1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* memcontrol.c - Memory Controller
4 * Copyright IBM Corporation, 2007
5 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
7 * Copyright 2007 OpenVZ SWsoft Inc
8 * Author: Pavel Emelianov <xemul@openvz.org>
11 * Copyright (C) 2009 Nokia Corporation
12 * Author: Kirill A. Shutemov
14 * Kernel Memory Controller
15 * Copyright (C) 2012 Parallels Inc. and Google Inc.
16 * Authors: Glauber Costa and Suleiman Souhlal
19 * Charge lifetime sanitation
20 * Lockless page tracking & accounting
21 * Unified hierarchy configuration model
22 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
25 #include <linux/page_counter.h>
26 #include <linux/memcontrol.h>
27 #include <linux/cgroup.h>
28 #include <linux/pagewalk.h>
29 #include <linux/sched/mm.h>
30 #include <linux/shmem_fs.h>
31 #include <linux/hugetlb.h>
32 #include <linux/pagemap.h>
33 #include <linux/vm_event_item.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/swap_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
59 #include <linux/tracehook.h>
60 #include <linux/psi.h>
61 #include <linux/seq_buf.h>
67 #include <linux/uaccess.h>
69 #include <trace/events/vmscan.h>
71 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
72 EXPORT_SYMBOL(memory_cgrp_subsys);
74 struct mem_cgroup *root_mem_cgroup __read_mostly;
76 #define MEM_CGROUP_RECLAIM_RETRIES 5
78 /* Socket memory accounting disabled? */
79 static bool cgroup_memory_nosocket;
81 /* Kernel memory accounting disabled? */
82 static bool cgroup_memory_nokmem;
84 /* Whether the swap controller is active */
85 #ifdef CONFIG_MEMCG_SWAP
86 int do_swap_account __read_mostly;
88 #define do_swap_account 0
91 #ifdef CONFIG_CGROUP_WRITEBACK
92 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
95 /* Whether legacy memory+swap accounting is active */
96 static bool do_memsw_account(void)
98 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && do_swap_account;
101 #define THRESHOLDS_EVENTS_TARGET 128
102 #define SOFTLIMIT_EVENTS_TARGET 1024
105 * Cgroups above their limits are maintained in a RB-Tree, independent of
106 * their hierarchy representation
109 struct mem_cgroup_tree_per_node {
110 struct rb_root rb_root;
111 struct rb_node *rb_rightmost;
115 struct mem_cgroup_tree {
116 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
119 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
122 struct mem_cgroup_eventfd_list {
123 struct list_head list;
124 struct eventfd_ctx *eventfd;
128 * cgroup_event represents events which userspace want to receive.
130 struct mem_cgroup_event {
132 * memcg which the event belongs to.
134 struct mem_cgroup *memcg;
136 * eventfd to signal userspace about the event.
138 struct eventfd_ctx *eventfd;
140 * Each of these stored in a list by the cgroup.
142 struct list_head list;
144 * register_event() callback will be used to add new userspace
145 * waiter for changes related to this event. Use eventfd_signal()
146 * on eventfd to send notification to userspace.
148 int (*register_event)(struct mem_cgroup *memcg,
149 struct eventfd_ctx *eventfd, const char *args);
151 * unregister_event() callback will be called when userspace closes
152 * the eventfd or on cgroup removing. This callback must be set,
153 * if you want provide notification functionality.
155 void (*unregister_event)(struct mem_cgroup *memcg,
156 struct eventfd_ctx *eventfd);
158 * All fields below needed to unregister event when
159 * userspace closes eventfd.
162 wait_queue_head_t *wqh;
163 wait_queue_entry_t wait;
164 struct work_struct remove;
167 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
168 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
170 /* Stuffs for move charges at task migration. */
172 * Types of charges to be moved.
174 #define MOVE_ANON 0x1U
175 #define MOVE_FILE 0x2U
176 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
178 /* "mc" and its members are protected by cgroup_mutex */
179 static struct move_charge_struct {
180 spinlock_t lock; /* for from, to */
181 struct mm_struct *mm;
182 struct mem_cgroup *from;
183 struct mem_cgroup *to;
185 unsigned long precharge;
186 unsigned long moved_charge;
187 unsigned long moved_swap;
188 struct task_struct *moving_task; /* a task moving charges */
189 wait_queue_head_t waitq; /* a waitq for other context */
191 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
192 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
196 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
197 * limit reclaim to prevent infinite loops, if they ever occur.
199 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
200 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
203 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
204 MEM_CGROUP_CHARGE_TYPE_ANON,
205 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
206 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
210 /* for encoding cft->private value on file */
219 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
220 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
221 #define MEMFILE_ATTR(val) ((val) & 0xffff)
222 /* Used for OOM nofiier */
223 #define OOM_CONTROL (0)
226 * Iteration constructs for visiting all cgroups (under a tree). If
227 * loops are exited prematurely (break), mem_cgroup_iter_break() must
228 * be used for reference counting.
230 #define for_each_mem_cgroup_tree(iter, root) \
231 for (iter = mem_cgroup_iter(root, NULL, NULL); \
233 iter = mem_cgroup_iter(root, iter, NULL))
235 #define for_each_mem_cgroup(iter) \
236 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
238 iter = mem_cgroup_iter(NULL, iter, NULL))
240 static inline bool should_force_charge(void)
242 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
243 (current->flags & PF_EXITING);
246 /* Some nice accessors for the vmpressure. */
247 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
250 memcg = root_mem_cgroup;
251 return &memcg->vmpressure;
254 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
256 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
259 #ifdef CONFIG_MEMCG_KMEM
261 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
262 * The main reason for not using cgroup id for this:
263 * this works better in sparse environments, where we have a lot of memcgs,
264 * but only a few kmem-limited. Or also, if we have, for instance, 200
265 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
266 * 200 entry array for that.
268 * The current size of the caches array is stored in memcg_nr_cache_ids. It
269 * will double each time we have to increase it.
271 static DEFINE_IDA(memcg_cache_ida);
272 int memcg_nr_cache_ids;
274 /* Protects memcg_nr_cache_ids */
275 static DECLARE_RWSEM(memcg_cache_ids_sem);
277 void memcg_get_cache_ids(void)
279 down_read(&memcg_cache_ids_sem);
282 void memcg_put_cache_ids(void)
284 up_read(&memcg_cache_ids_sem);
288 * MIN_SIZE is different than 1, because we would like to avoid going through
289 * the alloc/free process all the time. In a small machine, 4 kmem-limited
290 * cgroups is a reasonable guess. In the future, it could be a parameter or
291 * tunable, but that is strictly not necessary.
293 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
294 * this constant directly from cgroup, but it is understandable that this is
295 * better kept as an internal representation in cgroup.c. In any case, the
296 * cgrp_id space is not getting any smaller, and we don't have to necessarily
297 * increase ours as well if it increases.
299 #define MEMCG_CACHES_MIN_SIZE 4
300 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
303 * A lot of the calls to the cache allocation functions are expected to be
304 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
305 * conditional to this static branch, we'll have to allow modules that does
306 * kmem_cache_alloc and the such to see this symbol as well
308 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
309 EXPORT_SYMBOL(memcg_kmem_enabled_key);
311 struct workqueue_struct *memcg_kmem_cache_wq;
314 static int memcg_shrinker_map_size;
315 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
317 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
319 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
322 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
323 int size, int old_size)
325 struct memcg_shrinker_map *new, *old;
328 lockdep_assert_held(&memcg_shrinker_map_mutex);
331 old = rcu_dereference_protected(
332 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
333 /* Not yet online memcg */
337 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
341 /* Set all old bits, clear all new bits */
342 memset(new->map, (int)0xff, old_size);
343 memset((void *)new->map + old_size, 0, size - old_size);
345 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
346 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
352 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
354 struct mem_cgroup_per_node *pn;
355 struct memcg_shrinker_map *map;
358 if (mem_cgroup_is_root(memcg))
362 pn = mem_cgroup_nodeinfo(memcg, nid);
363 map = rcu_dereference_protected(pn->shrinker_map, true);
366 rcu_assign_pointer(pn->shrinker_map, NULL);
370 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
372 struct memcg_shrinker_map *map;
373 int nid, size, ret = 0;
375 if (mem_cgroup_is_root(memcg))
378 mutex_lock(&memcg_shrinker_map_mutex);
379 size = memcg_shrinker_map_size;
381 map = kvzalloc_node(sizeof(*map) + size, GFP_KERNEL, nid);
383 memcg_free_shrinker_maps(memcg);
387 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
389 mutex_unlock(&memcg_shrinker_map_mutex);
394 int memcg_expand_shrinker_maps(int new_id)
396 int size, old_size, ret = 0;
397 struct mem_cgroup *memcg;
399 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
400 old_size = memcg_shrinker_map_size;
401 if (size <= old_size)
404 mutex_lock(&memcg_shrinker_map_mutex);
405 if (!root_mem_cgroup)
408 for_each_mem_cgroup(memcg) {
409 if (mem_cgroup_is_root(memcg))
411 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
413 mem_cgroup_iter_break(NULL, memcg);
419 memcg_shrinker_map_size = size;
420 mutex_unlock(&memcg_shrinker_map_mutex);
424 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
426 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
427 struct memcg_shrinker_map *map;
430 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
431 /* Pairs with smp mb in shrink_slab() */
432 smp_mb__before_atomic();
433 set_bit(shrinker_id, map->map);
439 * mem_cgroup_css_from_page - css of the memcg associated with a page
440 * @page: page of interest
442 * If memcg is bound to the default hierarchy, css of the memcg associated
443 * with @page is returned. The returned css remains associated with @page
444 * until it is released.
446 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
449 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
451 struct mem_cgroup *memcg;
453 memcg = page->mem_cgroup;
455 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
456 memcg = root_mem_cgroup;
462 * page_cgroup_ino - return inode number of the memcg a page is charged to
465 * Look up the closest online ancestor of the memory cgroup @page is charged to
466 * and return its inode number or 0 if @page is not charged to any cgroup. It
467 * is safe to call this function without holding a reference to @page.
469 * Note, this function is inherently racy, because there is nothing to prevent
470 * the cgroup inode from getting torn down and potentially reallocated a moment
471 * after page_cgroup_ino() returns, so it only should be used by callers that
472 * do not care (such as procfs interfaces).
474 ino_t page_cgroup_ino(struct page *page)
476 struct mem_cgroup *memcg;
477 unsigned long ino = 0;
480 if (PageSlab(page) && !PageTail(page))
481 memcg = memcg_from_slab_page(page);
483 memcg = READ_ONCE(page->mem_cgroup);
484 while (memcg && !(memcg->css.flags & CSS_ONLINE))
485 memcg = parent_mem_cgroup(memcg);
487 ino = cgroup_ino(memcg->css.cgroup);
492 static struct mem_cgroup_per_node *
493 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
495 int nid = page_to_nid(page);
497 return memcg->nodeinfo[nid];
500 static struct mem_cgroup_tree_per_node *
501 soft_limit_tree_node(int nid)
503 return soft_limit_tree.rb_tree_per_node[nid];
506 static struct mem_cgroup_tree_per_node *
507 soft_limit_tree_from_page(struct page *page)
509 int nid = page_to_nid(page);
511 return soft_limit_tree.rb_tree_per_node[nid];
514 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
515 struct mem_cgroup_tree_per_node *mctz,
516 unsigned long new_usage_in_excess)
518 struct rb_node **p = &mctz->rb_root.rb_node;
519 struct rb_node *parent = NULL;
520 struct mem_cgroup_per_node *mz_node;
521 bool rightmost = true;
526 mz->usage_in_excess = new_usage_in_excess;
527 if (!mz->usage_in_excess)
531 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
533 if (mz->usage_in_excess < mz_node->usage_in_excess) {
539 * We can't avoid mem cgroups that are over their soft
540 * limit by the same amount
542 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
547 mctz->rb_rightmost = &mz->tree_node;
549 rb_link_node(&mz->tree_node, parent, p);
550 rb_insert_color(&mz->tree_node, &mctz->rb_root);
554 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
555 struct mem_cgroup_tree_per_node *mctz)
560 if (&mz->tree_node == mctz->rb_rightmost)
561 mctz->rb_rightmost = rb_prev(&mz->tree_node);
563 rb_erase(&mz->tree_node, &mctz->rb_root);
567 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
568 struct mem_cgroup_tree_per_node *mctz)
572 spin_lock_irqsave(&mctz->lock, flags);
573 __mem_cgroup_remove_exceeded(mz, mctz);
574 spin_unlock_irqrestore(&mctz->lock, flags);
577 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
579 unsigned long nr_pages = page_counter_read(&memcg->memory);
580 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
581 unsigned long excess = 0;
583 if (nr_pages > soft_limit)
584 excess = nr_pages - soft_limit;
589 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
591 unsigned long excess;
592 struct mem_cgroup_per_node *mz;
593 struct mem_cgroup_tree_per_node *mctz;
595 mctz = soft_limit_tree_from_page(page);
599 * Necessary to update all ancestors when hierarchy is used.
600 * because their event counter is not touched.
602 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
603 mz = mem_cgroup_page_nodeinfo(memcg, page);
604 excess = soft_limit_excess(memcg);
606 * We have to update the tree if mz is on RB-tree or
607 * mem is over its softlimit.
609 if (excess || mz->on_tree) {
612 spin_lock_irqsave(&mctz->lock, flags);
613 /* if on-tree, remove it */
615 __mem_cgroup_remove_exceeded(mz, mctz);
617 * Insert again. mz->usage_in_excess will be updated.
618 * If excess is 0, no tree ops.
620 __mem_cgroup_insert_exceeded(mz, mctz, excess);
621 spin_unlock_irqrestore(&mctz->lock, flags);
626 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
628 struct mem_cgroup_tree_per_node *mctz;
629 struct mem_cgroup_per_node *mz;
633 mz = mem_cgroup_nodeinfo(memcg, nid);
634 mctz = soft_limit_tree_node(nid);
636 mem_cgroup_remove_exceeded(mz, mctz);
640 static struct mem_cgroup_per_node *
641 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
643 struct mem_cgroup_per_node *mz;
647 if (!mctz->rb_rightmost)
648 goto done; /* Nothing to reclaim from */
650 mz = rb_entry(mctz->rb_rightmost,
651 struct mem_cgroup_per_node, tree_node);
653 * Remove the node now but someone else can add it back,
654 * we will to add it back at the end of reclaim to its correct
655 * position in the tree.
657 __mem_cgroup_remove_exceeded(mz, mctz);
658 if (!soft_limit_excess(mz->memcg) ||
659 !css_tryget(&mz->memcg->css))
665 static struct mem_cgroup_per_node *
666 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
668 struct mem_cgroup_per_node *mz;
670 spin_lock_irq(&mctz->lock);
671 mz = __mem_cgroup_largest_soft_limit_node(mctz);
672 spin_unlock_irq(&mctz->lock);
677 * __mod_memcg_state - update cgroup memory statistics
678 * @memcg: the memory cgroup
679 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
680 * @val: delta to add to the counter, can be negative
682 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
686 if (mem_cgroup_disabled())
689 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
690 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
691 struct mem_cgroup *mi;
694 * Batch local counters to keep them in sync with
695 * the hierarchical ones.
697 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
698 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
699 atomic_long_add(x, &mi->vmstats[idx]);
702 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
705 static struct mem_cgroup_per_node *
706 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
708 struct mem_cgroup *parent;
710 parent = parent_mem_cgroup(pn->memcg);
713 return mem_cgroup_nodeinfo(parent, nid);
717 * __mod_lruvec_state - update lruvec memory statistics
718 * @lruvec: the lruvec
719 * @idx: the stat item
720 * @val: delta to add to the counter, can be negative
722 * The lruvec is the intersection of the NUMA node and a cgroup. This
723 * function updates the all three counters that are affected by a
724 * change of state at this level: per-node, per-cgroup, per-lruvec.
726 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
729 pg_data_t *pgdat = lruvec_pgdat(lruvec);
730 struct mem_cgroup_per_node *pn;
731 struct mem_cgroup *memcg;
735 __mod_node_page_state(pgdat, idx, val);
737 if (mem_cgroup_disabled())
740 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
744 __mod_memcg_state(memcg, idx, val);
747 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
749 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
750 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
751 struct mem_cgroup_per_node *pi;
753 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
754 atomic_long_add(x, &pi->lruvec_stat[idx]);
757 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
760 void __mod_lruvec_slab_state(void *p, enum node_stat_item idx, int val)
762 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
763 struct mem_cgroup *memcg;
764 struct lruvec *lruvec;
767 memcg = mem_cgroup_from_obj(p);
769 /* Untracked pages have no memcg, no lruvec. Update only the node */
770 if (!memcg || memcg == root_mem_cgroup) {
771 __mod_node_page_state(pgdat, idx, val);
773 lruvec = mem_cgroup_lruvec(memcg, pgdat);
774 __mod_lruvec_state(lruvec, idx, val);
779 void mod_memcg_obj_state(void *p, int idx, int val)
781 struct mem_cgroup *memcg;
784 memcg = mem_cgroup_from_obj(p);
786 mod_memcg_state(memcg, idx, val);
791 * __count_memcg_events - account VM events in a cgroup
792 * @memcg: the memory cgroup
793 * @idx: the event item
794 * @count: the number of events that occured
796 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
801 if (mem_cgroup_disabled())
804 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
805 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
806 struct mem_cgroup *mi;
809 * Batch local counters to keep them in sync with
810 * the hierarchical ones.
812 __this_cpu_add(memcg->vmstats_local->events[idx], x);
813 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
814 atomic_long_add(x, &mi->vmevents[idx]);
817 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
820 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
822 return atomic_long_read(&memcg->vmevents[event]);
825 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
830 for_each_possible_cpu(cpu)
831 x += per_cpu(memcg->vmstats_local->events[event], cpu);
835 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
837 bool compound, int nr_pages)
840 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
841 * counted as CACHE even if it's on ANON LRU.
844 __mod_memcg_state(memcg, MEMCG_RSS, nr_pages);
846 __mod_memcg_state(memcg, MEMCG_CACHE, nr_pages);
847 if (PageSwapBacked(page))
848 __mod_memcg_state(memcg, NR_SHMEM, nr_pages);
852 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
853 __mod_memcg_state(memcg, MEMCG_RSS_HUGE, nr_pages);
856 /* pagein of a big page is an event. So, ignore page size */
858 __count_memcg_events(memcg, PGPGIN, 1);
860 __count_memcg_events(memcg, PGPGOUT, 1);
861 nr_pages = -nr_pages; /* for event */
864 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
867 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
868 enum mem_cgroup_events_target target)
870 unsigned long val, next;
872 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
873 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
874 /* from time_after() in jiffies.h */
875 if ((long)(next - val) < 0) {
877 case MEM_CGROUP_TARGET_THRESH:
878 next = val + THRESHOLDS_EVENTS_TARGET;
880 case MEM_CGROUP_TARGET_SOFTLIMIT:
881 next = val + SOFTLIMIT_EVENTS_TARGET;
886 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
893 * Check events in order.
896 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
898 /* threshold event is triggered in finer grain than soft limit */
899 if (unlikely(mem_cgroup_event_ratelimit(memcg,
900 MEM_CGROUP_TARGET_THRESH))) {
903 do_softlimit = mem_cgroup_event_ratelimit(memcg,
904 MEM_CGROUP_TARGET_SOFTLIMIT);
905 mem_cgroup_threshold(memcg);
906 if (unlikely(do_softlimit))
907 mem_cgroup_update_tree(memcg, page);
911 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
914 * mm_update_next_owner() may clear mm->owner to NULL
915 * if it races with swapoff, page migration, etc.
916 * So this can be called with p == NULL.
921 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
923 EXPORT_SYMBOL(mem_cgroup_from_task);
926 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
927 * @mm: mm from which memcg should be extracted. It can be NULL.
929 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
930 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
933 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
935 struct mem_cgroup *memcg;
937 if (mem_cgroup_disabled())
943 * Page cache insertions can happen withou an
944 * actual mm context, e.g. during disk probing
945 * on boot, loopback IO, acct() writes etc.
