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/smp.h>
43 #include <linux/page-flags.h>
44 #include <linux/backing-dev.h>
45 #include <linux/bit_spinlock.h>
46 #include <linux/rcupdate.h>
47 #include <linux/limits.h>
48 #include <linux/export.h>
49 #include <linux/mutex.h>
50 #include <linux/rbtree.h>
51 #include <linux/slab.h>
52 #include <linux/swap.h>
53 #include <linux/swapops.h>
54 #include <linux/spinlock.h>
55 #include <linux/eventfd.h>
56 #include <linux/poll.h>
57 #include <linux/sort.h>
59 #include <linux/seq_file.h>
60 #include <linux/vmpressure.h>
61 #include <linux/mm_inline.h>
62 #include <linux/swap_cgroup.h>
63 #include <linux/cpu.h>
64 #include <linux/oom.h>
65 #include <linux/lockdep.h>
66 #include <linux/file.h>
67 #include <linux/tracehook.h>
73 #include <linux/uaccess.h>
75 #include <trace/events/vmscan.h>
77 struct cgroup_subsys memory_cgrp_subsys __read_mostly
;
78 EXPORT_SYMBOL(memory_cgrp_subsys
);
80 struct mem_cgroup
*root_mem_cgroup __read_mostly
;
82 #define MEM_CGROUP_RECLAIM_RETRIES 5
84 /* Socket memory accounting disabled? */
85 static bool cgroup_memory_nosocket
;
87 /* Kernel memory accounting disabled? */
88 static bool cgroup_memory_nokmem
;
90 /* Whether the swap controller is active */
91 #ifdef CONFIG_MEMCG_SWAP
92 int do_swap_account __read_mostly
;
94 #define do_swap_account 0
97 /* Whether legacy memory+swap accounting is active */
98 static bool do_memsw_account(void)
100 return !cgroup_subsys_on_dfl(memory_cgrp_subsys
) && do_swap_account
;
103 static const char *const mem_cgroup_lru_names
[] = {
111 #define THRESHOLDS_EVENTS_TARGET 128
112 #define SOFTLIMIT_EVENTS_TARGET 1024
113 #define NUMAINFO_EVENTS_TARGET 1024
116 * Cgroups above their limits are maintained in a RB-Tree, independent of
117 * their hierarchy representation
120 struct mem_cgroup_tree_per_node
{
121 struct rb_root rb_root
;
122 struct rb_node
*rb_rightmost
;
126 struct mem_cgroup_tree
{
127 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
130 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
133 struct mem_cgroup_eventfd_list
{
134 struct list_head list
;
135 struct eventfd_ctx
*eventfd
;
139 * cgroup_event represents events which userspace want to receive.
141 struct mem_cgroup_event
{
143 * memcg which the event belongs to.
145 struct mem_cgroup
*memcg
;
147 * eventfd to signal userspace about the event.
149 struct eventfd_ctx
*eventfd
;
151 * Each of these stored in a list by the cgroup.
153 struct list_head list
;
155 * register_event() callback will be used to add new userspace
156 * waiter for changes related to this event. Use eventfd_signal()
157 * on eventfd to send notification to userspace.
159 int (*register_event
)(struct mem_cgroup
*memcg
,
160 struct eventfd_ctx
*eventfd
, const char *args
);
162 * unregister_event() callback will be called when userspace closes
163 * the eventfd or on cgroup removing. This callback must be set,
164 * if you want provide notification functionality.
166 void (*unregister_event
)(struct mem_cgroup
*memcg
,
167 struct eventfd_ctx
*eventfd
);
169 * All fields below needed to unregister event when
170 * userspace closes eventfd.
173 wait_queue_head_t
*wqh
;
174 wait_queue_entry_t wait
;
175 struct work_struct remove
;
178 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
179 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
181 /* Stuffs for move charges at task migration. */
183 * Types of charges to be moved.
185 #define MOVE_ANON 0x1U
186 #define MOVE_FILE 0x2U
187 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
189 /* "mc" and its members are protected by cgroup_mutex */
190 static struct move_charge_struct
{
191 spinlock_t lock
; /* for from, to */
192 struct mm_struct
*mm
;
193 struct mem_cgroup
*from
;
194 struct mem_cgroup
*to
;
196 unsigned long precharge
;
197 unsigned long moved_charge
;
198 unsigned long moved_swap
;
199 struct task_struct
*moving_task
; /* a task moving charges */
200 wait_queue_head_t waitq
; /* a waitq for other context */
202 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
203 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
207 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
208 * limit reclaim to prevent infinite loops, if they ever occur.
210 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
211 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
214 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
215 MEM_CGROUP_CHARGE_TYPE_ANON
,
216 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
217 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
221 /* for encoding cft->private value on file */
230 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
231 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
232 #define MEMFILE_ATTR(val) ((val) & 0xffff)
233 /* Used for OOM nofiier */
234 #define OOM_CONTROL (0)
236 /* Some nice accessors for the vmpressure. */
237 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
240 memcg
= root_mem_cgroup
;
241 return &memcg
->vmpressure
;
244 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
246 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
249 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
251 return (memcg
== root_mem_cgroup
);
256 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
257 * The main reason for not using cgroup id for this:
258 * this works better in sparse environments, where we have a lot of memcgs,
259 * but only a few kmem-limited. Or also, if we have, for instance, 200
260 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
261 * 200 entry array for that.
263 * The current size of the caches array is stored in memcg_nr_cache_ids. It
264 * will double each time we have to increase it.
266 static DEFINE_IDA(memcg_cache_ida
);
267 int memcg_nr_cache_ids
;
269 /* Protects memcg_nr_cache_ids */
270 static DECLARE_RWSEM(memcg_cache_ids_sem
);
272 void memcg_get_cache_ids(void)
274 down_read(&memcg_cache_ids_sem
);
277 void memcg_put_cache_ids(void)
279 up_read(&memcg_cache_ids_sem
);
283 * MIN_SIZE is different than 1, because we would like to avoid going through
284 * the alloc/free process all the time. In a small machine, 4 kmem-limited
285 * cgroups is a reasonable guess. In the future, it could be a parameter or
286 * tunable, but that is strictly not necessary.
288 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
289 * this constant directly from cgroup, but it is understandable that this is
290 * better kept as an internal representation in cgroup.c. In any case, the
291 * cgrp_id space is not getting any smaller, and we don't have to necessarily
292 * increase ours as well if it increases.
294 #define MEMCG_CACHES_MIN_SIZE 4
295 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
298 * A lot of the calls to the cache allocation functions are expected to be
299 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
300 * conditional to this static branch, we'll have to allow modules that does
301 * kmem_cache_alloc and the such to see this symbol as well
303 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key
);
304 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
306 struct workqueue_struct
*memcg_kmem_cache_wq
;
308 #endif /* !CONFIG_SLOB */
311 * mem_cgroup_css_from_page - css of the memcg associated with a page
312 * @page: page of interest
314 * If memcg is bound to the default hierarchy, css of the memcg associated
315 * with @page is returned. The returned css remains associated with @page
316 * until it is released.
318 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
321 struct cgroup_subsys_state
*mem_cgroup_css_from_page(struct page
*page
)
323 struct mem_cgroup
*memcg
;
325 memcg
= page
->mem_cgroup
;
327 if (!memcg
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
328 memcg
= root_mem_cgroup
;
334 * page_cgroup_ino - return inode number of the memcg a page is charged to
337 * Look up the closest online ancestor of the memory cgroup @page is charged to
338 * and return its inode number or 0 if @page is not charged to any cgroup. It
339 * is safe to call this function without holding a reference to @page.
341 * Note, this function is inherently racy, because there is nothing to prevent
342 * the cgroup inode from getting torn down and potentially reallocated a moment
343 * after page_cgroup_ino() returns, so it only should be used by callers that
344 * do not care (such as procfs interfaces).
346 ino_t
page_cgroup_ino(struct page
*page
)
348 struct mem_cgroup
*memcg
;
349 unsigned long ino
= 0;
352 memcg
= READ_ONCE(page
->mem_cgroup
);
353 while (memcg
&& !(memcg
->css
.flags
& CSS_ONLINE
))
354 memcg
= parent_mem_cgroup(memcg
);
356 ino
= cgroup_ino(memcg
->css
.cgroup
);
361 static struct mem_cgroup_per_node
*
362 mem_cgroup_page_nodeinfo(struct mem_cgroup
*memcg
, struct page
*page
)
364 int nid
= page_to_nid(page
);
366 return memcg
->nodeinfo
[nid
];
369 static struct mem_cgroup_tree_per_node
*
370 soft_limit_tree_node(int nid
)
372 return soft_limit_tree
.rb_tree_per_node
[nid
];
375 static struct mem_cgroup_tree_per_node
*
376 soft_limit_tree_from_page(struct page
*page
)
378 int nid
= page_to_nid(page
);
380 return soft_limit_tree
.rb_tree_per_node
[nid
];
383 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node
*mz
,
384 struct mem_cgroup_tree_per_node
*mctz
,
385 unsigned long new_usage_in_excess
)
387 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
388 struct rb_node
*parent
= NULL
;
389 struct mem_cgroup_per_node
*mz_node
;
390 bool rightmost
= true;
395 mz
->usage_in_excess
= new_usage_in_excess
;
396 if (!mz
->usage_in_excess
)
400 mz_node
= rb_entry(parent
, struct mem_cgroup_per_node
,
402 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
) {
408 * We can't avoid mem cgroups that are over their soft
409 * limit by the same amount
411 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
416 mctz
->rb_rightmost
= &mz
->tree_node
;
418 rb_link_node(&mz
->tree_node
, parent
, p
);
419 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
423 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
424 struct mem_cgroup_tree_per_node
*mctz
)
429 if (&mz
->tree_node
== mctz
->rb_rightmost
)
430 mctz
->rb_rightmost
= rb_prev(&mz
->tree_node
);
432 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
436 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
437 struct mem_cgroup_tree_per_node
*mctz
)
441 spin_lock_irqsave(&mctz
->lock
, flags
);
442 __mem_cgroup_remove_exceeded(mz
, mctz
);
443 spin_unlock_irqrestore(&mctz
->lock
, flags
);
446 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
448 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
449 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
450 unsigned long excess
= 0;
452 if (nr_pages
> soft_limit
)
453 excess
= nr_pages
- soft_limit
;
458 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
460 unsigned long excess
;
461 struct mem_cgroup_per_node
*mz
;
462 struct mem_cgroup_tree_per_node
*mctz
;
464 mctz
= soft_limit_tree_from_page(page
);
468 * Necessary to update all ancestors when hierarchy is used.
