
6405 usage = page_counter_read(&memcg->memory);
6407 return MEMCG_PROT_NONE;
6409 parent = parent_mem_cgroup(memcg);
6410 /* No parent means a non-hierarchical mode on v1 memcg */
6412 return MEMCG_PROT_NONE;
6414 if (parent == root) {
6415 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6416 memcg->memory.elow = memcg->memory.low;
6420 parent_usage = page_counter_read(&parent->memory);
6422 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6423 READ_ONCE(memcg->memory.min),
6424 READ_ONCE(parent->memory.emin),
6425 atomic_long_read(&parent->memory.children_min_usage)));
6427 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6428 memcg->memory.low, READ_ONCE(parent->memory.elow),
6429 atomic_long_read(&parent->memory.children_low_usage)));
6432 if (usage <= memcg->memory.emin)
6433 return MEMCG_PROT_MIN;
6434 else if (usage <= memcg->memory.elow)
6435 return MEMCG_PROT_LOW;
6437 return MEMCG_PROT_NONE;
6441 * mem_cgroup_try_charge - try charging a page
6442 * @page: page to charge
6443 * @mm: mm context of the victim
6444 * @gfp_mask: reclaim mode
6445 * @memcgp: charged memcg return
6446 * @compound: charge the page as compound or small page
6448 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6449 * pages according to @gfp_mask if necessary.
6451 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6452 * Otherwise, an error code is returned.
6454 * After page->mapping has been set up, the caller must finalize the
6455 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6456 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6458 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
6459 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6462 struct mem_cgroup *memcg = NULL;
6463 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6466 if (mem_cgroup_disabled())
6469 if (PageSwapCache(page)) {
6471 * Every swap fault against a single page tries to charge the
6472 * page, bail as early as possible. shmem_unuse() encounters
6473 * already charged pages, too. The USED bit is protected by
6474 * the page lock, which serializes swap cache removal, which
6475 * in turn serializes uncharging.
6477 VM_BUG_ON_PAGE(!PageLocked(page), page);
6478 if (compound_head(page)->mem_cgroup)
6481 if (do_swap_account) {
6482 swp_entry_t ent = { .val = page_private(page), };
6483 unsigned short id = lookup_swap_cgroup_id(ent);
6486 memcg = mem_cgroup_from_id(id);
6487 if (memcg && !css_tryget_online(&memcg->css))
6494 memcg = get_mem_cgroup_from_mm(mm);
6496 ret = try_charge(memcg, gfp_mask, nr_pages);
6498 css_put(&memcg->css);
6504 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
6505 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6508 struct mem_cgroup *memcg;
6511 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound);
6513 mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask);
6518 * mem_cgroup_commit_charge - commit a page charge
6519 * @page: page to charge
6520 * @memcg: memcg to charge the page to
6521 * @lrucare: page might be on LRU already
6522 * @compound: charge the page as compound or small page
6524 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6525 * after page->mapping has been set up. This must happen atomically
6526 * as part of the page instantiation, i.e. under the page table lock
6527 * for anonymous pages, under the page lock for page and swap cache.
6529 * In addition, the page must not be on the LRU during the commit, to
6530 * prevent racing with task migration. If it might be, use @lrucare.
6532 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6534 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
6535 bool lrucare, bool compound)
6537 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6539 VM_BUG_ON_PAGE(!page->mapping, page);
6540 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
6542 if (mem_cgroup_disabled())
6545 * Swap faults will attempt to charge the same page multiple
6546 * times. But reuse_swap_page() might have removed the page
6547 * from swapcache already, so we can't check PageSwapCache().
6552 commit_charge(page, memcg, lrucare);
6554 local_irq_disable();
6555 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
6556 memcg_check_events(memcg, page);
6559 if (do_memsw_account() && PageSwapCache(page)) {
6560 swp_entry_t entry = { .val = page_private(page) };
6562 * The swap entry might not get freed for a long time,
6563 * let's not wait for it. The page already received a
6564 * memory+swap charge, drop the swap entry duplicate.
