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_online(&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 if (!memcg || !css_tryget_online(&memcg->css))
976 memcg = root_mem_cgroup;
980 EXPORT_SYMBOL(get_mem_cgroup_from_page);
983 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
985 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
987 if (unlikely(current->active_memcg)) {
988 struct mem_cgroup *memcg = root_mem_cgroup;
991 if (css_tryget_online(¤t->active_memcg->css))
992 memcg = current->active_memcg;
996 return get_mem_cgroup_from_mm(current->mm);
1000 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1001 * @root: hierarchy root
1002 * @prev: previously returned memcg, NULL on first invocation
1003 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1005 * Returns references to children of the hierarchy below @root, or
1006 * @root itself, or %NULL after a full round-trip.
1008 * Caller must pass the return value in @prev on subsequent
1009 * invocations for reference counting, or use mem_cgroup_iter_break()
1010 * to cancel a hierarchy walk before the round-trip is complete.
1012 * Reclaimers can specify a node and a priority level in @reclaim to
1013 * divide up the memcgs in the hierarchy among all concurrent
1014 * reclaimers operating on the same node and priority.
1016 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1017 struct mem_cgroup *prev,
1018 struct mem_cgroup_reclaim_cookie *reclaim)
1020 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1021 struct cgroup_subsys_state *css = NULL;
1022 struct mem_cgroup *memcg = NULL;
1023 struct mem_cgroup *pos = NULL;
1025 if (mem_cgroup_disabled())
1029 root = root_mem_cgroup;
1031 if (prev && !reclaim)
1034 if (!root->use_hierarchy && root != root_mem_cgroup) {
1043 struct mem_cgroup_per_node *mz;
1045 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1048 if (prev && reclaim->generation != iter->generation)
1052 pos = READ_ONCE(iter->position);
1053 if (!pos || css_tryget(&pos->css))
1056 * css reference reached zero, so iter->position will
1057 * be cleared by ->css_released. However, we should not
1058 * rely on this happening soon, because ->css_released
1059 * is called from a work queue, and by busy-waiting we
1060 * might block it. So we clear iter->position right
1063 (void)cmpxchg(&iter->position, pos, NULL);
1071 css = css_next_descendant_pre(css, &root->css);
1074 * Reclaimers share the hierarchy walk, and a
1075 * new one might jump in right at the end of
1076 * the hierarchy - make sure they see at least
1077 * one group and restart from the beginning.
1085 * Verify the css and acquire a reference. The root
1086 * is provided by the caller, so we know it's alive
1087 * and kicking, and don't take an extra reference.
1089 memcg = mem_cgroup_from_css(css);
1091 if (css == &root->css)
1094 if (css_tryget(css))
1102 * The position could have already been updated by a competing
1103 * thread, so check that the value hasn't changed since we read
1104 * it to avoid reclaiming from the same cgroup twice.
1106 (void)cmpxchg(&iter->position, pos, memcg);
1114 reclaim->generation = iter->generation;
1120 if (prev && prev != root)
1121 css_put(&prev->css);
1127 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1128 * @root: hierarchy root
1129 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1131 void mem_cgroup_iter_break(struct mem_cgroup *root,
1132 struct mem_cgroup *prev)
1135 root = root_mem_cgroup;
1136 if (prev && prev != root)
1137 css_put(&prev->css);
1140 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1141 struct mem_cgroup *dead_memcg)
1143 struct mem_cgroup_reclaim_iter *iter;
1144 struct mem_cgroup_per_node *mz;
1147 for_each_node(nid) {
1148 mz = mem_cgroup_nodeinfo(from, nid);
1150 cmpxchg(&iter->position, dead_memcg, NULL);
1154 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1156 struct mem_cgroup *memcg = dead_memcg;
1157 struct mem_cgroup *last;
1160 __invalidate_reclaim_iterators(memcg, dead_memcg);
1162 } while ((memcg = parent_mem_cgroup(memcg)));
1165 * When cgruop1 non-hierarchy mode is used,
1166 * parent_mem_cgroup() does not walk all the way up to the
1167 * cgroup root (root_mem_cgroup). So we have to handle
1168 * dead_memcg from cgroup root separately.
1170 if (last != root_mem_cgroup)
1171 __invalidate_reclaim_iterators(root_mem_cgroup,
1176 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1177 * @memcg: hierarchy root
1178 * @fn: function to call for each task
1179 * @arg: argument passed to @fn
1181 * This function iterates over tasks attached to @memcg or to any of its
1182 * descendants and calls @fn for each task. If @fn returns a non-zero
1183 * value, the function breaks the iteration loop and returns the value.
1184 * Otherwise, it will iterate over all tasks and return 0.
1186 * This function must not be called for the root memory cgroup.
1188 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1189 int (*fn)(struct task_struct *, void *), void *arg)
1191 struct mem_cgroup *iter;
1194 BUG_ON(memcg == root_mem_cgroup);
1196 for_each_mem_cgroup_tree(iter, memcg) {
1197 struct css_task_iter it;
1198 struct task_struct *task;
1200 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1201 while (!ret && (task = css_task_iter_next(&it)))
1202 ret = fn(task, arg);
1203 css_task_iter_end(&it);
1205 mem_cgroup_iter_break(memcg, iter);
1213 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1215 * @pgdat: pgdat of the page
1217 * This function is only safe when following the LRU page isolation
1218 * and putback protocol: the LRU lock must be held, and the page must
1219 * either be PageLRU() or the caller must have isolated/allocated it.
1221 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1223 struct mem_cgroup_per_node *mz;
1224 struct mem_cgroup *memcg;
1225 struct lruvec *lruvec;
1227 if (mem_cgroup_disabled()) {
1228 lruvec = &pgdat->__lruvec;
1232 memcg = page->mem_cgroup;
1234 * Swapcache readahead pages are added to the LRU - and
1235 * possibly migrated - before they are charged.
1238 memcg = root_mem_cgroup;
1240 mz = mem_cgroup_page_nodeinfo(memcg, page);
1241 lruvec = &mz->lruvec;
1244 * Since a node can be onlined after the mem_cgroup was created,
1245 * we have to be prepared to initialize lruvec->zone here;
1246 * and if offlined then reonlined, we need to reinitialize it.
1248 if (unlikely(lruvec->pgdat != pgdat))
1249 lruvec->pgdat = pgdat;
1254 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1255 * @lruvec: mem_cgroup per zone lru vector
1256 * @lru: index of lru list the page is sitting on
1257 * @zid: zone id of the accounted pages
1258 * @nr_pages: positive when adding or negative when removing
1260 * This function must be called under lru_lock, just before a page is added
1261 * to or just after a page is removed from an lru list (that ordering being
1262 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1264 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1265 int zid, int nr_pages)
1267 struct mem_cgroup_per_node *mz;
1268 unsigned long *lru_size;
1271 if (mem_cgroup_disabled())
1274 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1275 lru_size = &mz->lru_zone_size[zid][lru];
1278 *lru_size += nr_pages;
1281 if (WARN_ONCE(size < 0,
1282 "%s(%p, %d, %d): lru_size %ld\n",
1283 __func__, lruvec, lru, nr_pages, size)) {
1289 *lru_size += nr_pages;
1293 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1294 * @memcg: the memory cgroup
1296 * Returns the maximum amount of memory @mem can be charged with, in
1299 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1301 unsigned long margin = 0;
1302 unsigned long count;
1303 unsigned long limit;
1305 count = page_counter_read(&memcg->memory);
1306 limit = READ_ONCE(memcg->memory.max);
1308 margin = limit - count;
1310 if (do_memsw_account()) {
1311 count = page_counter_read(&memcg->memsw);
1312 limit = READ_ONCE(memcg->memsw.max);
1314 margin = min(margin, limit - count);
1323 * A routine for checking "mem" is under move_account() or not.
1325 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1326 * moving cgroups. This is for waiting at high-memory pressure
1329 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1331 struct mem_cgroup *from;
1332 struct mem_cgroup *to;
1335 * Unlike task_move routines, we access mc.to, mc.from not under
1336 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1338 spin_lock(&mc.lock);
1344 ret = mem_cgroup_is_descendant(from, memcg) ||
1345 mem_cgroup_is_descendant(to, memcg);
1347 spin_unlock(&mc.lock);
1351 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1353 if (mc.moving_task && current != mc.moving_task) {
1354 if (mem_cgroup_under_move(memcg)) {
1356 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1357 /* moving charge context might have finished. */
1360 finish_wait(&mc.waitq, &wait);
1367 static char *memory_stat_format(struct mem_cgroup *memcg)
1372 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1377 * Provide statistics on the state of the memory subsystem as
1378 * well as cumulative event counters that show past behavior.
1380 * This list is ordered following a combination of these gradients:
1381 * 1) generic big picture -> specifics and details
1382 * 2) reflecting userspace activity -> reflecting kernel heuristics
1384 * Current memory state:
1387 seq_buf_printf(&s, "anon %llu\n",
1388 (u64)memcg_page_state(memcg, MEMCG_RSS) *
1390 seq_buf_printf(&s, "file %llu\n",
1391 (u64)memcg_page_state(memcg, MEMCG_CACHE) *
1393 seq_buf_printf(&s, "kernel_stack %llu\n",
1394 (u64)memcg_page_state(memcg, MEMCG_KERNEL_STACK_KB) *
1396 seq_buf_printf(&s, "slab %llu\n",
1397 (u64)(memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) +
1398 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE)) *
1400 seq_buf_printf(&s, "sock %llu\n",
1401 (u64)memcg_page_state(memcg, MEMCG_SOCK) *
1404 seq_buf_printf(&s, "shmem %llu\n",
1405 (u64)memcg_page_state(memcg, NR_SHMEM) *
1407 seq_buf_printf(&s, "file_mapped %llu\n",
1408 (u64)memcg_page_state(memcg, NR_FILE_MAPPED) *
1410 seq_buf_printf(&s, "file_dirty %llu\n",
1411 (u64)memcg_page_state(memcg, NR_FILE_DIRTY) *
1413 seq_buf_printf(&s, "file_writeback %llu\n",
1414 (u64)memcg_page_state(memcg, NR_WRITEBACK) *
1418 * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter
1419 * with the NR_ANON_THP vm counter, but right now it's a pain in the
1420 * arse because it requires migrating the work out of rmap to a place
1421 * where the page->mem_cgroup is set up and stable.
1423 seq_buf_printf(&s, "anon_thp %llu\n",
1424 (u64)memcg_page_state(memcg, MEMCG_RSS_HUGE) *
1427 for (i = 0; i < NR_LRU_LISTS; i++)
1428 seq_buf_printf(&s, "%s %llu\n", lru_list_name(i),
1429 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
1432 seq_buf_printf(&s, "slab_reclaimable %llu\n",
1433 (u64)memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) *
1435 seq_buf_printf(&s, "slab_unreclaimable %llu\n",
1436 (u64)memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE) *
1439 /* Accumulated memory events */
1441 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1442 memcg_events(memcg, PGFAULT));
1443 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1444 memcg_events(memcg, PGMAJFAULT));
1446 seq_buf_printf(&s, "workingset_refault %lu\n",
1447 memcg_page_state(memcg, WORKINGSET_REFAULT));
1448 seq_buf_printf(&s, "workingset_activate %lu\n",
1449 memcg_page_state(memcg, WORKINGSET_ACTIVATE));
1450 seq_buf_printf(&s, "workingset_nodereclaim %lu\n",
1451 memcg_page_state(memcg, WORKINGSET_NODERECLAIM));
1453 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1454 memcg_events(memcg, PGREFILL));
1455 seq_buf_printf(&s, "pgscan %lu\n",
1456 memcg_events(memcg, PGSCAN_KSWAPD) +
1457 memcg_events(memcg, PGSCAN_DIRECT));
1458 seq_buf_printf(&s, "pgsteal %lu\n",
1459 memcg_events(memcg, PGSTEAL_KSWAPD) +
1460 memcg_events(memcg, PGSTEAL_DIRECT));
1461 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1462 memcg_events(memcg, PGACTIVATE));
1463 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1464 memcg_events(memcg, PGDEACTIVATE));
1465 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1466 memcg_events(memcg, PGLAZYFREE));
1467 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1468 memcg_events(memcg, PGLAZYFREED));
1470 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1471 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1472 memcg_events(memcg, THP_FAULT_ALLOC));
1473 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1474 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1475 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1477 /* The above should easily fit into one page */
1478 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1483 #define K(x) ((x) << (PAGE_SHIFT-10))
1485 * mem_cgroup_print_oom_context: Print OOM information relevant to
1486 * memory controller.
1487 * @memcg: The memory cgroup that went over limit
1488 * @p: Task that is going to be killed
1490 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1493 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1498 pr_cont(",oom_memcg=");
1499 pr_cont_cgroup_path(memcg->css.cgroup);
1501 pr_cont(",global_oom");
1503 pr_cont(",task_memcg=");
1504 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1510 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1511 * memory controller.
1512 * @memcg: The memory cgroup that went over limit
1514 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1518 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1519 K((u64)page_counter_read(&memcg->memory)),
1520 K((u64)memcg->memory.max), memcg->memory.failcnt);
1521 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1522 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1523 K((u64)page_counter_read(&memcg->swap)),
1524 K((u64)memcg->swap.max), memcg->swap.failcnt);
1526 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1527 K((u64)page_counter_read(&memcg->memsw)),
1528 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1529 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1530 K((u64)page_counter_read(&memcg->kmem)),
1531 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1534 pr_info("Memory cgroup stats for ");
1535 pr_cont_cgroup_path(memcg->css.cgroup);
1537 buf = memory_stat_format(memcg);
1545 * Return the memory (and swap, if configured) limit for a memcg.
1547 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1551 max = memcg->memory.max;
1552 if (mem_cgroup_swappiness(memcg)) {
1553 unsigned long memsw_max;
1554 unsigned long swap_max;
1556 memsw_max = memcg->memsw.max;
1557 swap_max = memcg->swap.max;
1558 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1559 max = min(max + swap_max, memsw_max);
1564 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1566 return page_counter_read(&memcg->memory);
1569 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1572 struct oom_control oc = {
1576 .gfp_mask = gfp_mask,
1581 if (mutex_lock_killable(&oom_lock))
1584 * A few threads which were not waiting at mutex_lock_killable() can
1585 * fail to bail out. Therefore, check again after holding oom_lock.
1587 ret = should_force_charge() || out_of_memory(&oc);
1588 mutex_unlock(&oom_lock);
1592 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1595 unsigned long *total_scanned)
1597 struct mem_cgroup *victim = NULL;
1600 unsigned long excess;
1601 unsigned long nr_scanned;
1602 struct mem_cgroup_reclaim_cookie reclaim = {
1606 excess = soft_limit_excess(root_memcg);
1609 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1614 * If we have not been able to reclaim
1615 * anything, it might because there are
1616 * no reclaimable pages under this hierarchy
1621 * We want to do more targeted reclaim.
1622 * excess >> 2 is not to excessive so as to
1623 * reclaim too much, nor too less that we keep
1624 * coming back to reclaim from this cgroup
1626 if (total >= (excess >> 2) ||
1627 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1632 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1633 pgdat, &nr_scanned);
1634 *total_scanned += nr_scanned;
1635 if (!soft_limit_excess(root_memcg))
1638 mem_cgroup_iter_break(root_memcg, victim);
1642 #ifdef CONFIG_LOCKDEP
1643 static struct lockdep_map memcg_oom_lock_dep_map = {
1644 .name = "memcg_oom_lock",
1648 static DEFINE_SPINLOCK(memcg_oom_lock);
1651 * Check OOM-Killer is already running under our hierarchy.
1652 * If someone is running, return false.
1654 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1656 struct mem_cgroup *iter, *failed = NULL;
1658 spin_lock(&memcg_oom_lock);
1660 for_each_mem_cgroup_tree(iter, memcg) {
1661 if (iter->oom_lock) {
1663 * this subtree of our hierarchy is already locked
1664 * so we cannot give a lock.
1667 mem_cgroup_iter_break(memcg, iter);
1670 iter->oom_lock = true;
1675 * OK, we failed to lock the whole subtree so we have
1676 * to clean up what we set up to the failing subtree
1678 for_each_mem_cgroup_tree(iter, memcg) {
1679 if (iter == failed) {
1680 mem_cgroup_iter_break(memcg, iter);
1683 iter->oom_lock = false;
1686 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1688 spin_unlock(&memcg_oom_lock);
1693 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1695 struct mem_cgroup *iter;
1697 spin_lock(&memcg_oom_lock);
1698 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1699 for_each_mem_cgroup_tree(iter, memcg)
1700 iter->oom_lock = false;
1701 spin_unlock(&memcg_oom_lock);
1704 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1706 struct mem_cgroup *iter;
1708 spin_lock(&memcg_oom_lock);
1709 for_each_mem_cgroup_tree(iter, memcg)
1711 spin_unlock(&memcg_oom_lock);
1714 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1716 struct mem_cgroup *iter;
1719 * When a new child is created while the hierarchy is under oom,
1720 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1722 spin_lock(&memcg_oom_lock);
1723 for_each_mem_cgroup_tree(iter, memcg)
1724 if (iter->under_oom > 0)
1726 spin_unlock(&memcg_oom_lock);
1729 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1731 struct oom_wait_info {
1732 struct mem_cgroup *memcg;
1733 wait_queue_entry_t wait;
1736 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1737 unsigned mode, int sync, void *arg)
1739 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1740 struct mem_cgroup *oom_wait_memcg;
1741 struct oom_wait_info *oom_wait_info;
1743 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1744 oom_wait_memcg = oom_wait_info->memcg;
1746 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1747 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1749 return autoremove_wake_function(wait, mode, sync, arg);
1752 static void memcg_oom_recover(struct mem_cgroup *memcg)
1755 * For the following lockless ->under_oom test, the only required
1756 * guarantee is that it must see the state asserted by an OOM when
1757 * this function is called as a result of userland actions
1758 * triggered by the notification of the OOM. This is trivially
1759 * achieved by invoking mem_cgroup_mark_under_oom() before
1760 * triggering notification.