469 * because their event counter is not touched.
471 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
472 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
473 excess
= soft_limit_excess(memcg
);
475 * We have to update the tree if mz is on RB-tree or
476 * mem is over its softlimit.
478 if (excess
|| mz
->on_tree
) {
481 spin_lock_irqsave(&mctz
->lock
, flags
);
482 /* if on-tree, remove it */
484 __mem_cgroup_remove_exceeded(mz
, mctz
);
486 * Insert again. mz->usage_in_excess will be updated.
487 * If excess is 0, no tree ops.
489 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
490 spin_unlock_irqrestore(&mctz
->lock
, flags
);
495 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
497 struct mem_cgroup_tree_per_node
*mctz
;
498 struct mem_cgroup_per_node
*mz
;
502 mz
= mem_cgroup_nodeinfo(memcg
, nid
);
503 mctz
= soft_limit_tree_node(nid
);
505 mem_cgroup_remove_exceeded(mz
, mctz
);
509 static struct mem_cgroup_per_node
*
510 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
512 struct mem_cgroup_per_node
*mz
;
516 if (!mctz
->rb_rightmost
)
517 goto done
; /* Nothing to reclaim from */
519 mz
= rb_entry(mctz
->rb_rightmost
,
520 struct mem_cgroup_per_node
, tree_node
);
522 * Remove the node now but someone else can add it back,
523 * we will to add it back at the end of reclaim to its correct
524 * position in the tree.
526 __mem_cgroup_remove_exceeded(mz
, mctz
);
527 if (!soft_limit_excess(mz
->memcg
) ||
528 !css_tryget_online(&mz
->memcg
->css
))
534 static struct mem_cgroup_per_node
*
535 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
537 struct mem_cgroup_per_node
*mz
;
539 spin_lock_irq(&mctz
->lock
);
540 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
541 spin_unlock_irq(&mctz
->lock
);
545 static unsigned long memcg_sum_events(struct mem_cgroup
*memcg
,
548 return atomic_long_read(&memcg
->events
[event
]);
551 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
553 bool compound
, int nr_pages
)
556 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
557 * counted as CACHE even if it's on ANON LRU.
560 __mod_memcg_state(memcg
, MEMCG_RSS
, nr_pages
);
562 __mod_memcg_state(memcg
, MEMCG_CACHE
, nr_pages
);
563 if (PageSwapBacked(page
))
564 __mod_memcg_state(memcg
, NR_SHMEM
, nr_pages
);
568 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
569 __mod_memcg_state(memcg
, MEMCG_RSS_HUGE
, nr_pages
);
572 /* pagein of a big page is an event. So, ignore page size */
574 __count_memcg_events(memcg
, PGPGIN
, 1);
576 __count_memcg_events(memcg
, PGPGOUT
, 1);
577 nr_pages
= -nr_pages
; /* for event */
580 __this_cpu_add(memcg
->stat_cpu
->nr_page_events
, nr_pages
);
583 unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
584 int nid
, unsigned int lru_mask
)
586 struct lruvec
*lruvec
= mem_cgroup_lruvec(NODE_DATA(nid
), memcg
);
587 unsigned long nr
= 0;
590 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
593 if (!(BIT(lru
) & lru_mask
))
595 nr
+= mem_cgroup_get_lru_size(lruvec
, lru
);
600 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
601 unsigned int lru_mask
)
603 unsigned long nr
= 0;
606 for_each_node_state(nid
, N_MEMORY
)
607 nr
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
611 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
612 enum mem_cgroup_events_target target
)
614 unsigned long val
, next
;
616 val
= __this_cpu_read(memcg
->stat_cpu
->nr_page_events
);
617 next
= __this_cpu_read(memcg
->stat_cpu
->targets
[target
]);
618 /* from time_after() in jiffies.h */
619 if ((long)(next
- val
) < 0) {
621 case MEM_CGROUP_TARGET_THRESH
:
622 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
624 case MEM_CGROUP_TARGET_SOFTLIMIT
:
625 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
627 case MEM_CGROUP_TARGET_NUMAINFO
:
628 next
= val
+ NUMAINFO_EVENTS_TARGET
;
633 __this_cpu_write(memcg
->stat_cpu
->targets
[target
], next
);
640 * Check events in order.
643 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
645 /* threshold event is triggered in finer grain than soft limit */
646 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
647 MEM_CGROUP_TARGET_THRESH
))) {
649 bool do_numainfo __maybe_unused
;
651 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
652 MEM_CGROUP_TARGET_SOFTLIMIT
);
654 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
655 MEM_CGROUP_TARGET_NUMAINFO
);
657 mem_cgroup_threshold(memcg
);
658 if (unlikely(do_softlimit
))
659 mem_cgroup_update_tree(memcg
, page
);
661 if (unlikely(do_numainfo
))
662 atomic_inc(&memcg
->numainfo_events
);
667 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
670 * mm_update_next_owner() may clear mm->owner to NULL
671 * if it races with swapoff, page migration, etc.
672 * So this can be called with p == NULL.
677 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
679 EXPORT_SYMBOL(mem_cgroup_from_task
);
681 static struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
683 struct mem_cgroup
*memcg
= NULL
;
688 * Page cache insertions can happen withou an
689 * actual mm context, e.g. during disk probing
690 * on boot, loopback IO, acct() writes etc.
693 memcg
= root_mem_cgroup
;
695 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
696 if (unlikely(!memcg
))
697 memcg
= root_mem_cgroup
;
699 } while (!css_tryget_online(&memcg
->css
));
705 * mem_cgroup_iter - iterate over memory cgroup hierarchy
706 * @root: hierarchy root
707 * @prev: previously returned memcg, NULL on first invocation
708 * @reclaim: cookie for shared reclaim walks, NULL for full walks
710 * Returns references to children of the hierarchy below @root, or
711 * @root itself, or %NULL after a full round-trip.
713 * Caller must pass the return value in @prev on subsequent
714 * invocations for reference counting, or use mem_cgroup_iter_break()
715 * to cancel a hierarchy walk before the round-trip is complete.
717 * Reclaimers can specify a node and a priority level in @reclaim to
718 * divide up the memcgs in the hierarchy among all concurrent
719 * reclaimers operating on the same node and priority.
721 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
722 struct mem_cgroup
*prev
,
723 struct mem_cgroup_reclaim_cookie
*reclaim
)
725 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
726 struct cgroup_subsys_state
*css
= NULL
;
727 struct mem_cgroup
*memcg
= NULL
;
728 struct mem_cgroup
*pos
= NULL
;
730 if (mem_cgroup_disabled())
734 root
= root_mem_cgroup
;
736 if (prev
&& !reclaim
)
739 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
748 struct mem_cgroup_per_node
*mz
;
750 mz
= mem_cgroup_nodeinfo(root
, reclaim
->pgdat
->node_id
);
751 iter
= &mz
->iter
[reclaim
->priority
];
753 if (prev
&& reclaim
->generation
!= iter
->generation
)
757 pos
= READ_ONCE(iter
->position
);
758 if (!pos
|| css_tryget(&pos
->css
))
761 * css reference reached zero, so iter->position will
762 * be cleared by ->css_released. However, we should not
763 * rely on this happening soon, because ->css_released
764 * is called from a work queue, and by busy-waiting we
765 * might block it. So we clear iter->position right
768 (void)cmpxchg(&iter
->position
, pos
, NULL
);
776 css
= css_next_descendant_pre(css
, &root
->css
);
779 * Reclaimers share the hierarchy walk, and a
780 * new one might jump in right at the end of
781 * the hierarchy - make sure they see at least
782 * one group and restart from the beginning.
790 * Verify the css and acquire a reference. The root
791 * is provided by the caller, so we know it's alive
792 * and kicking, and don't take an extra reference.
794 memcg
= mem_cgroup_from_css(css
);
796 if (css
== &root
->css
)
807 * The position could have already been updated by a competing
808 * thread, so check that the value hasn't changed since we read
809 * it to avoid reclaiming from the same cgroup twice.
811 (void)cmpxchg(&iter
->position
, pos
, memcg
);
819 reclaim
->generation
= iter
->generation
;
825 if (prev
&& prev
!= root
)
832 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
833 * @root: hierarchy root
834 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
836 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
837 struct mem_cgroup
*prev
)
840 root
= root_mem_cgroup
;
841 if (prev
&& prev
!= root
)
845 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
847 struct mem_cgroup
*memcg
= dead_memcg
;
848 struct mem_cgroup_reclaim_iter
*iter
;
849 struct mem_cgroup_per_node
*mz
;
853 while ((memcg
= parent_mem_cgroup(memcg
))) {
855 mz
= mem_cgroup_nodeinfo(memcg
, nid
);
856 for (i
= 0; i
<= DEF_PRIORITY
; i
++) {
858 cmpxchg(&iter
->position
,
866 * Iteration constructs for visiting all cgroups (under a tree). If
867 * loops are exited prematurely (break), mem_cgroup_iter_break() must
868 * be used for reference counting.
870 #define for_each_mem_cgroup_tree(iter, root) \
871 for (iter = mem_cgroup_iter(root, NULL, NULL); \
873 iter = mem_cgroup_iter(root, iter, NULL))
875 #define for_each_mem_cgroup(iter) \
876 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
878 iter = mem_cgroup_iter(NULL, iter, NULL))
881 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
882 * @memcg: hierarchy root
883 * @fn: function to call for each task
884 * @arg: argument passed to @fn
886 * This function iterates over tasks attached to @memcg or to any of its
887 * descendants and calls @fn for each task. If @fn returns a non-zero
888 * value, the function breaks the iteration loop and returns the value.
889 * Otherwise, it will iterate over all tasks and return 0.
891 * This function must not be called for the root memory cgroup.