6566 mem_cgroup_uncharge_swap(entry, nr_pages);
6571 * mem_cgroup_cancel_charge - cancel a page charge
6572 * @page: page to charge
6573 * @memcg: memcg to charge the page to
6574 * @compound: charge the page as compound or small page
6576 * Cancel a charge transaction started by mem_cgroup_try_charge().
6578 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
6581 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6583 if (mem_cgroup_disabled())
6586 * Swap faults will attempt to charge the same page multiple
6587 * times. But reuse_swap_page() might have removed the page
6588 * from swapcache already, so we can't check PageSwapCache().
6593 cancel_charge(memcg, nr_pages);
6596 struct uncharge_gather {
6597 struct mem_cgroup *memcg;
6598 unsigned long pgpgout;
6599 unsigned long nr_anon;
6600 unsigned long nr_file;
6601 unsigned long nr_kmem;
6602 unsigned long nr_huge;
6603 unsigned long nr_shmem;
6604 struct page *dummy_page;
6607 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6609 memset(ug, 0, sizeof(*ug));
6612 static void uncharge_batch(const struct uncharge_gather *ug)
6614 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
6615 unsigned long flags;
6617 if (!mem_cgroup_is_root(ug->memcg)) {
6618 page_counter_uncharge(&ug->memcg->memory, nr_pages);
6619 if (do_memsw_account())
6620 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
6621 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6622 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6623 memcg_oom_recover(ug->memcg);
6626 local_irq_save(flags);
6627 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
6628 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
6629 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
6630 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
6631 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6632 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, nr_pages);
6633 memcg_check_events(ug->memcg, ug->dummy_page);
6634 local_irq_restore(flags);
6636 if (!mem_cgroup_is_root(ug->memcg))
6637 css_put_many(&ug->memcg->css, nr_pages);
6640 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6642 VM_BUG_ON_PAGE(PageLRU(page), page);
6643 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
6644 !PageHWPoison(page) , page);
6646 if (!page->mem_cgroup)
6650 * Nobody should be changing or seriously looking at
6651 * page->mem_cgroup at this point, we have fully
6652 * exclusive access to the page.
6655 if (ug->memcg != page->mem_cgroup) {
6658 uncharge_gather_clear(ug);
6660 ug->memcg = page->mem_cgroup;
6663 if (!PageKmemcg(page)) {
6664 unsigned int nr_pages = 1;
6666 if (PageTransHuge(page)) {
6667 nr_pages = compound_nr(page);
6668 ug->nr_huge += nr_pages;
6671 ug->nr_anon += nr_pages;
6673 ug->nr_file += nr_pages;
6674 if (PageSwapBacked(page))
6675 ug->nr_shmem += nr_pages;
6679 ug->nr_kmem += compound_nr(page);
6680 __ClearPageKmemcg(page);
6683 ug->dummy_page = page;
6684 page->mem_cgroup = NULL;
6687 static void uncharge_list(struct list_head *page_list)
6689 struct uncharge_gather ug;
6690 struct list_head *next;
6692 uncharge_gather_clear(&ug);
6695 * Note that the list can be a single page->lru; hence the
6696 * do-while loop instead of a simple list_for_each_entry().
6698 next = page_list->next;
6702 page = list_entry(next, struct page, lru);
6703 next = page->lru.next;
6705 uncharge_page(page, &ug);
6706 } while (next != page_list);
6709 uncharge_batch(&ug);
6713 * mem_cgroup_uncharge - uncharge a page
6714 * @page: page to uncharge
6716 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6717 * mem_cgroup_commit_charge().
