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(sizeof(*new) + size, GFP_KERNEL);
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(sizeof(*map) + size, GFP_KERNEL);
383 memcg_free_shrinker_maps(memcg);
387 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
389 mutex_unlock(&memcg_shrinker_map_mutex);
394 int memcg_expand_shrinker_maps(int new_id)
396 int size, old_size, ret = 0;
397 struct mem_cgroup *memcg;
399 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
400 old_size = memcg_shrinker_map_size;
401 if (size <= old_size)
404 mutex_lock(&memcg_shrinker_map_mutex);
405 if (!root_mem_cgroup)
408 for_each_mem_cgroup(memcg) {
409 if (mem_cgroup_is_root(memcg))
411 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
413 mem_cgroup_iter_break(NULL, memcg);
419 memcg_shrinker_map_size = size;
420 mutex_unlock(&memcg_shrinker_map_mutex);
424 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
426 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
427 struct memcg_shrinker_map *map;
430 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
431 /* Pairs with smp mb in shrink_slab() */
432 smp_mb__before_atomic();
433 set_bit(shrinker_id, map->map);
439 * mem_cgroup_css_from_page - css of the memcg associated with a page
440 * @page: page of interest
442 * If memcg is bound to the default hierarchy, css of the memcg associated
443 * with @page is returned. The returned css remains associated with @page
444 * until it is released.
446 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
449 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
451 struct mem_cgroup *memcg;
453 memcg = page->mem_cgroup;
455 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
456 memcg = root_mem_cgroup;
462 * page_cgroup_ino - return inode number of the memcg a page is charged to
465 * Look up the closest online ancestor of the memory cgroup @page is charged to
466 * and return its inode number or 0 if @page is not charged to any cgroup. It
467 * is safe to call this function without holding a reference to @page.
469 * Note, this function is inherently racy, because there is nothing to prevent
470 * the cgroup inode from getting torn down and potentially reallocated a moment
471 * after page_cgroup_ino() returns, so it only should be used by callers that
472 * do not care (such as procfs interfaces).
474 ino_t page_cgroup_ino(struct page *page)
476 struct mem_cgroup *memcg;
477 unsigned long ino = 0;
480 if (PageSlab(page) && !PageTail(page))
481 memcg = memcg_from_slab_page(page);
483 memcg = READ_ONCE(page->mem_cgroup);
484 while (memcg && !(memcg->css.flags & CSS_ONLINE))
485 memcg = parent_mem_cgroup(memcg);
487 ino = cgroup_ino(memcg->css.cgroup);
492 static struct mem_cgroup_per_node *
493 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
495 int nid = page_to_nid(page);
497 return memcg->nodeinfo[nid];
500 static struct mem_cgroup_tree_per_node *
501 soft_limit_tree_node(int nid)
503 return soft_limit_tree.rb_tree_per_node[nid];
506 static struct mem_cgroup_tree_per_node *
507 soft_limit_tree_from_page(struct page *page)
509 int nid = page_to_nid(page);
511 return soft_limit_tree.rb_tree_per_node[nid];
514 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
515 struct mem_cgroup_tree_per_node *mctz,
516 unsigned long new_usage_in_excess)
518 struct rb_node **p = &mctz->rb_root.rb_node;
519 struct rb_node *parent = NULL;
520 struct mem_cgroup_per_node *mz_node;
521 bool rightmost = true;
526 mz->usage_in_excess = new_usage_in_excess;
527 if (!mz->usage_in_excess)
531 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
533 if (mz->usage_in_excess < mz_node->usage_in_excess) {
539 * We can't avoid mem cgroups that are over their soft
540 * limit by the same amount
542 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
547 mctz->rb_rightmost = &mz->tree_node;
549 rb_link_node(&mz->tree_node, parent, p);
550 rb_insert_color(&mz->tree_node, &mctz->rb_root);
554 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
555 struct mem_cgroup_tree_per_node *mctz)
560 if (&mz->tree_node == mctz->rb_rightmost)
561 mctz->rb_rightmost = rb_prev(&mz->tree_node);
563 rb_erase(&mz->tree_node, &mctz->rb_root);
567 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
568 struct mem_cgroup_tree_per_node *mctz)
572 spin_lock_irqsave(&mctz->lock, flags);
573 __mem_cgroup_remove_exceeded(mz, mctz);
574 spin_unlock_irqrestore(&mctz->lock, flags);
577 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
579 unsigned long nr_pages = page_counter_read(&memcg->memory);
580 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
581 unsigned long excess = 0;
583 if (nr_pages > soft_limit)
584 excess = nr_pages - soft_limit;
589 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
591 unsigned long excess;
592 struct mem_cgroup_per_node *mz;
593 struct mem_cgroup_tree_per_node *mctz;
595 mctz = soft_limit_tree_from_page(page);
599 * Necessary to update all ancestors when hierarchy is used.
600 * because their event counter is not touched.
602 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
603 mz = mem_cgroup_page_nodeinfo(memcg, page);
604 excess = soft_limit_excess(memcg);
606 * We have to update the tree if mz is on RB-tree or
607 * mem is over its softlimit.
609 if (excess || mz->on_tree) {
612 spin_lock_irqsave(&mctz->lock, flags);
613 /* if on-tree, remove it */
615 __mem_cgroup_remove_exceeded(mz, mctz);
617 * Insert again. mz->usage_in_excess will be updated.
618 * If excess is 0, no tree ops.
620 __mem_cgroup_insert_exceeded(mz, mctz, excess);
621 spin_unlock_irqrestore(&mctz->lock, flags);
626 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
628 struct mem_cgroup_tree_per_node *mctz;
629 struct mem_cgroup_per_node *mz;
633 mz = mem_cgroup_nodeinfo(memcg, nid);
634 mctz = soft_limit_tree_node(nid);
636 mem_cgroup_remove_exceeded(mz, mctz);
640 static struct mem_cgroup_per_node *
641 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
643 struct mem_cgroup_per_node *mz;
647 if (!mctz->rb_rightmost)
648 goto done; /* Nothing to reclaim from */
650 mz = rb_entry(mctz->rb_rightmost,
651 struct mem_cgroup_per_node, tree_node);
653 * Remove the node now but someone else can add it back,
654 * we will to add it back at the end of reclaim to its correct
655 * position in the tree.
657 __mem_cgroup_remove_exceeded(mz, mctz);
658 if (!soft_limit_excess(mz->memcg) ||
659 !css_tryget_online(&mz->memcg->css))
665 static struct mem_cgroup_per_node *
666 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
668 struct mem_cgroup_per_node *mz;
670 spin_lock_irq(&mctz->lock);
671 mz = __mem_cgroup_largest_soft_limit_node(mctz);
672 spin_unlock_irq(&mctz->lock);
677 * __mod_memcg_state - update cgroup memory statistics
678 * @memcg: the memory cgroup
679 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
680 * @val: delta to add to the counter, can be negative
682 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
686 if (mem_cgroup_disabled())
689 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
690 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
691 struct mem_cgroup *mi;
694 * Batch local counters to keep them in sync with
695 * the hierarchical ones.
697 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
698 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
699 atomic_long_add(x, &mi->vmstats[idx]);
702 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
705 static struct mem_cgroup_per_node *
706 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
708 struct mem_cgroup *parent;
710 parent = parent_mem_cgroup(pn->memcg);
713 return mem_cgroup_nodeinfo(parent, nid);
717 * __mod_lruvec_state - update lruvec memory statistics
718 * @lruvec: the lruvec
719 * @idx: the stat item
720 * @val: delta to add to the counter, can be negative
722 * The lruvec is the intersection of the NUMA node and a cgroup. This
723 * function updates the all three counters that are affected by a
724 * change of state at this level: per-node, per-cgroup, per-lruvec.
726 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
729 pg_data_t *pgdat = lruvec_pgdat(lruvec);
730 struct mem_cgroup_per_node *pn;
731 struct mem_cgroup *memcg;
735 __mod_node_page_state(pgdat, idx, val);
737 if (mem_cgroup_disabled())
740 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
744 __mod_memcg_state(memcg, idx, val);
747 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
749 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
750 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
751 struct mem_cgroup_per_node *pi;
753 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
754 atomic_long_add(x, &pi->lruvec_stat[idx]);
757 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
760 void __mod_lruvec_slab_state(void *p, enum node_stat_item idx, int val)
762 struct page *page = virt_to_head_page(p);
763 pg_data_t *pgdat = page_pgdat(page);
764 struct mem_cgroup *memcg;
765 struct lruvec *lruvec;
768 memcg = memcg_from_slab_page(page);
770 /* Untracked pages have no memcg, no lruvec. Update only the node */
771 if (!memcg || memcg == root_mem_cgroup) {
772 __mod_node_page_state(pgdat, idx, val);
774 lruvec = mem_cgroup_lruvec(memcg, pgdat);
775 __mod_lruvec_state(lruvec, idx, val);
781 * __count_memcg_events - account VM events in a cgroup
782 * @memcg: the memory cgroup
783 * @idx: the event item
784 * @count: the number of events that occured
786 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
791 if (mem_cgroup_disabled())
794 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
795 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
796 struct mem_cgroup *mi;
799 * Batch local counters to keep them in sync with
800 * the hierarchical ones.
802 __this_cpu_add(memcg->vmstats_local->events[idx], x);
803 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
804 atomic_long_add(x, &mi->vmevents[idx]);
807 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
810 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
812 return atomic_long_read(&memcg->vmevents[event]);
815 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
820 for_each_possible_cpu(cpu)
821 x += per_cpu(memcg->vmstats_local->events[event], cpu);
825 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
827 bool compound, int nr_pages)
830 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
831 * counted as CACHE even if it's on ANON LRU.
834 __mod_memcg_state(memcg, MEMCG_RSS, nr_pages);
836 __mod_memcg_state(memcg, MEMCG_CACHE, nr_pages);
837 if (PageSwapBacked(page))
838 __mod_memcg_state(memcg, NR_SHMEM, nr_pages);
842 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
843 __mod_memcg_state(memcg, MEMCG_RSS_HUGE, nr_pages);
846 /* pagein of a big page is an event. So, ignore page size */
848 __count_memcg_events(memcg, PGPGIN, 1);
850 __count_memcg_events(memcg, PGPGOUT, 1);
851 nr_pages = -nr_pages; /* for event */
854 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
857 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
858 enum mem_cgroup_events_target target)
860 unsigned long val, next;
862 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
863 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
864 /* from time_after() in jiffies.h */
865 if ((long)(next - val) < 0) {
867 case MEM_CGROUP_TARGET_THRESH:
868 next = val + THRESHOLDS_EVENTS_TARGET;
870 case MEM_CGROUP_TARGET_SOFTLIMIT:
871 next = val + SOFTLIMIT_EVENTS_TARGET;
876 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
883 * Check events in order.
886 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
888 /* threshold event is triggered in finer grain than soft limit */
889 if (unlikely(mem_cgroup_event_ratelimit(memcg,
890 MEM_CGROUP_TARGET_THRESH))) {
893 do_softlimit = mem_cgroup_event_ratelimit(memcg,
894 MEM_CGROUP_TARGET_SOFTLIMIT);
895 mem_cgroup_threshold(memcg);
896 if (unlikely(do_softlimit))
897 mem_cgroup_update_tree(memcg, page);
901 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
904 * mm_update_next_owner() may clear mm->owner to NULL
905 * if it races with swapoff, page migration, etc.
906 * So this can be called with p == NULL.
911 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
913 EXPORT_SYMBOL(mem_cgroup_from_task);
916 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
917 * @mm: mm from which memcg should be extracted. It can be NULL.
919 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
920 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
923 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
925 struct mem_cgroup *memcg;
927 if (mem_cgroup_disabled())
933 * Page cache insertions can happen withou an
934 * actual mm context, e.g. during disk probing
935 * on boot, loopback IO, acct() writes etc.
938 memcg = root_mem_cgroup;
940 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
941 if (unlikely(!memcg))
942 memcg = root_mem_cgroup;
944 } while (!css_tryget(&memcg->css));
948 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
951 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
952 * @page: page from which memcg should be extracted.
954 * Obtain a reference on page->memcg and returns it if successful. Otherwise
955 * root_mem_cgroup is returned.
957 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
959 struct mem_cgroup *memcg = page->mem_cgroup;
961 if (mem_cgroup_disabled())
965 if (!memcg || !css_tryget_online(&memcg->css))
966 memcg = root_mem_cgroup;
970 EXPORT_SYMBOL(get_mem_cgroup_from_page);
973 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
975 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
977 if (unlikely(current->active_memcg)) {
978 struct mem_cgroup *memcg = root_mem_cgroup;
981 if (css_tryget_online(¤t->active_memcg->css))
982 memcg = current->active_memcg;
986 return get_mem_cgroup_from_mm(current->mm);
990 * mem_cgroup_iter - iterate over memory cgroup hierarchy
991 * @root: hierarchy root
992 * @prev: previously returned memcg, NULL on first invocation
993 * @reclaim: cookie for shared reclaim walks, NULL for full walks
995 * Returns references to children of the hierarchy below @root, or
996 * @root itself, or %NULL after a full round-trip.
998 * Caller must pass the return value in @prev on subsequent
999 * invocations for reference counting, or use mem_cgroup_iter_break()
1000 * to cancel a hierarchy walk before the round-trip is complete.
1002 * Reclaimers can specify a node and a priority level in @reclaim to
1003 * divide up the memcgs in the hierarchy among all concurrent
1004 * reclaimers operating on the same node and priority.
1006 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1007 struct mem_cgroup *prev,
1008 struct mem_cgroup_reclaim_cookie *reclaim)
1010 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1011 struct cgroup_subsys_state *css = NULL;
1012 struct mem_cgroup *memcg = NULL;
1013 struct mem_cgroup *pos = NULL;
1015 if (mem_cgroup_disabled())
1019 root = root_mem_cgroup;
1021 if (prev && !reclaim)
1024 if (!root->use_hierarchy && root != root_mem_cgroup) {
1033 struct mem_cgroup_per_node *mz;
1035 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1038 if (prev && reclaim->generation != iter->generation)
1042 pos = READ_ONCE(iter->position);
1043 if (!pos || css_tryget(&pos->css))
1046 * css reference reached zero, so iter->position will
1047 * be cleared by ->css_released. However, we should not
1048 * rely on this happening soon, because ->css_released
1049 * is called from a work queue, and by busy-waiting we
1050 * might block it. So we clear iter->position right
1053 (void)cmpxchg(&iter->position, pos, NULL);
1061 css = css_next_descendant_pre(css, &root->css);
1064 * Reclaimers share the hierarchy walk, and a
1065 * new one might jump in right at the end of
1066 * the hierarchy - make sure they see at least
1067 * one group and restart from the beginning.
1075 * Verify the css and acquire a reference. The root
1076 * is provided by the caller, so we know it's alive
1077 * and kicking, and don't take an extra reference.
1079 memcg = mem_cgroup_from_css(css);
1081 if (css == &root->css)
1084 if (css_tryget(css))
1092 * The position could have already been updated by a competing
1093 * thread, so check that the value hasn't changed since we read
1094 * it to avoid reclaiming from the same cgroup twice.
1096 (void)cmpxchg(&iter->position, pos, memcg);
1104 reclaim->generation = iter->generation;
1110 if (prev && prev != root)
1111 css_put(&prev->css);
1117 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1118 * @root: hierarchy root
1119 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1121 void mem_cgroup_iter_break(struct mem_cgroup *root,
1122 struct mem_cgroup *prev)
1125 root = root_mem_cgroup;
1126 if (prev && prev != root)
1127 css_put(&prev->css);
1130 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1131 struct mem_cgroup *dead_memcg)
1133 struct mem_cgroup_reclaim_iter *iter;
1134 struct mem_cgroup_per_node *mz;
1137 for_each_node(nid) {
1138 mz = mem_cgroup_nodeinfo(from, nid);
1140 cmpxchg(&iter->position, dead_memcg, NULL);
1144 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1146 struct mem_cgroup *memcg = dead_memcg;
1147 struct mem_cgroup *last;
1150 __invalidate_reclaim_iterators(memcg, dead_memcg);
1152 } while ((memcg = parent_mem_cgroup(memcg)));
1155 * When cgruop1 non-hierarchy mode is used,
1156 * parent_mem_cgroup() does not walk all the way up to the
1157 * cgroup root (root_mem_cgroup). So we have to handle
1158 * dead_memcg from cgroup root separately.
1160 if (last != root_mem_cgroup)
1161 __invalidate_reclaim_iterators(root_mem_cgroup,
1166 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1167 * @memcg: hierarchy root
1168 * @fn: function to call for each task
1169 * @arg: argument passed to @fn
1171 * This function iterates over tasks attached to @memcg or to any of its
1172 * descendants and calls @fn for each task. If @fn returns a non-zero
1173 * value, the function breaks the iteration loop and returns the value.
1174 * Otherwise, it will iterate over all tasks and return 0.
1176 * This function must not be called for the root memory cgroup.
1178 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1179 int (*fn)(struct task_struct *, void *), void *arg)
1181 struct mem_cgroup *iter;
1184 BUG_ON(memcg == root_mem_cgroup);
1186 for_each_mem_cgroup_tree(iter, memcg) {
1187 struct css_task_iter it;
1188 struct task_struct *task;
1190 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1191 while (!ret && (task = css_task_iter_next(&it)))
1192 ret = fn(task, arg);
1193 css_task_iter_end(&it);
1195 mem_cgroup_iter_break(memcg, iter);
1203 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1205 * @pgdat: pgdat of the page
1207 * This function is only safe when following the LRU page isolation
1208 * and putback protocol: the LRU lock must be held, and the page must
1209 * either be PageLRU() or the caller must have isolated/allocated it.
1211 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1213 struct mem_cgroup_per_node *mz;
1214 struct mem_cgroup *memcg;
1215 struct lruvec *lruvec;
1217 if (mem_cgroup_disabled()) {
1218 lruvec = &pgdat->__lruvec;
1222 memcg = page->mem_cgroup;
1224 * Swapcache readahead pages are added to the LRU - and
1225 * possibly migrated - before they are charged.
1228 memcg = root_mem_cgroup;
1230 mz = mem_cgroup_page_nodeinfo(memcg, page);
1231 lruvec = &mz->lruvec;
1234 * Since a node can be onlined after the mem_cgroup was created,
1235 * we have to be prepared to initialize lruvec->zone here;
1236 * and if offlined then reonlined, we need to reinitialize it.