893 int mem_cgroup_scan_tasks(struct mem_cgroup
*memcg
,
894 int (*fn
)(struct task_struct
*, void *), void *arg
)
896 struct mem_cgroup
*iter
;
899 BUG_ON(memcg
== root_mem_cgroup
);
901 for_each_mem_cgroup_tree(iter
, memcg
) {
902 struct css_task_iter it
;
903 struct task_struct
*task
;
905 css_task_iter_start(&iter
->css
, 0, &it
);
906 while (!ret
&& (task
= css_task_iter_next(&it
)))
908 css_task_iter_end(&it
);
910 mem_cgroup_iter_break(memcg
, iter
);
918 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
920 * @pgdat: pgdat of the page
922 * This function is only safe when following the LRU page isolation
923 * and putback protocol: the LRU lock must be held, and the page must
924 * either be PageLRU() or the caller must have isolated/allocated it.
926 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct pglist_data
*pgdat
)
928 struct mem_cgroup_per_node
*mz
;
929 struct mem_cgroup
*memcg
;
930 struct lruvec
*lruvec
;
932 if (mem_cgroup_disabled()) {
933 lruvec
= &pgdat
->lruvec
;
937 memcg
= page
->mem_cgroup
;
939 * Swapcache readahead pages are added to the LRU - and
940 * possibly migrated - before they are charged.
943 memcg
= root_mem_cgroup
;
945 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
946 lruvec
= &mz
->lruvec
;
949 * Since a node can be onlined after the mem_cgroup was created,
950 * we have to be prepared to initialize lruvec->zone here;
951 * and if offlined then reonlined, we need to reinitialize it.
953 if (unlikely(lruvec
->pgdat
!= pgdat
))
954 lruvec
->pgdat
= pgdat
;
959 * mem_cgroup_update_lru_size - account for adding or removing an lru page
960 * @lruvec: mem_cgroup per zone lru vector
961 * @lru: index of lru list the page is sitting on
962 * @zid: zone id of the accounted pages
963 * @nr_pages: positive when adding or negative when removing
965 * This function must be called under lru_lock, just before a page is added
966 * to or just after a page is removed from an lru list (that ordering being
967 * so as to allow it to check that lru_size 0 is consistent with list_empty).
969 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
970 int zid
, int nr_pages
)
972 struct mem_cgroup_per_node
*mz
;
973 unsigned long *lru_size
;
976 if (mem_cgroup_disabled())
979 mz
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
980 lru_size
= &mz
->lru_zone_size
[zid
][lru
];
983 *lru_size
+= nr_pages
;
986 if (WARN_ONCE(size
< 0,
987 "%s(%p, %d, %d): lru_size %ld\n",
988 __func__
, lruvec
, lru
, nr_pages
, size
)) {
994 *lru_size
+= nr_pages
;
997 bool task_in_mem_cgroup(struct task_struct
*task
, struct mem_cgroup
*memcg
)
999 struct mem_cgroup
*task_memcg
;
1000 struct task_struct
*p
;
1003 p
= find_lock_task_mm(task
);
1005 task_memcg
= get_mem_cgroup_from_mm(p
->mm
);
1009 * All threads may have already detached their mm's, but the oom
1010 * killer still needs to detect if they have already been oom
1011 * killed to prevent needlessly killing additional tasks.
1014 task_memcg
= mem_cgroup_from_task(task
);
1015 css_get(&task_memcg
->css
);
1018 ret
= mem_cgroup_is_descendant(task_memcg
, memcg
);
1019 css_put(&task_memcg
->css
);
1024 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1025 * @memcg: the memory cgroup
1027 * Returns the maximum amount of memory @mem can be charged with, in
1030 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1032 unsigned long margin
= 0;
1033 unsigned long count
;
1034 unsigned long limit
;
1036 count
= page_counter_read(&memcg
->memory
);
1037 limit
= READ_ONCE(memcg
->memory
.max
);
1039 margin
= limit
- count
;
1041 if (do_memsw_account()) {
1042 count
= page_counter_read(&memcg
->memsw
);
1043 limit
= READ_ONCE(memcg
->memsw
.max
);
1045 margin
= min(margin
, limit
- count
);
1054 * A routine for checking "mem" is under move_account() or not.
1056 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1057 * moving cgroups. This is for waiting at high-memory pressure
1060 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1062 struct mem_cgroup
*from
;
1063 struct mem_cgroup
*to
;
1066 * Unlike task_move routines, we access mc.to, mc.from not under
1067 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1069 spin_lock(&mc
.lock
);
1075 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1076 mem_cgroup_is_descendant(to
, memcg
);
1078 spin_unlock(&mc
.lock
);
1082 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1084 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1085 if (mem_cgroup_under_move(memcg
)) {
1087 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1088 /* moving charge context might have finished. */
1091 finish_wait(&mc
.waitq
, &wait
);
1098 static const unsigned int memcg1_stats
[] = {
1109 static const char *const memcg1_stat_names
[] = {
1120 #define K(x) ((x) << (PAGE_SHIFT-10))
1122 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1123 * @memcg: The memory cgroup that went over limit
1124 * @p: Task that is going to be killed
1126 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1129 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1131 struct mem_cgroup
*iter
;
1137 pr_info("Task in ");
1138 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1139 pr_cont(" killed as a result of limit of ");
1141 pr_info("Memory limit reached of cgroup ");
1144 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1149 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1150 K((u64
)page_counter_read(&memcg
->memory
)),
1151 K((u64
)memcg
->memory
.max
), memcg
->memory
.failcnt
);
1152 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1153 K((u64
)page_counter_read(&memcg
->memsw
)),
1154 K((u64
)memcg
->memsw
.max
), memcg
->memsw
.failcnt
);
1155 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1156 K((u64
)page_counter_read(&memcg
->kmem
)),
1157 K((u64
)memcg
->kmem
.max
), memcg
->kmem
.failcnt
);
1159 for_each_mem_cgroup_tree(iter
, memcg
) {
1160 pr_info("Memory cgroup stats for ");
1161 pr_cont_cgroup_path(iter
->css
.cgroup
);
1164 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
1165 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_swap_account
)
1167 pr_cont(" %s:%luKB", memcg1_stat_names
[i
],
1168 K(memcg_page_state(iter
, memcg1_stats
[i
])));
1171 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1172 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1173 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1180 * Return the memory (and swap, if configured) limit for a memcg.
1182 unsigned long mem_cgroup_get_max(struct mem_cgroup
*memcg
)
1186 max
= memcg
->memory
.max
;
1187 if (mem_cgroup_swappiness(memcg
)) {
1188 unsigned long memsw_max
;
1189 unsigned long swap_max
;
1191 memsw_max
= memcg
->memsw
.max
;
1192 swap_max
= memcg
->swap
.max
;
1193 swap_max
= min(swap_max
, (unsigned long)total_swap_pages
);
1194 max
= min(max
+ swap_max
, memsw_max
);
1199 static bool mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1202 struct oom_control oc
= {
1206 .gfp_mask
= gfp_mask
,
1211 mutex_lock(&oom_lock
);
1212 ret
= out_of_memory(&oc
);
1213 mutex_unlock(&oom_lock
);
1217 #if MAX_NUMNODES > 1
1220 * test_mem_cgroup_node_reclaimable
1221 * @memcg: the target memcg
1222 * @nid: the node ID to be checked.
1223 * @noswap : specify true here if the user wants flle only information.
1225 * This function returns whether the specified memcg contains any
1226 * reclaimable pages on a node. Returns true if there are any reclaimable
1227 * pages in the node.
1229 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1230 int nid
, bool noswap
)
1232 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1234 if (noswap
|| !total_swap_pages
)
1236 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1243 * Always updating the nodemask is not very good - even if we have an empty
1244 * list or the wrong list here, we can start from some node and traverse all
1245 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1248 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1252 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1253 * pagein/pageout changes since the last update.
1255 if (!atomic_read(&memcg
->numainfo_events
))
1257 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1260 /* make a nodemask where this memcg uses memory from */
1261 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1263 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1265 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1266 node_clear(nid
, memcg
->scan_nodes
);
1269 atomic_set(&memcg
->numainfo_events
, 0);
1270 atomic_set(&memcg
->numainfo_updating
, 0);
1274 * Selecting a node where we start reclaim from. Because what we need is just
1275 * reducing usage counter, start from anywhere is O,K. Considering
1276 * memory reclaim from current node, there are pros. and cons.
1278 * Freeing memory from current node means freeing memory from a node which
1279 * we'll use or we've used. So, it may make LRU bad. And if several threads
1280 * hit limits, it will see a contention on a node. But freeing from remote
1281 * node means more costs for memory reclaim because of memory latency.
1283 * Now, we use round-robin. Better algorithm is welcomed.
1285 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1289 mem_cgroup_may_update_nodemask(memcg
);
1290 node
= memcg
->last_scanned_node
;
1292 node
= next_node_in(node
, memcg
->scan_nodes
);
1294 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1295 * last time it really checked all the LRUs due to rate limiting.
1296 * Fallback to the current node in that case for simplicity.
1298 if (unlikely(node
== MAX_NUMNODES
))
1299 node
= numa_node_id();
1301 memcg
->last_scanned_node
= node
;
1305 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1311 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1314 unsigned long *total_scanned
)
1316 struct mem_cgroup
*victim
= NULL
;
1319 unsigned long excess
;
1320 unsigned long nr_scanned
;
1321 struct mem_cgroup_reclaim_cookie reclaim
= {
1326 excess
= soft_limit_excess(root_memcg
);
1329 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1334 * If we have not been able to reclaim
1335 * anything, it might because there are
1336 * no reclaimable pages under this hierarchy
1341 * We want to do more targeted reclaim.
1342 * excess >> 2 is not to excessive so as to
1343 * reclaim too much, nor too less that we keep
1344 * coming back to reclaim from this cgroup
1346 if (total
>= (excess
>> 2) ||
1347 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1352 total
+= mem_cgroup_shrink_node(victim
, gfp_mask
, false,
1353 pgdat
, &nr_scanned
);
1354 *total_scanned
+= nr_scanned
;
1355 if (!soft_limit_excess(root_memcg
))
1358 mem_cgroup_iter_break(root_memcg
, victim
);
1362 #ifdef CONFIG_LOCKDEP
1363 static struct lockdep_map memcg_oom_lock_dep_map
= {
1364 .name
= "memcg_oom_lock",
1368 static DEFINE_SPINLOCK(memcg_oom_lock
);
1371 * Check OOM-Killer is already running under our hierarchy.