1762 if (memcg && memcg->under_oom)
1763 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1773 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1775 enum oom_status ret;
1778 if (order > PAGE_ALLOC_COSTLY_ORDER)
1781 memcg_memory_event(memcg, MEMCG_OOM);
1784 * We are in the middle of the charge context here, so we
1785 * don't want to block when potentially sitting on a callstack
1786 * that holds all kinds of filesystem and mm locks.
1788 * cgroup1 allows disabling the OOM killer and waiting for outside
1789 * handling until the charge can succeed; remember the context and put
1790 * the task to sleep at the end of the page fault when all locks are
1793 * On the other hand, in-kernel OOM killer allows for an async victim
1794 * memory reclaim (oom_reaper) and that means that we are not solely
1795 * relying on the oom victim to make a forward progress and we can
1796 * invoke the oom killer here.
1798 * Please note that mem_cgroup_out_of_memory might fail to find a
1799 * victim and then we have to bail out from the charge path.
1801 if (memcg->oom_kill_disable) {
1802 if (!current->in_user_fault)
1804 css_get(&memcg->css);
1805 current->memcg_in_oom = memcg;
1806 current->memcg_oom_gfp_mask = mask;
1807 current->memcg_oom_order = order;
1812 mem_cgroup_mark_under_oom(memcg);
1814 locked = mem_cgroup_oom_trylock(memcg);
1817 mem_cgroup_oom_notify(memcg);
1819 mem_cgroup_unmark_under_oom(memcg);
1820 if (mem_cgroup_out_of_memory(memcg, mask, order))
1826 mem_cgroup_oom_unlock(memcg);
1832 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1833 * @handle: actually kill/wait or just clean up the OOM state
1835 * This has to be called at the end of a page fault if the memcg OOM
1836 * handler was enabled.
1838 * Memcg supports userspace OOM handling where failed allocations must
1839 * sleep on a waitqueue until the userspace task resolves the
1840 * situation. Sleeping directly in the charge context with all kinds
1841 * of locks held is not a good idea, instead we remember an OOM state
1842 * in the task and mem_cgroup_oom_synchronize() has to be called at
1843 * the end of the page fault to complete the OOM handling.
1845 * Returns %true if an ongoing memcg OOM situation was detected and
1846 * completed, %false otherwise.
1848 bool mem_cgroup_oom_synchronize(bool handle)
1850 struct mem_cgroup *memcg = current->memcg_in_oom;
1851 struct oom_wait_info owait;
1854 /* OOM is global, do not handle */
1861 owait.memcg = memcg;
1862 owait.wait.flags = 0;
1863 owait.wait.func = memcg_oom_wake_function;
1864 owait.wait.private = current;
1865 INIT_LIST_HEAD(&owait.wait.entry);
1867 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1868 mem_cgroup_mark_under_oom(memcg);
1870 locked = mem_cgroup_oom_trylock(memcg);
1873 mem_cgroup_oom_notify(memcg);
1875 if (locked && !memcg->oom_kill_disable) {
1876 mem_cgroup_unmark_under_oom(memcg);
1877 finish_wait(&memcg_oom_waitq, &owait.wait);
1878 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1879 current->memcg_oom_order);
1882 mem_cgroup_unmark_under_oom(memcg);
1883 finish_wait(&memcg_oom_waitq, &owait.wait);
1887 mem_cgroup_oom_unlock(memcg);
1889 * There is no guarantee that an OOM-lock contender
1890 * sees the wakeups triggered by the OOM kill
1891 * uncharges. Wake any sleepers explicitely.
1893 memcg_oom_recover(memcg);
1896 current->memcg_in_oom = NULL;
1897 css_put(&memcg->css);
1902 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1903 * @victim: task to be killed by the OOM killer
1904 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1906 * Returns a pointer to a memory cgroup, which has to be cleaned up
1907 * by killing all belonging OOM-killable tasks.
1909 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1911 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1912 struct mem_cgroup *oom_domain)
1914 struct mem_cgroup *oom_group = NULL;
1915 struct mem_cgroup *memcg;
1917 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
1921 oom_domain = root_mem_cgroup;
1925 memcg = mem_cgroup_from_task(victim);
1926 if (memcg == root_mem_cgroup)
1930 * Traverse the memory cgroup hierarchy from the victim task's
1931 * cgroup up to the OOMing cgroup (or root) to find the
1932 * highest-level memory cgroup with oom.group set.
1934 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1935 if (memcg->oom_group)
1938 if (memcg == oom_domain)
1943 css_get(&oom_group->css);
1950 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
1952 pr_info("Tasks in ");
1953 pr_cont_cgroup_path(memcg->css.cgroup);
1954 pr_cont(" are going to be killed due to memory.oom.group set\n");
1958 * lock_page_memcg - lock a page->mem_cgroup binding
1961 * This function protects unlocked LRU pages from being moved to
1964 * It ensures lifetime of the returned memcg. Caller is responsible
1965 * for the lifetime of the page; __unlock_page_memcg() is available
1966 * when @page might get freed inside the locked section.
1968 struct mem_cgroup *lock_page_memcg(struct page *page)
1970 struct mem_cgroup *memcg;
1971 unsigned long flags;
1974 * The RCU lock is held throughout the transaction. The fast
1975 * path can get away without acquiring the memcg->move_lock
1976 * because page moving starts with an RCU grace period.
1978 * The RCU lock also protects the memcg from being freed when
1979 * the page state that is going to change is the only thing
1980 * preventing the page itself from being freed. E.g. writeback
1981 * doesn't hold a page reference and relies on PG_writeback to
1982 * keep off truncation, migration and so forth.
1986 if (mem_cgroup_disabled())
1989 memcg = page->mem_cgroup;
1990 if (unlikely(!memcg))
1993 if (atomic_read(&memcg->moving_account) <= 0)
1996 spin_lock_irqsave(&memcg->move_lock, flags);
1997 if (memcg != page->mem_cgroup) {
1998 spin_unlock_irqrestore(&memcg->move_lock, flags);
2003 * When charge migration first begins, we can have locked and
2004 * unlocked page stat updates happening concurrently. Track
2005 * the task who has the lock for unlock_page_memcg().
2007 memcg->move_lock_task = current;
2008 memcg->move_lock_flags = flags;
2012 EXPORT_SYMBOL(lock_page_memcg);
2015 * __unlock_page_memcg - unlock and unpin a memcg
2018 * Unlock and unpin a memcg returned by lock_page_memcg().
2020 void __unlock_page_memcg(struct mem_cgroup *memcg)
2022 if (memcg && memcg->move_lock_task == current) {
2023 unsigned long flags = memcg->move_lock_flags;
2025 memcg->move_lock_task = NULL;
2026 memcg->move_lock_flags = 0;
2028 spin_unlock_irqrestore(&memcg->move_lock, flags);
2035 * unlock_page_memcg - unlock a page->mem_cgroup binding
2038 void unlock_page_memcg(struct page *page)
2040 __unlock_page_memcg(page->mem_cgroup);
2042 EXPORT_SYMBOL(unlock_page_memcg);
2044 struct memcg_stock_pcp {
2045 struct mem_cgroup *cached; /* this never be root cgroup */
2046 unsigned int nr_pages;
2047 struct work_struct work;
2048 unsigned long flags;
2049 #define FLUSHING_CACHED_CHARGE 0
2051 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2052 static DEFINE_MUTEX(percpu_charge_mutex);
2055 * consume_stock: Try to consume stocked charge on this cpu.
2056 * @memcg: memcg to consume from.
2057 * @nr_pages: how many pages to charge.
2059 * The charges will only happen if @memcg matches the current cpu's memcg
2060 * stock, and at least @nr_pages are available in that stock. Failure to
2061 * service an allocation will refill the stock.
2063 * returns true if successful, false otherwise.
2065 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2067 struct memcg_stock_pcp *stock;
2068 unsigned long flags;
2071 if (nr_pages > MEMCG_CHARGE_BATCH)
2074 local_irq_save(flags);
2076 stock = this_cpu_ptr(&memcg_stock);
2077 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2078 stock->nr_pages -= nr_pages;
2082 local_irq_restore(flags);
2088 * Returns stocks cached in percpu and reset cached information.
2090 static void drain_stock(struct memcg_stock_pcp *stock)
2092 struct mem_cgroup *old = stock->cached;
2094 if (stock->nr_pages) {
2095 page_counter_uncharge(&old->memory, stock->nr_pages);
2096 if (do_memsw_account())
2097 page_counter_uncharge(&old->memsw, stock->nr_pages);
2098 css_put_many(&old->css, stock->nr_pages);
2099 stock->nr_pages = 0;
2101 stock->cached = NULL;
2104 static void drain_local_stock(struct work_struct *dummy)
2106 struct memcg_stock_pcp *stock;
2107 unsigned long flags;
2110 * The only protection from memory hotplug vs. drain_stock races is
2111 * that we always operate on local CPU stock here with IRQ disabled
2113 local_irq_save(flags);
2115 stock = this_cpu_ptr(&memcg_stock);
2117 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2119 local_irq_restore(flags);
2123 * Cache charges(val) to local per_cpu area.
2124 * This will be consumed by consume_stock() function, later.
2126 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2128 struct memcg_stock_pcp *stock;
2129 unsigned long flags;
2131 local_irq_save(flags);
2133 stock = this_cpu_ptr(&memcg_stock);
2134 if (stock->cached != memcg) { /* reset if necessary */
2136 stock->cached = memcg;
2138 stock->nr_pages += nr_pages;
2140 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2143 local_irq_restore(flags);
2147 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2148 * of the hierarchy under it.
2150 static void drain_all_stock(struct mem_cgroup *root_memcg)
2154 /* If someone's already draining, avoid adding running more workers. */
2155 if (!mutex_trylock(&percpu_charge_mutex))
2158 * Notify other cpus that system-wide "drain" is running
2159 * We do not care about races with the cpu hotplug because cpu down
2160 * as well as workers from this path always operate on the local
2161 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2164 for_each_online_cpu(cpu) {
2165 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2166 struct mem_cgroup *memcg;
2170 memcg = stock->cached;
2171 if (memcg && stock->nr_pages &&
2172 mem_cgroup_is_descendant(memcg, root_memcg))
2177 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2179 drain_local_stock(&stock->work);
2181 schedule_work_on(cpu, &stock->work);
2185 mutex_unlock(&percpu_charge_mutex);
2188 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2190 struct memcg_stock_pcp *stock;
2191 struct mem_cgroup *memcg, *mi;
2193 stock = &per_cpu(memcg_stock, cpu);
2196 for_each_mem_cgroup(memcg) {
2199 for (i = 0; i < MEMCG_NR_STAT; i++) {
2203 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2205 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2206 atomic_long_add(x, &memcg->vmstats[i]);
2208 if (i >= NR_VM_NODE_STAT_ITEMS)
2211 for_each_node(nid) {
2212 struct mem_cgroup_per_node *pn;
2214 pn = mem_cgroup_nodeinfo(memcg, nid);
2215 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2218 atomic_long_add(x, &pn->lruvec_stat[i]);
2219 } while ((pn = parent_nodeinfo(pn, nid)));
2223 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2226 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2228 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2229 atomic_long_add(x, &memcg->vmevents[i]);
2236 static void reclaim_high(struct mem_cgroup *memcg,
2237 unsigned int nr_pages,
2241 if (page_counter_read(&memcg->memory) <= memcg->high)
2243 memcg_memory_event(memcg, MEMCG_HIGH);
2244 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2245 } while ((memcg = parent_mem_cgroup(memcg)));
2248 static void high_work_func(struct work_struct *work)
2250 struct mem_cgroup *memcg;
2252 memcg = container_of(work, struct mem_cgroup, high_work);
2253 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2257 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2258 * enough to still cause a significant slowdown in most cases, while still
2259 * allowing diagnostics and tracing to proceed without becoming stuck.
2261 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2264 * When calculating the delay, we use these either side of the exponentiation to
2265 * maintain precision and scale to a reasonable number of jiffies (see the table
2268 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2269 * overage ratio to a delay.
2270 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down down the
2271 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2272 * to produce a reasonable delay curve.
2274 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2275 * reasonable delay curve compared to precision-adjusted overage, not
2276 * penalising heavily at first, but still making sure that growth beyond the
2277 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2278 * example, with a high of 100 megabytes:
2280 * +-------+------------------------+
2281 * | usage | time to allocate in ms |
2282 * +-------+------------------------+
2304 * +-------+------------------------+
2306 #define MEMCG_DELAY_PRECISION_SHIFT 20
2307 #define MEMCG_DELAY_SCALING_SHIFT 14
2310 * Get the number of jiffies that we should penalise a mischievous cgroup which
2311 * is exceeding its memory.high by checking both it and its ancestors.
2313 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2314 unsigned int nr_pages)
2316 unsigned long penalty_jiffies;
2317 u64 max_overage = 0;
2320 unsigned long usage, high;
2323 usage = page_counter_read(&memcg->memory);
2324 high = READ_ONCE(memcg->high);
2327 * Prevent division by 0 in overage calculation by acting as if
2328 * it was a threshold of 1 page
2330 high = max(high, 1UL);
2332 overage = usage - high;
2333 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2334 overage = div64_u64(overage, high);
2336 if (overage > max_overage)
2337 max_overage = overage;
2338 } while ((memcg = parent_mem_cgroup(memcg)) &&
2339 !mem_cgroup_is_root(memcg));
2345 * We use overage compared to memory.high to calculate the number of
2346 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2347 * fairly lenient on small overages, and increasingly harsh when the
2348 * memcg in question makes it clear that it has no intention of stopping
2349 * its crazy behaviour, so we exponentially increase the delay based on
2352 penalty_jiffies = max_overage * max_overage * HZ;
2353 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2354 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2357 * Factor in the task's own contribution to the overage, such that four
2358 * N-sized allocations are throttled approximately the same as one
2359 * 4N-sized allocation.
2361 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2362 * larger the current charge patch is than that.
2364 penalty_jiffies = penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2367 * Clamp the max delay per usermode return so as to still keep the
2368 * application moving forwards and also permit diagnostics, albeit
2371 return min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2375 * Scheduled by try_charge() to be executed from the userland return path
2376 * and reclaims memory over the high limit.
2378 void mem_cgroup_handle_over_high(void)
2380 unsigned long penalty_jiffies;
2381 unsigned long pflags;
2382 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2383 struct mem_cgroup *memcg;
2385 if (likely(!nr_pages))
2388 memcg = get_mem_cgroup_from_mm(current->mm);
2389 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2390 current->memcg_nr_pages_over_high = 0;
2393 * memory.high is breached and reclaim is unable to keep up. Throttle
2394 * allocators proactively to slow down excessive growth.
2396 penalty_jiffies = calculate_high_delay(memcg, nr_pages);
2399 * Don't sleep if the amount of jiffies this memcg owes us is so low
2400 * that it's not even worth doing, in an attempt to be nice to those who
2401 * go only a small amount over their memory.high value and maybe haven't
2402 * been aggressively reclaimed enough yet.
2404 if (penalty_jiffies <= HZ / 100)
2408 * If we exit early, we're guaranteed to die (since
2409 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2410 * need to account for any ill-begotten jiffies to pay them off later.
2412 psi_memstall_enter(&pflags);
2413 schedule_timeout_killable(penalty_jiffies);
2414 psi_memstall_leave(&pflags);
2417 css_put(&memcg->css);
2420 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2421 unsigned int nr_pages)
2423 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2424 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2425 struct mem_cgroup *mem_over_limit;
2426 struct page_counter *counter;
2427 unsigned long nr_reclaimed;
2428 bool may_swap = true;
2429 bool drained = false;
2430 enum oom_status oom_status;
2432 if (mem_cgroup_is_root(memcg))
2435 if (consume_stock(memcg, nr_pages))
2438 if (!do_memsw_account() ||
2439 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2440 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2442 if (do_memsw_account())
2443 page_counter_uncharge(&memcg->memsw, batch);
2444 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2446 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2450 if (batch > nr_pages) {
2456 * Memcg doesn't have a dedicated reserve for atomic
2457 * allocations. But like the global atomic pool, we need to
2458 * put the burden of reclaim on regular allocation requests
2459 * and let these go through as privileged allocations.
2461 if (gfp_mask & __GFP_ATOMIC)
2465 * Unlike in global OOM situations, memcg is not in a physical
2466 * memory shortage. Allow dying and OOM-killed tasks to
2467 * bypass the last charges so that they can exit quickly and
2468 * free their memory.
2470 if (unlikely(should_force_charge()))
2474 * Prevent unbounded recursion when reclaim operations need to
2475 * allocate memory. This might exceed the limits temporarily,
2476 * but we prefer facilitating memory reclaim and getting back
2477 * under the limit over triggering OOM kills in these cases.
