1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
18 * Charge lifetime sanitation
19 * Lockless page tracking & accounting
20 * Unified hierarchy configuration model
21 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23 * This program is free software; you can redistribute it and/or modify
24 * it under the terms of the GNU General Public License as published by
25 * the Free Software Foundation; either version 2 of the License, or
26 * (at your option) any later version.
28 * This program is distributed in the hope that it will be useful,
29 * but WITHOUT ANY WARRANTY; without even the implied warranty of
30 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
31 * GNU General Public License for more details.
34 #include <linux/page_counter.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cgroup.h>
38 #include <linux/sched/mm.h>
39 #include <linux/shmem_fs.h>
40 #include <linux/hugetlb.h>
41 #include <linux/pagemap.h>
42 #include <linux/vm_event_item.h>
43 #include <linux/smp.h>
44 #include <linux/page-flags.h>
45 #include <linux/backing-dev.h>
46 #include <linux/bit_spinlock.h>
47 #include <linux/rcupdate.h>
48 #include <linux/limits.h>
49 #include <linux/export.h>
50 #include <linux/mutex.h>
51 #include <linux/rbtree.h>
52 #include <linux/slab.h>
53 #include <linux/swap.h>
54 #include <linux/swapops.h>
55 #include <linux/spinlock.h>
56 #include <linux/eventfd.h>
57 #include <linux/poll.h>
58 #include <linux/sort.h>
60 #include <linux/seq_file.h>
61 #include <linux/vmpressure.h>
62 #include <linux/mm_inline.h>
63 #include <linux/swap_cgroup.h>
64 #include <linux/cpu.h>
65 #include <linux/oom.h>
66 #include <linux/lockdep.h>
67 #include <linux/file.h>
68 #include <linux/tracehook.h>
74 #include <linux/uaccess.h>
76 #include <trace/events/vmscan.h>
78 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
79 EXPORT_SYMBOL(memory_cgrp_subsys);
81 struct mem_cgroup *root_mem_cgroup __read_mostly;
83 #define MEM_CGROUP_RECLAIM_RETRIES 5
85 /* Socket memory accounting disabled? */
86 static bool cgroup_memory_nosocket;
88 /* Kernel memory accounting disabled? */
89 static bool cgroup_memory_nokmem;
91 /* Whether the swap controller is active */
92 #ifdef CONFIG_MEMCG_SWAP
93 int do_swap_account __read_mostly;
95 #define do_swap_account 0
98 /* Whether legacy memory+swap accounting is active */
99 static bool do_memsw_account(void)
101 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && do_swap_account;
104 static const char *const mem_cgroup_lru_names[] = {
112 #define THRESHOLDS_EVENTS_TARGET 128
113 #define SOFTLIMIT_EVENTS_TARGET 1024
114 #define NUMAINFO_EVENTS_TARGET 1024
117 * Cgroups above their limits are maintained in a RB-Tree, independent of
118 * their hierarchy representation
121 struct mem_cgroup_tree_per_node {
122 struct rb_root rb_root;
123 struct rb_node *rb_rightmost;
127 struct mem_cgroup_tree {
128 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
131 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
134 struct mem_cgroup_eventfd_list {
135 struct list_head list;
136 struct eventfd_ctx *eventfd;
140 * cgroup_event represents events which userspace want to receive.
142 struct mem_cgroup_event {
144 * memcg which the event belongs to.
146 struct mem_cgroup *memcg;
148 * eventfd to signal userspace about the event.
150 struct eventfd_ctx *eventfd;
152 * Each of these stored in a list by the cgroup.
154 struct list_head list;
156 * register_event() callback will be used to add new userspace
157 * waiter for changes related to this event. Use eventfd_signal()
158 * on eventfd to send notification to userspace.
160 int (*register_event)(struct mem_cgroup *memcg,
161 struct eventfd_ctx *eventfd, const char *args);
163 * unregister_event() callback will be called when userspace closes
164 * the eventfd or on cgroup removing. This callback must be set,
165 * if you want provide notification functionality.
167 void (*unregister_event)(struct mem_cgroup *memcg,
168 struct eventfd_ctx *eventfd);
170 * All fields below needed to unregister event when
171 * userspace closes eventfd.
174 wait_queue_head_t *wqh;
175 wait_queue_entry_t wait;
176 struct work_struct remove;
179 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
180 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
182 /* Stuffs for move charges at task migration. */
184 * Types of charges to be moved.
186 #define MOVE_ANON 0x1U
187 #define MOVE_FILE 0x2U
188 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
190 /* "mc" and its members are protected by cgroup_mutex */
191 static struct move_charge_struct {
192 spinlock_t lock; /* for from, to */
193 struct mm_struct *mm;
194 struct mem_cgroup *from;
195 struct mem_cgroup *to;
197 unsigned long precharge;
198 unsigned long moved_charge;
199 unsigned long moved_swap;
200 struct task_struct *moving_task; /* a task moving charges */
201 wait_queue_head_t waitq; /* a waitq for other context */
203 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
204 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
208 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
209 * limit reclaim to prevent infinite loops, if they ever occur.
211 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
212 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
215 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
216 MEM_CGROUP_CHARGE_TYPE_ANON,
217 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
218 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
222 /* for encoding cft->private value on file */
231 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
232 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
233 #define MEMFILE_ATTR(val) ((val) & 0xffff)
234 /* Used for OOM nofiier */
235 #define OOM_CONTROL (0)
238 * Iteration constructs for visiting all cgroups (under a tree). If
239 * loops are exited prematurely (break), mem_cgroup_iter_break() must
240 * be used for reference counting.
242 #define for_each_mem_cgroup_tree(iter, root) \
243 for (iter = mem_cgroup_iter(root, NULL, NULL); \
245 iter = mem_cgroup_iter(root, iter, NULL))
247 #define for_each_mem_cgroup(iter) \
248 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
250 iter = mem_cgroup_iter(NULL, iter, NULL))
252 static inline bool should_force_charge(void)
254 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
255 (current->flags & PF_EXITING);
258 /* Some nice accessors for the vmpressure. */
259 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
262 memcg = root_mem_cgroup;
263 return &memcg->vmpressure;
266 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
268 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
271 #ifdef CONFIG_MEMCG_KMEM
273 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
274 * The main reason for not using cgroup id for this:
275 * this works better in sparse environments, where we have a lot of memcgs,
276 * but only a few kmem-limited. Or also, if we have, for instance, 200
277 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
278 * 200 entry array for that.
280 * The current size of the caches array is stored in memcg_nr_cache_ids. It
281 * will double each time we have to increase it.
283 static DEFINE_IDA(memcg_cache_ida);
284 int memcg_nr_cache_ids;
286 /* Protects memcg_nr_cache_ids */
287 static DECLARE_RWSEM(memcg_cache_ids_sem);
289 void memcg_get_cache_ids(void)
291 down_read(&memcg_cache_ids_sem);
294 void memcg_put_cache_ids(void)
296 up_read(&memcg_cache_ids_sem);
300 * MIN_SIZE is different than 1, because we would like to avoid going through
301 * the alloc/free process all the time. In a small machine, 4 kmem-limited
302 * cgroups is a reasonable guess. In the future, it could be a parameter or
303 * tunable, but that is strictly not necessary.
305 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
306 * this constant directly from cgroup, but it is understandable that this is
307 * better kept as an internal representation in cgroup.c. In any case, the
308 * cgrp_id space is not getting any smaller, and we don't have to necessarily
309 * increase ours as well if it increases.
311 #define MEMCG_CACHES_MIN_SIZE 4
312 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
315 * A lot of the calls to the cache allocation functions are expected to be
316 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
317 * conditional to this static branch, we'll have to allow modules that does
318 * kmem_cache_alloc and the such to see this symbol as well
320 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
321 EXPORT_SYMBOL(memcg_kmem_enabled_key);
323 struct workqueue_struct *memcg_kmem_cache_wq;
325 static int memcg_shrinker_map_size;
326 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
328 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
330 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
333 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
334 int size, int old_size)
336 struct memcg_shrinker_map *new, *old;
339 lockdep_assert_held(&memcg_shrinker_map_mutex);
342 old = rcu_dereference_protected(
343 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
344 /* Not yet online memcg */
348 new = kvmalloc(sizeof(*new) + size, GFP_KERNEL);
352 /* Set all old bits, clear all new bits */
353 memset(new->map, (int)0xff, old_size);
354 memset((void *)new->map + old_size, 0, size - old_size);
356 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
357 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
363 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
365 struct mem_cgroup_per_node *pn;
366 struct memcg_shrinker_map *map;
369 if (mem_cgroup_is_root(memcg))
373 pn = mem_cgroup_nodeinfo(memcg, nid);
374 map = rcu_dereference_protected(pn->shrinker_map, true);
377 rcu_assign_pointer(pn->shrinker_map, NULL);
381 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
383 struct memcg_shrinker_map *map;
384 int nid, size, ret = 0;
386 if (mem_cgroup_is_root(memcg))
389 mutex_lock(&memcg_shrinker_map_mutex);
390 size = memcg_shrinker_map_size;
392 map = kvzalloc(sizeof(*map) + size, GFP_KERNEL);
394 memcg_free_shrinker_maps(memcg);
398 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
400 mutex_unlock(&memcg_shrinker_map_mutex);
405 int memcg_expand_shrinker_maps(int new_id)
407 int size, old_size, ret = 0;
408 struct mem_cgroup *memcg;
410 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
411 old_size = memcg_shrinker_map_size;
412 if (size <= old_size)
415 mutex_lock(&memcg_shrinker_map_mutex);
416 if (!root_mem_cgroup)
419 for_each_mem_cgroup(memcg) {
420 if (mem_cgroup_is_root(memcg))
422 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
428 memcg_shrinker_map_size = size;
429 mutex_unlock(&memcg_shrinker_map_mutex);
433 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
435 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
436 struct memcg_shrinker_map *map;
439 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
440 /* Pairs with smp mb in shrink_slab() */
441 smp_mb__before_atomic();
442 set_bit(shrinker_id, map->map);
447 #else /* CONFIG_MEMCG_KMEM */
448 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
452 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg) { }
453 #endif /* CONFIG_MEMCG_KMEM */
456 * mem_cgroup_css_from_page - css of the memcg associated with a page
457 * @page: page of interest
459 * If memcg is bound to the default hierarchy, css of the memcg associated
460 * with @page is returned. The returned css remains associated with @page
461 * until it is released.
463 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
466 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
468 struct mem_cgroup *memcg;
470 memcg = page->mem_cgroup;
472 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
473 memcg = root_mem_cgroup;
479 * page_cgroup_ino - return inode number of the memcg a page is charged to
482 * Look up the closest online ancestor of the memory cgroup @page is charged to
483 * and return its inode number or 0 if @page is not charged to any cgroup. It
484 * is safe to call this function without holding a reference to @page.
486 * Note, this function is inherently racy, because there is nothing to prevent
487 * the cgroup inode from getting torn down and potentially reallocated a moment
488 * after page_cgroup_ino() returns, so it only should be used by callers that
489 * do not care (such as procfs interfaces).
491 ino_t page_cgroup_ino(struct page *page)
493 struct mem_cgroup *memcg;
494 unsigned long ino = 0;
497 memcg = READ_ONCE(page->mem_cgroup);
498 while (memcg && !(memcg->css.flags & CSS_ONLINE))
499 memcg = parent_mem_cgroup(memcg);
501 ino = cgroup_ino(memcg->css.cgroup);
506 static struct mem_cgroup_per_node *
507 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
509 int nid = page_to_nid(page);
511 return memcg->nodeinfo[nid];
514 static struct mem_cgroup_tree_per_node *
515 soft_limit_tree_node(int nid)
517 return soft_limit_tree.rb_tree_per_node[nid];
520 static struct mem_cgroup_tree_per_node *
521 soft_limit_tree_from_page(struct page *page)
523 int nid = page_to_nid(page);
525 return soft_limit_tree.rb_tree_per_node[nid];
528 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
529 struct mem_cgroup_tree_per_node *mctz,
530 unsigned long new_usage_in_excess)
532 struct rb_node **p = &mctz->rb_root.rb_node;
533 struct rb_node *parent = NULL;
534 struct mem_cgroup_per_node *mz_node;
535 bool rightmost = true;
540 mz->usage_in_excess = new_usage_in_excess;
541 if (!mz->usage_in_excess)
545 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
547 if (mz->usage_in_excess < mz_node->usage_in_excess) {
553 * We can't avoid mem cgroups that are over their soft
554 * limit by the same amount
556 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
561 mctz->rb_rightmost = &mz->tree_node;
563 rb_link_node(&mz->tree_node, parent, p);
564 rb_insert_color(&mz->tree_node, &mctz->rb_root);
568 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
569 struct mem_cgroup_tree_per_node *mctz)
574 if (&mz->tree_node == mctz->rb_rightmost)
575 mctz->rb_rightmost = rb_prev(&mz->tree_node);
577 rb_erase(&mz->tree_node, &mctz->rb_root);
581 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
582 struct mem_cgroup_tree_per_node *mctz)
586 spin_lock_irqsave(&mctz->lock, flags);
587 __mem_cgroup_remove_exceeded(mz, mctz);
588 spin_unlock_irqrestore(&mctz->lock, flags);
591 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
593 unsigned long nr_pages = page_counter_read(&memcg->memory);
594 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
595 unsigned long excess = 0;
597 if (nr_pages > soft_limit)
598 excess = nr_pages - soft_limit;
603 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
605 unsigned long excess;
606 struct mem_cgroup_per_node *mz;
607 struct mem_cgroup_tree_per_node *mctz;
609 mctz = soft_limit_tree_from_page(page);
613 * Necessary to update all ancestors when hierarchy is used.
614 * because their event counter is not touched.
616 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
617 mz = mem_cgroup_page_nodeinfo(memcg, page);
618 excess = soft_limit_excess(memcg);
620 * We have to update the tree if mz is on RB-tree or
621 * mem is over its softlimit.
623 if (excess || mz->on_tree) {
626 spin_lock_irqsave(&mctz->lock, flags);
627 /* if on-tree, remove it */
629 __mem_cgroup_remove_exceeded(mz, mctz);
631 * Insert again. mz->usage_in_excess will be updated.
632 * If excess is 0, no tree ops.
634 __mem_cgroup_insert_exceeded(mz, mctz, excess);
635 spin_unlock_irqrestore(&mctz->lock, flags);
640 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
642 struct mem_cgroup_tree_per_node *mctz;
643 struct mem_cgroup_per_node *mz;
647 mz = mem_cgroup_nodeinfo(memcg, nid);
648 mctz = soft_limit_tree_node(nid);
650 mem_cgroup_remove_exceeded(mz, mctz);
654 static struct mem_cgroup_per_node *
655 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
657 struct mem_cgroup_per_node *mz;
661 if (!mctz->rb_rightmost)
662 goto done; /* Nothing to reclaim from */
664 mz = rb_entry(mctz->rb_rightmost,
665 struct mem_cgroup_per_node, tree_node);
667 * Remove the node now but someone else can add it back,
668 * we will to add it back at the end of reclaim to its correct
669 * position in the tree.
671 __mem_cgroup_remove_exceeded(mz, mctz);
672 if (!soft_limit_excess(mz->memcg) ||
673 !css_tryget_online(&mz->memcg->css))
679 static struct mem_cgroup_per_node *
680 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
682 struct mem_cgroup_per_node *mz;
684 spin_lock_irq(&mctz->lock);
685 mz = __mem_cgroup_largest_soft_limit_node(mctz);
686 spin_unlock_irq(&mctz->lock);
691 * __mod_memcg_state - update cgroup memory statistics
692 * @memcg: the memory cgroup
693 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
694 * @val: delta to add to the counter, can be negative
696 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
700 if (mem_cgroup_disabled())
703 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
704 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
705 atomic_long_add(x, &memcg->vmstats[idx]);
708 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
712 * __mod_lruvec_state - update lruvec memory statistics
713 * @lruvec: the lruvec
714 * @idx: the stat item
715 * @val: delta to add to the counter, can be negative
717 * The lruvec is the intersection of the NUMA node and a cgroup. This
718 * function updates the all three counters that are affected by a
719 * change of state at this level: per-node, per-cgroup, per-lruvec.
721 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
724 struct mem_cgroup_per_node *pn;
728 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
730 if (mem_cgroup_disabled())
733 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
736 __mod_memcg_state(pn->memcg, idx, val);
739 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
740 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
741 atomic_long_add(x, &pn->lruvec_stat[idx]);
744 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
748 * __count_memcg_events - account VM events in a cgroup
749 * @memcg: the memory cgroup
750 * @idx: the event item
751 * @count: the number of events that occured
753 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
758 if (mem_cgroup_disabled())
761 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
762 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
763 atomic_long_add(x, &memcg->vmevents[idx]);
766 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
769 static unsigned long memcg_events_local(struct mem_cgroup *memcg,
772 return atomic_long_read(&memcg->vmevents[event]);
775 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
777 bool compound, int nr_pages)
780 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
781 * counted as CACHE even if it's on ANON LRU.
784 __mod_memcg_state(memcg, MEMCG_RSS, nr_pages);
786 __mod_memcg_state(memcg, MEMCG_CACHE, nr_pages);
787 if (PageSwapBacked(page))
788 __mod_memcg_state(memcg, NR_SHMEM, nr_pages);
792 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
793 __mod_memcg_state(memcg, MEMCG_RSS_HUGE, nr_pages);
796 /* pagein of a big page is an event. So, ignore page size */
798 __count_memcg_events(memcg, PGPGIN, 1);
800 __count_memcg_events(memcg, PGPGOUT, 1);
801 nr_pages = -nr_pages; /* for event */
804 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
807 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
808 enum mem_cgroup_events_target target)
810 unsigned long val, next;
812 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
813 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
814 /* from time_after() in jiffies.h */
815 if ((long)(next - val) < 0) {
817 case MEM_CGROUP_TARGET_THRESH:
818 next = val + THRESHOLDS_EVENTS_TARGET;
820 case MEM_CGROUP_TARGET_SOFTLIMIT:
821 next = val + SOFTLIMIT_EVENTS_TARGET;
823 case MEM_CGROUP_TARGET_NUMAINFO:
824 next = val + NUMAINFO_EVENTS_TARGET;
829 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
836 * Check events in order.
839 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
841 /* threshold event is triggered in finer grain than soft limit */
842 if (unlikely(mem_cgroup_event_ratelimit(memcg,
843 MEM_CGROUP_TARGET_THRESH))) {
845 bool do_numainfo __maybe_unused;
847 do_softlimit = mem_cgroup_event_ratelimit(memcg,
848 MEM_CGROUP_TARGET_SOFTLIMIT);
850 do_numainfo = mem_cgroup_event_ratelimit(memcg,
851 MEM_CGROUP_TARGET_NUMAINFO);
853 mem_cgroup_threshold(memcg);
854 if (unlikely(do_softlimit))
855 mem_cgroup_update_tree(memcg, page);
857 if (unlikely(do_numainfo))
858 atomic_inc(&memcg->numainfo_events);
863 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
866 * mm_update_next_owner() may clear mm->owner to NULL
867 * if it races with swapoff, page migration, etc.
868 * So this can be called with p == NULL.