948 memcg = root_mem_cgroup;
950 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
951 if (unlikely(!memcg))
952 memcg = root_mem_cgroup;
954 } while (!css_tryget(&memcg->css));
958 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
961 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
962 * @page: page from which memcg should be extracted.
964 * Obtain a reference on page->memcg and returns it if successful. Otherwise
965 * root_mem_cgroup is returned.
967 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
969 struct mem_cgroup *memcg = page->mem_cgroup;
971 if (mem_cgroup_disabled())
975 /* Page should not get uncharged and freed memcg under us. */
976 if (!memcg || WARN_ON_ONCE(!css_tryget(&memcg->css)))
977 memcg = root_mem_cgroup;
981 EXPORT_SYMBOL(get_mem_cgroup_from_page);
984 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
986 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
988 if (unlikely(current->active_memcg)) {
989 struct mem_cgroup *memcg;
992 /* current->active_memcg must hold a ref. */
993 if (WARN_ON_ONCE(!css_tryget(¤t->active_memcg->css)))
994 memcg = root_mem_cgroup;
996 memcg = current->active_memcg;
1000 return get_mem_cgroup_from_mm(current->mm);
1004 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1005 * @root: hierarchy root
1006 * @prev: previously returned memcg, NULL on first invocation
1007 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1009 * Returns references to children of the hierarchy below @root, or
1010 * @root itself, or %NULL after a full round-trip.
1012 * Caller must pass the return value in @prev on subsequent
1013 * invocations for reference counting, or use mem_cgroup_iter_break()
1014 * to cancel a hierarchy walk before the round-trip is complete.
1016 * Reclaimers can specify a node and a priority level in @reclaim to
1017 * divide up the memcgs in the hierarchy among all concurrent
1018 * reclaimers operating on the same node and priority.
1020 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1021 struct mem_cgroup *prev,
1022 struct mem_cgroup_reclaim_cookie *reclaim)
1024 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1025 struct cgroup_subsys_state *css = NULL;
1026 struct mem_cgroup *memcg = NULL;
1027 struct mem_cgroup *pos = NULL;
1029 if (mem_cgroup_disabled())
1033 root = root_mem_cgroup;
1035 if (prev && !reclaim)
1038 if (!root->use_hierarchy && root != root_mem_cgroup) {
1047 struct mem_cgroup_per_node *mz;
1049 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1052 if (prev && reclaim->generation != iter->generation)
1056 pos = READ_ONCE(iter->position);
1057 if (!pos || css_tryget(&pos->css))
1060 * css reference reached zero, so iter->position will
1061 * be cleared by ->css_released. However, we should not
1062 * rely on this happening soon, because ->css_released
1063 * is called from a work queue, and by busy-waiting we
1064 * might block it. So we clear iter->position right
1067 (void)cmpxchg(&iter->position, pos, NULL);
1075 css = css_next_descendant_pre(css, &root->css);
1078 * Reclaimers share the hierarchy walk, and a
1079 * new one might jump in right at the end of
1080 * the hierarchy - make sure they see at least
1081 * one group and restart from the beginning.
1089 * Verify the css and acquire a reference. The root
1090 * is provided by the caller, so we know it's alive
1091 * and kicking, and don't take an extra reference.
1093 memcg = mem_cgroup_from_css(css);
1095 if (css == &root->css)
1098 if (css_tryget(css))
1106 * The position could have already been updated by a competing
1107 * thread, so check that the value hasn't changed since we read
1108 * it to avoid reclaiming from the same cgroup twice.
1110 (void)cmpxchg(&iter->position, pos, memcg);
1118 reclaim->generation = iter->generation;
1124 if (prev && prev != root)
1125 css_put(&prev->css);
1131 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1132 * @root: hierarchy root
1133 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1135 void mem_cgroup_iter_break(struct mem_cgroup *root,
1136 struct mem_cgroup *prev)
1139 root = root_mem_cgroup;
1140 if (prev && prev != root)
1141 css_put(&prev->css);
1144 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1145 struct mem_cgroup *dead_memcg)
1147 struct mem_cgroup_reclaim_iter *iter;
1148 struct mem_cgroup_per_node *mz;
1151 for_each_node(nid) {
1152 mz = mem_cgroup_nodeinfo(from, nid);
1154 cmpxchg(&iter->position, dead_memcg, NULL);
1158 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1160 struct mem_cgroup *memcg = dead_memcg;
1161 struct mem_cgroup *last;
1164 __invalidate_reclaim_iterators(memcg, dead_memcg);
1166 } while ((memcg = parent_mem_cgroup(memcg)));
1169 * When cgruop1 non-hierarchy mode is used,
1170 * parent_mem_cgroup() does not walk all the way up to the
1171 * cgroup root (root_mem_cgroup). So we have to handle
1172 * dead_memcg from cgroup root separately.
1174 if (last != root_mem_cgroup)
1175 __invalidate_reclaim_iterators(root_mem_cgroup,
1180 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1181 * @memcg: hierarchy root
1182 * @fn: function to call for each task
1183 * @arg: argument passed to @fn
1185 * This function iterates over tasks attached to @memcg or to any of its
1186 * descendants and calls @fn for each task. If @fn returns a non-zero
1187 * value, the function breaks the iteration loop and returns the value.
1188 * Otherwise, it will iterate over all tasks and return 0.
1190 * This function must not be called for the root memory cgroup.
1192 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1193 int (*fn)(struct task_struct *, void *), void *arg)
1195 struct mem_cgroup *iter;
1198 BUG_ON(memcg == root_mem_cgroup);
1200 for_each_mem_cgroup_tree(iter, memcg) {
1201 struct css_task_iter it;
1202 struct task_struct *task;
1204 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1205 while (!ret && (task = css_task_iter_next(&it)))
1206 ret = fn(task, arg);
1207 css_task_iter_end(&it);
1209 mem_cgroup_iter_break(memcg, iter);
1217 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1219 * @pgdat: pgdat of the page
1221 * This function is only safe when following the LRU page isolation
1222 * and putback protocol: the LRU lock must be held, and the page must
1223 * either be PageLRU() or the caller must have isolated/allocated it.
1225 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1227 struct mem_cgroup_per_node *mz;
1228 struct mem_cgroup *memcg;
1229 struct lruvec *lruvec;
1231 if (mem_cgroup_disabled()) {
1232 lruvec = &pgdat->__lruvec;
1236 memcg = page->mem_cgroup;
1238 * Swapcache readahead pages are added to the LRU - and
1239 * possibly migrated - before they are charged.
1242 memcg = root_mem_cgroup;
1244 mz = mem_cgroup_page_nodeinfo(memcg, page);
1245 lruvec = &mz->lruvec;
1248 * Since a node can be onlined after the mem_cgroup was created,
1249 * we have to be prepared to initialize lruvec->zone here;
1250 * and if offlined then reonlined, we need to reinitialize it.
1252 if (unlikely(lruvec->pgdat != pgdat))
1253 lruvec->pgdat = pgdat;
1258 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1259 * @lruvec: mem_cgroup per zone lru vector
1260 * @lru: index of lru list the page is sitting on
1261 * @zid: zone id of the accounted pages
1262 * @nr_pages: positive when adding or negative when removing
1264 * This function must be called under lru_lock, just before a page is added
1265 * to or just after a page is removed from an lru list (that ordering being
1266 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1268 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1269 int zid, int nr_pages)
1271 struct mem_cgroup_per_node *mz;
1272 unsigned long *lru_size;
1275 if (mem_cgroup_disabled())
1278 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1279 lru_size = &mz->lru_zone_size[zid][lru];
1282 *lru_size += nr_pages;
1285 if (WARN_ONCE(size < 0,
1286 "%s(%p, %d, %d): lru_size %ld\n",
1287 __func__, lruvec, lru, nr_pages, size)) {
1293 *lru_size += nr_pages;
1297 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1298 * @memcg: the memory cgroup
1300 * Returns the maximum amount of memory @mem can be charged with, in
1303 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1305 unsigned long margin = 0;
1306 unsigned long count;
1307 unsigned long limit;
1309 count = page_counter_read(&memcg->memory);
1310 limit = READ_ONCE(memcg->memory.max);
1312 margin = limit - count;
1314 if (do_memsw_account()) {
1315 count = page_counter_read(&memcg->memsw);
1316 limit = READ_ONCE(memcg->memsw.max);
1318 margin = min(margin, limit - count);
1327 * A routine for checking "mem" is under move_account() or not.
1329 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1330 * moving cgroups. This is for waiting at high-memory pressure
1333 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1335 struct mem_cgroup *from;
1336 struct mem_cgroup *to;
1339 * Unlike task_move routines, we access mc.to, mc.from not under
1340 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1342 spin_lock(&mc.lock);
1348 ret = mem_cgroup_is_descendant(from, memcg) ||
1349 mem_cgroup_is_descendant(to, memcg);
1351 spin_unlock(&mc.lock);
1355 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1357 if (mc.moving_task && current != mc.moving_task) {
1358 if (mem_cgroup_under_move(memcg)) {
1360 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1361 /* moving charge context might have finished. */
1364 finish_wait(&mc.waitq, &wait);
1371 static char *memory_stat_format(struct mem_cgroup *memcg)
1376 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1381 * Provide statistics on the state of the memory subsystem as
1382 * well as cumulative event counters that show past behavior.
1384 * This list is ordered following a combination of these gradients:
1385 * 1) generic big picture -> specifics and details
1386 * 2) reflecting userspace activity -> reflecting kernel heuristics
1388 * Current memory state:
1391 seq_buf_printf(&s, "anon %llu\n",
1392 (u64)memcg_page_state(memcg, MEMCG_RSS) *
1394 seq_buf_printf(&s, "file %llu\n",
1395 (u64)memcg_page_state(memcg, MEMCG_CACHE) *
1397 seq_buf_printf(&s, "kernel_stack %llu\n",
1398 (u64)memcg_page_state(memcg, MEMCG_KERNEL_STACK_KB) *
1400 seq_buf_printf(&s, "slab %llu\n",
1401 (u64)(memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) +
1402 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE)) *
1404 seq_buf_printf(&s, "sock %llu\n",
1405 (u64)memcg_page_state(memcg, MEMCG_SOCK) *
1408 seq_buf_printf(&s, "shmem %llu\n",
1409 (u64)memcg_page_state(memcg, NR_SHMEM) *
1411 seq_buf_printf(&s, "file_mapped %llu\n",
1412 (u64)memcg_page_state(memcg, NR_FILE_MAPPED) *
1414 seq_buf_printf(&s, "file_dirty %llu\n",
1415 (u64)memcg_page_state(memcg, NR_FILE_DIRTY) *
1417 seq_buf_printf(&s, "file_writeback %llu\n",
1418 (u64)memcg_page_state(memcg, NR_WRITEBACK) *
1422 * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter
1423 * with the NR_ANON_THP vm counter, but right now it's a pain in the
1424 * arse because it requires migrating the work out of rmap to a place
1425 * where the page->mem_cgroup is set up and stable.
1427 seq_buf_printf(&s, "anon_thp %llu\n",
1428 (u64)memcg_page_state(memcg, MEMCG_RSS_HUGE) *
1431 for (i = 0; i < NR_LRU_LISTS; i++)
1432 seq_buf_printf(&s, "%s %llu\n", lru_list_name(i),
1433 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
1436 seq_buf_printf(&s, "slab_reclaimable %llu\n",
1437 (u64)memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) *
1439 seq_buf_printf(&s, "slab_unreclaimable %llu\n",
1440 (u64)memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE) *
1443 /* Accumulated memory events */
1445 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1446 memcg_events(memcg, PGFAULT));
1447 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1448 memcg_events(memcg, PGMAJFAULT));
1450 seq_buf_printf(&s, "workingset_refault %lu\n",
1451 memcg_page_state(memcg, WORKINGSET_REFAULT));
1452 seq_buf_printf(&s, "workingset_activate %lu\n",
1453 memcg_page_state(memcg, WORKINGSET_ACTIVATE));
1454 seq_buf_printf(&s, "workingset_restore %lu\n",
1455 memcg_page_state(memcg, WORKINGSET_RESTORE));
1456 seq_buf_printf(&s, "workingset_nodereclaim %lu\n",
1457 memcg_page_state(memcg, WORKINGSET_NODERECLAIM));
1459 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1460 memcg_events(memcg, PGREFILL));
1461 seq_buf_printf(&s, "pgscan %lu\n",
1462 memcg_events(memcg, PGSCAN_KSWAPD) +
1463 memcg_events(memcg, PGSCAN_DIRECT));
1464 seq_buf_printf(&s, "pgsteal %lu\n",
1465 memcg_events(memcg, PGSTEAL_KSWAPD) +
1466 memcg_events(memcg, PGSTEAL_DIRECT));
1467 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1468 memcg_events(memcg, PGACTIVATE));
1469 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1470 memcg_events(memcg, PGDEACTIVATE));
1471 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1472 memcg_events(memcg, PGLAZYFREE));
1473 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1474 memcg_events(memcg, PGLAZYFREED));
1476 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1477 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1478 memcg_events(memcg, THP_FAULT_ALLOC));
1479 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1480 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1481 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1483 /* The above should easily fit into one page */
1484 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1489 #define K(x) ((x) << (PAGE_SHIFT-10))
1491 * mem_cgroup_print_oom_context: Print OOM information relevant to
1492 * memory controller.
1493 * @memcg: The memory cgroup that went over limit
1494 * @p: Task that is going to be killed
1496 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1499 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1504 pr_cont(",oom_memcg=");
1505 pr_cont_cgroup_path(memcg->css.cgroup);
1507 pr_cont(",global_oom");
1509 pr_cont(",task_memcg=");
1510 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1516 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1517 * memory controller.
1518 * @memcg: The memory cgroup that went over limit
1520 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1524 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1525 K((u64)page_counter_read(&memcg->memory)),
1526 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1527 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1528 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1529 K((u64)page_counter_read(&memcg->swap)),
1530 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1532 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1533 K((u64)page_counter_read(&memcg->memsw)),
1534 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1535 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1536 K((u64)page_counter_read(&memcg->kmem)),
1537 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1540 pr_info("Memory cgroup stats for ");
1541 pr_cont_cgroup_path(memcg->css.cgroup);
1543 buf = memory_stat_format(memcg);
1551 * Return the memory (and swap, if configured) limit for a memcg.
1553 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1557 max = READ_ONCE(memcg->memory.max);
1558 if (mem_cgroup_swappiness(memcg)) {
1559 unsigned long memsw_max;
1560 unsigned long swap_max;
1562 memsw_max = memcg->memsw.max;
1563 swap_max = READ_ONCE(memcg->swap.max);
1564 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1565 max = min(max + swap_max, memsw_max);
1570 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1572 return page_counter_read(&memcg->memory);
1575 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1578 struct oom_control oc = {
1582 .gfp_mask = gfp_mask,
1587 if (mutex_lock_killable(&oom_lock))
1590 * A few threads which were not waiting at mutex_lock_killable() can
1591 * fail to bail out. Therefore, check again after holding oom_lock.
1593 ret = should_force_charge() || out_of_memory(&oc);
1594 mutex_unlock(&oom_lock);
1598 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1601 unsigned long *total_scanned)
1603 struct mem_cgroup *victim = NULL;
1606 unsigned long excess;
1607 unsigned long nr_scanned;
1608 struct mem_cgroup_reclaim_cookie reclaim = {
1612 excess = soft_limit_excess(root_memcg);
1615 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1620 * If we have not been able to reclaim
1621 * anything, it might because there are
1622 * no reclaimable pages under this hierarchy
1627 * We want to do more targeted reclaim.
1628 * excess >> 2 is not to excessive so as to
1629 * reclaim too much, nor too less that we keep
1630 * coming back to reclaim from this cgroup
1632 if (total >= (excess >> 2) ||
1633 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1638 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1639 pgdat, &nr_scanned);
1640 *total_scanned += nr_scanned;
1641 if (!soft_limit_excess(root_memcg))
1644 mem_cgroup_iter_break(root_memcg, victim);
1648 #ifdef CONFIG_LOCKDEP
1649 static struct lockdep_map memcg_oom_lock_dep_map = {
1650 .name = "memcg_oom_lock",
1654 static DEFINE_SPINLOCK(memcg_oom_lock);
1657 * Check OOM-Killer is already running under our hierarchy.
1658 * If someone is running, return false.
1660 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1662 struct mem_cgroup *iter, *failed = NULL;
1664 spin_lock(&memcg_oom_lock);
1666 for_each_mem_cgroup_tree(iter, memcg) {
1667 if (iter->oom_lock) {
1669 * this subtree of our hierarchy is already locked
1670 * so we cannot give a lock.
1673 mem_cgroup_iter_break(memcg, iter);
1676 iter->oom_lock = true;
1681 * OK, we failed to lock the whole subtree so we have
1682 * to clean up what we set up to the failing subtree
1684 for_each_mem_cgroup_tree(iter, memcg) {
1685 if (iter == failed) {
1686 mem_cgroup_iter_break(memcg, iter);
1689 iter->oom_lock = false;
1692 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1694 spin_unlock(&memcg_oom_lock);
1699 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1701 struct mem_cgroup *iter;
1703 spin_lock(&memcg_oom_lock);
1704 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1705 for_each_mem_cgroup_tree(iter, memcg)
1706 iter->oom_lock = false;
1707 spin_unlock(&memcg_oom_lock);
1710 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1712 struct mem_cgroup *iter;
1714 spin_lock(&memcg_oom_lock);
1715 for_each_mem_cgroup_tree(iter, memcg)
1717 spin_unlock(&memcg_oom_lock);
1720 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1722 struct mem_cgroup *iter;
1725 * When a new child is created while the hierarchy is under oom,
1726 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1728 spin_lock(&memcg_oom_lock);
1729 for_each_mem_cgroup_tree(iter, memcg)
1730 if (iter->under_oom > 0)
1732 spin_unlock(&memcg_oom_lock);
1735 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1737 struct oom_wait_info {
1738 struct mem_cgroup *memcg;
1739 wait_queue_entry_t wait;
1742 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1743 unsigned mode, int sync, void *arg)
1745 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1746 struct mem_cgroup *oom_wait_memcg;
1747 struct oom_wait_info *oom_wait_info;
1749 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1750 oom_wait_memcg = oom_wait_info->memcg;
1752 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1753 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1755 return autoremove_wake_function(wait, mode, sync, arg);
1758 static void memcg_oom_recover(struct mem_cgroup *memcg)
1761 * For the following lockless ->under_oom test, the only required
1762 * guarantee is that it must see the state asserted by an OOM when
1763 * this function is called as a result of userland actions
1764 * triggered by the notification of the OOM. This is trivially
1765 * achieved by invoking mem_cgroup_mark_under_oom() before
1766 * triggering notification.
1768 if (memcg && memcg->under_oom)
1769 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1779 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1781 enum oom_status ret;
1784 if (order > PAGE_ALLOC_COSTLY_ORDER)
1787 memcg_memory_event(memcg, MEMCG_OOM);
1790 * We are in the middle of the charge context here, so we
1791 * don't want to block when potentially sitting on a callstack
1792 * that holds all kinds of filesystem and mm locks.
1794 * cgroup1 allows disabling the OOM killer and waiting for outside
1795 * handling until the charge can succeed; remember the context and put
1796 * the task to sleep at the end of the page fault when all locks are
1799 * On the other hand, in-kernel OOM killer allows for an async victim
1800 * memory reclaim (oom_reaper) and that means that we are not solely
1801 * relying on the oom victim to make a forward progress and we can
1802 * invoke the oom killer here.
1804 * Please note that mem_cgroup_out_of_memory might fail to find a
1805 * victim and then we have to bail out from the charge path.
1807 if (memcg->oom_kill_disable) {
1808 if (!current->in_user_fault)
1810 css_get(&memcg->css);
1811 current->memcg_in_oom = memcg;
1812 current->memcg_oom_gfp_mask = mask;
1813 current->memcg_oom_order = order;
1818 mem_cgroup_mark_under_oom(memcg);
1820 locked = mem_cgroup_oom_trylock(memcg);
1823 mem_cgroup_oom_notify(memcg);
1825 mem_cgroup_unmark_under_oom(memcg);
1826 if (mem_cgroup_out_of_memory(memcg, mask, order))
1832 mem_cgroup_oom_unlock(memcg);
1838 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1839 * @handle: actually kill/wait or just clean up the OOM state
1841 * This has to be called at the end of a page fault if the memcg OOM
1842 * handler was enabled.