2479 if (unlikely(current->flags & PF_MEMALLOC))
2482 if (unlikely(task_in_memcg_oom(current)))
2485 if (!gfpflags_allow_blocking(gfp_mask))
2488 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2490 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2491 gfp_mask, may_swap);
2493 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2497 drain_all_stock(mem_over_limit);
2502 if (gfp_mask & __GFP_NORETRY)
2505 * Even though the limit is exceeded at this point, reclaim
2506 * may have been able to free some pages. Retry the charge
2507 * before killing the task.
2509 * Only for regular pages, though: huge pages are rather
2510 * unlikely to succeed so close to the limit, and we fall back
2511 * to regular pages anyway in case of failure.
2513 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2516 * At task move, charge accounts can be doubly counted. So, it's
2517 * better to wait until the end of task_move if something is going on.
2519 if (mem_cgroup_wait_acct_move(mem_over_limit))
2525 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2528 if (gfp_mask & __GFP_NOFAIL)
2531 if (fatal_signal_pending(current))
2535 * keep retrying as long as the memcg oom killer is able to make
2536 * a forward progress or bypass the charge if the oom killer
2537 * couldn't make any progress.
2539 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2540 get_order(nr_pages * PAGE_SIZE));
2541 switch (oom_status) {
2543 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2551 if (!(gfp_mask & __GFP_NOFAIL))
2555 * The allocation either can't fail or will lead to more memory
2556 * being freed very soon. Allow memory usage go over the limit
2557 * temporarily by force charging it.
2559 page_counter_charge(&memcg->memory, nr_pages);
2560 if (do_memsw_account())
2561 page_counter_charge(&memcg->memsw, nr_pages);
2562 css_get_many(&memcg->css, nr_pages);
2567 css_get_many(&memcg->css, batch);
2568 if (batch > nr_pages)
2569 refill_stock(memcg, batch - nr_pages);
2572 * If the hierarchy is above the normal consumption range, schedule
2573 * reclaim on returning to userland. We can perform reclaim here
2574 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2575 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2576 * not recorded as it most likely matches current's and won't
2577 * change in the meantime. As high limit is checked again before
2578 * reclaim, the cost of mismatch is negligible.
2581 if (page_counter_read(&memcg->memory) > memcg->high) {
2582 /* Don't bother a random interrupted task */
2583 if (in_interrupt()) {
2584 schedule_work(&memcg->high_work);
2587 current->memcg_nr_pages_over_high += batch;
2588 set_notify_resume(current);
2591 } while ((memcg = parent_mem_cgroup(memcg)));
2596 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2598 if (mem_cgroup_is_root(memcg))
2601 page_counter_uncharge(&memcg->memory, nr_pages);
2602 if (do_memsw_account())
2603 page_counter_uncharge(&memcg->memsw, nr_pages);
2605 css_put_many(&memcg->css, nr_pages);
2608 static void lock_page_lru(struct page *page, int *isolated)
2610 pg_data_t *pgdat = page_pgdat(page);
2612 spin_lock_irq(&pgdat->lru_lock);
2613 if (PageLRU(page)) {
2614 struct lruvec *lruvec;
2616 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2618 del_page_from_lru_list(page, lruvec, page_lru(page));
2624 static void unlock_page_lru(struct page *page, int isolated)
2626 pg_data_t *pgdat = page_pgdat(page);
2629 struct lruvec *lruvec;
2631 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2632 VM_BUG_ON_PAGE(PageLRU(page), page);
2634 add_page_to_lru_list(page, lruvec, page_lru(page));
2636 spin_unlock_irq(&pgdat->lru_lock);
2639 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2644 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2647 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2648 * may already be on some other mem_cgroup's LRU. Take care of it.
2651 lock_page_lru(page, &isolated);
2654 * Nobody should be changing or seriously looking at
2655 * page->mem_cgroup at this point:
2657 * - the page is uncharged
2659 * - the page is off-LRU
2661 * - an anonymous fault has exclusive page access, except for
2662 * a locked page table
2664 * - a page cache insertion, a swapin fault, or a migration
2665 * have the page locked
2667 page->mem_cgroup = memcg;
2670 unlock_page_lru(page, isolated);
2673 #ifdef CONFIG_MEMCG_KMEM
2675 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2677 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2678 * cgroup_mutex, etc.
2680 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2684 if (mem_cgroup_disabled())
2687 page = virt_to_head_page(p);
2690 * Slab pages don't have page->mem_cgroup set because corresponding
2691 * kmem caches can be reparented during the lifetime. That's why
2692 * memcg_from_slab_page() should be used instead.
2695 return memcg_from_slab_page(page);
2697 /* All other pages use page->mem_cgroup */
2698 return page->mem_cgroup;
2701 static int memcg_alloc_cache_id(void)
2706 id = ida_simple_get(&memcg_cache_ida,
2707 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2711 if (id < memcg_nr_cache_ids)
2715 * There's no space for the new id in memcg_caches arrays,
2716 * so we have to grow them.
2718 down_write(&memcg_cache_ids_sem);
2720 size = 2 * (id + 1);
2721 if (size < MEMCG_CACHES_MIN_SIZE)
2722 size = MEMCG_CACHES_MIN_SIZE;
2723 else if (size > MEMCG_CACHES_MAX_SIZE)
2724 size = MEMCG_CACHES_MAX_SIZE;
2726 err = memcg_update_all_caches(size);
2728 err = memcg_update_all_list_lrus(size);
2730 memcg_nr_cache_ids = size;
2732 up_write(&memcg_cache_ids_sem);
2735 ida_simple_remove(&memcg_cache_ida, id);
2741 static void memcg_free_cache_id(int id)
2743 ida_simple_remove(&memcg_cache_ida, id);
2746 struct memcg_kmem_cache_create_work {
2747 struct mem_cgroup *memcg;
2748 struct kmem_cache *cachep;
2749 struct work_struct work;
2752 static void memcg_kmem_cache_create_func(struct work_struct *w)
2754 struct memcg_kmem_cache_create_work *cw =
2755 container_of(w, struct memcg_kmem_cache_create_work, work);
2756 struct mem_cgroup *memcg = cw->memcg;
2757 struct kmem_cache *cachep = cw->cachep;
2759 memcg_create_kmem_cache(memcg, cachep);
2761 css_put(&memcg->css);
2766 * Enqueue the creation of a per-memcg kmem_cache.
2768 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2769 struct kmem_cache *cachep)
2771 struct memcg_kmem_cache_create_work *cw;
2773 if (!css_tryget_online(&memcg->css))
2776 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2781 cw->cachep = cachep;
2782 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2784 queue_work(memcg_kmem_cache_wq, &cw->work);
2787 static inline bool memcg_kmem_bypass(void)
2789 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2795 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2796 * @cachep: the original global kmem cache
2798 * Return the kmem_cache we're supposed to use for a slab allocation.
2799 * We try to use the current memcg's version of the cache.
2801 * If the cache does not exist yet, if we are the first user of it, we
2802 * create it asynchronously in a workqueue and let the current allocation
2803 * go through with the original cache.
2805 * This function takes a reference to the cache it returns to assure it
2806 * won't get destroyed while we are working with it. Once the caller is
2807 * done with it, memcg_kmem_put_cache() must be called to release the
2810 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2812 struct mem_cgroup *memcg;
2813 struct kmem_cache *memcg_cachep;
2814 struct memcg_cache_array *arr;
2817 VM_BUG_ON(!is_root_cache(cachep));
2819 if (memcg_kmem_bypass())
2824 if (unlikely(current->active_memcg))
2825 memcg = current->active_memcg;
2827 memcg = mem_cgroup_from_task(current);
2829 if (!memcg || memcg == root_mem_cgroup)
2832 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2836 arr = rcu_dereference(cachep->memcg_params.memcg_caches);
2839 * Make sure we will access the up-to-date value. The code updating
2840 * memcg_caches issues a write barrier to match the data dependency
2841 * barrier inside READ_ONCE() (see memcg_create_kmem_cache()).
2843 memcg_cachep = READ_ONCE(arr->entries[kmemcg_id]);
2846 * If we are in a safe context (can wait, and not in interrupt
2847 * context), we could be be predictable and return right away.
2848 * This would guarantee that the allocation being performed
2849 * already belongs in the new cache.
2851 * However, there are some clashes that can arrive from locking.
2852 * For instance, because we acquire the slab_mutex while doing
2853 * memcg_create_kmem_cache, this means no further allocation
2854 * could happen with the slab_mutex held. So it's better to
2857 * If the memcg is dying or memcg_cache is about to be released,
2858 * don't bother creating new kmem_caches. Because memcg_cachep
2859 * is ZEROed as the fist step of kmem offlining, we don't need
2860 * percpu_ref_tryget_live() here. css_tryget_online() check in
2861 * memcg_schedule_kmem_cache_create() will prevent us from
2862 * creation of a new kmem_cache.
2864 if (unlikely(!memcg_cachep))
2865 memcg_schedule_kmem_cache_create(memcg, cachep);
2866 else if (percpu_ref_tryget(&memcg_cachep->memcg_params.refcnt))
2867 cachep = memcg_cachep;
2874 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2875 * @cachep: the cache returned by memcg_kmem_get_cache
2877 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2879 if (!is_root_cache(cachep))
2880 percpu_ref_put(&cachep->memcg_params.refcnt);
2884 * __memcg_kmem_charge_memcg: charge a kmem page
2885 * @memcg: memory cgroup to charge
2886 * @gfp: reclaim mode
2887 * @order: allocation order
2889 * Returns 0 on success, an error code on failure.
2891 int __memcg_kmem_charge_memcg(struct mem_cgroup *memcg, gfp_t gfp, int order)
2893 unsigned int nr_pages = 1 << order;
2894 struct page_counter *counter;
2897 ret = try_charge(memcg, gfp, nr_pages);
2901 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2902 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2905 * Enforce __GFP_NOFAIL allocation because callers are not
2906 * prepared to see failures and likely do not have any failure
2909 if (gfp & __GFP_NOFAIL) {
2910 page_counter_charge(&memcg->kmem, nr_pages);
2913 cancel_charge(memcg, nr_pages);
2920 * __memcg_kmem_charge: charge a kmem page to the current memory cgroup
2921 * @page: page to charge
2922 * @gfp: reclaim mode
2923 * @order: allocation order
2925 * Returns 0 on success, an error code on failure.
2927 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2929 struct mem_cgroup *memcg;
2932 if (memcg_kmem_bypass())
2935 memcg = get_mem_cgroup_from_current();
2936 if (!mem_cgroup_is_root(memcg)) {
2937 ret = __memcg_kmem_charge_memcg(memcg, gfp, order);
2939 page->mem_cgroup = memcg;
2940 __SetPageKmemcg(page);
2943 css_put(&memcg->css);
2948 * __memcg_kmem_uncharge_memcg: uncharge a kmem page
2949 * @memcg: memcg to uncharge
2950 * @nr_pages: number of pages to uncharge
2952 void __memcg_kmem_uncharge_memcg(struct mem_cgroup *memcg,
2953 unsigned int nr_pages)
2955 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2956 page_counter_uncharge(&memcg->kmem, nr_pages);
2958 page_counter_uncharge(&memcg->memory, nr_pages);
2959 if (do_memsw_account())
2960 page_counter_uncharge(&memcg->memsw, nr_pages);
2963 * __memcg_kmem_uncharge: uncharge a kmem page
2964 * @page: page to uncharge
2965 * @order: allocation order
2967 void __memcg_kmem_uncharge(struct page *page, int order)
2969 struct mem_cgroup *memcg = page->mem_cgroup;
2970 unsigned int nr_pages = 1 << order;
2975 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2976 __memcg_kmem_uncharge_memcg(memcg, nr_pages);
2977 page->mem_cgroup = NULL;
2979 /* slab pages do not have PageKmemcg flag set */
2980 if (PageKmemcg(page))
2981 __ClearPageKmemcg(page);
2983 css_put_many(&memcg->css, nr_pages);
2985 #endif /* CONFIG_MEMCG_KMEM */
2987 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2990 * Because tail pages are not marked as "used", set it. We're under
2991 * pgdat->lru_lock and migration entries setup in all page mappings.
2993 void mem_cgroup_split_huge_fixup(struct page *head)
2997 if (mem_cgroup_disabled())
3000 for (i = 1; i < HPAGE_PMD_NR; i++)
3001 head[i].mem_cgroup = head->mem_cgroup;
3003 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
3005 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3007 #ifdef CONFIG_MEMCG_SWAP
3009 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3010 * @entry: swap entry to be moved
3011 * @from: mem_cgroup which the entry is moved from
3012 * @to: mem_cgroup which the entry is moved to
3014 * It succeeds only when the swap_cgroup's record for this entry is the same
3015 * as the mem_cgroup's id of @from.
3017 * Returns 0 on success, -EINVAL on failure.
3019 * The caller must have charged to @to, IOW, called page_counter_charge() about
3020 * both res and memsw, and called css_get().
3022 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3023 struct mem_cgroup *from, struct mem_cgroup *to)
3025 unsigned short old_id, new_id;
3027 old_id = mem_cgroup_id(from);
3028 new_id = mem_cgroup_id(to);
3030 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3031 mod_memcg_state(from, MEMCG_SWAP, -1);
3032 mod_memcg_state(to, MEMCG_SWAP, 1);
3038 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3039 struct mem_cgroup *from, struct mem_cgroup *to)
3045 static DEFINE_MUTEX(memcg_max_mutex);
3047 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3048 unsigned long max, bool memsw)
3050 bool enlarge = false;
3051 bool drained = false;
3053 bool limits_invariant;
3054 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3057 if (signal_pending(current)) {
3062 mutex_lock(&memcg_max_mutex);
3064 * Make sure that the new limit (memsw or memory limit) doesn't
3065 * break our basic invariant rule memory.max <= memsw.max.
3067 limits_invariant = memsw ? max >= memcg->memory.max :
3068 max <= memcg->memsw.max;
3069 if (!limits_invariant) {
3070 mutex_unlock(&memcg_max_mutex);
3074 if (max > counter->max)
3076 ret = page_counter_set_max(counter, max);
3077 mutex_unlock(&memcg_max_mutex);
3083 drain_all_stock(memcg);
3088 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3089 GFP_KERNEL, !memsw)) {
3095 if (!ret && enlarge)
3096 memcg_oom_recover(memcg);
3101 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3103 unsigned long *total_scanned)
3105 unsigned long nr_reclaimed = 0;
3106 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3107 unsigned long reclaimed;
3109 struct mem_cgroup_tree_per_node *mctz;
3110 unsigned long excess;
3111 unsigned long nr_scanned;
3116 mctz = soft_limit_tree_node(pgdat->node_id);
3119 * Do not even bother to check the largest node if the root
3120 * is empty. Do it lockless to prevent lock bouncing. Races
3121 * are acceptable as soft limit is best effort anyway.
3123 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3127 * This loop can run a while, specially if mem_cgroup's continuously
3128 * keep exceeding their soft limit and putting the system under
3135 mz = mem_cgroup_largest_soft_limit_node(mctz);
3140 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3141 gfp_mask, &nr_scanned);
3142 nr_reclaimed += reclaimed;
3143 *total_scanned += nr_scanned;
3144 spin_lock_irq(&mctz->lock);
3145 __mem_cgroup_remove_exceeded(mz, mctz);
3148 * If we failed to reclaim anything from this memory cgroup
3149 * it is time to move on to the next cgroup
3153 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3155 excess = soft_limit_excess(mz->memcg);
3157 * One school of thought says that we should not add
3158 * back the node to the tree if reclaim returns 0.
3159 * But our reclaim could return 0, simply because due
3160 * to priority we are exposing a smaller subset of
3161 * memory to reclaim from. Consider this as a longer
3164 /* If excess == 0, no tree ops */
3165 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3166 spin_unlock_irq(&mctz->lock);
3167 css_put(&mz->memcg->css);
3170 * Could not reclaim anything and there are no more
3171 * mem cgroups to try or we seem to be looping without
3172 * reclaiming anything.
3174 if (!nr_reclaimed &&
3176 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3178 } while (!nr_reclaimed);
3180 css_put(&next_mz->memcg->css);
3181 return nr_reclaimed;
3185 * Test whether @memcg has children, dead or alive. Note that this
3186 * function doesn't care whether @memcg has use_hierarchy enabled and
3187 * returns %true if there are child csses according to the cgroup
3188 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3190 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3195 ret = css_next_child(NULL, &memcg->css);
3201 * Reclaims as many pages from the given memcg as possible.
3203 * Caller is responsible for holding css reference for memcg.
3205 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3207 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3209 /* we call try-to-free pages for make this cgroup empty */
3210 lru_add_drain_all();
3212 drain_all_stock(memcg);
3214 /* try to free all pages in this cgroup */
3215 while (nr_retries && page_counter_read(&memcg->memory)) {
3218 if (signal_pending(current))
3221 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3225 /* maybe some writeback is necessary */
3226 congestion_wait(BLK_RW_ASYNC, HZ/10);
3234 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3235 char *buf, size_t nbytes,
3238 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3240 if (mem_cgroup_is_root(memcg))
3242 return mem_cgroup_force_empty(memcg) ?: nbytes;
3245 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3248 return mem_cgroup_from_css(css)->use_hierarchy;
3251 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3252 struct cftype *cft, u64 val)
3255 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3256 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3258 if (memcg->use_hierarchy == val)
3262 * If parent's use_hierarchy is set, we can't make any modifications
3263 * in the child subtrees. If it is unset, then the change can
3264 * occur, provided the current cgroup has no children.