1238 if (unlikely(lruvec->pgdat != pgdat))
1239 lruvec->pgdat = pgdat;
1244 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1245 * @lruvec: mem_cgroup per zone lru vector
1246 * @lru: index of lru list the page is sitting on
1247 * @zid: zone id of the accounted pages
1248 * @nr_pages: positive when adding or negative when removing
1250 * This function must be called under lru_lock, just before a page is added
1251 * to or just after a page is removed from an lru list (that ordering being
1252 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1254 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1255 int zid, int nr_pages)
1257 struct mem_cgroup_per_node *mz;
1258 unsigned long *lru_size;
1261 if (mem_cgroup_disabled())
1264 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1265 lru_size = &mz->lru_zone_size[zid][lru];
1268 *lru_size += nr_pages;
1271 if (WARN_ONCE(size < 0,
1272 "%s(%p, %d, %d): lru_size %ld\n",
1273 __func__, lruvec, lru, nr_pages, size)) {
1279 *lru_size += nr_pages;
1283 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1284 * @memcg: the memory cgroup
1286 * Returns the maximum amount of memory @mem can be charged with, in
1289 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1291 unsigned long margin = 0;
1292 unsigned long count;
1293 unsigned long limit;
1295 count = page_counter_read(&memcg->memory);
1296 limit = READ_ONCE(memcg->memory.max);
1298 margin = limit - count;
1300 if (do_memsw_account()) {
1301 count = page_counter_read(&memcg->memsw);
1302 limit = READ_ONCE(memcg->memsw.max);
1304 margin = min(margin, limit - count);
1313 * A routine for checking "mem" is under move_account() or not.
1315 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1316 * moving cgroups. This is for waiting at high-memory pressure
1319 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1321 struct mem_cgroup *from;
1322 struct mem_cgroup *to;
1325 * Unlike task_move routines, we access mc.to, mc.from not under
1326 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1328 spin_lock(&mc.lock);
1334 ret = mem_cgroup_is_descendant(from, memcg) ||
1335 mem_cgroup_is_descendant(to, memcg);
1337 spin_unlock(&mc.lock);
1341 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1343 if (mc.moving_task && current != mc.moving_task) {
1344 if (mem_cgroup_under_move(memcg)) {
1346 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1347 /* moving charge context might have finished. */
1350 finish_wait(&mc.waitq, &wait);
1357 static char *memory_stat_format(struct mem_cgroup *memcg)
1362 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1367 * Provide statistics on the state of the memory subsystem as
1368 * well as cumulative event counters that show past behavior.
1370 * This list is ordered following a combination of these gradients:
1371 * 1) generic big picture -> specifics and details
1372 * 2) reflecting userspace activity -> reflecting kernel heuristics
1374 * Current memory state:
1377 seq_buf_printf(&s, "anon %llu\n",
1378 (u64)memcg_page_state(memcg, MEMCG_RSS) *
1380 seq_buf_printf(&s, "file %llu\n",
1381 (u64)memcg_page_state(memcg, MEMCG_CACHE) *
1383 seq_buf_printf(&s, "kernel_stack %llu\n",
1384 (u64)memcg_page_state(memcg, MEMCG_KERNEL_STACK_KB) *
1386 seq_buf_printf(&s, "slab %llu\n",
1387 (u64)(memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) +
1388 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE)) *
1390 seq_buf_printf(&s, "sock %llu\n",
1391 (u64)memcg_page_state(memcg, MEMCG_SOCK) *
1394 seq_buf_printf(&s, "shmem %llu\n",
1395 (u64)memcg_page_state(memcg, NR_SHMEM) *
1397 seq_buf_printf(&s, "file_mapped %llu\n",
1398 (u64)memcg_page_state(memcg, NR_FILE_MAPPED) *
1400 seq_buf_printf(&s, "file_dirty %llu\n",
1401 (u64)memcg_page_state(memcg, NR_FILE_DIRTY) *
1403 seq_buf_printf(&s, "file_writeback %llu\n",
1404 (u64)memcg_page_state(memcg, NR_WRITEBACK) *
1408 * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter
1409 * with the NR_ANON_THP vm counter, but right now it's a pain in the
1410 * arse because it requires migrating the work out of rmap to a place
1411 * where the page->mem_cgroup is set up and stable.
1413 seq_buf_printf(&s, "anon_thp %llu\n",
1414 (u64)memcg_page_state(memcg, MEMCG_RSS_HUGE) *
1417 for (i = 0; i < NR_LRU_LISTS; i++)
1418 seq_buf_printf(&s, "%s %llu\n", lru_list_name(i),
1419 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
1422 seq_buf_printf(&s, "slab_reclaimable %llu\n",
1423 (u64)memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) *
1425 seq_buf_printf(&s, "slab_unreclaimable %llu\n",
1426 (u64)memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE) *
1429 /* Accumulated memory events */
1431 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1432 memcg_events(memcg, PGFAULT));
1433 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1434 memcg_events(memcg, PGMAJFAULT));
1436 seq_buf_printf(&s, "workingset_refault %lu\n",
1437 memcg_page_state(memcg, WORKINGSET_REFAULT));
1438 seq_buf_printf(&s, "workingset_activate %lu\n",
1439 memcg_page_state(memcg, WORKINGSET_ACTIVATE));
1440 seq_buf_printf(&s, "workingset_nodereclaim %lu\n",
1441 memcg_page_state(memcg, WORKINGSET_NODERECLAIM));
1443 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1444 memcg_events(memcg, PGREFILL));
1445 seq_buf_printf(&s, "pgscan %lu\n",
1446 memcg_events(memcg, PGSCAN_KSWAPD) +
1447 memcg_events(memcg, PGSCAN_DIRECT));
1448 seq_buf_printf(&s, "pgsteal %lu\n",
1449 memcg_events(memcg, PGSTEAL_KSWAPD) +
1450 memcg_events(memcg, PGSTEAL_DIRECT));
1451 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1452 memcg_events(memcg, PGACTIVATE));
1453 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1454 memcg_events(memcg, PGDEACTIVATE));
1455 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1456 memcg_events(memcg, PGLAZYFREE));
1457 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1458 memcg_events(memcg, PGLAZYFREED));
1460 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1461 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1462 memcg_events(memcg, THP_FAULT_ALLOC));
1463 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1464 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1465 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1467 /* The above should easily fit into one page */
1468 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1473 #define K(x) ((x) << (PAGE_SHIFT-10))
1475 * mem_cgroup_print_oom_context: Print OOM information relevant to
1476 * memory controller.
1477 * @memcg: The memory cgroup that went over limit
1478 * @p: Task that is going to be killed
1480 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1483 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1488 pr_cont(",oom_memcg=");
1489 pr_cont_cgroup_path(memcg->css.cgroup);
1491 pr_cont(",global_oom");
1493 pr_cont(",task_memcg=");
1494 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1500 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1501 * memory controller.
1502 * @memcg: The memory cgroup that went over limit
1504 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1508 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1509 K((u64)page_counter_read(&memcg->memory)),
1510 K((u64)memcg->memory.max), memcg->memory.failcnt);
1511 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1512 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1513 K((u64)page_counter_read(&memcg->swap)),
1514 K((u64)memcg->swap.max), memcg->swap.failcnt);
1516 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1517 K((u64)page_counter_read(&memcg->memsw)),
1518 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1519 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1520 K((u64)page_counter_read(&memcg->kmem)),
1521 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1524 pr_info("Memory cgroup stats for ");
1525 pr_cont_cgroup_path(memcg->css.cgroup);
1527 buf = memory_stat_format(memcg);
1535 * Return the memory (and swap, if configured) limit for a memcg.
1537 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1541 max = memcg->memory.max;
1542 if (mem_cgroup_swappiness(memcg)) {
1543 unsigned long memsw_max;
1544 unsigned long swap_max;
1546 memsw_max = memcg->memsw.max;
1547 swap_max = memcg->swap.max;
1548 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1549 max = min(max + swap_max, memsw_max);
1554 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1556 return page_counter_read(&memcg->memory);
1559 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1562 struct oom_control oc = {
1566 .gfp_mask = gfp_mask,
1571 if (mutex_lock_killable(&oom_lock))
1574 * A few threads which were not waiting at mutex_lock_killable() can
1575 * fail to bail out. Therefore, check again after holding oom_lock.
1577 ret = should_force_charge() || out_of_memory(&oc);
1578 mutex_unlock(&oom_lock);
1582 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1585 unsigned long *total_scanned)
1587 struct mem_cgroup *victim = NULL;
1590 unsigned long excess;
1591 unsigned long nr_scanned;
1592 struct mem_cgroup_reclaim_cookie reclaim = {
1596 excess = soft_limit_excess(root_memcg);
1599 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1604 * If we have not been able to reclaim
1605 * anything, it might because there are
1606 * no reclaimable pages under this hierarchy
1611 * We want to do more targeted reclaim.
1612 * excess >> 2 is not to excessive so as to
1613 * reclaim too much, nor too less that we keep
1614 * coming back to reclaim from this cgroup
1616 if (total >= (excess >> 2) ||
1617 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1622 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1623 pgdat, &nr_scanned);
1624 *total_scanned += nr_scanned;
1625 if (!soft_limit_excess(root_memcg))
1628 mem_cgroup_iter_break(root_memcg, victim);
1632 #ifdef CONFIG_LOCKDEP
1633 static struct lockdep_map memcg_oom_lock_dep_map = {
1634 .name = "memcg_oom_lock",
1638 static DEFINE_SPINLOCK(memcg_oom_lock);
1641 * Check OOM-Killer is already running under our hierarchy.
1642 * If someone is running, return false.
1644 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1646 struct mem_cgroup *iter, *failed = NULL;
1648 spin_lock(&memcg_oom_lock);
1650 for_each_mem_cgroup_tree(iter, memcg) {
1651 if (iter->oom_lock) {
1653 * this subtree of our hierarchy is already locked
1654 * so we cannot give a lock.
1657 mem_cgroup_iter_break(memcg, iter);
1660 iter->oom_lock = true;
1665 * OK, we failed to lock the whole subtree so we have
1666 * to clean up what we set up to the failing subtree
1668 for_each_mem_cgroup_tree(iter, memcg) {
1669 if (iter == failed) {
1670 mem_cgroup_iter_break(memcg, iter);
1673 iter->oom_lock = false;
1676 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1678 spin_unlock(&memcg_oom_lock);
1683 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1685 struct mem_cgroup *iter;
1687 spin_lock(&memcg_oom_lock);
1688 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1689 for_each_mem_cgroup_tree(iter, memcg)
1690 iter->oom_lock = false;
1691 spin_unlock(&memcg_oom_lock);
1694 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1696 struct mem_cgroup *iter;
1698 spin_lock(&memcg_oom_lock);
1699 for_each_mem_cgroup_tree(iter, memcg)
1701 spin_unlock(&memcg_oom_lock);
1704 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1706 struct mem_cgroup *iter;
1709 * When a new child is created while the hierarchy is under oom,
1710 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1712 spin_lock(&memcg_oom_lock);
1713 for_each_mem_cgroup_tree(iter, memcg)
1714 if (iter->under_oom > 0)
1716 spin_unlock(&memcg_oom_lock);
1719 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1721 struct oom_wait_info {
1722 struct mem_cgroup *memcg;
1723 wait_queue_entry_t wait;
1726 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1727 unsigned mode, int sync, void *arg)
1729 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1730 struct mem_cgroup *oom_wait_memcg;
1731 struct oom_wait_info *oom_wait_info;
1733 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1734 oom_wait_memcg = oom_wait_info->memcg;
1736 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1737 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1739 return autoremove_wake_function(wait, mode, sync, arg);
1742 static void memcg_oom_recover(struct mem_cgroup *memcg)
1745 * For the following lockless ->under_oom test, the only required
1746 * guarantee is that it must see the state asserted by an OOM when
1747 * this function is called as a result of userland actions
1748 * triggered by the notification of the OOM. This is trivially
1749 * achieved by invoking mem_cgroup_mark_under_oom() before
1750 * triggering notification.
1752 if (memcg && memcg->under_oom)
1753 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1763 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1765 enum oom_status ret;
1768 if (order > PAGE_ALLOC_COSTLY_ORDER)
1771 memcg_memory_event(memcg, MEMCG_OOM);
1774 * We are in the middle of the charge context here, so we
1775 * don't want to block when potentially sitting on a callstack
1776 * that holds all kinds of filesystem and mm locks.
1778 * cgroup1 allows disabling the OOM killer and waiting for outside
1779 * handling until the charge can succeed; remember the context and put
1780 * the task to sleep at the end of the page fault when all locks are
1783 * On the other hand, in-kernel OOM killer allows for an async victim
1784 * memory reclaim (oom_reaper) and that means that we are not solely
1785 * relying on the oom victim to make a forward progress and we can
1786 * invoke the oom killer here.
1788 * Please note that mem_cgroup_out_of_memory might fail to find a
1789 * victim and then we have to bail out from the charge path.
1791 if (memcg->oom_kill_disable) {
1792 if (!current->in_user_fault)
1794 css_get(&memcg->css);
1795 current->memcg_in_oom = memcg;
1796 current->memcg_oom_gfp_mask = mask;
1797 current->memcg_oom_order = order;
1802 mem_cgroup_mark_under_oom(memcg);
1804 locked = mem_cgroup_oom_trylock(memcg);
1807 mem_cgroup_oom_notify(memcg);
1809 mem_cgroup_unmark_under_oom(memcg);
1810 if (mem_cgroup_out_of_memory(memcg, mask, order))
1816 mem_cgroup_oom_unlock(memcg);
1822 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1823 * @handle: actually kill/wait or just clean up the OOM state
1825 * This has to be called at the end of a page fault if the memcg OOM
1826 * handler was enabled.
1828 * Memcg supports userspace OOM handling where failed allocations must
1829 * sleep on a waitqueue until the userspace task resolves the
1830 * situation. Sleeping directly in the charge context with all kinds
1831 * of locks held is not a good idea, instead we remember an OOM state
1832 * in the task and mem_cgroup_oom_synchronize() has to be called at
1833 * the end of the page fault to complete the OOM handling.
1835 * Returns %true if an ongoing memcg OOM situation was detected and
1836 * completed, %false otherwise.
1838 bool mem_cgroup_oom_synchronize(bool handle)
1840 struct mem_cgroup *memcg = current->memcg_in_oom;
1841 struct oom_wait_info owait;
1844 /* OOM is global, do not handle */
1851 owait.memcg = memcg;
1852 owait.wait.flags = 0;
1853 owait.wait.func = memcg_oom_wake_function;
1854 owait.wait.private = current;
1855 INIT_LIST_HEAD(&owait.wait.entry);
1857 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1858 mem_cgroup_mark_under_oom(memcg);
1860 locked = mem_cgroup_oom_trylock(memcg);
1863 mem_cgroup_oom_notify(memcg);
1865 if (locked && !memcg->oom_kill_disable) {
1866 mem_cgroup_unmark_under_oom(memcg);
1867 finish_wait(&memcg_oom_waitq, &owait.wait);
1868 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1869 current->memcg_oom_order);
1872 mem_cgroup_unmark_under_oom(memcg);
1873 finish_wait(&memcg_oom_waitq, &owait.wait);
1877 mem_cgroup_oom_unlock(memcg);
1879 * There is no guarantee that an OOM-lock contender
1880 * sees the wakeups triggered by the OOM kill
1881 * uncharges. Wake any sleepers explicitely.
1883 memcg_oom_recover(memcg);
1886 current->memcg_in_oom = NULL;
1887 css_put(&memcg->css);
1892 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1893 * @victim: task to be killed by the OOM killer
1894 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1896 * Returns a pointer to a memory cgroup, which has to be cleaned up
1897 * by killing all belonging OOM-killable tasks.
1899 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1901 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1902 struct mem_cgroup *oom_domain)
1904 struct mem_cgroup *oom_group = NULL;
1905 struct mem_cgroup *memcg;
1907 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
1911 oom_domain = root_mem_cgroup;
1915 memcg = mem_cgroup_from_task(victim);
1916 if (memcg == root_mem_cgroup)
1920 * Traverse the memory cgroup hierarchy from the victim task's
1921 * cgroup up to the OOMing cgroup (or root) to find the
1922 * highest-level memory cgroup with oom.group set.
1924 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1925 if (memcg->oom_group)
1928 if (memcg == oom_domain)
1933 css_get(&oom_group->css);
1940 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
1942 pr_info("Tasks in ");
1943 pr_cont_cgroup_path(memcg->css.cgroup);
1944 pr_cont(" are going to be killed due to memory.oom.group set\n");
1948 * lock_page_memcg - lock a page->mem_cgroup binding
1951 * This function protects unlocked LRU pages from being moved to
1954 * It ensures lifetime of the returned memcg. Caller is responsible
1955 * for the lifetime of the page; __unlock_page_memcg() is available
1956 * when @page might get freed inside the locked section.
1958 struct mem_cgroup *lock_page_memcg(struct page *page)
1960 struct mem_cgroup *memcg;
1961 unsigned long flags;
1964 * The RCU lock is held throughout the transaction. The fast
1965 * path can get away without acquiring the memcg->move_lock
1966 * because page moving starts with an RCU grace period.
1968 * The RCU lock also protects the memcg from being freed when
1969 * the page state that is going to change is the only thing
1970 * preventing the page itself from being freed. E.g. writeback
1971 * doesn't hold a page reference and relies on PG_writeback to
1972 * keep off truncation, migration and so forth.
1976 if (mem_cgroup_disabled())
1979 memcg = page->mem_cgroup;
1980 if (unlikely(!memcg))
1983 if (atomic_read(&memcg->moving_account) <= 0)
1986 spin_lock_irqsave(&memcg->move_lock, flags);
1987 if (memcg != page->mem_cgroup) {
1988 spin_unlock_irqrestore(&memcg->move_lock, flags);
1993 * When charge migration first begins, we can have locked and
1994 * unlocked page stat updates happening concurrently. Track
1995 * the task who has the lock for unlock_page_memcg().
1997 memcg->move_lock_task = current;
1998 memcg->move_lock_flags = flags;
2002 EXPORT_SYMBOL(lock_page_memcg);
2005 * __unlock_page_memcg - unlock and unpin a memcg
2008 * Unlock and unpin a memcg returned by lock_page_memcg().
2010 void __unlock_page_memcg(struct mem_cgroup *memcg)
2012 if (memcg && memcg->move_lock_task == current) {
2013 unsigned long flags = memcg->move_lock_flags;
2015 memcg->move_lock_task = NULL;
2016 memcg->move_lock_flags = 0;
2018 spin_unlock_irqrestore(&memcg->move_lock, flags);
2025 * unlock_page_memcg - unlock a page->mem_cgroup binding
2028 void unlock_page_memcg(struct page *page)
2030 __unlock_page_memcg(page->mem_cgroup);
2032 EXPORT_SYMBOL(unlock_page_memcg);
2034 struct memcg_stock_pcp {
2035 struct mem_cgroup *cached; /* this never be root cgroup */
2036 unsigned int nr_pages;
2037 struct work_struct work;
2038 unsigned long flags;
2039 #define FLUSHING_CACHED_CHARGE 0
2041 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2042 static DEFINE_MUTEX(percpu_charge_mutex);
2045 * consume_stock: Try to consume stocked charge on this cpu.