1844 * Memcg supports userspace OOM handling where failed allocations must
1845 * sleep on a waitqueue until the userspace task resolves the
1846 * situation. Sleeping directly in the charge context with all kinds
1847 * of locks held is not a good idea, instead we remember an OOM state
1848 * in the task and mem_cgroup_oom_synchronize() has to be called at
1849 * the end of the page fault to complete the OOM handling.
1851 * Returns %true if an ongoing memcg OOM situation was detected and
1852 * completed, %false otherwise.
1854 bool mem_cgroup_oom_synchronize(bool handle)
1856 struct mem_cgroup *memcg = current->memcg_in_oom;
1857 struct oom_wait_info owait;
1860 /* OOM is global, do not handle */
1867 owait.memcg = memcg;
1868 owait.wait.flags = 0;
1869 owait.wait.func = memcg_oom_wake_function;
1870 owait.wait.private = current;
1871 INIT_LIST_HEAD(&owait.wait.entry);
1873 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1874 mem_cgroup_mark_under_oom(memcg);
1876 locked = mem_cgroup_oom_trylock(memcg);
1879 mem_cgroup_oom_notify(memcg);
1881 if (locked && !memcg->oom_kill_disable) {
1882 mem_cgroup_unmark_under_oom(memcg);
1883 finish_wait(&memcg_oom_waitq, &owait.wait);
1884 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1885 current->memcg_oom_order);
1888 mem_cgroup_unmark_under_oom(memcg);
1889 finish_wait(&memcg_oom_waitq, &owait.wait);
1893 mem_cgroup_oom_unlock(memcg);
1895 * There is no guarantee that an OOM-lock contender
1896 * sees the wakeups triggered by the OOM kill
1897 * uncharges. Wake any sleepers explicitely.
1899 memcg_oom_recover(memcg);
1902 current->memcg_in_oom = NULL;
1903 css_put(&memcg->css);
1908 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1909 * @victim: task to be killed by the OOM killer
1910 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1912 * Returns a pointer to a memory cgroup, which has to be cleaned up
1913 * by killing all belonging OOM-killable tasks.
1915 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1917 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1918 struct mem_cgroup *oom_domain)
1920 struct mem_cgroup *oom_group = NULL;
1921 struct mem_cgroup *memcg;
1923 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
1927 oom_domain = root_mem_cgroup;
1931 memcg = mem_cgroup_from_task(victim);
1932 if (memcg == root_mem_cgroup)
1936 * If the victim task has been asynchronously moved to a different
1937 * memory cgroup, we might end up killing tasks outside oom_domain.
1938 * In this case it's better to ignore memory.group.oom.
1940 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
1944 * Traverse the memory cgroup hierarchy from the victim task's
1945 * cgroup up to the OOMing cgroup (or root) to find the
1946 * highest-level memory cgroup with oom.group set.
1948 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1949 if (memcg->oom_group)
1952 if (memcg == oom_domain)
1957 css_get(&oom_group->css);
1964 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
1966 pr_info("Tasks in ");
1967 pr_cont_cgroup_path(memcg->css.cgroup);
1968 pr_cont(" are going to be killed due to memory.oom.group set\n");
1972 * lock_page_memcg - lock a page->mem_cgroup binding
1975 * This function protects unlocked LRU pages from being moved to
1978 * It ensures lifetime of the returned memcg. Caller is responsible
1979 * for the lifetime of the page; __unlock_page_memcg() is available
1980 * when @page might get freed inside the locked section.
1982 struct mem_cgroup *lock_page_memcg(struct page *page)
1984 struct mem_cgroup *memcg;
1985 unsigned long flags;
1988 * The RCU lock is held throughout the transaction. The fast
1989 * path can get away without acquiring the memcg->move_lock
1990 * because page moving starts with an RCU grace period.
1992 * The RCU lock also protects the memcg from being freed when
1993 * the page state that is going to change is the only thing
1994 * preventing the page itself from being freed. E.g. writeback
1995 * doesn't hold a page reference and relies on PG_writeback to
1996 * keep off truncation, migration and so forth.
2000 if (mem_cgroup_disabled())
2003 memcg = page->mem_cgroup;
2004 if (unlikely(!memcg))
2007 if (atomic_read(&memcg->moving_account) <= 0)
2010 spin_lock_irqsave(&memcg->move_lock, flags);
2011 if (memcg != page->mem_cgroup) {
2012 spin_unlock_irqrestore(&memcg->move_lock, flags);
2017 * When charge migration first begins, we can have locked and
2018 * unlocked page stat updates happening concurrently. Track
2019 * the task who has the lock for unlock_page_memcg().
2021 memcg->move_lock_task = current;
2022 memcg->move_lock_flags = flags;
2026 EXPORT_SYMBOL(lock_page_memcg);
2029 * __unlock_page_memcg - unlock and unpin a memcg
2032 * Unlock and unpin a memcg returned by lock_page_memcg().
2034 void __unlock_page_memcg(struct mem_cgroup *memcg)
2036 if (memcg && memcg->move_lock_task == current) {
2037 unsigned long flags = memcg->move_lock_flags;
2039 memcg->move_lock_task = NULL;
2040 memcg->move_lock_flags = 0;
2042 spin_unlock_irqrestore(&memcg->move_lock, flags);
2049 * unlock_page_memcg - unlock a page->mem_cgroup binding
2052 void unlock_page_memcg(struct page *page)
2054 __unlock_page_memcg(page->mem_cgroup);
2056 EXPORT_SYMBOL(unlock_page_memcg);
2058 struct memcg_stock_pcp {
2059 struct mem_cgroup *cached; /* this never be root cgroup */
2060 unsigned int nr_pages;
2061 struct work_struct work;
2062 unsigned long flags;
2063 #define FLUSHING_CACHED_CHARGE 0
2065 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2066 static DEFINE_MUTEX(percpu_charge_mutex);
2069 * consume_stock: Try to consume stocked charge on this cpu.
2070 * @memcg: memcg to consume from.
2071 * @nr_pages: how many pages to charge.
2073 * The charges will only happen if @memcg matches the current cpu's memcg
2074 * stock, and at least @nr_pages are available in that stock. Failure to
2075 * service an allocation will refill the stock.
2077 * returns true if successful, false otherwise.
2079 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2081 struct memcg_stock_pcp *stock;
2082 unsigned long flags;
2085 if (nr_pages > MEMCG_CHARGE_BATCH)
2088 local_irq_save(flags);
2090 stock = this_cpu_ptr(&memcg_stock);
2091 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2092 stock->nr_pages -= nr_pages;
2096 local_irq_restore(flags);
2102 * Returns stocks cached in percpu and reset cached information.
2104 static void drain_stock(struct memcg_stock_pcp *stock)
2106 struct mem_cgroup *old = stock->cached;
2108 if (stock->nr_pages) {
2109 page_counter_uncharge(&old->memory, stock->nr_pages);
2110 if (do_memsw_account())
2111 page_counter_uncharge(&old->memsw, stock->nr_pages);
2112 css_put_many(&old->css, stock->nr_pages);
2113 stock->nr_pages = 0;
2115 stock->cached = NULL;
2118 static void drain_local_stock(struct work_struct *dummy)
2120 struct memcg_stock_pcp *stock;
2121 unsigned long flags;
2124 * The only protection from memory hotplug vs. drain_stock races is
2125 * that we always operate on local CPU stock here with IRQ disabled
2127 local_irq_save(flags);
2129 stock = this_cpu_ptr(&memcg_stock);
2131 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2133 local_irq_restore(flags);
2137 * Cache charges(val) to local per_cpu area.
2138 * This will be consumed by consume_stock() function, later.
2140 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2142 struct memcg_stock_pcp *stock;
2143 unsigned long flags;
2145 local_irq_save(flags);
2147 stock = this_cpu_ptr(&memcg_stock);
2148 if (stock->cached != memcg) { /* reset if necessary */
2150 stock->cached = memcg;
2152 stock->nr_pages += nr_pages;
2154 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2157 local_irq_restore(flags);
2161 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2162 * of the hierarchy under it.
2164 static void drain_all_stock(struct mem_cgroup *root_memcg)
2168 /* If someone's already draining, avoid adding running more workers. */
2169 if (!mutex_trylock(&percpu_charge_mutex))
2172 * Notify other cpus that system-wide "drain" is running
2173 * We do not care about races with the cpu hotplug because cpu down
2174 * as well as workers from this path always operate on the local
2175 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2178 for_each_online_cpu(cpu) {
2179 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2180 struct mem_cgroup *memcg;
2184 memcg = stock->cached;
2185 if (memcg && stock->nr_pages &&
2186 mem_cgroup_is_descendant(memcg, root_memcg))
2191 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2193 drain_local_stock(&stock->work);
2195 schedule_work_on(cpu, &stock->work);
2199 mutex_unlock(&percpu_charge_mutex);
2202 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2204 struct memcg_stock_pcp *stock;
2205 struct mem_cgroup *memcg, *mi;
2207 stock = &per_cpu(memcg_stock, cpu);
2210 for_each_mem_cgroup(memcg) {
2213 for (i = 0; i < MEMCG_NR_STAT; i++) {
2217 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2219 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2220 atomic_long_add(x, &memcg->vmstats[i]);
2222 if (i >= NR_VM_NODE_STAT_ITEMS)
2225 for_each_node(nid) {
2226 struct mem_cgroup_per_node *pn;
2228 pn = mem_cgroup_nodeinfo(memcg, nid);
2229 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2232 atomic_long_add(x, &pn->lruvec_stat[i]);
2233 } while ((pn = parent_nodeinfo(pn, nid)));
2237 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2240 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2242 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2243 atomic_long_add(x, &memcg->vmevents[i]);
2250 static void reclaim_high(struct mem_cgroup *memcg,
2251 unsigned int nr_pages,
2255 if (page_counter_read(&memcg->memory) <= READ_ONCE(memcg->high))
2257 memcg_memory_event(memcg, MEMCG_HIGH);
2258 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2259 } while ((memcg = parent_mem_cgroup(memcg)) &&
2260 !mem_cgroup_is_root(memcg));
2263 static void high_work_func(struct work_struct *work)
2265 struct mem_cgroup *memcg;
2267 memcg = container_of(work, struct mem_cgroup, high_work);
2268 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2272 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2273 * enough to still cause a significant slowdown in most cases, while still
2274 * allowing diagnostics and tracing to proceed without becoming stuck.
2276 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2279 * When calculating the delay, we use these either side of the exponentiation to
2280 * maintain precision and scale to a reasonable number of jiffies (see the table
2283 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2284 * overage ratio to a delay.
2285 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down down the
2286 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2287 * to produce a reasonable delay curve.
2289 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2290 * reasonable delay curve compared to precision-adjusted overage, not
2291 * penalising heavily at first, but still making sure that growth beyond the
2292 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2293 * example, with a high of 100 megabytes:
2295 * +-------+------------------------+
2296 * | usage | time to allocate in ms |
2297 * +-------+------------------------+
2319 * +-------+------------------------+
2321 #define MEMCG_DELAY_PRECISION_SHIFT 20
2322 #define MEMCG_DELAY_SCALING_SHIFT 14
2325 * Get the number of jiffies that we should penalise a mischievous cgroup which
2326 * is exceeding its memory.high by checking both it and its ancestors.
2328 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2329 unsigned int nr_pages)
2331 unsigned long penalty_jiffies;
2332 u64 max_overage = 0;
2335 unsigned long usage, high;
2338 usage = page_counter_read(&memcg->memory);
2339 high = READ_ONCE(memcg->high);
2345 * Prevent division by 0 in overage calculation by acting as if
2346 * it was a threshold of 1 page
2348 high = max(high, 1UL);
2350 overage = usage - high;
2351 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2352 overage = div64_u64(overage, high);
2354 if (overage > max_overage)
2355 max_overage = overage;
2356 } while ((memcg = parent_mem_cgroup(memcg)) &&
2357 !mem_cgroup_is_root(memcg));
2363 * We use overage compared to memory.high to calculate the number of
2364 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2365 * fairly lenient on small overages, and increasingly harsh when the
2366 * memcg in question makes it clear that it has no intention of stopping
2367 * its crazy behaviour, so we exponentially increase the delay based on
2370 penalty_jiffies = max_overage * max_overage * HZ;
2371 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2372 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2375 * Factor in the task's own contribution to the overage, such that four
2376 * N-sized allocations are throttled approximately the same as one
2377 * 4N-sized allocation.
2379 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2380 * larger the current charge patch is than that.
2382 penalty_jiffies = penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2385 * Clamp the max delay per usermode return so as to still keep the
2386 * application moving forwards and also permit diagnostics, albeit
2389 return min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2393 * Scheduled by try_charge() to be executed from the userland return path
2394 * and reclaims memory over the high limit.
2396 void mem_cgroup_handle_over_high(void)
2398 unsigned long penalty_jiffies;
2399 unsigned long pflags;
2400 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2401 struct mem_cgroup *memcg;
2403 if (likely(!nr_pages))
2406 memcg = get_mem_cgroup_from_mm(current->mm);
2407 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2408 current->memcg_nr_pages_over_high = 0;
2411 * memory.high is breached and reclaim is unable to keep up. Throttle
2412 * allocators proactively to slow down excessive growth.
2414 penalty_jiffies = calculate_high_delay(memcg, nr_pages);
2417 * Don't sleep if the amount of jiffies this memcg owes us is so low
2418 * that it's not even worth doing, in an attempt to be nice to those who
2419 * go only a small amount over their memory.high value and maybe haven't
2420 * been aggressively reclaimed enough yet.
2422 if (penalty_jiffies <= HZ / 100)
2426 * If we exit early, we're guaranteed to die (since
2427 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2428 * need to account for any ill-begotten jiffies to pay them off later.
2430 psi_memstall_enter(&pflags);
2431 schedule_timeout_killable(penalty_jiffies);
2432 psi_memstall_leave(&pflags);
2435 css_put(&memcg->css);
2438 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2439 unsigned int nr_pages)
2441 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2442 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2443 struct mem_cgroup *mem_over_limit;
2444 struct page_counter *counter;
2445 unsigned long nr_reclaimed;
2446 bool may_swap = true;
2447 bool drained = false;
2448 enum oom_status oom_status;
2450 if (mem_cgroup_is_root(memcg))
2453 if (consume_stock(memcg, nr_pages))
2456 if (!do_memsw_account() ||
2457 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2458 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2460 if (do_memsw_account())
2461 page_counter_uncharge(&memcg->memsw, batch);
2462 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2464 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2468 if (batch > nr_pages) {
2474 * Memcg doesn't have a dedicated reserve for atomic
2475 * allocations. But like the global atomic pool, we need to
2476 * put the burden of reclaim on regular allocation requests
2477 * and let these go through as privileged allocations.
2479 if (gfp_mask & __GFP_ATOMIC)
2483 * Unlike in global OOM situations, memcg is not in a physical
2484 * memory shortage. Allow dying and OOM-killed tasks to
2485 * bypass the last charges so that they can exit quickly and
2486 * free their memory.
2488 if (unlikely(should_force_charge()))
2492 * Prevent unbounded recursion when reclaim operations need to
2493 * allocate memory. This might exceed the limits temporarily,
2494 * but we prefer facilitating memory reclaim and getting back
2495 * under the limit over triggering OOM kills in these cases.
2497 if (unlikely(current->flags & PF_MEMALLOC))
2500 if (unlikely(task_in_memcg_oom(current)))
2503 if (!gfpflags_allow_blocking(gfp_mask))
2506 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2508 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2509 gfp_mask, may_swap);
2511 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2515 drain_all_stock(mem_over_limit);
2520 if (gfp_mask & __GFP_NORETRY)
2523 * Even though the limit is exceeded at this point, reclaim
2524 * may have been able to free some pages. Retry the charge
2525 * before killing the task.
2527 * Only for regular pages, though: huge pages are rather
2528 * unlikely to succeed so close to the limit, and we fall back
2529 * to regular pages anyway in case of failure.
2531 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2534 * At task move, charge accounts can be doubly counted. So, it's
2535 * better to wait until the end of task_move if something is going on.
2537 if (mem_cgroup_wait_acct_move(mem_over_limit))
2543 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2546 if (gfp_mask & __GFP_NOFAIL)
2549 if (fatal_signal_pending(current))
2553 * keep retrying as long as the memcg oom killer is able to make
2554 * a forward progress or bypass the charge if the oom killer
2555 * couldn't make any progress.
2557 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2558 get_order(nr_pages * PAGE_SIZE));
2559 switch (oom_status) {
2561 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2569 if (!(gfp_mask & __GFP_NOFAIL))
2573 * The allocation either can't fail or will lead to more memory
2574 * being freed very soon. Allow memory usage go over the limit
2575 * temporarily by force charging it.
2577 page_counter_charge(&memcg->memory, nr_pages);
2578 if (do_memsw_account())
2579 page_counter_charge(&memcg->memsw, nr_pages);
2580 css_get_many(&memcg->css, nr_pages);
2585 css_get_many(&memcg->css, batch);
2586 if (batch > nr_pages)
2587 refill_stock(memcg, batch - nr_pages);
2590 * If the hierarchy is above the normal consumption range, schedule
2591 * reclaim on returning to userland. We can perform reclaim here
2592 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2593 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2594 * not recorded as it most likely matches current's and won't
2595 * change in the meantime. As high limit is checked again before
2596 * reclaim, the cost of mismatch is negligible.
2599 if (page_counter_read(&memcg->memory) > READ_ONCE(memcg->high)) {
2600 /* Don't bother a random interrupted task */
2601 if (in_interrupt()) {
2602 schedule_work(&memcg->high_work);
2605 current->memcg_nr_pages_over_high += batch;
2606 set_notify_resume(current);
2609 } while ((memcg = parent_mem_cgroup(memcg)));
2614 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2616 if (mem_cgroup_is_root(memcg))
2619 page_counter_uncharge(&memcg->memory, nr_pages);
2620 if (do_memsw_account())
2621 page_counter_uncharge(&memcg->memsw, nr_pages);
2623 css_put_many(&memcg->css, nr_pages);
2626 static void lock_page_lru(struct page *page, int *isolated)
2628 pg_data_t *pgdat = page_pgdat(page);
2630 spin_lock_irq(&pgdat->lru_lock);
2631 if (PageLRU(page)) {
2632 struct lruvec *lruvec;
2634 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2636 del_page_from_lru_list(page, lruvec, page_lru(page));
2642 static void unlock_page_lru(struct page *page, int isolated)
2644 pg_data_t *pgdat = page_pgdat(page);
2647 struct lruvec *lruvec;
2649 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2650 VM_BUG_ON_PAGE(PageLRU(page), page);
2652 add_page_to_lru_list(page, lruvec, page_lru(page));
2654 spin_unlock_irq(&pgdat->lru_lock);
2657 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2662 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2665 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2666 * may already be on some other mem_cgroup's LRU. Take care of it.
2669 lock_page_lru(page, &isolated);
2672 * Nobody should be changing or seriously looking at
2673 * page->mem_cgroup at this point:
2675 * - the page is uncharged
2677 * - the page is off-LRU
2679 * - an anonymous fault has exclusive page access, except for
2680 * a locked page table
2682 * - a page cache insertion, a swapin fault, or a migration
2683 * have the page locked
2685 page->mem_cgroup = memcg;
2688 unlock_page_lru(page, isolated);
2691 #ifdef CONFIG_MEMCG_KMEM
2693 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2695 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2696 * cgroup_mutex, etc.
2698 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2702 if (mem_cgroup_disabled())
2705 page = virt_to_head_page(p);
2708 * Slab pages don't have page->mem_cgroup set because corresponding
2709 * kmem caches can be reparented during the lifetime. That's why
2710 * memcg_from_slab_page() should be used instead.