3266 * For the root cgroup, parent_mem is NULL, we allow value to be
3267 * set if there are no children.
3269 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3270 (val == 1 || val == 0)) {
3271 if (!memcg_has_children(memcg))
3272 memcg->use_hierarchy = val;
3281 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3285 if (mem_cgroup_is_root(memcg)) {
3286 val = memcg_page_state(memcg, MEMCG_CACHE) +
3287 memcg_page_state(memcg, MEMCG_RSS);
3289 val += memcg_page_state(memcg, MEMCG_SWAP);
3292 val = page_counter_read(&memcg->memory);
3294 val = page_counter_read(&memcg->memsw);
3307 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3310 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3311 struct page_counter *counter;
3313 switch (MEMFILE_TYPE(cft->private)) {
3315 counter = &memcg->memory;
3318 counter = &memcg->memsw;
3321 counter = &memcg->kmem;
3324 counter = &memcg->tcpmem;
3330 switch (MEMFILE_ATTR(cft->private)) {
3332 if (counter == &memcg->memory)
3333 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3334 if (counter == &memcg->memsw)
3335 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3336 return (u64)page_counter_read(counter) * PAGE_SIZE;
3338 return (u64)counter->max * PAGE_SIZE;
3340 return (u64)counter->watermark * PAGE_SIZE;
3342 return counter->failcnt;
3343 case RES_SOFT_LIMIT:
3344 return (u64)memcg->soft_limit * PAGE_SIZE;
3350 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3352 unsigned long stat[MEMCG_NR_STAT] = {0};
3353 struct mem_cgroup *mi;
3356 for_each_online_cpu(cpu)
3357 for (i = 0; i < MEMCG_NR_STAT; i++)
3358 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3360 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3361 for (i = 0; i < MEMCG_NR_STAT; i++)
3362 atomic_long_add(stat[i], &mi->vmstats[i]);
3364 for_each_node(node) {
3365 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3366 struct mem_cgroup_per_node *pi;
3368 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3371 for_each_online_cpu(cpu)
3372 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3374 pn->lruvec_stat_cpu->count[i], cpu);
3376 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3377 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3378 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3382 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3384 unsigned long events[NR_VM_EVENT_ITEMS];
3385 struct mem_cgroup *mi;
3388 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3391 for_each_online_cpu(cpu)
3392 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3393 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3396 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3397 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3398 atomic_long_add(events[i], &mi->vmevents[i]);
3401 #ifdef CONFIG_MEMCG_KMEM
3402 static int memcg_online_kmem(struct mem_cgroup *memcg)
3406 if (cgroup_memory_nokmem)
3409 BUG_ON(memcg->kmemcg_id >= 0);
3410 BUG_ON(memcg->kmem_state);
3412 memcg_id = memcg_alloc_cache_id();
3416 static_branch_inc(&memcg_kmem_enabled_key);
3418 * A memory cgroup is considered kmem-online as soon as it gets
3419 * kmemcg_id. Setting the id after enabling static branching will
3420 * guarantee no one starts accounting before all call sites are
3423 memcg->kmemcg_id = memcg_id;
3424 memcg->kmem_state = KMEM_ONLINE;
3425 INIT_LIST_HEAD(&memcg->kmem_caches);
3430 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3432 struct cgroup_subsys_state *css;
3433 struct mem_cgroup *parent, *child;
3436 if (memcg->kmem_state != KMEM_ONLINE)
3439 * Clear the online state before clearing memcg_caches array
3440 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3441 * guarantees that no cache will be created for this cgroup
3442 * after we are done (see memcg_create_kmem_cache()).
3444 memcg->kmem_state = KMEM_ALLOCATED;
3446 parent = parent_mem_cgroup(memcg);
3448 parent = root_mem_cgroup;
3451 * Deactivate and reparent kmem_caches.
3453 memcg_deactivate_kmem_caches(memcg, parent);
3455 kmemcg_id = memcg->kmemcg_id;
3456 BUG_ON(kmemcg_id < 0);
3459 * Change kmemcg_id of this cgroup and all its descendants to the
3460 * parent's id, and then move all entries from this cgroup's list_lrus
3461 * to ones of the parent. After we have finished, all list_lrus
3462 * corresponding to this cgroup are guaranteed to remain empty. The
3463 * ordering is imposed by list_lru_node->lock taken by
3464 * memcg_drain_all_list_lrus().
3466 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3467 css_for_each_descendant_pre(css, &memcg->css) {
3468 child = mem_cgroup_from_css(css);
3469 BUG_ON(child->kmemcg_id != kmemcg_id);
3470 child->kmemcg_id = parent->kmemcg_id;
3471 if (!memcg->use_hierarchy)
3476 memcg_drain_all_list_lrus(kmemcg_id, parent);
3478 memcg_free_cache_id(kmemcg_id);
3481 static void memcg_free_kmem(struct mem_cgroup *memcg)
3483 /* css_alloc() failed, offlining didn't happen */
3484 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3485 memcg_offline_kmem(memcg);
3487 if (memcg->kmem_state == KMEM_ALLOCATED) {
3488 WARN_ON(!list_empty(&memcg->kmem_caches));
3489 static_branch_dec(&memcg_kmem_enabled_key);
3493 static int memcg_online_kmem(struct mem_cgroup *memcg)
3497 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3500 static void memcg_free_kmem(struct mem_cgroup *memcg)
3503 #endif /* CONFIG_MEMCG_KMEM */
3505 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3510 mutex_lock(&memcg_max_mutex);
3511 ret = page_counter_set_max(&memcg->kmem, max);
3512 mutex_unlock(&memcg_max_mutex);
3516 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3520 mutex_lock(&memcg_max_mutex);
3522 ret = page_counter_set_max(&memcg->tcpmem, max);
3526 if (!memcg->tcpmem_active) {
3528 * The active flag needs to be written after the static_key
3529 * update. This is what guarantees that the socket activation
3530 * function is the last one to run. See mem_cgroup_sk_alloc()
3531 * for details, and note that we don't mark any socket as
3532 * belonging to this memcg until that flag is up.
3534 * We need to do this, because static_keys will span multiple
3535 * sites, but we can't control their order. If we mark a socket
3536 * as accounted, but the accounting functions are not patched in
3537 * yet, we'll lose accounting.
3539 * We never race with the readers in mem_cgroup_sk_alloc(),
3540 * because when this value change, the code to process it is not
3543 static_branch_inc(&memcg_sockets_enabled_key);
3544 memcg->tcpmem_active = true;
3547 mutex_unlock(&memcg_max_mutex);
3552 * The user of this function is...
3555 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3556 char *buf, size_t nbytes, loff_t off)
3558 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3559 unsigned long nr_pages;
3562 buf = strstrip(buf);
3563 ret = page_counter_memparse(buf, "-1", &nr_pages);
3567 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3569 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3573 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3575 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3578 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3581 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3582 "Please report your usecase to linux-mm@kvack.org if you "
3583 "depend on this functionality.\n");
3584 ret = memcg_update_kmem_max(memcg, nr_pages);
3587 ret = memcg_update_tcp_max(memcg, nr_pages);
3591 case RES_SOFT_LIMIT:
3592 memcg->soft_limit = nr_pages;
3596 return ret ?: nbytes;
3599 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3600 size_t nbytes, loff_t off)
3602 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3603 struct page_counter *counter;
3605 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3607 counter = &memcg->memory;
3610 counter = &memcg->memsw;
3613 counter = &memcg->kmem;
3616 counter = &memcg->tcpmem;
3622 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3624 page_counter_reset_watermark(counter);
3627 counter->failcnt = 0;
3636 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3639 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3643 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3644 struct cftype *cft, u64 val)
3646 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3648 if (val & ~MOVE_MASK)
3652 * No kind of locking is needed in here, because ->can_attach() will
3653 * check this value once in the beginning of the process, and then carry
3654 * on with stale data. This means that changes to this value will only
3655 * affect task migrations starting after the change.
3657 memcg->move_charge_at_immigrate = val;
3661 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3662 struct cftype *cft, u64 val)
3670 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3671 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3672 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3674 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3675 int nid, unsigned int lru_mask)
3677 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3678 unsigned long nr = 0;
3681 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3684 if (!(BIT(lru) & lru_mask))
3686 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3691 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3692 unsigned int lru_mask)
3694 unsigned long nr = 0;
3698 if (!(BIT(lru) & lru_mask))
3700 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3705 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3709 unsigned int lru_mask;
3712 static const struct numa_stat stats[] = {
3713 { "total", LRU_ALL },
3714 { "file", LRU_ALL_FILE },
3715 { "anon", LRU_ALL_ANON },
3716 { "unevictable", BIT(LRU_UNEVICTABLE) },
3718 const struct numa_stat *stat;
3721 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3723 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3724 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3725 seq_printf(m, "%s=%lu", stat->name, nr);
3726 for_each_node_state(nid, N_MEMORY) {
3727 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3729 seq_printf(m, " N%d=%lu", nid, nr);
3734 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3735 struct mem_cgroup *iter;
3738 for_each_mem_cgroup_tree(iter, memcg)
3739 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3740 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3741 for_each_node_state(nid, N_MEMORY) {
3743 for_each_mem_cgroup_tree(iter, memcg)
3744 nr += mem_cgroup_node_nr_lru_pages(
3745 iter, nid, stat->lru_mask);
3746 seq_printf(m, " N%d=%lu", nid, nr);
3753 #endif /* CONFIG_NUMA */
3755 static const unsigned int memcg1_stats[] = {
3766 static const char *const memcg1_stat_names[] = {
3777 /* Universal VM events cgroup1 shows, original sort order */
3778 static const unsigned int memcg1_events[] = {
3785 static int memcg_stat_show(struct seq_file *m, void *v)
3787 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3788 unsigned long memory, memsw;
3789 struct mem_cgroup *mi;
3792 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3794 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3795 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3797 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3798 memcg_page_state_local(memcg, memcg1_stats[i]) *
3802 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3803 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
3804 memcg_events_local(memcg, memcg1_events[i]));
3806 for (i = 0; i < NR_LRU_LISTS; i++)
3807 seq_printf(m, "%s %lu\n", lru_list_name(i),
3808 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3811 /* Hierarchical information */
3812 memory = memsw = PAGE_COUNTER_MAX;
3813 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3814 memory = min(memory, mi->memory.max);
3815 memsw = min(memsw, mi->memsw.max);
3817 seq_printf(m, "hierarchical_memory_limit %llu\n",
3818 (u64)memory * PAGE_SIZE);
3819 if (do_memsw_account())
3820 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3821 (u64)memsw * PAGE_SIZE);
3823 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3824 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3826 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3827 (u64)memcg_page_state(memcg, memcg1_stats[i]) *
3831 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3832 seq_printf(m, "total_%s %llu\n",
3833 vm_event_name(memcg1_events[i]),
3834 (u64)memcg_events(memcg, memcg1_events[i]));
3836 for (i = 0; i < NR_LRU_LISTS; i++)
3837 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
3838 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
3841 #ifdef CONFIG_DEBUG_VM
3844 struct mem_cgroup_per_node *mz;
3845 struct zone_reclaim_stat *rstat;
3846 unsigned long recent_rotated[2] = {0, 0};
3847 unsigned long recent_scanned[2] = {0, 0};
3849 for_each_online_pgdat(pgdat) {
3850 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3851 rstat = &mz->lruvec.reclaim_stat;
3853 recent_rotated[0] += rstat->recent_rotated[0];
3854 recent_rotated[1] += rstat->recent_rotated[1];
3855 recent_scanned[0] += rstat->recent_scanned[0];
3856 recent_scanned[1] += rstat->recent_scanned[1];
3858 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3859 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3860 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3861 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3868 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3871 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3873 return mem_cgroup_swappiness(memcg);
3876 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3877 struct cftype *cft, u64 val)
3879 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3885 memcg->swappiness = val;
3887 vm_swappiness = val;
3892 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3894 struct mem_cgroup_threshold_ary *t;
3895 unsigned long usage;
3900 t = rcu_dereference(memcg->thresholds.primary);
3902 t = rcu_dereference(memcg->memsw_thresholds.primary);
3907 usage = mem_cgroup_usage(memcg, swap);
3910 * current_threshold points to threshold just below or equal to usage.
3911 * If it's not true, a threshold was crossed after last
3912 * call of __mem_cgroup_threshold().
3914 i = t->current_threshold;
3917 * Iterate backward over array of thresholds starting from
3918 * current_threshold and check if a threshold is crossed.
3919 * If none of thresholds below usage is crossed, we read
3920 * only one element of the array here.
3922 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3923 eventfd_signal(t->entries[i].eventfd, 1);
3925 /* i = current_threshold + 1 */
3929 * Iterate forward over array of thresholds starting from
3930 * current_threshold+1 and check if a threshold is crossed.
3931 * If none of thresholds above usage is crossed, we read
3932 * only one element of the array here.
3934 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3935 eventfd_signal(t->entries[i].eventfd, 1);
3937 /* Update current_threshold */
3938 t->current_threshold = i - 1;
3943 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3946 __mem_cgroup_threshold(memcg, false);
3947 if (do_memsw_account())
3948 __mem_cgroup_threshold(memcg, true);
3950 memcg = parent_mem_cgroup(memcg);
3954 static int compare_thresholds(const void *a, const void *b)
3956 const struct mem_cgroup_threshold *_a = a;
3957 const struct mem_cgroup_threshold *_b = b;
3959 if (_a->threshold > _b->threshold)
3962 if (_a->threshold < _b->threshold)
3968 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3970 struct mem_cgroup_eventfd_list *ev;
3972 spin_lock(&memcg_oom_lock);
3974 list_for_each_entry(ev, &memcg->oom_notify, list)
3975 eventfd_signal(ev->eventfd, 1);
3977 spin_unlock(&memcg_oom_lock);
3981 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3983 struct mem_cgroup *iter;
3985 for_each_mem_cgroup_tree(iter, memcg)
3986 mem_cgroup_oom_notify_cb(iter);
3989 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3990 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3992 struct mem_cgroup_thresholds *thresholds;
3993 struct mem_cgroup_threshold_ary *new;
3994 unsigned long threshold;
3995 unsigned long usage;
3998 ret = page_counter_memparse(args, "-1", &threshold);
4002 mutex_lock(&memcg->thresholds_lock);
4005 thresholds = &memcg->thresholds;
4006 usage = mem_cgroup_usage(memcg, false);
4007 } else if (type == _MEMSWAP) {
4008 thresholds = &memcg->memsw_thresholds;
4009 usage = mem_cgroup_usage(memcg, true);
4013 /* Check if a threshold crossed before adding a new one */
4014 if (thresholds->primary)
4015 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4017 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4019 /* Allocate memory for new array of thresholds */
4020 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4027 /* Copy thresholds (if any) to new array */
4028 if (thresholds->primary) {
4029 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4030 sizeof(struct mem_cgroup_threshold));
4033 /* Add new threshold */
4034 new->entries[size - 1].eventfd = eventfd;
4035 new->entries[size - 1].threshold = threshold;
4037 /* Sort thresholds. Registering of new threshold isn't time-critical */
4038 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4039 compare_thresholds, NULL);
4041 /* Find current threshold */
4042 new->current_threshold = -1;
4043 for (i = 0; i < size; i++) {
4044 if (new->entries[i].threshold <= usage) {
4046 * new->current_threshold will not be used until
4047 * rcu_assign_pointer(), so it's safe to increment
4050 ++new->current_threshold;
4055 /* Free old spare buffer and save old primary buffer as spare */
4056 kfree(thresholds->spare);
4057 thresholds->spare = thresholds->primary;
4059 rcu_assign_pointer(thresholds->primary, new);
4061 /* To be sure that nobody uses thresholds */
4065 mutex_unlock(&memcg->thresholds_lock);
4070 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4071 struct eventfd_ctx *eventfd, const char *args)
4073 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4076 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4077 struct eventfd_ctx *eventfd, const char *args)
4079 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4082 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4083 struct eventfd_ctx *eventfd, enum res_type type)
4085 struct mem_cgroup_thresholds *thresholds;
4086 struct mem_cgroup_threshold_ary *new;
4087 unsigned long usage;
4088 int i, j, size, entries;
4090 mutex_lock(&memcg->thresholds_lock);
4093 thresholds = &memcg->thresholds;
4094 usage = mem_cgroup_usage(memcg, false);
4095 } else if (type == _MEMSWAP) {
4096 thresholds = &memcg->memsw_thresholds;
4097 usage = mem_cgroup_usage(memcg, true);
4101 if (!thresholds->primary)
4104 /* Check if a threshold crossed before removing */
4105 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4107 /* Calculate new number of threshold */
4109 for (i = 0; i < thresholds->primary->size; i++) {
4110 if (thresholds->primary->entries[i].eventfd != eventfd)
4116 new = thresholds->spare;
4118 /* If no items related to eventfd have been cleared, nothing to do */
4122 /* Set thresholds array to NULL if we don't have thresholds */
4131 /* Copy thresholds and find current threshold */
4132 new->current_threshold = -1;
4133 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4134 if (thresholds->primary->entries[i].eventfd == eventfd)
4137 new->entries[j] = thresholds->primary->entries[i];
4138 if (new->entries[j].threshold <= usage) {
4140 * new->current_threshold will not be used
4141 * until rcu_assign_pointer(), so it's safe to increment
4144 ++new->current_threshold;
4150 /* Swap primary and spare array */
4151 thresholds->spare = thresholds->primary;
4153 rcu_assign_pointer(thresholds->primary, new);
4155 /* To be sure that nobody uses thresholds */
4158 /* If all events are unregistered, free the spare array */
4160 kfree(thresholds->spare);
4161 thresholds->spare = NULL;
4164 mutex_unlock(&memcg->thresholds_lock);
4167 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4168 struct eventfd_ctx *eventfd)
4170 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4173 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4174 struct eventfd_ctx *eventfd)
4176 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4179 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4180 struct eventfd_ctx *eventfd, const char *args)
4182 struct mem_cgroup_eventfd_list *event;
4184 event = kmalloc(sizeof(*event), GFP_KERNEL);
4188 spin_lock(&memcg_oom_lock);
4190 event->eventfd = eventfd;
4191 list_add(&event->list, &memcg->oom_notify);
4193 /* already in OOM ? */
4194 if (memcg->under_oom)
4195 eventfd_signal(eventfd, 1);
4196 spin_unlock(&memcg_oom_lock);
4201 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4202 struct eventfd_ctx *eventfd)
4204 struct mem_cgroup_eventfd_list *ev, *tmp;
4206 spin_lock(&memcg_oom_lock);
4208 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4209 if (ev->eventfd == eventfd) {
4210 list_del(&ev->list);
4215 spin_unlock(&memcg_oom_lock);
4218 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4220 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4222 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4223 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4224 seq_printf(sf, "oom_kill %lu\n",
4225 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4229 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4230 struct cftype *cft, u64 val)
4232 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4234 /* cannot set to root cgroup and only 0 and 1 are allowed */
4235 if (!css->parent || !((val == 0) || (val == 1)))
4238 memcg->oom_kill_disable = val;
4240 memcg_oom_recover(memcg);
4245 #ifdef CONFIG_CGROUP_WRITEBACK
4247 #include <trace/events/writeback.h>
4249 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4251 return wb_domain_init(&memcg->cgwb_domain, gfp);
4254 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4256 wb_domain_exit(&memcg->cgwb_domain);
4259 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4261 wb_domain_size_changed(&memcg->cgwb_domain);
4264 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4266 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4268 if (!memcg->css.parent)
4271 return &memcg->cgwb_domain;
4275 * idx can be of type enum memcg_stat_item or node_stat_item.