2046 * @memcg: memcg to consume from.
2047 * @nr_pages: how many pages to charge.
2049 * The charges will only happen if @memcg matches the current cpu's memcg
2050 * stock, and at least @nr_pages are available in that stock. Failure to
2051 * service an allocation will refill the stock.
2053 * returns true if successful, false otherwise.
2055 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2057 struct memcg_stock_pcp *stock;
2058 unsigned long flags;
2061 if (nr_pages > MEMCG_CHARGE_BATCH)
2064 local_irq_save(flags);
2066 stock = this_cpu_ptr(&memcg_stock);
2067 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2068 stock->nr_pages -= nr_pages;
2072 local_irq_restore(flags);
2078 * Returns stocks cached in percpu and reset cached information.
2080 static void drain_stock(struct memcg_stock_pcp *stock)
2082 struct mem_cgroup *old = stock->cached;
2084 if (stock->nr_pages) {
2085 page_counter_uncharge(&old->memory, stock->nr_pages);
2086 if (do_memsw_account())
2087 page_counter_uncharge(&old->memsw, stock->nr_pages);
2088 css_put_many(&old->css, stock->nr_pages);
2089 stock->nr_pages = 0;
2091 stock->cached = NULL;
2094 static void drain_local_stock(struct work_struct *dummy)
2096 struct memcg_stock_pcp *stock;
2097 unsigned long flags;
2100 * The only protection from memory hotplug vs. drain_stock races is
2101 * that we always operate on local CPU stock here with IRQ disabled
2103 local_irq_save(flags);
2105 stock = this_cpu_ptr(&memcg_stock);
2107 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2109 local_irq_restore(flags);
2113 * Cache charges(val) to local per_cpu area.
2114 * This will be consumed by consume_stock() function, later.
2116 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2118 struct memcg_stock_pcp *stock;
2119 unsigned long flags;
2121 local_irq_save(flags);
2123 stock = this_cpu_ptr(&memcg_stock);
2124 if (stock->cached != memcg) { /* reset if necessary */
2126 stock->cached = memcg;
2128 stock->nr_pages += nr_pages;
2130 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2133 local_irq_restore(flags);
2137 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2138 * of the hierarchy under it.
2140 static void drain_all_stock(struct mem_cgroup *root_memcg)
2144 /* If someone's already draining, avoid adding running more workers. */
2145 if (!mutex_trylock(&percpu_charge_mutex))
2148 * Notify other cpus that system-wide "drain" is running
2149 * We do not care about races with the cpu hotplug because cpu down
2150 * as well as workers from this path always operate on the local
2151 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2154 for_each_online_cpu(cpu) {
2155 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2156 struct mem_cgroup *memcg;
2160 memcg = stock->cached;
2161 if (memcg && stock->nr_pages &&
2162 mem_cgroup_is_descendant(memcg, root_memcg))
2167 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2169 drain_local_stock(&stock->work);
2171 schedule_work_on(cpu, &stock->work);
2175 mutex_unlock(&percpu_charge_mutex);
2178 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2180 struct memcg_stock_pcp *stock;
2181 struct mem_cgroup *memcg, *mi;
2183 stock = &per_cpu(memcg_stock, cpu);
2186 for_each_mem_cgroup(memcg) {
2189 for (i = 0; i < MEMCG_NR_STAT; i++) {
2193 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2195 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2196 atomic_long_add(x, &memcg->vmstats[i]);
2198 if (i >= NR_VM_NODE_STAT_ITEMS)
2201 for_each_node(nid) {
2202 struct mem_cgroup_per_node *pn;
2204 pn = mem_cgroup_nodeinfo(memcg, nid);
2205 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2208 atomic_long_add(x, &pn->lruvec_stat[i]);
2209 } while ((pn = parent_nodeinfo(pn, nid)));
2213 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2216 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2218 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2219 atomic_long_add(x, &memcg->vmevents[i]);
2226 static void reclaim_high(struct mem_cgroup *memcg,
2227 unsigned int nr_pages,
2231 if (page_counter_read(&memcg->memory) <= memcg->high)
2233 memcg_memory_event(memcg, MEMCG_HIGH);
2234 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2235 } while ((memcg = parent_mem_cgroup(memcg)));
2238 static void high_work_func(struct work_struct *work)
2240 struct mem_cgroup *memcg;
2242 memcg = container_of(work, struct mem_cgroup, high_work);
2243 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2247 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2248 * enough to still cause a significant slowdown in most cases, while still
2249 * allowing diagnostics and tracing to proceed without becoming stuck.
2251 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2254 * When calculating the delay, we use these either side of the exponentiation to
2255 * maintain precision and scale to a reasonable number of jiffies (see the table
2258 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2259 * overage ratio to a delay.
2260 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down down the
2261 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2262 * to produce a reasonable delay curve.
2264 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2265 * reasonable delay curve compared to precision-adjusted overage, not
2266 * penalising heavily at first, but still making sure that growth beyond the
2267 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2268 * example, with a high of 100 megabytes:
2270 * +-------+------------------------+
2271 * | usage | time to allocate in ms |
2272 * +-------+------------------------+
2294 * +-------+------------------------+
2296 #define MEMCG_DELAY_PRECISION_SHIFT 20
2297 #define MEMCG_DELAY_SCALING_SHIFT 14
2300 * Scheduled by try_charge() to be executed from the userland return path
2301 * and reclaims memory over the high limit.
2303 void mem_cgroup_handle_over_high(void)
2305 unsigned long usage, high, clamped_high;
2306 unsigned long pflags;
2307 unsigned long penalty_jiffies, overage;
2308 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2309 struct mem_cgroup *memcg;
2311 if (likely(!nr_pages))
2314 memcg = get_mem_cgroup_from_mm(current->mm);
2315 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2316 current->memcg_nr_pages_over_high = 0;
2319 * memory.high is breached and reclaim is unable to keep up. Throttle
2320 * allocators proactively to slow down excessive growth.
2322 * We use overage compared to memory.high to calculate the number of
2323 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2324 * fairly lenient on small overages, and increasingly harsh when the
2325 * memcg in question makes it clear that it has no intention of stopping
2326 * its crazy behaviour, so we exponentially increase the delay based on
2330 usage = page_counter_read(&memcg->memory);
2331 high = READ_ONCE(memcg->high);
2337 * Prevent division by 0 in overage calculation by acting as if it was a
2338 * threshold of 1 page
2340 clamped_high = max(high, 1UL);
2342 overage = div64_u64((u64)(usage - high) << MEMCG_DELAY_PRECISION_SHIFT,
2345 penalty_jiffies = ((u64)overage * overage * HZ)
2346 >> (MEMCG_DELAY_PRECISION_SHIFT + MEMCG_DELAY_SCALING_SHIFT);
2349 * Factor in the task's own contribution to the overage, such that four
2350 * N-sized allocations are throttled approximately the same as one
2351 * 4N-sized allocation.
2353 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2354 * larger the current charge patch is than that.
2356 penalty_jiffies = penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2359 * Clamp the max delay per usermode return so as to still keep the
2360 * application moving forwards and also permit diagnostics, albeit
2363 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2366 * Don't sleep if the amount of jiffies this memcg owes us is so low
2367 * that it's not even worth doing, in an attempt to be nice to those who
2368 * go only a small amount over their memory.high value and maybe haven't
2369 * been aggressively reclaimed enough yet.
2371 if (penalty_jiffies <= HZ / 100)
2375 * If we exit early, we're guaranteed to die (since
2376 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2377 * need to account for any ill-begotten jiffies to pay them off later.
2379 psi_memstall_enter(&pflags);
2380 schedule_timeout_killable(penalty_jiffies);
2381 psi_memstall_leave(&pflags);
2384 css_put(&memcg->css);
2387 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2388 unsigned int nr_pages)
2390 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2391 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2392 struct mem_cgroup *mem_over_limit;
2393 struct page_counter *counter;
2394 unsigned long nr_reclaimed;
2395 bool may_swap = true;
2396 bool drained = false;
2397 enum oom_status oom_status;
2399 if (mem_cgroup_is_root(memcg))
2402 if (consume_stock(memcg, nr_pages))
2405 if (!do_memsw_account() ||
2406 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2407 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2409 if (do_memsw_account())
2410 page_counter_uncharge(&memcg->memsw, batch);
2411 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2413 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2417 if (batch > nr_pages) {
2423 * Memcg doesn't have a dedicated reserve for atomic
2424 * allocations. But like the global atomic pool, we need to
2425 * put the burden of reclaim on regular allocation requests
2426 * and let these go through as privileged allocations.
2428 if (gfp_mask & __GFP_ATOMIC)
2432 * Unlike in global OOM situations, memcg is not in a physical
2433 * memory shortage. Allow dying and OOM-killed tasks to
2434 * bypass the last charges so that they can exit quickly and
2435 * free their memory.
2437 if (unlikely(should_force_charge()))
2441 * Prevent unbounded recursion when reclaim operations need to
2442 * allocate memory. This might exceed the limits temporarily,
2443 * but we prefer facilitating memory reclaim and getting back
2444 * under the limit over triggering OOM kills in these cases.
2446 if (unlikely(current->flags & PF_MEMALLOC))
2449 if (unlikely(task_in_memcg_oom(current)))
2452 if (!gfpflags_allow_blocking(gfp_mask))
2455 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2457 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2458 gfp_mask, may_swap);
2460 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2464 drain_all_stock(mem_over_limit);
2469 if (gfp_mask & __GFP_NORETRY)
2472 * Even though the limit is exceeded at this point, reclaim
2473 * may have been able to free some pages. Retry the charge
2474 * before killing the task.
2476 * Only for regular pages, though: huge pages are rather
2477 * unlikely to succeed so close to the limit, and we fall back
2478 * to regular pages anyway in case of failure.
2480 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2483 * At task move, charge accounts can be doubly counted. So, it's
2484 * better to wait until the end of task_move if something is going on.
2486 if (mem_cgroup_wait_acct_move(mem_over_limit))
2492 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2495 if (gfp_mask & __GFP_NOFAIL)
2498 if (fatal_signal_pending(current))
2502 * keep retrying as long as the memcg oom killer is able to make
2503 * a forward progress or bypass the charge if the oom killer
2504 * couldn't make any progress.
2506 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2507 get_order(nr_pages * PAGE_SIZE));
2508 switch (oom_status) {
2510 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2518 if (!(gfp_mask & __GFP_NOFAIL))
2522 * The allocation either can't fail or will lead to more memory
2523 * being freed very soon. Allow memory usage go over the limit
2524 * temporarily by force charging it.
2526 page_counter_charge(&memcg->memory, nr_pages);
2527 if (do_memsw_account())
2528 page_counter_charge(&memcg->memsw, nr_pages);
2529 css_get_many(&memcg->css, nr_pages);
2534 css_get_many(&memcg->css, batch);
2535 if (batch > nr_pages)
2536 refill_stock(memcg, batch - nr_pages);
2539 * If the hierarchy is above the normal consumption range, schedule
2540 * reclaim on returning to userland. We can perform reclaim here
2541 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2542 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2543 * not recorded as it most likely matches current's and won't
2544 * change in the meantime. As high limit is checked again before
2545 * reclaim, the cost of mismatch is negligible.
2548 if (page_counter_read(&memcg->memory) > memcg->high) {
2549 /* Don't bother a random interrupted task */
2550 if (in_interrupt()) {
2551 schedule_work(&memcg->high_work);
2554 current->memcg_nr_pages_over_high += batch;
2555 set_notify_resume(current);
2558 } while ((memcg = parent_mem_cgroup(memcg)));
2563 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2565 if (mem_cgroup_is_root(memcg))
2568 page_counter_uncharge(&memcg->memory, nr_pages);
2569 if (do_memsw_account())
2570 page_counter_uncharge(&memcg->memsw, nr_pages);
2572 css_put_many(&memcg->css, nr_pages);
2575 static void lock_page_lru(struct page *page, int *isolated)
2577 pg_data_t *pgdat = page_pgdat(page);
2579 spin_lock_irq(&pgdat->lru_lock);
2580 if (PageLRU(page)) {
2581 struct lruvec *lruvec;
2583 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2585 del_page_from_lru_list(page, lruvec, page_lru(page));
2591 static void unlock_page_lru(struct page *page, int isolated)
2593 pg_data_t *pgdat = page_pgdat(page);
2596 struct lruvec *lruvec;
2598 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2599 VM_BUG_ON_PAGE(PageLRU(page), page);
2601 add_page_to_lru_list(page, lruvec, page_lru(page));
2603 spin_unlock_irq(&pgdat->lru_lock);
2606 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2611 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2614 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2615 * may already be on some other mem_cgroup's LRU. Take care of it.
2618 lock_page_lru(page, &isolated);
2621 * Nobody should be changing or seriously looking at
2622 * page->mem_cgroup at this point:
2624 * - the page is uncharged
2626 * - the page is off-LRU
2628 * - an anonymous fault has exclusive page access, except for
2629 * a locked page table
2631 * - a page cache insertion, a swapin fault, or a migration
2632 * have the page locked
2634 page->mem_cgroup = memcg;
2637 unlock_page_lru(page, isolated);
2640 #ifdef CONFIG_MEMCG_KMEM
2641 static int memcg_alloc_cache_id(void)
2646 id = ida_simple_get(&memcg_cache_ida,
2647 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2651 if (id < memcg_nr_cache_ids)
2655 * There's no space for the new id in memcg_caches arrays,
2656 * so we have to grow them.
2658 down_write(&memcg_cache_ids_sem);
2660 size = 2 * (id + 1);
2661 if (size < MEMCG_CACHES_MIN_SIZE)
2662 size = MEMCG_CACHES_MIN_SIZE;
2663 else if (size > MEMCG_CACHES_MAX_SIZE)
2664 size = MEMCG_CACHES_MAX_SIZE;
2666 err = memcg_update_all_caches(size);
2668 err = memcg_update_all_list_lrus(size);
2670 memcg_nr_cache_ids = size;
2672 up_write(&memcg_cache_ids_sem);
2675 ida_simple_remove(&memcg_cache_ida, id);
2681 static void memcg_free_cache_id(int id)
2683 ida_simple_remove(&memcg_cache_ida, id);
2686 struct memcg_kmem_cache_create_work {
2687 struct mem_cgroup *memcg;
2688 struct kmem_cache *cachep;
2689 struct work_struct work;
2692 static void memcg_kmem_cache_create_func(struct work_struct *w)
2694 struct memcg_kmem_cache_create_work *cw =
2695 container_of(w, struct memcg_kmem_cache_create_work, work);
2696 struct mem_cgroup *memcg = cw->memcg;
2697 struct kmem_cache *cachep = cw->cachep;
2699 memcg_create_kmem_cache(memcg, cachep);
2701 css_put(&memcg->css);
2706 * Enqueue the creation of a per-memcg kmem_cache.
2708 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2709 struct kmem_cache *cachep)
2711 struct memcg_kmem_cache_create_work *cw;
2713 if (!css_tryget_online(&memcg->css))
2716 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2721 cw->cachep = cachep;
2722 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2724 queue_work(memcg_kmem_cache_wq, &cw->work);
2727 static inline bool memcg_kmem_bypass(void)
2729 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2735 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2736 * @cachep: the original global kmem cache
2738 * Return the kmem_cache we're supposed to use for a slab allocation.
2739 * We try to use the current memcg's version of the cache.
2741 * If the cache does not exist yet, if we are the first user of it, we
2742 * create it asynchronously in a workqueue and let the current allocation
2743 * go through with the original cache.
2745 * This function takes a reference to the cache it returns to assure it
2746 * won't get destroyed while we are working with it. Once the caller is
2747 * done with it, memcg_kmem_put_cache() must be called to release the
2750 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2752 struct mem_cgroup *memcg;
2753 struct kmem_cache *memcg_cachep;
2754 struct memcg_cache_array *arr;
2757 VM_BUG_ON(!is_root_cache(cachep));
2759 if (memcg_kmem_bypass())
2764 if (unlikely(current->active_memcg))
2765 memcg = current->active_memcg;
2767 memcg = mem_cgroup_from_task(current);
2769 if (!memcg || memcg == root_mem_cgroup)
2772 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2776 arr = rcu_dereference(cachep->memcg_params.memcg_caches);
2779 * Make sure we will access the up-to-date value. The code updating
2780 * memcg_caches issues a write barrier to match the data dependency
2781 * barrier inside READ_ONCE() (see memcg_create_kmem_cache()).
2783 memcg_cachep = READ_ONCE(arr->entries[kmemcg_id]);
2786 * If we are in a safe context (can wait, and not in interrupt
2787 * context), we could be be predictable and return right away.
2788 * This would guarantee that the allocation being performed
2789 * already belongs in the new cache.
2791 * However, there are some clashes that can arrive from locking.
2792 * For instance, because we acquire the slab_mutex while doing
2793 * memcg_create_kmem_cache, this means no further allocation
2794 * could happen with the slab_mutex held. So it's better to
2797 * If the memcg is dying or memcg_cache is about to be released,
2798 * don't bother creating new kmem_caches. Because memcg_cachep
2799 * is ZEROed as the fist step of kmem offlining, we don't need
2800 * percpu_ref_tryget_live() here. css_tryget_online() check in
2801 * memcg_schedule_kmem_cache_create() will prevent us from
2802 * creation of a new kmem_cache.
2804 if (unlikely(!memcg_cachep))
2805 memcg_schedule_kmem_cache_create(memcg, cachep);
2806 else if (percpu_ref_tryget(&memcg_cachep->memcg_params.refcnt))
2807 cachep = memcg_cachep;
2814 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2815 * @cachep: the cache returned by memcg_kmem_get_cache
2817 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2819 if (!is_root_cache(cachep))
2820 percpu_ref_put(&cachep->memcg_params.refcnt);
2824 * __memcg_kmem_charge_memcg: charge a kmem page
2825 * @page: page to charge
2826 * @gfp: reclaim mode
2827 * @order: allocation order
2828 * @memcg: memory cgroup to charge
2830 * Returns 0 on success, an error code on failure.
2832 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2833 struct mem_cgroup *memcg)
2835 unsigned int nr_pages = 1 << order;
2836 struct page_counter *counter;
2839 ret = try_charge(memcg, gfp, nr_pages);
2843 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2844 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2847 * Enforce __GFP_NOFAIL allocation because callers are not
2848 * prepared to see failures and likely do not have any failure
2851 if (gfp & __GFP_NOFAIL) {
2852 page_counter_charge(&memcg->kmem, nr_pages);
2855 cancel_charge(memcg, nr_pages);
2862 * __memcg_kmem_charge: charge a kmem page to the current memory cgroup
2863 * @page: page to charge
2864 * @gfp: reclaim mode
2865 * @order: allocation order
2867 * Returns 0 on success, an error code on failure.