873 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
875 EXPORT_SYMBOL(mem_cgroup_from_task);
878 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
879 * @mm: mm from which memcg should be extracted. It can be NULL.
881 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
882 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
885 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
887 struct mem_cgroup *memcg;
889 if (mem_cgroup_disabled())
895 * Page cache insertions can happen withou an
896 * actual mm context, e.g. during disk probing
897 * on boot, loopback IO, acct() writes etc.
900 memcg = root_mem_cgroup;
902 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
903 if (unlikely(!memcg))
904 memcg = root_mem_cgroup;
906 } while (!css_tryget_online(&memcg->css));
910 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
913 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
914 * @page: page from which memcg should be extracted.
916 * Obtain a reference on page->memcg and returns it if successful. Otherwise
917 * root_mem_cgroup is returned.
919 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
921 struct mem_cgroup *memcg = page->mem_cgroup;
923 if (mem_cgroup_disabled())
927 if (!memcg || !css_tryget_online(&memcg->css))
928 memcg = root_mem_cgroup;
932 EXPORT_SYMBOL(get_mem_cgroup_from_page);
935 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
937 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
939 if (unlikely(current->active_memcg)) {
940 struct mem_cgroup *memcg = root_mem_cgroup;
943 if (css_tryget_online(¤t->active_memcg->css))
944 memcg = current->active_memcg;
948 return get_mem_cgroup_from_mm(current->mm);
952 * mem_cgroup_iter - iterate over memory cgroup hierarchy
953 * @root: hierarchy root
954 * @prev: previously returned memcg, NULL on first invocation
955 * @reclaim: cookie for shared reclaim walks, NULL for full walks
957 * Returns references to children of the hierarchy below @root, or
958 * @root itself, or %NULL after a full round-trip.
960 * Caller must pass the return value in @prev on subsequent
961 * invocations for reference counting, or use mem_cgroup_iter_break()
962 * to cancel a hierarchy walk before the round-trip is complete.
964 * Reclaimers can specify a node and a priority level in @reclaim to
965 * divide up the memcgs in the hierarchy among all concurrent
966 * reclaimers operating on the same node and priority.
968 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
969 struct mem_cgroup *prev,
970 struct mem_cgroup_reclaim_cookie *reclaim)
972 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
973 struct cgroup_subsys_state *css = NULL;
974 struct mem_cgroup *memcg = NULL;
975 struct mem_cgroup *pos = NULL;
977 if (mem_cgroup_disabled())
981 root = root_mem_cgroup;
983 if (prev && !reclaim)
986 if (!root->use_hierarchy && root != root_mem_cgroup) {
995 struct mem_cgroup_per_node *mz;
997 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
998 iter = &mz->iter[reclaim->priority];
1000 if (prev && reclaim->generation != iter->generation)
1004 pos = READ_ONCE(iter->position);
1005 if (!pos || css_tryget(&pos->css))
1008 * css reference reached zero, so iter->position will
1009 * be cleared by ->css_released. However, we should not
1010 * rely on this happening soon, because ->css_released
1011 * is called from a work queue, and by busy-waiting we
1012 * might block it. So we clear iter->position right
1015 (void)cmpxchg(&iter->position, pos, NULL);
1023 css = css_next_descendant_pre(css, &root->css);
1026 * Reclaimers share the hierarchy walk, and a
1027 * new one might jump in right at the end of
1028 * the hierarchy - make sure they see at least
1029 * one group and restart from the beginning.
1037 * Verify the css and acquire a reference. The root
1038 * is provided by the caller, so we know it's alive
1039 * and kicking, and don't take an extra reference.
1041 memcg = mem_cgroup_from_css(css);
1043 if (css == &root->css)
1046 if (css_tryget(css))
1054 * The position could have already been updated by a competing
1055 * thread, so check that the value hasn't changed since we read
1056 * it to avoid reclaiming from the same cgroup twice.
1058 (void)cmpxchg(&iter->position, pos, memcg);
1066 reclaim->generation = iter->generation;
1072 if (prev && prev != root)
1073 css_put(&prev->css);
1079 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1080 * @root: hierarchy root
1081 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1083 void mem_cgroup_iter_break(struct mem_cgroup *root,
1084 struct mem_cgroup *prev)
1087 root = root_mem_cgroup;
1088 if (prev && prev != root)
1089 css_put(&prev->css);
1092 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1094 struct mem_cgroup *memcg = dead_memcg;
1095 struct mem_cgroup_reclaim_iter *iter;
1096 struct mem_cgroup_per_node *mz;
1100 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1101 for_each_node(nid) {
1102 mz = mem_cgroup_nodeinfo(memcg, nid);
1103 for (i = 0; i <= DEF_PRIORITY; i++) {
1104 iter = &mz->iter[i];
1105 cmpxchg(&iter->position,
1113 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1114 * @memcg: hierarchy root
1115 * @fn: function to call for each task
1116 * @arg: argument passed to @fn
1118 * This function iterates over tasks attached to @memcg or to any of its
1119 * descendants and calls @fn for each task. If @fn returns a non-zero
1120 * value, the function breaks the iteration loop and returns the value.
1121 * Otherwise, it will iterate over all tasks and return 0.
1123 * This function must not be called for the root memory cgroup.
1125 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1126 int (*fn)(struct task_struct *, void *), void *arg)
1128 struct mem_cgroup *iter;
1131 BUG_ON(memcg == root_mem_cgroup);
1133 for_each_mem_cgroup_tree(iter, memcg) {
1134 struct css_task_iter it;
1135 struct task_struct *task;
1137 css_task_iter_start(&iter->css, 0, &it);
1138 while (!ret && (task = css_task_iter_next(&it)))
1139 ret = fn(task, arg);
1140 css_task_iter_end(&it);
1142 mem_cgroup_iter_break(memcg, iter);
1150 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1152 * @pgdat: pgdat of the page
1154 * This function is only safe when following the LRU page isolation
1155 * and putback protocol: the LRU lock must be held, and the page must
1156 * either be PageLRU() or the caller must have isolated/allocated it.
1158 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1160 struct mem_cgroup_per_node *mz;
1161 struct mem_cgroup *memcg;
1162 struct lruvec *lruvec;
1164 if (mem_cgroup_disabled()) {
1165 lruvec = &pgdat->lruvec;
1169 memcg = page->mem_cgroup;
1171 * Swapcache readahead pages are added to the LRU - and
1172 * possibly migrated - before they are charged.
1175 memcg = root_mem_cgroup;
1177 mz = mem_cgroup_page_nodeinfo(memcg, page);
1178 lruvec = &mz->lruvec;
1181 * Since a node can be onlined after the mem_cgroup was created,
1182 * we have to be prepared to initialize lruvec->zone here;
1183 * and if offlined then reonlined, we need to reinitialize it.
1185 if (unlikely(lruvec->pgdat != pgdat))
1186 lruvec->pgdat = pgdat;
1191 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1192 * @lruvec: mem_cgroup per zone lru vector
1193 * @lru: index of lru list the page is sitting on
1194 * @zid: zone id of the accounted pages
1195 * @nr_pages: positive when adding or negative when removing
1197 * This function must be called under lru_lock, just before a page is added
1198 * to or just after a page is removed from an lru list (that ordering being
1199 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1201 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1202 int zid, int nr_pages)
1204 struct mem_cgroup_per_node *mz;
1205 unsigned long *lru_size;
1208 if (mem_cgroup_disabled())
1211 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1212 lru_size = &mz->lru_zone_size[zid][lru];
1215 *lru_size += nr_pages;
1218 if (WARN_ONCE(size < 0,
1219 "%s(%p, %d, %d): lru_size %ld\n",
1220 __func__, lruvec, lru, nr_pages, size)) {
1226 *lru_size += nr_pages;
1229 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1231 struct mem_cgroup *task_memcg;
1232 struct task_struct *p;
1235 p = find_lock_task_mm(task);
1237 task_memcg = get_mem_cgroup_from_mm(p->mm);
1241 * All threads may have already detached their mm's, but the oom
1242 * killer still needs to detect if they have already been oom
1243 * killed to prevent needlessly killing additional tasks.
1246 task_memcg = mem_cgroup_from_task(task);
1247 css_get(&task_memcg->css);
1250 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1251 css_put(&task_memcg->css);
1256 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1257 * @memcg: the memory cgroup
1259 * Returns the maximum amount of memory @mem can be charged with, in
1262 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1264 unsigned long margin = 0;
1265 unsigned long count;
1266 unsigned long limit;
1268 count = page_counter_read(&memcg->memory);
1269 limit = READ_ONCE(memcg->memory.max);
1271 margin = limit - count;
1273 if (do_memsw_account()) {
1274 count = page_counter_read(&memcg->memsw);
1275 limit = READ_ONCE(memcg->memsw.max);
1277 margin = min(margin, limit - count);
1286 * A routine for checking "mem" is under move_account() or not.
1288 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1289 * moving cgroups. This is for waiting at high-memory pressure
1292 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1294 struct mem_cgroup *from;
1295 struct mem_cgroup *to;
1298 * Unlike task_move routines, we access mc.to, mc.from not under
1299 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1301 spin_lock(&mc.lock);
1307 ret = mem_cgroup_is_descendant(from, memcg) ||
1308 mem_cgroup_is_descendant(to, memcg);
1310 spin_unlock(&mc.lock);
1314 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1316 if (mc.moving_task && current != mc.moving_task) {
1317 if (mem_cgroup_under_move(memcg)) {
1319 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1320 /* moving charge context might have finished. */
1323 finish_wait(&mc.waitq, &wait);
1330 static const unsigned int memcg1_stats[] = {
1341 static const char *const memcg1_stat_names[] = {
1352 #define K(x) ((x) << (PAGE_SHIFT-10))
1354 * mem_cgroup_print_oom_context: Print OOM information relevant to
1355 * memory controller.
1356 * @memcg: The memory cgroup that went over limit
1357 * @p: Task that is going to be killed
1359 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1362 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1367 pr_cont(",oom_memcg=");
1368 pr_cont_cgroup_path(memcg->css.cgroup);
1370 pr_cont(",global_oom");
1372 pr_cont(",task_memcg=");
1373 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1379 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1380 * memory controller.
1381 * @memcg: The memory cgroup that went over limit
1383 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1385 struct mem_cgroup *iter;
1388 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1389 K((u64)page_counter_read(&memcg->memory)),
1390 K((u64)memcg->memory.max), memcg->memory.failcnt);
1391 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1392 K((u64)page_counter_read(&memcg->memsw)),
1393 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1394 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1395 K((u64)page_counter_read(&memcg->kmem)),
1396 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1398 for_each_mem_cgroup_tree(iter, memcg) {
1399 pr_info("Memory cgroup stats for ");
1400 pr_cont_cgroup_path(iter->css.cgroup);
1403 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
1404 if (memcg1_stats[i] == MEMCG_SWAP && !do_swap_account)
1406 pr_cont(" %s:%luKB", memcg1_stat_names[i],
1407 K(memcg_page_state_local(iter,
1411 for (i = 0; i < NR_LRU_LISTS; i++)
1412 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1413 K(memcg_page_state_local(iter,
1421 * Return the memory (and swap, if configured) limit for a memcg.
1423 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1427 max = memcg->memory.max;
1428 if (mem_cgroup_swappiness(memcg)) {
1429 unsigned long memsw_max;
1430 unsigned long swap_max;
1432 memsw_max = memcg->memsw.max;
1433 swap_max = memcg->swap.max;
1434 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1435 max = min(max + swap_max, memsw_max);
1440 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1443 struct oom_control oc = {
1447 .gfp_mask = gfp_mask,
1452 if (mutex_lock_killable(&oom_lock))
1455 * A few threads which were not waiting at mutex_lock_killable() can
1456 * fail to bail out. Therefore, check again after holding oom_lock.
1458 ret = should_force_charge() || out_of_memory(&oc);
1459 mutex_unlock(&oom_lock);
1463 #if MAX_NUMNODES > 1
1466 * test_mem_cgroup_node_reclaimable
1467 * @memcg: the target memcg
1468 * @nid: the node ID to be checked.
1469 * @noswap : specify true here if the user wants flle only information.
1471 * This function returns whether the specified memcg contains any
1472 * reclaimable pages on a node. Returns true if there are any reclaimable
1473 * pages in the node.
1475 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1476 int nid, bool noswap)
1478 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
1480 if (lruvec_page_state_local(lruvec, NR_INACTIVE_FILE) ||
1481 lruvec_page_state_local(lruvec, NR_ACTIVE_FILE))
1483 if (noswap || !total_swap_pages)
1485 if (lruvec_page_state_local(lruvec, NR_INACTIVE_ANON) ||
1486 lruvec_page_state_local(lruvec, NR_ACTIVE_ANON))
1493 * Always updating the nodemask is not very good - even if we have an empty
1494 * list or the wrong list here, we can start from some node and traverse all
1495 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1498 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1502 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1503 * pagein/pageout changes since the last update.
1505 if (!atomic_read(&memcg->numainfo_events))
1507 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1510 /* make a nodemask where this memcg uses memory from */
1511 memcg->scan_nodes = node_states[N_MEMORY];
1513 for_each_node_mask(nid, node_states[N_MEMORY]) {
1515 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1516 node_clear(nid, memcg->scan_nodes);
1519 atomic_set(&memcg->numainfo_events, 0);
1520 atomic_set(&memcg->numainfo_updating, 0);
1524 * Selecting a node where we start reclaim from. Because what we need is just
1525 * reducing usage counter, start from anywhere is O,K. Considering
1526 * memory reclaim from current node, there are pros. and cons.
1528 * Freeing memory from current node means freeing memory from a node which
1529 * we'll use or we've used. So, it may make LRU bad. And if several threads
1530 * hit limits, it will see a contention on a node. But freeing from remote
1531 * node means more costs for memory reclaim because of memory latency.
1533 * Now, we use round-robin. Better algorithm is welcomed.
1535 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1539 mem_cgroup_may_update_nodemask(memcg);
1540 node = memcg->last_scanned_node;
1542 node = next_node_in(node, memcg->scan_nodes);
1544 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1545 * last time it really checked all the LRUs due to rate limiting.
1546 * Fallback to the current node in that case for simplicity.
1548 if (unlikely(node == MAX_NUMNODES))
1549 node = numa_node_id();
1551 memcg->last_scanned_node = node;
1555 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1561 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1564 unsigned long *total_scanned)
1566 struct mem_cgroup *victim = NULL;
1569 unsigned long excess;
1570 unsigned long nr_scanned;
1571 struct mem_cgroup_reclaim_cookie reclaim = {
1576 excess = soft_limit_excess(root_memcg);
1579 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1584 * If we have not been able to reclaim
1585 * anything, it might because there are
1586 * no reclaimable pages under this hierarchy
1591 * We want to do more targeted reclaim.
1592 * excess >> 2 is not to excessive so as to
1593 * reclaim too much, nor too less that we keep
1594 * coming back to reclaim from this cgroup
1596 if (total >= (excess >> 2) ||
1597 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1602 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1603 pgdat, &nr_scanned);
1604 *total_scanned += nr_scanned;
1605 if (!soft_limit_excess(root_memcg))
1608 mem_cgroup_iter_break(root_memcg, victim);
1612 #ifdef CONFIG_LOCKDEP
1613 static struct lockdep_map memcg_oom_lock_dep_map = {
1614 .name = "memcg_oom_lock",
1618 static DEFINE_SPINLOCK(memcg_oom_lock);
1621 * Check OOM-Killer is already running under our hierarchy.
1622 * If someone is running, return false.
1624 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1626 struct mem_cgroup *iter, *failed = NULL;
1628 spin_lock(&memcg_oom_lock);
1630 for_each_mem_cgroup_tree(iter, memcg) {
1631 if (iter->oom_lock) {
1633 * this subtree of our hierarchy is already locked
1634 * so we cannot give a lock.
1637 mem_cgroup_iter_break(memcg, iter);
1640 iter->oom_lock = true;
1645 * OK, we failed to lock the whole subtree so we have
1646 * to clean up what we set up to the failing subtree
1648 for_each_mem_cgroup_tree(iter, memcg) {
1649 if (iter == failed) {
1650 mem_cgroup_iter_break(memcg, iter);
1653 iter->oom_lock = false;
1656 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1658 spin_unlock(&memcg_oom_lock);
1663 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1665 struct mem_cgroup *iter;
1667 spin_lock(&memcg_oom_lock);
1668 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1669 for_each_mem_cgroup_tree(iter, memcg)
1670 iter->oom_lock = false;
1671 spin_unlock(&memcg_oom_lock);
1674 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1676 struct mem_cgroup *iter;
1678 spin_lock(&memcg_oom_lock);
1679 for_each_mem_cgroup_tree(iter, memcg)
1681 spin_unlock(&memcg_oom_lock);
1684 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1686 struct mem_cgroup *iter;
1689 * When a new child is created while the hierarchy is under oom,
1690 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1692 spin_lock(&memcg_oom_lock);
1693 for_each_mem_cgroup_tree(iter, memcg)
1694 if (iter->under_oom > 0)
1696 spin_unlock(&memcg_oom_lock);
1699 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1701 struct oom_wait_info {
1702 struct mem_cgroup *memcg;
1703 wait_queue_entry_t wait;
1706 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1707 unsigned mode, int sync, void *arg)
1709 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1710 struct mem_cgroup *oom_wait_memcg;
1711 struct oom_wait_info *oom_wait_info;
1713 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1714 oom_wait_memcg = oom_wait_info->memcg;
1716 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1717 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1719 return autoremove_wake_function(wait, mode, sync, arg);
1722 static void memcg_oom_recover(struct mem_cgroup *memcg)
1725 * For the following lockless ->under_oom test, the only required
1726 * guarantee is that it must see the state asserted by an OOM when
1727 * this function is called as a result of userland actions
1728 * triggered by the notification of the OOM. This is trivially
1729 * achieved by invoking mem_cgroup_mark_under_oom() before
1730 * triggering notification.
1732 if (memcg && memcg->under_oom)
1733 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1743 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1745 enum oom_status ret;
1748 if (order > PAGE_ALLOC_COSTLY_ORDER)
1751 memcg_memory_event(memcg, MEMCG_OOM);
1754 * We are in the middle of the charge context here, so we
1755 * don't want to block when potentially sitting on a callstack
1756 * that holds all kinds of filesystem and mm locks.
1758 * cgroup1 allows disabling the OOM killer and waiting for outside
1759 * handling until the charge can succeed; remember the context and put
1760 * the task to sleep at the end of the page fault when all locks are
1763 * On the other hand, in-kernel OOM killer allows for an async victim
1764 * memory reclaim (oom_reaper) and that means that we are not solely
1765 * relying on the oom victim to make a forward progress and we can
1766 * invoke the oom killer here.
1768 * Please note that mem_cgroup_out_of_memory might fail to find a
1769 * victim and then we have to bail out from the charge path.