4276 * Keep in sync with memcg_exact_page().
4278 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4280 long x = atomic_long_read(&memcg->vmstats[idx]);
4283 for_each_online_cpu(cpu)
4284 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4291 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4292 * @wb: bdi_writeback in question
4293 * @pfilepages: out parameter for number of file pages
4294 * @pheadroom: out parameter for number of allocatable pages according to memcg
4295 * @pdirty: out parameter for number of dirty pages
4296 * @pwriteback: out parameter for number of pages under writeback
4298 * Determine the numbers of file, headroom, dirty, and writeback pages in
4299 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4300 * is a bit more involved.
4302 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4303 * headroom is calculated as the lowest headroom of itself and the
4304 * ancestors. Note that this doesn't consider the actual amount of
4305 * available memory in the system. The caller should further cap
4306 * *@pheadroom accordingly.
4308 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4309 unsigned long *pheadroom, unsigned long *pdirty,
4310 unsigned long *pwriteback)
4312 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4313 struct mem_cgroup *parent;
4315 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4317 /* this should eventually include NR_UNSTABLE_NFS */
4318 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4319 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4320 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4321 *pheadroom = PAGE_COUNTER_MAX;
4323 while ((parent = parent_mem_cgroup(memcg))) {
4324 unsigned long ceiling = min(memcg->memory.max, memcg->high);
4325 unsigned long used = page_counter_read(&memcg->memory);
4327 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4333 * Foreign dirty flushing
4335 * There's an inherent mismatch between memcg and writeback. The former
4336 * trackes ownership per-page while the latter per-inode. This was a
4337 * deliberate design decision because honoring per-page ownership in the
4338 * writeback path is complicated, may lead to higher CPU and IO overheads
4339 * and deemed unnecessary given that write-sharing an inode across
4340 * different cgroups isn't a common use-case.
4342 * Combined with inode majority-writer ownership switching, this works well
4343 * enough in most cases but there are some pathological cases. For
4344 * example, let's say there are two cgroups A and B which keep writing to
4345 * different but confined parts of the same inode. B owns the inode and
4346 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4347 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4348 * triggering background writeback. A will be slowed down without a way to
4349 * make writeback of the dirty pages happen.
4351 * Conditions like the above can lead to a cgroup getting repatedly and
4352 * severely throttled after making some progress after each
4353 * dirty_expire_interval while the underyling IO device is almost
4356 * Solving this problem completely requires matching the ownership tracking
4357 * granularities between memcg and writeback in either direction. However,
4358 * the more egregious behaviors can be avoided by simply remembering the
4359 * most recent foreign dirtying events and initiating remote flushes on
4360 * them when local writeback isn't enough to keep the memory clean enough.
4362 * The following two functions implement such mechanism. When a foreign
4363 * page - a page whose memcg and writeback ownerships don't match - is
4364 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4365 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4366 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4367 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4368 * foreign bdi_writebacks which haven't expired. Both the numbers of
4369 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4370 * limited to MEMCG_CGWB_FRN_CNT.
4372 * The mechanism only remembers IDs and doesn't hold any object references.
4373 * As being wrong occasionally doesn't matter, updates and accesses to the
4374 * records are lockless and racy.
4376 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4377 struct bdi_writeback *wb)
4379 struct mem_cgroup *memcg = page->mem_cgroup;
4380 struct memcg_cgwb_frn *frn;
4381 u64 now = get_jiffies_64();
4382 u64 oldest_at = now;
4386 trace_track_foreign_dirty(page, wb);
4389 * Pick the slot to use. If there is already a slot for @wb, keep
4390 * using it. If not replace the oldest one which isn't being
4393 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4394 frn = &memcg->cgwb_frn[i];
4395 if (frn->bdi_id == wb->bdi->id &&
4396 frn->memcg_id == wb->memcg_css->id)
4398 if (time_before64(frn->at, oldest_at) &&
4399 atomic_read(&frn->done.cnt) == 1) {
4401 oldest_at = frn->at;
4405 if (i < MEMCG_CGWB_FRN_CNT) {
4407 * Re-using an existing one. Update timestamp lazily to
4408 * avoid making the cacheline hot. We want them to be
4409 * reasonably up-to-date and significantly shorter than
4410 * dirty_expire_interval as that's what expires the record.
4411 * Use the shorter of 1s and dirty_expire_interval / 8.
4413 unsigned long update_intv =
4414 min_t(unsigned long, HZ,
4415 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4417 if (time_before64(frn->at, now - update_intv))
4419 } else if (oldest >= 0) {
4420 /* replace the oldest free one */
4421 frn = &memcg->cgwb_frn[oldest];
4422 frn->bdi_id = wb->bdi->id;
4423 frn->memcg_id = wb->memcg_css->id;
4428 /* issue foreign writeback flushes for recorded foreign dirtying events */
4429 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4431 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4432 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4433 u64 now = jiffies_64;
4436 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4437 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4440 * If the record is older than dirty_expire_interval,
4441 * writeback on it has already started. No need to kick it
4442 * off again. Also, don't start a new one if there's
4443 * already one in flight.
4445 if (time_after64(frn->at, now - intv) &&
4446 atomic_read(&frn->done.cnt) == 1) {
4448 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4449 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4450 WB_REASON_FOREIGN_FLUSH,
4456 #else /* CONFIG_CGROUP_WRITEBACK */
4458 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4463 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4467 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4471 #endif /* CONFIG_CGROUP_WRITEBACK */
4474 * DO NOT USE IN NEW FILES.
4476 * "cgroup.event_control" implementation.
4478 * This is way over-engineered. It tries to support fully configurable
4479 * events for each user. Such level of flexibility is completely
4480 * unnecessary especially in the light of the planned unified hierarchy.
4482 * Please deprecate this and replace with something simpler if at all
4487 * Unregister event and free resources.
4489 * Gets called from workqueue.
4491 static void memcg_event_remove(struct work_struct *work)
4493 struct mem_cgroup_event *event =
4494 container_of(work, struct mem_cgroup_event, remove);
4495 struct mem_cgroup *memcg = event->memcg;
4497 remove_wait_queue(event->wqh, &event->wait);
4499 event->unregister_event(memcg, event->eventfd);
4501 /* Notify userspace the event is going away. */
4502 eventfd_signal(event->eventfd, 1);
4504 eventfd_ctx_put(event->eventfd);
4506 css_put(&memcg->css);
4510 * Gets called on EPOLLHUP on eventfd when user closes it.
4512 * Called with wqh->lock held and interrupts disabled.
4514 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4515 int sync, void *key)
4517 struct mem_cgroup_event *event =
4518 container_of(wait, struct mem_cgroup_event, wait);
4519 struct mem_cgroup *memcg = event->memcg;
4520 __poll_t flags = key_to_poll(key);
4522 if (flags & EPOLLHUP) {
4524 * If the event has been detached at cgroup removal, we
4525 * can simply return knowing the other side will cleanup
4528 * We can't race against event freeing since the other
4529 * side will require wqh->lock via remove_wait_queue(),
4532 spin_lock(&memcg->event_list_lock);
4533 if (!list_empty(&event->list)) {
4534 list_del_init(&event->list);
4536 * We are in atomic context, but cgroup_event_remove()
4537 * may sleep, so we have to call it in workqueue.
4539 schedule_work(&event->remove);
4541 spin_unlock(&memcg->event_list_lock);
4547 static void memcg_event_ptable_queue_proc(struct file *file,
4548 wait_queue_head_t *wqh, poll_table *pt)
4550 struct mem_cgroup_event *event =
4551 container_of(pt, struct mem_cgroup_event, pt);
4554 add_wait_queue(wqh, &event->wait);
4558 * DO NOT USE IN NEW FILES.
4560 * Parse input and register new cgroup event handler.
4562 * Input must be in format '<event_fd> <control_fd> <args>'.
4563 * Interpretation of args is defined by control file implementation.
4565 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4566 char *buf, size_t nbytes, loff_t off)
4568 struct cgroup_subsys_state *css = of_css(of);
4569 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4570 struct mem_cgroup_event *event;
4571 struct cgroup_subsys_state *cfile_css;
4572 unsigned int efd, cfd;
4579 buf = strstrip(buf);
4581 efd = simple_strtoul(buf, &endp, 10);
4586 cfd = simple_strtoul(buf, &endp, 10);
4587 if ((*endp != ' ') && (*endp != '\0'))
4591 event = kzalloc(sizeof(*event), GFP_KERNEL);
4595 event->memcg = memcg;
4596 INIT_LIST_HEAD(&event->list);
4597 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4598 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4599 INIT_WORK(&event->remove, memcg_event_remove);
4607 event->eventfd = eventfd_ctx_fileget(efile.file);
4608 if (IS_ERR(event->eventfd)) {
4609 ret = PTR_ERR(event->eventfd);
4616 goto out_put_eventfd;
4619 /* the process need read permission on control file */
4620 /* AV: shouldn't we check that it's been opened for read instead? */
4621 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4626 * Determine the event callbacks and set them in @event. This used
4627 * to be done via struct cftype but cgroup core no longer knows
4628 * about these events. The following is crude but the whole thing
4629 * is for compatibility anyway.
4631 * DO NOT ADD NEW FILES.
4633 name = cfile.file->f_path.dentry->d_name.name;
4635 if (!strcmp(name, "memory.usage_in_bytes")) {
4636 event->register_event = mem_cgroup_usage_register_event;
4637 event->unregister_event = mem_cgroup_usage_unregister_event;
4638 } else if (!strcmp(name, "memory.oom_control")) {
4639 event->register_event = mem_cgroup_oom_register_event;
4640 event->unregister_event = mem_cgroup_oom_unregister_event;
4641 } else if (!strcmp(name, "memory.pressure_level")) {
4642 event->register_event = vmpressure_register_event;
4643 event->unregister_event = vmpressure_unregister_event;
4644 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4645 event->register_event = memsw_cgroup_usage_register_event;
4646 event->unregister_event = memsw_cgroup_usage_unregister_event;
4653 * Verify @cfile should belong to @css. Also, remaining events are
4654 * automatically removed on cgroup destruction but the removal is
4655 * asynchronous, so take an extra ref on @css.
4657 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4658 &memory_cgrp_subsys);
4660 if (IS_ERR(cfile_css))
4662 if (cfile_css != css) {
4667 ret = event->register_event(memcg, event->eventfd, buf);
4671 vfs_poll(efile.file, &event->pt);
4673 spin_lock(&memcg->event_list_lock);
4674 list_add(&event->list, &memcg->event_list);
4675 spin_unlock(&memcg->event_list_lock);
4687 eventfd_ctx_put(event->eventfd);
4696 static struct cftype mem_cgroup_legacy_files[] = {
4698 .name = "usage_in_bytes",
4699 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4700 .read_u64 = mem_cgroup_read_u64,
4703 .name = "max_usage_in_bytes",
4704 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4705 .write = mem_cgroup_reset,
4706 .read_u64 = mem_cgroup_read_u64,
4709 .name = "limit_in_bytes",
4710 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4711 .write = mem_cgroup_write,
4712 .read_u64 = mem_cgroup_read_u64,
4715 .name = "soft_limit_in_bytes",
4716 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4717 .write = mem_cgroup_write,
4718 .read_u64 = mem_cgroup_read_u64,
4722 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4723 .write = mem_cgroup_reset,
4724 .read_u64 = mem_cgroup_read_u64,
4728 .seq_show = memcg_stat_show,
4731 .name = "force_empty",
4732 .write = mem_cgroup_force_empty_write,
4735 .name = "use_hierarchy",
4736 .write_u64 = mem_cgroup_hierarchy_write,
4737 .read_u64 = mem_cgroup_hierarchy_read,
4740 .name = "cgroup.event_control", /* XXX: for compat */
4741 .write = memcg_write_event_control,
4742 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4745 .name = "swappiness",
4746 .read_u64 = mem_cgroup_swappiness_read,
4747 .write_u64 = mem_cgroup_swappiness_write,
4750 .name = "move_charge_at_immigrate",
4751 .read_u64 = mem_cgroup_move_charge_read,
4752 .write_u64 = mem_cgroup_move_charge_write,
4755 .name = "oom_control",
4756 .seq_show = mem_cgroup_oom_control_read,
4757 .write_u64 = mem_cgroup_oom_control_write,
4758 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4761 .name = "pressure_level",
4765 .name = "numa_stat",
4766 .seq_show = memcg_numa_stat_show,
4770 .name = "kmem.limit_in_bytes",
4771 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4772 .write = mem_cgroup_write,
4773 .read_u64 = mem_cgroup_read_u64,
4776 .name = "kmem.usage_in_bytes",
4777 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4778 .read_u64 = mem_cgroup_read_u64,
4781 .name = "kmem.failcnt",
4782 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4783 .write = mem_cgroup_reset,
4784 .read_u64 = mem_cgroup_read_u64,
4787 .name = "kmem.max_usage_in_bytes",
4788 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4789 .write = mem_cgroup_reset,
4790 .read_u64 = mem_cgroup_read_u64,
4792 #if defined(CONFIG_MEMCG_KMEM) && \
4793 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
4795 .name = "kmem.slabinfo",
4796 .seq_start = memcg_slab_start,
4797 .seq_next = memcg_slab_next,
4798 .seq_stop = memcg_slab_stop,
4799 .seq_show = memcg_slab_show,
4803 .name = "kmem.tcp.limit_in_bytes",
4804 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4805 .write = mem_cgroup_write,
4806 .read_u64 = mem_cgroup_read_u64,
4809 .name = "kmem.tcp.usage_in_bytes",
4810 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4811 .read_u64 = mem_cgroup_read_u64,
4814 .name = "kmem.tcp.failcnt",
4815 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4816 .write = mem_cgroup_reset,
4817 .read_u64 = mem_cgroup_read_u64,
4820 .name = "kmem.tcp.max_usage_in_bytes",
4821 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4822 .write = mem_cgroup_reset,
4823 .read_u64 = mem_cgroup_read_u64,
4825 { }, /* terminate */
4829 * Private memory cgroup IDR
4831 * Swap-out records and page cache shadow entries need to store memcg
4832 * references in constrained space, so we maintain an ID space that is
4833 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4834 * memory-controlled cgroups to 64k.
4836 * However, there usually are many references to the oflline CSS after
4837 * the cgroup has been destroyed, such as page cache or reclaimable
4838 * slab objects, that don't need to hang on to the ID. We want to keep
4839 * those dead CSS from occupying IDs, or we might quickly exhaust the
4840 * relatively small ID space and prevent the creation of new cgroups
4841 * even when there are much fewer than 64k cgroups - possibly none.
4843 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4844 * be freed and recycled when it's no longer needed, which is usually
4845 * when the CSS is offlined.
4847 * The only exception to that are records of swapped out tmpfs/shmem
4848 * pages that need to be attributed to live ancestors on swapin. But
4849 * those references are manageable from userspace.