2869 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2871 struct mem_cgroup *memcg;
2874 if (memcg_kmem_bypass())
2877 memcg = get_mem_cgroup_from_current();
2878 if (!mem_cgroup_is_root(memcg)) {
2879 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2881 page->mem_cgroup = memcg;
2882 __SetPageKmemcg(page);
2885 css_put(&memcg->css);
2890 * __memcg_kmem_uncharge_memcg: uncharge a kmem page
2891 * @memcg: memcg to uncharge
2892 * @nr_pages: number of pages to uncharge
2894 void __memcg_kmem_uncharge_memcg(struct mem_cgroup *memcg,
2895 unsigned int nr_pages)
2897 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2898 page_counter_uncharge(&memcg->kmem, nr_pages);
2900 page_counter_uncharge(&memcg->memory, nr_pages);
2901 if (do_memsw_account())
2902 page_counter_uncharge(&memcg->memsw, nr_pages);
2905 * __memcg_kmem_uncharge: uncharge a kmem page
2906 * @page: page to uncharge
2907 * @order: allocation order
2909 void __memcg_kmem_uncharge(struct page *page, int order)
2911 struct mem_cgroup *memcg = page->mem_cgroup;
2912 unsigned int nr_pages = 1 << order;
2917 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2918 __memcg_kmem_uncharge_memcg(memcg, nr_pages);
2919 page->mem_cgroup = NULL;
2921 /* slab pages do not have PageKmemcg flag set */
2922 if (PageKmemcg(page))
2923 __ClearPageKmemcg(page);
2925 css_put_many(&memcg->css, nr_pages);
2927 #endif /* CONFIG_MEMCG_KMEM */
2929 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2932 * Because tail pages are not marked as "used", set it. We're under
2933 * pgdat->lru_lock and migration entries setup in all page mappings.
2935 void mem_cgroup_split_huge_fixup(struct page *head)
2939 if (mem_cgroup_disabled())
2942 for (i = 1; i < HPAGE_PMD_NR; i++)
2943 head[i].mem_cgroup = head->mem_cgroup;
2945 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
2947 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2949 #ifdef CONFIG_MEMCG_SWAP
2951 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2952 * @entry: swap entry to be moved
2953 * @from: mem_cgroup which the entry is moved from
2954 * @to: mem_cgroup which the entry is moved to
2956 * It succeeds only when the swap_cgroup's record for this entry is the same
2957 * as the mem_cgroup's id of @from.
2959 * Returns 0 on success, -EINVAL on failure.
2961 * The caller must have charged to @to, IOW, called page_counter_charge() about
2962 * both res and memsw, and called css_get().
2964 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2965 struct mem_cgroup *from, struct mem_cgroup *to)
2967 unsigned short old_id, new_id;
2969 old_id = mem_cgroup_id(from);
2970 new_id = mem_cgroup_id(to);
2972 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2973 mod_memcg_state(from, MEMCG_SWAP, -1);
2974 mod_memcg_state(to, MEMCG_SWAP, 1);
2980 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2981 struct mem_cgroup *from, struct mem_cgroup *to)
2987 static DEFINE_MUTEX(memcg_max_mutex);
2989 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
2990 unsigned long max, bool memsw)
2992 bool enlarge = false;
2993 bool drained = false;
2995 bool limits_invariant;
2996 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
2999 if (signal_pending(current)) {
3004 mutex_lock(&memcg_max_mutex);
3006 * Make sure that the new limit (memsw or memory limit) doesn't
3007 * break our basic invariant rule memory.max <= memsw.max.
3009 limits_invariant = memsw ? max >= memcg->memory.max :
3010 max <= memcg->memsw.max;
3011 if (!limits_invariant) {
3012 mutex_unlock(&memcg_max_mutex);
3016 if (max > counter->max)
3018 ret = page_counter_set_max(counter, max);
3019 mutex_unlock(&memcg_max_mutex);
3025 drain_all_stock(memcg);
3030 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3031 GFP_KERNEL, !memsw)) {
3037 if (!ret && enlarge)
3038 memcg_oom_recover(memcg);
3043 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3045 unsigned long *total_scanned)
3047 unsigned long nr_reclaimed = 0;
3048 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3049 unsigned long reclaimed;
3051 struct mem_cgroup_tree_per_node *mctz;
3052 unsigned long excess;
3053 unsigned long nr_scanned;
3058 mctz = soft_limit_tree_node(pgdat->node_id);
3061 * Do not even bother to check the largest node if the root
3062 * is empty. Do it lockless to prevent lock bouncing. Races
3063 * are acceptable as soft limit is best effort anyway.
3065 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3069 * This loop can run a while, specially if mem_cgroup's continuously
3070 * keep exceeding their soft limit and putting the system under
3077 mz = mem_cgroup_largest_soft_limit_node(mctz);
3082 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3083 gfp_mask, &nr_scanned);
3084 nr_reclaimed += reclaimed;
3085 *total_scanned += nr_scanned;
3086 spin_lock_irq(&mctz->lock);
3087 __mem_cgroup_remove_exceeded(mz, mctz);
3090 * If we failed to reclaim anything from this memory cgroup
3091 * it is time to move on to the next cgroup
3095 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3097 excess = soft_limit_excess(mz->memcg);
3099 * One school of thought says that we should not add
3100 * back the node to the tree if reclaim returns 0.
3101 * But our reclaim could return 0, simply because due
3102 * to priority we are exposing a smaller subset of
3103 * memory to reclaim from. Consider this as a longer
3106 /* If excess == 0, no tree ops */
3107 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3108 spin_unlock_irq(&mctz->lock);
3109 css_put(&mz->memcg->css);
3112 * Could not reclaim anything and there are no more
3113 * mem cgroups to try or we seem to be looping without
3114 * reclaiming anything.
3116 if (!nr_reclaimed &&
3118 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3120 } while (!nr_reclaimed);
3122 css_put(&next_mz->memcg->css);
3123 return nr_reclaimed;
3127 * Test whether @memcg has children, dead or alive. Note that this
3128 * function doesn't care whether @memcg has use_hierarchy enabled and
3129 * returns %true if there are child csses according to the cgroup
3130 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3132 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3137 ret = css_next_child(NULL, &memcg->css);
3143 * Reclaims as many pages from the given memcg as possible.
3145 * Caller is responsible for holding css reference for memcg.
3147 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3149 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3151 /* we call try-to-free pages for make this cgroup empty */
3152 lru_add_drain_all();
3154 drain_all_stock(memcg);
3156 /* try to free all pages in this cgroup */
3157 while (nr_retries && page_counter_read(&memcg->memory)) {
3160 if (signal_pending(current))
3163 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3167 /* maybe some writeback is necessary */
3168 congestion_wait(BLK_RW_ASYNC, HZ/10);
3176 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3177 char *buf, size_t nbytes,
3180 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3182 if (mem_cgroup_is_root(memcg))
3184 return mem_cgroup_force_empty(memcg) ?: nbytes;
3187 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3190 return mem_cgroup_from_css(css)->use_hierarchy;
3193 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3194 struct cftype *cft, u64 val)
3197 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3198 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3200 if (memcg->use_hierarchy == val)
3204 * If parent's use_hierarchy is set, we can't make any modifications
3205 * in the child subtrees. If it is unset, then the change can
3206 * occur, provided the current cgroup has no children.
3208 * For the root cgroup, parent_mem is NULL, we allow value to be
3209 * set if there are no children.
3211 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3212 (val == 1 || val == 0)) {
3213 if (!memcg_has_children(memcg))
3214 memcg->use_hierarchy = val;
3223 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3227 if (mem_cgroup_is_root(memcg)) {
3228 val = memcg_page_state(memcg, MEMCG_CACHE) +
3229 memcg_page_state(memcg, MEMCG_RSS);
3231 val += memcg_page_state(memcg, MEMCG_SWAP);
3234 val = page_counter_read(&memcg->memory);
3236 val = page_counter_read(&memcg->memsw);
3249 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3252 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3253 struct page_counter *counter;
3255 switch (MEMFILE_TYPE(cft->private)) {
3257 counter = &memcg->memory;
3260 counter = &memcg->memsw;
3263 counter = &memcg->kmem;
3266 counter = &memcg->tcpmem;
3272 switch (MEMFILE_ATTR(cft->private)) {
3274 if (counter == &memcg->memory)
3275 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3276 if (counter == &memcg->memsw)
3277 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3278 return (u64)page_counter_read(counter) * PAGE_SIZE;
3280 return (u64)counter->max * PAGE_SIZE;
3282 return (u64)counter->watermark * PAGE_SIZE;
3284 return counter->failcnt;
3285 case RES_SOFT_LIMIT:
3286 return (u64)memcg->soft_limit * PAGE_SIZE;
3292 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3294 unsigned long stat[MEMCG_NR_STAT] = {0};
3295 struct mem_cgroup *mi;
3298 for_each_online_cpu(cpu)
3299 for (i = 0; i < MEMCG_NR_STAT; i++)
3300 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3302 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3303 for (i = 0; i < MEMCG_NR_STAT; i++)
3304 atomic_long_add(stat[i], &mi->vmstats[i]);
3306 for_each_node(node) {
3307 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3308 struct mem_cgroup_per_node *pi;
3310 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3313 for_each_online_cpu(cpu)
3314 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3316 pn->lruvec_stat_cpu->count[i], cpu);
3318 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3319 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3320 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3324 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3326 unsigned long events[NR_VM_EVENT_ITEMS];
3327 struct mem_cgroup *mi;
3330 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3333 for_each_online_cpu(cpu)
3334 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3335 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3338 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3339 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3340 atomic_long_add(events[i], &mi->vmevents[i]);
3343 #ifdef CONFIG_MEMCG_KMEM
3344 static int memcg_online_kmem(struct mem_cgroup *memcg)
3348 if (cgroup_memory_nokmem)
3351 BUG_ON(memcg->kmemcg_id >= 0);
3352 BUG_ON(memcg->kmem_state);
3354 memcg_id = memcg_alloc_cache_id();
3358 static_branch_inc(&memcg_kmem_enabled_key);
3360 * A memory cgroup is considered kmem-online as soon as it gets
3361 * kmemcg_id. Setting the id after enabling static branching will
3362 * guarantee no one starts accounting before all call sites are
3365 memcg->kmemcg_id = memcg_id;
3366 memcg->kmem_state = KMEM_ONLINE;
3367 INIT_LIST_HEAD(&memcg->kmem_caches);
3372 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3374 struct cgroup_subsys_state *css;
3375 struct mem_cgroup *parent, *child;
3378 if (memcg->kmem_state != KMEM_ONLINE)
3381 * Clear the online state before clearing memcg_caches array
3382 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3383 * guarantees that no cache will be created for this cgroup
3384 * after we are done (see memcg_create_kmem_cache()).
3386 memcg->kmem_state = KMEM_ALLOCATED;
3388 parent = parent_mem_cgroup(memcg);
3390 parent = root_mem_cgroup;
3393 * Deactivate and reparent kmem_caches.
3395 memcg_deactivate_kmem_caches(memcg, parent);
3397 kmemcg_id = memcg->kmemcg_id;
3398 BUG_ON(kmemcg_id < 0);
3401 * Change kmemcg_id of this cgroup and all its descendants to the
3402 * parent's id, and then move all entries from this cgroup's list_lrus
3403 * to ones of the parent. After we have finished, all list_lrus
3404 * corresponding to this cgroup are guaranteed to remain empty. The
3405 * ordering is imposed by list_lru_node->lock taken by
3406 * memcg_drain_all_list_lrus().
3408 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3409 css_for_each_descendant_pre(css, &memcg->css) {
3410 child = mem_cgroup_from_css(css);
3411 BUG_ON(child->kmemcg_id != kmemcg_id);
3412 child->kmemcg_id = parent->kmemcg_id;
3413 if (!memcg->use_hierarchy)
3418 memcg_drain_all_list_lrus(kmemcg_id, parent);
3420 memcg_free_cache_id(kmemcg_id);
3423 static void memcg_free_kmem(struct mem_cgroup *memcg)
3425 /* css_alloc() failed, offlining didn't happen */
3426 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3427 memcg_offline_kmem(memcg);
3429 if (memcg->kmem_state == KMEM_ALLOCATED) {
3430 WARN_ON(!list_empty(&memcg->kmem_caches));
3431 static_branch_dec(&memcg_kmem_enabled_key);
3435 static int memcg_online_kmem(struct mem_cgroup *memcg)
3439 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3442 static void memcg_free_kmem(struct mem_cgroup *memcg)
3445 #endif /* CONFIG_MEMCG_KMEM */
3447 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3452 mutex_lock(&memcg_max_mutex);
3453 ret = page_counter_set_max(&memcg->kmem, max);
3454 mutex_unlock(&memcg_max_mutex);
3458 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3462 mutex_lock(&memcg_max_mutex);
3464 ret = page_counter_set_max(&memcg->tcpmem, max);
3468 if (!memcg->tcpmem_active) {
3470 * The active flag needs to be written after the static_key
3471 * update. This is what guarantees that the socket activation
3472 * function is the last one to run. See mem_cgroup_sk_alloc()
3473 * for details, and note that we don't mark any socket as
3474 * belonging to this memcg until that flag is up.
3476 * We need to do this, because static_keys will span multiple
3477 * sites, but we can't control their order. If we mark a socket
3478 * as accounted, but the accounting functions are not patched in
3479 * yet, we'll lose accounting.
3481 * We never race with the readers in mem_cgroup_sk_alloc(),
3482 * because when this value change, the code to process it is not
3485 static_branch_inc(&memcg_sockets_enabled_key);
3486 memcg->tcpmem_active = true;
3489 mutex_unlock(&memcg_max_mutex);
3494 * The user of this function is...
3497 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3498 char *buf, size_t nbytes, loff_t off)
3500 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3501 unsigned long nr_pages;
3504 buf = strstrip(buf);
3505 ret = page_counter_memparse(buf, "-1", &nr_pages);
3509 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3511 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3515 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3517 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3520 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3523 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3524 "Please report your usecase to linux-mm@kvack.org if you "
3525 "depend on this functionality.\n");
3526 ret = memcg_update_kmem_max(memcg, nr_pages);
3529 ret = memcg_update_tcp_max(memcg, nr_pages);
3533 case RES_SOFT_LIMIT:
3534 memcg->soft_limit = nr_pages;
3538 return ret ?: nbytes;
3541 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3542 size_t nbytes, loff_t off)
3544 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3545 struct page_counter *counter;
3547 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3549 counter = &memcg->memory;
3552 counter = &memcg->memsw;
3555 counter = &memcg->kmem;
3558 counter = &memcg->tcpmem;
3564 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3566 page_counter_reset_watermark(counter);
3569 counter->failcnt = 0;
3578 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3581 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3585 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3586 struct cftype *cft, u64 val)
3588 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3590 if (val & ~MOVE_MASK)
3594 * No kind of locking is needed in here, because ->can_attach() will
3595 * check this value once in the beginning of the process, and then carry
3596 * on with stale data. This means that changes to this value will only
3597 * affect task migrations starting after the change.
3599 memcg->move_charge_at_immigrate = val;
3603 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3604 struct cftype *cft, u64 val)
3612 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3613 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3614 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3616 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3617 int nid, unsigned int lru_mask)
3619 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3620 unsigned long nr = 0;
3623 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3626 if (!(BIT(lru) & lru_mask))
3628 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3633 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3634 unsigned int lru_mask)
3636 unsigned long nr = 0;
3640 if (!(BIT(lru) & lru_mask))
3642 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3647 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3651 unsigned int lru_mask;
3654 static const struct numa_stat stats[] = {
3655 { "total", LRU_ALL },
3656 { "file", LRU_ALL_FILE },
3657 { "anon", LRU_ALL_ANON },
3658 { "unevictable", BIT(LRU_UNEVICTABLE) },
3660 const struct numa_stat *stat;
3663 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3665 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3666 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3667 seq_printf(m, "%s=%lu", stat->name, nr);
3668 for_each_node_state(nid, N_MEMORY) {
3669 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3671 seq_printf(m, " N%d=%lu", nid, nr);
3676 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3677 struct mem_cgroup *iter;
3680 for_each_mem_cgroup_tree(iter, memcg)
3681 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3682 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3683 for_each_node_state(nid, N_MEMORY) {
3685 for_each_mem_cgroup_tree(iter, memcg)
3686 nr += mem_cgroup_node_nr_lru_pages(
3687 iter, nid, stat->lru_mask);
3688 seq_printf(m, " N%d=%lu", nid, nr);
3695 #endif /* CONFIG_NUMA */
3697 static const unsigned int memcg1_stats[] = {
3708 static const char *const memcg1_stat_names[] = {
3719 /* Universal VM events cgroup1 shows, original sort order */
3720 static const unsigned int memcg1_events[] = {
3727 static int memcg_stat_show(struct seq_file *m, void *v)
3729 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3730 unsigned long memory, memsw;
3731 struct mem_cgroup *mi;
3734 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3736 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3737 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3739 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3740 memcg_page_state_local(memcg, memcg1_stats[i]) *
3744 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3745 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
3746 memcg_events_local(memcg, memcg1_events[i]));
3748 for (i = 0; i < NR_LRU_LISTS; i++)
3749 seq_printf(m, "%s %lu\n", lru_list_name(i),
3750 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3753 /* Hierarchical information */
3754 memory = memsw = PAGE_COUNTER_MAX;
3755 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3756 memory = min(memory, mi->memory.max);
3757 memsw = min(memsw, mi->memsw.max);
3759 seq_printf(m, "hierarchical_memory_limit %llu\n",
3760 (u64)memory * PAGE_SIZE);
3761 if (do_memsw_account())
3762 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3763 (u64)memsw * PAGE_SIZE);
3765 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3766 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3768 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3769 (u64)memcg_page_state(memcg, memcg1_stats[i]) *
3773 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3774 seq_printf(m, "total_%s %llu\n",
3775 vm_event_name(memcg1_events[i]),
3776 (u64)memcg_events(memcg, memcg1_events[i]));
3778 for (i = 0; i < NR_LRU_LISTS; i++)
3779 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
3780 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
3783 #ifdef CONFIG_DEBUG_VM
3786 struct mem_cgroup_per_node *mz;
3787 struct zone_reclaim_stat *rstat;
3788 unsigned long recent_rotated[2] = {0, 0};
3789 unsigned long recent_scanned[2] = {0, 0};
3791 for_each_online_pgdat(pgdat) {
3792 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3793 rstat = &mz->lruvec.reclaim_stat;
3795 recent_rotated[0] += rstat->recent_rotated[0];
3796 recent_rotated[1] += rstat->recent_rotated[1];
3797 recent_scanned[0] += rstat->recent_scanned[0];
3798 recent_scanned[1] += rstat->recent_scanned[1];
3800 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3801 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3802 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3803 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3810 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3813 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3815 return mem_cgroup_swappiness(memcg);
3818 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3819 struct cftype *cft, u64 val)
3821 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3827 memcg->swappiness = val;
3829 vm_swappiness = val;
3834 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3836 struct mem_cgroup_threshold_ary *t;
3837 unsigned long usage;
3842 t = rcu_dereference(memcg->thresholds.primary);
3844 t = rcu_dereference(memcg->memsw_thresholds.primary);
3849 usage = mem_cgroup_usage(memcg, swap);
3852 * current_threshold points to threshold just below or equal to usage.