1372 * If someone is running, return false.
1374 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1376 struct mem_cgroup
*iter
, *failed
= NULL
;
1378 spin_lock(&memcg_oom_lock
);
1380 for_each_mem_cgroup_tree(iter
, memcg
) {
1381 if (iter
->oom_lock
) {
1383 * this subtree of our hierarchy is already locked
1384 * so we cannot give a lock.
1387 mem_cgroup_iter_break(memcg
, iter
);
1390 iter
->oom_lock
= true;
1395 * OK, we failed to lock the whole subtree so we have
1396 * to clean up what we set up to the failing subtree
1398 for_each_mem_cgroup_tree(iter
, memcg
) {
1399 if (iter
== failed
) {
1400 mem_cgroup_iter_break(memcg
, iter
);
1403 iter
->oom_lock
= false;
1406 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1408 spin_unlock(&memcg_oom_lock
);
1413 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1415 struct mem_cgroup
*iter
;
1417 spin_lock(&memcg_oom_lock
);
1418 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1419 for_each_mem_cgroup_tree(iter
, memcg
)
1420 iter
->oom_lock
= false;
1421 spin_unlock(&memcg_oom_lock
);
1424 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1426 struct mem_cgroup
*iter
;
1428 spin_lock(&memcg_oom_lock
);
1429 for_each_mem_cgroup_tree(iter
, memcg
)
1431 spin_unlock(&memcg_oom_lock
);
1434 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1436 struct mem_cgroup
*iter
;
1439 * When a new child is created while the hierarchy is under oom,
1440 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1442 spin_lock(&memcg_oom_lock
);
1443 for_each_mem_cgroup_tree(iter
, memcg
)
1444 if (iter
->under_oom
> 0)
1446 spin_unlock(&memcg_oom_lock
);
1449 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1451 struct oom_wait_info
{
1452 struct mem_cgroup
*memcg
;
1453 wait_queue_entry_t wait
;
1456 static int memcg_oom_wake_function(wait_queue_entry_t
*wait
,
1457 unsigned mode
, int sync
, void *arg
)
1459 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1460 struct mem_cgroup
*oom_wait_memcg
;
1461 struct oom_wait_info
*oom_wait_info
;
1463 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1464 oom_wait_memcg
= oom_wait_info
->memcg
;
1466 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1467 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1469 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1472 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1475 * For the following lockless ->under_oom test, the only required
1476 * guarantee is that it must see the state asserted by an OOM when
1477 * this function is called as a result of userland actions
1478 * triggered by the notification of the OOM. This is trivially
1479 * achieved by invoking mem_cgroup_mark_under_oom() before
1480 * triggering notification.
1482 if (memcg
&& memcg
->under_oom
)
1483 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1486 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1488 if (!current
->memcg_may_oom
|| order
> PAGE_ALLOC_COSTLY_ORDER
)
1491 * We are in the middle of the charge context here, so we
1492 * don't want to block when potentially sitting on a callstack
1493 * that holds all kinds of filesystem and mm locks.
1495 * Also, the caller may handle a failed allocation gracefully
1496 * (like optional page cache readahead) and so an OOM killer
1497 * invocation might not even be necessary.
1499 * That's why we don't do anything here except remember the
1500 * OOM context and then deal with it at the end of the page
1501 * fault when the stack is unwound, the locks are released,
1502 * and when we know whether the fault was overall successful.
1504 css_get(&memcg
->css
);
1505 current
->memcg_in_oom
= memcg
;
1506 current
->memcg_oom_gfp_mask
= mask
;
1507 current
->memcg_oom_order
= order
;
1511 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1512 * @handle: actually kill/wait or just clean up the OOM state
1514 * This has to be called at the end of a page fault if the memcg OOM
1515 * handler was enabled.
1517 * Memcg supports userspace OOM handling where failed allocations must
1518 * sleep on a waitqueue until the userspace task resolves the
1519 * situation. Sleeping directly in the charge context with all kinds
1520 * of locks held is not a good idea, instead we remember an OOM state
1521 * in the task and mem_cgroup_oom_synchronize() has to be called at
1522 * the end of the page fault to complete the OOM handling.
1524 * Returns %true if an ongoing memcg OOM situation was detected and
1525 * completed, %false otherwise.
1527 bool mem_cgroup_oom_synchronize(bool handle
)
1529 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1530 struct oom_wait_info owait
;
1533 /* OOM is global, do not handle */
1540 owait
.memcg
= memcg
;
1541 owait
.wait
.flags
= 0;
1542 owait
.wait
.func
= memcg_oom_wake_function
;
1543 owait
.wait
.private = current
;
1544 INIT_LIST_HEAD(&owait
.wait
.entry
);
1546 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1547 mem_cgroup_mark_under_oom(memcg
);
1549 locked
= mem_cgroup_oom_trylock(memcg
);
1552 mem_cgroup_oom_notify(memcg
);
1554 if (locked
&& !memcg
->oom_kill_disable
) {
1555 mem_cgroup_unmark_under_oom(memcg
);
1556 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1557 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1558 current
->memcg_oom_order
);
1561 mem_cgroup_unmark_under_oom(memcg
);
1562 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1566 mem_cgroup_oom_unlock(memcg
);
1568 * There is no guarantee that an OOM-lock contender
1569 * sees the wakeups triggered by the OOM kill
1570 * uncharges. Wake any sleepers explicitely.
1572 memcg_oom_recover(memcg
);
1575 current
->memcg_in_oom
= NULL
;
1576 css_put(&memcg
->css
);
1581 * lock_page_memcg - lock a page->mem_cgroup binding
1584 * This function protects unlocked LRU pages from being moved to
1587 * It ensures lifetime of the returned memcg. Caller is responsible
1588 * for the lifetime of the page; __unlock_page_memcg() is available
1589 * when @page might get freed inside the locked section.
1591 struct mem_cgroup
*lock_page_memcg(struct page
*page
)
1593 struct mem_cgroup
*memcg
;
1594 unsigned long flags
;
1597 * The RCU lock is held throughout the transaction. The fast
1598 * path can get away without acquiring the memcg->move_lock
1599 * because page moving starts with an RCU grace period.
1601 * The RCU lock also protects the memcg from being freed when
1602 * the page state that is going to change is the only thing
1603 * preventing the page itself from being freed. E.g. writeback
1604 * doesn't hold a page reference and relies on PG_writeback to
1605 * keep off truncation, migration and so forth.
1609 if (mem_cgroup_disabled())
1612 memcg
= page
->mem_cgroup
;
1613 if (unlikely(!memcg
))
1616 if (atomic_read(&memcg
->moving_account
) <= 0)
1619 spin_lock_irqsave(&memcg
->move_lock
, flags
);
1620 if (memcg
!= page
->mem_cgroup
) {
1621 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1626 * When charge migration first begins, we can have locked and
1627 * unlocked page stat updates happening concurrently. Track
1628 * the task who has the lock for unlock_page_memcg().
1630 memcg
->move_lock_task
= current
;
1631 memcg
->move_lock_flags
= flags
;
1635 EXPORT_SYMBOL(lock_page_memcg
);
1638 * __unlock_page_memcg - unlock and unpin a memcg
1641 * Unlock and unpin a memcg returned by lock_page_memcg().
1643 void __unlock_page_memcg(struct mem_cgroup
*memcg
)
1645 if (memcg
&& memcg
->move_lock_task
== current
) {
1646 unsigned long flags
= memcg
->move_lock_flags
;
1648 memcg
->move_lock_task
= NULL
;
1649 memcg
->move_lock_flags
= 0;
1651 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1658 * unlock_page_memcg - unlock a page->mem_cgroup binding
1661 void unlock_page_memcg(struct page
*page
)
1663 __unlock_page_memcg(page
->mem_cgroup
);
1665 EXPORT_SYMBOL(unlock_page_memcg
);
1667 struct memcg_stock_pcp
{
1668 struct mem_cgroup
*cached
; /* this never be root cgroup */
1669 unsigned int nr_pages
;
1670 struct work_struct work
;
1671 unsigned long flags
;
1672 #define FLUSHING_CACHED_CHARGE 0
1674 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
1675 static DEFINE_MUTEX(percpu_charge_mutex
);
1678 * consume_stock: Try to consume stocked charge on this cpu.
1679 * @memcg: memcg to consume from.
1680 * @nr_pages: how many pages to charge.
1682 * The charges will only happen if @memcg matches the current cpu's memcg
1683 * stock, and at least @nr_pages are available in that stock. Failure to
1684 * service an allocation will refill the stock.
1686 * returns true if successful, false otherwise.
1688 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1690 struct memcg_stock_pcp
*stock
;
1691 unsigned long flags
;
1694 if (nr_pages
> MEMCG_CHARGE_BATCH
)
1697 local_irq_save(flags
);
1699 stock
= this_cpu_ptr(&memcg_stock
);
1700 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
1701 stock
->nr_pages
-= nr_pages
;
1705 local_irq_restore(flags
);
1711 * Returns stocks cached in percpu and reset cached information.
1713 static void drain_stock(struct memcg_stock_pcp
*stock
)
1715 struct mem_cgroup
*old
= stock
->cached
;
1717 if (stock
->nr_pages
) {
1718 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
1719 if (do_memsw_account())
1720 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
1721 css_put_many(&old
->css
, stock
->nr_pages
);
1722 stock
->nr_pages
= 0;
1724 stock
->cached
= NULL
;
1727 static void drain_local_stock(struct work_struct
*dummy
)
1729 struct memcg_stock_pcp
*stock
;
1730 unsigned long flags
;
1733 * The only protection from memory hotplug vs. drain_stock races is
1734 * that we always operate on local CPU stock here with IRQ disabled
1736 local_irq_save(flags
);
1738 stock
= this_cpu_ptr(&memcg_stock
);
1740 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
1742 local_irq_restore(flags
);
1746 * Cache charges(val) to local per_cpu area.
1747 * This will be consumed by consume_stock() function, later.