4852 static DEFINE_IDR(mem_cgroup_idr);
4854 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4856 if (memcg->id.id > 0) {
4857 idr_remove(&mem_cgroup_idr, memcg->id.id);
4862 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4864 refcount_add(n, &memcg->id.ref);
4867 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4869 if (refcount_sub_and_test(n, &memcg->id.ref)) {
4870 mem_cgroup_id_remove(memcg);
4872 /* Memcg ID pins CSS */
4873 css_put(&memcg->css);
4877 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4879 mem_cgroup_id_put_many(memcg, 1);
4883 * mem_cgroup_from_id - look up a memcg from a memcg id
4884 * @id: the memcg id to look up
4886 * Caller must hold rcu_read_lock().
4888 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4890 WARN_ON_ONCE(!rcu_read_lock_held());
4891 return idr_find(&mem_cgroup_idr, id);
4894 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4896 struct mem_cgroup_per_node *pn;
4899 * This routine is called against possible nodes.
4900 * But it's BUG to call kmalloc() against offline node.
4902 * TODO: this routine can waste much memory for nodes which will
4903 * never be onlined. It's better to use memory hotplug callback
4906 if (!node_state(node, N_NORMAL_MEMORY))
4908 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4912 pn->lruvec_stat_local = alloc_percpu(struct lruvec_stat);
4913 if (!pn->lruvec_stat_local) {
4918 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
4919 if (!pn->lruvec_stat_cpu) {
4920 free_percpu(pn->lruvec_stat_local);
4925 lruvec_init(&pn->lruvec);
4926 pn->usage_in_excess = 0;
4927 pn->on_tree = false;
4930 memcg->nodeinfo[node] = pn;
4934 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4936 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
4941 free_percpu(pn->lruvec_stat_cpu);
4942 free_percpu(pn->lruvec_stat_local);
4946 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4951 free_mem_cgroup_per_node_info(memcg, node);
4952 free_percpu(memcg->vmstats_percpu);
4953 free_percpu(memcg->vmstats_local);
4957 static void mem_cgroup_free(struct mem_cgroup *memcg)
4959 memcg_wb_domain_exit(memcg);
4961 * Flush percpu vmstats and vmevents to guarantee the value correctness
4962 * on parent's and all ancestor levels.
4964 memcg_flush_percpu_vmstats(memcg);
4965 memcg_flush_percpu_vmevents(memcg);
4966 __mem_cgroup_free(memcg);
4969 static struct mem_cgroup *mem_cgroup_alloc(void)
4971 struct mem_cgroup *memcg;
4974 int __maybe_unused i;
4976 size = sizeof(struct mem_cgroup);
4977 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4979 memcg = kzalloc(size, GFP_KERNEL);
4983 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4984 1, MEM_CGROUP_ID_MAX,
4986 if (memcg->id.id < 0)
4989 memcg->vmstats_local = alloc_percpu(struct memcg_vmstats_percpu);
4990 if (!memcg->vmstats_local)
4993 memcg->vmstats_percpu = alloc_percpu(struct memcg_vmstats_percpu);
4994 if (!memcg->vmstats_percpu)
4998 if (alloc_mem_cgroup_per_node_info(memcg, node))
5001 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5004 INIT_WORK(&memcg->high_work, high_work_func);
5005 INIT_LIST_HEAD(&memcg->oom_notify);
5006 mutex_init(&memcg->thresholds_lock);
5007 spin_lock_init(&memcg->move_lock);
5008 vmpressure_init(&memcg->vmpressure);
5009 INIT_LIST_HEAD(&memcg->event_list);
5010 spin_lock_init(&memcg->event_list_lock);
5011 memcg->socket_pressure = jiffies;
5012 #ifdef CONFIG_MEMCG_KMEM
5013 memcg->kmemcg_id = -1;
5015 #ifdef CONFIG_CGROUP_WRITEBACK
5016 INIT_LIST_HEAD(&memcg->cgwb_list);
5017 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5018 memcg->cgwb_frn[i].done =
5019 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5021 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5022 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5023 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5024 memcg->deferred_split_queue.split_queue_len = 0;
5026 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5029 mem_cgroup_id_remove(memcg);
5030 __mem_cgroup_free(memcg);
5034 static struct cgroup_subsys_state * __ref
5035 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5037 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5038 struct mem_cgroup *memcg;
5039 long error = -ENOMEM;
5041 memcg = mem_cgroup_alloc();
5043 return ERR_PTR(error);
5045 memcg->high = PAGE_COUNTER_MAX;
5046 memcg->soft_limit = PAGE_COUNTER_MAX;
5048 memcg->swappiness = mem_cgroup_swappiness(parent);
5049 memcg->oom_kill_disable = parent->oom_kill_disable;
5051 if (parent && parent->use_hierarchy) {
5052 memcg->use_hierarchy = true;
5053 page_counter_init(&memcg->memory, &parent->memory);
5054 page_counter_init(&memcg->swap, &parent->swap);
5055 page_counter_init(&memcg->memsw, &parent->memsw);
5056 page_counter_init(&memcg->kmem, &parent->kmem);
5057 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5059 page_counter_init(&memcg->memory, NULL);
5060 page_counter_init(&memcg->swap, NULL);
5061 page_counter_init(&memcg->memsw, NULL);
5062 page_counter_init(&memcg->kmem, NULL);
5063 page_counter_init(&memcg->tcpmem, NULL);
5065 * Deeper hierachy with use_hierarchy == false doesn't make
5066 * much sense so let cgroup subsystem know about this
5067 * unfortunate state in our controller.
5069 if (parent != root_mem_cgroup)
5070 memory_cgrp_subsys.broken_hierarchy = true;
5073 /* The following stuff does not apply to the root */
5075 #ifdef CONFIG_MEMCG_KMEM
5076 INIT_LIST_HEAD(&memcg->kmem_caches);
5078 root_mem_cgroup = memcg;
5082 error = memcg_online_kmem(memcg);
5086 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5087 static_branch_inc(&memcg_sockets_enabled_key);
5091 mem_cgroup_id_remove(memcg);
5092 mem_cgroup_free(memcg);
5093 return ERR_PTR(-ENOMEM);
5096 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5098 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5101 * A memcg must be visible for memcg_expand_shrinker_maps()
5102 * by the time the maps are allocated. So, we allocate maps
5103 * here, when for_each_mem_cgroup() can't skip it.
5105 if (memcg_alloc_shrinker_maps(memcg)) {
5106 mem_cgroup_id_remove(memcg);
5110 /* Online state pins memcg ID, memcg ID pins CSS */
5111 refcount_set(&memcg->id.ref, 1);
5116 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5118 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5119 struct mem_cgroup_event *event, *tmp;
5122 * Unregister events and notify userspace.
5123 * Notify userspace about cgroup removing only after rmdir of cgroup
5124 * directory to avoid race between userspace and kernelspace.
5126 spin_lock(&memcg->event_list_lock);
5127 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5128 list_del_init(&event->list);
5129 schedule_work(&event->remove);
5131 spin_unlock(&memcg->event_list_lock);
5133 page_counter_set_min(&memcg->memory, 0);
5134 page_counter_set_low(&memcg->memory, 0);
5136 memcg_offline_kmem(memcg);
5137 wb_memcg_offline(memcg);
5139 drain_all_stock(memcg);
5141 mem_cgroup_id_put(memcg);
5144 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5146 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5148 invalidate_reclaim_iterators(memcg);
5151 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5153 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5154 int __maybe_unused i;
5156 #ifdef CONFIG_CGROUP_WRITEBACK
5157 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5158 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5160 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5161 static_branch_dec(&memcg_sockets_enabled_key);
5163 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5164 static_branch_dec(&memcg_sockets_enabled_key);
5166 vmpressure_cleanup(&memcg->vmpressure);
5167 cancel_work_sync(&memcg->high_work);
5168 mem_cgroup_remove_from_trees(memcg);
5169 memcg_free_shrinker_maps(memcg);
5170 memcg_free_kmem(memcg);
5171 mem_cgroup_free(memcg);
5175 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5176 * @css: the target css
5178 * Reset the states of the mem_cgroup associated with @css. This is
5179 * invoked when the userland requests disabling on the default hierarchy
5180 * but the memcg is pinned through dependency. The memcg should stop
5181 * applying policies and should revert to the vanilla state as it may be
5182 * made visible again.
5184 * The current implementation only resets the essential configurations.
5185 * This needs to be expanded to cover all the visible parts.
5187 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5189 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5191 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5192 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5193 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
5194 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5195 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5196 page_counter_set_min(&memcg->memory, 0);
5197 page_counter_set_low(&memcg->memory, 0);
5198 memcg->high = PAGE_COUNTER_MAX;
5199 memcg->soft_limit = PAGE_COUNTER_MAX;
5200 memcg_wb_domain_size_changed(memcg);
5204 /* Handlers for move charge at task migration. */
5205 static int mem_cgroup_do_precharge(unsigned long count)
5209 /* Try a single bulk charge without reclaim first, kswapd may wake */
5210 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5212 mc.precharge += count;
5216 /* Try charges one by one with reclaim, but do not retry */
5218 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5232 enum mc_target_type {
5239 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5240 unsigned long addr, pte_t ptent)
5242 struct page *page = vm_normal_page(vma, addr, ptent);
5244 if (!page || !page_mapped(page))
5246 if (PageAnon(page)) {
5247 if (!(mc.flags & MOVE_ANON))
5250 if (!(mc.flags & MOVE_FILE))
5253 if (!get_page_unless_zero(page))
5259 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5260 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5261 pte_t ptent, swp_entry_t *entry)
5263 struct page *page = NULL;
5264 swp_entry_t ent = pte_to_swp_entry(ptent);
5266 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
5270 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5271 * a device and because they are not accessible by CPU they are store
5272 * as special swap entry in the CPU page table.
5274 if (is_device_private_entry(ent)) {
5275 page = device_private_entry_to_page(ent);
5277 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5278 * a refcount of 1 when free (unlike normal page)
5280 if (!page_ref_add_unless(page, 1, 1))
5286 * Because lookup_swap_cache() updates some statistics counter,
5287 * we call find_get_page() with swapper_space directly.
5289 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5290 if (do_memsw_account())
5291 entry->val = ent.val;
5296 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5297 pte_t ptent, swp_entry_t *entry)
5303 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5304 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5306 struct page *page = NULL;
5307 struct address_space *mapping;
5310 if (!vma->vm_file) /* anonymous vma */
5312 if (!(mc.flags & MOVE_FILE))
5315 mapping = vma->vm_file->f_mapping;
5316 pgoff = linear_page_index(vma, addr);
5318 /* page is moved even if it's not RSS of this task(page-faulted). */
5320 /* shmem/tmpfs may report page out on swap: account for that too. */
5321 if (shmem_mapping(mapping)) {
5322 page = find_get_entry(mapping, pgoff);
5323 if (xa_is_value(page)) {
5324 swp_entry_t swp = radix_to_swp_entry(page);
5325 if (do_memsw_account())
5327 page = find_get_page(swap_address_space(swp),
5331 page = find_get_page(mapping, pgoff);
5333 page = find_get_page(mapping, pgoff);
5339 * mem_cgroup_move_account - move account of the page
5341 * @compound: charge the page as compound or small page
5342 * @from: mem_cgroup which the page is moved from.
5343 * @to: mem_cgroup which the page is moved to. @from != @to.
5345 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5347 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5350 static int mem_cgroup_move_account(struct page *page,
5352 struct mem_cgroup *from,
5353 struct mem_cgroup *to)
5355 struct lruvec *from_vec, *to_vec;
5356 struct pglist_data *pgdat;
5357 unsigned long flags;
5358 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5362 VM_BUG_ON(from == to);
5363 VM_BUG_ON_PAGE(PageLRU(page), page);
5364 VM_BUG_ON(compound && !PageTransHuge(page));
5367 * Prevent mem_cgroup_migrate() from looking at
5368 * page->mem_cgroup of its source page while we change it.
5371 if (!trylock_page(page))
5375 if (page->mem_cgroup != from)
5378 anon = PageAnon(page);
5380 pgdat = page_pgdat(page);
5381 from_vec = mem_cgroup_lruvec(from, pgdat);
5382 to_vec = mem_cgroup_lruvec(to, pgdat);
5384 spin_lock_irqsave(&from->move_lock, flags);
5386 if (!anon && page_mapped(page)) {
5387 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5388 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5392 * move_lock grabbed above and caller set from->moving_account, so
5393 * mod_memcg_page_state will serialize updates to PageDirty.
5394 * So mapping should be stable for dirty pages.
5396 if (!anon && PageDirty(page)) {
5397 struct address_space *mapping = page_mapping(page);
5399 if (mapping_cap_account_dirty(mapping)) {
5400 __mod_lruvec_state(from_vec, NR_FILE_DIRTY, -nr_pages);
5401 __mod_lruvec_state(to_vec, NR_FILE_DIRTY, nr_pages);
5405 if (PageWriteback(page)) {
5406 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5407 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5411 * It is safe to change page->mem_cgroup here because the page
5412 * is referenced, charged, and isolated - we can't race with
5413 * uncharging, charging, migration, or LRU putback.
5416 /* caller should have done css_get */
5417 page->mem_cgroup = to;
5419 spin_unlock_irqrestore(&from->move_lock, flags);
5423 local_irq_disable();
5424 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
5425 memcg_check_events(to, page);
5426 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
5427 memcg_check_events(from, page);
5436 * get_mctgt_type - get target type of moving charge
5437 * @vma: the vma the pte to be checked belongs
5438 * @addr: the address corresponding to the pte to be checked
5439 * @ptent: the pte to be checked
5440 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5443 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5444 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5445 * move charge. if @target is not NULL, the page is stored in target->page
5446 * with extra refcnt got(Callers should handle it).
5447 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5448 * target for charge migration. if @target is not NULL, the entry is stored
5450 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5451 * (so ZONE_DEVICE page and thus not on the lru).
5452 * For now we such page is charge like a regular page would be as for all
5453 * intent and purposes it is just special memory taking the place of a
5456 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5458 * Called with pte lock held.
5461 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5462 unsigned long addr, pte_t ptent, union mc_target *target)
5464 struct page *page = NULL;
5465 enum mc_target_type ret = MC_TARGET_NONE;
5466 swp_entry_t ent = { .val = 0 };
5468 if (pte_present(ptent))
5469 page = mc_handle_present_pte(vma, addr, ptent);
5470 else if (is_swap_pte(ptent))
5471 page = mc_handle_swap_pte(vma, ptent, &ent);
5472 else if (pte_none(ptent))
5473 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5475 if (!page && !ent.val)
5479 * Do only loose check w/o serialization.
5480 * mem_cgroup_move_account() checks the page is valid or
5481 * not under LRU exclusion.
5483 if (page->mem_cgroup == mc.from) {
5484 ret = MC_TARGET_PAGE;
5485 if (is_device_private_page(page))
5486 ret = MC_TARGET_DEVICE;
5488 target->page = page;
5490 if (!ret || !target)
5494 * There is a swap entry and a page doesn't exist or isn't charged.
5495 * But we cannot move a tail-page in a THP.
5497 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5498 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5499 ret = MC_TARGET_SWAP;
5506 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5508 * We don't consider PMD mapped swapping or file mapped pages because THP does
5509 * not support them for now.
5510 * Caller should make sure that pmd_trans_huge(pmd) is true.
5512 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5513 unsigned long addr, pmd_t pmd, union mc_target *target)
5515 struct page *page = NULL;
5516 enum mc_target_type ret = MC_TARGET_NONE;
5518 if (unlikely(is_swap_pmd(pmd))) {
5519 VM_BUG_ON(thp_migration_supported() &&
5520 !is_pmd_migration_entry(pmd));
5523 page = pmd_page(pmd);
5524 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5525 if (!(mc.flags & MOVE_ANON))
5527 if (page->mem_cgroup == mc.from) {
5528 ret = MC_TARGET_PAGE;
5531 target->page = page;
5537 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5538 unsigned long addr, pmd_t pmd, union mc_target *target)
5540 return MC_TARGET_NONE;
5544 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5545 unsigned long addr, unsigned long end,
5546 struct mm_walk *walk)
5548 struct vm_area_struct *vma = walk->vma;
5552 ptl = pmd_trans_huge_lock(pmd, vma);
5555 * Note their can not be MC_TARGET_DEVICE for now as we do not
5556 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5557 * this might change.
5559 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5560 mc.precharge += HPAGE_PMD_NR;
5565 if (pmd_trans_unstable(pmd))
5567 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5568 for (; addr != end; pte++, addr += PAGE_SIZE)
5569 if (get_mctgt_type(vma, addr, *pte, NULL))
5570 mc.precharge++; /* increment precharge temporarily */
5571 pte_unmap_unlock(pte - 1, ptl);
5577 static const struct mm_walk_ops precharge_walk_ops = {
5578 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5581 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5583 unsigned long precharge;
5585 down_read(&mm->mmap_sem);
5586 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5587 up_read(&mm->mmap_sem);
5589 precharge = mc.precharge;
5595 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5597 unsigned long precharge = mem_cgroup_count_precharge(mm);
5599 VM_BUG_ON(mc.moving_task);
5600 mc.moving_task = current;
5601 return mem_cgroup_do_precharge(precharge);
5604 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5605 static void __mem_cgroup_clear_mc(void)
5607 struct mem_cgroup *from = mc.from;
5608 struct mem_cgroup *to = mc.to;
5610 /* we must uncharge all the leftover precharges from mc.to */
5612 cancel_charge(mc.to, mc.precharge);
5616 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5617 * we must uncharge here.