3853 * If it's not true, a threshold was crossed after last
3854 * call of __mem_cgroup_threshold().
3856 i = t->current_threshold;
3859 * Iterate backward over array of thresholds starting from
3860 * current_threshold and check if a threshold is crossed.
3861 * If none of thresholds below usage is crossed, we read
3862 * only one element of the array here.
3864 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3865 eventfd_signal(t->entries[i].eventfd, 1);
3867 /* i = current_threshold + 1 */
3871 * Iterate forward over array of thresholds starting from
3872 * current_threshold+1 and check if a threshold is crossed.
3873 * If none of thresholds above usage is crossed, we read
3874 * only one element of the array here.
3876 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3877 eventfd_signal(t->entries[i].eventfd, 1);
3879 /* Update current_threshold */
3880 t->current_threshold = i - 1;
3885 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3888 __mem_cgroup_threshold(memcg, false);
3889 if (do_memsw_account())
3890 __mem_cgroup_threshold(memcg, true);
3892 memcg = parent_mem_cgroup(memcg);
3896 static int compare_thresholds(const void *a, const void *b)
3898 const struct mem_cgroup_threshold *_a = a;
3899 const struct mem_cgroup_threshold *_b = b;
3901 if (_a->threshold > _b->threshold)
3904 if (_a->threshold < _b->threshold)
3910 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3912 struct mem_cgroup_eventfd_list *ev;
3914 spin_lock(&memcg_oom_lock);
3916 list_for_each_entry(ev, &memcg->oom_notify, list)
3917 eventfd_signal(ev->eventfd, 1);
3919 spin_unlock(&memcg_oom_lock);
3923 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3925 struct mem_cgroup *iter;
3927 for_each_mem_cgroup_tree(iter, memcg)
3928 mem_cgroup_oom_notify_cb(iter);
3931 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3932 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3934 struct mem_cgroup_thresholds *thresholds;
3935 struct mem_cgroup_threshold_ary *new;
3936 unsigned long threshold;
3937 unsigned long usage;
3940 ret = page_counter_memparse(args, "-1", &threshold);
3944 mutex_lock(&memcg->thresholds_lock);
3947 thresholds = &memcg->thresholds;
3948 usage = mem_cgroup_usage(memcg, false);
3949 } else if (type == _MEMSWAP) {
3950 thresholds = &memcg->memsw_thresholds;
3951 usage = mem_cgroup_usage(memcg, true);
3955 /* Check if a threshold crossed before adding a new one */
3956 if (thresholds->primary)
3957 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3959 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3961 /* Allocate memory for new array of thresholds */
3962 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
3969 /* Copy thresholds (if any) to new array */
3970 if (thresholds->primary) {
3971 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3972 sizeof(struct mem_cgroup_threshold));
3975 /* Add new threshold */
3976 new->entries[size - 1].eventfd = eventfd;
3977 new->entries[size - 1].threshold = threshold;
3979 /* Sort thresholds. Registering of new threshold isn't time-critical */
3980 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3981 compare_thresholds, NULL);
3983 /* Find current threshold */
3984 new->current_threshold = -1;
3985 for (i = 0; i < size; i++) {
3986 if (new->entries[i].threshold <= usage) {
3988 * new->current_threshold will not be used until
3989 * rcu_assign_pointer(), so it's safe to increment
3992 ++new->current_threshold;
3997 /* Free old spare buffer and save old primary buffer as spare */
3998 kfree(thresholds->spare);
3999 thresholds->spare = thresholds->primary;
4001 rcu_assign_pointer(thresholds->primary, new);
4003 /* To be sure that nobody uses thresholds */
4007 mutex_unlock(&memcg->thresholds_lock);
4012 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4013 struct eventfd_ctx *eventfd, const char *args)
4015 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4018 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4019 struct eventfd_ctx *eventfd, const char *args)
4021 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4024 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4025 struct eventfd_ctx *eventfd, enum res_type type)
4027 struct mem_cgroup_thresholds *thresholds;
4028 struct mem_cgroup_threshold_ary *new;
4029 unsigned long usage;
4030 int i, j, size, entries;
4032 mutex_lock(&memcg->thresholds_lock);
4035 thresholds = &memcg->thresholds;
4036 usage = mem_cgroup_usage(memcg, false);
4037 } else if (type == _MEMSWAP) {
4038 thresholds = &memcg->memsw_thresholds;
4039 usage = mem_cgroup_usage(memcg, true);
4043 if (!thresholds->primary)
4046 /* Check if a threshold crossed before removing */
4047 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4049 /* Calculate new number of threshold */
4051 for (i = 0; i < thresholds->primary->size; i++) {
4052 if (thresholds->primary->entries[i].eventfd != eventfd)
4058 new = thresholds->spare;
4060 /* If no items related to eventfd have been cleared, nothing to do */
4064 /* Set thresholds array to NULL if we don't have thresholds */
4073 /* Copy thresholds and find current threshold */
4074 new->current_threshold = -1;
4075 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4076 if (thresholds->primary->entries[i].eventfd == eventfd)
4079 new->entries[j] = thresholds->primary->entries[i];
4080 if (new->entries[j].threshold <= usage) {
4082 * new->current_threshold will not be used
4083 * until rcu_assign_pointer(), so it's safe to increment
4086 ++new->current_threshold;
4092 /* Swap primary and spare array */
4093 thresholds->spare = thresholds->primary;
4095 rcu_assign_pointer(thresholds->primary, new);
4097 /* To be sure that nobody uses thresholds */
4100 /* If all events are unregistered, free the spare array */
4102 kfree(thresholds->spare);
4103 thresholds->spare = NULL;
4106 mutex_unlock(&memcg->thresholds_lock);
4109 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4110 struct eventfd_ctx *eventfd)
4112 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4115 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4116 struct eventfd_ctx *eventfd)
4118 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4121 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4122 struct eventfd_ctx *eventfd, const char *args)
4124 struct mem_cgroup_eventfd_list *event;
4126 event = kmalloc(sizeof(*event), GFP_KERNEL);
4130 spin_lock(&memcg_oom_lock);
4132 event->eventfd = eventfd;
4133 list_add(&event->list, &memcg->oom_notify);
4135 /* already in OOM ? */
4136 if (memcg->under_oom)
4137 eventfd_signal(eventfd, 1);
4138 spin_unlock(&memcg_oom_lock);
4143 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4144 struct eventfd_ctx *eventfd)
4146 struct mem_cgroup_eventfd_list *ev, *tmp;
4148 spin_lock(&memcg_oom_lock);
4150 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4151 if (ev->eventfd == eventfd) {
4152 list_del(&ev->list);
4157 spin_unlock(&memcg_oom_lock);
4160 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4162 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4164 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4165 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4166 seq_printf(sf, "oom_kill %lu\n",
4167 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4171 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4172 struct cftype *cft, u64 val)
4174 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4176 /* cannot set to root cgroup and only 0 and 1 are allowed */
4177 if (!css->parent || !((val == 0) || (val == 1)))
4180 memcg->oom_kill_disable = val;
4182 memcg_oom_recover(memcg);
4187 #ifdef CONFIG_CGROUP_WRITEBACK
4189 #include <trace/events/writeback.h>
4191 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4193 return wb_domain_init(&memcg->cgwb_domain, gfp);
4196 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4198 wb_domain_exit(&memcg->cgwb_domain);
4201 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4203 wb_domain_size_changed(&memcg->cgwb_domain);
4206 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4208 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4210 if (!memcg->css.parent)
4213 return &memcg->cgwb_domain;
4217 * idx can be of type enum memcg_stat_item or node_stat_item.
4218 * Keep in sync with memcg_exact_page().
4220 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4222 long x = atomic_long_read(&memcg->vmstats[idx]);
4225 for_each_online_cpu(cpu)
4226 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4233 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4234 * @wb: bdi_writeback in question
4235 * @pfilepages: out parameter for number of file pages
4236 * @pheadroom: out parameter for number of allocatable pages according to memcg
4237 * @pdirty: out parameter for number of dirty pages
4238 * @pwriteback: out parameter for number of pages under writeback
4240 * Determine the numbers of file, headroom, dirty, and writeback pages in
4241 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4242 * is a bit more involved.
4244 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4245 * headroom is calculated as the lowest headroom of itself and the
4246 * ancestors. Note that this doesn't consider the actual amount of
4247 * available memory in the system. The caller should further cap
4248 * *@pheadroom accordingly.
4250 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4251 unsigned long *pheadroom, unsigned long *pdirty,
4252 unsigned long *pwriteback)
4254 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4255 struct mem_cgroup *parent;
4257 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4259 /* this should eventually include NR_UNSTABLE_NFS */
4260 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4261 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4262 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4263 *pheadroom = PAGE_COUNTER_MAX;
4265 while ((parent = parent_mem_cgroup(memcg))) {
4266 unsigned long ceiling = min(memcg->memory.max, memcg->high);
4267 unsigned long used = page_counter_read(&memcg->memory);
4269 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4275 * Foreign dirty flushing
4277 * There's an inherent mismatch between memcg and writeback. The former
4278 * trackes ownership per-page while the latter per-inode. This was a
4279 * deliberate design decision because honoring per-page ownership in the
4280 * writeback path is complicated, may lead to higher CPU and IO overheads
4281 * and deemed unnecessary given that write-sharing an inode across
4282 * different cgroups isn't a common use-case.
4284 * Combined with inode majority-writer ownership switching, this works well
4285 * enough in most cases but there are some pathological cases. For
4286 * example, let's say there are two cgroups A and B which keep writing to
4287 * different but confined parts of the same inode. B owns the inode and
4288 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4289 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4290 * triggering background writeback. A will be slowed down without a way to
4291 * make writeback of the dirty pages happen.
4293 * Conditions like the above can lead to a cgroup getting repatedly and
4294 * severely throttled after making some progress after each
4295 * dirty_expire_interval while the underyling IO device is almost
4298 * Solving this problem completely requires matching the ownership tracking
4299 * granularities between memcg and writeback in either direction. However,
4300 * the more egregious behaviors can be avoided by simply remembering the
4301 * most recent foreign dirtying events and initiating remote flushes on
4302 * them when local writeback isn't enough to keep the memory clean enough.
4304 * The following two functions implement such mechanism. When a foreign
4305 * page - a page whose memcg and writeback ownerships don't match - is
4306 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4307 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4308 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4309 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4310 * foreign bdi_writebacks which haven't expired. Both the numbers of
4311 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4312 * limited to MEMCG_CGWB_FRN_CNT.
4314 * The mechanism only remembers IDs and doesn't hold any object references.
4315 * As being wrong occasionally doesn't matter, updates and accesses to the
4316 * records are lockless and racy.
4318 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4319 struct bdi_writeback *wb)
4321 struct mem_cgroup *memcg = page->mem_cgroup;
4322 struct memcg_cgwb_frn *frn;
4323 u64 now = get_jiffies_64();
4324 u64 oldest_at = now;
4328 trace_track_foreign_dirty(page, wb);
4331 * Pick the slot to use. If there is already a slot for @wb, keep
4332 * using it. If not replace the oldest one which isn't being
4335 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4336 frn = &memcg->cgwb_frn[i];
4337 if (frn->bdi_id == wb->bdi->id &&
4338 frn->memcg_id == wb->memcg_css->id)
4340 if (time_before64(frn->at, oldest_at) &&
4341 atomic_read(&frn->done.cnt) == 1) {
4343 oldest_at = frn->at;
4347 if (i < MEMCG_CGWB_FRN_CNT) {
4349 * Re-using an existing one. Update timestamp lazily to
4350 * avoid making the cacheline hot. We want them to be
4351 * reasonably up-to-date and significantly shorter than
4352 * dirty_expire_interval as that's what expires the record.
4353 * Use the shorter of 1s and dirty_expire_interval / 8.
4355 unsigned long update_intv =
4356 min_t(unsigned long, HZ,
4357 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4359 if (time_before64(frn->at, now - update_intv))
4361 } else if (oldest >= 0) {
4362 /* replace the oldest free one */
4363 frn = &memcg->cgwb_frn[oldest];
4364 frn->bdi_id = wb->bdi->id;
4365 frn->memcg_id = wb->memcg_css->id;
4370 /* issue foreign writeback flushes for recorded foreign dirtying events */
4371 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4373 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4374 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4375 u64 now = jiffies_64;
4378 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4379 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4382 * If the record is older than dirty_expire_interval,
4383 * writeback on it has already started. No need to kick it
4384 * off again. Also, don't start a new one if there's
4385 * already one in flight.
4387 if (time_after64(frn->at, now - intv) &&
4388 atomic_read(&frn->done.cnt) == 1) {
4390 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4391 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4392 WB_REASON_FOREIGN_FLUSH,
4398 #else /* CONFIG_CGROUP_WRITEBACK */
4400 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4405 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4409 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4413 #endif /* CONFIG_CGROUP_WRITEBACK */
4416 * DO NOT USE IN NEW FILES.
4418 * "cgroup.event_control" implementation.
4420 * This is way over-engineered. It tries to support fully configurable
4421 * events for each user. Such level of flexibility is completely
4422 * unnecessary especially in the light of the planned unified hierarchy.
4424 * Please deprecate this and replace with something simpler if at all
4429 * Unregister event and free resources.
4431 * Gets called from workqueue.
4433 static void memcg_event_remove(struct work_struct *work)
4435 struct mem_cgroup_event *event =
4436 container_of(work, struct mem_cgroup_event, remove);
4437 struct mem_cgroup *memcg = event->memcg;
4439 remove_wait_queue(event->wqh, &event->wait);
4441 event->unregister_event(memcg, event->eventfd);
4443 /* Notify userspace the event is going away. */
4444 eventfd_signal(event->eventfd, 1);
4446 eventfd_ctx_put(event->eventfd);
4448 css_put(&memcg->css);
4452 * Gets called on EPOLLHUP on eventfd when user closes it.
4454 * Called with wqh->lock held and interrupts disabled.
4456 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4457 int sync, void *key)
4459 struct mem_cgroup_event *event =
4460 container_of(wait, struct mem_cgroup_event, wait);
4461 struct mem_cgroup *memcg = event->memcg;
4462 __poll_t flags = key_to_poll(key);
4464 if (flags & EPOLLHUP) {
4466 * If the event has been detached at cgroup removal, we
4467 * can simply return knowing the other side will cleanup
4470 * We can't race against event freeing since the other
4471 * side will require wqh->lock via remove_wait_queue(),
4474 spin_lock(&memcg->event_list_lock);
4475 if (!list_empty(&event->list)) {
4476 list_del_init(&event->list);
4478 * We are in atomic context, but cgroup_event_remove()
4479 * may sleep, so we have to call it in workqueue.
4481 schedule_work(&event->remove);
4483 spin_unlock(&memcg->event_list_lock);
4489 static void memcg_event_ptable_queue_proc(struct file *file,
4490 wait_queue_head_t *wqh, poll_table *pt)
4492 struct mem_cgroup_event *event =
4493 container_of(pt, struct mem_cgroup_event, pt);
4496 add_wait_queue(wqh, &event->wait);
4500 * DO NOT USE IN NEW FILES.
4502 * Parse input and register new cgroup event handler.
4504 * Input must be in format '<event_fd> <control_fd> <args>'.
4505 * Interpretation of args is defined by control file implementation.
4507 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4508 char *buf, size_t nbytes, loff_t off)
4510 struct cgroup_subsys_state *css = of_css(of);
4511 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4512 struct mem_cgroup_event *event;
4513 struct cgroup_subsys_state *cfile_css;
4514 unsigned int efd, cfd;
4521 buf = strstrip(buf);
4523 efd = simple_strtoul(buf, &endp, 10);
4528 cfd = simple_strtoul(buf, &endp, 10);
4529 if ((*endp != ' ') && (*endp != '\0'))
4533 event = kzalloc(sizeof(*event), GFP_KERNEL);
4537 event->memcg = memcg;
4538 INIT_LIST_HEAD(&event->list);
4539 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4540 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4541 INIT_WORK(&event->remove, memcg_event_remove);
4549 event->eventfd = eventfd_ctx_fileget(efile.file);
4550 if (IS_ERR(event->eventfd)) {
4551 ret = PTR_ERR(event->eventfd);
4558 goto out_put_eventfd;
4561 /* the process need read permission on control file */
4562 /* AV: shouldn't we check that it's been opened for read instead? */
4563 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4568 * Determine the event callbacks and set them in @event. This used
4569 * to be done via struct cftype but cgroup core no longer knows
4570 * about these events. The following is crude but the whole thing
4571 * is for compatibility anyway.
4573 * DO NOT ADD NEW FILES.
4575 name = cfile.file->f_path.dentry->d_name.name;
4577 if (!strcmp(name, "memory.usage_in_bytes")) {
4578 event->register_event = mem_cgroup_usage_register_event;
4579 event->unregister_event = mem_cgroup_usage_unregister_event;
4580 } else if (!strcmp(name, "memory.oom_control")) {
4581 event->register_event = mem_cgroup_oom_register_event;
4582 event->unregister_event = mem_cgroup_oom_unregister_event;
4583 } else if (!strcmp(name, "memory.pressure_level")) {
4584 event->register_event = vmpressure_register_event;
4585 event->unregister_event = vmpressure_unregister_event;
4586 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4587 event->register_event = memsw_cgroup_usage_register_event;
4588 event->unregister_event = memsw_cgroup_usage_unregister_event;
4595 * Verify @cfile should belong to @css. Also, remaining events are
4596 * automatically removed on cgroup destruction but the removal is
4597 * asynchronous, so take an extra ref on @css.