2713 return memcg_from_slab_page(page);
2715 /* All other pages use page->mem_cgroup */
2716 return page->mem_cgroup;
2719 static int memcg_alloc_cache_id(void)
2724 id = ida_simple_get(&memcg_cache_ida,
2725 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2729 if (id < memcg_nr_cache_ids)
2733 * There's no space for the new id in memcg_caches arrays,
2734 * so we have to grow them.
2736 down_write(&memcg_cache_ids_sem);
2738 size = 2 * (id + 1);
2739 if (size < MEMCG_CACHES_MIN_SIZE)
2740 size = MEMCG_CACHES_MIN_SIZE;
2741 else if (size > MEMCG_CACHES_MAX_SIZE)
2742 size = MEMCG_CACHES_MAX_SIZE;
2744 err = memcg_update_all_caches(size);
2746 err = memcg_update_all_list_lrus(size);
2748 memcg_nr_cache_ids = size;
2750 up_write(&memcg_cache_ids_sem);
2753 ida_simple_remove(&memcg_cache_ida, id);
2759 static void memcg_free_cache_id(int id)
2761 ida_simple_remove(&memcg_cache_ida, id);
2764 struct memcg_kmem_cache_create_work {
2765 struct mem_cgroup *memcg;
2766 struct kmem_cache *cachep;
2767 struct work_struct work;
2770 static void memcg_kmem_cache_create_func(struct work_struct *w)
2772 struct memcg_kmem_cache_create_work *cw =
2773 container_of(w, struct memcg_kmem_cache_create_work, work);
2774 struct mem_cgroup *memcg = cw->memcg;
2775 struct kmem_cache *cachep = cw->cachep;
2777 memcg_create_kmem_cache(memcg, cachep);
2779 css_put(&memcg->css);
2784 * Enqueue the creation of a per-memcg kmem_cache.
2786 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2787 struct kmem_cache *cachep)
2789 struct memcg_kmem_cache_create_work *cw;
2791 if (!css_tryget_online(&memcg->css))
2794 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2799 cw->cachep = cachep;
2800 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2802 queue_work(memcg_kmem_cache_wq, &cw->work);
2805 static inline bool memcg_kmem_bypass(void)
2807 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2813 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2814 * @cachep: the original global kmem cache
2816 * Return the kmem_cache we're supposed to use for a slab allocation.
2817 * We try to use the current memcg's version of the cache.
2819 * If the cache does not exist yet, if we are the first user of it, we
2820 * create it asynchronously in a workqueue and let the current allocation
2821 * go through with the original cache.
2823 * This function takes a reference to the cache it returns to assure it
2824 * won't get destroyed while we are working with it. Once the caller is
2825 * done with it, memcg_kmem_put_cache() must be called to release the
2828 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2830 struct mem_cgroup *memcg;
2831 struct kmem_cache *memcg_cachep;
2832 struct memcg_cache_array *arr;
2835 VM_BUG_ON(!is_root_cache(cachep));
2837 if (memcg_kmem_bypass())
2842 if (unlikely(current->active_memcg))
2843 memcg = current->active_memcg;
2845 memcg = mem_cgroup_from_task(current);
2847 if (!memcg || memcg == root_mem_cgroup)
2850 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2854 arr = rcu_dereference(cachep->memcg_params.memcg_caches);
2857 * Make sure we will access the up-to-date value. The code updating
2858 * memcg_caches issues a write barrier to match the data dependency
2859 * barrier inside READ_ONCE() (see memcg_create_kmem_cache()).
2861 memcg_cachep = READ_ONCE(arr->entries[kmemcg_id]);
2864 * If we are in a safe context (can wait, and not in interrupt
2865 * context), we could be be predictable and return right away.
2866 * This would guarantee that the allocation being performed
2867 * already belongs in the new cache.
2869 * However, there are some clashes that can arrive from locking.
2870 * For instance, because we acquire the slab_mutex while doing
2871 * memcg_create_kmem_cache, this means no further allocation
2872 * could happen with the slab_mutex held. So it's better to
2875 * If the memcg is dying or memcg_cache is about to be released,
2876 * don't bother creating new kmem_caches. Because memcg_cachep
2877 * is ZEROed as the fist step of kmem offlining, we don't need
2878 * percpu_ref_tryget_live() here. css_tryget_online() check in
2879 * memcg_schedule_kmem_cache_create() will prevent us from
2880 * creation of a new kmem_cache.
2882 if (unlikely(!memcg_cachep))
2883 memcg_schedule_kmem_cache_create(memcg, cachep);
2884 else if (percpu_ref_tryget(&memcg_cachep->memcg_params.refcnt))
2885 cachep = memcg_cachep;
2892 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2893 * @cachep: the cache returned by memcg_kmem_get_cache
2895 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2897 if (!is_root_cache(cachep))
2898 percpu_ref_put(&cachep->memcg_params.refcnt);
2902 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
2903 * @memcg: memory cgroup to charge
2904 * @gfp: reclaim mode
2905 * @nr_pages: number of pages to charge
2907 * Returns 0 on success, an error code on failure.
2909 int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
2910 unsigned int nr_pages)
2912 struct page_counter *counter;
2915 ret = try_charge(memcg, gfp, nr_pages);
2919 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2920 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2923 * Enforce __GFP_NOFAIL allocation because callers are not
2924 * prepared to see failures and likely do not have any failure
2927 if (gfp & __GFP_NOFAIL) {
2928 page_counter_charge(&memcg->kmem, nr_pages);
2931 cancel_charge(memcg, nr_pages);
2938 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
2939 * @memcg: memcg to uncharge
2940 * @nr_pages: number of pages to uncharge
2942 void __memcg_kmem_uncharge(struct mem_cgroup *memcg, unsigned int nr_pages)
2944 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2945 page_counter_uncharge(&memcg->kmem, nr_pages);
2947 page_counter_uncharge(&memcg->memory, nr_pages);
2948 if (do_memsw_account())
2949 page_counter_uncharge(&memcg->memsw, nr_pages);
2953 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
2954 * @page: page to charge
2955 * @gfp: reclaim mode
2956 * @order: allocation order
2958 * Returns 0 on success, an error code on failure.
2960 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
2962 struct mem_cgroup *memcg;
2965 if (memcg_kmem_bypass())
2968 memcg = get_mem_cgroup_from_current();
2969 if (!mem_cgroup_is_root(memcg)) {
2970 ret = __memcg_kmem_charge(memcg, gfp, 1 << order);
2972 page->mem_cgroup = memcg;
2973 __SetPageKmemcg(page);
2976 css_put(&memcg->css);
2981 * __memcg_kmem_uncharge_page: uncharge a kmem page
2982 * @page: page to uncharge
2983 * @order: allocation order
2985 void __memcg_kmem_uncharge_page(struct page *page, int order)
2987 struct mem_cgroup *memcg = page->mem_cgroup;
2988 unsigned int nr_pages = 1 << order;
2993 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2994 __memcg_kmem_uncharge(memcg, nr_pages);
2995 page->mem_cgroup = NULL;
2997 /* slab pages do not have PageKmemcg flag set */
2998 if (PageKmemcg(page))
2999 __ClearPageKmemcg(page);
3001 css_put_many(&memcg->css, nr_pages);
3003 #endif /* CONFIG_MEMCG_KMEM */
3005 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3008 * Because tail pages are not marked as "used", set it. We're under
3009 * pgdat->lru_lock and migration entries setup in all page mappings.
3011 void mem_cgroup_split_huge_fixup(struct page *head)
3015 if (mem_cgroup_disabled())
3018 for (i = 1; i < HPAGE_PMD_NR; i++)
3019 head[i].mem_cgroup = head->mem_cgroup;
3021 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
3023 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3025 #ifdef CONFIG_MEMCG_SWAP
3027 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3028 * @entry: swap entry to be moved
3029 * @from: mem_cgroup which the entry is moved from
3030 * @to: mem_cgroup which the entry is moved to
3032 * It succeeds only when the swap_cgroup's record for this entry is the same
3033 * as the mem_cgroup's id of @from.
3035 * Returns 0 on success, -EINVAL on failure.
3037 * The caller must have charged to @to, IOW, called page_counter_charge() about
3038 * both res and memsw, and called css_get().
3040 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3041 struct mem_cgroup *from, struct mem_cgroup *to)
3043 unsigned short old_id, new_id;
3045 old_id = mem_cgroup_id(from);
3046 new_id = mem_cgroup_id(to);
3048 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3049 mod_memcg_state(from, MEMCG_SWAP, -1);
3050 mod_memcg_state(to, MEMCG_SWAP, 1);
3056 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3057 struct mem_cgroup *from, struct mem_cgroup *to)
3063 static DEFINE_MUTEX(memcg_max_mutex);
3065 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3066 unsigned long max, bool memsw)
3068 bool enlarge = false;
3069 bool drained = false;
3071 bool limits_invariant;
3072 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3075 if (signal_pending(current)) {
3080 mutex_lock(&memcg_max_mutex);
3082 * Make sure that the new limit (memsw or memory limit) doesn't
3083 * break our basic invariant rule memory.max <= memsw.max.
3085 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3086 max <= memcg->memsw.max;
3087 if (!limits_invariant) {
3088 mutex_unlock(&memcg_max_mutex);
3092 if (max > counter->max)
3094 ret = page_counter_set_max(counter, max);
3095 mutex_unlock(&memcg_max_mutex);
3101 drain_all_stock(memcg);
3106 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3107 GFP_KERNEL, !memsw)) {
3113 if (!ret && enlarge)
3114 memcg_oom_recover(memcg);
3119 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3121 unsigned long *total_scanned)
3123 unsigned long nr_reclaimed = 0;
3124 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3125 unsigned long reclaimed;
3127 struct mem_cgroup_tree_per_node *mctz;
3128 unsigned long excess;
3129 unsigned long nr_scanned;
3134 mctz = soft_limit_tree_node(pgdat->node_id);
3137 * Do not even bother to check the largest node if the root
3138 * is empty. Do it lockless to prevent lock bouncing. Races
3139 * are acceptable as soft limit is best effort anyway.
3141 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3145 * This loop can run a while, specially if mem_cgroup's continuously
3146 * keep exceeding their soft limit and putting the system under
3153 mz = mem_cgroup_largest_soft_limit_node(mctz);
3158 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3159 gfp_mask, &nr_scanned);
3160 nr_reclaimed += reclaimed;
3161 *total_scanned += nr_scanned;
3162 spin_lock_irq(&mctz->lock);
3163 __mem_cgroup_remove_exceeded(mz, mctz);
3166 * If we failed to reclaim anything from this memory cgroup
3167 * it is time to move on to the next cgroup
3171 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3173 excess = soft_limit_excess(mz->memcg);
3175 * One school of thought says that we should not add
3176 * back the node to the tree if reclaim returns 0.
3177 * But our reclaim could return 0, simply because due
3178 * to priority we are exposing a smaller subset of
3179 * memory to reclaim from. Consider this as a longer
3182 /* If excess == 0, no tree ops */
3183 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3184 spin_unlock_irq(&mctz->lock);
3185 css_put(&mz->memcg->css);
3188 * Could not reclaim anything and there are no more
3189 * mem cgroups to try or we seem to be looping without
3190 * reclaiming anything.
3192 if (!nr_reclaimed &&
3194 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3196 } while (!nr_reclaimed);
3198 css_put(&next_mz->memcg->css);
3199 return nr_reclaimed;
3203 * Test whether @memcg has children, dead or alive. Note that this
3204 * function doesn't care whether @memcg has use_hierarchy enabled and
3205 * returns %true if there are child csses according to the cgroup
3206 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3208 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3213 ret = css_next_child(NULL, &memcg->css);
3219 * Reclaims as many pages from the given memcg as possible.
3221 * Caller is responsible for holding css reference for memcg.
3223 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3225 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3227 /* we call try-to-free pages for make this cgroup empty */
3228 lru_add_drain_all();
3230 drain_all_stock(memcg);
3232 /* try to free all pages in this cgroup */
3233 while (nr_retries && page_counter_read(&memcg->memory)) {
3236 if (signal_pending(current))
3239 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3243 /* maybe some writeback is necessary */
3244 congestion_wait(BLK_RW_ASYNC, HZ/10);
3252 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3253 char *buf, size_t nbytes,
3256 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3258 if (mem_cgroup_is_root(memcg))
3260 return mem_cgroup_force_empty(memcg) ?: nbytes;
3263 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3266 return mem_cgroup_from_css(css)->use_hierarchy;
3269 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3270 struct cftype *cft, u64 val)
3273 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3274 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3276 if (memcg->use_hierarchy == val)
3280 * If parent's use_hierarchy is set, we can't make any modifications
3281 * in the child subtrees. If it is unset, then the change can
3282 * occur, provided the current cgroup has no children.
3284 * For the root cgroup, parent_mem is NULL, we allow value to be
3285 * set if there are no children.
3287 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3288 (val == 1 || val == 0)) {
3289 if (!memcg_has_children(memcg))
3290 memcg->use_hierarchy = val;
3299 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3303 if (mem_cgroup_is_root(memcg)) {
3304 val = memcg_page_state(memcg, MEMCG_CACHE) +
3305 memcg_page_state(memcg, MEMCG_RSS);
3307 val += memcg_page_state(memcg, MEMCG_SWAP);
3310 val = page_counter_read(&memcg->memory);
3312 val = page_counter_read(&memcg->memsw);
3325 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3328 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3329 struct page_counter *counter;
3331 switch (MEMFILE_TYPE(cft->private)) {
3333 counter = &memcg->memory;
3336 counter = &memcg->memsw;
3339 counter = &memcg->kmem;
3342 counter = &memcg->tcpmem;
3348 switch (MEMFILE_ATTR(cft->private)) {
3350 if (counter == &memcg->memory)
3351 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3352 if (counter == &memcg->memsw)
3353 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3354 return (u64)page_counter_read(counter) * PAGE_SIZE;
3356 return (u64)counter->max * PAGE_SIZE;
3358 return (u64)counter->watermark * PAGE_SIZE;
3360 return counter->failcnt;
3361 case RES_SOFT_LIMIT:
3362 return (u64)memcg->soft_limit * PAGE_SIZE;
3368 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3370 unsigned long stat[MEMCG_NR_STAT] = {0};
3371 struct mem_cgroup *mi;
3374 for_each_online_cpu(cpu)
3375 for (i = 0; i < MEMCG_NR_STAT; i++)
3376 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3378 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3379 for (i = 0; i < MEMCG_NR_STAT; i++)
3380 atomic_long_add(stat[i], &mi->vmstats[i]);
3382 for_each_node(node) {
3383 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3384 struct mem_cgroup_per_node *pi;
3386 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3389 for_each_online_cpu(cpu)
3390 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3392 pn->lruvec_stat_cpu->count[i], cpu);
3394 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3395 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3396 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3400 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3402 unsigned long events[NR_VM_EVENT_ITEMS];
3403 struct mem_cgroup *mi;
3406 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3409 for_each_online_cpu(cpu)
3410 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3411 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3414 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3415 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3416 atomic_long_add(events[i], &mi->vmevents[i]);
3419 #ifdef CONFIG_MEMCG_KMEM
3420 static int memcg_online_kmem(struct mem_cgroup *memcg)
3424 if (cgroup_memory_nokmem)
3427 BUG_ON(memcg->kmemcg_id >= 0);
3428 BUG_ON(memcg->kmem_state);
3430 memcg_id = memcg_alloc_cache_id();
3434 static_branch_inc(&memcg_kmem_enabled_key);
3436 * A memory cgroup is considered kmem-online as soon as it gets
3437 * kmemcg_id. Setting the id after enabling static branching will
3438 * guarantee no one starts accounting before all call sites are
3441 memcg->kmemcg_id = memcg_id;
3442 memcg->kmem_state = KMEM_ONLINE;
3443 INIT_LIST_HEAD(&memcg->kmem_caches);
3448 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3450 struct cgroup_subsys_state *css;
3451 struct mem_cgroup *parent, *child;
3454 if (memcg->kmem_state != KMEM_ONLINE)
3457 * Clear the online state before clearing memcg_caches array
3458 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3459 * guarantees that no cache will be created for this cgroup
3460 * after we are done (see memcg_create_kmem_cache()).
3462 memcg->kmem_state = KMEM_ALLOCATED;
3464 parent = parent_mem_cgroup(memcg);
3466 parent = root_mem_cgroup;
3469 * Deactivate and reparent kmem_caches.
3471 memcg_deactivate_kmem_caches(memcg, parent);
3473 kmemcg_id = memcg->kmemcg_id;
3474 BUG_ON(kmemcg_id < 0);
3477 * Change kmemcg_id of this cgroup and all its descendants to the
3478 * parent's id, and then move all entries from this cgroup's list_lrus
3479 * to ones of the parent. After we have finished, all list_lrus
3480 * corresponding to this cgroup are guaranteed to remain empty. The
3481 * ordering is imposed by list_lru_node->lock taken by
3482 * memcg_drain_all_list_lrus().
3484 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3485 css_for_each_descendant_pre(css, &memcg->css) {
3486 child = mem_cgroup_from_css(css);
3487 BUG_ON(child->kmemcg_id != kmemcg_id);
3488 child->kmemcg_id = parent->kmemcg_id;
3489 if (!memcg->use_hierarchy)
3494 memcg_drain_all_list_lrus(kmemcg_id, parent);
3496 memcg_free_cache_id(kmemcg_id);
3499 static void memcg_free_kmem(struct mem_cgroup *memcg)
3501 /* css_alloc() failed, offlining didn't happen */
3502 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3503 memcg_offline_kmem(memcg);
3505 if (memcg->kmem_state == KMEM_ALLOCATED) {
3506 WARN_ON(!list_empty(&memcg->kmem_caches));
3507 static_branch_dec(&memcg_kmem_enabled_key);
3511 static int memcg_online_kmem(struct mem_cgroup *memcg)
3515 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3518 static void memcg_free_kmem(struct mem_cgroup *memcg)
3521 #endif /* CONFIG_MEMCG_KMEM */
3523 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3528 mutex_lock(&memcg_max_mutex);
3529 ret = page_counter_set_max(&memcg->kmem, max);
3530 mutex_unlock(&memcg_max_mutex);
3534 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3538 mutex_lock(&memcg_max_mutex);
3540 ret = page_counter_set_max(&memcg->tcpmem, max);
3544 if (!memcg->tcpmem_active) {
3546 * The active flag needs to be written after the static_key
3547 * update. This is what guarantees that the socket activation
3548 * function is the last one to run. See mem_cgroup_sk_alloc()
3549 * for details, and note that we don't mark any socket as
3550 * belonging to this memcg until that flag is up.
3552 * We need to do this, because static_keys will span multiple
3553 * sites, but we can't control their order. If we mark a socket
3554 * as accounted, but the accounting functions are not patched in
3555 * yet, we'll lose accounting.
3557 * We never race with the readers in mem_cgroup_sk_alloc(),
3558 * because when this value change, the code to process it is not
3561 static_branch_inc(&memcg_sockets_enabled_key);
3562 memcg->tcpmem_active = true;
3565 mutex_unlock(&memcg_max_mutex);
3570 * The user of this function is...
3573 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3574 char *buf, size_t nbytes, loff_t off)
3576 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3577 unsigned long nr_pages;
3580 buf = strstrip(buf);
3581 ret = page_counter_memparse(buf, "-1", &nr_pages);
3585 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3587 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3591 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3593 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3596 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3599 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3600 "Please report your usecase to linux-mm@kvack.org if you "
3601 "depend on this functionality.\n");
3602 ret = memcg_update_kmem_max(memcg, nr_pages);
3605 ret = memcg_update_tcp_max(memcg, nr_pages);
3609 case RES_SOFT_LIMIT:
3610 memcg->soft_limit = nr_pages;
3614 return ret ?: nbytes;
3617 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3618 size_t nbytes, loff_t off)
3620 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3621 struct page_counter *counter;
3623 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3625 counter = &memcg->memory;
3628 counter = &memcg->memsw;
3631 counter = &memcg->kmem;
3634 counter = &memcg->tcpmem;
3640 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3642 page_counter_reset_watermark(counter);
3645 counter->failcnt = 0;
3654 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3657 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3661 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3662 struct cftype *cft, u64 val)
3664 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3666 if (val & ~MOVE_MASK)
3670 * No kind of locking is needed in here, because ->can_attach() will
3671 * check this value once in the beginning of the process, and then carry
3672 * on with stale data. This means that changes to this value will only
3673 * affect task migrations starting after the change.