1771 if (memcg->oom_kill_disable) {
1772 if (!current->in_user_fault)
1774 css_get(&memcg->css);
1775 current->memcg_in_oom = memcg;
1776 current->memcg_oom_gfp_mask = mask;
1777 current->memcg_oom_order = order;
1782 mem_cgroup_mark_under_oom(memcg);
1784 locked = mem_cgroup_oom_trylock(memcg);
1787 mem_cgroup_oom_notify(memcg);
1789 mem_cgroup_unmark_under_oom(memcg);
1790 if (mem_cgroup_out_of_memory(memcg, mask, order))
1796 mem_cgroup_oom_unlock(memcg);
1802 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1803 * @handle: actually kill/wait or just clean up the OOM state
1805 * This has to be called at the end of a page fault if the memcg OOM
1806 * handler was enabled.
1808 * Memcg supports userspace OOM handling where failed allocations must
1809 * sleep on a waitqueue until the userspace task resolves the
1810 * situation. Sleeping directly in the charge context with all kinds
1811 * of locks held is not a good idea, instead we remember an OOM state
1812 * in the task and mem_cgroup_oom_synchronize() has to be called at
1813 * the end of the page fault to complete the OOM handling.
1815 * Returns %true if an ongoing memcg OOM situation was detected and
1816 * completed, %false otherwise.
1818 bool mem_cgroup_oom_synchronize(bool handle)
1820 struct mem_cgroup *memcg = current->memcg_in_oom;
1821 struct oom_wait_info owait;
1824 /* OOM is global, do not handle */
1831 owait.memcg = memcg;
1832 owait.wait.flags = 0;
1833 owait.wait.func = memcg_oom_wake_function;
1834 owait.wait.private = current;
1835 INIT_LIST_HEAD(&owait.wait.entry);
1837 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1838 mem_cgroup_mark_under_oom(memcg);
1840 locked = mem_cgroup_oom_trylock(memcg);
1843 mem_cgroup_oom_notify(memcg);
1845 if (locked && !memcg->oom_kill_disable) {
1846 mem_cgroup_unmark_under_oom(memcg);
1847 finish_wait(&memcg_oom_waitq, &owait.wait);
1848 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1849 current->memcg_oom_order);
1852 mem_cgroup_unmark_under_oom(memcg);
1853 finish_wait(&memcg_oom_waitq, &owait.wait);
1857 mem_cgroup_oom_unlock(memcg);
1859 * There is no guarantee that an OOM-lock contender
1860 * sees the wakeups triggered by the OOM kill
1861 * uncharges. Wake any sleepers explicitely.
1863 memcg_oom_recover(memcg);
1866 current->memcg_in_oom = NULL;
1867 css_put(&memcg->css);
1872 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1873 * @victim: task to be killed by the OOM killer
1874 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1876 * Returns a pointer to a memory cgroup, which has to be cleaned up
1877 * by killing all belonging OOM-killable tasks.
1879 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1881 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1882 struct mem_cgroup *oom_domain)
1884 struct mem_cgroup *oom_group = NULL;
1885 struct mem_cgroup *memcg;
1887 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
1891 oom_domain = root_mem_cgroup;
1895 memcg = mem_cgroup_from_task(victim);
1896 if (memcg == root_mem_cgroup)
1900 * Traverse the memory cgroup hierarchy from the victim task's
1901 * cgroup up to the OOMing cgroup (or root) to find the
1902 * highest-level memory cgroup with oom.group set.
1904 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1905 if (memcg->oom_group)
1908 if (memcg == oom_domain)
1913 css_get(&oom_group->css);
1920 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
1922 pr_info("Tasks in ");
1923 pr_cont_cgroup_path(memcg->css.cgroup);
1924 pr_cont(" are going to be killed due to memory.oom.group set\n");
1928 * lock_page_memcg - lock a page->mem_cgroup binding
1931 * This function protects unlocked LRU pages from being moved to
1934 * It ensures lifetime of the returned memcg. Caller is responsible
1935 * for the lifetime of the page; __unlock_page_memcg() is available
1936 * when @page might get freed inside the locked section.
1938 struct mem_cgroup *lock_page_memcg(struct page *page)
1940 struct mem_cgroup *memcg;
1941 unsigned long flags;
1944 * The RCU lock is held throughout the transaction. The fast
1945 * path can get away without acquiring the memcg->move_lock
1946 * because page moving starts with an RCU grace period.
1948 * The RCU lock also protects the memcg from being freed when
1949 * the page state that is going to change is the only thing
1950 * preventing the page itself from being freed. E.g. writeback
1951 * doesn't hold a page reference and relies on PG_writeback to
1952 * keep off truncation, migration and so forth.
1956 if (mem_cgroup_disabled())
1959 memcg = page->mem_cgroup;
1960 if (unlikely(!memcg))
1963 if (atomic_read(&memcg->moving_account) <= 0)
1966 spin_lock_irqsave(&memcg->move_lock, flags);
1967 if (memcg != page->mem_cgroup) {
1968 spin_unlock_irqrestore(&memcg->move_lock, flags);
1973 * When charge migration first begins, we can have locked and
1974 * unlocked page stat updates happening concurrently. Track
1975 * the task who has the lock for unlock_page_memcg().
1977 memcg->move_lock_task = current;
1978 memcg->move_lock_flags = flags;
1982 EXPORT_SYMBOL(lock_page_memcg);
1985 * __unlock_page_memcg - unlock and unpin a memcg
1988 * Unlock and unpin a memcg returned by lock_page_memcg().
1990 void __unlock_page_memcg(struct mem_cgroup *memcg)
1992 if (memcg && memcg->move_lock_task == current) {
1993 unsigned long flags = memcg->move_lock_flags;
1995 memcg->move_lock_task = NULL;
1996 memcg->move_lock_flags = 0;
1998 spin_unlock_irqrestore(&memcg->move_lock, flags);
2005 * unlock_page_memcg - unlock a page->mem_cgroup binding
2008 void unlock_page_memcg(struct page *page)
2010 __unlock_page_memcg(page->mem_cgroup);
2012 EXPORT_SYMBOL(unlock_page_memcg);
2014 struct memcg_stock_pcp {
2015 struct mem_cgroup *cached; /* this never be root cgroup */
2016 unsigned int nr_pages;
2017 struct work_struct work;
2018 unsigned long flags;
2019 #define FLUSHING_CACHED_CHARGE 0
2021 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2022 static DEFINE_MUTEX(percpu_charge_mutex);
2025 * consume_stock: Try to consume stocked charge on this cpu.
2026 * @memcg: memcg to consume from.
2027 * @nr_pages: how many pages to charge.
2029 * The charges will only happen if @memcg matches the current cpu's memcg
2030 * stock, and at least @nr_pages are available in that stock. Failure to
2031 * service an allocation will refill the stock.
2033 * returns true if successful, false otherwise.
2035 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2037 struct memcg_stock_pcp *stock;
2038 unsigned long flags;
2041 if (nr_pages > MEMCG_CHARGE_BATCH)
2044 local_irq_save(flags);
2046 stock = this_cpu_ptr(&memcg_stock);
2047 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2048 stock->nr_pages -= nr_pages;
2052 local_irq_restore(flags);
2058 * Returns stocks cached in percpu and reset cached information.
2060 static void drain_stock(struct memcg_stock_pcp *stock)
2062 struct mem_cgroup *old = stock->cached;
2064 if (stock->nr_pages) {
2065 page_counter_uncharge(&old->memory, stock->nr_pages);
2066 if (do_memsw_account())
2067 page_counter_uncharge(&old->memsw, stock->nr_pages);
2068 css_put_many(&old->css, stock->nr_pages);
2069 stock->nr_pages = 0;
2071 stock->cached = NULL;
2074 static void drain_local_stock(struct work_struct *dummy)
2076 struct memcg_stock_pcp *stock;
2077 unsigned long flags;
2080 * The only protection from memory hotplug vs. drain_stock races is
2081 * that we always operate on local CPU stock here with IRQ disabled
2083 local_irq_save(flags);
2085 stock = this_cpu_ptr(&memcg_stock);
2087 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2089 local_irq_restore(flags);
2093 * Cache charges(val) to local per_cpu area.
2094 * This will be consumed by consume_stock() function, later.
2096 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2098 struct memcg_stock_pcp *stock;
2099 unsigned long flags;
2101 local_irq_save(flags);
2103 stock = this_cpu_ptr(&memcg_stock);
2104 if (stock->cached != memcg) { /* reset if necessary */
2106 stock->cached = memcg;
2108 stock->nr_pages += nr_pages;
2110 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2113 local_irq_restore(flags);
2117 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2118 * of the hierarchy under it.
2120 static void drain_all_stock(struct mem_cgroup *root_memcg)
2124 /* If someone's already draining, avoid adding running more workers. */
2125 if (!mutex_trylock(&percpu_charge_mutex))
2128 * Notify other cpus that system-wide "drain" is running
2129 * We do not care about races with the cpu hotplug because cpu down
2130 * as well as workers from this path always operate on the local
2131 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2134 for_each_online_cpu(cpu) {
2135 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2136 struct mem_cgroup *memcg;
2138 memcg = stock->cached;
2139 if (!memcg || !stock->nr_pages || !css_tryget(&memcg->css))
2141 if (!mem_cgroup_is_descendant(memcg, root_memcg)) {
2142 css_put(&memcg->css);
2145 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2147 drain_local_stock(&stock->work);
2149 schedule_work_on(cpu, &stock->work);
2151 css_put(&memcg->css);
2154 mutex_unlock(&percpu_charge_mutex);
2157 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2159 struct memcg_stock_pcp *stock;
2160 struct mem_cgroup *memcg;
2162 stock = &per_cpu(memcg_stock, cpu);
2165 for_each_mem_cgroup(memcg) {
2168 for (i = 0; i < MEMCG_NR_STAT; i++) {
2172 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2174 atomic_long_add(x, &memcg->vmstats[i]);
2176 if (i >= NR_VM_NODE_STAT_ITEMS)
2179 for_each_node(nid) {
2180 struct mem_cgroup_per_node *pn;
2182 pn = mem_cgroup_nodeinfo(memcg, nid);
2183 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2185 atomic_long_add(x, &pn->lruvec_stat[i]);
2189 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2192 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2194 atomic_long_add(x, &memcg->vmevents[i]);
2201 static void reclaim_high(struct mem_cgroup *memcg,
2202 unsigned int nr_pages,
2206 if (page_counter_read(&memcg->memory) <= memcg->high)
2208 memcg_memory_event(memcg, MEMCG_HIGH);
2209 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2210 } while ((memcg = parent_mem_cgroup(memcg)));
2213 static void high_work_func(struct work_struct *work)
2215 struct mem_cgroup *memcg;
2217 memcg = container_of(work, struct mem_cgroup, high_work);
2218 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2222 * Scheduled by try_charge() to be executed from the userland return path
2223 * and reclaims memory over the high limit.
2225 void mem_cgroup_handle_over_high(void)
2227 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2228 struct mem_cgroup *memcg;
2230 if (likely(!nr_pages))
2233 memcg = get_mem_cgroup_from_mm(current->mm);
2234 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2235 css_put(&memcg->css);
2236 current->memcg_nr_pages_over_high = 0;
2239 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2240 unsigned int nr_pages)
2242 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2243 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2244 struct mem_cgroup *mem_over_limit;
2245 struct page_counter *counter;
2246 unsigned long nr_reclaimed;
2247 bool may_swap = true;
2248 bool drained = false;
2250 enum oom_status oom_status;
2252 if (mem_cgroup_is_root(memcg))
2255 if (consume_stock(memcg, nr_pages))
2258 if (!do_memsw_account() ||
2259 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2260 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2262 if (do_memsw_account())
2263 page_counter_uncharge(&memcg->memsw, batch);
2264 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2266 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2270 if (batch > nr_pages) {
2276 * Unlike in global OOM situations, memcg is not in a physical
2277 * memory shortage. Allow dying and OOM-killed tasks to
2278 * bypass the last charges so that they can exit quickly and
2279 * free their memory.
2281 if (unlikely(should_force_charge()))
2285 * Prevent unbounded recursion when reclaim operations need to
2286 * allocate memory. This might exceed the limits temporarily,
2287 * but we prefer facilitating memory reclaim and getting back
2288 * under the limit over triggering OOM kills in these cases.
2290 if (unlikely(current->flags & PF_MEMALLOC))
2293 if (unlikely(task_in_memcg_oom(current)))
2296 if (!gfpflags_allow_blocking(gfp_mask))
2299 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2301 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2302 gfp_mask, may_swap);
2304 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2308 drain_all_stock(mem_over_limit);
2313 if (gfp_mask & __GFP_NORETRY)
2316 * Even though the limit is exceeded at this point, reclaim
2317 * may have been able to free some pages. Retry the charge
2318 * before killing the task.
2320 * Only for regular pages, though: huge pages are rather
2321 * unlikely to succeed so close to the limit, and we fall back
2322 * to regular pages anyway in case of failure.
2324 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2327 * At task move, charge accounts can be doubly counted. So, it's
2328 * better to wait until the end of task_move if something is going on.
2330 if (mem_cgroup_wait_acct_move(mem_over_limit))
2336 if (gfp_mask & __GFP_RETRY_MAYFAIL && oomed)
2339 if (gfp_mask & __GFP_NOFAIL)
2342 if (fatal_signal_pending(current))
2346 * keep retrying as long as the memcg oom killer is able to make
2347 * a forward progress or bypass the charge if the oom killer
2348 * couldn't make any progress.
2350 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2351 get_order(nr_pages * PAGE_SIZE));
2352 switch (oom_status) {
2354 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2363 if (!(gfp_mask & __GFP_NOFAIL))
2367 * The allocation either can't fail or will lead to more memory
2368 * being freed very soon. Allow memory usage go over the limit
2369 * temporarily by force charging it.
2371 page_counter_charge(&memcg->memory, nr_pages);
2372 if (do_memsw_account())
2373 page_counter_charge(&memcg->memsw, nr_pages);
2374 css_get_many(&memcg->css, nr_pages);
2379 css_get_many(&memcg->css, batch);
2380 if (batch > nr_pages)
2381 refill_stock(memcg, batch - nr_pages);
2384 * If the hierarchy is above the normal consumption range, schedule
2385 * reclaim on returning to userland. We can perform reclaim here
2386 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2387 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2388 * not recorded as it most likely matches current's and won't
2389 * change in the meantime. As high limit is checked again before
2390 * reclaim, the cost of mismatch is negligible.
2393 if (page_counter_read(&memcg->memory) > memcg->high) {
2394 /* Don't bother a random interrupted task */
2395 if (in_interrupt()) {
2396 schedule_work(&memcg->high_work);
2399 current->memcg_nr_pages_over_high += batch;
2400 set_notify_resume(current);
2403 } while ((memcg = parent_mem_cgroup(memcg)));
2408 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2410 if (mem_cgroup_is_root(memcg))
2413 page_counter_uncharge(&memcg->memory, nr_pages);
2414 if (do_memsw_account())
2415 page_counter_uncharge(&memcg->memsw, nr_pages);
2417 css_put_many(&memcg->css, nr_pages);
2420 static void lock_page_lru(struct page *page, int *isolated)
2422 pg_data_t *pgdat = page_pgdat(page);
2424 spin_lock_irq(&pgdat->lru_lock);
2425 if (PageLRU(page)) {
2426 struct lruvec *lruvec;
2428 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2430 del_page_from_lru_list(page, lruvec, page_lru(page));
2436 static void unlock_page_lru(struct page *page, int isolated)
2438 pg_data_t *pgdat = page_pgdat(page);
2441 struct lruvec *lruvec;
2443 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2444 VM_BUG_ON_PAGE(PageLRU(page), page);
2446 add_page_to_lru_list(page, lruvec, page_lru(page));
2448 spin_unlock_irq(&pgdat->lru_lock);
2451 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2456 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2459 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2460 * may already be on some other mem_cgroup's LRU. Take care of it.
2463 lock_page_lru(page, &isolated);
2466 * Nobody should be changing or seriously looking at
2467 * page->mem_cgroup at this point:
2469 * - the page is uncharged
2471 * - the page is off-LRU
2473 * - an anonymous fault has exclusive page access, except for
2474 * a locked page table
2476 * - a page cache insertion, a swapin fault, or a migration
2477 * have the page locked
2479 page->mem_cgroup = memcg;
2482 unlock_page_lru(page, isolated);
2485 #ifdef CONFIG_MEMCG_KMEM
2486 static int memcg_alloc_cache_id(void)
2491 id = ida_simple_get(&memcg_cache_ida,
2492 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2496 if (id < memcg_nr_cache_ids)
2500 * There's no space for the new id in memcg_caches arrays,
2501 * so we have to grow them.
2503 down_write(&memcg_cache_ids_sem);
2505 size = 2 * (id + 1);
2506 if (size < MEMCG_CACHES_MIN_SIZE)
2507 size = MEMCG_CACHES_MIN_SIZE;
2508 else if (size > MEMCG_CACHES_MAX_SIZE)
2509 size = MEMCG_CACHES_MAX_SIZE;
2511 err = memcg_update_all_caches(size);
2513 err = memcg_update_all_list_lrus(size);
2515 memcg_nr_cache_ids = size;
2517 up_write(&memcg_cache_ids_sem);
2520 ida_simple_remove(&memcg_cache_ida, id);
2526 static void memcg_free_cache_id(int id)
2528 ida_simple_remove(&memcg_cache_ida, id);
2531 struct memcg_kmem_cache_create_work {
2532 struct mem_cgroup *memcg;
2533 struct kmem_cache *cachep;
2534 struct work_struct work;
2537 static void memcg_kmem_cache_create_func(struct work_struct *w)
2539 struct memcg_kmem_cache_create_work *cw =
2540 container_of(w, struct memcg_kmem_cache_create_work, work);
2541 struct mem_cgroup *memcg = cw->memcg;
2542 struct kmem_cache *cachep = cw->cachep;
2544 memcg_create_kmem_cache(memcg, cachep);
2546 css_put(&memcg->css);
2551 * Enqueue the creation of a per-memcg kmem_cache.
2553 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2554 struct kmem_cache *cachep)
2556 struct memcg_kmem_cache_create_work *cw;
2558 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2562 css_get(&memcg->css);
2565 cw->cachep = cachep;
2566 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2568 queue_work(memcg_kmem_cache_wq, &cw->work);
2571 static inline bool memcg_kmem_bypass(void)
2573 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2579 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2580 * @cachep: the original global kmem cache
2582 * Return the kmem_cache we're supposed to use for a slab allocation.
2583 * We try to use the current memcg's version of the cache.
2585 * If the cache does not exist yet, if we are the first user of it, we
2586 * create it asynchronously in a workqueue and let the current allocation
2587 * go through with the original cache.
2589 * This function takes a reference to the cache it returns to assure it
2590 * won't get destroyed while we are working with it. Once the caller is
2591 * done with it, memcg_kmem_put_cache() must be called to release the
2594 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2596 struct mem_cgroup *memcg;
2597 struct kmem_cache *memcg_cachep;
2600 VM_BUG_ON(!is_root_cache(cachep));
2602 if (memcg_kmem_bypass())
2605 memcg = get_mem_cgroup_from_current();
2606 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2610 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2611 if (likely(memcg_cachep))
2612 return memcg_cachep;
2615 * If we are in a safe context (can wait, and not in interrupt
2616 * context), we could be be predictable and return right away.
2617 * This would guarantee that the allocation being performed
2618 * already belongs in the new cache.
2620 * However, there are some clashes that can arrive from locking.