4599 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4600 &memory_cgrp_subsys);
4602 if (IS_ERR(cfile_css))
4604 if (cfile_css != css) {
4609 ret = event->register_event(memcg, event->eventfd, buf);
4613 vfs_poll(efile.file, &event->pt);
4615 spin_lock(&memcg->event_list_lock);
4616 list_add(&event->list, &memcg->event_list);
4617 spin_unlock(&memcg->event_list_lock);
4629 eventfd_ctx_put(event->eventfd);
4638 static struct cftype mem_cgroup_legacy_files[] = {
4640 .name = "usage_in_bytes",
4641 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4642 .read_u64 = mem_cgroup_read_u64,
4645 .name = "max_usage_in_bytes",
4646 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4647 .write = mem_cgroup_reset,
4648 .read_u64 = mem_cgroup_read_u64,
4651 .name = "limit_in_bytes",
4652 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4653 .write = mem_cgroup_write,
4654 .read_u64 = mem_cgroup_read_u64,
4657 .name = "soft_limit_in_bytes",
4658 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4659 .write = mem_cgroup_write,
4660 .read_u64 = mem_cgroup_read_u64,
4664 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4665 .write = mem_cgroup_reset,
4666 .read_u64 = mem_cgroup_read_u64,
4670 .seq_show = memcg_stat_show,
4673 .name = "force_empty",
4674 .write = mem_cgroup_force_empty_write,
4677 .name = "use_hierarchy",
4678 .write_u64 = mem_cgroup_hierarchy_write,
4679 .read_u64 = mem_cgroup_hierarchy_read,
4682 .name = "cgroup.event_control", /* XXX: for compat */
4683 .write = memcg_write_event_control,
4684 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4687 .name = "swappiness",
4688 .read_u64 = mem_cgroup_swappiness_read,
4689 .write_u64 = mem_cgroup_swappiness_write,
4692 .name = "move_charge_at_immigrate",
4693 .read_u64 = mem_cgroup_move_charge_read,
4694 .write_u64 = mem_cgroup_move_charge_write,
4697 .name = "oom_control",
4698 .seq_show = mem_cgroup_oom_control_read,
4699 .write_u64 = mem_cgroup_oom_control_write,
4700 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4703 .name = "pressure_level",
4707 .name = "numa_stat",
4708 .seq_show = memcg_numa_stat_show,
4712 .name = "kmem.limit_in_bytes",
4713 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4714 .write = mem_cgroup_write,
4715 .read_u64 = mem_cgroup_read_u64,
4718 .name = "kmem.usage_in_bytes",
4719 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4720 .read_u64 = mem_cgroup_read_u64,
4723 .name = "kmem.failcnt",
4724 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4725 .write = mem_cgroup_reset,
4726 .read_u64 = mem_cgroup_read_u64,
4729 .name = "kmem.max_usage_in_bytes",
4730 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4731 .write = mem_cgroup_reset,
4732 .read_u64 = mem_cgroup_read_u64,
4734 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4736 .name = "kmem.slabinfo",
4737 .seq_start = memcg_slab_start,
4738 .seq_next = memcg_slab_next,
4739 .seq_stop = memcg_slab_stop,
4740 .seq_show = memcg_slab_show,
4744 .name = "kmem.tcp.limit_in_bytes",
4745 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4746 .write = mem_cgroup_write,
4747 .read_u64 = mem_cgroup_read_u64,
4750 .name = "kmem.tcp.usage_in_bytes",
4751 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4752 .read_u64 = mem_cgroup_read_u64,
4755 .name = "kmem.tcp.failcnt",
4756 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4757 .write = mem_cgroup_reset,
4758 .read_u64 = mem_cgroup_read_u64,
4761 .name = "kmem.tcp.max_usage_in_bytes",
4762 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4763 .write = mem_cgroup_reset,
4764 .read_u64 = mem_cgroup_read_u64,
4766 { }, /* terminate */
4770 * Private memory cgroup IDR
4772 * Swap-out records and page cache shadow entries need to store memcg
4773 * references in constrained space, so we maintain an ID space that is
4774 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4775 * memory-controlled cgroups to 64k.
4777 * However, there usually are many references to the oflline CSS after
4778 * the cgroup has been destroyed, such as page cache or reclaimable
4779 * slab objects, that don't need to hang on to the ID. We want to keep
4780 * those dead CSS from occupying IDs, or we might quickly exhaust the
4781 * relatively small ID space and prevent the creation of new cgroups
4782 * even when there are much fewer than 64k cgroups - possibly none.
4784 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4785 * be freed and recycled when it's no longer needed, which is usually
4786 * when the CSS is offlined.
4788 * The only exception to that are records of swapped out tmpfs/shmem
4789 * pages that need to be attributed to live ancestors on swapin. But
4790 * those references are manageable from userspace.
4793 static DEFINE_IDR(mem_cgroup_idr);
4795 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4797 if (memcg->id.id > 0) {
4798 idr_remove(&mem_cgroup_idr, memcg->id.id);
4803 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4805 refcount_add(n, &memcg->id.ref);
4808 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4810 if (refcount_sub_and_test(n, &memcg->id.ref)) {
4811 mem_cgroup_id_remove(memcg);
4813 /* Memcg ID pins CSS */
4814 css_put(&memcg->css);
4818 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4820 mem_cgroup_id_put_many(memcg, 1);
4824 * mem_cgroup_from_id - look up a memcg from a memcg id
4825 * @id: the memcg id to look up
4827 * Caller must hold rcu_read_lock().
4829 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4831 WARN_ON_ONCE(!rcu_read_lock_held());
4832 return idr_find(&mem_cgroup_idr, id);
4835 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4837 struct mem_cgroup_per_node *pn;
4840 * This routine is called against possible nodes.
4841 * But it's BUG to call kmalloc() against offline node.
4843 * TODO: this routine can waste much memory for nodes which will
4844 * never be onlined. It's better to use memory hotplug callback
4847 if (!node_state(node, N_NORMAL_MEMORY))
4849 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4853 pn->lruvec_stat_local = alloc_percpu(struct lruvec_stat);
4854 if (!pn->lruvec_stat_local) {
4859 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
4860 if (!pn->lruvec_stat_cpu) {
4861 free_percpu(pn->lruvec_stat_local);
4866 lruvec_init(&pn->lruvec);
4867 pn->usage_in_excess = 0;
4868 pn->on_tree = false;
4871 memcg->nodeinfo[node] = pn;
4875 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4877 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
4882 free_percpu(pn->lruvec_stat_cpu);
4883 free_percpu(pn->lruvec_stat_local);
4887 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4892 free_mem_cgroup_per_node_info(memcg, node);
4893 free_percpu(memcg->vmstats_percpu);
4894 free_percpu(memcg->vmstats_local);
4898 static void mem_cgroup_free(struct mem_cgroup *memcg)
4900 memcg_wb_domain_exit(memcg);
4902 * Flush percpu vmstats and vmevents to guarantee the value correctness
4903 * on parent's and all ancestor levels.
4905 memcg_flush_percpu_vmstats(memcg);
4906 memcg_flush_percpu_vmevents(memcg);
4907 __mem_cgroup_free(memcg);
4910 static struct mem_cgroup *mem_cgroup_alloc(void)
4912 struct mem_cgroup *memcg;
4915 int __maybe_unused i;
4917 size = sizeof(struct mem_cgroup);
4918 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4920 memcg = kzalloc(size, GFP_KERNEL);
4924 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4925 1, MEM_CGROUP_ID_MAX,
4927 if (memcg->id.id < 0)
4930 memcg->vmstats_local = alloc_percpu(struct memcg_vmstats_percpu);
4931 if (!memcg->vmstats_local)
4934 memcg->vmstats_percpu = alloc_percpu(struct memcg_vmstats_percpu);
4935 if (!memcg->vmstats_percpu)
4939 if (alloc_mem_cgroup_per_node_info(memcg, node))
4942 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4945 INIT_WORK(&memcg->high_work, high_work_func);
4946 INIT_LIST_HEAD(&memcg->oom_notify);
4947 mutex_init(&memcg->thresholds_lock);
4948 spin_lock_init(&memcg->move_lock);
4949 vmpressure_init(&memcg->vmpressure);
4950 INIT_LIST_HEAD(&memcg->event_list);
4951 spin_lock_init(&memcg->event_list_lock);
4952 memcg->socket_pressure = jiffies;
4953 #ifdef CONFIG_MEMCG_KMEM
4954 memcg->kmemcg_id = -1;
4956 #ifdef CONFIG_CGROUP_WRITEBACK
4957 INIT_LIST_HEAD(&memcg->cgwb_list);
4958 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
4959 memcg->cgwb_frn[i].done =
4960 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
4962 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4963 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
4964 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
4965 memcg->deferred_split_queue.split_queue_len = 0;
4967 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4970 mem_cgroup_id_remove(memcg);
4971 __mem_cgroup_free(memcg);
4975 static struct cgroup_subsys_state * __ref
4976 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4978 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4979 struct mem_cgroup *memcg;
4980 long error = -ENOMEM;
4982 memcg = mem_cgroup_alloc();
4984 return ERR_PTR(error);
4986 memcg->high = PAGE_COUNTER_MAX;
4987 memcg->soft_limit = PAGE_COUNTER_MAX;
4989 memcg->swappiness = mem_cgroup_swappiness(parent);
4990 memcg->oom_kill_disable = parent->oom_kill_disable;
4992 if (parent && parent->use_hierarchy) {
4993 memcg->use_hierarchy = true;
4994 page_counter_init(&memcg->memory, &parent->memory);
4995 page_counter_init(&memcg->swap, &parent->swap);
4996 page_counter_init(&memcg->memsw, &parent->memsw);
4997 page_counter_init(&memcg->kmem, &parent->kmem);
4998 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5000 page_counter_init(&memcg->memory, NULL);
5001 page_counter_init(&memcg->swap, NULL);
5002 page_counter_init(&memcg->memsw, NULL);
5003 page_counter_init(&memcg->kmem, NULL);
5004 page_counter_init(&memcg->tcpmem, NULL);
5006 * Deeper hierachy with use_hierarchy == false doesn't make
5007 * much sense so let cgroup subsystem know about this
5008 * unfortunate state in our controller.
5010 if (parent != root_mem_cgroup)
5011 memory_cgrp_subsys.broken_hierarchy = true;
5014 /* The following stuff does not apply to the root */
5016 #ifdef CONFIG_MEMCG_KMEM
5017 INIT_LIST_HEAD(&memcg->kmem_caches);
5019 root_mem_cgroup = memcg;
5023 error = memcg_online_kmem(memcg);
5027 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5028 static_branch_inc(&memcg_sockets_enabled_key);
5032 mem_cgroup_id_remove(memcg);
5033 mem_cgroup_free(memcg);
5034 return ERR_PTR(-ENOMEM);
5037 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5039 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5042 * A memcg must be visible for memcg_expand_shrinker_maps()
5043 * by the time the maps are allocated. So, we allocate maps
5044 * here, when for_each_mem_cgroup() can't skip it.
5046 if (memcg_alloc_shrinker_maps(memcg)) {
5047 mem_cgroup_id_remove(memcg);
5051 /* Online state pins memcg ID, memcg ID pins CSS */
5052 refcount_set(&memcg->id.ref, 1);
5057 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5059 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5060 struct mem_cgroup_event *event, *tmp;
5063 * Unregister events and notify userspace.
5064 * Notify userspace about cgroup removing only after rmdir of cgroup
5065 * directory to avoid race between userspace and kernelspace.
5067 spin_lock(&memcg->event_list_lock);
5068 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5069 list_del_init(&event->list);
5070 schedule_work(&event->remove);
5072 spin_unlock(&memcg->event_list_lock);
5074 page_counter_set_min(&memcg->memory, 0);
5075 page_counter_set_low(&memcg->memory, 0);
5077 memcg_offline_kmem(memcg);
5078 wb_memcg_offline(memcg);
5080 drain_all_stock(memcg);
5082 mem_cgroup_id_put(memcg);
5085 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5087 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5089 invalidate_reclaim_iterators(memcg);
5092 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5094 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5095 int __maybe_unused i;
5097 #ifdef CONFIG_CGROUP_WRITEBACK
5098 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5099 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5101 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5102 static_branch_dec(&memcg_sockets_enabled_key);
5104 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5105 static_branch_dec(&memcg_sockets_enabled_key);
5107 vmpressure_cleanup(&memcg->vmpressure);
5108 cancel_work_sync(&memcg->high_work);
5109 mem_cgroup_remove_from_trees(memcg);
5110 memcg_free_shrinker_maps(memcg);
5111 memcg_free_kmem(memcg);
5112 mem_cgroup_free(memcg);
5116 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5117 * @css: the target css
5119 * Reset the states of the mem_cgroup associated with @css. This is
5120 * invoked when the userland requests disabling on the default hierarchy
5121 * but the memcg is pinned through dependency. The memcg should stop
5122 * applying policies and should revert to the vanilla state as it may be
5123 * made visible again.
5125 * The current implementation only resets the essential configurations.
5126 * This needs to be expanded to cover all the visible parts.
5128 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5130 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5132 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5133 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5134 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
5135 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5136 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5137 page_counter_set_min(&memcg->memory, 0);
5138 page_counter_set_low(&memcg->memory, 0);
5139 memcg->high = PAGE_COUNTER_MAX;
5140 memcg->soft_limit = PAGE_COUNTER_MAX;
5141 memcg_wb_domain_size_changed(memcg);
5145 /* Handlers for move charge at task migration. */
5146 static int mem_cgroup_do_precharge(unsigned long count)
5150 /* Try a single bulk charge without reclaim first, kswapd may wake */
5151 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5153 mc.precharge += count;
5157 /* Try charges one by one with reclaim, but do not retry */
5159 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5173 enum mc_target_type {
5180 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5181 unsigned long addr, pte_t ptent)
5183 struct page *page = vm_normal_page(vma, addr, ptent);
5185 if (!page || !page_mapped(page))
5187 if (PageAnon(page)) {
5188 if (!(mc.flags & MOVE_ANON))
5191 if (!(mc.flags & MOVE_FILE))
5194 if (!get_page_unless_zero(page))
5200 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5201 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5202 pte_t ptent, swp_entry_t *entry)
5204 struct page *page = NULL;
5205 swp_entry_t ent = pte_to_swp_entry(ptent);
5207 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
5211 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5212 * a device and because they are not accessible by CPU they are store
5213 * as special swap entry in the CPU page table.
5215 if (is_device_private_entry(ent)) {
5216 page = device_private_entry_to_page(ent);
5218 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5219 * a refcount of 1 when free (unlike normal page)
5221 if (!page_ref_add_unless(page, 1, 1))
5227 * Because lookup_swap_cache() updates some statistics counter,
5228 * we call find_get_page() with swapper_space directly.
5230 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5231 if (do_memsw_account())
5232 entry->val = ent.val;
5237 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5238 pte_t ptent, swp_entry_t *entry)
5244 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5245 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5247 struct page *page = NULL;
5248 struct address_space *mapping;
5251 if (!vma->vm_file) /* anonymous vma */
5253 if (!(mc.flags & MOVE_FILE))
5256 mapping = vma->vm_file->f_mapping;
5257 pgoff = linear_page_index(vma, addr);
5259 /* page is moved even if it's not RSS of this task(page-faulted). */
5261 /* shmem/tmpfs may report page out on swap: account for that too. */
5262 if (shmem_mapping(mapping)) {
5263 page = find_get_entry(mapping, pgoff);
5264 if (xa_is_value(page)) {
5265 swp_entry_t swp = radix_to_swp_entry(page);
5266 if (do_memsw_account())
5268 page = find_get_page(swap_address_space(swp),
5272 page = find_get_page(mapping, pgoff);
5274 page = find_get_page(mapping, pgoff);
5280 * mem_cgroup_move_account - move account of the page
5282 * @compound: charge the page as compound or small page
5283 * @from: mem_cgroup which the page is moved from.
5284 * @to: mem_cgroup which the page is moved to. @from != @to.
5286 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5288 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5291 static int mem_cgroup_move_account(struct page *page,
5293 struct mem_cgroup *from,
5294 struct mem_cgroup *to)
5296 struct lruvec *from_vec, *to_vec;
5297 struct pglist_data *pgdat;
5298 unsigned long flags;
5299 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5303 VM_BUG_ON(from == to);
5304 VM_BUG_ON_PAGE(PageLRU(page), page);
5305 VM_BUG_ON(compound && !PageTransHuge(page));
5308 * Prevent mem_cgroup_migrate() from looking at
5309 * page->mem_cgroup of its source page while we change it.
5312 if (!trylock_page(page))
5316 if (page->mem_cgroup != from)
5319 anon = PageAnon(page);
5321 pgdat = page_pgdat(page);
5322 from_vec = mem_cgroup_lruvec(from, pgdat);
5323 to_vec = mem_cgroup_lruvec(to, pgdat);
5325 spin_lock_irqsave(&from->move_lock, flags);
5327 if (!anon && page_mapped(page)) {
5328 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5329 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5333 * move_lock grabbed above and caller set from->moving_account, so
5334 * mod_memcg_page_state will serialize updates to PageDirty.
5335 * So mapping should be stable for dirty pages.
5337 if (!anon && PageDirty(page)) {
5338 struct address_space *mapping = page_mapping(page);
5340 if (mapping_cap_account_dirty(mapping)) {
5341 __mod_lruvec_state(from_vec, NR_FILE_DIRTY, -nr_pages);
5342 __mod_lruvec_state(to_vec, NR_FILE_DIRTY, nr_pages);
5346 if (PageWriteback(page)) {
5347 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5348 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5352 * It is safe to change page->mem_cgroup here because the page
5353 * is referenced, charged, and isolated - we can't race with
5354 * uncharging, charging, migration, or LRU putback.
5357 /* caller should have done css_get */
5358 page->mem_cgroup = to;
5360 spin_unlock_irqrestore(&from->move_lock, flags);
5364 local_irq_disable();
5365 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
5366 memcg_check_events(to, page);
5367 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
5368 memcg_check_events(from, page);
5377 * get_mctgt_type - get target type of moving charge
5378 * @vma: the vma the pte to be checked belongs
5379 * @addr: the address corresponding to the pte to be checked
5380 * @ptent: the pte to be checked
5381 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5384 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5385 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5386 * move charge. if @target is not NULL, the page is stored in target->page
5387 * with extra refcnt got(Callers should handle it).
5388 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5389 * target for charge migration. if @target is not NULL, the entry is stored
5391 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5392 * (so ZONE_DEVICE page and thus not on the lru).
5393 * For now we such page is charge like a regular page would be as for all
5394 * intent and purposes it is just special memory taking the place of a
5397 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5399 * Called with pte lock held.
5402 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5403 unsigned long addr, pte_t ptent, union mc_target *target)
5405 struct page *page = NULL;
5406 enum mc_target_type ret = MC_TARGET_NONE;
5407 swp_entry_t ent = { .val = 0 };
5409 if (pte_present(ptent))
5410 page = mc_handle_present_pte(vma, addr, ptent);
5411 else if (is_swap_pte(ptent))
5412 page = mc_handle_swap_pte(vma, ptent, &ent);
5413 else if (pte_none(ptent))
5414 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5416 if (!page && !ent.val)
5420 * Do only loose check w/o serialization.
5421 * mem_cgroup_move_account() checks the page is valid or
5422 * not under LRU exclusion.
5424 if (page->mem_cgroup == mc.from) {
5425 ret = MC_TARGET_PAGE;
5426 if (is_device_private_page(page))
5427 ret = MC_TARGET_DEVICE;
5429 target->page = page;
5431 if (!ret || !target)
5435 * There is a swap entry and a page doesn't exist or isn't charged.