1749 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1751 struct memcg_stock_pcp
*stock
;
1752 unsigned long flags
;
1754 local_irq_save(flags
);
1756 stock
= this_cpu_ptr(&memcg_stock
);
1757 if (stock
->cached
!= memcg
) { /* reset if necessary */
1759 stock
->cached
= memcg
;
1761 stock
->nr_pages
+= nr_pages
;
1763 if (stock
->nr_pages
> MEMCG_CHARGE_BATCH
)
1766 local_irq_restore(flags
);
1770 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1771 * of the hierarchy under it.
1773 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
1777 /* If someone's already draining, avoid adding running more workers. */
1778 if (!mutex_trylock(&percpu_charge_mutex
))
1781 * Notify other cpus that system-wide "drain" is running
1782 * We do not care about races with the cpu hotplug because cpu down
1783 * as well as workers from this path always operate on the local
1784 * per-cpu data. CPU up doesn't touch memcg_stock at all.
1787 for_each_online_cpu(cpu
) {
1788 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
1789 struct mem_cgroup
*memcg
;
1791 memcg
= stock
->cached
;
1792 if (!memcg
|| !stock
->nr_pages
|| !css_tryget(&memcg
->css
))
1794 if (!mem_cgroup_is_descendant(memcg
, root_memcg
)) {
1795 css_put(&memcg
->css
);
1798 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
1800 drain_local_stock(&stock
->work
);
1802 schedule_work_on(cpu
, &stock
->work
);
1804 css_put(&memcg
->css
);
1807 mutex_unlock(&percpu_charge_mutex
);
1810 static int memcg_hotplug_cpu_dead(unsigned int cpu
)
1812 struct memcg_stock_pcp
*stock
;
1813 struct mem_cgroup
*memcg
;
1815 stock
= &per_cpu(memcg_stock
, cpu
);
1818 for_each_mem_cgroup(memcg
) {
1821 for (i
= 0; i
< MEMCG_NR_STAT
; i
++) {
1825 x
= this_cpu_xchg(memcg
->stat_cpu
->count
[i
], 0);
1827 atomic_long_add(x
, &memcg
->stat
[i
]);
1829 if (i
>= NR_VM_NODE_STAT_ITEMS
)
1832 for_each_node(nid
) {
1833 struct mem_cgroup_per_node
*pn
;
1835 pn
= mem_cgroup_nodeinfo(memcg
, nid
);
1836 x
= this_cpu_xchg(pn
->lruvec_stat_cpu
->count
[i
], 0);
1838 atomic_long_add(x
, &pn
->lruvec_stat
[i
]);
1842 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++) {
1845 x
= this_cpu_xchg(memcg
->stat_cpu
->events
[i
], 0);
1847 atomic_long_add(x
, &memcg
->events
[i
]);
1854 static void reclaim_high(struct mem_cgroup
*memcg
,
1855 unsigned int nr_pages
,
1859 if (page_counter_read(&memcg
->memory
) <= memcg
->high
)
1861 memcg_memory_event(memcg
, MEMCG_HIGH
);
1862 try_to_free_mem_cgroup_pages(memcg
, nr_pages
, gfp_mask
, true);
1863 } while ((memcg
= parent_mem_cgroup(memcg
)));
1866 static void high_work_func(struct work_struct
*work
)
1868 struct mem_cgroup
*memcg
;
1870 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
1871 reclaim_high(memcg
, MEMCG_CHARGE_BATCH
, GFP_KERNEL
);
1875 * Scheduled by try_charge() to be executed from the userland return path
1876 * and reclaims memory over the high limit.
1878 void mem_cgroup_handle_over_high(void)
1880 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
1881 struct mem_cgroup
*memcg
;
1883 if (likely(!nr_pages
))
1886 memcg
= get_mem_cgroup_from_mm(current
->mm
);
1887 reclaim_high(memcg
, nr_pages
, GFP_KERNEL
);
1888 css_put(&memcg
->css
);
1889 current
->memcg_nr_pages_over_high
= 0;
1892 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1893 unsigned int nr_pages
)
1895 unsigned int batch
= max(MEMCG_CHARGE_BATCH
, nr_pages
);
1896 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
1897 struct mem_cgroup
*mem_over_limit
;
1898 struct page_counter
*counter
;
1899 unsigned long nr_reclaimed
;
1900 bool may_swap
= true;
1901 bool drained
= false;
1903 if (mem_cgroup_is_root(memcg
))
1906 if (consume_stock(memcg
, nr_pages
))
1909 if (!do_memsw_account() ||
1910 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
1911 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
1913 if (do_memsw_account())
1914 page_counter_uncharge(&memcg
->memsw
, batch
);
1915 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
1917 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
1921 if (batch
> nr_pages
) {
1927 * Unlike in global OOM situations, memcg is not in a physical
1928 * memory shortage. Allow dying and OOM-killed tasks to
1929 * bypass the last charges so that they can exit quickly and
1930 * free their memory.
1932 if (unlikely(tsk_is_oom_victim(current
) ||
1933 fatal_signal_pending(current
) ||
1934 current
->flags
& PF_EXITING
))
1938 * Prevent unbounded recursion when reclaim operations need to
1939 * allocate memory. This might exceed the limits temporarily,
1940 * but we prefer facilitating memory reclaim and getting back
1941 * under the limit over triggering OOM kills in these cases.
1943 if (unlikely(current
->flags
& PF_MEMALLOC
))
1946 if (unlikely(task_in_memcg_oom(current
)))
1949 if (!gfpflags_allow_blocking(gfp_mask
))
1952 memcg_memory_event(mem_over_limit
, MEMCG_MAX
);
1954 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
1955 gfp_mask
, may_swap
);
1957 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
1961 drain_all_stock(mem_over_limit
);
1966 if (gfp_mask
& __GFP_NORETRY
)
1969 * Even though the limit is exceeded at this point, reclaim
1970 * may have been able to free some pages. Retry the charge
1971 * before killing the task.
1973 * Only for regular pages, though: huge pages are rather
1974 * unlikely to succeed so close to the limit, and we fall back
1975 * to regular pages anyway in case of failure.
1977 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
1980 * At task move, charge accounts can be doubly counted. So, it's
1981 * better to wait until the end of task_move if something is going on.
1983 if (mem_cgroup_wait_acct_move(mem_over_limit
))
1989 if (gfp_mask
& __GFP_NOFAIL
)
1992 if (fatal_signal_pending(current
))
1995 memcg_memory_event(mem_over_limit
, MEMCG_OOM
);
1997 mem_cgroup_oom(mem_over_limit
, gfp_mask
,
1998 get_order(nr_pages
* PAGE_SIZE
));
2000 if (!(gfp_mask
& __GFP_NOFAIL
))
2004 * The allocation either can't fail or will lead to more memory
2005 * being freed very soon. Allow memory usage go over the limit
2006 * temporarily by force charging it.
2008 page_counter_charge(&memcg
->memory
, nr_pages
);
2009 if (do_memsw_account())
2010 page_counter_charge(&memcg
->memsw
, nr_pages
);
2011 css_get_many(&memcg
->css
, nr_pages
);
2016 css_get_many(&memcg
->css
, batch
);
2017 if (batch
> nr_pages
)
2018 refill_stock(memcg
, batch
- nr_pages
);
2021 * If the hierarchy is above the normal consumption range, schedule
2022 * reclaim on returning to userland. We can perform reclaim here
2023 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2024 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2025 * not recorded as it most likely matches current's and won't
2026 * change in the meantime. As high limit is checked again before
2027 * reclaim, the cost of mismatch is negligible.
2030 if (page_counter_read(&memcg
->memory
) > memcg
->high
) {
2031 /* Don't bother a random interrupted task */
2032 if (in_interrupt()) {
2033 schedule_work(&memcg
->high_work
);
2036 current
->memcg_nr_pages_over_high
+= batch
;
2037 set_notify_resume(current
);
2040 } while ((memcg
= parent_mem_cgroup(memcg
)));
2045 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2047 if (mem_cgroup_is_root(memcg
))
2050 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2051 if (do_memsw_account())
2052 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2054 css_put_many(&memcg
->css
, nr_pages
);
2057 static void lock_page_lru(struct page
*page
, int *isolated
)
2059 struct zone
*zone
= page_zone(page
);
2061 spin_lock_irq(zone_lru_lock(zone
));
2062 if (PageLRU(page
)) {
2063 struct lruvec
*lruvec
;
2065 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
2067 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2073 static void unlock_page_lru(struct page
*page
, int isolated
)
2075 struct zone
*zone
= page_zone(page
);
2078 struct lruvec
*lruvec
;
2080 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
2081 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2083 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2085 spin_unlock_irq(zone_lru_lock(zone
));
2088 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2093 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2096 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2097 * may already be on some other mem_cgroup's LRU. Take care of it.
2100 lock_page_lru(page
, &isolated
);
2103 * Nobody should be changing or seriously looking at
2104 * page->mem_cgroup at this point:
2106 * - the page is uncharged
2108 * - the page is off-LRU
2110 * - an anonymous fault has exclusive page access, except for
2111 * a locked page table
2113 * - a page cache insertion, a swapin fault, or a migration
2114 * have the page locked
2116 page
->mem_cgroup
= memcg
;
2119 unlock_page_lru(page
, isolated
);
2123 static int memcg_alloc_cache_id(void)
2128 id
= ida_simple_get(&memcg_cache_ida
,
2129 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2133 if (id
< memcg_nr_cache_ids
)
2137 * There's no space for the new id in memcg_caches arrays,
2138 * so we have to grow them.
2140 down_write(&memcg_cache_ids_sem
);
2142 size
= 2 * (id
+ 1);
2143 if (size
< MEMCG_CACHES_MIN_SIZE
)
2144 size
= MEMCG_CACHES_MIN_SIZE
;
2145 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2146 size
= MEMCG_CACHES_MAX_SIZE
;
2148 err
= memcg_update_all_caches(size
);
2150 err
= memcg_update_all_list_lrus(size
);
2152 memcg_nr_cache_ids
= size
;
2154 up_write(&memcg_cache_ids_sem
);
2157 ida_simple_remove(&memcg_cache_ida
, id
);
2163 static void memcg_free_cache_id(int id
)
2165 ida_simple_remove(&memcg_cache_ida
, id
);
2168 struct memcg_kmem_cache_create_work
{
2169 struct mem_cgroup
*memcg
;
2170 struct kmem_cache
*cachep
;
2171 struct work_struct work
;
2174 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2176 struct memcg_kmem_cache_create_work
*cw
=
2177 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2178 struct mem_cgroup
*memcg
= cw
->memcg
;
2179 struct kmem_cache
*cachep
= cw
->cachep
;
2181 memcg_create_kmem_cache(memcg
, cachep
);
2183 css_put(&memcg
->css
);
2188 * Enqueue the creation of a per-memcg kmem_cache.