2621 * For instance, because we acquire the slab_mutex while doing
2622 * memcg_create_kmem_cache, this means no further allocation
2623 * could happen with the slab_mutex held. So it's better to
2626 memcg_schedule_kmem_cache_create(memcg, cachep);
2628 css_put(&memcg->css);
2633 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2634 * @cachep: the cache returned by memcg_kmem_get_cache
2636 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2638 if (!is_root_cache(cachep))
2639 css_put(&cachep->memcg_params.memcg->css);
2643 * __memcg_kmem_charge_memcg: charge a kmem page
2644 * @page: page to charge
2645 * @gfp: reclaim mode
2646 * @order: allocation order
2647 * @memcg: memory cgroup to charge
2649 * Returns 0 on success, an error code on failure.
2651 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2652 struct mem_cgroup *memcg)
2654 unsigned int nr_pages = 1 << order;
2655 struct page_counter *counter;
2658 ret = try_charge(memcg, gfp, nr_pages);
2662 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2663 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2664 cancel_charge(memcg, nr_pages);
2668 page->mem_cgroup = memcg;
2674 * __memcg_kmem_charge: charge a kmem page to the current memory cgroup
2675 * @page: page to charge
2676 * @gfp: reclaim mode
2677 * @order: allocation order
2679 * Returns 0 on success, an error code on failure.
2681 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2683 struct mem_cgroup *memcg;
2686 if (memcg_kmem_bypass())
2689 memcg = get_mem_cgroup_from_current();
2690 if (!mem_cgroup_is_root(memcg)) {
2691 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2693 __SetPageKmemcg(page);
2695 css_put(&memcg->css);
2699 * __memcg_kmem_uncharge: uncharge a kmem page
2700 * @page: page to uncharge
2701 * @order: allocation order
2703 void __memcg_kmem_uncharge(struct page *page, int order)
2705 struct mem_cgroup *memcg = page->mem_cgroup;
2706 unsigned int nr_pages = 1 << order;
2711 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2713 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2714 page_counter_uncharge(&memcg->kmem, nr_pages);
2716 page_counter_uncharge(&memcg->memory, nr_pages);
2717 if (do_memsw_account())
2718 page_counter_uncharge(&memcg->memsw, nr_pages);
2720 page->mem_cgroup = NULL;
2722 /* slab pages do not have PageKmemcg flag set */
2723 if (PageKmemcg(page))
2724 __ClearPageKmemcg(page);
2726 css_put_many(&memcg->css, nr_pages);
2728 #endif /* CONFIG_MEMCG_KMEM */
2730 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2733 * Because tail pages are not marked as "used", set it. We're under
2734 * pgdat->lru_lock and migration entries setup in all page mappings.
2736 void mem_cgroup_split_huge_fixup(struct page *head)
2740 if (mem_cgroup_disabled())
2743 for (i = 1; i < HPAGE_PMD_NR; i++)
2744 head[i].mem_cgroup = head->mem_cgroup;
2746 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
2748 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2750 #ifdef CONFIG_MEMCG_SWAP
2752 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2753 * @entry: swap entry to be moved
2754 * @from: mem_cgroup which the entry is moved from
2755 * @to: mem_cgroup which the entry is moved to
2757 * It succeeds only when the swap_cgroup's record for this entry is the same
2758 * as the mem_cgroup's id of @from.
2760 * Returns 0 on success, -EINVAL on failure.
2762 * The caller must have charged to @to, IOW, called page_counter_charge() about
2763 * both res and memsw, and called css_get().
2765 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2766 struct mem_cgroup *from, struct mem_cgroup *to)
2768 unsigned short old_id, new_id;
2770 old_id = mem_cgroup_id(from);
2771 new_id = mem_cgroup_id(to);
2773 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2774 mod_memcg_state(from, MEMCG_SWAP, -1);
2775 mod_memcg_state(to, MEMCG_SWAP, 1);
2781 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2782 struct mem_cgroup *from, struct mem_cgroup *to)
2788 static DEFINE_MUTEX(memcg_max_mutex);
2790 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
2791 unsigned long max, bool memsw)
2793 bool enlarge = false;
2794 bool drained = false;
2796 bool limits_invariant;
2797 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
2800 if (signal_pending(current)) {
2805 mutex_lock(&memcg_max_mutex);
2807 * Make sure that the new limit (memsw or memory limit) doesn't
2808 * break our basic invariant rule memory.max <= memsw.max.
2810 limits_invariant = memsw ? max >= memcg->memory.max :
2811 max <= memcg->memsw.max;
2812 if (!limits_invariant) {
2813 mutex_unlock(&memcg_max_mutex);
2817 if (max > counter->max)
2819 ret = page_counter_set_max(counter, max);
2820 mutex_unlock(&memcg_max_mutex);
2826 drain_all_stock(memcg);
2831 if (!try_to_free_mem_cgroup_pages(memcg, 1,
2832 GFP_KERNEL, !memsw)) {
2838 if (!ret && enlarge)
2839 memcg_oom_recover(memcg);
2844 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
2846 unsigned long *total_scanned)
2848 unsigned long nr_reclaimed = 0;
2849 struct mem_cgroup_per_node *mz, *next_mz = NULL;
2850 unsigned long reclaimed;
2852 struct mem_cgroup_tree_per_node *mctz;
2853 unsigned long excess;
2854 unsigned long nr_scanned;
2859 mctz = soft_limit_tree_node(pgdat->node_id);
2862 * Do not even bother to check the largest node if the root
2863 * is empty. Do it lockless to prevent lock bouncing. Races
2864 * are acceptable as soft limit is best effort anyway.
2866 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
2870 * This loop can run a while, specially if mem_cgroup's continuously
2871 * keep exceeding their soft limit and putting the system under
2878 mz = mem_cgroup_largest_soft_limit_node(mctz);
2883 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
2884 gfp_mask, &nr_scanned);
2885 nr_reclaimed += reclaimed;
2886 *total_scanned += nr_scanned;
2887 spin_lock_irq(&mctz->lock);
2888 __mem_cgroup_remove_exceeded(mz, mctz);
2891 * If we failed to reclaim anything from this memory cgroup
2892 * it is time to move on to the next cgroup
2896 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2898 excess = soft_limit_excess(mz->memcg);
2900 * One school of thought says that we should not add
2901 * back the node to the tree if reclaim returns 0.
2902 * But our reclaim could return 0, simply because due
2903 * to priority we are exposing a smaller subset of
2904 * memory to reclaim from. Consider this as a longer
2907 /* If excess == 0, no tree ops */
2908 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2909 spin_unlock_irq(&mctz->lock);
2910 css_put(&mz->memcg->css);
2913 * Could not reclaim anything and there are no more
2914 * mem cgroups to try or we seem to be looping without
2915 * reclaiming anything.
2917 if (!nr_reclaimed &&
2919 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2921 } while (!nr_reclaimed);
2923 css_put(&next_mz->memcg->css);
2924 return nr_reclaimed;
2928 * Test whether @memcg has children, dead or alive. Note that this
2929 * function doesn't care whether @memcg has use_hierarchy enabled and
2930 * returns %true if there are child csses according to the cgroup
2931 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2933 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2938 ret = css_next_child(NULL, &memcg->css);
2944 * Reclaims as many pages from the given memcg as possible.
2946 * Caller is responsible for holding css reference for memcg.
2948 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2950 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2952 /* we call try-to-free pages for make this cgroup empty */
2953 lru_add_drain_all();
2955 drain_all_stock(memcg);
2957 /* try to free all pages in this cgroup */
2958 while (nr_retries && page_counter_read(&memcg->memory)) {
2961 if (signal_pending(current))
2964 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2968 /* maybe some writeback is necessary */
2969 congestion_wait(BLK_RW_ASYNC, HZ/10);
2977 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2978 char *buf, size_t nbytes,
2981 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2983 if (mem_cgroup_is_root(memcg))
2985 return mem_cgroup_force_empty(memcg) ?: nbytes;
2988 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2991 return mem_cgroup_from_css(css)->use_hierarchy;
2994 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2995 struct cftype *cft, u64 val)
2998 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2999 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3001 if (memcg->use_hierarchy == val)
3005 * If parent's use_hierarchy is set, we can't make any modifications
3006 * in the child subtrees. If it is unset, then the change can
3007 * occur, provided the current cgroup has no children.
3009 * For the root cgroup, parent_mem is NULL, we allow value to be
3010 * set if there are no children.
3012 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3013 (val == 1 || val == 0)) {
3014 if (!memcg_has_children(memcg))
3015 memcg->use_hierarchy = val;
3024 struct accumulated_vmstats {
3025 unsigned long vmstats[MEMCG_NR_STAT];
3026 unsigned long vmevents[NR_VM_EVENT_ITEMS];
3027 unsigned long lru_pages[NR_LRU_LISTS];
3029 /* overrides for v1 */
3030 const unsigned int *vmstats_array;
3031 const unsigned int *vmevents_array;
3037 static void accumulate_vmstats(struct mem_cgroup *memcg,
3038 struct accumulated_vmstats *acc)
3040 struct mem_cgroup *mi;
3043 for_each_mem_cgroup_tree(mi, memcg) {
3044 for (i = 0; i < acc->vmstats_size; i++)
3045 acc->vmstats[i] += memcg_page_state_local(mi,
3046 acc->vmstats_array ? acc->vmstats_array[i] : i);
3048 for (i = 0; i < acc->vmevents_size; i++)
3049 acc->vmevents[i] += memcg_events_local(mi,
3051 ? acc->vmevents_array[i] : i);
3053 for (i = 0; i < NR_LRU_LISTS; i++)
3054 acc->lru_pages[i] += memcg_page_state_local(mi,
3059 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3061 unsigned long val = 0;
3063 if (mem_cgroup_is_root(memcg)) {
3064 struct mem_cgroup *iter;
3066 for_each_mem_cgroup_tree(iter, memcg) {
3067 val += memcg_page_state_local(iter, MEMCG_CACHE);
3068 val += memcg_page_state_local(iter, MEMCG_RSS);
3070 val += memcg_page_state_local(iter, MEMCG_SWAP);
3074 val = page_counter_read(&memcg->memory);
3076 val = page_counter_read(&memcg->memsw);
3089 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3092 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3093 struct page_counter *counter;
3095 switch (MEMFILE_TYPE(cft->private)) {
3097 counter = &memcg->memory;
3100 counter = &memcg->memsw;
3103 counter = &memcg->kmem;
3106 counter = &memcg->tcpmem;
3112 switch (MEMFILE_ATTR(cft->private)) {
3114 if (counter == &memcg->memory)
3115 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3116 if (counter == &memcg->memsw)
3117 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3118 return (u64)page_counter_read(counter) * PAGE_SIZE;
3120 return (u64)counter->max * PAGE_SIZE;
3122 return (u64)counter->watermark * PAGE_SIZE;
3124 return counter->failcnt;
3125 case RES_SOFT_LIMIT:
3126 return (u64)memcg->soft_limit * PAGE_SIZE;
3132 #ifdef CONFIG_MEMCG_KMEM
3133 static int memcg_online_kmem(struct mem_cgroup *memcg)
3137 if (cgroup_memory_nokmem)
3140 BUG_ON(memcg->kmemcg_id >= 0);
3141 BUG_ON(memcg->kmem_state);
3143 memcg_id = memcg_alloc_cache_id();
3147 static_branch_inc(&memcg_kmem_enabled_key);
3149 * A memory cgroup is considered kmem-online as soon as it gets
3150 * kmemcg_id. Setting the id after enabling static branching will
3151 * guarantee no one starts accounting before all call sites are
3154 memcg->kmemcg_id = memcg_id;
3155 memcg->kmem_state = KMEM_ONLINE;
3156 INIT_LIST_HEAD(&memcg->kmem_caches);
3161 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3163 struct cgroup_subsys_state *css;
3164 struct mem_cgroup *parent, *child;
3167 if (memcg->kmem_state != KMEM_ONLINE)
3170 * Clear the online state before clearing memcg_caches array
3171 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3172 * guarantees that no cache will be created for this cgroup
3173 * after we are done (see memcg_create_kmem_cache()).
3175 memcg->kmem_state = KMEM_ALLOCATED;
3177 memcg_deactivate_kmem_caches(memcg);
3179 kmemcg_id = memcg->kmemcg_id;
3180 BUG_ON(kmemcg_id < 0);
3182 parent = parent_mem_cgroup(memcg);
3184 parent = root_mem_cgroup;
3187 * Change kmemcg_id of this cgroup and all its descendants to the
3188 * parent's id, and then move all entries from this cgroup's list_lrus
3189 * to ones of the parent. After we have finished, all list_lrus
3190 * corresponding to this cgroup are guaranteed to remain empty. The
3191 * ordering is imposed by list_lru_node->lock taken by
3192 * memcg_drain_all_list_lrus().
3194 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3195 css_for_each_descendant_pre(css, &memcg->css) {
3196 child = mem_cgroup_from_css(css);
3197 BUG_ON(child->kmemcg_id != kmemcg_id);
3198 child->kmemcg_id = parent->kmemcg_id;
3199 if (!memcg->use_hierarchy)
3204 memcg_drain_all_list_lrus(kmemcg_id, parent);
3206 memcg_free_cache_id(kmemcg_id);
3209 static void memcg_free_kmem(struct mem_cgroup *memcg)
3211 /* css_alloc() failed, offlining didn't happen */
3212 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3213 memcg_offline_kmem(memcg);
3215 if (memcg->kmem_state == KMEM_ALLOCATED) {
3216 memcg_destroy_kmem_caches(memcg);
3217 static_branch_dec(&memcg_kmem_enabled_key);
3218 WARN_ON(page_counter_read(&memcg->kmem));
3222 static int memcg_online_kmem(struct mem_cgroup *memcg)
3226 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3229 static void memcg_free_kmem(struct mem_cgroup *memcg)
3232 #endif /* CONFIG_MEMCG_KMEM */
3234 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3239 mutex_lock(&memcg_max_mutex);
3240 ret = page_counter_set_max(&memcg->kmem, max);
3241 mutex_unlock(&memcg_max_mutex);
3245 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3249 mutex_lock(&memcg_max_mutex);
3251 ret = page_counter_set_max(&memcg->tcpmem, max);
3255 if (!memcg->tcpmem_active) {
3257 * The active flag needs to be written after the static_key
3258 * update. This is what guarantees that the socket activation
3259 * function is the last one to run. See mem_cgroup_sk_alloc()
3260 * for details, and note that we don't mark any socket as
3261 * belonging to this memcg until that flag is up.
3263 * We need to do this, because static_keys will span multiple
3264 * sites, but we can't control their order. If we mark a socket
3265 * as accounted, but the accounting functions are not patched in
3266 * yet, we'll lose accounting.
3268 * We never race with the readers in mem_cgroup_sk_alloc(),
3269 * because when this value change, the code to process it is not
3272 static_branch_inc(&memcg_sockets_enabled_key);
3273 memcg->tcpmem_active = true;
3276 mutex_unlock(&memcg_max_mutex);
3281 * The user of this function is...
3284 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3285 char *buf, size_t nbytes, loff_t off)
3287 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3288 unsigned long nr_pages;
3291 buf = strstrip(buf);
3292 ret = page_counter_memparse(buf, "-1", &nr_pages);
3296 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3298 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3302 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3304 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3307 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3310 ret = memcg_update_kmem_max(memcg, nr_pages);
3313 ret = memcg_update_tcp_max(memcg, nr_pages);
3317 case RES_SOFT_LIMIT:
3318 memcg->soft_limit = nr_pages;
3322 return ret ?: nbytes;
3325 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3326 size_t nbytes, loff_t off)
3328 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3329 struct page_counter *counter;
3331 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3333 counter = &memcg->memory;
3336 counter = &memcg->memsw;
3339 counter = &memcg->kmem;
3342 counter = &memcg->tcpmem;
3348 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3350 page_counter_reset_watermark(counter);
3353 counter->failcnt = 0;
3362 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3365 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3369 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3370 struct cftype *cft, u64 val)
3372 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3374 if (val & ~MOVE_MASK)
3378 * No kind of locking is needed in here, because ->can_attach() will
3379 * check this value once in the beginning of the process, and then carry
3380 * on with stale data. This means that changes to this value will only
3381 * affect task migrations starting after the change.
3383 memcg->move_charge_at_immigrate = val;
3387 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3388 struct cftype *cft, u64 val)
3396 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3397 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3398 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3400 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3401 int nid, unsigned int lru_mask)
3403 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
3404 unsigned long nr = 0;
3407 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3410 if (!(BIT(lru) & lru_mask))
3412 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3417 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3418 unsigned int lru_mask)
3420 unsigned long nr = 0;
3424 if (!(BIT(lru) & lru_mask))
3426 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3431 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3435 unsigned int lru_mask;
3438 static const struct numa_stat stats[] = {
3439 { "total", LRU_ALL },
3440 { "file", LRU_ALL_FILE },
3441 { "anon", LRU_ALL_ANON },
3442 { "unevictable", BIT(LRU_UNEVICTABLE) },
3444 const struct numa_stat *stat;
3447 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3449 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3450 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3451 seq_printf(m, "%s=%lu", stat->name, nr);
3452 for_each_node_state(nid, N_MEMORY) {
3453 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3455 seq_printf(m, " N%d=%lu", nid, nr);
3460 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3461 struct mem_cgroup *iter;
3464 for_each_mem_cgroup_tree(iter, memcg)
3465 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3466 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3467 for_each_node_state(nid, N_MEMORY) {
3469 for_each_mem_cgroup_tree(iter, memcg)
3470 nr += mem_cgroup_node_nr_lru_pages(
3471 iter, nid, stat->lru_mask);
3472 seq_printf(m, " N%d=%lu", nid, nr);
3479 #endif /* CONFIG_NUMA */
3481 /* Universal VM events cgroup1 shows, original sort order */
3482 static const unsigned int memcg1_events[] = {
3489 static const char *const memcg1_event_names[] = {
3496 static int memcg_stat_show(struct seq_file *m, void *v)
3498 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3499 unsigned long memory, memsw;
3500 struct mem_cgroup *mi;
3502 struct accumulated_vmstats acc;
3504 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3505 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3507 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3508 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3510 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3511 memcg_page_state_local(memcg, memcg1_stats[i]) *
3515 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3516 seq_printf(m, "%s %lu\n", memcg1_event_names[i],
3517 memcg_events_local(memcg, memcg1_events[i]));
3519 for (i = 0; i < NR_LRU_LISTS; i++)
3520 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3521 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3524 /* Hierarchical information */
3525 memory = memsw = PAGE_COUNTER_MAX;
3526 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3527 memory = min(memory, mi->memory.max);
3528 memsw = min(memsw, mi->memsw.max);
3530 seq_printf(m, "hierarchical_memory_limit %llu\n",
3531 (u64)memory * PAGE_SIZE);
3532 if (do_memsw_account())
3533 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3534 (u64)memsw * PAGE_SIZE);
3536 memset(&acc, 0, sizeof(acc));
3537 acc.vmstats_size = ARRAY_SIZE(memcg1_stats);
3538 acc.vmstats_array = memcg1_stats;
3539 acc.vmevents_size = ARRAY_SIZE(memcg1_events);
3540 acc.vmevents_array = memcg1_events;
3541 accumulate_vmstats(memcg, &acc);
3543 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3544 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3546 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3547 (u64)acc.vmstats[i] * PAGE_SIZE);
3550 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3551 seq_printf(m, "total_%s %llu\n", memcg1_event_names[i],
3552 (u64)acc.vmevents[i]);
3554 for (i = 0; i < NR_LRU_LISTS; i++)
3555 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i],
3556 (u64)acc.lru_pages[i] * PAGE_SIZE);
3558 #ifdef CONFIG_DEBUG_VM
3561 struct mem_cgroup_per_node *mz;
3562 struct zone_reclaim_stat *rstat;
3563 unsigned long recent_rotated[2] = {0, 0};
3564 unsigned long recent_scanned[2] = {0, 0};
3566 for_each_online_pgdat(pgdat) {
3567 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3568 rstat = &mz->lruvec.reclaim_stat;
3570 recent_rotated[0] += rstat->recent_rotated[0];
3571 recent_rotated[1] += rstat->recent_rotated[1];
3572 recent_scanned[0] += rstat->recent_scanned[0];
3573 recent_scanned[1] += rstat->recent_scanned[1];
3575 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3576 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3577 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3578 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3585 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3588 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3590 return mem_cgroup_swappiness(memcg);
3593 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3594 struct cftype *cft, u64 val)
3596 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3602 memcg->swappiness = val;
3604 vm_swappiness = val;
3609 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3611 struct mem_cgroup_threshold_ary *t;
3612 unsigned long usage;
3617 t = rcu_dereference(memcg->thresholds.primary);
3619 t = rcu_dereference(memcg->memsw_thresholds.primary);
3624 usage = mem_cgroup_usage(memcg, swap);
3627 * current_threshold points to threshold just below or equal to usage.