5619 if (mc.moved_charge) {
5620 cancel_charge(mc.from, mc.moved_charge);
5621 mc.moved_charge = 0;
5623 /* we must fixup refcnts and charges */
5624 if (mc.moved_swap) {
5625 /* uncharge swap account from the old cgroup */
5626 if (!mem_cgroup_is_root(mc.from))
5627 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5629 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5632 * we charged both to->memory and to->memsw, so we
5633 * should uncharge to->memory.
5635 if (!mem_cgroup_is_root(mc.to))
5636 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5638 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
5639 css_put_many(&mc.to->css, mc.moved_swap);
5643 memcg_oom_recover(from);
5644 memcg_oom_recover(to);
5645 wake_up_all(&mc.waitq);
5648 static void mem_cgroup_clear_mc(void)
5650 struct mm_struct *mm = mc.mm;
5653 * we must clear moving_task before waking up waiters at the end of
5656 mc.moving_task = NULL;
5657 __mem_cgroup_clear_mc();
5658 spin_lock(&mc.lock);
5662 spin_unlock(&mc.lock);
5667 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5669 struct cgroup_subsys_state *css;
5670 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5671 struct mem_cgroup *from;
5672 struct task_struct *leader, *p;
5673 struct mm_struct *mm;
5674 unsigned long move_flags;
5677 /* charge immigration isn't supported on the default hierarchy */
5678 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5682 * Multi-process migrations only happen on the default hierarchy
5683 * where charge immigration is not used. Perform charge
5684 * immigration if @tset contains a leader and whine if there are
5688 cgroup_taskset_for_each_leader(leader, css, tset) {
5691 memcg = mem_cgroup_from_css(css);
5697 * We are now commited to this value whatever it is. Changes in this
5698 * tunable will only affect upcoming migrations, not the current one.
5699 * So we need to save it, and keep it going.
5701 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5705 from = mem_cgroup_from_task(p);
5707 VM_BUG_ON(from == memcg);
5709 mm = get_task_mm(p);
5712 /* We move charges only when we move a owner of the mm */
5713 if (mm->owner == p) {
5716 VM_BUG_ON(mc.precharge);
5717 VM_BUG_ON(mc.moved_charge);
5718 VM_BUG_ON(mc.moved_swap);
5720 spin_lock(&mc.lock);
5724 mc.flags = move_flags;
5725 spin_unlock(&mc.lock);
5726 /* We set mc.moving_task later */
5728 ret = mem_cgroup_precharge_mc(mm);
5730 mem_cgroup_clear_mc();
5737 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5740 mem_cgroup_clear_mc();
5743 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5744 unsigned long addr, unsigned long end,
5745 struct mm_walk *walk)
5748 struct vm_area_struct *vma = walk->vma;
5751 enum mc_target_type target_type;
5752 union mc_target target;
5755 ptl = pmd_trans_huge_lock(pmd, vma);
5757 if (mc.precharge < HPAGE_PMD_NR) {
5761 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5762 if (target_type == MC_TARGET_PAGE) {
5764 if (!isolate_lru_page(page)) {
5765 if (!mem_cgroup_move_account(page, true,
5767 mc.precharge -= HPAGE_PMD_NR;
5768 mc.moved_charge += HPAGE_PMD_NR;
5770 putback_lru_page(page);
5773 } else if (target_type == MC_TARGET_DEVICE) {
5775 if (!mem_cgroup_move_account(page, true,
5777 mc.precharge -= HPAGE_PMD_NR;
5778 mc.moved_charge += HPAGE_PMD_NR;
5786 if (pmd_trans_unstable(pmd))
5789 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5790 for (; addr != end; addr += PAGE_SIZE) {
5791 pte_t ptent = *(pte++);
5792 bool device = false;
5798 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5799 case MC_TARGET_DEVICE:
5802 case MC_TARGET_PAGE:
5805 * We can have a part of the split pmd here. Moving it
5806 * can be done but it would be too convoluted so simply
5807 * ignore such a partial THP and keep it in original
5808 * memcg. There should be somebody mapping the head.
5810 if (PageTransCompound(page))
5812 if (!device && isolate_lru_page(page))
5814 if (!mem_cgroup_move_account(page, false,
5817 /* we uncharge from mc.from later. */
5821 putback_lru_page(page);
5822 put: /* get_mctgt_type() gets the page */
5825 case MC_TARGET_SWAP:
5827 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5829 /* we fixup refcnts and charges later. */
5837 pte_unmap_unlock(pte - 1, ptl);
5842 * We have consumed all precharges we got in can_attach().
5843 * We try charge one by one, but don't do any additional
5844 * charges to mc.to if we have failed in charge once in attach()
5847 ret = mem_cgroup_do_precharge(1);
5855 static const struct mm_walk_ops charge_walk_ops = {
5856 .pmd_entry = mem_cgroup_move_charge_pte_range,
5859 static void mem_cgroup_move_charge(void)
5861 lru_add_drain_all();
5863 * Signal lock_page_memcg() to take the memcg's move_lock
5864 * while we're moving its pages to another memcg. Then wait
5865 * for already started RCU-only updates to finish.
5867 atomic_inc(&mc.from->moving_account);
5870 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5872 * Someone who are holding the mmap_sem might be waiting in
5873 * waitq. So we cancel all extra charges, wake up all waiters,
5874 * and retry. Because we cancel precharges, we might not be able
5875 * to move enough charges, but moving charge is a best-effort
5876 * feature anyway, so it wouldn't be a big problem.
5878 __mem_cgroup_clear_mc();
5883 * When we have consumed all precharges and failed in doing
5884 * additional charge, the page walk just aborts.
5886 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
5889 up_read(&mc.mm->mmap_sem);
5890 atomic_dec(&mc.from->moving_account);
5893 static void mem_cgroup_move_task(void)
5896 mem_cgroup_move_charge();
5897 mem_cgroup_clear_mc();
5900 #else /* !CONFIG_MMU */
5901 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5905 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5908 static void mem_cgroup_move_task(void)
5914 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5915 * to verify whether we're attached to the default hierarchy on each mount
5918 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5921 * use_hierarchy is forced on the default hierarchy. cgroup core
5922 * guarantees that @root doesn't have any children, so turning it
5923 * on for the root memcg is enough.
5925 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5926 root_mem_cgroup->use_hierarchy = true;
5928 root_mem_cgroup->use_hierarchy = false;
5931 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
5933 if (value == PAGE_COUNTER_MAX)
5934 seq_puts(m, "max\n");
5936 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
5941 static u64 memory_current_read(struct cgroup_subsys_state *css,
5944 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5946 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5949 static int memory_min_show(struct seq_file *m, void *v)
5951 return seq_puts_memcg_tunable(m,
5952 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
5955 static ssize_t memory_min_write(struct kernfs_open_file *of,
5956 char *buf, size_t nbytes, loff_t off)
5958 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5962 buf = strstrip(buf);
5963 err = page_counter_memparse(buf, "max", &min);
5967 page_counter_set_min(&memcg->memory, min);
5972 static int memory_low_show(struct seq_file *m, void *v)
5974 return seq_puts_memcg_tunable(m,
5975 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
5978 static ssize_t memory_low_write(struct kernfs_open_file *of,
5979 char *buf, size_t nbytes, loff_t off)
5981 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5985 buf = strstrip(buf);
5986 err = page_counter_memparse(buf, "max", &low);
5990 page_counter_set_low(&memcg->memory, low);
5995 static int memory_high_show(struct seq_file *m, void *v)
5997 return seq_puts_memcg_tunable(m, READ_ONCE(mem_cgroup_from_seq(m)->high));
6000 static ssize_t memory_high_write(struct kernfs_open_file *of,
6001 char *buf, size_t nbytes, loff_t off)
6003 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6004 unsigned int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
6005 bool drained = false;
6009 buf = strstrip(buf);
6010 err = page_counter_memparse(buf, "max", &high);
6017 unsigned long nr_pages = page_counter_read(&memcg->memory);
6018 unsigned long reclaimed;
6020 if (nr_pages <= high)
6023 if (signal_pending(current))
6027 drain_all_stock(memcg);
6032 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6035 if (!reclaimed && !nr_retries--)
6042 static int memory_max_show(struct seq_file *m, void *v)
6044 return seq_puts_memcg_tunable(m,
6045 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6048 static ssize_t memory_max_write(struct kernfs_open_file *of,
6049 char *buf, size_t nbytes, loff_t off)
6051 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6052 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
6053 bool drained = false;
6057 buf = strstrip(buf);
6058 err = page_counter_memparse(buf, "max", &max);
6062 xchg(&memcg->memory.max, max);
6065 unsigned long nr_pages = page_counter_read(&memcg->memory);
6067 if (nr_pages <= max)
6070 if (signal_pending(current))
6074 drain_all_stock(memcg);
6080 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6086 memcg_memory_event(memcg, MEMCG_OOM);
6087 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6091 memcg_wb_domain_size_changed(memcg);
6095 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6097 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6098 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6099 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6100 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6101 seq_printf(m, "oom_kill %lu\n",
6102 atomic_long_read(&events[MEMCG_OOM_KILL]));
6105 static int memory_events_show(struct seq_file *m, void *v)
6107 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6109 __memory_events_show(m, memcg->memory_events);
6113 static int memory_events_local_show(struct seq_file *m, void *v)
6115 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6117 __memory_events_show(m, memcg->memory_events_local);
6121 static int memory_stat_show(struct seq_file *m, void *v)
6123 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6126 buf = memory_stat_format(memcg);
6134 static int memory_oom_group_show(struct seq_file *m, void *v)
6136 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6138 seq_printf(m, "%d\n", memcg->oom_group);
6143 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6144 char *buf, size_t nbytes, loff_t off)
6146 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6149 buf = strstrip(buf);
6153 ret = kstrtoint(buf, 0, &oom_group);
6157 if (oom_group != 0 && oom_group != 1)
6160 memcg->oom_group = oom_group;
6165 static struct cftype memory_files[] = {
6168 .flags = CFTYPE_NOT_ON_ROOT,
6169 .read_u64 = memory_current_read,
6173 .flags = CFTYPE_NOT_ON_ROOT,
6174 .seq_show = memory_min_show,
6175 .write = memory_min_write,
6179 .flags = CFTYPE_NOT_ON_ROOT,
6180 .seq_show = memory_low_show,
6181 .write = memory_low_write,
6185 .flags = CFTYPE_NOT_ON_ROOT,
6186 .seq_show = memory_high_show,
6187 .write = memory_high_write,
6191 .flags = CFTYPE_NOT_ON_ROOT,
6192 .seq_show = memory_max_show,
6193 .write = memory_max_write,
6197 .flags = CFTYPE_NOT_ON_ROOT,
6198 .file_offset = offsetof(struct mem_cgroup, events_file),
6199 .seq_show = memory_events_show,
6202 .name = "events.local",
6203 .flags = CFTYPE_NOT_ON_ROOT,
6204 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6205 .seq_show = memory_events_local_show,
6209 .flags = CFTYPE_NOT_ON_ROOT,
6210 .seq_show = memory_stat_show,
6213 .name = "oom.group",
6214 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6215 .seq_show = memory_oom_group_show,
6216 .write = memory_oom_group_write,
6221 struct cgroup_subsys memory_cgrp_subsys = {
6222 .css_alloc = mem_cgroup_css_alloc,
6223 .css_online = mem_cgroup_css_online,
6224 .css_offline = mem_cgroup_css_offline,
6225 .css_released = mem_cgroup_css_released,
6226 .css_free = mem_cgroup_css_free,
6227 .css_reset = mem_cgroup_css_reset,
6228 .can_attach = mem_cgroup_can_attach,
6229 .cancel_attach = mem_cgroup_cancel_attach,
6230 .post_attach = mem_cgroup_move_task,
6231 .bind = mem_cgroup_bind,
6232 .dfl_cftypes = memory_files,
6233 .legacy_cftypes = mem_cgroup_legacy_files,
6238 * mem_cgroup_protected - check if memory consumption is in the normal range
6239 * @root: the top ancestor of the sub-tree being checked
6240 * @memcg: the memory cgroup to check
6242 * WARNING: This function is not stateless! It can only be used as part
6243 * of a top-down tree iteration, not for isolated queries.
6245 * Returns one of the following:
6246 * MEMCG_PROT_NONE: cgroup memory is not protected
6247 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
6248 * an unprotected supply of reclaimable memory from other cgroups.
6249 * MEMCG_PROT_MIN: cgroup memory is protected
6251 * @root is exclusive; it is never protected when looked at directly
6253 * To provide a proper hierarchical behavior, effective memory.min/low values
6254 * are used. Below is the description of how effective memory.low is calculated.
6255 * Effective memory.min values is calculated in the same way.
6257 * Effective memory.low is always equal or less than the original memory.low.
6258 * If there is no memory.low overcommittment (which is always true for
6259 * top-level memory cgroups), these two values are equal.
6260 * Otherwise, it's a part of parent's effective memory.low,
6261 * calculated as a cgroup's memory.low usage divided by sum of sibling's
6262 * memory.low usages, where memory.low usage is the size of actually
6266 * elow = min( memory.low, parent->elow * ------------------ ),
6267 * siblings_low_usage
6269 * | memory.current, if memory.current < memory.low
6274 * Such definition of the effective memory.low provides the expected
6275 * hierarchical behavior: parent's memory.low value is limiting
6276 * children, unprotected memory is reclaimed first and cgroups,
6277 * which are not using their guarantee do not affect actual memory
6280 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
6282 * A A/memory.low = 2G, A/memory.current = 6G
6284 * BC DE B/memory.low = 3G B/memory.current = 2G
6285 * C/memory.low = 1G C/memory.current = 2G
6286 * D/memory.low = 0 D/memory.current = 2G
6287 * E/memory.low = 10G E/memory.current = 0
6289 * and the memory pressure is applied, the following memory distribution
6290 * is expected (approximately):
6292 * A/memory.current = 2G
6294 * B/memory.current = 1.3G
6295 * C/memory.current = 0.6G
6296 * D/memory.current = 0
6297 * E/memory.current = 0
6299 * These calculations require constant tracking of the actual low usages
6300 * (see propagate_protected_usage()), as well as recursive calculation of
6301 * effective memory.low values. But as we do call mem_cgroup_protected()
6302 * path for each memory cgroup top-down from the reclaim,
6303 * it's possible to optimize this part, and save calculated elow
6304 * for next usage. This part is intentionally racy, but it's ok,
6305 * as memory.low is a best-effort mechanism.
6307 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
6308 struct mem_cgroup *memcg)
6310 struct mem_cgroup *parent;
6311 unsigned long emin, parent_emin;
6312 unsigned long elow, parent_elow;
6313 unsigned long usage;
6315 if (mem_cgroup_disabled())
6316 return MEMCG_PROT_NONE;
6319 root = root_mem_cgroup;
6321 return MEMCG_PROT_NONE;
6323 usage = page_counter_read(&memcg->memory);
6325 return MEMCG_PROT_NONE;
6327 emin = memcg->memory.min;
6328 elow = memcg->memory.low;
6330 parent = parent_mem_cgroup(memcg);
6331 /* No parent means a non-hierarchical mode on v1 memcg */
6333 return MEMCG_PROT_NONE;
6338 parent_emin = READ_ONCE(parent->memory.emin);
6339 emin = min(emin, parent_emin);
6340 if (emin && parent_emin) {
6341 unsigned long min_usage, siblings_min_usage;
6343 min_usage = min(usage, memcg->memory.min);
6344 siblings_min_usage = atomic_long_read(
6345 &parent->memory.children_min_usage);
6347 if (min_usage && siblings_min_usage)
6348 emin = min(emin, parent_emin * min_usage /
6349 siblings_min_usage);
6352 parent_elow = READ_ONCE(parent->memory.elow);
6353 elow = min(elow, parent_elow);
6354 if (elow && parent_elow) {
6355 unsigned long low_usage, siblings_low_usage;
6357 low_usage = min(usage, memcg->memory.low);
6358 siblings_low_usage = atomic_long_read(
6359 &parent->memory.children_low_usage);
6361 if (low_usage && siblings_low_usage)
6362 elow = min(elow, parent_elow * low_usage /
6363 siblings_low_usage);
6367 memcg->memory.emin = emin;
6368 memcg->memory.elow = elow;
6371 return MEMCG_PROT_MIN;
6372 else if (usage <= elow)
6373 return MEMCG_PROT_LOW;
6375 return MEMCG_PROT_NONE;
6379 * mem_cgroup_try_charge - try charging a page
6380 * @page: page to charge
6381 * @mm: mm context of the victim
6382 * @gfp_mask: reclaim mode
6383 * @memcgp: charged memcg return
6384 * @compound: charge the page as compound or small page
6386 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6387 * pages according to @gfp_mask if necessary.
6389 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6390 * Otherwise, an error code is returned.
6392 * After page->mapping has been set up, the caller must finalize the
6393 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6394 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6396 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
6397 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6400 struct mem_cgroup *memcg = NULL;
6401 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6404 if (mem_cgroup_disabled())
6407 if (PageSwapCache(page)) {
6409 * Every swap fault against a single page tries to charge the
6410 * page, bail as early as possible. shmem_unuse() encounters
6411 * already charged pages, too. The USED bit is protected by
6412 * the page lock, which serializes swap cache removal, which
6413 * in turn serializes uncharging.