6719 void mem_cgroup_uncharge(struct page *page)
6721 struct uncharge_gather ug;
6723 if (mem_cgroup_disabled())
6726 /* Don't touch page->lru of any random page, pre-check: */
6727 if (!page->mem_cgroup)
6730 uncharge_gather_clear(&ug);
6731 uncharge_page(page, &ug);
6732 uncharge_batch(&ug);
6736 * mem_cgroup_uncharge_list - uncharge a list of page
6737 * @page_list: list of pages to uncharge
6739 * Uncharge a list of pages previously charged with
6740 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6742 void mem_cgroup_uncharge_list(struct list_head *page_list)
6744 if (mem_cgroup_disabled())
6747 if (!list_empty(page_list))
6748 uncharge_list(page_list);
6752 * mem_cgroup_migrate - charge a page's replacement
6753 * @oldpage: currently circulating page
6754 * @newpage: replacement page
6756 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6757 * be uncharged upon free.
6759 * Both pages must be locked, @newpage->mapping must be set up.
6761 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6763 struct mem_cgroup *memcg;
6764 unsigned int nr_pages;
6765 unsigned long flags;
6767 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6768 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6769 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6770 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6773 if (mem_cgroup_disabled())
6776 /* Page cache replacement: new page already charged? */
6777 if (newpage->mem_cgroup)
6780 /* Swapcache readahead pages can get replaced before being charged */
6781 memcg = oldpage->mem_cgroup;
6785 /* Force-charge the new page. The old one will be freed soon */
6786 nr_pages = hpage_nr_pages(newpage);
6788 page_counter_charge(&memcg->memory, nr_pages);
6789 if (do_memsw_account())
6790 page_counter_charge(&memcg->memsw, nr_pages);
6791 css_get_many(&memcg->css, nr_pages);
6793 commit_charge(newpage, memcg, false);
6795 local_irq_save(flags);
6796 mem_cgroup_charge_statistics(memcg, newpage, PageTransHuge(newpage),
6798 memcg_check_events(memcg, newpage);
6799 local_irq_restore(flags);
6802 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6803 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6805 void mem_cgroup_sk_alloc(struct sock *sk)
6807 struct mem_cgroup *memcg;
6809 if (!mem_cgroup_sockets_enabled)
6812 /* Do not associate the sock with unrelated interrupted task's memcg. */
6817 memcg = mem_cgroup_from_task(current);
6818 if (memcg == root_mem_cgroup)
6820 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6822 if (css_tryget(&memcg->css))
6823 sk->sk_memcg = memcg;
6828 void mem_cgroup_sk_free(struct sock *sk)
6831 css_put(&sk->sk_memcg->css);
6835 * mem_cgroup_charge_skmem - charge socket memory
6836 * @memcg: memcg to charge
6837 * @nr_pages: number of pages to charge
6839 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6840 * @memcg's configured limit, %false if the charge had to be forced.
6842 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6844 gfp_t gfp_mask = GFP_KERNEL;
6846 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6847 struct page_counter *fail;
6849 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6850 memcg->tcpmem_pressure = 0;
6853 page_counter_charge(&memcg->tcpmem, nr_pages);
6854 memcg->tcpmem_pressure = 1;
6858 /* Don't block in the packet receive path */
6860 gfp_mask = GFP_NOWAIT;
6862 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6864 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6867 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6872 * mem_cgroup_uncharge_skmem - uncharge socket memory
6873 * @memcg: memcg to uncharge
6874 * @nr_pages: number of pages to uncharge
6876 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6878 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6879 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6883 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6885 refill_stock(memcg, nr_pages);
6888 static int __init cgroup_memory(char *s)
6892 while ((token = strsep(&s, ",")) != NULL) {
6895 if (!strcmp(token, "nosocket"))
6896 cgroup_memory_nosocket = true;
6897 if (!strcmp(token, "nokmem"))
6898 cgroup_memory_nokmem = true;
6902 __setup("cgroup.memory=", cgroup_memory);
6905 * subsys_initcall() for memory controller.
6907 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6908 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6909 * basically everything that doesn't depend on a specific mem_cgroup structure
6910 * should be initialized from here.
6912 static int __init mem_cgroup_init(void)
6916 #ifdef CONFIG_MEMCG_KMEM
6918 * Kmem cache creation is mostly done with the slab_mutex held,
6919 * so use a workqueue with limited concurrency to avoid stalling
6920 * all worker threads in case lots of cgroups are created and
6921 * destroyed simultaneously.