3675 memcg->move_charge_at_immigrate = val;
3679 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3680 struct cftype *cft, u64 val)
3688 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3689 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3690 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3692 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3693 int nid, unsigned int lru_mask)
3695 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3696 unsigned long nr = 0;
3699 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3702 if (!(BIT(lru) & lru_mask))
3704 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3709 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3710 unsigned int lru_mask)
3712 unsigned long nr = 0;
3716 if (!(BIT(lru) & lru_mask))
3718 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3723 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3727 unsigned int lru_mask;
3730 static const struct numa_stat stats[] = {
3731 { "total", LRU_ALL },
3732 { "file", LRU_ALL_FILE },
3733 { "anon", LRU_ALL_ANON },
3734 { "unevictable", BIT(LRU_UNEVICTABLE) },
3736 const struct numa_stat *stat;
3739 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3741 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3742 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3743 seq_printf(m, "%s=%lu", stat->name, nr);
3744 for_each_node_state(nid, N_MEMORY) {
3745 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3747 seq_printf(m, " N%d=%lu", nid, nr);
3752 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3753 struct mem_cgroup *iter;
3756 for_each_mem_cgroup_tree(iter, memcg)
3757 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3758 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3759 for_each_node_state(nid, N_MEMORY) {
3761 for_each_mem_cgroup_tree(iter, memcg)
3762 nr += mem_cgroup_node_nr_lru_pages(
3763 iter, nid, stat->lru_mask);
3764 seq_printf(m, " N%d=%lu", nid, nr);
3771 #endif /* CONFIG_NUMA */
3773 static const unsigned int memcg1_stats[] = {
3784 static const char *const memcg1_stat_names[] = {
3795 /* Universal VM events cgroup1 shows, original sort order */
3796 static const unsigned int memcg1_events[] = {
3803 static int memcg_stat_show(struct seq_file *m, void *v)
3805 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3806 unsigned long memory, memsw;
3807 struct mem_cgroup *mi;
3810 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3812 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3813 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3815 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3816 memcg_page_state_local(memcg, memcg1_stats[i]) *
3820 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3821 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
3822 memcg_events_local(memcg, memcg1_events[i]));
3824 for (i = 0; i < NR_LRU_LISTS; i++)
3825 seq_printf(m, "%s %lu\n", lru_list_name(i),
3826 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3829 /* Hierarchical information */
3830 memory = memsw = PAGE_COUNTER_MAX;
3831 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3832 memory = min(memory, READ_ONCE(mi->memory.max));
3833 memsw = min(memsw, READ_ONCE(mi->memsw.max));
3835 seq_printf(m, "hierarchical_memory_limit %llu\n",
3836 (u64)memory * PAGE_SIZE);
3837 if (do_memsw_account())
3838 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3839 (u64)memsw * PAGE_SIZE);
3841 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3842 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3844 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3845 (u64)memcg_page_state(memcg, memcg1_stats[i]) *
3849 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3850 seq_printf(m, "total_%s %llu\n",
3851 vm_event_name(memcg1_events[i]),
3852 (u64)memcg_events(memcg, memcg1_events[i]));
3854 for (i = 0; i < NR_LRU_LISTS; i++)
3855 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
3856 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
3859 #ifdef CONFIG_DEBUG_VM
3862 struct mem_cgroup_per_node *mz;
3863 struct zone_reclaim_stat *rstat;
3864 unsigned long recent_rotated[2] = {0, 0};
3865 unsigned long recent_scanned[2] = {0, 0};
3867 for_each_online_pgdat(pgdat) {
3868 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3869 rstat = &mz->lruvec.reclaim_stat;
3871 recent_rotated[0] += rstat->recent_rotated[0];
3872 recent_rotated[1] += rstat->recent_rotated[1];
3873 recent_scanned[0] += rstat->recent_scanned[0];
3874 recent_scanned[1] += rstat->recent_scanned[1];
3876 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3877 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3878 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3879 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3886 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3889 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3891 return mem_cgroup_swappiness(memcg);
3894 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3895 struct cftype *cft, u64 val)
3897 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3903 memcg->swappiness = val;
3905 vm_swappiness = val;
3910 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3912 struct mem_cgroup_threshold_ary *t;
3913 unsigned long usage;
3918 t = rcu_dereference(memcg->thresholds.primary);
3920 t = rcu_dereference(memcg->memsw_thresholds.primary);
3925 usage = mem_cgroup_usage(memcg, swap);
3928 * current_threshold points to threshold just below or equal to usage.
3929 * If it's not true, a threshold was crossed after last
3930 * call of __mem_cgroup_threshold().
3932 i = t->current_threshold;
3935 * Iterate backward over array of thresholds starting from
3936 * current_threshold and check if a threshold is crossed.
3937 * If none of thresholds below usage is crossed, we read
3938 * only one element of the array here.
3940 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3941 eventfd_signal(t->entries[i].eventfd, 1);
3943 /* i = current_threshold + 1 */
3947 * Iterate forward over array of thresholds starting from
3948 * current_threshold+1 and check if a threshold is crossed.
3949 * If none of thresholds above usage is crossed, we read
3950 * only one element of the array here.
3952 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3953 eventfd_signal(t->entries[i].eventfd, 1);
3955 /* Update current_threshold */
3956 t->current_threshold = i - 1;
3961 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3964 __mem_cgroup_threshold(memcg, false);
3965 if (do_memsw_account())
3966 __mem_cgroup_threshold(memcg, true);
3968 memcg = parent_mem_cgroup(memcg);
3972 static int compare_thresholds(const void *a, const void *b)
3974 const struct mem_cgroup_threshold *_a = a;
3975 const struct mem_cgroup_threshold *_b = b;
3977 if (_a->threshold > _b->threshold)
3980 if (_a->threshold < _b->threshold)
3986 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3988 struct mem_cgroup_eventfd_list *ev;
3990 spin_lock(&memcg_oom_lock);
3992 list_for_each_entry(ev, &memcg->oom_notify, list)
3993 eventfd_signal(ev->eventfd, 1);
3995 spin_unlock(&memcg_oom_lock);
3999 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4001 struct mem_cgroup *iter;
4003 for_each_mem_cgroup_tree(iter, memcg)
4004 mem_cgroup_oom_notify_cb(iter);
4007 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4008 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4010 struct mem_cgroup_thresholds *thresholds;
4011 struct mem_cgroup_threshold_ary *new;
4012 unsigned long threshold;
4013 unsigned long usage;
4016 ret = page_counter_memparse(args, "-1", &threshold);
4020 mutex_lock(&memcg->thresholds_lock);
4023 thresholds = &memcg->thresholds;
4024 usage = mem_cgroup_usage(memcg, false);
4025 } else if (type == _MEMSWAP) {
4026 thresholds = &memcg->memsw_thresholds;
4027 usage = mem_cgroup_usage(memcg, true);
4031 /* Check if a threshold crossed before adding a new one */
4032 if (thresholds->primary)
4033 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4035 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4037 /* Allocate memory for new array of thresholds */
4038 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4045 /* Copy thresholds (if any) to new array */
4046 if (thresholds->primary) {
4047 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4048 sizeof(struct mem_cgroup_threshold));
4051 /* Add new threshold */
4052 new->entries[size - 1].eventfd = eventfd;
4053 new->entries[size - 1].threshold = threshold;
4055 /* Sort thresholds. Registering of new threshold isn't time-critical */
4056 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4057 compare_thresholds, NULL);
4059 /* Find current threshold */
4060 new->current_threshold = -1;
4061 for (i = 0; i < size; i++) {
4062 if (new->entries[i].threshold <= usage) {
4064 * new->current_threshold will not be used until
4065 * rcu_assign_pointer(), so it's safe to increment
4068 ++new->current_threshold;
4073 /* Free old spare buffer and save old primary buffer as spare */
4074 kfree(thresholds->spare);
4075 thresholds->spare = thresholds->primary;
4077 rcu_assign_pointer(thresholds->primary, new);
4079 /* To be sure that nobody uses thresholds */
4083 mutex_unlock(&memcg->thresholds_lock);
4088 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4089 struct eventfd_ctx *eventfd, const char *args)
4091 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4094 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4095 struct eventfd_ctx *eventfd, const char *args)
4097 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4100 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4101 struct eventfd_ctx *eventfd, enum res_type type)
4103 struct mem_cgroup_thresholds *thresholds;
4104 struct mem_cgroup_threshold_ary *new;
4105 unsigned long usage;
4106 int i, j, size, entries;
4108 mutex_lock(&memcg->thresholds_lock);
4111 thresholds = &memcg->thresholds;
4112 usage = mem_cgroup_usage(memcg, false);
4113 } else if (type == _MEMSWAP) {
4114 thresholds = &memcg->memsw_thresholds;
4115 usage = mem_cgroup_usage(memcg, true);
4119 if (!thresholds->primary)
4122 /* Check if a threshold crossed before removing */
4123 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4125 /* Calculate new number of threshold */
4127 for (i = 0; i < thresholds->primary->size; i++) {
4128 if (thresholds->primary->entries[i].eventfd != eventfd)
4134 new = thresholds->spare;
4136 /* If no items related to eventfd have been cleared, nothing to do */
4140 /* Set thresholds array to NULL if we don't have thresholds */
4149 /* Copy thresholds and find current threshold */
4150 new->current_threshold = -1;
4151 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4152 if (thresholds->primary->entries[i].eventfd == eventfd)
4155 new->entries[j] = thresholds->primary->entries[i];
4156 if (new->entries[j].threshold <= usage) {
4158 * new->current_threshold will not be used
4159 * until rcu_assign_pointer(), so it's safe to increment
4162 ++new->current_threshold;
4168 /* Swap primary and spare array */
4169 thresholds->spare = thresholds->primary;
4171 rcu_assign_pointer(thresholds->primary, new);
4173 /* To be sure that nobody uses thresholds */
4176 /* If all events are unregistered, free the spare array */
4178 kfree(thresholds->spare);
4179 thresholds->spare = NULL;
4182 mutex_unlock(&memcg->thresholds_lock);
4185 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4186 struct eventfd_ctx *eventfd)
4188 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4191 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4192 struct eventfd_ctx *eventfd)
4194 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4197 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4198 struct eventfd_ctx *eventfd, const char *args)
4200 struct mem_cgroup_eventfd_list *event;
4202 event = kmalloc(sizeof(*event), GFP_KERNEL);
4206 spin_lock(&memcg_oom_lock);
4208 event->eventfd = eventfd;
4209 list_add(&event->list, &memcg->oom_notify);
4211 /* already in OOM ? */
4212 if (memcg->under_oom)
4213 eventfd_signal(eventfd, 1);
4214 spin_unlock(&memcg_oom_lock);
4219 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4220 struct eventfd_ctx *eventfd)
4222 struct mem_cgroup_eventfd_list *ev, *tmp;
4224 spin_lock(&memcg_oom_lock);
4226 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4227 if (ev->eventfd == eventfd) {
4228 list_del(&ev->list);
4233 spin_unlock(&memcg_oom_lock);
4236 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4238 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4240 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4241 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4242 seq_printf(sf, "oom_kill %lu\n",
4243 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4247 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4248 struct cftype *cft, u64 val)
4250 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4252 /* cannot set to root cgroup and only 0 and 1 are allowed */
4253 if (!css->parent || !((val == 0) || (val == 1)))
4256 memcg->oom_kill_disable = val;
4258 memcg_oom_recover(memcg);
4263 #ifdef CONFIG_CGROUP_WRITEBACK
4265 #include <trace/events/writeback.h>
4267 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4269 return wb_domain_init(&memcg->cgwb_domain, gfp);
4272 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4274 wb_domain_exit(&memcg->cgwb_domain);
4277 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4279 wb_domain_size_changed(&memcg->cgwb_domain);
4282 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4284 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4286 if (!memcg->css.parent)
4289 return &memcg->cgwb_domain;
4293 * idx can be of type enum memcg_stat_item or node_stat_item.
4294 * Keep in sync with memcg_exact_page().
4296 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4298 long x = atomic_long_read(&memcg->vmstats[idx]);
4301 for_each_online_cpu(cpu)
4302 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4309 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4310 * @wb: bdi_writeback in question
4311 * @pfilepages: out parameter for number of file pages
4312 * @pheadroom: out parameter for number of allocatable pages according to memcg
4313 * @pdirty: out parameter for number of dirty pages
4314 * @pwriteback: out parameter for number of pages under writeback
4316 * Determine the numbers of file, headroom, dirty, and writeback pages in
4317 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4318 * is a bit more involved.
4320 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4321 * headroom is calculated as the lowest headroom of itself and the
4322 * ancestors. Note that this doesn't consider the actual amount of
4323 * available memory in the system. The caller should further cap
4324 * *@pheadroom accordingly.
4326 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4327 unsigned long *pheadroom, unsigned long *pdirty,
4328 unsigned long *pwriteback)
4330 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4331 struct mem_cgroup *parent;
4333 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4335 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4336 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4337 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4338 *pheadroom = PAGE_COUNTER_MAX;
4340 while ((parent = parent_mem_cgroup(memcg))) {
4341 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4342 READ_ONCE(memcg->high));
4343 unsigned long used = page_counter_read(&memcg->memory);
4345 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4351 * Foreign dirty flushing
4353 * There's an inherent mismatch between memcg and writeback. The former
4354 * trackes ownership per-page while the latter per-inode. This was a
4355 * deliberate design decision because honoring per-page ownership in the
4356 * writeback path is complicated, may lead to higher CPU and IO overheads
4357 * and deemed unnecessary given that write-sharing an inode across
4358 * different cgroups isn't a common use-case.
4360 * Combined with inode majority-writer ownership switching, this works well
4361 * enough in most cases but there are some pathological cases. For
4362 * example, let's say there are two cgroups A and B which keep writing to
4363 * different but confined parts of the same inode. B owns the inode and
4364 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4365 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4366 * triggering background writeback. A will be slowed down without a way to
4367 * make writeback of the dirty pages happen.
4369 * Conditions like the above can lead to a cgroup getting repatedly and
4370 * severely throttled after making some progress after each
4371 * dirty_expire_interval while the underyling IO device is almost
4374 * Solving this problem completely requires matching the ownership tracking
4375 * granularities between memcg and writeback in either direction. However,
4376 * the more egregious behaviors can be avoided by simply remembering the
4377 * most recent foreign dirtying events and initiating remote flushes on
4378 * them when local writeback isn't enough to keep the memory clean enough.
4380 * The following two functions implement such mechanism. When a foreign
4381 * page - a page whose memcg and writeback ownerships don't match - is
4382 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4383 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4384 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4385 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4386 * foreign bdi_writebacks which haven't expired. Both the numbers of
4387 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4388 * limited to MEMCG_CGWB_FRN_CNT.
4390 * The mechanism only remembers IDs and doesn't hold any object references.
4391 * As being wrong occasionally doesn't matter, updates and accesses to the
4392 * records are lockless and racy.
4394 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4395 struct bdi_writeback *wb)
4397 struct mem_cgroup *memcg = page->mem_cgroup;
4398 struct memcg_cgwb_frn *frn;
4399 u64 now = get_jiffies_64();
4400 u64 oldest_at = now;
4404 trace_track_foreign_dirty(page, wb);
4407 * Pick the slot to use. If there is already a slot for @wb, keep
4408 * using it. If not replace the oldest one which isn't being
4411 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4412 frn = &memcg->cgwb_frn[i];
4413 if (frn->bdi_id == wb->bdi->id &&
4414 frn->memcg_id == wb->memcg_css->id)
4416 if (time_before64(frn->at, oldest_at) &&
4417 atomic_read(&frn->done.cnt) == 1) {
4419 oldest_at = frn->at;
4423 if (i < MEMCG_CGWB_FRN_CNT) {
4425 * Re-using an existing one. Update timestamp lazily to
4426 * avoid making the cacheline hot. We want them to be
4427 * reasonably up-to-date and significantly shorter than
4428 * dirty_expire_interval as that's what expires the record.
4429 * Use the shorter of 1s and dirty_expire_interval / 8.
4431 unsigned long update_intv =
4432 min_t(unsigned long, HZ,
4433 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4435 if (time_before64(frn->at, now - update_intv))
4437 } else if (oldest >= 0) {
4438 /* replace the oldest free one */
4439 frn = &memcg->cgwb_frn[oldest];
4440 frn->bdi_id = wb->bdi->id;
4441 frn->memcg_id = wb->memcg_css->id;
4446 /* issue foreign writeback flushes for recorded foreign dirtying events */
4447 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4449 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4450 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4451 u64 now = jiffies_64;
4454 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4455 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4458 * If the record is older than dirty_expire_interval,
4459 * writeback on it has already started. No need to kick it
4460 * off again. Also, don't start a new one if there's
4461 * already one in flight.
4463 if (time_after64(frn->at, now - intv) &&
4464 atomic_read(&frn->done.cnt) == 1) {
4466 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4467 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4468 WB_REASON_FOREIGN_FLUSH,
4474 #else /* CONFIG_CGROUP_WRITEBACK */
4476 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4481 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4485 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4489 #endif /* CONFIG_CGROUP_WRITEBACK */
4492 * DO NOT USE IN NEW FILES.
4494 * "cgroup.event_control" implementation.
4496 * This is way over-engineered. It tries to support fully configurable
4497 * events for each user. Such level of flexibility is completely
4498 * unnecessary especially in the light of the planned unified hierarchy.
4500 * Please deprecate this and replace with something simpler if at all
4505 * Unregister event and free resources.
4507 * Gets called from workqueue.
4509 static void memcg_event_remove(struct work_struct *work)
4511 struct mem_cgroup_event *event =
4512 container_of(work, struct mem_cgroup_event, remove);
4513 struct mem_cgroup *memcg = event->memcg;
4515 remove_wait_queue(event->wqh, &event->wait);
4517 event->unregister_event(memcg, event->eventfd);
4519 /* Notify userspace the event is going away. */
4520 eventfd_signal(event->eventfd, 1);
4522 eventfd_ctx_put(event->eventfd);
4524 css_put(&memcg->css);
4528 * Gets called on EPOLLHUP on eventfd when user closes it.
4530 * Called with wqh->lock held and interrupts disabled.
4532 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4533 int sync, void *key)
4535 struct mem_cgroup_event *event =
4536 container_of(wait, struct mem_cgroup_event, wait);
4537 struct mem_cgroup *memcg = event->memcg;
4538 __poll_t flags = key_to_poll(key);
4540 if (flags & EPOLLHUP) {
4542 * If the event has been detached at cgroup removal, we
4543 * can simply return knowing the other side will cleanup
4546 * We can't race against event freeing since the other
4547 * side will require wqh->lock via remove_wait_queue(),
4550 spin_lock(&memcg->event_list_lock);
4551 if (!list_empty(&event->list)) {
4552 list_del_init(&event->list);
4554 * We are in atomic context, but cgroup_event_remove()
4555 * may sleep, so we have to call it in workqueue.
4557 schedule_work(&event->remove);
4559 spin_unlock(&memcg->event_list_lock);
4565 static void memcg_event_ptable_queue_proc(struct file *file,
4566 wait_queue_head_t *wqh, poll_table *pt)
4568 struct mem_cgroup_event *event =
4569 container_of(pt, struct mem_cgroup_event, pt);
4572 add_wait_queue(wqh, &event->wait);
4576 * DO NOT USE IN NEW FILES.
4578 * Parse input and register new cgroup event handler.
4580 * Input must be in format '<event_fd> <control_fd> <args>'.
4581 * Interpretation of args is defined by control file implementation.