3628 * If it's not true, a threshold was crossed after last
3629 * call of __mem_cgroup_threshold().
3631 i = t->current_threshold;
3634 * Iterate backward over array of thresholds starting from
3635 * current_threshold and check if a threshold is crossed.
3636 * If none of thresholds below usage is crossed, we read
3637 * only one element of the array here.
3639 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3640 eventfd_signal(t->entries[i].eventfd, 1);
3642 /* i = current_threshold + 1 */
3646 * Iterate forward over array of thresholds starting from
3647 * current_threshold+1 and check if a threshold is crossed.
3648 * If none of thresholds above usage is crossed, we read
3649 * only one element of the array here.
3651 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3652 eventfd_signal(t->entries[i].eventfd, 1);
3654 /* Update current_threshold */
3655 t->current_threshold = i - 1;
3660 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3663 __mem_cgroup_threshold(memcg, false);
3664 if (do_memsw_account())
3665 __mem_cgroup_threshold(memcg, true);
3667 memcg = parent_mem_cgroup(memcg);
3671 static int compare_thresholds(const void *a, const void *b)
3673 const struct mem_cgroup_threshold *_a = a;
3674 const struct mem_cgroup_threshold *_b = b;
3676 if (_a->threshold > _b->threshold)
3679 if (_a->threshold < _b->threshold)
3685 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3687 struct mem_cgroup_eventfd_list *ev;
3689 spin_lock(&memcg_oom_lock);
3691 list_for_each_entry(ev, &memcg->oom_notify, list)
3692 eventfd_signal(ev->eventfd, 1);
3694 spin_unlock(&memcg_oom_lock);
3698 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3700 struct mem_cgroup *iter;
3702 for_each_mem_cgroup_tree(iter, memcg)
3703 mem_cgroup_oom_notify_cb(iter);
3706 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3707 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3709 struct mem_cgroup_thresholds *thresholds;
3710 struct mem_cgroup_threshold_ary *new;
3711 unsigned long threshold;
3712 unsigned long usage;
3715 ret = page_counter_memparse(args, "-1", &threshold);
3719 mutex_lock(&memcg->thresholds_lock);
3722 thresholds = &memcg->thresholds;
3723 usage = mem_cgroup_usage(memcg, false);
3724 } else if (type == _MEMSWAP) {
3725 thresholds = &memcg->memsw_thresholds;
3726 usage = mem_cgroup_usage(memcg, true);
3730 /* Check if a threshold crossed before adding a new one */
3731 if (thresholds->primary)
3732 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3734 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3736 /* Allocate memory for new array of thresholds */
3737 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
3744 /* Copy thresholds (if any) to new array */
3745 if (thresholds->primary) {
3746 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3747 sizeof(struct mem_cgroup_threshold));
3750 /* Add new threshold */
3751 new->entries[size - 1].eventfd = eventfd;
3752 new->entries[size - 1].threshold = threshold;
3754 /* Sort thresholds. Registering of new threshold isn't time-critical */
3755 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3756 compare_thresholds, NULL);
3758 /* Find current threshold */
3759 new->current_threshold = -1;
3760 for (i = 0; i < size; i++) {
3761 if (new->entries[i].threshold <= usage) {
3763 * new->current_threshold will not be used until
3764 * rcu_assign_pointer(), so it's safe to increment
3767 ++new->current_threshold;
3772 /* Free old spare buffer and save old primary buffer as spare */
3773 kfree(thresholds->spare);
3774 thresholds->spare = thresholds->primary;
3776 rcu_assign_pointer(thresholds->primary, new);
3778 /* To be sure that nobody uses thresholds */
3782 mutex_unlock(&memcg->thresholds_lock);
3787 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3788 struct eventfd_ctx *eventfd, const char *args)
3790 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3793 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3794 struct eventfd_ctx *eventfd, const char *args)
3796 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3799 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3800 struct eventfd_ctx *eventfd, enum res_type type)
3802 struct mem_cgroup_thresholds *thresholds;
3803 struct mem_cgroup_threshold_ary *new;
3804 unsigned long usage;
3807 mutex_lock(&memcg->thresholds_lock);
3810 thresholds = &memcg->thresholds;
3811 usage = mem_cgroup_usage(memcg, false);
3812 } else if (type == _MEMSWAP) {
3813 thresholds = &memcg->memsw_thresholds;
3814 usage = mem_cgroup_usage(memcg, true);
3818 if (!thresholds->primary)
3821 /* Check if a threshold crossed before removing */
3822 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3824 /* Calculate new number of threshold */
3826 for (i = 0; i < thresholds->primary->size; i++) {
3827 if (thresholds->primary->entries[i].eventfd != eventfd)
3831 new = thresholds->spare;
3833 /* Set thresholds array to NULL if we don't have thresholds */
3842 /* Copy thresholds and find current threshold */
3843 new->current_threshold = -1;
3844 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3845 if (thresholds->primary->entries[i].eventfd == eventfd)
3848 new->entries[j] = thresholds->primary->entries[i];
3849 if (new->entries[j].threshold <= usage) {
3851 * new->current_threshold will not be used
3852 * until rcu_assign_pointer(), so it's safe to increment
3855 ++new->current_threshold;
3861 /* Swap primary and spare array */
3862 thresholds->spare = thresholds->primary;
3864 rcu_assign_pointer(thresholds->primary, new);
3866 /* To be sure that nobody uses thresholds */
3869 /* If all events are unregistered, free the spare array */
3871 kfree(thresholds->spare);
3872 thresholds->spare = NULL;
3875 mutex_unlock(&memcg->thresholds_lock);
3878 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3879 struct eventfd_ctx *eventfd)
3881 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3884 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3885 struct eventfd_ctx *eventfd)
3887 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3890 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3891 struct eventfd_ctx *eventfd, const char *args)
3893 struct mem_cgroup_eventfd_list *event;
3895 event = kmalloc(sizeof(*event), GFP_KERNEL);
3899 spin_lock(&memcg_oom_lock);
3901 event->eventfd = eventfd;
3902 list_add(&event->list, &memcg->oom_notify);
3904 /* already in OOM ? */
3905 if (memcg->under_oom)
3906 eventfd_signal(eventfd, 1);
3907 spin_unlock(&memcg_oom_lock);
3912 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3913 struct eventfd_ctx *eventfd)
3915 struct mem_cgroup_eventfd_list *ev, *tmp;
3917 spin_lock(&memcg_oom_lock);
3919 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3920 if (ev->eventfd == eventfd) {
3921 list_del(&ev->list);
3926 spin_unlock(&memcg_oom_lock);
3929 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3931 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
3933 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3934 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3935 seq_printf(sf, "oom_kill %lu\n",
3936 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
3940 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3941 struct cftype *cft, u64 val)
3943 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3945 /* cannot set to root cgroup and only 0 and 1 are allowed */
3946 if (!css->parent || !((val == 0) || (val == 1)))
3949 memcg->oom_kill_disable = val;
3951 memcg_oom_recover(memcg);
3956 #ifdef CONFIG_CGROUP_WRITEBACK
3958 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3960 return wb_domain_init(&memcg->cgwb_domain, gfp);
3963 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3965 wb_domain_exit(&memcg->cgwb_domain);
3968 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3970 wb_domain_size_changed(&memcg->cgwb_domain);
3973 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3975 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3977 if (!memcg->css.parent)
3980 return &memcg->cgwb_domain;
3984 * idx can be of type enum memcg_stat_item or node_stat_item.
3985 * Keep in sync with memcg_exact_page().
3987 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
3989 long x = atomic_long_read(&memcg->vmstats[idx]);
3992 for_each_online_cpu(cpu)
3993 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4000 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4001 * @wb: bdi_writeback in question
4002 * @pfilepages: out parameter for number of file pages
4003 * @pheadroom: out parameter for number of allocatable pages according to memcg
4004 * @pdirty: out parameter for number of dirty pages
4005 * @pwriteback: out parameter for number of pages under writeback
4007 * Determine the numbers of file, headroom, dirty, and writeback pages in
4008 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4009 * is a bit more involved.
4011 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4012 * headroom is calculated as the lowest headroom of itself and the
4013 * ancestors. Note that this doesn't consider the actual amount of
4014 * available memory in the system. The caller should further cap
4015 * *@pheadroom accordingly.
4017 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4018 unsigned long *pheadroom, unsigned long *pdirty,
4019 unsigned long *pwriteback)
4021 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4022 struct mem_cgroup *parent;
4024 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4026 /* this should eventually include NR_UNSTABLE_NFS */
4027 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4028 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4029 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4030 *pheadroom = PAGE_COUNTER_MAX;
4032 while ((parent = parent_mem_cgroup(memcg))) {
4033 unsigned long ceiling = min(memcg->memory.max, memcg->high);
4034 unsigned long used = page_counter_read(&memcg->memory);
4036 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4041 #else /* CONFIG_CGROUP_WRITEBACK */
4043 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4048 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4052 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4056 #endif /* CONFIG_CGROUP_WRITEBACK */
4059 * DO NOT USE IN NEW FILES.
4061 * "cgroup.event_control" implementation.
4063 * This is way over-engineered. It tries to support fully configurable
4064 * events for each user. Such level of flexibility is completely
4065 * unnecessary especially in the light of the planned unified hierarchy.
4067 * Please deprecate this and replace with something simpler if at all
4072 * Unregister event and free resources.
4074 * Gets called from workqueue.
4076 static void memcg_event_remove(struct work_struct *work)
4078 struct mem_cgroup_event *event =
4079 container_of(work, struct mem_cgroup_event, remove);
4080 struct mem_cgroup *memcg = event->memcg;
4082 remove_wait_queue(event->wqh, &event->wait);
4084 event->unregister_event(memcg, event->eventfd);
4086 /* Notify userspace the event is going away. */
4087 eventfd_signal(event->eventfd, 1);
4089 eventfd_ctx_put(event->eventfd);
4091 css_put(&memcg->css);
4095 * Gets called on EPOLLHUP on eventfd when user closes it.
4097 * Called with wqh->lock held and interrupts disabled.
4099 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4100 int sync, void *key)
4102 struct mem_cgroup_event *event =
4103 container_of(wait, struct mem_cgroup_event, wait);
4104 struct mem_cgroup *memcg = event->memcg;
4105 __poll_t flags = key_to_poll(key);
4107 if (flags & EPOLLHUP) {
4109 * If the event has been detached at cgroup removal, we
4110 * can simply return knowing the other side will cleanup
4113 * We can't race against event freeing since the other
4114 * side will require wqh->lock via remove_wait_queue(),
4117 spin_lock(&memcg->event_list_lock);
4118 if (!list_empty(&event->list)) {
4119 list_del_init(&event->list);
4121 * We are in atomic context, but cgroup_event_remove()
4122 * may sleep, so we have to call it in workqueue.
4124 schedule_work(&event->remove);
4126 spin_unlock(&memcg->event_list_lock);
4132 static void memcg_event_ptable_queue_proc(struct file *file,
4133 wait_queue_head_t *wqh, poll_table *pt)
4135 struct mem_cgroup_event *event =
4136 container_of(pt, struct mem_cgroup_event, pt);
4139 add_wait_queue(wqh, &event->wait);
4143 * DO NOT USE IN NEW FILES.
4145 * Parse input and register new cgroup event handler.
4147 * Input must be in format '<event_fd> <control_fd> <args>'.
4148 * Interpretation of args is defined by control file implementation.
4150 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4151 char *buf, size_t nbytes, loff_t off)
4153 struct cgroup_subsys_state *css = of_css(of);
4154 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4155 struct mem_cgroup_event *event;
4156 struct cgroup_subsys_state *cfile_css;
4157 unsigned int efd, cfd;
4164 buf = strstrip(buf);
4166 efd = simple_strtoul(buf, &endp, 10);
4171 cfd = simple_strtoul(buf, &endp, 10);
4172 if ((*endp != ' ') && (*endp != '\0'))
4176 event = kzalloc(sizeof(*event), GFP_KERNEL);
4180 event->memcg = memcg;
4181 INIT_LIST_HEAD(&event->list);
4182 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4183 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4184 INIT_WORK(&event->remove, memcg_event_remove);
4192 event->eventfd = eventfd_ctx_fileget(efile.file);
4193 if (IS_ERR(event->eventfd)) {
4194 ret = PTR_ERR(event->eventfd);
4201 goto out_put_eventfd;
4204 /* the process need read permission on control file */
4205 /* AV: shouldn't we check that it's been opened for read instead? */
4206 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4211 * Determine the event callbacks and set them in @event. This used
4212 * to be done via struct cftype but cgroup core no longer knows
4213 * about these events. The following is crude but the whole thing
4214 * is for compatibility anyway.
4216 * DO NOT ADD NEW FILES.
4218 name = cfile.file->f_path.dentry->d_name.name;
4220 if (!strcmp(name, "memory.usage_in_bytes")) {
4221 event->register_event = mem_cgroup_usage_register_event;
4222 event->unregister_event = mem_cgroup_usage_unregister_event;
4223 } else if (!strcmp(name, "memory.oom_control")) {
4224 event->register_event = mem_cgroup_oom_register_event;
4225 event->unregister_event = mem_cgroup_oom_unregister_event;
4226 } else if (!strcmp(name, "memory.pressure_level")) {
4227 event->register_event = vmpressure_register_event;
4228 event->unregister_event = vmpressure_unregister_event;
4229 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4230 event->register_event = memsw_cgroup_usage_register_event;
4231 event->unregister_event = memsw_cgroup_usage_unregister_event;
4238 * Verify @cfile should belong to @css. Also, remaining events are
4239 * automatically removed on cgroup destruction but the removal is
4240 * asynchronous, so take an extra ref on @css.
4242 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4243 &memory_cgrp_subsys);
4245 if (IS_ERR(cfile_css))
4247 if (cfile_css != css) {
4252 ret = event->register_event(memcg, event->eventfd, buf);
4256 vfs_poll(efile.file, &event->pt);
4258 spin_lock(&memcg->event_list_lock);
4259 list_add(&event->list, &memcg->event_list);
4260 spin_unlock(&memcg->event_list_lock);
4272 eventfd_ctx_put(event->eventfd);
4281 static struct cftype mem_cgroup_legacy_files[] = {
4283 .name = "usage_in_bytes",
4284 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4285 .read_u64 = mem_cgroup_read_u64,
4288 .name = "max_usage_in_bytes",
4289 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4290 .write = mem_cgroup_reset,
4291 .read_u64 = mem_cgroup_read_u64,
4294 .name = "limit_in_bytes",
4295 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4296 .write = mem_cgroup_write,
4297 .read_u64 = mem_cgroup_read_u64,
4300 .name = "soft_limit_in_bytes",
4301 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4302 .write = mem_cgroup_write,
4303 .read_u64 = mem_cgroup_read_u64,
4307 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4308 .write = mem_cgroup_reset,
4309 .read_u64 = mem_cgroup_read_u64,
4313 .seq_show = memcg_stat_show,
4316 .name = "force_empty",
4317 .write = mem_cgroup_force_empty_write,
4320 .name = "use_hierarchy",
4321 .write_u64 = mem_cgroup_hierarchy_write,
4322 .read_u64 = mem_cgroup_hierarchy_read,
4325 .name = "cgroup.event_control", /* XXX: for compat */
4326 .write = memcg_write_event_control,
4327 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4330 .name = "swappiness",
4331 .read_u64 = mem_cgroup_swappiness_read,
4332 .write_u64 = mem_cgroup_swappiness_write,
4335 .name = "move_charge_at_immigrate",
4336 .read_u64 = mem_cgroup_move_charge_read,
4337 .write_u64 = mem_cgroup_move_charge_write,
4340 .name = "oom_control",
4341 .seq_show = mem_cgroup_oom_control_read,
4342 .write_u64 = mem_cgroup_oom_control_write,
4343 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4346 .name = "pressure_level",
4350 .name = "numa_stat",
4351 .seq_show = memcg_numa_stat_show,
4355 .name = "kmem.limit_in_bytes",
4356 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4357 .write = mem_cgroup_write,
4358 .read_u64 = mem_cgroup_read_u64,
4361 .name = "kmem.usage_in_bytes",
4362 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4363 .read_u64 = mem_cgroup_read_u64,
4366 .name = "kmem.failcnt",
4367 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4368 .write = mem_cgroup_reset,
4369 .read_u64 = mem_cgroup_read_u64,
4372 .name = "kmem.max_usage_in_bytes",
4373 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4374 .write = mem_cgroup_reset,
4375 .read_u64 = mem_cgroup_read_u64,
4377 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4379 .name = "kmem.slabinfo",
4380 .seq_start = memcg_slab_start,
4381 .seq_next = memcg_slab_next,
4382 .seq_stop = memcg_slab_stop,
4383 .seq_show = memcg_slab_show,
4387 .name = "kmem.tcp.limit_in_bytes",
4388 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4389 .write = mem_cgroup_write,
4390 .read_u64 = mem_cgroup_read_u64,
4393 .name = "kmem.tcp.usage_in_bytes",
4394 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4395 .read_u64 = mem_cgroup_read_u64,
4398 .name = "kmem.tcp.failcnt",
4399 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4400 .write = mem_cgroup_reset,
4401 .read_u64 = mem_cgroup_read_u64,
4404 .name = "kmem.tcp.max_usage_in_bytes",
4405 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4406 .write = mem_cgroup_reset,
4407 .read_u64 = mem_cgroup_read_u64,
4409 { }, /* terminate */
4413 * Private memory cgroup IDR
4415 * Swap-out records and page cache shadow entries need to store memcg
4416 * references in constrained space, so we maintain an ID space that is
4417 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4418 * memory-controlled cgroups to 64k.