6923 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
6924 BUG_ON(!memcg_kmem_cache_wq);
6927 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6928 memcg_hotplug_cpu_dead);
6930 for_each_possible_cpu(cpu)
6931 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
6934 for_each_node(node) {
6935 struct mem_cgroup_tree_per_node *rtpn;
6937 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
6938 node_online(node) ? node : NUMA_NO_NODE);
6940 rtpn->rb_root = RB_ROOT;
6941 rtpn->rb_rightmost = NULL;
6942 spin_lock_init(&rtpn->lock);
6943 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6948 subsys_initcall(mem_cgroup_init);
6950 #ifdef CONFIG_MEMCG_SWAP
6951 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
6953 while (!refcount_inc_not_zero(&memcg->id.ref)) {
6955 * The root cgroup cannot be destroyed, so it's refcount must
6958 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
6962 memcg = parent_mem_cgroup(memcg);
6964 memcg = root_mem_cgroup;
6970 * mem_cgroup_swapout - transfer a memsw charge to swap
6971 * @page: page whose memsw charge to transfer
6972 * @entry: swap entry to move the charge to
6974 * Transfer the memsw charge of @page to @entry.
6976 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6978 struct mem_cgroup *memcg, *swap_memcg;
6979 unsigned int nr_entries;
6980 unsigned short oldid;
6982 VM_BUG_ON_PAGE(PageLRU(page), page);
6983 VM_BUG_ON_PAGE(page_count(page), page);
6985 if (!do_memsw_account())
6988 memcg = page->mem_cgroup;
6990 /* Readahead page, never charged */
6995 * In case the memcg owning these pages has been offlined and doesn't
6996 * have an ID allocated to it anymore, charge the closest online
6997 * ancestor for the swap instead and transfer the memory+swap charge.
6999 swap_memcg = mem_cgroup_id_get_online(memcg);
7000 nr_entries = hpage_nr_pages(page);
7001 /* Get references for the tail pages, too */
7003 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7004 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7006 VM_BUG_ON_PAGE(oldid, page);
7007 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7009 page->mem_cgroup = NULL;
7011 if (!mem_cgroup_is_root(memcg))
7012 page_counter_uncharge(&memcg->memory, nr_entries);
7014 if (memcg != swap_memcg) {
7015 if (!mem_cgroup_is_root(swap_memcg))
7016 page_counter_charge(&swap_memcg->memsw, nr_entries);
7017 page_counter_uncharge(&memcg->memsw, nr_entries);
7021 * Interrupts should be disabled here because the caller holds the
7022 * i_pages lock which is taken with interrupts-off. It is
7023 * important here to have the interrupts disabled because it is the
7024 * only synchronisation we have for updating the per-CPU variables.
7026 VM_BUG_ON(!irqs_disabled());
7027 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
7029 memcg_check_events(memcg, page);
7031 if (!mem_cgroup_is_root(memcg))
7032 css_put_many(&memcg->css, nr_entries);
7036 * mem_cgroup_try_charge_swap - try charging swap space for a page
7037 * @page: page being added to swap
7038 * @entry: swap entry to charge
7040 * Try to charge @page's memcg for the swap space at @entry.
7042 * Returns 0 on success, -ENOMEM on failure.