4583 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4584 char *buf, size_t nbytes, loff_t off)
4586 struct cgroup_subsys_state *css = of_css(of);
4587 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4588 struct mem_cgroup_event *event;
4589 struct cgroup_subsys_state *cfile_css;
4590 unsigned int efd, cfd;
4597 buf = strstrip(buf);
4599 efd = simple_strtoul(buf, &endp, 10);
4604 cfd = simple_strtoul(buf, &endp, 10);
4605 if ((*endp != ' ') && (*endp != '\0'))
4609 event = kzalloc(sizeof(*event), GFP_KERNEL);
4613 event->memcg = memcg;
4614 INIT_LIST_HEAD(&event->list);
4615 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4616 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4617 INIT_WORK(&event->remove, memcg_event_remove);
4625 event->eventfd = eventfd_ctx_fileget(efile.file);
4626 if (IS_ERR(event->eventfd)) {
4627 ret = PTR_ERR(event->eventfd);
4634 goto out_put_eventfd;
4637 /* the process need read permission on control file */
4638 /* AV: shouldn't we check that it's been opened for read instead? */
4639 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4644 * Determine the event callbacks and set them in @event. This used
4645 * to be done via struct cftype but cgroup core no longer knows
4646 * about these events. The following is crude but the whole thing
4647 * is for compatibility anyway.
4649 * DO NOT ADD NEW FILES.
4651 name = cfile.file->f_path.dentry->d_name.name;
4653 if (!strcmp(name, "memory.usage_in_bytes")) {
4654 event->register_event = mem_cgroup_usage_register_event;
4655 event->unregister_event = mem_cgroup_usage_unregister_event;
4656 } else if (!strcmp(name, "memory.oom_control")) {
4657 event->register_event = mem_cgroup_oom_register_event;
4658 event->unregister_event = mem_cgroup_oom_unregister_event;
4659 } else if (!strcmp(name, "memory.pressure_level")) {
4660 event->register_event = vmpressure_register_event;
4661 event->unregister_event = vmpressure_unregister_event;
4662 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4663 event->register_event = memsw_cgroup_usage_register_event;
4664 event->unregister_event = memsw_cgroup_usage_unregister_event;
4671 * Verify @cfile should belong to @css. Also, remaining events are
4672 * automatically removed on cgroup destruction but the removal is
4673 * asynchronous, so take an extra ref on @css.
4675 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4676 &memory_cgrp_subsys);
4678 if (IS_ERR(cfile_css))
4680 if (cfile_css != css) {
4685 ret = event->register_event(memcg, event->eventfd, buf);
4689 vfs_poll(efile.file, &event->pt);
4691 spin_lock(&memcg->event_list_lock);
4692 list_add(&event->list, &memcg->event_list);
4693 spin_unlock(&memcg->event_list_lock);
4705 eventfd_ctx_put(event->eventfd);
4714 static struct cftype mem_cgroup_legacy_files[] = {
4716 .name = "usage_in_bytes",
4717 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4718 .read_u64 = mem_cgroup_read_u64,
4721 .name = "max_usage_in_bytes",
4722 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4723 .write = mem_cgroup_reset,
4724 .read_u64 = mem_cgroup_read_u64,
4727 .name = "limit_in_bytes",
4728 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4729 .write = mem_cgroup_write,
4730 .read_u64 = mem_cgroup_read_u64,
4733 .name = "soft_limit_in_bytes",
4734 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4735 .write = mem_cgroup_write,
4736 .read_u64 = mem_cgroup_read_u64,
4740 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4741 .write = mem_cgroup_reset,
4742 .read_u64 = mem_cgroup_read_u64,
4746 .seq_show = memcg_stat_show,
4749 .name = "force_empty",
4750 .write = mem_cgroup_force_empty_write,
4753 .name = "use_hierarchy",
4754 .write_u64 = mem_cgroup_hierarchy_write,
4755 .read_u64 = mem_cgroup_hierarchy_read,
4758 .name = "cgroup.event_control", /* XXX: for compat */
4759 .write = memcg_write_event_control,
4760 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4763 .name = "swappiness",
4764 .read_u64 = mem_cgroup_swappiness_read,
4765 .write_u64 = mem_cgroup_swappiness_write,
4768 .name = "move_charge_at_immigrate",
4769 .read_u64 = mem_cgroup_move_charge_read,
4770 .write_u64 = mem_cgroup_move_charge_write,
4773 .name = "oom_control",
4774 .seq_show = mem_cgroup_oom_control_read,
4775 .write_u64 = mem_cgroup_oom_control_write,
4776 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4779 .name = "pressure_level",
4783 .name = "numa_stat",
4784 .seq_show = memcg_numa_stat_show,
4788 .name = "kmem.limit_in_bytes",
4789 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4790 .write = mem_cgroup_write,
4791 .read_u64 = mem_cgroup_read_u64,
4794 .name = "kmem.usage_in_bytes",
4795 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4796 .read_u64 = mem_cgroup_read_u64,
4799 .name = "kmem.failcnt",
4800 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4801 .write = mem_cgroup_reset,
4802 .read_u64 = mem_cgroup_read_u64,
4805 .name = "kmem.max_usage_in_bytes",
4806 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4807 .write = mem_cgroup_reset,
4808 .read_u64 = mem_cgroup_read_u64,
4810 #if defined(CONFIG_MEMCG_KMEM) && \
4811 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
4813 .name = "kmem.slabinfo",
4814 .seq_start = memcg_slab_start,
4815 .seq_next = memcg_slab_next,
4816 .seq_stop = memcg_slab_stop,
4817 .seq_show = memcg_slab_show,
4821 .name = "kmem.tcp.limit_in_bytes",
4822 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4823 .write = mem_cgroup_write,
4824 .read_u64 = mem_cgroup_read_u64,
4827 .name = "kmem.tcp.usage_in_bytes",
4828 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4829 .read_u64 = mem_cgroup_read_u64,
4832 .name = "kmem.tcp.failcnt",
4833 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4834 .write = mem_cgroup_reset,
4835 .read_u64 = mem_cgroup_read_u64,
4838 .name = "kmem.tcp.max_usage_in_bytes",
4839 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4840 .write = mem_cgroup_reset,
4841 .read_u64 = mem_cgroup_read_u64,
4843 { }, /* terminate */
4847 * Private memory cgroup IDR
4849 * Swap-out records and page cache shadow entries need to store memcg
4850 * references in constrained space, so we maintain an ID space that is
4851 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4852 * memory-controlled cgroups to 64k.
4854 * However, there usually are many references to the oflline CSS after
4855 * the cgroup has been destroyed, such as page cache or reclaimable
4856 * slab objects, that don't need to hang on to the ID. We want to keep
4857 * those dead CSS from occupying IDs, or we might quickly exhaust the
4858 * relatively small ID space and prevent the creation of new cgroups
4859 * even when there are much fewer than 64k cgroups - possibly none.
4861 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4862 * be freed and recycled when it's no longer needed, which is usually
4863 * when the CSS is offlined.
4865 * The only exception to that are records of swapped out tmpfs/shmem
4866 * pages that need to be attributed to live ancestors on swapin. But
4867 * those references are manageable from userspace.
4870 static DEFINE_IDR(mem_cgroup_idr);
4872 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4874 if (memcg->id.id > 0) {
4875 idr_remove(&mem_cgroup_idr, memcg->id.id);
4880 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
4883 refcount_add(n, &memcg->id.ref);
4886 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4888 if (refcount_sub_and_test(n, &memcg->id.ref)) {
4889 mem_cgroup_id_remove(memcg);
4891 /* Memcg ID pins CSS */
4892 css_put(&memcg->css);
4896 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4898 mem_cgroup_id_put_many(memcg, 1);
4902 * mem_cgroup_from_id - look up a memcg from a memcg id
4903 * @id: the memcg id to look up
4905 * Caller must hold rcu_read_lock().
4907 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4909 WARN_ON_ONCE(!rcu_read_lock_held());
4910 return idr_find(&mem_cgroup_idr, id);
4913 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4915 struct mem_cgroup_per_node *pn;
4918 * This routine is called against possible nodes.
4919 * But it's BUG to call kmalloc() against offline node.
4921 * TODO: this routine can waste much memory for nodes which will
4922 * never be onlined. It's better to use memory hotplug callback
4925 if (!node_state(node, N_NORMAL_MEMORY))
4927 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4931 pn->lruvec_stat_local = alloc_percpu(struct lruvec_stat);
4932 if (!pn->lruvec_stat_local) {
4937 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
4938 if (!pn->lruvec_stat_cpu) {
4939 free_percpu(pn->lruvec_stat_local);
4944 lruvec_init(&pn->lruvec);
4945 pn->usage_in_excess = 0;
4946 pn->on_tree = false;
4949 memcg->nodeinfo[node] = pn;
4953 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4955 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
4960 free_percpu(pn->lruvec_stat_cpu);
4961 free_percpu(pn->lruvec_stat_local);
4965 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4970 free_mem_cgroup_per_node_info(memcg, node);
4971 free_percpu(memcg->vmstats_percpu);
4972 free_percpu(memcg->vmstats_local);
4976 static void mem_cgroup_free(struct mem_cgroup *memcg)
4978 memcg_wb_domain_exit(memcg);
4980 * Flush percpu vmstats and vmevents to guarantee the value correctness
4981 * on parent's and all ancestor levels.
4983 memcg_flush_percpu_vmstats(memcg);
4984 memcg_flush_percpu_vmevents(memcg);
4985 __mem_cgroup_free(memcg);
4988 static struct mem_cgroup *mem_cgroup_alloc(void)
4990 struct mem_cgroup *memcg;
4993 int __maybe_unused i;
4994 long error = -ENOMEM;
4996 size = sizeof(struct mem_cgroup);
4997 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4999 memcg = kzalloc(size, GFP_KERNEL);
5001 return ERR_PTR(error);
5003 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5004 1, MEM_CGROUP_ID_MAX,
5006 if (memcg->id.id < 0) {
5007 error = memcg->id.id;
5011 memcg->vmstats_local = alloc_percpu(struct memcg_vmstats_percpu);
5012 if (!memcg->vmstats_local)
5015 memcg->vmstats_percpu = alloc_percpu(struct memcg_vmstats_percpu);
5016 if (!memcg->vmstats_percpu)
5020 if (alloc_mem_cgroup_per_node_info(memcg, node))
5023 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5026 INIT_WORK(&memcg->high_work, high_work_func);
5027 INIT_LIST_HEAD(&memcg->oom_notify);
5028 mutex_init(&memcg->thresholds_lock);
5029 spin_lock_init(&memcg->move_lock);
5030 vmpressure_init(&memcg->vmpressure);
5031 INIT_LIST_HEAD(&memcg->event_list);
5032 spin_lock_init(&memcg->event_list_lock);
5033 memcg->socket_pressure = jiffies;
5034 #ifdef CONFIG_MEMCG_KMEM
5035 memcg->kmemcg_id = -1;
5037 #ifdef CONFIG_CGROUP_WRITEBACK
5038 INIT_LIST_HEAD(&memcg->cgwb_list);
5039 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5040 memcg->cgwb_frn[i].done =
5041 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5043 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5044 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5045 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5046 memcg->deferred_split_queue.split_queue_len = 0;
5048 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5051 mem_cgroup_id_remove(memcg);
5052 __mem_cgroup_free(memcg);
5053 return ERR_PTR(error);
5056 static struct cgroup_subsys_state * __ref
5057 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5059 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5060 struct mem_cgroup *memcg;
5061 long error = -ENOMEM;
5063 memcg = mem_cgroup_alloc();
5065 return ERR_CAST(memcg);
5067 WRITE_ONCE(memcg->high, PAGE_COUNTER_MAX);
5068 memcg->soft_limit = PAGE_COUNTER_MAX;
5070 memcg->swappiness = mem_cgroup_swappiness(parent);
5071 memcg->oom_kill_disable = parent->oom_kill_disable;
5073 if (parent && parent->use_hierarchy) {
5074 memcg->use_hierarchy = true;
5075 page_counter_init(&memcg->memory, &parent->memory);
5076 page_counter_init(&memcg->swap, &parent->swap);
5077 page_counter_init(&memcg->memsw, &parent->memsw);
5078 page_counter_init(&memcg->kmem, &parent->kmem);
5079 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5081 page_counter_init(&memcg->memory, NULL);
5082 page_counter_init(&memcg->swap, NULL);
5083 page_counter_init(&memcg->memsw, NULL);
5084 page_counter_init(&memcg->kmem, NULL);
5085 page_counter_init(&memcg->tcpmem, NULL);
5087 * Deeper hierachy with use_hierarchy == false doesn't make
5088 * much sense so let cgroup subsystem know about this
5089 * unfortunate state in our controller.
5091 if (parent != root_mem_cgroup)
5092 memory_cgrp_subsys.broken_hierarchy = true;
5095 /* The following stuff does not apply to the root */
5097 #ifdef CONFIG_MEMCG_KMEM
5098 INIT_LIST_HEAD(&memcg->kmem_caches);
5100 root_mem_cgroup = memcg;
5104 error = memcg_online_kmem(memcg);
5108 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5109 static_branch_inc(&memcg_sockets_enabled_key);
5113 mem_cgroup_id_remove(memcg);
5114 mem_cgroup_free(memcg);
5115 return ERR_PTR(error);
5118 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5120 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5123 * A memcg must be visible for memcg_expand_shrinker_maps()
5124 * by the time the maps are allocated. So, we allocate maps
5125 * here, when for_each_mem_cgroup() can't skip it.
5127 if (memcg_alloc_shrinker_maps(memcg)) {
5128 mem_cgroup_id_remove(memcg);
5132 /* Online state pins memcg ID, memcg ID pins CSS */
5133 refcount_set(&memcg->id.ref, 1);
5138 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5140 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5141 struct mem_cgroup_event *event, *tmp;
5144 * Unregister events and notify userspace.
5145 * Notify userspace about cgroup removing only after rmdir of cgroup
5146 * directory to avoid race between userspace and kernelspace.
5148 spin_lock(&memcg->event_list_lock);
5149 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5150 list_del_init(&event->list);
5151 schedule_work(&event->remove);
5153 spin_unlock(&memcg->event_list_lock);
5155 page_counter_set_min(&memcg->memory, 0);
5156 page_counter_set_low(&memcg->memory, 0);
5158 memcg_offline_kmem(memcg);
5159 wb_memcg_offline(memcg);
5161 drain_all_stock(memcg);
5163 mem_cgroup_id_put(memcg);
5166 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5168 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5170 invalidate_reclaim_iterators(memcg);
5173 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5175 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5176 int __maybe_unused i;
5178 #ifdef CONFIG_CGROUP_WRITEBACK
5179 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5180 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5182 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5183 static_branch_dec(&memcg_sockets_enabled_key);
5185 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5186 static_branch_dec(&memcg_sockets_enabled_key);
5188 vmpressure_cleanup(&memcg->vmpressure);
5189 cancel_work_sync(&memcg->high_work);
5190 mem_cgroup_remove_from_trees(memcg);
5191 memcg_free_shrinker_maps(memcg);
5192 memcg_free_kmem(memcg);
5193 mem_cgroup_free(memcg);
5197 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5198 * @css: the target css
5200 * Reset the states of the mem_cgroup associated with @css. This is
5201 * invoked when the userland requests disabling on the default hierarchy
5202 * but the memcg is pinned through dependency. The memcg should stop
5203 * applying policies and should revert to the vanilla state as it may be
5204 * made visible again.
5206 * The current implementation only resets the essential configurations.
5207 * This needs to be expanded to cover all the visible parts.
5209 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5211 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5213 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5214 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5215 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
5216 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5217 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5218 page_counter_set_min(&memcg->memory, 0);
5219 page_counter_set_low(&memcg->memory, 0);
5220 WRITE_ONCE(memcg->high, PAGE_COUNTER_MAX);
5221 memcg->soft_limit = PAGE_COUNTER_MAX;
5222 memcg_wb_domain_size_changed(memcg);
5226 /* Handlers for move charge at task migration. */
5227 static int mem_cgroup_do_precharge(unsigned long count)
5231 /* Try a single bulk charge without reclaim first, kswapd may wake */
5232 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5234 mc.precharge += count;
5238 /* Try charges one by one with reclaim, but do not retry */
5240 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5254 enum mc_target_type {
5261 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5262 unsigned long addr, pte_t ptent)
5264 struct page *page = vm_normal_page(vma, addr, ptent);
5266 if (!page || !page_mapped(page))
5268 if (PageAnon(page)) {
5269 if (!(mc.flags & MOVE_ANON))
5272 if (!(mc.flags & MOVE_FILE))
5275 if (!get_page_unless_zero(page))
5281 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5282 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5283 pte_t ptent, swp_entry_t *entry)
5285 struct page *page = NULL;
5286 swp_entry_t ent = pte_to_swp_entry(ptent);
5288 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
5292 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5293 * a device and because they are not accessible by CPU they are store
5294 * as special swap entry in the CPU page table.
5296 if (is_device_private_entry(ent)) {
5297 page = device_private_entry_to_page(ent);
5299 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5300 * a refcount of 1 when free (unlike normal page)
5302 if (!page_ref_add_unless(page, 1, 1))
5308 * Because lookup_swap_cache() updates some statistics counter,
5309 * we call find_get_page() with swapper_space directly.
5311 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5312 if (do_memsw_account())
5313 entry->val = ent.val;
5318 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5319 pte_t ptent, swp_entry_t *entry)
5325 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5326 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5328 struct page *page = NULL;
5329 struct address_space *mapping;
5332 if (!vma->vm_file) /* anonymous vma */
5334 if (!(mc.flags & MOVE_FILE))
5337 mapping = vma->vm_file->f_mapping;
5338 pgoff = linear_page_index(vma, addr);
5340 /* page is moved even if it's not RSS of this task(page-faulted). */
5342 /* shmem/tmpfs may report page out on swap: account for that too. */
5343 if (shmem_mapping(mapping)) {
5344 page = find_get_entry(mapping, pgoff);
5345 if (xa_is_value(page)) {
5346 swp_entry_t swp = radix_to_swp_entry(page);
5347 if (do_memsw_account())
5349 page = find_get_page(swap_address_space(swp),
5353 page = find_get_page(mapping, pgoff);
5355 page = find_get_page(mapping, pgoff);
5361 * mem_cgroup_move_account - move account of the page
5363 * @compound: charge the page as compound or small page
5364 * @from: mem_cgroup which the page is moved from.
5365 * @to: mem_cgroup which the page is moved to. @from != @to.
5367 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5369 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5372 static int mem_cgroup_move_account(struct page *page,
5374 struct mem_cgroup *from,
5375 struct mem_cgroup *to)
5377 struct lruvec *from_vec, *to_vec;
5378 struct pglist_data *pgdat;
5379 unsigned long flags;
5380 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5384 VM_BUG_ON(from == to);
5385 VM_BUG_ON_PAGE(PageLRU(page), page);
5386 VM_BUG_ON(compound && !PageTransHuge(page));
5389 * Prevent mem_cgroup_migrate() from looking at
5390 * page->mem_cgroup of its source page while we change it.
5393 if (!trylock_page(page))
5397 if (page->mem_cgroup != from)
5400 anon = PageAnon(page);
5402 pgdat = page_pgdat(page);
5403 from_vec = mem_cgroup_lruvec(from, pgdat);
5404 to_vec = mem_cgroup_lruvec(to, pgdat);
5406 spin_lock_irqsave(&from->move_lock, flags);
5408 if (!anon && page_mapped(page)) {
5409 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5410 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5414 * move_lock grabbed above and caller set from->moving_account, so
5415 * mod_memcg_page_state will serialize updates to PageDirty.
5416 * So mapping should be stable for dirty pages.
5418 if (!anon && PageDirty(page)) {
5419 struct address_space *mapping = page_mapping(page);
5421 if (mapping_cap_account_dirty(mapping)) {
5422 __mod_lruvec_state(from_vec, NR_FILE_DIRTY, -nr_pages);
5423 __mod_lruvec_state(to_vec, NR_FILE_DIRTY, nr_pages);
5427 if (PageWriteback(page)) {
5428 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5429 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5433 * It is safe to change page->mem_cgroup here because the page
5434 * is referenced, charged, and isolated - we can't race with
5435 * uncharging, charging, migration, or LRU putback.