2190 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2191 struct kmem_cache
*cachep
)
2193 struct memcg_kmem_cache_create_work
*cw
;
2195 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
| __GFP_NOWARN
);
2199 css_get(&memcg
->css
);
2202 cw
->cachep
= cachep
;
2203 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2205 queue_work(memcg_kmem_cache_wq
, &cw
->work
);
2208 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2209 struct kmem_cache
*cachep
)
2212 * We need to stop accounting when we kmalloc, because if the
2213 * corresponding kmalloc cache is not yet created, the first allocation
2214 * in __memcg_schedule_kmem_cache_create will recurse.
2216 * However, it is better to enclose the whole function. Depending on
2217 * the debugging options enabled, INIT_WORK(), for instance, can
2218 * trigger an allocation. This too, will make us recurse. Because at
2219 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2220 * the safest choice is to do it like this, wrapping the whole function.
2222 current
->memcg_kmem_skip_account
= 1;
2223 __memcg_schedule_kmem_cache_create(memcg
, cachep
);
2224 current
->memcg_kmem_skip_account
= 0;
2227 static inline bool memcg_kmem_bypass(void)
2229 if (in_interrupt() || !current
->mm
|| (current
->flags
& PF_KTHREAD
))
2235 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2236 * @cachep: the original global kmem cache
2238 * Return the kmem_cache we're supposed to use for a slab allocation.
2239 * We try to use the current memcg's version of the cache.
2241 * If the cache does not exist yet, if we are the first user of it, we
2242 * create it asynchronously in a workqueue and let the current allocation
2243 * go through with the original cache.
2245 * This function takes a reference to the cache it returns to assure it
2246 * won't get destroyed while we are working with it. Once the caller is
2247 * done with it, memcg_kmem_put_cache() must be called to release the
2250 struct kmem_cache
*memcg_kmem_get_cache(struct kmem_cache
*cachep
)
2252 struct mem_cgroup
*memcg
;
2253 struct kmem_cache
*memcg_cachep
;
2256 VM_BUG_ON(!is_root_cache(cachep
));
2258 if (memcg_kmem_bypass())
2261 if (current
->memcg_kmem_skip_account
)
2264 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2265 kmemcg_id
= READ_ONCE(memcg
->kmemcg_id
);
2269 memcg_cachep
= cache_from_memcg_idx(cachep
, kmemcg_id
);
2270 if (likely(memcg_cachep
))
2271 return memcg_cachep
;
2274 * If we are in a safe context (can wait, and not in interrupt
2275 * context), we could be be predictable and return right away.
2276 * This would guarantee that the allocation being performed
2277 * already belongs in the new cache.
2279 * However, there are some clashes that can arrive from locking.
2280 * For instance, because we acquire the slab_mutex while doing
2281 * memcg_create_kmem_cache, this means no further allocation
2282 * could happen with the slab_mutex held. So it's better to
2285 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2287 css_put(&memcg
->css
);
2292 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2293 * @cachep: the cache returned by memcg_kmem_get_cache
2295 void memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2297 if (!is_root_cache(cachep
))
2298 css_put(&cachep
->memcg_params
.memcg
->css
);
2302 * memcg_kmem_charge_memcg: charge a kmem page
2303 * @page: page to charge
2304 * @gfp: reclaim mode
2305 * @order: allocation order
2306 * @memcg: memory cgroup to charge
2308 * Returns 0 on success, an error code on failure.
2310 int memcg_kmem_charge_memcg(struct page
*page
, gfp_t gfp
, int order
,
2311 struct mem_cgroup
*memcg
)
2313 unsigned int nr_pages
= 1 << order
;
2314 struct page_counter
*counter
;
2317 ret
= try_charge(memcg
, gfp
, nr_pages
);
2321 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
2322 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
2323 cancel_charge(memcg
, nr_pages
);
2327 page
->mem_cgroup
= memcg
;
2333 * memcg_kmem_charge: charge a kmem page to the current memory cgroup
2334 * @page: page to charge
2335 * @gfp: reclaim mode
2336 * @order: allocation order
2338 * Returns 0 on success, an error code on failure.
2340 int memcg_kmem_charge(struct page
*page
, gfp_t gfp
, int order
)
2342 struct mem_cgroup
*memcg
;
2345 if (memcg_kmem_bypass())
2348 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2349 if (!mem_cgroup_is_root(memcg
)) {
2350 ret
= memcg_kmem_charge_memcg(page
, gfp
, order
, memcg
);
2352 __SetPageKmemcg(page
);
2354 css_put(&memcg
->css
);
2358 * memcg_kmem_uncharge: uncharge a kmem page
2359 * @page: page to uncharge
2360 * @order: allocation order
2362 void memcg_kmem_uncharge(struct page
*page
, int order
)
2364 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2365 unsigned int nr_pages
= 1 << order
;
2370 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2372 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2373 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2375 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2376 if (do_memsw_account())
2377 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2379 page
->mem_cgroup
= NULL
;
2381 /* slab pages do not have PageKmemcg flag set */
2382 if (PageKmemcg(page
))
2383 __ClearPageKmemcg(page
);
2385 css_put_many(&memcg
->css
, nr_pages
);
2387 #endif /* !CONFIG_SLOB */
2389 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2392 * Because tail pages are not marked as "used", set it. We're under
2393 * zone_lru_lock and migration entries setup in all page mappings.
2395 void mem_cgroup_split_huge_fixup(struct page
*head
)
2399 if (mem_cgroup_disabled())
2402 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
2403 head
[i
].mem_cgroup
= head
->mem_cgroup
;
2405 __mod_memcg_state(head
->mem_cgroup
, MEMCG_RSS_HUGE
, -HPAGE_PMD_NR
);
2407 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2409 #ifdef CONFIG_MEMCG_SWAP
2411 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2412 * @entry: swap entry to be moved
2413 * @from: mem_cgroup which the entry is moved from
2414 * @to: mem_cgroup which the entry is moved to
2416 * It succeeds only when the swap_cgroup's record for this entry is the same
2417 * as the mem_cgroup's id of @from.
2419 * Returns 0 on success, -EINVAL on failure.
2421 * The caller must have charged to @to, IOW, called page_counter_charge() about
2422 * both res and memsw, and called css_get().
2424 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2425 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2427 unsigned short old_id
, new_id
;
2429 old_id
= mem_cgroup_id(from
);
2430 new_id
= mem_cgroup_id(to
);
2432 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2433 mod_memcg_state(from
, MEMCG_SWAP
, -1);
2434 mod_memcg_state(to
, MEMCG_SWAP
, 1);
2440 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2441 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2447 static DEFINE_MUTEX(memcg_max_mutex
);
2449 static int mem_cgroup_resize_max(struct mem_cgroup
*memcg
,
2450 unsigned long max
, bool memsw
)
2452 bool enlarge
= false;
2454 bool limits_invariant
;
2455 struct page_counter
*counter
= memsw
? &memcg
->memsw
: &memcg
->memory
;
2458 if (signal_pending(current
)) {
2463 mutex_lock(&memcg_max_mutex
);
2465 * Make sure that the new limit (memsw or memory limit) doesn't
2466 * break our basic invariant rule memory.max <= memsw.max.
2468 limits_invariant
= memsw
? max
>= memcg
->memory
.max
:
2469 max
<= memcg
->memsw
.max
;
2470 if (!limits_invariant
) {
2471 mutex_unlock(&memcg_max_mutex
);
2475 if (max
> counter
->max
)
2477 ret
= page_counter_set_max(counter
, max
);
2478 mutex_unlock(&memcg_max_mutex
);
2483 if (!try_to_free_mem_cgroup_pages(memcg
, 1,
2484 GFP_KERNEL
, !memsw
)) {
2490 if (!ret
&& enlarge
)
2491 memcg_oom_recover(memcg
);
2496 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t
*pgdat
, int order
,
2498 unsigned long *total_scanned
)
2500 unsigned long nr_reclaimed
= 0;
2501 struct mem_cgroup_per_node
*mz
, *next_mz
= NULL
;
2502 unsigned long reclaimed
;
2504 struct mem_cgroup_tree_per_node
*mctz
;
2505 unsigned long excess
;
2506 unsigned long nr_scanned
;
2511 mctz
= soft_limit_tree_node(pgdat
->node_id
);
2514 * Do not even bother to check the largest node if the root
2515 * is empty. Do it lockless to prevent lock bouncing. Races
2516 * are acceptable as soft limit is best effort anyway.
2518 if (!mctz
|| RB_EMPTY_ROOT(&mctz
->rb_root
))
2522 * This loop can run a while, specially if mem_cgroup's continuously
2523 * keep exceeding their soft limit and putting the system under
2530 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
2535 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, pgdat
,
2536 gfp_mask
, &nr_scanned
);
2537 nr_reclaimed
+= reclaimed
;
2538 *total_scanned
+= nr_scanned
;
2539 spin_lock_irq(&mctz
->lock
);
2540 __mem_cgroup_remove_exceeded(mz
, mctz
);
2543 * If we failed to reclaim anything from this memory cgroup
2544 * it is time to move on to the next cgroup
2548 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
2550 excess
= soft_limit_excess(mz
->memcg
);
2552 * One school of thought says that we should not add
2553 * back the node to the tree if reclaim returns 0.
2554 * But our reclaim could return 0, simply because due
2555 * to priority we are exposing a smaller subset of
2556 * memory to reclaim from. Consider this as a longer
2559 /* If excess == 0, no tree ops */
2560 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
2561 spin_unlock_irq(&mctz
->lock
);
2562 css_put(&mz
->memcg
->css
);
2565 * Could not reclaim anything and there are no more
2566 * mem cgroups to try or we seem to be looping without
2567 * reclaiming anything.