7044 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7046 unsigned int nr_pages = hpage_nr_pages(page);
7047 struct page_counter *counter;
7048 struct mem_cgroup *memcg;
7049 unsigned short oldid;
7051 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
7054 memcg = page->mem_cgroup;
7056 /* Readahead page, never charged */
7061 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7065 memcg = mem_cgroup_id_get_online(memcg);
7067 if (!mem_cgroup_is_root(memcg) &&
7068 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7069 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7070 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7071 mem_cgroup_id_put(memcg);
7075 /* Get references for the tail pages, too */
7077 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7078 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7079 VM_BUG_ON_PAGE(oldid, page);
7080 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7086 * mem_cgroup_uncharge_swap - uncharge swap space
7087 * @entry: swap entry to uncharge
7088 * @nr_pages: the amount of swap space to uncharge
7090 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7092 struct mem_cgroup *memcg;
7095 if (!do_swap_account)
7098 id = swap_cgroup_record(entry, 0, nr_pages);
7100 memcg = mem_cgroup_from_id(id);
7102 if (!mem_cgroup_is_root(memcg)) {
7103 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7104 page_counter_uncharge(&memcg->swap, nr_pages);
7106 page_counter_uncharge(&memcg->memsw, nr_pages);
7108 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7109 mem_cgroup_id_put_many(memcg, nr_pages);
7114 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7116 long nr_swap_pages = get_nr_swap_pages();
7118 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7119 return nr_swap_pages;
7120 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7121 nr_swap_pages = min_t(long, nr_swap_pages,
7122 READ_ONCE(memcg->swap.max) -
7123 page_counter_read(&memcg->swap));
7124 return nr_swap_pages;
7127 bool mem_cgroup_swap_full(struct page *page)
7129 struct mem_cgroup *memcg;
7131 VM_BUG_ON_PAGE(!PageLocked(page), page);
7135 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7138 memcg = page->mem_cgroup;
7142 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7143 if (page_counter_read(&memcg->swap) * 2 >=
7144 READ_ONCE(memcg->swap.max))
7150 /* for remember boot option*/
7151 #ifdef CONFIG_MEMCG_SWAP_ENABLED
7152 static int really_do_swap_account __initdata = 1;
7154 static int really_do_swap_account __initdata;
7157 static int __init enable_swap_account(char *s)
7159 if (!strcmp(s, "1"))
7160 really_do_swap_account = 1;
7161 else if (!strcmp(s, "0"))
7162 really_do_swap_account = 0;
7165 __setup("swapaccount=", enable_swap_account);
7167 static u64 swap_current_read(struct cgroup_subsys_state *css,
7170 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7172 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7175 static int swap_max_show(struct seq_file *m, void *v)
7177 return seq_puts_memcg_tunable(m,
7178 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7181 static ssize_t swap_max_write(struct kernfs_open_file *of,
7182 char *buf, size_t nbytes, loff_t off)
7184 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7188 buf = strstrip(buf);
7189 err = page_counter_memparse(buf, "max", &max);
7193 xchg(&memcg->swap.max, max);
7198 static int swap_events_show(struct seq_file *m, void *v)
7200 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7202 seq_printf(m, "max %lu\n",
7203 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7204 seq_printf(m, "fail %lu\n",
7205 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7210 static struct cftype swap_files[] = {
7212 .name = "swap.current",
7213 .flags = CFTYPE_NOT_ON_ROOT,
7214 .read_u64 = swap_current_read,
7218 .flags = CFTYPE_NOT_ON_ROOT,
7219 .seq_show = swap_max_show,
7220 .write = swap_max_write,
7223 .name = "swap.events",
7224 .flags = CFTYPE_NOT_ON_ROOT,
7225 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7226 .seq_show = swap_events_show,
7231 static struct cftype memsw_cgroup_files[] = {
7233 .name = "memsw.usage_in_bytes",
7234 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7235 .read_u64 = mem_cgroup_read_u64,
7238 .name = "memsw.max_usage_in_bytes",
7239 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7240 .write = mem_cgroup_reset,
7241 .read_u64 = mem_cgroup_read_u64,
7244 .name = "memsw.limit_in_bytes",
7245 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7246 .write = mem_cgroup_write,
7247 .read_u64 = mem_cgroup_read_u64,
7250 .name = "memsw.failcnt",
7251 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7252 .write = mem_cgroup_reset,
7253 .read_u64 = mem_cgroup_read_u64,
7255 { }, /* terminate */
7258 static int __init mem_cgroup_swap_init(void)
7260 if (!mem_cgroup_disabled() && really_do_swap_account) {
7261 do_swap_account = 1;
7262 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
7264 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
7265 memsw_cgroup_files));
7269 subsys_initcall(mem_cgroup_swap_init);
7271 #endif /* CONFIG_MEMCG_SWAP */