5438 /* caller should have done css_get */
5439 page->mem_cgroup = to;
5441 spin_unlock_irqrestore(&from->move_lock, flags);
5445 local_irq_disable();
5446 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
5447 memcg_check_events(to, page);
5448 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
5449 memcg_check_events(from, page);
5458 * get_mctgt_type - get target type of moving charge
5459 * @vma: the vma the pte to be checked belongs
5460 * @addr: the address corresponding to the pte to be checked
5461 * @ptent: the pte to be checked
5462 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5465 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5466 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5467 * move charge. if @target is not NULL, the page is stored in target->page
5468 * with extra refcnt got(Callers should handle it).
5469 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5470 * target for charge migration. if @target is not NULL, the entry is stored
5472 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5473 * (so ZONE_DEVICE page and thus not on the lru).
5474 * For now we such page is charge like a regular page would be as for all
5475 * intent and purposes it is just special memory taking the place of a
5478 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5480 * Called with pte lock held.
5483 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5484 unsigned long addr, pte_t ptent, union mc_target *target)
5486 struct page *page = NULL;
5487 enum mc_target_type ret = MC_TARGET_NONE;
5488 swp_entry_t ent = { .val = 0 };
5490 if (pte_present(ptent))
5491 page = mc_handle_present_pte(vma, addr, ptent);
5492 else if (is_swap_pte(ptent))
5493 page = mc_handle_swap_pte(vma, ptent, &ent);
5494 else if (pte_none(ptent))
5495 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5497 if (!page && !ent.val)
5501 * Do only loose check w/o serialization.
5502 * mem_cgroup_move_account() checks the page is valid or
5503 * not under LRU exclusion.
5505 if (page->mem_cgroup == mc.from) {
5506 ret = MC_TARGET_PAGE;
5507 if (is_device_private_page(page))
5508 ret = MC_TARGET_DEVICE;
5510 target->page = page;
5512 if (!ret || !target)
5516 * There is a swap entry and a page doesn't exist or isn't charged.
5517 * But we cannot move a tail-page in a THP.
5519 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5520 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5521 ret = MC_TARGET_SWAP;
5528 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5530 * We don't consider PMD mapped swapping or file mapped pages because THP does
5531 * not support them for now.
5532 * Caller should make sure that pmd_trans_huge(pmd) is true.
5534 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5535 unsigned long addr, pmd_t pmd, union mc_target *target)
5537 struct page *page = NULL;
5538 enum mc_target_type ret = MC_TARGET_NONE;
5540 if (unlikely(is_swap_pmd(pmd))) {
5541 VM_BUG_ON(thp_migration_supported() &&
5542 !is_pmd_migration_entry(pmd));
5545 page = pmd_page(pmd);
5546 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5547 if (!(mc.flags & MOVE_ANON))
5549 if (page->mem_cgroup == mc.from) {
5550 ret = MC_TARGET_PAGE;
5553 target->page = page;
5559 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5560 unsigned long addr, pmd_t pmd, union mc_target *target)
5562 return MC_TARGET_NONE;
5566 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5567 unsigned long addr, unsigned long end,
5568 struct mm_walk *walk)
5570 struct vm_area_struct *vma = walk->vma;
5574 ptl = pmd_trans_huge_lock(pmd, vma);
5577 * Note their can not be MC_TARGET_DEVICE for now as we do not
5578 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5579 * this might change.
5581 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5582 mc.precharge += HPAGE_PMD_NR;
5587 if (pmd_trans_unstable(pmd))
5589 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5590 for (; addr != end; pte++, addr += PAGE_SIZE)
5591 if (get_mctgt_type(vma, addr, *pte, NULL))
5592 mc.precharge++; /* increment precharge temporarily */
5593 pte_unmap_unlock(pte - 1, ptl);
5599 static const struct mm_walk_ops precharge_walk_ops = {
5600 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5603 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5605 unsigned long precharge;
5607 down_read(&mm->mmap_sem);
5608 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5609 up_read(&mm->mmap_sem);
5611 precharge = mc.precharge;
5617 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5619 unsigned long precharge = mem_cgroup_count_precharge(mm);
5621 VM_BUG_ON(mc.moving_task);
5622 mc.moving_task = current;
5623 return mem_cgroup_do_precharge(precharge);
5626 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5627 static void __mem_cgroup_clear_mc(void)
5629 struct mem_cgroup *from = mc.from;
5630 struct mem_cgroup *to = mc.to;
5632 /* we must uncharge all the leftover precharges from mc.to */
5634 cancel_charge(mc.to, mc.precharge);
5638 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5639 * we must uncharge here.
5641 if (mc.moved_charge) {
5642 cancel_charge(mc.from, mc.moved_charge);
5643 mc.moved_charge = 0;
5645 /* we must fixup refcnts and charges */
5646 if (mc.moved_swap) {
5647 /* uncharge swap account from the old cgroup */
5648 if (!mem_cgroup_is_root(mc.from))
5649 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5651 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5654 * we charged both to->memory and to->memsw, so we
5655 * should uncharge to->memory.
5657 if (!mem_cgroup_is_root(mc.to))
5658 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5660 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
5661 css_put_many(&mc.to->css, mc.moved_swap);
5665 memcg_oom_recover(from);
5666 memcg_oom_recover(to);
5667 wake_up_all(&mc.waitq);
5670 static void mem_cgroup_clear_mc(void)
5672 struct mm_struct *mm = mc.mm;
5675 * we must clear moving_task before waking up waiters at the end of
5678 mc.moving_task = NULL;
5679 __mem_cgroup_clear_mc();
5680 spin_lock(&mc.lock);
5684 spin_unlock(&mc.lock);
5689 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5691 struct cgroup_subsys_state *css;
5692 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5693 struct mem_cgroup *from;
5694 struct task_struct *leader, *p;
5695 struct mm_struct *mm;
5696 unsigned long move_flags;
5699 /* charge immigration isn't supported on the default hierarchy */
5700 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5704 * Multi-process migrations only happen on the default hierarchy
5705 * where charge immigration is not used. Perform charge
5706 * immigration if @tset contains a leader and whine if there are
5710 cgroup_taskset_for_each_leader(leader, css, tset) {
5713 memcg = mem_cgroup_from_css(css);
5719 * We are now commited to this value whatever it is. Changes in this
5720 * tunable will only affect upcoming migrations, not the current one.
5721 * So we need to save it, and keep it going.
5723 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5727 from = mem_cgroup_from_task(p);
5729 VM_BUG_ON(from == memcg);
5731 mm = get_task_mm(p);
5734 /* We move charges only when we move a owner of the mm */
5735 if (mm->owner == p) {
5738 VM_BUG_ON(mc.precharge);
5739 VM_BUG_ON(mc.moved_charge);
5740 VM_BUG_ON(mc.moved_swap);
5742 spin_lock(&mc.lock);
5746 mc.flags = move_flags;
5747 spin_unlock(&mc.lock);
5748 /* We set mc.moving_task later */
5750 ret = mem_cgroup_precharge_mc(mm);
5752 mem_cgroup_clear_mc();
5759 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5762 mem_cgroup_clear_mc();
5765 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5766 unsigned long addr, unsigned long end,
5767 struct mm_walk *walk)
5770 struct vm_area_struct *vma = walk->vma;
5773 enum mc_target_type target_type;
5774 union mc_target target;
5777 ptl = pmd_trans_huge_lock(pmd, vma);
5779 if (mc.precharge < HPAGE_PMD_NR) {
5783 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5784 if (target_type == MC_TARGET_PAGE) {
5786 if (!isolate_lru_page(page)) {
5787 if (!mem_cgroup_move_account(page, true,
5789 mc.precharge -= HPAGE_PMD_NR;
5790 mc.moved_charge += HPAGE_PMD_NR;
5792 putback_lru_page(page);
5795 } else if (target_type == MC_TARGET_DEVICE) {
5797 if (!mem_cgroup_move_account(page, true,
5799 mc.precharge -= HPAGE_PMD_NR;
5800 mc.moved_charge += HPAGE_PMD_NR;
5808 if (pmd_trans_unstable(pmd))
5811 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5812 for (; addr != end; addr += PAGE_SIZE) {
5813 pte_t ptent = *(pte++);
5814 bool device = false;
5820 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5821 case MC_TARGET_DEVICE:
5824 case MC_TARGET_PAGE:
5827 * We can have a part of the split pmd here. Moving it
5828 * can be done but it would be too convoluted so simply
5829 * ignore such a partial THP and keep it in original
5830 * memcg. There should be somebody mapping the head.
5832 if (PageTransCompound(page))
5834 if (!device && isolate_lru_page(page))
5836 if (!mem_cgroup_move_account(page, false,
5839 /* we uncharge from mc.from later. */
5843 putback_lru_page(page);
5844 put: /* get_mctgt_type() gets the page */
5847 case MC_TARGET_SWAP:
5849 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5851 /* we fixup refcnts and charges later. */
5859 pte_unmap_unlock(pte - 1, ptl);
5864 * We have consumed all precharges we got in can_attach().
5865 * We try charge one by one, but don't do any additional
5866 * charges to mc.to if we have failed in charge once in attach()
5869 ret = mem_cgroup_do_precharge(1);
5877 static const struct mm_walk_ops charge_walk_ops = {
5878 .pmd_entry = mem_cgroup_move_charge_pte_range,
5881 static void mem_cgroup_move_charge(void)
5883 lru_add_drain_all();
5885 * Signal lock_page_memcg() to take the memcg's move_lock
5886 * while we're moving its pages to another memcg. Then wait
5887 * for already started RCU-only updates to finish.
5889 atomic_inc(&mc.from->moving_account);
5892 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5894 * Someone who are holding the mmap_sem might be waiting in
5895 * waitq. So we cancel all extra charges, wake up all waiters,
5896 * and retry. Because we cancel precharges, we might not be able
5897 * to move enough charges, but moving charge is a best-effort
5898 * feature anyway, so it wouldn't be a big problem.
5900 __mem_cgroup_clear_mc();
5905 * When we have consumed all precharges and failed in doing
5906 * additional charge, the page walk just aborts.
5908 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
5911 up_read(&mc.mm->mmap_sem);
5912 atomic_dec(&mc.from->moving_account);
5915 static void mem_cgroup_move_task(void)
5918 mem_cgroup_move_charge();
5919 mem_cgroup_clear_mc();
5922 #else /* !CONFIG_MMU */
5923 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5927 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5930 static void mem_cgroup_move_task(void)
5936 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5937 * to verify whether we're attached to the default hierarchy on each mount
5940 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5943 * use_hierarchy is forced on the default hierarchy. cgroup core
5944 * guarantees that @root doesn't have any children, so turning it
5945 * on for the root memcg is enough.
5947 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5948 root_mem_cgroup->use_hierarchy = true;
5950 root_mem_cgroup->use_hierarchy = false;
5953 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
5955 if (value == PAGE_COUNTER_MAX)
5956 seq_puts(m, "max\n");
5958 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
5963 static u64 memory_current_read(struct cgroup_subsys_state *css,
5966 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5968 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5971 static int memory_min_show(struct seq_file *m, void *v)
5973 return seq_puts_memcg_tunable(m,
5974 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
5977 static ssize_t memory_min_write(struct kernfs_open_file *of,
5978 char *buf, size_t nbytes, loff_t off)
5980 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5984 buf = strstrip(buf);
5985 err = page_counter_memparse(buf, "max", &min);
5989 page_counter_set_min(&memcg->memory, min);
5994 static int memory_low_show(struct seq_file *m, void *v)
5996 return seq_puts_memcg_tunable(m,
5997 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6000 static ssize_t memory_low_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));
6007 buf = strstrip(buf);
6008 err = page_counter_memparse(buf, "max", &low);
6012 page_counter_set_low(&memcg->memory, low);
6017 static int memory_high_show(struct seq_file *m, void *v)
6019 return seq_puts_memcg_tunable(m, READ_ONCE(mem_cgroup_from_seq(m)->high));
6022 static ssize_t memory_high_write(struct kernfs_open_file *of,
6023 char *buf, size_t nbytes, loff_t off)
6025 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6026 unsigned int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
6027 bool drained = false;
6031 buf = strstrip(buf);
6032 err = page_counter_memparse(buf, "max", &high);
6036 WRITE_ONCE(memcg->high, high);
6039 unsigned long nr_pages = page_counter_read(&memcg->memory);
6040 unsigned long reclaimed;
6042 if (nr_pages <= high)
6045 if (signal_pending(current))
6049 drain_all_stock(memcg);
6054 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6057 if (!reclaimed && !nr_retries--)
6064 static int memory_max_show(struct seq_file *m, void *v)
6066 return seq_puts_memcg_tunable(m,
6067 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6070 static ssize_t memory_max_write(struct kernfs_open_file *of,
6071 char *buf, size_t nbytes, loff_t off)
6073 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6074 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
6075 bool drained = false;
6079 buf = strstrip(buf);
6080 err = page_counter_memparse(buf, "max", &max);
6084 xchg(&memcg->memory.max, max);
6087 unsigned long nr_pages = page_counter_read(&memcg->memory);
6089 if (nr_pages <= max)
6092 if (signal_pending(current))
6096 drain_all_stock(memcg);
6102 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6108 memcg_memory_event(memcg, MEMCG_OOM);
6109 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6113 memcg_wb_domain_size_changed(memcg);
6117 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6119 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6120 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6121 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6122 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6123 seq_printf(m, "oom_kill %lu\n",
6124 atomic_long_read(&events[MEMCG_OOM_KILL]));
6127 static int memory_events_show(struct seq_file *m, void *v)
6129 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6131 __memory_events_show(m, memcg->memory_events);
6135 static int memory_events_local_show(struct seq_file *m, void *v)
6137 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6139 __memory_events_show(m, memcg->memory_events_local);
6143 static int memory_stat_show(struct seq_file *m, void *v)
6145 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6148 buf = memory_stat_format(memcg);
6156 static int memory_oom_group_show(struct seq_file *m, void *v)
6158 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6160 seq_printf(m, "%d\n", memcg->oom_group);
6165 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6166 char *buf, size_t nbytes, loff_t off)
6168 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6171 buf = strstrip(buf);
6175 ret = kstrtoint(buf, 0, &oom_group);
6179 if (oom_group != 0 && oom_group != 1)
6182 memcg->oom_group = oom_group;
6187 static struct cftype memory_files[] = {
6190 .flags = CFTYPE_NOT_ON_ROOT,
6191 .read_u64 = memory_current_read,
6195 .flags = CFTYPE_NOT_ON_ROOT,
6196 .seq_show = memory_min_show,
6197 .write = memory_min_write,
6201 .flags = CFTYPE_NOT_ON_ROOT,
6202 .seq_show = memory_low_show,
6203 .write = memory_low_write,
6207 .flags = CFTYPE_NOT_ON_ROOT,
6208 .seq_show = memory_high_show,
6209 .write = memory_high_write,
6213 .flags = CFTYPE_NOT_ON_ROOT,
6214 .seq_show = memory_max_show,
6215 .write = memory_max_write,
6219 .flags = CFTYPE_NOT_ON_ROOT,
6220 .file_offset = offsetof(struct mem_cgroup, events_file),
6221 .seq_show = memory_events_show,
6224 .name = "events.local",
6225 .flags = CFTYPE_NOT_ON_ROOT,
6226 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6227 .seq_show = memory_events_local_show,
6231 .seq_show = memory_stat_show,
6234 .name = "oom.group",
6235 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6236 .seq_show = memory_oom_group_show,
6237 .write = memory_oom_group_write,
6242 struct cgroup_subsys memory_cgrp_subsys = {
6243 .css_alloc = mem_cgroup_css_alloc,
6244 .css_online = mem_cgroup_css_online,
6245 .css_offline = mem_cgroup_css_offline,
6246 .css_released = mem_cgroup_css_released,
6247 .css_free = mem_cgroup_css_free,
6248 .css_reset = mem_cgroup_css_reset,
6249 .can_attach = mem_cgroup_can_attach,
6250 .cancel_attach = mem_cgroup_cancel_attach,
6251 .post_attach = mem_cgroup_move_task,
6252 .bind = mem_cgroup_bind,
6253 .dfl_cftypes = memory_files,
6254 .legacy_cftypes = mem_cgroup_legacy_files,
6259 * This function calculates an individual cgroup's effective
6260 * protection which is derived from its own memory.min/low, its
6261 * parent's and siblings' settings, as well as the actual memory
6262 * distribution in the tree.
6264 * The following rules apply to the effective protection values:
6266 * 1. At the first level of reclaim, effective protection is equal to
6267 * the declared protection in memory.min and memory.low.
6269 * 2. To enable safe delegation of the protection configuration, at
6270 * subsequent levels the effective protection is capped to the
6271 * parent's effective protection.
6273 * 3. To make complex and dynamic subtrees easier to configure, the
6274 * user is allowed to overcommit the declared protection at a given
6275 * level. If that is the case, the parent's effective protection is
6276 * distributed to the children in proportion to how much protection
6277 * they have declared and how much of it they are utilizing.
6279 * This makes distribution proportional, but also work-conserving:
6280 * if one cgroup claims much more protection than it uses memory,
6281 * the unused remainder is available to its siblings.
6283 * 4. Conversely, when the declared protection is undercommitted at a
6284 * given level, the distribution of the larger parental protection
6285 * budget is NOT proportional. A cgroup's protection from a sibling
6286 * is capped to its own memory.min/low setting.
6288 * 5. However, to allow protecting recursive subtrees from each other
6289 * without having to declare each individual cgroup's fixed share
6290 * of the ancestor's claim to protection, any unutilized -
6291 * "floating" - protection from up the tree is distributed in
6292 * proportion to each cgroup's *usage*. This makes the protection
6293 * neutral wrt sibling cgroups and lets them compete freely over
6294 * the shared parental protection budget, but it protects the
6295 * subtree as a whole from neighboring subtrees.
6297 * Note that 4. and 5. are not in conflict: 4. is about protecting
6298 * against immediate siblings whereas 5. is about protecting against
6299 * neighboring subtrees.
6301 static unsigned long effective_protection(unsigned long usage,
6302 unsigned long parent_usage,
6303 unsigned long setting,
6304 unsigned long parent_effective,
6305 unsigned long siblings_protected)
6307 unsigned long protected;
6310 protected = min(usage, setting);
6312 * If all cgroups at this level combined claim and use more
6313 * protection then what the parent affords them, distribute
6314 * shares in proportion to utilization.
6316 * We are using actual utilization rather than the statically
6317 * claimed protection in order to be work-conserving: claimed
6318 * but unused protection is available to siblings that would
6319 * otherwise get a smaller chunk than what they claimed.
6321 if (siblings_protected > parent_effective)
6322 return protected * parent_effective / siblings_protected;
6325 * Ok, utilized protection of all children is within what the
6326 * parent affords them, so we know whatever this child claims
6327 * and utilizes is effectively protected.
6329 * If there is unprotected usage beyond this value, reclaim
6330 * will apply pressure in proportion to that amount.
6332 * If there is unutilized protection, the cgroup will be fully
6333 * shielded from reclaim, but we do return a smaller value for
6334 * protection than what the group could enjoy in theory. This
6335 * is okay. With the overcommit distribution above, effective
6336 * protection is always dependent on how memory is actually
6337 * consumed among the siblings anyway.
6342 * If the children aren't claiming (all of) the protection
6343 * afforded to them by the parent, distribute the remainder in
6344 * proportion to the (unprotected) memory of each cgroup. That
6345 * way, cgroups that aren't explicitly prioritized wrt each
6346 * other compete freely over the allowance, but they are
6347 * collectively protected from neighboring trees.
6349 * We're using unprotected memory for the weight so that if
6350 * some cgroups DO claim explicit protection, we don't protect
6351 * the same bytes twice.
6353 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6356 if (parent_effective > siblings_protected && usage > protected) {
6357 unsigned long unclaimed;
6359 unclaimed = parent_effective - siblings_protected;
6360 unclaimed *= usage - protected;
6361 unclaimed /= parent_usage - siblings_protected;
6370 * mem_cgroup_protected - check if memory consumption is in the normal range
6371 * @root: the top ancestor of the sub-tree being checked
6372 * @memcg: the memory cgroup to check
6374 * WARNING: This function is not stateless! It can only be used as part
6375 * of a top-down tree iteration, not for isolated queries.