2569 if (!nr_reclaimed
&&
2571 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
2573 } while (!nr_reclaimed
);
2575 css_put(&next_mz
->memcg
->css
);
2576 return nr_reclaimed
;
2580 * Test whether @memcg has children, dead or alive. Note that this
2581 * function doesn't care whether @memcg has use_hierarchy enabled and
2582 * returns %true if there are child csses according to the cgroup
2583 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2585 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
2590 ret
= css_next_child(NULL
, &memcg
->css
);
2596 * Reclaims as many pages from the given memcg as possible.
2598 * Caller is responsible for holding css reference for memcg.
2600 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
2602 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2604 /* we call try-to-free pages for make this cgroup empty */
2605 lru_add_drain_all();
2606 /* try to free all pages in this cgroup */
2607 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
2610 if (signal_pending(current
))
2613 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
2617 /* maybe some writeback is necessary */
2618 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2626 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
2627 char *buf
, size_t nbytes
,
2630 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2632 if (mem_cgroup_is_root(memcg
))
2634 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
2637 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
2640 return mem_cgroup_from_css(css
)->use_hierarchy
;
2643 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
2644 struct cftype
*cft
, u64 val
)
2647 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2648 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
2650 if (memcg
->use_hierarchy
== val
)
2654 * If parent's use_hierarchy is set, we can't make any modifications
2655 * in the child subtrees. If it is unset, then the change can
2656 * occur, provided the current cgroup has no children.
2658 * For the root cgroup, parent_mem is NULL, we allow value to be
2659 * set if there are no children.
2661 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
2662 (val
== 1 || val
== 0)) {
2663 if (!memcg_has_children(memcg
))
2664 memcg
->use_hierarchy
= val
;
2673 static void tree_stat(struct mem_cgroup
*memcg
, unsigned long *stat
)
2675 struct mem_cgroup
*iter
;
2678 memset(stat
, 0, sizeof(*stat
) * MEMCG_NR_STAT
);
2680 for_each_mem_cgroup_tree(iter
, memcg
) {
2681 for (i
= 0; i
< MEMCG_NR_STAT
; i
++)
2682 stat
[i
] += memcg_page_state(iter
, i
);
2686 static void tree_events(struct mem_cgroup
*memcg
, unsigned long *events
)
2688 struct mem_cgroup
*iter
;
2691 memset(events
, 0, sizeof(*events
) * NR_VM_EVENT_ITEMS
);
2693 for_each_mem_cgroup_tree(iter
, memcg
) {
2694 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++)
2695 events
[i
] += memcg_sum_events(iter
, i
);
2699 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
2701 unsigned long val
= 0;
2703 if (mem_cgroup_is_root(memcg
)) {
2704 struct mem_cgroup
*iter
;
2706 for_each_mem_cgroup_tree(iter
, memcg
) {
2707 val
+= memcg_page_state(iter
, MEMCG_CACHE
);
2708 val
+= memcg_page_state(iter
, MEMCG_RSS
);
2710 val
+= memcg_page_state(iter
, MEMCG_SWAP
);
2714 val
= page_counter_read(&memcg
->memory
);
2716 val
= page_counter_read(&memcg
->memsw
);
2729 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
2732 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2733 struct page_counter
*counter
;
2735 switch (MEMFILE_TYPE(cft
->private)) {
2737 counter
= &memcg
->memory
;
2740 counter
= &memcg
->memsw
;
2743 counter
= &memcg
->kmem
;
2746 counter
= &memcg
->tcpmem
;
2752 switch (MEMFILE_ATTR(cft
->private)) {
2754 if (counter
== &memcg
->memory
)
2755 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
2756 if (counter
== &memcg
->memsw
)
2757 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
2758 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
2760 return (u64
)counter
->max
* PAGE_SIZE
;
2762 return (u64
)counter
->watermark
* PAGE_SIZE
;
2764 return counter
->failcnt
;
2765 case RES_SOFT_LIMIT
:
2766 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
2773 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2777 if (cgroup_memory_nokmem
)
2780 BUG_ON(memcg
->kmemcg_id
>= 0);
2781 BUG_ON(memcg
->kmem_state
);
2783 memcg_id
= memcg_alloc_cache_id();
2787 static_branch_inc(&memcg_kmem_enabled_key
);
2789 * A memory cgroup is considered kmem-online as soon as it gets
2790 * kmemcg_id. Setting the id after enabling static branching will
2791 * guarantee no one starts accounting before all call sites are
2794 memcg
->kmemcg_id
= memcg_id
;
2795 memcg
->kmem_state
= KMEM_ONLINE
;
2796 INIT_LIST_HEAD(&memcg
->kmem_caches
);
2801 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2803 struct cgroup_subsys_state
*css
;
2804 struct mem_cgroup
*parent
, *child
;
2807 if (memcg
->kmem_state
!= KMEM_ONLINE
)
2810 * Clear the online state before clearing memcg_caches array
2811 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2812 * guarantees that no cache will be created for this cgroup
2813 * after we are done (see memcg_create_kmem_cache()).
2815 memcg
->kmem_state
= KMEM_ALLOCATED
;
2817 memcg_deactivate_kmem_caches(memcg
);
2819 kmemcg_id
= memcg
->kmemcg_id
;
2820 BUG_ON(kmemcg_id
< 0);
2822 parent
= parent_mem_cgroup(memcg
);
2824 parent
= root_mem_cgroup
;
2827 * Change kmemcg_id of this cgroup and all its descendants to the
2828 * parent's id, and then move all entries from this cgroup's list_lrus
2829 * to ones of the parent. After we have finished, all list_lrus
2830 * corresponding to this cgroup are guaranteed to remain empty. The
2831 * ordering is imposed by list_lru_node->lock taken by
2832 * memcg_drain_all_list_lrus().
2834 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
2835 css_for_each_descendant_pre(css
, &memcg
->css
) {
2836 child
= mem_cgroup_from_css(css
);
2837 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
2838 child
->kmemcg_id
= parent
->kmemcg_id
;
2839 if (!memcg
->use_hierarchy
)
2844 memcg_drain_all_list_lrus(kmemcg_id
, parent
->kmemcg_id
);
2846 memcg_free_cache_id(kmemcg_id
);
2849 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2851 /* css_alloc() failed, offlining didn't happen */
2852 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
2853 memcg_offline_kmem(memcg
);
2855 if (memcg
->kmem_state
== KMEM_ALLOCATED
) {
2856 memcg_destroy_kmem_caches(memcg
);
2857 static_branch_dec(&memcg_kmem_enabled_key
);
2858 WARN_ON(page_counter_read(&memcg
->kmem
));
2862 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2866 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2869 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2872 #endif /* !CONFIG_SLOB */
2874 static int memcg_update_kmem_max(struct mem_cgroup
*memcg
,
2879 mutex_lock(&memcg_max_mutex
);
2880 ret
= page_counter_set_max(&memcg
->kmem
, max
);
2881 mutex_unlock(&memcg_max_mutex
);
2885 static int memcg_update_tcp_max(struct mem_cgroup
*memcg
, unsigned long max
)
2889 mutex_lock(&memcg_max_mutex
);
2891 ret
= page_counter_set_max(&memcg
->tcpmem
, max
);
2895 if (!memcg
->tcpmem_active
) {
2897 * The active flag needs to be written after the static_key
2898 * update. This is what guarantees that the socket activation
2899 * function is the last one to run. See mem_cgroup_sk_alloc()
2900 * for details, and note that we don't mark any socket as
2901 * belonging to this memcg until that flag is up.
2903 * We need to do this, because static_keys will span multiple
2904 * sites, but we can't control their order. If we mark a socket
2905 * as accounted, but the accounting functions are not patched in
2906 * yet, we'll lose accounting.
2908 * We never race with the readers in mem_cgroup_sk_alloc(),
2909 * because when this value change, the code to process it is not
2912 static_branch_inc(&memcg_sockets_enabled_key
);
2913 memcg
->tcpmem_active
= true;
2916 mutex_unlock(&memcg_max_mutex
);
2921 * The user of this function is...
2924 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
2925 char *buf
, size_t nbytes
, loff_t off
)
2927 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2928 unsigned long nr_pages
;
2931 buf
= strstrip(buf
);
2932 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
2936 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
2938 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
2942 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
2944 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, false);
2947 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, true);
2950 ret
= memcg_update_kmem_max(memcg
, nr_pages
);
2953 ret
= memcg_update_tcp_max(memcg
, nr_pages
);
2957 case RES_SOFT_LIMIT
:
2958 memcg
->soft_limit
= nr_pages
;
2962 return ret
?: nbytes
;
2965 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
2966 size_t nbytes
, loff_t off
)
2968 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2969 struct page_counter
*counter
;
2971 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
2973 counter
= &memcg
->memory
;
2976 counter
= &memcg
->memsw
;
2979 counter
= &memcg
->kmem
;
2982 counter
= &memcg
->tcpmem
;
2988 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
2990 page_counter_reset_watermark(counter
);
2993 counter
->failcnt
= 0;
3002 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3005 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3009 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3010 struct cftype
*cft
, u64 val
)
3012 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3014 if (val
& ~MOVE_MASK
)
3018 * No kind of locking is needed in here, because ->can_attach() will
3019 * check this value once in the beginning of the process, and then carry
3020 * on with stale data. This means that changes to this value will only
3021 * affect task migrations starting after the change.