5436 * But we cannot move a tail-page in a THP.
5438 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5439 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5440 ret = MC_TARGET_SWAP;
5447 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5449 * We don't consider PMD mapped swapping or file mapped pages because THP does
5450 * not support them for now.
5451 * Caller should make sure that pmd_trans_huge(pmd) is true.
5453 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5454 unsigned long addr, pmd_t pmd, union mc_target *target)
5456 struct page *page = NULL;
5457 enum mc_target_type ret = MC_TARGET_NONE;
5459 if (unlikely(is_swap_pmd(pmd))) {
5460 VM_BUG_ON(thp_migration_supported() &&
5461 !is_pmd_migration_entry(pmd));
5464 page = pmd_page(pmd);
5465 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5466 if (!(mc.flags & MOVE_ANON))
5468 if (page->mem_cgroup == mc.from) {
5469 ret = MC_TARGET_PAGE;
5472 target->page = page;
5478 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5479 unsigned long addr, pmd_t pmd, union mc_target *target)
5481 return MC_TARGET_NONE;
5485 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5486 unsigned long addr, unsigned long end,
5487 struct mm_walk *walk)
5489 struct vm_area_struct *vma = walk->vma;
5493 ptl = pmd_trans_huge_lock(pmd, vma);
5496 * Note their can not be MC_TARGET_DEVICE for now as we do not
5497 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5498 * this might change.
5500 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5501 mc.precharge += HPAGE_PMD_NR;
5506 if (pmd_trans_unstable(pmd))
5508 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5509 for (; addr != end; pte++, addr += PAGE_SIZE)
5510 if (get_mctgt_type(vma, addr, *pte, NULL))
5511 mc.precharge++; /* increment precharge temporarily */
5512 pte_unmap_unlock(pte - 1, ptl);
5518 static const struct mm_walk_ops precharge_walk_ops = {
5519 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5522 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5524 unsigned long precharge;
5526 down_read(&mm->mmap_sem);
5527 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5528 up_read(&mm->mmap_sem);
5530 precharge = mc.precharge;
5536 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5538 unsigned long precharge = mem_cgroup_count_precharge(mm);
5540 VM_BUG_ON(mc.moving_task);
5541 mc.moving_task = current;
5542 return mem_cgroup_do_precharge(precharge);
5545 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5546 static void __mem_cgroup_clear_mc(void)
5548 struct mem_cgroup *from = mc.from;
5549 struct mem_cgroup *to = mc.to;
5551 /* we must uncharge all the leftover precharges from mc.to */
5553 cancel_charge(mc.to, mc.precharge);
5557 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5558 * we must uncharge here.
5560 if (mc.moved_charge) {
5561 cancel_charge(mc.from, mc.moved_charge);
5562 mc.moved_charge = 0;
5564 /* we must fixup refcnts and charges */
5565 if (mc.moved_swap) {
5566 /* uncharge swap account from the old cgroup */
5567 if (!mem_cgroup_is_root(mc.from))
5568 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5570 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5573 * we charged both to->memory and to->memsw, so we
5574 * should uncharge to->memory.
5576 if (!mem_cgroup_is_root(mc.to))
5577 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5579 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
5580 css_put_many(&mc.to->css, mc.moved_swap);
5584 memcg_oom_recover(from);
5585 memcg_oom_recover(to);
5586 wake_up_all(&mc.waitq);
5589 static void mem_cgroup_clear_mc(void)
5591 struct mm_struct *mm = mc.mm;
5594 * we must clear moving_task before waking up waiters at the end of
5597 mc.moving_task = NULL;
5598 __mem_cgroup_clear_mc();
5599 spin_lock(&mc.lock);
5603 spin_unlock(&mc.lock);
5608 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5610 struct cgroup_subsys_state *css;
5611 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5612 struct mem_cgroup *from;
5613 struct task_struct *leader, *p;
5614 struct mm_struct *mm;
5615 unsigned long move_flags;
5618 /* charge immigration isn't supported on the default hierarchy */
5619 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5623 * Multi-process migrations only happen on the default hierarchy
5624 * where charge immigration is not used. Perform charge
5625 * immigration if @tset contains a leader and whine if there are
5629 cgroup_taskset_for_each_leader(leader, css, tset) {
5632 memcg = mem_cgroup_from_css(css);
5638 * We are now commited to this value whatever it is. Changes in this
5639 * tunable will only affect upcoming migrations, not the current one.
5640 * So we need to save it, and keep it going.
5642 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5646 from = mem_cgroup_from_task(p);
5648 VM_BUG_ON(from == memcg);
5650 mm = get_task_mm(p);
5653 /* We move charges only when we move a owner of the mm */
5654 if (mm->owner == p) {
5657 VM_BUG_ON(mc.precharge);
5658 VM_BUG_ON(mc.moved_charge);
5659 VM_BUG_ON(mc.moved_swap);
5661 spin_lock(&mc.lock);
5665 mc.flags = move_flags;
5666 spin_unlock(&mc.lock);
5667 /* We set mc.moving_task later */
5669 ret = mem_cgroup_precharge_mc(mm);
5671 mem_cgroup_clear_mc();
5678 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5681 mem_cgroup_clear_mc();
5684 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5685 unsigned long addr, unsigned long end,
5686 struct mm_walk *walk)
5689 struct vm_area_struct *vma = walk->vma;
5692 enum mc_target_type target_type;
5693 union mc_target target;
5696 ptl = pmd_trans_huge_lock(pmd, vma);
5698 if (mc.precharge < HPAGE_PMD_NR) {
5702 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5703 if (target_type == MC_TARGET_PAGE) {
5705 if (!isolate_lru_page(page)) {
5706 if (!mem_cgroup_move_account(page, true,
5708 mc.precharge -= HPAGE_PMD_NR;
5709 mc.moved_charge += HPAGE_PMD_NR;
5711 putback_lru_page(page);
5714 } else if (target_type == MC_TARGET_DEVICE) {
5716 if (!mem_cgroup_move_account(page, true,
5718 mc.precharge -= HPAGE_PMD_NR;
5719 mc.moved_charge += HPAGE_PMD_NR;
5727 if (pmd_trans_unstable(pmd))
5730 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5731 for (; addr != end; addr += PAGE_SIZE) {
5732 pte_t ptent = *(pte++);
5733 bool device = false;
5739 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5740 case MC_TARGET_DEVICE:
5743 case MC_TARGET_PAGE:
5746 * We can have a part of the split pmd here. Moving it
5747 * can be done but it would be too convoluted so simply
5748 * ignore such a partial THP and keep it in original
5749 * memcg. There should be somebody mapping the head.
5751 if (PageTransCompound(page))
5753 if (!device && isolate_lru_page(page))
5755 if (!mem_cgroup_move_account(page, false,
5758 /* we uncharge from mc.from later. */
5762 putback_lru_page(page);
5763 put: /* get_mctgt_type() gets the page */
5766 case MC_TARGET_SWAP:
5768 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5770 /* we fixup refcnts and charges later. */
5778 pte_unmap_unlock(pte - 1, ptl);
5783 * We have consumed all precharges we got in can_attach().
5784 * We try charge one by one, but don't do any additional
5785 * charges to mc.to if we have failed in charge once in attach()
5788 ret = mem_cgroup_do_precharge(1);
5796 static const struct mm_walk_ops charge_walk_ops = {
5797 .pmd_entry = mem_cgroup_move_charge_pte_range,
5800 static void mem_cgroup_move_charge(void)
5802 lru_add_drain_all();
5804 * Signal lock_page_memcg() to take the memcg's move_lock
5805 * while we're moving its pages to another memcg. Then wait
5806 * for already started RCU-only updates to finish.
5808 atomic_inc(&mc.from->moving_account);
5811 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5813 * Someone who are holding the mmap_sem might be waiting in
5814 * waitq. So we cancel all extra charges, wake up all waiters,
5815 * and retry. Because we cancel precharges, we might not be able
5816 * to move enough charges, but moving charge is a best-effort
5817 * feature anyway, so it wouldn't be a big problem.
5819 __mem_cgroup_clear_mc();
5824 * When we have consumed all precharges and failed in doing
5825 * additional charge, the page walk just aborts.
5827 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
5830 up_read(&mc.mm->mmap_sem);
5831 atomic_dec(&mc.from->moving_account);
5834 static void mem_cgroup_move_task(void)
5837 mem_cgroup_move_charge();
5838 mem_cgroup_clear_mc();
5841 #else /* !CONFIG_MMU */
5842 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5846 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5849 static void mem_cgroup_move_task(void)
5855 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5856 * to verify whether we're attached to the default hierarchy on each mount
5859 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5862 * use_hierarchy is forced on the default hierarchy. cgroup core
5863 * guarantees that @root doesn't have any children, so turning it
5864 * on for the root memcg is enough.
5866 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5867 root_mem_cgroup->use_hierarchy = true;
5869 root_mem_cgroup->use_hierarchy = false;
5872 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
5874 if (value == PAGE_COUNTER_MAX)
5875 seq_puts(m, "max\n");
5877 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
5882 static u64 memory_current_read(struct cgroup_subsys_state *css,
5885 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5887 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5890 static int memory_min_show(struct seq_file *m, void *v)
5892 return seq_puts_memcg_tunable(m,
5893 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
5896 static ssize_t memory_min_write(struct kernfs_open_file *of,
5897 char *buf, size_t nbytes, loff_t off)
5899 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5903 buf = strstrip(buf);
5904 err = page_counter_memparse(buf, "max", &min);
5908 page_counter_set_min(&memcg->memory, min);
5913 static int memory_low_show(struct seq_file *m, void *v)
5915 return seq_puts_memcg_tunable(m,
5916 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
5919 static ssize_t memory_low_write(struct kernfs_open_file *of,
5920 char *buf, size_t nbytes, loff_t off)
5922 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5926 buf = strstrip(buf);
5927 err = page_counter_memparse(buf, "max", &low);
5931 page_counter_set_low(&memcg->memory, low);
5936 static int memory_high_show(struct seq_file *m, void *v)
5938 return seq_puts_memcg_tunable(m, READ_ONCE(mem_cgroup_from_seq(m)->high));
5941 static ssize_t memory_high_write(struct kernfs_open_file *of,
5942 char *buf, size_t nbytes, loff_t off)
5944 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5945 unsigned int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
5946 bool drained = false;
5950 buf = strstrip(buf);
5951 err = page_counter_memparse(buf, "max", &high);
5958 unsigned long nr_pages = page_counter_read(&memcg->memory);
5959 unsigned long reclaimed;
5961 if (nr_pages <= high)
5964 if (signal_pending(current))
5968 drain_all_stock(memcg);
5973 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5976 if (!reclaimed && !nr_retries--)
5983 static int memory_max_show(struct seq_file *m, void *v)
5985 return seq_puts_memcg_tunable(m,
5986 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
5989 static ssize_t memory_max_write(struct kernfs_open_file *of,
5990 char *buf, size_t nbytes, loff_t off)
5992 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5993 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5994 bool drained = false;
5998 buf = strstrip(buf);
5999 err = page_counter_memparse(buf, "max", &max);
6003 xchg(&memcg->memory.max, max);
6006 unsigned long nr_pages = page_counter_read(&memcg->memory);
6008 if (nr_pages <= max)
6011 if (signal_pending(current))
6015 drain_all_stock(memcg);
6021 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6027 memcg_memory_event(memcg, MEMCG_OOM);
6028 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6032 memcg_wb_domain_size_changed(memcg);
6036 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6038 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6039 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6040 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6041 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6042 seq_printf(m, "oom_kill %lu\n",
6043 atomic_long_read(&events[MEMCG_OOM_KILL]));
6046 static int memory_events_show(struct seq_file *m, void *v)
6048 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6050 __memory_events_show(m, memcg->memory_events);
6054 static int memory_events_local_show(struct seq_file *m, void *v)
6056 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6058 __memory_events_show(m, memcg->memory_events_local);
6062 static int memory_stat_show(struct seq_file *m, void *v)
6064 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6067 buf = memory_stat_format(memcg);
6075 static int memory_oom_group_show(struct seq_file *m, void *v)
6077 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6079 seq_printf(m, "%d\n", memcg->oom_group);
6084 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6085 char *buf, size_t nbytes, loff_t off)
6087 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6090 buf = strstrip(buf);
6094 ret = kstrtoint(buf, 0, &oom_group);
6098 if (oom_group != 0 && oom_group != 1)
6101 memcg->oom_group = oom_group;
6106 static struct cftype memory_files[] = {
6109 .flags = CFTYPE_NOT_ON_ROOT,
6110 .read_u64 = memory_current_read,
6114 .flags = CFTYPE_NOT_ON_ROOT,
6115 .seq_show = memory_min_show,
6116 .write = memory_min_write,
6120 .flags = CFTYPE_NOT_ON_ROOT,
6121 .seq_show = memory_low_show,
6122 .write = memory_low_write,
6126 .flags = CFTYPE_NOT_ON_ROOT,
6127 .seq_show = memory_high_show,
6128 .write = memory_high_write,
6132 .flags = CFTYPE_NOT_ON_ROOT,
6133 .seq_show = memory_max_show,
6134 .write = memory_max_write,
6138 .flags = CFTYPE_NOT_ON_ROOT,
6139 .file_offset = offsetof(struct mem_cgroup, events_file),
6140 .seq_show = memory_events_show,
6143 .name = "events.local",
6144 .flags = CFTYPE_NOT_ON_ROOT,
6145 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6146 .seq_show = memory_events_local_show,
6150 .flags = CFTYPE_NOT_ON_ROOT,
6151 .seq_show = memory_stat_show,
6154 .name = "oom.group",
6155 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6156 .seq_show = memory_oom_group_show,
6157 .write = memory_oom_group_write,
6162 struct cgroup_subsys memory_cgrp_subsys = {
6163 .css_alloc = mem_cgroup_css_alloc,
6164 .css_online = mem_cgroup_css_online,
6165 .css_offline = mem_cgroup_css_offline,
6166 .css_released = mem_cgroup_css_released,
6167 .css_free = mem_cgroup_css_free,
6168 .css_reset = mem_cgroup_css_reset,
6169 .can_attach = mem_cgroup_can_attach,
6170 .cancel_attach = mem_cgroup_cancel_attach,
6171 .post_attach = mem_cgroup_move_task,
6172 .bind = mem_cgroup_bind,
6173 .dfl_cftypes = memory_files,
6174 .legacy_cftypes = mem_cgroup_legacy_files,
6179 * mem_cgroup_protected - check if memory consumption is in the normal range
6180 * @root: the top ancestor of the sub-tree being checked
6181 * @memcg: the memory cgroup to check
6183 * WARNING: This function is not stateless! It can only be used as part
6184 * of a top-down tree iteration, not for isolated queries.
6186 * Returns one of the following:
6187 * MEMCG_PROT_NONE: cgroup memory is not protected
6188 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
6189 * an unprotected supply of reclaimable memory from other cgroups.
6190 * MEMCG_PROT_MIN: cgroup memory is protected
6192 * @root is exclusive; it is never protected when looked at directly
6194 * To provide a proper hierarchical behavior, effective memory.min/low values
6195 * are used. Below is the description of how effective memory.low is calculated.
6196 * Effective memory.min values is calculated in the same way.
6198 * Effective memory.low is always equal or less than the original memory.low.
6199 * If there is no memory.low overcommittment (which is always true for
6200 * top-level memory cgroups), these two values are equal.
6201 * Otherwise, it's a part of parent's effective memory.low,
6202 * calculated as a cgroup's memory.low usage divided by sum of sibling's
6203 * memory.low usages, where memory.low usage is the size of actually
6207 * elow = min( memory.low, parent->elow * ------------------ ),
6208 * siblings_low_usage
6210 * | memory.current, if memory.current < memory.low
6215 * Such definition of the effective memory.low provides the expected
6216 * hierarchical behavior: parent's memory.low value is limiting
6217 * children, unprotected memory is reclaimed first and cgroups,
6218 * which are not using their guarantee do not affect actual memory
6221 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
6223 * A A/memory.low = 2G, A/memory.current = 6G
6225 * BC DE B/memory.low = 3G B/memory.current = 2G
6226 * C/memory.low = 1G C/memory.current = 2G
6227 * D/memory.low = 0 D/memory.current = 2G
6228 * E/memory.low = 10G E/memory.current = 0
6230 * and the memory pressure is applied, the following memory distribution
6231 * is expected (approximately):
6233 * A/memory.current = 2G
6235 * B/memory.current = 1.3G
6236 * C/memory.current = 0.6G
6237 * D/memory.current = 0
6238 * E/memory.current = 0
6240 * These calculations require constant tracking of the actual low usages
6241 * (see propagate_protected_usage()), as well as recursive calculation of
6242 * effective memory.low values. But as we do call mem_cgroup_protected()
6243 * path for each memory cgroup top-down from the reclaim,
6244 * it's possible to optimize this part, and save calculated elow
6245 * for next usage. This part is intentionally racy, but it's ok,
6246 * as memory.low is a best-effort mechanism.
6248 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
6249 struct mem_cgroup *memcg)
6251 struct mem_cgroup *parent;
6252 unsigned long emin, parent_emin;
6253 unsigned long elow, parent_elow;
6254 unsigned long usage;
6256 if (mem_cgroup_disabled())
6257 return MEMCG_PROT_NONE;
6260 root = root_mem_cgroup;
6262 return MEMCG_PROT_NONE;
6264 usage = page_counter_read(&memcg->memory);
6266 return MEMCG_PROT_NONE;
6268 emin = memcg->memory.min;
6269 elow = memcg->memory.low;
6271 parent = parent_mem_cgroup(memcg);
6272 /* No parent means a non-hierarchical mode on v1 memcg */
6274 return MEMCG_PROT_NONE;
6279 parent_emin = READ_ONCE(parent->memory.emin);
6280 emin = min(emin, parent_emin);
6281 if (emin && parent_emin) {
6282 unsigned long min_usage, siblings_min_usage;
6284 min_usage = min(usage, memcg->memory.min);
6285 siblings_min_usage = atomic_long_read(
6286 &parent->memory.children_min_usage);
6288 if (min_usage && siblings_min_usage)
6289 emin = min(emin, parent_emin * min_usage /
6290 siblings_min_usage);
6293 parent_elow = READ_ONCE(parent->memory.elow);
6294 elow = min(elow, parent_elow);
6295 if (elow && parent_elow) {
6296 unsigned long low_usage, siblings_low_usage;
6298 low_usage = min(usage, memcg->memory.low);
6299 siblings_low_usage = atomic_long_read(
6300 &parent->memory.children_low_usage);
6302 if (low_usage && siblings_low_usage)
6303 elow = min(elow, parent_elow * low_usage /
6304 siblings_low_usage);
6308 memcg->memory.emin = emin;
6309 memcg->memory.elow = elow;
6312 return MEMCG_PROT_MIN;
6313 else if (usage <= elow)
6314 return MEMCG_PROT_LOW;
6316 return MEMCG_PROT_NONE;
6320 * mem_cgroup_try_charge - try charging a page
6321 * @page: page to charge
6322 * @mm: mm context of the victim
6323 * @gfp_mask: reclaim mode
6324 * @memcgp: charged memcg return
6325 * @compound: charge the page as compound or small page
6327 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6328 * pages according to @gfp_mask if necessary.