4420 * However, there usually are many references to the oflline CSS after
4421 * the cgroup has been destroyed, such as page cache or reclaimable
4422 * slab objects, that don't need to hang on to the ID. We want to keep
4423 * those dead CSS from occupying IDs, or we might quickly exhaust the
4424 * relatively small ID space and prevent the creation of new cgroups
4425 * even when there are much fewer than 64k cgroups - possibly none.
4427 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4428 * be freed and recycled when it's no longer needed, which is usually
4429 * when the CSS is offlined.
4431 * The only exception to that are records of swapped out tmpfs/shmem
4432 * pages that need to be attributed to live ancestors on swapin. But
4433 * those references are manageable from userspace.
4436 static DEFINE_IDR(mem_cgroup_idr);
4438 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4440 if (memcg->id.id > 0) {
4441 idr_remove(&mem_cgroup_idr, memcg->id.id);
4446 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4448 refcount_add(n, &memcg->id.ref);
4451 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4453 if (refcount_sub_and_test(n, &memcg->id.ref)) {
4454 mem_cgroup_id_remove(memcg);
4456 /* Memcg ID pins CSS */
4457 css_put(&memcg->css);
4461 static inline void mem_cgroup_id_get(struct mem_cgroup *memcg)
4463 mem_cgroup_id_get_many(memcg, 1);
4466 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4468 mem_cgroup_id_put_many(memcg, 1);
4472 * mem_cgroup_from_id - look up a memcg from a memcg id
4473 * @id: the memcg id to look up
4475 * Caller must hold rcu_read_lock().
4477 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4479 WARN_ON_ONCE(!rcu_read_lock_held());
4480 return idr_find(&mem_cgroup_idr, id);
4483 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4485 struct mem_cgroup_per_node *pn;
4488 * This routine is called against possible nodes.
4489 * But it's BUG to call kmalloc() against offline node.
4491 * TODO: this routine can waste much memory for nodes which will
4492 * never be onlined. It's better to use memory hotplug callback
4495 if (!node_state(node, N_NORMAL_MEMORY))
4497 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4501 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
4502 if (!pn->lruvec_stat_cpu) {
4507 lruvec_init(&pn->lruvec);
4508 pn->usage_in_excess = 0;
4509 pn->on_tree = false;
4512 memcg->nodeinfo[node] = pn;
4516 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4518 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
4523 free_percpu(pn->lruvec_stat_cpu);
4527 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4532 free_mem_cgroup_per_node_info(memcg, node);
4533 free_percpu(memcg->vmstats_percpu);
4537 static void mem_cgroup_free(struct mem_cgroup *memcg)
4539 memcg_wb_domain_exit(memcg);
4540 __mem_cgroup_free(memcg);
4543 static struct mem_cgroup *mem_cgroup_alloc(void)
4545 struct mem_cgroup *memcg;
4549 size = sizeof(struct mem_cgroup);
4550 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4552 memcg = kzalloc(size, GFP_KERNEL);
4556 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4557 1, MEM_CGROUP_ID_MAX,
4559 if (memcg->id.id < 0)
4562 memcg->vmstats_percpu = alloc_percpu(struct memcg_vmstats_percpu);
4563 if (!memcg->vmstats_percpu)
4567 if (alloc_mem_cgroup_per_node_info(memcg, node))
4570 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4573 INIT_WORK(&memcg->high_work, high_work_func);
4574 memcg->last_scanned_node = MAX_NUMNODES;
4575 INIT_LIST_HEAD(&memcg->oom_notify);
4576 mutex_init(&memcg->thresholds_lock);
4577 spin_lock_init(&memcg->move_lock);
4578 vmpressure_init(&memcg->vmpressure);
4579 INIT_LIST_HEAD(&memcg->event_list);
4580 spin_lock_init(&memcg->event_list_lock);
4581 memcg->socket_pressure = jiffies;
4582 #ifdef CONFIG_MEMCG_KMEM
4583 memcg->kmemcg_id = -1;
4585 #ifdef CONFIG_CGROUP_WRITEBACK
4586 INIT_LIST_HEAD(&memcg->cgwb_list);
4588 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4591 mem_cgroup_id_remove(memcg);
4592 __mem_cgroup_free(memcg);
4596 static struct cgroup_subsys_state * __ref
4597 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4599 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4600 struct mem_cgroup *memcg;
4601 long error = -ENOMEM;
4603 memcg = mem_cgroup_alloc();
4605 return ERR_PTR(error);
4607 memcg->high = PAGE_COUNTER_MAX;
4608 memcg->soft_limit = PAGE_COUNTER_MAX;
4610 memcg->swappiness = mem_cgroup_swappiness(parent);
4611 memcg->oom_kill_disable = parent->oom_kill_disable;
4613 if (parent && parent->use_hierarchy) {
4614 memcg->use_hierarchy = true;
4615 page_counter_init(&memcg->memory, &parent->memory);
4616 page_counter_init(&memcg->swap, &parent->swap);
4617 page_counter_init(&memcg->memsw, &parent->memsw);
4618 page_counter_init(&memcg->kmem, &parent->kmem);
4619 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4621 page_counter_init(&memcg->memory, NULL);
4622 page_counter_init(&memcg->swap, NULL);
4623 page_counter_init(&memcg->memsw, NULL);
4624 page_counter_init(&memcg->kmem, NULL);
4625 page_counter_init(&memcg->tcpmem, NULL);
4627 * Deeper hierachy with use_hierarchy == false doesn't make
4628 * much sense so let cgroup subsystem know about this
4629 * unfortunate state in our controller.
4631 if (parent != root_mem_cgroup)
4632 memory_cgrp_subsys.broken_hierarchy = true;
4635 /* The following stuff does not apply to the root */
4637 root_mem_cgroup = memcg;
4641 error = memcg_online_kmem(memcg);
4645 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4646 static_branch_inc(&memcg_sockets_enabled_key);
4650 mem_cgroup_id_remove(memcg);
4651 mem_cgroup_free(memcg);
4652 return ERR_PTR(-ENOMEM);
4655 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
4657 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4660 * A memcg must be visible for memcg_expand_shrinker_maps()
4661 * by the time the maps are allocated. So, we allocate maps
4662 * here, when for_each_mem_cgroup() can't skip it.
4664 if (memcg_alloc_shrinker_maps(memcg)) {
4665 mem_cgroup_id_remove(memcg);
4669 /* Online state pins memcg ID, memcg ID pins CSS */
4670 refcount_set(&memcg->id.ref, 1);
4675 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4677 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4678 struct mem_cgroup_event *event, *tmp;
4681 * Unregister events and notify userspace.
4682 * Notify userspace about cgroup removing only after rmdir of cgroup
4683 * directory to avoid race between userspace and kernelspace.
4685 spin_lock(&memcg->event_list_lock);
4686 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4687 list_del_init(&event->list);
4688 schedule_work(&event->remove);
4690 spin_unlock(&memcg->event_list_lock);
4692 page_counter_set_min(&memcg->memory, 0);
4693 page_counter_set_low(&memcg->memory, 0);
4695 memcg_offline_kmem(memcg);
4696 wb_memcg_offline(memcg);
4698 drain_all_stock(memcg);
4700 mem_cgroup_id_put(memcg);
4703 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4705 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4707 invalidate_reclaim_iterators(memcg);
4710 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4712 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4714 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4715 static_branch_dec(&memcg_sockets_enabled_key);
4717 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4718 static_branch_dec(&memcg_sockets_enabled_key);
4720 vmpressure_cleanup(&memcg->vmpressure);
4721 cancel_work_sync(&memcg->high_work);
4722 mem_cgroup_remove_from_trees(memcg);
4723 memcg_free_shrinker_maps(memcg);
4724 memcg_free_kmem(memcg);
4725 mem_cgroup_free(memcg);
4729 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4730 * @css: the target css
4732 * Reset the states of the mem_cgroup associated with @css. This is
4733 * invoked when the userland requests disabling on the default hierarchy
4734 * but the memcg is pinned through dependency. The memcg should stop
4735 * applying policies and should revert to the vanilla state as it may be
4736 * made visible again.
4738 * The current implementation only resets the essential configurations.
4739 * This needs to be expanded to cover all the visible parts.
4741 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4743 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4745 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
4746 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
4747 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
4748 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
4749 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
4750 page_counter_set_min(&memcg->memory, 0);
4751 page_counter_set_low(&memcg->memory, 0);
4752 memcg->high = PAGE_COUNTER_MAX;
4753 memcg->soft_limit = PAGE_COUNTER_MAX;
4754 memcg_wb_domain_size_changed(memcg);
4758 /* Handlers for move charge at task migration. */
4759 static int mem_cgroup_do_precharge(unsigned long count)
4763 /* Try a single bulk charge without reclaim first, kswapd may wake */
4764 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4766 mc.precharge += count;
4770 /* Try charges one by one with reclaim, but do not retry */
4772 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
4786 enum mc_target_type {
4793 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4794 unsigned long addr, pte_t ptent)
4796 struct page *page = _vm_normal_page(vma, addr, ptent, true);
4798 if (!page || !page_mapped(page))
4800 if (PageAnon(page)) {
4801 if (!(mc.flags & MOVE_ANON))
4804 if (!(mc.flags & MOVE_FILE))
4807 if (!get_page_unless_zero(page))
4813 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
4814 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4815 pte_t ptent, swp_entry_t *entry)
4817 struct page *page = NULL;
4818 swp_entry_t ent = pte_to_swp_entry(ptent);
4820 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4824 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
4825 * a device and because they are not accessible by CPU they are store
4826 * as special swap entry in the CPU page table.
4828 if (is_device_private_entry(ent)) {
4829 page = device_private_entry_to_page(ent);
4831 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
4832 * a refcount of 1 when free (unlike normal page)
4834 if (!page_ref_add_unless(page, 1, 1))
4840 * Because lookup_swap_cache() updates some statistics counter,
4841 * we call find_get_page() with swapper_space directly.
4843 page = find_get_page(swap_address_space(ent), swp_offset(ent));
4844 if (do_memsw_account())
4845 entry->val = ent.val;
4850 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4851 pte_t ptent, swp_entry_t *entry)
4857 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4858 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4860 struct page *page = NULL;
4861 struct address_space *mapping;
4864 if (!vma->vm_file) /* anonymous vma */
4866 if (!(mc.flags & MOVE_FILE))
4869 mapping = vma->vm_file->f_mapping;
4870 pgoff = linear_page_index(vma, addr);
4872 /* page is moved even if it's not RSS of this task(page-faulted). */
4874 /* shmem/tmpfs may report page out on swap: account for that too. */
4875 if (shmem_mapping(mapping)) {
4876 page = find_get_entry(mapping, pgoff);
4877 if (xa_is_value(page)) {
4878 swp_entry_t swp = radix_to_swp_entry(page);
4879 if (do_memsw_account())
4881 page = find_get_page(swap_address_space(swp),
4885 page = find_get_page(mapping, pgoff);
4887 page = find_get_page(mapping, pgoff);
4893 * mem_cgroup_move_account - move account of the page
4895 * @compound: charge the page as compound or small page
4896 * @from: mem_cgroup which the page is moved from.
4897 * @to: mem_cgroup which the page is moved to. @from != @to.
4899 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4901 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4904 static int mem_cgroup_move_account(struct page *page,
4906 struct mem_cgroup *from,
4907 struct mem_cgroup *to)
4909 unsigned long flags;
4910 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4914 VM_BUG_ON(from == to);
4915 VM_BUG_ON_PAGE(PageLRU(page), page);
4916 VM_BUG_ON(compound && !PageTransHuge(page));
4919 * Prevent mem_cgroup_migrate() from looking at
4920 * page->mem_cgroup of its source page while we change it.
4923 if (!trylock_page(page))
4927 if (page->mem_cgroup != from)
4930 anon = PageAnon(page);
4932 spin_lock_irqsave(&from->move_lock, flags);
4934 if (!anon && page_mapped(page)) {
4935 __mod_memcg_state(from, NR_FILE_MAPPED, -nr_pages);
4936 __mod_memcg_state(to, NR_FILE_MAPPED, nr_pages);
4940 * move_lock grabbed above and caller set from->moving_account, so
4941 * mod_memcg_page_state will serialize updates to PageDirty.
4942 * So mapping should be stable for dirty pages.
4944 if (!anon && PageDirty(page)) {
4945 struct address_space *mapping = page_mapping(page);
4947 if (mapping_cap_account_dirty(mapping)) {
4948 __mod_memcg_state(from, NR_FILE_DIRTY, -nr_pages);
4949 __mod_memcg_state(to, NR_FILE_DIRTY, nr_pages);
4953 if (PageWriteback(page)) {
4954 __mod_memcg_state(from, NR_WRITEBACK, -nr_pages);
4955 __mod_memcg_state(to, NR_WRITEBACK, nr_pages);
4959 * It is safe to change page->mem_cgroup here because the page
4960 * is referenced, charged, and isolated - we can't race with
4961 * uncharging, charging, migration, or LRU putback.
4964 /* caller should have done css_get */
4965 page->mem_cgroup = to;
4966 spin_unlock_irqrestore(&from->move_lock, flags);
4970 local_irq_disable();
4971 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4972 memcg_check_events(to, page);
4973 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4974 memcg_check_events(from, page);
4983 * get_mctgt_type - get target type of moving charge
4984 * @vma: the vma the pte to be checked belongs
4985 * @addr: the address corresponding to the pte to be checked
4986 * @ptent: the pte to be checked
4987 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4990 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4991 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4992 * move charge. if @target is not NULL, the page is stored in target->page
4993 * with extra refcnt got(Callers should handle it).
4994 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4995 * target for charge migration. if @target is not NULL, the entry is stored
4997 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PUBLIC
4998 * or MEMORY_DEVICE_PRIVATE (so ZONE_DEVICE page and thus not on the lru).
4999 * For now we such page is charge like a regular page would be as for all
5000 * intent and purposes it is just special memory taking the place of a
5003 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5005 * Called with pte lock held.
5008 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5009 unsigned long addr, pte_t ptent, union mc_target *target)
5011 struct page *page = NULL;
5012 enum mc_target_type ret = MC_TARGET_NONE;
5013 swp_entry_t ent = { .val = 0 };
5015 if (pte_present(ptent))
5016 page = mc_handle_present_pte(vma, addr, ptent);
5017 else if (is_swap_pte(ptent))
5018 page = mc_handle_swap_pte(vma, ptent, &ent);
5019 else if (pte_none(ptent))
5020 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5022 if (!page && !ent.val)
5026 * Do only loose check w/o serialization.
5027 * mem_cgroup_move_account() checks the page is valid or
5028 * not under LRU exclusion.
5030 if (page->mem_cgroup == mc.from) {
5031 ret = MC_TARGET_PAGE;
5032 if (is_device_private_page(page) ||
5033 is_device_public_page(page))
5034 ret = MC_TARGET_DEVICE;
5036 target->page = page;
5038 if (!ret || !target)
5042 * There is a swap entry and a page doesn't exist or isn't charged.
5043 * But we cannot move a tail-page in a THP.
5045 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5046 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5047 ret = MC_TARGET_SWAP;
5054 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5056 * We don't consider PMD mapped swapping or file mapped pages because THP does
5057 * not support them for now.
5058 * Caller should make sure that pmd_trans_huge(pmd) is true.
5060 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5061 unsigned long addr, pmd_t pmd, union mc_target *target)
5063 struct page *page = NULL;
5064 enum mc_target_type ret = MC_TARGET_NONE;
5066 if (unlikely(is_swap_pmd(pmd))) {
5067 VM_BUG_ON(thp_migration_supported() &&
5068 !is_pmd_migration_entry(pmd));
5071 page = pmd_page(pmd);
5072 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5073 if (!(mc.flags & MOVE_ANON))
5075 if (page->mem_cgroup == mc.from) {
5076 ret = MC_TARGET_PAGE;
5079 target->page = page;
5085 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5086 unsigned long addr, pmd_t pmd, union mc_target *target)
5088 return MC_TARGET_NONE;
5092 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5093 unsigned long addr, unsigned long end,
5094 struct mm_walk *walk)
5096 struct vm_area_struct *vma = walk->vma;
5100 ptl = pmd_trans_huge_lock(pmd, vma);
5103 * Note their can not be MC_TARGET_DEVICE for now as we do not
5104 * support transparent huge page with MEMORY_DEVICE_PUBLIC or
5105 * MEMORY_DEVICE_PRIVATE but this might change.
5107 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5108 mc.precharge += HPAGE_PMD_NR;
5113 if (pmd_trans_unstable(pmd))
5115 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5116 for (; addr != end; pte++, addr += PAGE_SIZE)
5117 if (get_mctgt_type(vma, addr, *pte, NULL))
5118 mc.precharge++; /* increment precharge temporarily */
5119 pte_unmap_unlock(pte - 1, ptl);
5125 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5127 unsigned long precharge;
5129 struct mm_walk mem_cgroup_count_precharge_walk = {
5130 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5133 down_read(&mm->mmap_sem);
5134 walk_page_range(0, mm->highest_vm_end,
5135 &mem_cgroup_count_precharge_walk);
5136 up_read(&mm->mmap_sem);
5138 precharge = mc.precharge;
5144 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5146 unsigned long precharge = mem_cgroup_count_precharge(mm);
5148 VM_BUG_ON(mc.moving_task);
5149 mc.moving_task = current;
5150 return mem_cgroup_do_precharge(precharge);
5153 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5154 static void __mem_cgroup_clear_mc(void)
5156 struct mem_cgroup *from = mc.from;
5157 struct mem_cgroup *to = mc.to;
5159 /* we must uncharge all the leftover precharges from mc.to */
5161 cancel_charge(mc.to, mc.precharge);
5165 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5166 * we must uncharge here.
5168 if (mc.moved_charge) {
5169 cancel_charge(mc.from, mc.moved_charge);
5170 mc.moved_charge = 0;
5172 /* we must fixup refcnts and charges */
5173 if (mc.moved_swap) {
5174 /* uncharge swap account from the old cgroup */
5175 if (!mem_cgroup_is_root(mc.from))
5176 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5178 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5181 * we charged both to->memory and to->memsw, so we
5182 * should uncharge to->memory.
5184 if (!mem_cgroup_is_root(mc.to))
5185 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5187 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
5188 css_put_many(&mc.to->css, mc.moved_swap);
5192 memcg_oom_recover(from);
5193 memcg_oom_recover(to);
5194 wake_up_all(&mc.waitq);
5197 static void mem_cgroup_clear_mc(void)
5199 struct mm_struct *mm = mc.mm;
5202 * we must clear moving_task before waking up waiters at the end of
5205 mc.moving_task = NULL;
5206 __mem_cgroup_clear_mc();
5207 spin_lock(&mc.lock);
5211 spin_unlock(&mc.lock);
5216 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5218 struct cgroup_subsys_state *css;
5219 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5220 struct mem_cgroup *from;
5221 struct task_struct *leader, *p;
5222 struct mm_struct *mm;
5223 unsigned long move_flags;
5226 /* charge immigration isn't supported on the default hierarchy */
5227 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5231 * Multi-process migrations only happen on the default hierarchy
5232 * where charge immigration is not used. Perform charge
5233 * immigration if @tset contains a leader and whine if there are
5237 cgroup_taskset_for_each_leader(leader, css, tset) {
5240 memcg = mem_cgroup_from_css(css);
5246 * We are now commited to this value whatever it is. Changes in this
5247 * tunable will only affect upcoming migrations, not the current one.