6415 VM_BUG_ON_PAGE(!PageLocked(page), page);
6416 if (compound_head(page)->mem_cgroup)
6419 if (do_swap_account) {
6420 swp_entry_t ent = { .val = page_private(page), };
6421 unsigned short id = lookup_swap_cgroup_id(ent);
6424 memcg = mem_cgroup_from_id(id);
6425 if (memcg && !css_tryget_online(&memcg->css))
6432 memcg = get_mem_cgroup_from_mm(mm);
6434 ret = try_charge(memcg, gfp_mask, nr_pages);
6436 css_put(&memcg->css);
6442 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
6443 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6446 struct mem_cgroup *memcg;
6449 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound);
6451 mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask);
6456 * mem_cgroup_commit_charge - commit a page charge
6457 * @page: page to charge
6458 * @memcg: memcg to charge the page to
6459 * @lrucare: page might be on LRU already
6460 * @compound: charge the page as compound or small page
6462 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6463 * after page->mapping has been set up. This must happen atomically
6464 * as part of the page instantiation, i.e. under the page table lock
6465 * for anonymous pages, under the page lock for page and swap cache.
6467 * In addition, the page must not be on the LRU during the commit, to
6468 * prevent racing with task migration. If it might be, use @lrucare.
6470 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6472 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
6473 bool lrucare, bool compound)
6475 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6477 VM_BUG_ON_PAGE(!page->mapping, page);
6478 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
6480 if (mem_cgroup_disabled())
6483 * Swap faults will attempt to charge the same page multiple
6484 * times. But reuse_swap_page() might have removed the page
6485 * from swapcache already, so we can't check PageSwapCache().
6490 commit_charge(page, memcg, lrucare);
6492 local_irq_disable();
6493 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
6494 memcg_check_events(memcg, page);
6497 if (do_memsw_account() && PageSwapCache(page)) {
6498 swp_entry_t entry = { .val = page_private(page) };
6500 * The swap entry might not get freed for a long time,
6501 * let's not wait for it. The page already received a
6502 * memory+swap charge, drop the swap entry duplicate.
6504 mem_cgroup_uncharge_swap(entry, nr_pages);
6509 * mem_cgroup_cancel_charge - cancel a page charge
6510 * @page: page to charge
6511 * @memcg: memcg to charge the page to
6512 * @compound: charge the page as compound or small page
6514 * Cancel a charge transaction started by mem_cgroup_try_charge().
6516 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
6519 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6521 if (mem_cgroup_disabled())
6524 * Swap faults will attempt to charge the same page multiple
6525 * times. But reuse_swap_page() might have removed the page
6526 * from swapcache already, so we can't check PageSwapCache().
6531 cancel_charge(memcg, nr_pages);
6534 struct uncharge_gather {
6535 struct mem_cgroup *memcg;
6536 unsigned long pgpgout;
6537 unsigned long nr_anon;
6538 unsigned long nr_file;
6539 unsigned long nr_kmem;
6540 unsigned long nr_huge;
6541 unsigned long nr_shmem;
6542 struct page *dummy_page;
6545 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6547 memset(ug, 0, sizeof(*ug));
6550 static void uncharge_batch(const struct uncharge_gather *ug)
6552 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
6553 unsigned long flags;
6555 if (!mem_cgroup_is_root(ug->memcg)) {
6556 page_counter_uncharge(&ug->memcg->memory, nr_pages);
6557 if (do_memsw_account())
6558 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
6559 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6560 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6561 memcg_oom_recover(ug->memcg);
6564 local_irq_save(flags);
6565 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
6566 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
6567 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
6568 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
6569 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6570 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, nr_pages);
6571 memcg_check_events(ug->memcg, ug->dummy_page);
6572 local_irq_restore(flags);
6574 if (!mem_cgroup_is_root(ug->memcg))
6575 css_put_many(&ug->memcg->css, nr_pages);
6578 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6580 VM_BUG_ON_PAGE(PageLRU(page), page);
6581 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
6582 !PageHWPoison(page) , page);
6584 if (!page->mem_cgroup)
6588 * Nobody should be changing or seriously looking at
6589 * page->mem_cgroup at this point, we have fully
6590 * exclusive access to the page.
6593 if (ug->memcg != page->mem_cgroup) {
6596 uncharge_gather_clear(ug);
6598 ug->memcg = page->mem_cgroup;
6601 if (!PageKmemcg(page)) {
6602 unsigned int nr_pages = 1;
6604 if (PageTransHuge(page)) {
6605 nr_pages = compound_nr(page);
6606 ug->nr_huge += nr_pages;
6609 ug->nr_anon += nr_pages;
6611 ug->nr_file += nr_pages;
6612 if (PageSwapBacked(page))
6613 ug->nr_shmem += nr_pages;
6617 ug->nr_kmem += compound_nr(page);
6618 __ClearPageKmemcg(page);
6621 ug->dummy_page = page;
6622 page->mem_cgroup = NULL;
6625 static void uncharge_list(struct list_head *page_list)
6627 struct uncharge_gather ug;
6628 struct list_head *next;
6630 uncharge_gather_clear(&ug);
6633 * Note that the list can be a single page->lru; hence the
6634 * do-while loop instead of a simple list_for_each_entry().
6636 next = page_list->next;
6640 page = list_entry(next, struct page, lru);
6641 next = page->lru.next;
6643 uncharge_page(page, &ug);
6644 } while (next != page_list);
6647 uncharge_batch(&ug);
6651 * mem_cgroup_uncharge - uncharge a page
6652 * @page: page to uncharge
6654 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6655 * mem_cgroup_commit_charge().
6657 void mem_cgroup_uncharge(struct page *page)
6659 struct uncharge_gather ug;
6661 if (mem_cgroup_disabled())
6664 /* Don't touch page->lru of any random page, pre-check: */
6665 if (!page->mem_cgroup)
6668 uncharge_gather_clear(&ug);
6669 uncharge_page(page, &ug);
6670 uncharge_batch(&ug);
6674 * mem_cgroup_uncharge_list - uncharge a list of page
6675 * @page_list: list of pages to uncharge
6677 * Uncharge a list of pages previously charged with
6678 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6680 void mem_cgroup_uncharge_list(struct list_head *page_list)
6682 if (mem_cgroup_disabled())
6685 if (!list_empty(page_list))
6686 uncharge_list(page_list);
6690 * mem_cgroup_migrate - charge a page's replacement
6691 * @oldpage: currently circulating page
6692 * @newpage: replacement page
6694 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6695 * be uncharged upon free.
6697 * Both pages must be locked, @newpage->mapping must be set up.
6699 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6701 struct mem_cgroup *memcg;
6702 unsigned int nr_pages;
6703 unsigned long flags;
6705 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6706 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6707 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6708 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6711 if (mem_cgroup_disabled())
6714 /* Page cache replacement: new page already charged? */
6715 if (newpage->mem_cgroup)
6718 /* Swapcache readahead pages can get replaced before being charged */
6719 memcg = oldpage->mem_cgroup;
6723 /* Force-charge the new page. The old one will be freed soon */
6724 nr_pages = hpage_nr_pages(newpage);
6726 page_counter_charge(&memcg->memory, nr_pages);
6727 if (do_memsw_account())
6728 page_counter_charge(&memcg->memsw, nr_pages);
6729 css_get_many(&memcg->css, nr_pages);
6731 commit_charge(newpage, memcg, false);
6733 local_irq_save(flags);
6734 mem_cgroup_charge_statistics(memcg, newpage, PageTransHuge(newpage),
6736 memcg_check_events(memcg, newpage);
6737 local_irq_restore(flags);
6740 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6741 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6743 void mem_cgroup_sk_alloc(struct sock *sk)
6745 struct mem_cgroup *memcg;
6747 if (!mem_cgroup_sockets_enabled)
6750 /* Do not associate the sock with unrelated interrupted task's memcg. */
6755 memcg = mem_cgroup_from_task(current);
6756 if (memcg == root_mem_cgroup)
6758 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6760 if (css_tryget_online(&memcg->css))
6761 sk->sk_memcg = memcg;
6766 void mem_cgroup_sk_free(struct sock *sk)
6769 css_put(&sk->sk_memcg->css);
6773 * mem_cgroup_charge_skmem - charge socket memory
6774 * @memcg: memcg to charge
6775 * @nr_pages: number of pages to charge
6777 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6778 * @memcg's configured limit, %false if the charge had to be forced.
6780 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6782 gfp_t gfp_mask = GFP_KERNEL;
6784 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6785 struct page_counter *fail;
6787 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6788 memcg->tcpmem_pressure = 0;
6791 page_counter_charge(&memcg->tcpmem, nr_pages);
6792 memcg->tcpmem_pressure = 1;
6796 /* Don't block in the packet receive path */
6798 gfp_mask = GFP_NOWAIT;
6800 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6802 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6805 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6810 * mem_cgroup_uncharge_skmem - uncharge socket memory
6811 * @memcg: memcg to uncharge
6812 * @nr_pages: number of pages to uncharge
6814 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6816 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6817 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6821 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6823 refill_stock(memcg, nr_pages);
6826 static int __init cgroup_memory(char *s)
6830 while ((token = strsep(&s, ",")) != NULL) {
6833 if (!strcmp(token, "nosocket"))
6834 cgroup_memory_nosocket = true;
6835 if (!strcmp(token, "nokmem"))
6836 cgroup_memory_nokmem = true;
6840 __setup("cgroup.memory=", cgroup_memory);
6843 * subsys_initcall() for memory controller.
6845 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6846 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6847 * basically everything that doesn't depend on a specific mem_cgroup structure
6848 * should be initialized from here.
6850 static int __init mem_cgroup_init(void)
6854 #ifdef CONFIG_MEMCG_KMEM
6856 * Kmem cache creation is mostly done with the slab_mutex held,
6857 * so use a workqueue with limited concurrency to avoid stalling
6858 * all worker threads in case lots of cgroups are created and
6859 * destroyed simultaneously.
6861 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
6862 BUG_ON(!memcg_kmem_cache_wq);
6865 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6866 memcg_hotplug_cpu_dead);
6868 for_each_possible_cpu(cpu)
6869 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
6872 for_each_node(node) {
6873 struct mem_cgroup_tree_per_node *rtpn;
6875 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
6876 node_online(node) ? node : NUMA_NO_NODE);
6878 rtpn->rb_root = RB_ROOT;
6879 rtpn->rb_rightmost = NULL;
6880 spin_lock_init(&rtpn->lock);
6881 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6886 subsys_initcall(mem_cgroup_init);
6888 #ifdef CONFIG_MEMCG_SWAP
6889 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
6891 while (!refcount_inc_not_zero(&memcg->id.ref)) {
6893 * The root cgroup cannot be destroyed, so it's refcount must
6896 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
6900 memcg = parent_mem_cgroup(memcg);
6902 memcg = root_mem_cgroup;
6908 * mem_cgroup_swapout - transfer a memsw charge to swap
6909 * @page: page whose memsw charge to transfer
6910 * @entry: swap entry to move the charge to
6912 * Transfer the memsw charge of @page to @entry.
6914 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6916 struct mem_cgroup *memcg, *swap_memcg;
6917 unsigned int nr_entries;
6918 unsigned short oldid;
6920 VM_BUG_ON_PAGE(PageLRU(page), page);
6921 VM_BUG_ON_PAGE(page_count(page), page);
6923 if (!do_memsw_account())
6926 memcg = page->mem_cgroup;
6928 /* Readahead page, never charged */
6933 * In case the memcg owning these pages has been offlined and doesn't
6934 * have an ID allocated to it anymore, charge the closest online
6935 * ancestor for the swap instead and transfer the memory+swap charge.
6937 swap_memcg = mem_cgroup_id_get_online(memcg);
6938 nr_entries = hpage_nr_pages(page);
6939 /* Get references for the tail pages, too */
6941 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
6942 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
6944 VM_BUG_ON_PAGE(oldid, page);
6945 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
6947 page->mem_cgroup = NULL;
6949 if (!mem_cgroup_is_root(memcg))
6950 page_counter_uncharge(&memcg->memory, nr_entries);
6952 if (memcg != swap_memcg) {
6953 if (!mem_cgroup_is_root(swap_memcg))
6954 page_counter_charge(&swap_memcg->memsw, nr_entries);
6955 page_counter_uncharge(&memcg->memsw, nr_entries);
6959 * Interrupts should be disabled here because the caller holds the
6960 * i_pages lock which is taken with interrupts-off. It is
6961 * important here to have the interrupts disabled because it is the
6962 * only synchronisation we have for updating the per-CPU variables.
6964 VM_BUG_ON(!irqs_disabled());
6965 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
6967 memcg_check_events(memcg, page);
6969 if (!mem_cgroup_is_root(memcg))
6970 css_put_many(&memcg->css, nr_entries);
6974 * mem_cgroup_try_charge_swap - try charging swap space for a page
6975 * @page: page being added to swap
6976 * @entry: swap entry to charge
6978 * Try to charge @page's memcg for the swap space at @entry.
6980 * Returns 0 on success, -ENOMEM on failure.
6982 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
6984 unsigned int nr_pages = hpage_nr_pages(page);
6985 struct page_counter *counter;
6986 struct mem_cgroup *memcg;
6987 unsigned short oldid;
6989 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
6992 memcg = page->mem_cgroup;
6994 /* Readahead page, never charged */
6999 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7003 memcg = mem_cgroup_id_get_online(memcg);
7005 if (!mem_cgroup_is_root(memcg) &&
7006 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7007 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7008 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7009 mem_cgroup_id_put(memcg);
7013 /* Get references for the tail pages, too */
7015 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7016 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7017 VM_BUG_ON_PAGE(oldid, page);
7018 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7024 * mem_cgroup_uncharge_swap - uncharge swap space
7025 * @entry: swap entry to uncharge
7026 * @nr_pages: the amount of swap space to uncharge
7028 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7030 struct mem_cgroup *memcg;
7033 if (!do_swap_account)
7036 id = swap_cgroup_record(entry, 0, nr_pages);
7038 memcg = mem_cgroup_from_id(id);
7040 if (!mem_cgroup_is_root(memcg)) {
7041 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7042 page_counter_uncharge(&memcg->swap, nr_pages);
7044 page_counter_uncharge(&memcg->memsw, nr_pages);
7046 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7047 mem_cgroup_id_put_many(memcg, nr_pages);
7052 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7054 long nr_swap_pages = get_nr_swap_pages();
7056 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7057 return nr_swap_pages;
7058 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7059 nr_swap_pages = min_t(long, nr_swap_pages,
7060 READ_ONCE(memcg->swap.max) -
7061 page_counter_read(&memcg->swap));
7062 return nr_swap_pages;
7065 bool mem_cgroup_swap_full(struct page *page)
7067 struct mem_cgroup *memcg;
7069 VM_BUG_ON_PAGE(!PageLocked(page), page);
7073 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7076 memcg = page->mem_cgroup;
7080 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7081 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.max)
7087 /* for remember boot option*/
7088 #ifdef CONFIG_MEMCG_SWAP_ENABLED
7089 static int really_do_swap_account __initdata = 1;
7091 static int really_do_swap_account __initdata;
7094 static int __init enable_swap_account(char *s)
7096 if (!strcmp(s, "1"))
7097 really_do_swap_account = 1;
7098 else if (!strcmp(s, "0"))
7099 really_do_swap_account = 0;
7102 __setup("swapaccount=", enable_swap_account);
7104 static u64 swap_current_read(struct cgroup_subsys_state *css,
7107 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7109 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7112 static int swap_max_show(struct seq_file *m, void *v)
7114 return seq_puts_memcg_tunable(m,
7115 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7118 static ssize_t swap_max_write(struct kernfs_open_file *of,
7119 char *buf, size_t nbytes, loff_t off)
7121 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7125 buf = strstrip(buf);
7126 err = page_counter_memparse(buf, "max", &max);
7130 xchg(&memcg->swap.max, max);
7135 static int swap_events_show(struct seq_file *m, void *v)
7137 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7139 seq_printf(m, "max %lu\n",
7140 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7141 seq_printf(m, "fail %lu\n",
7142 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7147 static struct cftype swap_files[] = {
7149 .name = "swap.current",
7150 .flags = CFTYPE_NOT_ON_ROOT,
7151 .read_u64 = swap_current_read,
7155 .flags = CFTYPE_NOT_ON_ROOT,
7156 .seq_show = swap_max_show,
7157 .write = swap_max_write,
7160 .name = "swap.events",
7161 .flags = CFTYPE_NOT_ON_ROOT,
7162 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7163 .seq_show = swap_events_show,
7168 static struct cftype memsw_cgroup_files[] = {
7170 .name = "memsw.usage_in_bytes",
7171 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7172 .read_u64 = mem_cgroup_read_u64,
7175 .name = "memsw.max_usage_in_bytes",
7176 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7177 .write = mem_cgroup_reset,
7178 .read_u64 = mem_cgroup_read_u64,
7181 .name = "memsw.limit_in_bytes",
7182 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7183 .write = mem_cgroup_write,
7184 .read_u64 = mem_cgroup_read_u64,
7187 .name = "memsw.failcnt",
7188 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7189 .write = mem_cgroup_reset,
7190 .read_u64 = mem_cgroup_read_u64,
7192 { }, /* terminate */
7195 static int __init mem_cgroup_swap_init(void)
7197 if (!mem_cgroup_disabled() && really_do_swap_account) {
7198 do_swap_account = 1;
7199 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
7201 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
7202 memsw_cgroup_files));
7206 subsys_initcall(mem_cgroup_swap_init);
7208 #endif /* CONFIG_MEMCG_SWAP */