6330 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6331 * Otherwise, an error code is returned.
6333 * After page->mapping has been set up, the caller must finalize the
6334 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6335 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6337 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
6338 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6341 struct mem_cgroup *memcg = NULL;
6342 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6345 if (mem_cgroup_disabled())
6348 if (PageSwapCache(page)) {
6350 * Every swap fault against a single page tries to charge the
6351 * page, bail as early as possible. shmem_unuse() encounters
6352 * already charged pages, too. The USED bit is protected by
6353 * the page lock, which serializes swap cache removal, which
6354 * in turn serializes uncharging.
6356 VM_BUG_ON_PAGE(!PageLocked(page), page);
6357 if (compound_head(page)->mem_cgroup)
6360 if (do_swap_account) {
6361 swp_entry_t ent = { .val = page_private(page), };
6362 unsigned short id = lookup_swap_cgroup_id(ent);
6365 memcg = mem_cgroup_from_id(id);
6366 if (memcg && !css_tryget_online(&memcg->css))
6373 memcg = get_mem_cgroup_from_mm(mm);
6375 ret = try_charge(memcg, gfp_mask, nr_pages);
6377 css_put(&memcg->css);
6383 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
6384 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6387 struct mem_cgroup *memcg;
6390 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound);
6392 mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask);
6397 * mem_cgroup_commit_charge - commit a page charge
6398 * @page: page to charge
6399 * @memcg: memcg to charge the page to
6400 * @lrucare: page might be on LRU already
6401 * @compound: charge the page as compound or small page
6403 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6404 * after page->mapping has been set up. This must happen atomically
6405 * as part of the page instantiation, i.e. under the page table lock
6406 * for anonymous pages, under the page lock for page and swap cache.
6408 * In addition, the page must not be on the LRU during the commit, to
6409 * prevent racing with task migration. If it might be, use @lrucare.
6411 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6413 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
6414 bool lrucare, bool compound)
6416 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6418 VM_BUG_ON_PAGE(!page->mapping, page);
6419 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
6421 if (mem_cgroup_disabled())
6424 * Swap faults will attempt to charge the same page multiple
6425 * times. But reuse_swap_page() might have removed the page
6426 * from swapcache already, so we can't check PageSwapCache().
6431 commit_charge(page, memcg, lrucare);
6433 local_irq_disable();
6434 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
6435 memcg_check_events(memcg, page);
6438 if (do_memsw_account() && PageSwapCache(page)) {
6439 swp_entry_t entry = { .val = page_private(page) };
6441 * The swap entry might not get freed for a long time,
6442 * let's not wait for it. The page already received a
6443 * memory+swap charge, drop the swap entry duplicate.
6445 mem_cgroup_uncharge_swap(entry, nr_pages);
6450 * mem_cgroup_cancel_charge - cancel a page charge
6451 * @page: page to charge
6452 * @memcg: memcg to charge the page to
6453 * @compound: charge the page as compound or small page
6455 * Cancel a charge transaction started by mem_cgroup_try_charge().
6457 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
6460 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6462 if (mem_cgroup_disabled())
6465 * Swap faults will attempt to charge the same page multiple
6466 * times. But reuse_swap_page() might have removed the page
6467 * from swapcache already, so we can't check PageSwapCache().
6472 cancel_charge(memcg, nr_pages);
6475 struct uncharge_gather {
6476 struct mem_cgroup *memcg;
6477 unsigned long pgpgout;
6478 unsigned long nr_anon;
6479 unsigned long nr_file;
6480 unsigned long nr_kmem;
6481 unsigned long nr_huge;
6482 unsigned long nr_shmem;
6483 struct page *dummy_page;
6486 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6488 memset(ug, 0, sizeof(*ug));
6491 static void uncharge_batch(const struct uncharge_gather *ug)
6493 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
6494 unsigned long flags;
6496 if (!mem_cgroup_is_root(ug->memcg)) {
6497 page_counter_uncharge(&ug->memcg->memory, nr_pages);
6498 if (do_memsw_account())
6499 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
6500 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6501 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6502 memcg_oom_recover(ug->memcg);
6505 local_irq_save(flags);
6506 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
6507 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
6508 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
6509 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
6510 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6511 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, nr_pages);
6512 memcg_check_events(ug->memcg, ug->dummy_page);
6513 local_irq_restore(flags);
6515 if (!mem_cgroup_is_root(ug->memcg))
6516 css_put_many(&ug->memcg->css, nr_pages);
6519 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6521 VM_BUG_ON_PAGE(PageLRU(page), page);
6522 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
6523 !PageHWPoison(page) , page);
6525 if (!page->mem_cgroup)
6529 * Nobody should be changing or seriously looking at
6530 * page->mem_cgroup at this point, we have fully
6531 * exclusive access to the page.
6534 if (ug->memcg != page->mem_cgroup) {
6537 uncharge_gather_clear(ug);
6539 ug->memcg = page->mem_cgroup;
6542 if (!PageKmemcg(page)) {
6543 unsigned int nr_pages = 1;
6545 if (PageTransHuge(page)) {
6546 nr_pages = compound_nr(page);
6547 ug->nr_huge += nr_pages;
6550 ug->nr_anon += nr_pages;
6552 ug->nr_file += nr_pages;
6553 if (PageSwapBacked(page))
6554 ug->nr_shmem += nr_pages;
6558 ug->nr_kmem += compound_nr(page);
6559 __ClearPageKmemcg(page);
6562 ug->dummy_page = page;
6563 page->mem_cgroup = NULL;
6566 static void uncharge_list(struct list_head *page_list)
6568 struct uncharge_gather ug;
6569 struct list_head *next;
6571 uncharge_gather_clear(&ug);
6574 * Note that the list can be a single page->lru; hence the
6575 * do-while loop instead of a simple list_for_each_entry().
6577 next = page_list->next;
6581 page = list_entry(next, struct page, lru);
6582 next = page->lru.next;
6584 uncharge_page(page, &ug);
6585 } while (next != page_list);
6588 uncharge_batch(&ug);
6592 * mem_cgroup_uncharge - uncharge a page
6593 * @page: page to uncharge
6595 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6596 * mem_cgroup_commit_charge().
6598 void mem_cgroup_uncharge(struct page *page)
6600 struct uncharge_gather ug;
6602 if (mem_cgroup_disabled())
6605 /* Don't touch page->lru of any random page, pre-check: */
6606 if (!page->mem_cgroup)
6609 uncharge_gather_clear(&ug);
6610 uncharge_page(page, &ug);
6611 uncharge_batch(&ug);
6615 * mem_cgroup_uncharge_list - uncharge a list of page
6616 * @page_list: list of pages to uncharge
6618 * Uncharge a list of pages previously charged with
6619 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6621 void mem_cgroup_uncharge_list(struct list_head *page_list)
6623 if (mem_cgroup_disabled())
6626 if (!list_empty(page_list))
6627 uncharge_list(page_list);
6631 * mem_cgroup_migrate - charge a page's replacement
6632 * @oldpage: currently circulating page
6633 * @newpage: replacement page
6635 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6636 * be uncharged upon free.
6638 * Both pages must be locked, @newpage->mapping must be set up.
6640 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6642 struct mem_cgroup *memcg;
6643 unsigned int nr_pages;
6644 unsigned long flags;
6646 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6647 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6648 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6649 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6652 if (mem_cgroup_disabled())
6655 /* Page cache replacement: new page already charged? */
6656 if (newpage->mem_cgroup)
6659 /* Swapcache readahead pages can get replaced before being charged */
6660 memcg = oldpage->mem_cgroup;
6664 /* Force-charge the new page. The old one will be freed soon */
6665 nr_pages = hpage_nr_pages(newpage);
6667 page_counter_charge(&memcg->memory, nr_pages);
6668 if (do_memsw_account())
6669 page_counter_charge(&memcg->memsw, nr_pages);
6670 css_get_many(&memcg->css, nr_pages);
6672 commit_charge(newpage, memcg, false);
6674 local_irq_save(flags);
6675 mem_cgroup_charge_statistics(memcg, newpage, PageTransHuge(newpage),
6677 memcg_check_events(memcg, newpage);
6678 local_irq_restore(flags);
6681 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6682 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6684 void mem_cgroup_sk_alloc(struct sock *sk)
6686 struct mem_cgroup *memcg;
6688 if (!mem_cgroup_sockets_enabled)
6691 /* Do not associate the sock with unrelated interrupted task's memcg. */
6696 memcg = mem_cgroup_from_task(current);
6697 if (memcg == root_mem_cgroup)
6699 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6701 if (css_tryget_online(&memcg->css))
6702 sk->sk_memcg = memcg;
6707 void mem_cgroup_sk_free(struct sock *sk)
6710 css_put(&sk->sk_memcg->css);
6714 * mem_cgroup_charge_skmem - charge socket memory
6715 * @memcg: memcg to charge
6716 * @nr_pages: number of pages to charge
6718 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6719 * @memcg's configured limit, %false if the charge had to be forced.
6721 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6723 gfp_t gfp_mask = GFP_KERNEL;
6725 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6726 struct page_counter *fail;
6728 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6729 memcg->tcpmem_pressure = 0;
6732 page_counter_charge(&memcg->tcpmem, nr_pages);
6733 memcg->tcpmem_pressure = 1;
6737 /* Don't block in the packet receive path */
6739 gfp_mask = GFP_NOWAIT;
6741 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6743 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6746 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6751 * mem_cgroup_uncharge_skmem - uncharge socket memory
6752 * @memcg: memcg to uncharge
6753 * @nr_pages: number of pages to uncharge
6755 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6757 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6758 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6762 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6764 refill_stock(memcg, nr_pages);
6767 static int __init cgroup_memory(char *s)
6771 while ((token = strsep(&s, ",")) != NULL) {
6774 if (!strcmp(token, "nosocket"))
6775 cgroup_memory_nosocket = true;
6776 if (!strcmp(token, "nokmem"))
6777 cgroup_memory_nokmem = true;
6781 __setup("cgroup.memory=", cgroup_memory);
6784 * subsys_initcall() for memory controller.
6786 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6787 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6788 * basically everything that doesn't depend on a specific mem_cgroup structure
6789 * should be initialized from here.
6791 static int __init mem_cgroup_init(void)
6795 #ifdef CONFIG_MEMCG_KMEM
6797 * Kmem cache creation is mostly done with the slab_mutex held,
6798 * so use a workqueue with limited concurrency to avoid stalling
6799 * all worker threads in case lots of cgroups are created and
6800 * destroyed simultaneously.
6802 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
6803 BUG_ON(!memcg_kmem_cache_wq);
6806 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6807 memcg_hotplug_cpu_dead);
6809 for_each_possible_cpu(cpu)
6810 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
6813 for_each_node(node) {
6814 struct mem_cgroup_tree_per_node *rtpn;
6816 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
6817 node_online(node) ? node : NUMA_NO_NODE);
6819 rtpn->rb_root = RB_ROOT;
6820 rtpn->rb_rightmost = NULL;
6821 spin_lock_init(&rtpn->lock);
6822 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6827 subsys_initcall(mem_cgroup_init);
6829 #ifdef CONFIG_MEMCG_SWAP
6830 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
6832 while (!refcount_inc_not_zero(&memcg->id.ref)) {
6834 * The root cgroup cannot be destroyed, so it's refcount must
6837 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
6841 memcg = parent_mem_cgroup(memcg);
6843 memcg = root_mem_cgroup;
6849 * mem_cgroup_swapout - transfer a memsw charge to swap
6850 * @page: page whose memsw charge to transfer
6851 * @entry: swap entry to move the charge to
6853 * Transfer the memsw charge of @page to @entry.
6855 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6857 struct mem_cgroup *memcg, *swap_memcg;
6858 unsigned int nr_entries;
6859 unsigned short oldid;
6861 VM_BUG_ON_PAGE(PageLRU(page), page);
6862 VM_BUG_ON_PAGE(page_count(page), page);
6864 if (!do_memsw_account())
6867 memcg = page->mem_cgroup;
6869 /* Readahead page, never charged */
6874 * In case the memcg owning these pages has been offlined and doesn't
6875 * have an ID allocated to it anymore, charge the closest online
6876 * ancestor for the swap instead and transfer the memory+swap charge.
6878 swap_memcg = mem_cgroup_id_get_online(memcg);
6879 nr_entries = hpage_nr_pages(page);
6880 /* Get references for the tail pages, too */
6882 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
6883 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
6885 VM_BUG_ON_PAGE(oldid, page);
6886 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
6888 page->mem_cgroup = NULL;
6890 if (!mem_cgroup_is_root(memcg))
6891 page_counter_uncharge(&memcg->memory, nr_entries);
6893 if (memcg != swap_memcg) {
6894 if (!mem_cgroup_is_root(swap_memcg))
6895 page_counter_charge(&swap_memcg->memsw, nr_entries);
6896 page_counter_uncharge(&memcg->memsw, nr_entries);
6900 * Interrupts should be disabled here because the caller holds the
6901 * i_pages lock which is taken with interrupts-off. It is
6902 * important here to have the interrupts disabled because it is the
6903 * only synchronisation we have for updating the per-CPU variables.
6905 VM_BUG_ON(!irqs_disabled());
6906 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
6908 memcg_check_events(memcg, page);
6910 if (!mem_cgroup_is_root(memcg))
6911 css_put_many(&memcg->css, nr_entries);
6915 * mem_cgroup_try_charge_swap - try charging swap space for a page
6916 * @page: page being added to swap
6917 * @entry: swap entry to charge
6919 * Try to charge @page's memcg for the swap space at @entry.
6921 * Returns 0 on success, -ENOMEM on failure.
6923 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
6925 unsigned int nr_pages = hpage_nr_pages(page);
6926 struct page_counter *counter;
6927 struct mem_cgroup *memcg;
6928 unsigned short oldid;
6930 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
6933 memcg = page->mem_cgroup;
6935 /* Readahead page, never charged */
6940 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6944 memcg = mem_cgroup_id_get_online(memcg);
6946 if (!mem_cgroup_is_root(memcg) &&
6947 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
6948 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
6949 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6950 mem_cgroup_id_put(memcg);
6954 /* Get references for the tail pages, too */
6956 mem_cgroup_id_get_many(memcg, nr_pages - 1);
6957 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
6958 VM_BUG_ON_PAGE(oldid, page);
6959 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
6965 * mem_cgroup_uncharge_swap - uncharge swap space
6966 * @entry: swap entry to uncharge
6967 * @nr_pages: the amount of swap space to uncharge
6969 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
6971 struct mem_cgroup *memcg;
6974 if (!do_swap_account)
6977 id = swap_cgroup_record(entry, 0, nr_pages);
6979 memcg = mem_cgroup_from_id(id);
6981 if (!mem_cgroup_is_root(memcg)) {
6982 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6983 page_counter_uncharge(&memcg->swap, nr_pages);
6985 page_counter_uncharge(&memcg->memsw, nr_pages);
6987 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
6988 mem_cgroup_id_put_many(memcg, nr_pages);
6993 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
6995 long nr_swap_pages = get_nr_swap_pages();
6997 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6998 return nr_swap_pages;
6999 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7000 nr_swap_pages = min_t(long, nr_swap_pages,
7001 READ_ONCE(memcg->swap.max) -
7002 page_counter_read(&memcg->swap));
7003 return nr_swap_pages;
7006 bool mem_cgroup_swap_full(struct page *page)
7008 struct mem_cgroup *memcg;
7010 VM_BUG_ON_PAGE(!PageLocked(page), page);
7014 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7017 memcg = page->mem_cgroup;
7021 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7022 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.max)
7028 /* for remember boot option*/
7029 #ifdef CONFIG_MEMCG_SWAP_ENABLED
7030 static int really_do_swap_account __initdata = 1;
7032 static int really_do_swap_account __initdata;
7035 static int __init enable_swap_account(char *s)
7037 if (!strcmp(s, "1"))
7038 really_do_swap_account = 1;
7039 else if (!strcmp(s, "0"))
7040 really_do_swap_account = 0;
7043 __setup("swapaccount=", enable_swap_account);
7045 static u64 swap_current_read(struct cgroup_subsys_state *css,
7048 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7050 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7053 static int swap_max_show(struct seq_file *m, void *v)
7055 return seq_puts_memcg_tunable(m,
7056 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7059 static ssize_t swap_max_write(struct kernfs_open_file *of,
7060 char *buf, size_t nbytes, loff_t off)
7062 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7066 buf = strstrip(buf);
7067 err = page_counter_memparse(buf, "max", &max);
7071 xchg(&memcg->swap.max, max);
7076 static int swap_events_show(struct seq_file *m, void *v)
7078 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7080 seq_printf(m, "max %lu\n",
7081 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7082 seq_printf(m, "fail %lu\n",
7083 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7088 static struct cftype swap_files[] = {
7090 .name = "swap.current",
7091 .flags = CFTYPE_NOT_ON_ROOT,
7092 .read_u64 = swap_current_read,
7096 .flags = CFTYPE_NOT_ON_ROOT,
7097 .seq_show = swap_max_show,
7098 .write = swap_max_write,
7101 .name = "swap.events",
7102 .flags = CFTYPE_NOT_ON_ROOT,
7103 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7104 .seq_show = swap_events_show,
7109 static struct cftype memsw_cgroup_files[] = {
7111 .name = "memsw.usage_in_bytes",
7112 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7113 .read_u64 = mem_cgroup_read_u64,
7116 .name = "memsw.max_usage_in_bytes",
7117 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7118 .write = mem_cgroup_reset,
7119 .read_u64 = mem_cgroup_read_u64,
7122 .name = "memsw.limit_in_bytes",
7123 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7124 .write = mem_cgroup_write,
7125 .read_u64 = mem_cgroup_read_u64,
7128 .name = "memsw.failcnt",
7129 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7130 .write = mem_cgroup_reset,
7131 .read_u64 = mem_cgroup_read_u64,
7133 { }, /* terminate */
7136 static int __init mem_cgroup_swap_init(void)
7138 if (!mem_cgroup_disabled() && really_do_swap_account) {
7139 do_swap_account = 1;
7140 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
7142 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
7143 memsw_cgroup_files));
7147 subsys_initcall(mem_cgroup_swap_init);
7149 #endif /* CONFIG_MEMCG_SWAP */