5248 * So we need to save it, and keep it going.
5250 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5254 from = mem_cgroup_from_task(p);
5256 VM_BUG_ON(from == memcg);
5258 mm = get_task_mm(p);
5261 /* We move charges only when we move a owner of the mm */
5262 if (mm->owner == p) {
5265 VM_BUG_ON(mc.precharge);
5266 VM_BUG_ON(mc.moved_charge);
5267 VM_BUG_ON(mc.moved_swap);
5269 spin_lock(&mc.lock);
5273 mc.flags = move_flags;
5274 spin_unlock(&mc.lock);
5275 /* We set mc.moving_task later */
5277 ret = mem_cgroup_precharge_mc(mm);
5279 mem_cgroup_clear_mc();
5286 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5289 mem_cgroup_clear_mc();
5292 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5293 unsigned long addr, unsigned long end,
5294 struct mm_walk *walk)
5297 struct vm_area_struct *vma = walk->vma;
5300 enum mc_target_type target_type;
5301 union mc_target target;
5304 ptl = pmd_trans_huge_lock(pmd, vma);
5306 if (mc.precharge < HPAGE_PMD_NR) {
5310 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5311 if (target_type == MC_TARGET_PAGE) {
5313 if (!isolate_lru_page(page)) {
5314 if (!mem_cgroup_move_account(page, true,
5316 mc.precharge -= HPAGE_PMD_NR;
5317 mc.moved_charge += HPAGE_PMD_NR;
5319 putback_lru_page(page);
5322 } else if (target_type == MC_TARGET_DEVICE) {
5324 if (!mem_cgroup_move_account(page, true,
5326 mc.precharge -= HPAGE_PMD_NR;
5327 mc.moved_charge += HPAGE_PMD_NR;
5335 if (pmd_trans_unstable(pmd))
5338 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5339 for (; addr != end; addr += PAGE_SIZE) {
5340 pte_t ptent = *(pte++);
5341 bool device = false;
5347 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5348 case MC_TARGET_DEVICE:
5351 case MC_TARGET_PAGE:
5354 * We can have a part of the split pmd here. Moving it
5355 * can be done but it would be too convoluted so simply
5356 * ignore such a partial THP and keep it in original
5357 * memcg. There should be somebody mapping the head.
5359 if (PageTransCompound(page))
5361 if (!device && isolate_lru_page(page))
5363 if (!mem_cgroup_move_account(page, false,
5366 /* we uncharge from mc.from later. */
5370 putback_lru_page(page);
5371 put: /* get_mctgt_type() gets the page */
5374 case MC_TARGET_SWAP:
5376 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5378 /* we fixup refcnts and charges later. */
5386 pte_unmap_unlock(pte - 1, ptl);
5391 * We have consumed all precharges we got in can_attach().
5392 * We try charge one by one, but don't do any additional
5393 * charges to mc.to if we have failed in charge once in attach()
5396 ret = mem_cgroup_do_precharge(1);
5404 static void mem_cgroup_move_charge(void)
5406 struct mm_walk mem_cgroup_move_charge_walk = {
5407 .pmd_entry = mem_cgroup_move_charge_pte_range,
5411 lru_add_drain_all();
5413 * Signal lock_page_memcg() to take the memcg's move_lock
5414 * while we're moving its pages to another memcg. Then wait
5415 * for already started RCU-only updates to finish.
5417 atomic_inc(&mc.from->moving_account);
5420 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5422 * Someone who are holding the mmap_sem might be waiting in
5423 * waitq. So we cancel all extra charges, wake up all waiters,
5424 * and retry. Because we cancel precharges, we might not be able
5425 * to move enough charges, but moving charge is a best-effort
5426 * feature anyway, so it wouldn't be a big problem.
5428 __mem_cgroup_clear_mc();
5433 * When we have consumed all precharges and failed in doing
5434 * additional charge, the page walk just aborts.
5436 walk_page_range(0, mc.mm->highest_vm_end, &mem_cgroup_move_charge_walk);
5438 up_read(&mc.mm->mmap_sem);
5439 atomic_dec(&mc.from->moving_account);
5442 static void mem_cgroup_move_task(void)
5445 mem_cgroup_move_charge();
5446 mem_cgroup_clear_mc();
5449 #else /* !CONFIG_MMU */
5450 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5454 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5457 static void mem_cgroup_move_task(void)
5463 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5464 * to verify whether we're attached to the default hierarchy on each mount
5467 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5470 * use_hierarchy is forced on the default hierarchy. cgroup core
5471 * guarantees that @root doesn't have any children, so turning it
5472 * on for the root memcg is enough.
5474 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5475 root_mem_cgroup->use_hierarchy = true;
5477 root_mem_cgroup->use_hierarchy = false;
5480 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
5482 if (value == PAGE_COUNTER_MAX)
5483 seq_puts(m, "max\n");
5485 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
5490 static u64 memory_current_read(struct cgroup_subsys_state *css,
5493 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5495 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5498 static int memory_min_show(struct seq_file *m, void *v)
5500 return seq_puts_memcg_tunable(m,
5501 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
5504 static ssize_t memory_min_write(struct kernfs_open_file *of,
5505 char *buf, size_t nbytes, loff_t off)
5507 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5511 buf = strstrip(buf);
5512 err = page_counter_memparse(buf, "max", &min);
5516 page_counter_set_min(&memcg->memory, min);
5521 static int memory_low_show(struct seq_file *m, void *v)
5523 return seq_puts_memcg_tunable(m,
5524 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
5527 static ssize_t memory_low_write(struct kernfs_open_file *of,
5528 char *buf, size_t nbytes, loff_t off)
5530 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5534 buf = strstrip(buf);
5535 err = page_counter_memparse(buf, "max", &low);
5539 page_counter_set_low(&memcg->memory, low);
5544 static int memory_high_show(struct seq_file *m, void *v)
5546 return seq_puts_memcg_tunable(m, READ_ONCE(mem_cgroup_from_seq(m)->high));
5549 static ssize_t memory_high_write(struct kernfs_open_file *of,
5550 char *buf, size_t nbytes, loff_t off)
5552 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5553 unsigned long nr_pages;
5557 buf = strstrip(buf);
5558 err = page_counter_memparse(buf, "max", &high);
5564 nr_pages = page_counter_read(&memcg->memory);
5565 if (nr_pages > high)
5566 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5569 memcg_wb_domain_size_changed(memcg);
5573 static int memory_max_show(struct seq_file *m, void *v)
5575 return seq_puts_memcg_tunable(m,
5576 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
5579 static ssize_t memory_max_write(struct kernfs_open_file *of,
5580 char *buf, size_t nbytes, loff_t off)
5582 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5583 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5584 bool drained = false;
5588 buf = strstrip(buf);
5589 err = page_counter_memparse(buf, "max", &max);
5593 xchg(&memcg->memory.max, max);
5596 unsigned long nr_pages = page_counter_read(&memcg->memory);
5598 if (nr_pages <= max)
5601 if (signal_pending(current)) {
5607 drain_all_stock(memcg);
5613 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5619 memcg_memory_event(memcg, MEMCG_OOM);
5620 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5624 memcg_wb_domain_size_changed(memcg);
5628 static int memory_events_show(struct seq_file *m, void *v)
5630 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5632 seq_printf(m, "low %lu\n",
5633 atomic_long_read(&memcg->memory_events[MEMCG_LOW]));
5634 seq_printf(m, "high %lu\n",
5635 atomic_long_read(&memcg->memory_events[MEMCG_HIGH]));
5636 seq_printf(m, "max %lu\n",
5637 atomic_long_read(&memcg->memory_events[MEMCG_MAX]));
5638 seq_printf(m, "oom %lu\n",
5639 atomic_long_read(&memcg->memory_events[MEMCG_OOM]));
5640 seq_printf(m, "oom_kill %lu\n",
5641 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
5646 static int memory_stat_show(struct seq_file *m, void *v)
5648 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5649 struct accumulated_vmstats acc;
5653 * Provide statistics on the state of the memory subsystem as
5654 * well as cumulative event counters that show past behavior.
5656 * This list is ordered following a combination of these gradients:
5657 * 1) generic big picture -> specifics and details
5658 * 2) reflecting userspace activity -> reflecting kernel heuristics
5660 * Current memory state:
5663 memset(&acc, 0, sizeof(acc));
5664 acc.vmstats_size = MEMCG_NR_STAT;
5665 acc.vmevents_size = NR_VM_EVENT_ITEMS;
5666 accumulate_vmstats(memcg, &acc);
5668 seq_printf(m, "anon %llu\n",
5669 (u64)acc.vmstats[MEMCG_RSS] * PAGE_SIZE);
5670 seq_printf(m, "file %llu\n",
5671 (u64)acc.vmstats[MEMCG_CACHE] * PAGE_SIZE);
5672 seq_printf(m, "kernel_stack %llu\n",
5673 (u64)acc.vmstats[MEMCG_KERNEL_STACK_KB] * 1024);
5674 seq_printf(m, "slab %llu\n",
5675 (u64)(acc.vmstats[NR_SLAB_RECLAIMABLE] +
5676 acc.vmstats[NR_SLAB_UNRECLAIMABLE]) * PAGE_SIZE);
5677 seq_printf(m, "sock %llu\n",
5678 (u64)acc.vmstats[MEMCG_SOCK] * PAGE_SIZE);
5680 seq_printf(m, "shmem %llu\n",
5681 (u64)acc.vmstats[NR_SHMEM] * PAGE_SIZE);
5682 seq_printf(m, "file_mapped %llu\n",
5683 (u64)acc.vmstats[NR_FILE_MAPPED] * PAGE_SIZE);
5684 seq_printf(m, "file_dirty %llu\n",
5685 (u64)acc.vmstats[NR_FILE_DIRTY] * PAGE_SIZE);
5686 seq_printf(m, "file_writeback %llu\n",
5687 (u64)acc.vmstats[NR_WRITEBACK] * PAGE_SIZE);
5690 * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter
5691 * with the NR_ANON_THP vm counter, but right now it's a pain in the
5692 * arse because it requires migrating the work out of rmap to a place
5693 * where the page->mem_cgroup is set up and stable.
5695 seq_printf(m, "anon_thp %llu\n",
5696 (u64)acc.vmstats[MEMCG_RSS_HUGE] * PAGE_SIZE);
5698 for (i = 0; i < NR_LRU_LISTS; i++)
5699 seq_printf(m, "%s %llu\n", mem_cgroup_lru_names[i],
5700 (u64)acc.lru_pages[i] * PAGE_SIZE);
5702 seq_printf(m, "slab_reclaimable %llu\n",
5703 (u64)acc.vmstats[NR_SLAB_RECLAIMABLE] * PAGE_SIZE);
5704 seq_printf(m, "slab_unreclaimable %llu\n",
5705 (u64)acc.vmstats[NR_SLAB_UNRECLAIMABLE] * PAGE_SIZE);
5707 /* Accumulated memory events */
5709 seq_printf(m, "pgfault %lu\n", acc.vmevents[PGFAULT]);
5710 seq_printf(m, "pgmajfault %lu\n", acc.vmevents[PGMAJFAULT]);
5712 seq_printf(m, "workingset_refault %lu\n",
5713 acc.vmstats[WORKINGSET_REFAULT]);
5714 seq_printf(m, "workingset_activate %lu\n",
5715 acc.vmstats[WORKINGSET_ACTIVATE]);
5716 seq_printf(m, "workingset_nodereclaim %lu\n",
5717 acc.vmstats[WORKINGSET_NODERECLAIM]);
5719 seq_printf(m, "pgrefill %lu\n", acc.vmevents[PGREFILL]);
5720 seq_printf(m, "pgscan %lu\n", acc.vmevents[PGSCAN_KSWAPD] +
5721 acc.vmevents[PGSCAN_DIRECT]);
5722 seq_printf(m, "pgsteal %lu\n", acc.vmevents[PGSTEAL_KSWAPD] +
5723 acc.vmevents[PGSTEAL_DIRECT]);
5724 seq_printf(m, "pgactivate %lu\n", acc.vmevents[PGACTIVATE]);
5725 seq_printf(m, "pgdeactivate %lu\n", acc.vmevents[PGDEACTIVATE]);
5726 seq_printf(m, "pglazyfree %lu\n", acc.vmevents[PGLAZYFREE]);
5727 seq_printf(m, "pglazyfreed %lu\n", acc.vmevents[PGLAZYFREED]);
5729 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5730 seq_printf(m, "thp_fault_alloc %lu\n", acc.vmevents[THP_FAULT_ALLOC]);
5731 seq_printf(m, "thp_collapse_alloc %lu\n",
5732 acc.vmevents[THP_COLLAPSE_ALLOC]);
5733 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
5738 static int memory_oom_group_show(struct seq_file *m, void *v)
5740 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5742 seq_printf(m, "%d\n", memcg->oom_group);
5747 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
5748 char *buf, size_t nbytes, loff_t off)
5750 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5753 buf = strstrip(buf);
5757 ret = kstrtoint(buf, 0, &oom_group);
5761 if (oom_group != 0 && oom_group != 1)
5764 memcg->oom_group = oom_group;
5769 static struct cftype memory_files[] = {
5772 .flags = CFTYPE_NOT_ON_ROOT,
5773 .read_u64 = memory_current_read,
5777 .flags = CFTYPE_NOT_ON_ROOT,
5778 .seq_show = memory_min_show,
5779 .write = memory_min_write,
5783 .flags = CFTYPE_NOT_ON_ROOT,
5784 .seq_show = memory_low_show,
5785 .write = memory_low_write,
5789 .flags = CFTYPE_NOT_ON_ROOT,
5790 .seq_show = memory_high_show,
5791 .write = memory_high_write,
5795 .flags = CFTYPE_NOT_ON_ROOT,
5796 .seq_show = memory_max_show,
5797 .write = memory_max_write,
5801 .flags = CFTYPE_NOT_ON_ROOT,
5802 .file_offset = offsetof(struct mem_cgroup, events_file),
5803 .seq_show = memory_events_show,
5807 .flags = CFTYPE_NOT_ON_ROOT,
5808 .seq_show = memory_stat_show,
5811 .name = "oom.group",
5812 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
5813 .seq_show = memory_oom_group_show,
5814 .write = memory_oom_group_write,
5819 struct cgroup_subsys memory_cgrp_subsys = {
5820 .css_alloc = mem_cgroup_css_alloc,
5821 .css_online = mem_cgroup_css_online,
5822 .css_offline = mem_cgroup_css_offline,
5823 .css_released = mem_cgroup_css_released,
5824 .css_free = mem_cgroup_css_free,
5825 .css_reset = mem_cgroup_css_reset,
5826 .can_attach = mem_cgroup_can_attach,
5827 .cancel_attach = mem_cgroup_cancel_attach,
5828 .post_attach = mem_cgroup_move_task,
5829 .bind = mem_cgroup_bind,
5830 .dfl_cftypes = memory_files,
5831 .legacy_cftypes = mem_cgroup_legacy_files,
5836 * mem_cgroup_protected - check if memory consumption is in the normal range
5837 * @root: the top ancestor of the sub-tree being checked
5838 * @memcg: the memory cgroup to check
5840 * WARNING: This function is not stateless! It can only be used as part
5841 * of a top-down tree iteration, not for isolated queries.
5843 * Returns one of the following:
5844 * MEMCG_PROT_NONE: cgroup memory is not protected
5845 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
5846 * an unprotected supply of reclaimable memory from other cgroups.
5847 * MEMCG_PROT_MIN: cgroup memory is protected
5849 * @root is exclusive; it is never protected when looked at directly
5851 * To provide a proper hierarchical behavior, effective memory.min/low values
5852 * are used. Below is the description of how effective memory.low is calculated.
5853 * Effective memory.min values is calculated in the same way.
5855 * Effective memory.low is always equal or less than the original memory.low.
5856 * If there is no memory.low overcommittment (which is always true for
5857 * top-level memory cgroups), these two values are equal.
5858 * Otherwise, it's a part of parent's effective memory.low,
5859 * calculated as a cgroup's memory.low usage divided by sum of sibling's
5860 * memory.low usages, where memory.low usage is the size of actually
5864 * elow = min( memory.low, parent->elow * ------------------ ),
5865 * siblings_low_usage
5867 * | memory.current, if memory.current < memory.low
5872 * Such definition of the effective memory.low provides the expected
5873 * hierarchical behavior: parent's memory.low value is limiting
5874 * children, unprotected memory is reclaimed first and cgroups,
5875 * which are not using their guarantee do not affect actual memory
5878 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
5880 * A A/memory.low = 2G, A/memory.current = 6G
5882 * BC DE B/memory.low = 3G B/memory.current = 2G
5883 * C/memory.low = 1G C/memory.current = 2G
5884 * D/memory.low = 0 D/memory.current = 2G
5885 * E/memory.low = 10G E/memory.current = 0
5887 * and the memory pressure is applied, the following memory distribution
5888 * is expected (approximately):
5890 * A/memory.current = 2G
5892 * B/memory.current = 1.3G
5893 * C/memory.current = 0.6G
5894 * D/memory.current = 0
5895 * E/memory.current = 0
5897 * These calculations require constant tracking of the actual low usages
5898 * (see propagate_protected_usage()), as well as recursive calculation of
5899 * effective memory.low values. But as we do call mem_cgroup_protected()
5900 * path for each memory cgroup top-down from the reclaim,
5901 * it's possible to optimize this part, and save calculated elow
5902 * for next usage. This part is intentionally racy, but it's ok,
5903 * as memory.low is a best-effort mechanism.
5905 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
5906 struct mem_cgroup *memcg)
5908 struct mem_cgroup *parent;
5909 unsigned long emin, parent_emin;
5910 unsigned long elow, parent_elow;
5911 unsigned long usage;
5913 if (mem_cgroup_disabled())
5914 return MEMCG_PROT_NONE;
5917 root = root_mem_cgroup;
5919 return MEMCG_PROT_NONE;
5921 usage = page_counter_read(&memcg->memory);
5923 return MEMCG_PROT_NONE;
5925 emin = memcg->memory.min;
5926 elow = memcg->memory.low;
5928 parent = parent_mem_cgroup(memcg);
5929 /* No parent means a non-hierarchical mode on v1 memcg */
5931 return MEMCG_PROT_NONE;
5936 parent_emin = READ_ONCE(parent->memory.emin);
5937 emin = min(emin, parent_emin);
5938 if (emin && parent_emin) {
5939 unsigned long min_usage, siblings_min_usage;
5941 min_usage = min(usage, memcg->memory.min);
5942 siblings_min_usage = atomic_long_read(
5943 &parent->memory.children_min_usage);
5945 if (min_usage && siblings_min_usage)
5946 emin = min(emin, parent_emin * min_usage /
5947 siblings_min_usage);
5950 parent_elow = READ_ONCE(parent->memory.elow);
5951 elow = min(elow, parent_elow);
5952 if (elow && parent_elow) {
5953 unsigned long low_usage, siblings_low_usage;
5955 low_usage = min(usage, memcg->memory.low);
5956 siblings_low_usage = atomic_long_read(
5957 &parent->memory.children_low_usage);
5959 if (low_usage && siblings_low_usage)
5960 elow = min(elow, parent_elow * low_usage /
5961 siblings_low_usage);
5965 memcg->memory.emin = emin;
5966 memcg->memory.elow = elow;
5969 return MEMCG_PROT_MIN;
5970 else if (usage <= elow)
5971 return MEMCG_PROT_LOW;
5973 return MEMCG_PROT_NONE;
5977 * mem_cgroup_try_charge - try charging a page
5978 * @page: page to charge
5979 * @mm: mm context of the victim
5980 * @gfp_mask: reclaim mode
5981 * @memcgp: charged memcg return
5982 * @compound: charge the page as compound or small page
5984 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5985 * pages according to @gfp_mask if necessary.