6377 * Returns one of the following:
6378 * MEMCG_PROT_NONE: cgroup memory is not protected
6379 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
6380 * an unprotected supply of reclaimable memory from other cgroups.
6381 * MEMCG_PROT_MIN: cgroup memory is protected
6383 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
6384 struct mem_cgroup *memcg)
6386 unsigned long usage, parent_usage;
6387 struct mem_cgroup *parent;
6389 if (mem_cgroup_disabled())
6390 return MEMCG_PROT_NONE;
6393 root = root_mem_cgroup;
6395 return MEMCG_PROT_NONE;
6397 usage = page_counter_read(&memcg->memory);
6399 return MEMCG_PROT_NONE;
6401 parent = parent_mem_cgroup(memcg);
6402 /* No parent means a non-hierarchical mode on v1 memcg */
6404 return MEMCG_PROT_NONE;
6406 if (parent == root) {
6407 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6408 memcg->memory.elow = memcg->memory.low;
6412 parent_usage = page_counter_read(&parent->memory);
6414 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6415 READ_ONCE(memcg->memory.min),
6416 READ_ONCE(parent->memory.emin),
6417 atomic_long_read(&parent->memory.children_min_usage)));
6419 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6420 memcg->memory.low, READ_ONCE(parent->memory.elow),
6421 atomic_long_read(&parent->memory.children_low_usage)));
6424 if (usage <= memcg->memory.emin)
6425 return MEMCG_PROT_MIN;
6426 else if (usage <= memcg->memory.elow)
6427 return MEMCG_PROT_LOW;
6429 return MEMCG_PROT_NONE;
6433 * mem_cgroup_try_charge - try charging a page
6434 * @page: page to charge
6435 * @mm: mm context of the victim
6436 * @gfp_mask: reclaim mode
6437 * @memcgp: charged memcg return
6438 * @compound: charge the page as compound or small page
6440 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6441 * pages according to @gfp_mask if necessary.
6443 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6444 * Otherwise, an error code is returned.
6446 * After page->mapping has been set up, the caller must finalize the
6447 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6448 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6450 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
6451 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6454 struct mem_cgroup *memcg = NULL;
6455 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6458 if (mem_cgroup_disabled())
6461 if (PageSwapCache(page)) {
6463 * Every swap fault against a single page tries to charge the
6464 * page, bail as early as possible. shmem_unuse() encounters
6465 * already charged pages, too. The USED bit is protected by
6466 * the page lock, which serializes swap cache removal, which
6467 * in turn serializes uncharging.
6469 VM_BUG_ON_PAGE(!PageLocked(page), page);
6470 if (compound_head(page)->mem_cgroup)
6473 if (do_swap_account) {
6474 swp_entry_t ent = { .val = page_private(page), };
6475 unsigned short id = lookup_swap_cgroup_id(ent);
6478 memcg = mem_cgroup_from_id(id);
6479 if (memcg && !css_tryget_online(&memcg->css))
6486 memcg = get_mem_cgroup_from_mm(mm);
6488 ret = try_charge(memcg, gfp_mask, nr_pages);
6490 css_put(&memcg->css);
6496 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
6497 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6500 struct mem_cgroup *memcg;
6503 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound);
6505 mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask);
6510 * mem_cgroup_commit_charge - commit a page charge
6511 * @page: page to charge
6512 * @memcg: memcg to charge the page to
6513 * @lrucare: page might be on LRU already
6514 * @compound: charge the page as compound or small page
6516 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6517 * after page->mapping has been set up. This must happen atomically
6518 * as part of the page instantiation, i.e. under the page table lock
6519 * for anonymous pages, under the page lock for page and swap cache.
6521 * In addition, the page must not be on the LRU during the commit, to
6522 * prevent racing with task migration. If it might be, use @lrucare.
6524 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6526 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
6527 bool lrucare, bool compound)
6529 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6531 VM_BUG_ON_PAGE(!page->mapping, page);
6532 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
6534 if (mem_cgroup_disabled())
6537 * Swap faults will attempt to charge the same page multiple
6538 * times. But reuse_swap_page() might have removed the page
6539 * from swapcache already, so we can't check PageSwapCache().
6544 commit_charge(page, memcg, lrucare);
6546 local_irq_disable();
6547 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
6548 memcg_check_events(memcg, page);
6551 if (do_memsw_account() && PageSwapCache(page)) {
6552 swp_entry_t entry = { .val = page_private(page) };
6554 * The swap entry might not get freed for a long time,
6555 * let's not wait for it. The page already received a
6556 * memory+swap charge, drop the swap entry duplicate.
6558 mem_cgroup_uncharge_swap(entry, nr_pages);
6563 * mem_cgroup_cancel_charge - cancel a page charge
6564 * @page: page to charge
6565 * @memcg: memcg to charge the page to
6566 * @compound: charge the page as compound or small page
6568 * Cancel a charge transaction started by mem_cgroup_try_charge().
6570 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
6573 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6575 if (mem_cgroup_disabled())
6578 * Swap faults will attempt to charge the same page multiple
6579 * times. But reuse_swap_page() might have removed the page
6580 * from swapcache already, so we can't check PageSwapCache().
6585 cancel_charge(memcg, nr_pages);
6588 struct uncharge_gather {
6589 struct mem_cgroup *memcg;
6590 unsigned long pgpgout;
6591 unsigned long nr_anon;
6592 unsigned long nr_file;
6593 unsigned long nr_kmem;
6594 unsigned long nr_huge;
6595 unsigned long nr_shmem;
6596 struct page *dummy_page;
6599 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6601 memset(ug, 0, sizeof(*ug));
6604 static void uncharge_batch(const struct uncharge_gather *ug)
6606 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
6607 unsigned long flags;
6609 if (!mem_cgroup_is_root(ug->memcg)) {
6610 page_counter_uncharge(&ug->memcg->memory, nr_pages);
6611 if (do_memsw_account())
6612 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
6613 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6614 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6615 memcg_oom_recover(ug->memcg);
6618 local_irq_save(flags);
6619 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
6620 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
6621 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
6622 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
6623 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6624 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, nr_pages);
6625 memcg_check_events(ug->memcg, ug->dummy_page);
6626 local_irq_restore(flags);
6628 if (!mem_cgroup_is_root(ug->memcg))
6629 css_put_many(&ug->memcg->css, nr_pages);
6632 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6634 VM_BUG_ON_PAGE(PageLRU(page), page);
6635 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
6636 !PageHWPoison(page) , page);
6638 if (!page->mem_cgroup)
6642 * Nobody should be changing or seriously looking at
6643 * page->mem_cgroup at this point, we have fully
6644 * exclusive access to the page.
6647 if (ug->memcg != page->mem_cgroup) {
6650 uncharge_gather_clear(ug);
6652 ug->memcg = page->mem_cgroup;
6655 if (!PageKmemcg(page)) {
6656 unsigned int nr_pages = 1;
6658 if (PageTransHuge(page)) {
6659 nr_pages = compound_nr(page);
6660 ug->nr_huge += nr_pages;
6663 ug->nr_anon += nr_pages;
6665 ug->nr_file += nr_pages;
6666 if (PageSwapBacked(page))
6667 ug->nr_shmem += nr_pages;
6671 ug->nr_kmem += compound_nr(page);
6672 __ClearPageKmemcg(page);
6675 ug->dummy_page = page;
6676 page->mem_cgroup = NULL;
6679 static void uncharge_list(struct list_head *page_list)
6681 struct uncharge_gather ug;
6682 struct list_head *next;
6684 uncharge_gather_clear(&ug);
6687 * Note that the list can be a single page->lru; hence the
6688 * do-while loop instead of a simple list_for_each_entry().
6690 next = page_list->next;
6694 page = list_entry(next, struct page, lru);
6695 next = page->lru.next;
6697 uncharge_page(page, &ug);
6698 } while (next != page_list);
6701 uncharge_batch(&ug);
6705 * mem_cgroup_uncharge - uncharge a page
6706 * @page: page to uncharge
6708 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6709 * mem_cgroup_commit_charge().
6711 void mem_cgroup_uncharge(struct page *page)
6713 struct uncharge_gather ug;
6715 if (mem_cgroup_disabled())
6718 /* Don't touch page->lru of any random page, pre-check: */
6719 if (!page->mem_cgroup)
6722 uncharge_gather_clear(&ug);
6723 uncharge_page(page, &ug);
6724 uncharge_batch(&ug);
6728 * mem_cgroup_uncharge_list - uncharge a list of page
6729 * @page_list: list of pages to uncharge
6731 * Uncharge a list of pages previously charged with
6732 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6734 void mem_cgroup_uncharge_list(struct list_head *page_list)
6736 if (mem_cgroup_disabled())
6739 if (!list_empty(page_list))
6740 uncharge_list(page_list);
6744 * mem_cgroup_migrate - charge a page's replacement
6745 * @oldpage: currently circulating page
6746 * @newpage: replacement page
6748 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6749 * be uncharged upon free.
6751 * Both pages must be locked, @newpage->mapping must be set up.
6753 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6755 struct mem_cgroup *memcg;
6756 unsigned int nr_pages;
6757 unsigned long flags;
6759 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6760 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6761 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6762 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6765 if (mem_cgroup_disabled())
6768 /* Page cache replacement: new page already charged? */
6769 if (newpage->mem_cgroup)
6772 /* Swapcache readahead pages can get replaced before being charged */
6773 memcg = oldpage->mem_cgroup;
6777 /* Force-charge the new page. The old one will be freed soon */
6778 nr_pages = hpage_nr_pages(newpage);
6780 page_counter_charge(&memcg->memory, nr_pages);
6781 if (do_memsw_account())
6782 page_counter_charge(&memcg->memsw, nr_pages);
6783 css_get_many(&memcg->css, nr_pages);
6785 commit_charge(newpage, memcg, false);
6787 local_irq_save(flags);
6788 mem_cgroup_charge_statistics(memcg, newpage, PageTransHuge(newpage),
6790 memcg_check_events(memcg, newpage);
6791 local_irq_restore(flags);
6794 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6795 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6797 void mem_cgroup_sk_alloc(struct sock *sk)
6799 struct mem_cgroup *memcg;
6801 if (!mem_cgroup_sockets_enabled)
6804 /* Do not associate the sock with unrelated interrupted task's memcg. */
6809 memcg = mem_cgroup_from_task(current);
6810 if (memcg == root_mem_cgroup)
6812 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6814 if (css_tryget(&memcg->css))
6815 sk->sk_memcg = memcg;
6820 void mem_cgroup_sk_free(struct sock *sk)
6823 css_put(&sk->sk_memcg->css);
6827 * mem_cgroup_charge_skmem - charge socket memory
6828 * @memcg: memcg to charge
6829 * @nr_pages: number of pages to charge
6831 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6832 * @memcg's configured limit, %false if the charge had to be forced.
6834 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6836 gfp_t gfp_mask = GFP_KERNEL;
6838 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6839 struct page_counter *fail;
6841 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6842 memcg->tcpmem_pressure = 0;
6845 page_counter_charge(&memcg->tcpmem, nr_pages);
6846 memcg->tcpmem_pressure = 1;
6850 /* Don't block in the packet receive path */
6852 gfp_mask = GFP_NOWAIT;
6854 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6856 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6859 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6864 * mem_cgroup_uncharge_skmem - uncharge socket memory
6865 * @memcg: memcg to uncharge
6866 * @nr_pages: number of pages to uncharge
6868 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6870 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6871 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6875 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6877 refill_stock(memcg, nr_pages);
6880 static int __init cgroup_memory(char *s)
6884 while ((token = strsep(&s, ",")) != NULL) {
6887 if (!strcmp(token, "nosocket"))
6888 cgroup_memory_nosocket = true;
6889 if (!strcmp(token, "nokmem"))
6890 cgroup_memory_nokmem = true;
6894 __setup("cgroup.memory=", cgroup_memory);
6897 * subsys_initcall() for memory controller.
6899 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6900 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6901 * basically everything that doesn't depend on a specific mem_cgroup structure
6902 * should be initialized from here.
6904 static int __init mem_cgroup_init(void)
6908 #ifdef CONFIG_MEMCG_KMEM
6910 * Kmem cache creation is mostly done with the slab_mutex held,
6911 * so use a workqueue with limited concurrency to avoid stalling
6912 * all worker threads in case lots of cgroups are created and
6913 * destroyed simultaneously.
6915 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
6916 BUG_ON(!memcg_kmem_cache_wq);
6919 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6920 memcg_hotplug_cpu_dead);
6922 for_each_possible_cpu(cpu)
6923 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
6926 for_each_node(node) {
6927 struct mem_cgroup_tree_per_node *rtpn;
6929 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
6930 node_online(node) ? node : NUMA_NO_NODE);
6932 rtpn->rb_root = RB_ROOT;
6933 rtpn->rb_rightmost = NULL;
6934 spin_lock_init(&rtpn->lock);
6935 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6940 subsys_initcall(mem_cgroup_init);
6942 #ifdef CONFIG_MEMCG_SWAP
6943 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
6945 while (!refcount_inc_not_zero(&memcg->id.ref)) {
6947 * The root cgroup cannot be destroyed, so it's refcount must
6950 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
6954 memcg = parent_mem_cgroup(memcg);
6956 memcg = root_mem_cgroup;
6962 * mem_cgroup_swapout - transfer a memsw charge to swap
6963 * @page: page whose memsw charge to transfer
6964 * @entry: swap entry to move the charge to
6966 * Transfer the memsw charge of @page to @entry.
6968 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6970 struct mem_cgroup *memcg, *swap_memcg;
6971 unsigned int nr_entries;
6972 unsigned short oldid;
6974 VM_BUG_ON_PAGE(PageLRU(page), page);
6975 VM_BUG_ON_PAGE(page_count(page), page);
6977 if (!do_memsw_account())
6980 memcg = page->mem_cgroup;
6982 /* Readahead page, never charged */
6987 * In case the memcg owning these pages has been offlined and doesn't
6988 * have an ID allocated to it anymore, charge the closest online
6989 * ancestor for the swap instead and transfer the memory+swap charge.
6991 swap_memcg = mem_cgroup_id_get_online(memcg);
6992 nr_entries = hpage_nr_pages(page);
6993 /* Get references for the tail pages, too */
6995 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
6996 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
6998 VM_BUG_ON_PAGE(oldid, page);
6999 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7001 page->mem_cgroup = NULL;
7003 if (!mem_cgroup_is_root(memcg))
7004 page_counter_uncharge(&memcg->memory, nr_entries);
7006 if (memcg != swap_memcg) {
7007 if (!mem_cgroup_is_root(swap_memcg))
7008 page_counter_charge(&swap_memcg->memsw, nr_entries);
7009 page_counter_uncharge(&memcg->memsw, nr_entries);
7013 * Interrupts should be disabled here because the caller holds the
7014 * i_pages lock which is taken with interrupts-off. It is
7015 * important here to have the interrupts disabled because it is the
7016 * only synchronisation we have for updating the per-CPU variables.
7018 VM_BUG_ON(!irqs_disabled());
7019 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
7021 memcg_check_events(memcg, page);
7023 if (!mem_cgroup_is_root(memcg))
7024 css_put_many(&memcg->css, nr_entries);
7028 * mem_cgroup_try_charge_swap - try charging swap space for a page
7029 * @page: page being added to swap
7030 * @entry: swap entry to charge
7032 * Try to charge @page's memcg for the swap space at @entry.
7034 * Returns 0 on success, -ENOMEM on failure.
7036 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7038 unsigned int nr_pages = hpage_nr_pages(page);
7039 struct page_counter *counter;
7040 struct mem_cgroup *memcg;
7041 unsigned short oldid;
7043 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
7046 memcg = page->mem_cgroup;
7048 /* Readahead page, never charged */
7053 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7057 memcg = mem_cgroup_id_get_online(memcg);
7059 if (!mem_cgroup_is_root(memcg) &&
7060 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7061 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7062 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7063 mem_cgroup_id_put(memcg);
7067 /* Get references for the tail pages, too */
7069 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7070 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7071 VM_BUG_ON_PAGE(oldid, page);
7072 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7078 * mem_cgroup_uncharge_swap - uncharge swap space
7079 * @entry: swap entry to uncharge
7080 * @nr_pages: the amount of swap space to uncharge
7082 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7084 struct mem_cgroup *memcg;
7087 if (!do_swap_account)
7090 id = swap_cgroup_record(entry, 0, nr_pages);
7092 memcg = mem_cgroup_from_id(id);
7094 if (!mem_cgroup_is_root(memcg)) {
7095 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7096 page_counter_uncharge(&memcg->swap, nr_pages);
7098 page_counter_uncharge(&memcg->memsw, nr_pages);
7100 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7101 mem_cgroup_id_put_many(memcg, nr_pages);
7106 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7108 long nr_swap_pages = get_nr_swap_pages();
7110 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7111 return nr_swap_pages;
7112 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7113 nr_swap_pages = min_t(long, nr_swap_pages,
7114 READ_ONCE(memcg->swap.max) -
7115 page_counter_read(&memcg->swap));
7116 return nr_swap_pages;
7119 bool mem_cgroup_swap_full(struct page *page)
7121 struct mem_cgroup *memcg;
7123 VM_BUG_ON_PAGE(!PageLocked(page), page);
7127 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7130 memcg = page->mem_cgroup;
7134 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7135 if (page_counter_read(&memcg->swap) * 2 >=
7136 READ_ONCE(memcg->swap.max))
7142 /* for remember boot option*/
7143 #ifdef CONFIG_MEMCG_SWAP_ENABLED
7144 static int really_do_swap_account __initdata = 1;
7146 static int really_do_swap_account __initdata;
7149 static int __init enable_swap_account(char *s)
7151 if (!strcmp(s, "1"))
7152 really_do_swap_account = 1;
7153 else if (!strcmp(s, "0"))
7154 really_do_swap_account = 0;
7157 __setup("swapaccount=", enable_swap_account);
7159 static u64 swap_current_read(struct cgroup_subsys_state *css,
7162 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7164 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7167 static int swap_max_show(struct seq_file *m, void *v)
7169 return seq_puts_memcg_tunable(m,
7170 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7173 static ssize_t swap_max_write(struct kernfs_open_file *of,
7174 char *buf, size_t nbytes, loff_t off)
7176 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7180 buf = strstrip(buf);
7181 err = page_counter_memparse(buf, "max", &max);
7185 xchg(&memcg->swap.max, max);
7190 static int swap_events_show(struct seq_file *m, void *v)
7192 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7194 seq_printf(m, "max %lu\n",
7195 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7196 seq_printf(m, "fail %lu\n",
7197 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7202 static struct cftype swap_files[] = {
7204 .name = "swap.current",
7205 .flags = CFTYPE_NOT_ON_ROOT,
7206 .read_u64 = swap_current_read,
7210 .flags = CFTYPE_NOT_ON_ROOT,
7211 .seq_show = swap_max_show,
7212 .write = swap_max_write,
7215 .name = "swap.events",
7216 .flags = CFTYPE_NOT_ON_ROOT,
7217 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7218 .seq_show = swap_events_show,
7223 static struct cftype memsw_cgroup_files[] = {
7225 .name = "memsw.usage_in_bytes",
7226 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7227 .read_u64 = mem_cgroup_read_u64,
7230 .name = "memsw.max_usage_in_bytes",
7231 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7232 .write = mem_cgroup_reset,
7233 .read_u64 = mem_cgroup_read_u64,
7236 .name = "memsw.limit_in_bytes",
7237 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7238 .write = mem_cgroup_write,
7239 .read_u64 = mem_cgroup_read_u64,
7242 .name = "memsw.failcnt",
7243 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7244 .write = mem_cgroup_reset,
7245 .read_u64 = mem_cgroup_read_u64,
7247 { }, /* terminate */
7250 static int __init mem_cgroup_swap_init(void)
7252 if (!mem_cgroup_disabled() && really_do_swap_account) {
7253 do_swap_account = 1;
7254 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
7256 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
7257 memsw_cgroup_files));
7261 subsys_initcall(mem_cgroup_swap_init);
7263 #endif /* CONFIG_MEMCG_SWAP */