5987 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5988 * Otherwise, an error code is returned.
5990 * After page->mapping has been set up, the caller must finalize the
5991 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5992 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5994 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5995 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5998 struct mem_cgroup *memcg = NULL;
5999 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6002 if (mem_cgroup_disabled())
6005 if (PageSwapCache(page)) {
6007 * Every swap fault against a single page tries to charge the
6008 * page, bail as early as possible. shmem_unuse() encounters
6009 * already charged pages, too. The USED bit is protected by
6010 * the page lock, which serializes swap cache removal, which
6011 * in turn serializes uncharging.
6013 VM_BUG_ON_PAGE(!PageLocked(page), page);
6014 if (compound_head(page)->mem_cgroup)
6017 if (do_swap_account) {
6018 swp_entry_t ent = { .val = page_private(page), };
6019 unsigned short id = lookup_swap_cgroup_id(ent);
6022 memcg = mem_cgroup_from_id(id);
6023 if (memcg && !css_tryget_online(&memcg->css))
6030 memcg = get_mem_cgroup_from_mm(mm);
6032 ret = try_charge(memcg, gfp_mask, nr_pages);
6034 css_put(&memcg->css);
6040 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
6041 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6044 struct mem_cgroup *memcg;
6047 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound);
6049 mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask);
6054 * mem_cgroup_commit_charge - commit a page charge
6055 * @page: page to charge
6056 * @memcg: memcg to charge the page to
6057 * @lrucare: page might be on LRU already
6058 * @compound: charge the page as compound or small page
6060 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6061 * after page->mapping has been set up. This must happen atomically
6062 * as part of the page instantiation, i.e. under the page table lock
6063 * for anonymous pages, under the page lock for page and swap cache.
6065 * In addition, the page must not be on the LRU during the commit, to
6066 * prevent racing with task migration. If it might be, use @lrucare.
6068 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6070 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
6071 bool lrucare, bool compound)
6073 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6075 VM_BUG_ON_PAGE(!page->mapping, page);
6076 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
6078 if (mem_cgroup_disabled())
6081 * Swap faults will attempt to charge the same page multiple
6082 * times. But reuse_swap_page() might have removed the page
6083 * from swapcache already, so we can't check PageSwapCache().
6088 commit_charge(page, memcg, lrucare);
6090 local_irq_disable();
6091 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
6092 memcg_check_events(memcg, page);
6095 if (do_memsw_account() && PageSwapCache(page)) {
6096 swp_entry_t entry = { .val = page_private(page) };
6098 * The swap entry might not get freed for a long time,
6099 * let's not wait for it. The page already received a
6100 * memory+swap charge, drop the swap entry duplicate.
6102 mem_cgroup_uncharge_swap(entry, nr_pages);
6107 * mem_cgroup_cancel_charge - cancel a page charge
6108 * @page: page to charge
6109 * @memcg: memcg to charge the page to
6110 * @compound: charge the page as compound or small page
6112 * Cancel a charge transaction started by mem_cgroup_try_charge().
6114 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
6117 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6119 if (mem_cgroup_disabled())
6122 * Swap faults will attempt to charge the same page multiple
6123 * times. But reuse_swap_page() might have removed the page
6124 * from swapcache already, so we can't check PageSwapCache().
6129 cancel_charge(memcg, nr_pages);
6132 struct uncharge_gather {
6133 struct mem_cgroup *memcg;
6134 unsigned long pgpgout;
6135 unsigned long nr_anon;
6136 unsigned long nr_file;
6137 unsigned long nr_kmem;
6138 unsigned long nr_huge;
6139 unsigned long nr_shmem;
6140 struct page *dummy_page;
6143 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6145 memset(ug, 0, sizeof(*ug));
6148 static void uncharge_batch(const struct uncharge_gather *ug)
6150 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
6151 unsigned long flags;
6153 if (!mem_cgroup_is_root(ug->memcg)) {
6154 page_counter_uncharge(&ug->memcg->memory, nr_pages);
6155 if (do_memsw_account())
6156 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
6157 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6158 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6159 memcg_oom_recover(ug->memcg);
6162 local_irq_save(flags);
6163 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
6164 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
6165 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
6166 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
6167 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6168 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, nr_pages);
6169 memcg_check_events(ug->memcg, ug->dummy_page);
6170 local_irq_restore(flags);
6172 if (!mem_cgroup_is_root(ug->memcg))
6173 css_put_many(&ug->memcg->css, nr_pages);
6176 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6178 VM_BUG_ON_PAGE(PageLRU(page), page);
6179 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
6180 !PageHWPoison(page) , page);
6182 if (!page->mem_cgroup)
6186 * Nobody should be changing or seriously looking at
6187 * page->mem_cgroup at this point, we have fully
6188 * exclusive access to the page.
6191 if (ug->memcg != page->mem_cgroup) {
6194 uncharge_gather_clear(ug);
6196 ug->memcg = page->mem_cgroup;
6199 if (!PageKmemcg(page)) {
6200 unsigned int nr_pages = 1;
6202 if (PageTransHuge(page)) {
6203 nr_pages <<= compound_order(page);
6204 ug->nr_huge += nr_pages;
6207 ug->nr_anon += nr_pages;
6209 ug->nr_file += nr_pages;
6210 if (PageSwapBacked(page))
6211 ug->nr_shmem += nr_pages;
6215 ug->nr_kmem += 1 << compound_order(page);
6216 __ClearPageKmemcg(page);
6219 ug->dummy_page = page;
6220 page->mem_cgroup = NULL;
6223 static void uncharge_list(struct list_head *page_list)
6225 struct uncharge_gather ug;
6226 struct list_head *next;
6228 uncharge_gather_clear(&ug);
6231 * Note that the list can be a single page->lru; hence the
6232 * do-while loop instead of a simple list_for_each_entry().
6234 next = page_list->next;
6238 page = list_entry(next, struct page, lru);
6239 next = page->lru.next;
6241 uncharge_page(page, &ug);
6242 } while (next != page_list);
6245 uncharge_batch(&ug);
6249 * mem_cgroup_uncharge - uncharge a page
6250 * @page: page to uncharge
6252 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6253 * mem_cgroup_commit_charge().
6255 void mem_cgroup_uncharge(struct page *page)
6257 struct uncharge_gather ug;
6259 if (mem_cgroup_disabled())
6262 /* Don't touch page->lru of any random page, pre-check: */
6263 if (!page->mem_cgroup)
6266 uncharge_gather_clear(&ug);
6267 uncharge_page(page, &ug);
6268 uncharge_batch(&ug);
6272 * mem_cgroup_uncharge_list - uncharge a list of page
6273 * @page_list: list of pages to uncharge
6275 * Uncharge a list of pages previously charged with
6276 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6278 void mem_cgroup_uncharge_list(struct list_head *page_list)
6280 if (mem_cgroup_disabled())
6283 if (!list_empty(page_list))
6284 uncharge_list(page_list);
6288 * mem_cgroup_migrate - charge a page's replacement
6289 * @oldpage: currently circulating page
6290 * @newpage: replacement page
6292 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6293 * be uncharged upon free.
6295 * Both pages must be locked, @newpage->mapping must be set up.
6297 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6299 struct mem_cgroup *memcg;
6300 unsigned int nr_pages;
6302 unsigned long flags;
6304 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6305 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6306 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6307 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6310 if (mem_cgroup_disabled())
6313 /* Page cache replacement: new page already charged? */
6314 if (newpage->mem_cgroup)
6317 /* Swapcache readahead pages can get replaced before being charged */
6318 memcg = oldpage->mem_cgroup;
6322 /* Force-charge the new page. The old one will be freed soon */
6323 compound = PageTransHuge(newpage);
6324 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
6326 page_counter_charge(&memcg->memory, nr_pages);
6327 if (do_memsw_account())
6328 page_counter_charge(&memcg->memsw, nr_pages);
6329 css_get_many(&memcg->css, nr_pages);
6331 commit_charge(newpage, memcg, false);
6333 local_irq_save(flags);
6334 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
6335 memcg_check_events(memcg, newpage);
6336 local_irq_restore(flags);
6339 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6340 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6342 void mem_cgroup_sk_alloc(struct sock *sk)
6344 struct mem_cgroup *memcg;
6346 if (!mem_cgroup_sockets_enabled)
6350 * Socket cloning can throw us here with sk_memcg already
6351 * filled. It won't however, necessarily happen from
6352 * process context. So the test for root memcg given
6353 * the current task's memcg won't help us in this case.
6355 * Respecting the original socket's memcg is a better
6356 * decision in this case.
6359 css_get(&sk->sk_memcg->css);
6364 memcg = mem_cgroup_from_task(current);
6365 if (memcg == root_mem_cgroup)
6367 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6369 if (css_tryget_online(&memcg->css))
6370 sk->sk_memcg = memcg;
6375 void mem_cgroup_sk_free(struct sock *sk)
6378 css_put(&sk->sk_memcg->css);
6382 * mem_cgroup_charge_skmem - charge socket memory
6383 * @memcg: memcg to charge
6384 * @nr_pages: number of pages to charge
6386 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6387 * @memcg's configured limit, %false if the charge had to be forced.
6389 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6391 gfp_t gfp_mask = GFP_KERNEL;
6393 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6394 struct page_counter *fail;
6396 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6397 memcg->tcpmem_pressure = 0;
6400 page_counter_charge(&memcg->tcpmem, nr_pages);
6401 memcg->tcpmem_pressure = 1;
6405 /* Don't block in the packet receive path */
6407 gfp_mask = GFP_NOWAIT;
6409 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6411 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6414 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6419 * mem_cgroup_uncharge_skmem - uncharge socket memory
6420 * @memcg: memcg to uncharge
6421 * @nr_pages: number of pages to uncharge
6423 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6425 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6426 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6430 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6432 refill_stock(memcg, nr_pages);
6435 static int __init cgroup_memory(char *s)
6439 while ((token = strsep(&s, ",")) != NULL) {
6442 if (!strcmp(token, "nosocket"))
6443 cgroup_memory_nosocket = true;
6444 if (!strcmp(token, "nokmem"))
6445 cgroup_memory_nokmem = true;
6449 __setup("cgroup.memory=", cgroup_memory);
6452 * subsys_initcall() for memory controller.
6454 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6455 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6456 * basically everything that doesn't depend on a specific mem_cgroup structure
6457 * should be initialized from here.
6459 static int __init mem_cgroup_init(void)
6463 #ifdef CONFIG_MEMCG_KMEM
6465 * Kmem cache creation is mostly done with the slab_mutex held,
6466 * so use a workqueue with limited concurrency to avoid stalling
6467 * all worker threads in case lots of cgroups are created and
6468 * destroyed simultaneously.
6470 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
6471 BUG_ON(!memcg_kmem_cache_wq);
6474 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6475 memcg_hotplug_cpu_dead);
6477 for_each_possible_cpu(cpu)
6478 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
6481 for_each_node(node) {
6482 struct mem_cgroup_tree_per_node *rtpn;
6484 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
6485 node_online(node) ? node : NUMA_NO_NODE);
6487 rtpn->rb_root = RB_ROOT;
6488 rtpn->rb_rightmost = NULL;
6489 spin_lock_init(&rtpn->lock);
6490 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6495 subsys_initcall(mem_cgroup_init);
6497 #ifdef CONFIG_MEMCG_SWAP
6498 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
6500 while (!refcount_inc_not_zero(&memcg->id.ref)) {
6502 * The root cgroup cannot be destroyed, so it's refcount must
6505 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
6509 memcg = parent_mem_cgroup(memcg);
6511 memcg = root_mem_cgroup;
6517 * mem_cgroup_swapout - transfer a memsw charge to swap
6518 * @page: page whose memsw charge to transfer
6519 * @entry: swap entry to move the charge to
6521 * Transfer the memsw charge of @page to @entry.
6523 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6525 struct mem_cgroup *memcg, *swap_memcg;
6526 unsigned int nr_entries;
6527 unsigned short oldid;
6529 VM_BUG_ON_PAGE(PageLRU(page), page);
6530 VM_BUG_ON_PAGE(page_count(page), page);
6532 if (!do_memsw_account())
6535 memcg = page->mem_cgroup;
6537 /* Readahead page, never charged */
6542 * In case the memcg owning these pages has been offlined and doesn't
6543 * have an ID allocated to it anymore, charge the closest online
6544 * ancestor for the swap instead and transfer the memory+swap charge.
6546 swap_memcg = mem_cgroup_id_get_online(memcg);
6547 nr_entries = hpage_nr_pages(page);
6548 /* Get references for the tail pages, too */
6550 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
6551 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
6553 VM_BUG_ON_PAGE(oldid, page);
6554 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
6556 page->mem_cgroup = NULL;
6558 if (!mem_cgroup_is_root(memcg))
6559 page_counter_uncharge(&memcg->memory, nr_entries);
6561 if (memcg != swap_memcg) {
6562 if (!mem_cgroup_is_root(swap_memcg))
6563 page_counter_charge(&swap_memcg->memsw, nr_entries);
6564 page_counter_uncharge(&memcg->memsw, nr_entries);
6568 * Interrupts should be disabled here because the caller holds the
6569 * i_pages lock which is taken with interrupts-off. It is
6570 * important here to have the interrupts disabled because it is the
6571 * only synchronisation we have for updating the per-CPU variables.
6573 VM_BUG_ON(!irqs_disabled());
6574 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
6576 memcg_check_events(memcg, page);
6578 if (!mem_cgroup_is_root(memcg))
6579 css_put_many(&memcg->css, nr_entries);
6583 * mem_cgroup_try_charge_swap - try charging swap space for a page
6584 * @page: page being added to swap
6585 * @entry: swap entry to charge
6587 * Try to charge @page's memcg for the swap space at @entry.
6589 * Returns 0 on success, -ENOMEM on failure.
6591 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
6593 unsigned int nr_pages = hpage_nr_pages(page);
6594 struct page_counter *counter;
6595 struct mem_cgroup *memcg;
6596 unsigned short oldid;
6598 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
6601 memcg = page->mem_cgroup;
6603 /* Readahead page, never charged */
6608 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6612 memcg = mem_cgroup_id_get_online(memcg);
6614 if (!mem_cgroup_is_root(memcg) &&
6615 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
6616 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
6617 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6618 mem_cgroup_id_put(memcg);
6622 /* Get references for the tail pages, too */
6624 mem_cgroup_id_get_many(memcg, nr_pages - 1);
6625 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
6626 VM_BUG_ON_PAGE(oldid, page);
6627 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
6633 * mem_cgroup_uncharge_swap - uncharge swap space
6634 * @entry: swap entry to uncharge
6635 * @nr_pages: the amount of swap space to uncharge
6637 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
6639 struct mem_cgroup *memcg;
6642 if (!do_swap_account)
6645 id = swap_cgroup_record(entry, 0, nr_pages);
6647 memcg = mem_cgroup_from_id(id);
6649 if (!mem_cgroup_is_root(memcg)) {
6650 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6651 page_counter_uncharge(&memcg->swap, nr_pages);
6653 page_counter_uncharge(&memcg->memsw, nr_pages);
6655 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
6656 mem_cgroup_id_put_many(memcg, nr_pages);
6661 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
6663 long nr_swap_pages = get_nr_swap_pages();
6665 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6666 return nr_swap_pages;
6667 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6668 nr_swap_pages = min_t(long, nr_swap_pages,
6669 READ_ONCE(memcg->swap.max) -
6670 page_counter_read(&memcg->swap));
6671 return nr_swap_pages;
6674 bool mem_cgroup_swap_full(struct page *page)
6676 struct mem_cgroup *memcg;
6678 VM_BUG_ON_PAGE(!PageLocked(page), page);
6682 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6685 memcg = page->mem_cgroup;
6689 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6690 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.max)
6696 /* for remember boot option*/
6697 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6698 static int really_do_swap_account __initdata = 1;
6700 static int really_do_swap_account __initdata;
6703 static int __init enable_swap_account(char *s)
6705 if (!strcmp(s, "1"))
6706 really_do_swap_account = 1;
6707 else if (!strcmp(s, "0"))
6708 really_do_swap_account = 0;
6711 __setup("swapaccount=", enable_swap_account);
6713 static u64 swap_current_read(struct cgroup_subsys_state *css,
6716 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6718 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
6721 static int swap_max_show(struct seq_file *m, void *v)
6723 return seq_puts_memcg_tunable(m,
6724 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
6727 static ssize_t swap_max_write(struct kernfs_open_file *of,
6728 char *buf, size_t nbytes, loff_t off)
6730 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6734 buf = strstrip(buf);
6735 err = page_counter_memparse(buf, "max", &max);
6739 xchg(&memcg->swap.max, max);
6744 static int swap_events_show(struct seq_file *m, void *v)
6746 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6748 seq_printf(m, "max %lu\n",
6749 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
6750 seq_printf(m, "fail %lu\n",
6751 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
6756 static struct cftype swap_files[] = {
6758 .name = "swap.current",
6759 .flags = CFTYPE_NOT_ON_ROOT,
6760 .read_u64 = swap_current_read,
6764 .flags = CFTYPE_NOT_ON_ROOT,
6765 .seq_show = swap_max_show,
6766 .write = swap_max_write,
6769 .name = "swap.events",
6770 .flags = CFTYPE_NOT_ON_ROOT,
6771 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
6772 .seq_show = swap_events_show,
6777 static struct cftype memsw_cgroup_files[] = {
6779 .name = "memsw.usage_in_bytes",
6780 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6781 .read_u64 = mem_cgroup_read_u64,
6784 .name = "memsw.max_usage_in_bytes",
6785 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6786 .write = mem_cgroup_reset,
6787 .read_u64 = mem_cgroup_read_u64,
6790 .name = "memsw.limit_in_bytes",
6791 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6792 .write = mem_cgroup_write,
6793 .read_u64 = mem_cgroup_read_u64,
6796 .name = "memsw.failcnt",
6797 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6798 .write = mem_cgroup_reset,
6799 .read_u64 = mem_cgroup_read_u64,
6801 { }, /* terminate */
6804 static int __init mem_cgroup_swap_init(void)
6806 if (!mem_cgroup_disabled() && really_do_swap_account) {
6807 do_swap_account = 1;
6808 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
6810 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6811 memsw_cgroup_files));
6815 subsys_initcall(mem_cgroup_swap_init);
6817 #endif /* CONFIG_MEMCG_SWAP */