3023 memcg
->move_charge_at_immigrate
= val
;
3027 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3028 struct cftype
*cft
, u64 val
)
3035 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3039 unsigned int lru_mask
;
3042 static const struct numa_stat stats
[] = {
3043 { "total", LRU_ALL
},
3044 { "file", LRU_ALL_FILE
},
3045 { "anon", LRU_ALL_ANON
},
3046 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3048 const struct numa_stat
*stat
;
3051 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3053 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3054 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3055 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3056 for_each_node_state(nid
, N_MEMORY
) {
3057 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3059 seq_printf(m
, " N%d=%lu", nid
, nr
);
3064 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3065 struct mem_cgroup
*iter
;
3068 for_each_mem_cgroup_tree(iter
, memcg
)
3069 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3070 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3071 for_each_node_state(nid
, N_MEMORY
) {
3073 for_each_mem_cgroup_tree(iter
, memcg
)
3074 nr
+= mem_cgroup_node_nr_lru_pages(
3075 iter
, nid
, stat
->lru_mask
);
3076 seq_printf(m
, " N%d=%lu", nid
, nr
);
3083 #endif /* CONFIG_NUMA */
3085 /* Universal VM events cgroup1 shows, original sort order */
3086 unsigned int memcg1_events
[] = {
3093 static const char *const memcg1_event_names
[] = {
3100 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3102 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3103 unsigned long memory
, memsw
;
3104 struct mem_cgroup
*mi
;
3107 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names
) != ARRAY_SIZE(memcg1_stats
));
3108 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3110 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
3111 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
3113 seq_printf(m
, "%s %lu\n", memcg1_stat_names
[i
],
3114 memcg_page_state(memcg
, memcg1_stats
[i
]) *
3118 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
3119 seq_printf(m
, "%s %lu\n", memcg1_event_names
[i
],
3120 memcg_sum_events(memcg
, memcg1_events
[i
]));
3122 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3123 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3124 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
3126 /* Hierarchical information */
3127 memory
= memsw
= PAGE_COUNTER_MAX
;
3128 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3129 memory
= min(memory
, mi
->memory
.max
);
3130 memsw
= min(memsw
, mi
->memsw
.max
);
3132 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3133 (u64
)memory
* PAGE_SIZE
);
3134 if (do_memsw_account())
3135 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3136 (u64
)memsw
* PAGE_SIZE
);
3138 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
3139 unsigned long long val
= 0;
3141 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
3143 for_each_mem_cgroup_tree(mi
, memcg
)
3144 val
+= memcg_page_state(mi
, memcg1_stats
[i
]) *
3146 seq_printf(m
, "total_%s %llu\n", memcg1_stat_names
[i
], val
);
3149 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++) {
3150 unsigned long long val
= 0;
3152 for_each_mem_cgroup_tree(mi
, memcg
)
3153 val
+= memcg_sum_events(mi
, memcg1_events
[i
]);
3154 seq_printf(m
, "total_%s %llu\n", memcg1_event_names
[i
], val
);
3157 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
3158 unsigned long long val
= 0;
3160 for_each_mem_cgroup_tree(mi
, memcg
)
3161 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
3162 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
3165 #ifdef CONFIG_DEBUG_VM
3168 struct mem_cgroup_per_node
*mz
;
3169 struct zone_reclaim_stat
*rstat
;
3170 unsigned long recent_rotated
[2] = {0, 0};
3171 unsigned long recent_scanned
[2] = {0, 0};
3173 for_each_online_pgdat(pgdat
) {
3174 mz
= mem_cgroup_nodeinfo(memcg
, pgdat
->node_id
);
3175 rstat
= &mz
->lruvec
.reclaim_stat
;
3177 recent_rotated
[0] += rstat
->recent_rotated
[0];
3178 recent_rotated
[1] += rstat
->recent_rotated
[1];
3179 recent_scanned
[0] += rstat
->recent_scanned
[0];
3180 recent_scanned
[1] += rstat
->recent_scanned
[1];
3182 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3183 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3184 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3185 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3192 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3195 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3197 return mem_cgroup_swappiness(memcg
);
3200 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3201 struct cftype
*cft
, u64 val
)
3203 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3209 memcg
->swappiness
= val
;
3211 vm_swappiness
= val
;
3216 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3218 struct mem_cgroup_threshold_ary
*t
;
3219 unsigned long usage
;
3224 t
= rcu_dereference(memcg
->thresholds
.primary
);
3226 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3231 usage
= mem_cgroup_usage(memcg
, swap
);
3234 * current_threshold points to threshold just below or equal to usage.
3235 * If it's not true, a threshold was crossed after last
3236 * call of __mem_cgroup_threshold().
3238 i
= t
->current_threshold
;
3241 * Iterate backward over array of thresholds starting from
3242 * current_threshold and check if a threshold is crossed.
3243 * If none of thresholds below usage is crossed, we read
3244 * only one element of the array here.
3246 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3247 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3249 /* i = current_threshold + 1 */
3253 * Iterate forward over array of thresholds starting from
3254 * current_threshold+1 and check if a threshold is crossed.
3255 * If none of thresholds above usage is crossed, we read
3256 * only one element of the array here.
3258 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3259 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3261 /* Update current_threshold */
3262 t
->current_threshold
= i
- 1;
3267 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3270 __mem_cgroup_threshold(memcg
, false);
3271 if (do_memsw_account())
3272 __mem_cgroup_threshold(memcg
, true);
3274 memcg
= parent_mem_cgroup(memcg
);
3278 static int compare_thresholds(const void *a
, const void *b
)
3280 const struct mem_cgroup_threshold
*_a
= a
;
3281 const struct mem_cgroup_threshold
*_b
= b
;
3283 if (_a
->threshold
> _b
->threshold
)
3286 if (_a
->threshold
< _b
->threshold
)
3292 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3294 struct mem_cgroup_eventfd_list
*ev
;
3296 spin_lock(&memcg_oom_lock
);
3298 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3299 eventfd_signal(ev
->eventfd
, 1);
3301 spin_unlock(&memcg_oom_lock
);
3305 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3307 struct mem_cgroup
*iter
;
3309 for_each_mem_cgroup_tree(iter
, memcg
)
3310 mem_cgroup_oom_notify_cb(iter
);
3313 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3314 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
3316 struct mem_cgroup_thresholds
*thresholds
;
3317 struct mem_cgroup_threshold_ary
*new;
3318 unsigned long threshold
;
3319 unsigned long usage
;
3322 ret
= page_counter_memparse(args
, "-1", &threshold
);
3326 mutex_lock(&memcg
->thresholds_lock
);
3329 thresholds
= &memcg
->thresholds
;
3330 usage
= mem_cgroup_usage(memcg
, false);
3331 } else if (type
== _MEMSWAP
) {
3332 thresholds
= &memcg
->memsw_thresholds
;
3333 usage
= mem_cgroup_usage(memcg
, true);
3337 /* Check if a threshold crossed before adding a new one */
3338 if (thresholds
->primary
)
3339 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3341 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3343 /* Allocate memory for new array of thresholds */
3344 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
3352 /* Copy thresholds (if any) to new array */
3353 if (thresholds
->primary
) {
3354 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3355 sizeof(struct mem_cgroup_threshold
));
3358 /* Add new threshold */
3359 new->entries
[size
- 1].eventfd
= eventfd
;
3360 new->entries
[size
- 1].threshold
= threshold
;
3362 /* Sort thresholds. Registering of new threshold isn't time-critical */
3363 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3364 compare_thresholds
, NULL
);
3366 /* Find current threshold */
3367 new->current_threshold
= -1;
3368 for (i
= 0; i
< size
; i
++) {
3369 if (new->entries
[i
].threshold
<= usage
) {
3371 * new->current_threshold will not be used until
3372 * rcu_assign_pointer(), so it's safe to increment
3375 ++new->current_threshold
;
3380 /* Free old spare buffer and save old primary buffer as spare */
3381 kfree(thresholds
->spare
);
3382 thresholds
->spare
= thresholds
->primary
;
3384 rcu_assign_pointer(thresholds
->primary
, new);
3386 /* To be sure that nobody uses thresholds */
3390 mutex_unlock(&memcg
->thresholds_lock
);
3395 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3396 struct eventfd_ctx
*eventfd
, const char *args
)
3398 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
3401 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3402 struct eventfd_ctx
*eventfd
, const char *args
)
3404 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
3407 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3408 struct eventfd_ctx
*eventfd
, enum res_type type
)
3410 struct mem_cgroup_thresholds
*thresholds
;
3411 struct mem_cgroup_threshold_ary
*new;
3412 unsigned long usage
;
3415 mutex_lock(&memcg
->thresholds_lock
);
3418 thresholds
= &memcg
->thresholds
;
3419 usage
= mem_cgroup_usage(memcg
, false);
3420 } else if (type
== _MEMSWAP
) {
3421 thresholds
= &memcg
->memsw_thresholds
;
3422 usage
= mem_cgroup_usage(memcg
, true);
3426 if (!thresholds
->primary
)
3429 /* Check if a threshold crossed before removing */
3430 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3432 /* Calculate new number of threshold */
3434 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
3435 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
3439 new = thresholds
->spare
;
3441 /* Set thresholds array to NULL if we don't have thresholds */
3450 /* Copy thresholds and find current threshold */
3451 new->current_threshold
= -1;
3452 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
3453 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
3456 new->entries
[j
] = thresholds
->primary
->entries
[i
];
3457 if (new->entries
[j
].threshold
<= usage
) {
3459 * new->current_threshold will not be used
3460 * until rcu_assign_pointer(), so it's safe to increment
3463 ++new->current_threshold
;
3469 /* Swap primary and spare array */
3470 thresholds
->spare
= thresholds
->primary
;
3472 rcu_assign_pointer(thresholds
->primary
, new);
3474 /* To be sure that nobody uses thresholds */
3477 /* If all events are unregistered, free the spare array */
3479 kfree(thresholds
->spare
);
3480 thresholds
->spare
= NULL
;
3483 mutex_unlock(&memcg
->thresholds_lock
);
3486 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3487 struct eventfd_ctx
*eventfd
)
3489 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
3492 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3493 struct eventfd_ctx
*eventfd
)
3495 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
3498 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
3499 struct eventfd_ctx
*eventfd
, const char *args
)
3501 struct mem_cgroup_eventfd_list
*event
;
3503 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
3507 spin_lock(&memcg_oom_lock
);
3509 event
->eventfd
= eventfd
;
3510 list_add(&event
->list
, &memcg
->oom_notify
);
3512 /* already in OOM ? */
3513 if (memcg
->under_oom
)
3514 eventfd_signal(eventfd
, 1);
3515 spin_unlock(&memcg_oom_lock
);
3520 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
3521 struct eventfd_ctx
*eventfd
)
3523 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
3525 spin_lock(&memcg_oom_lock
);
3527 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
3528 if (ev
->eventfd
== eventfd
) {
3529 list_del(&ev
->list
);
3534 spin_unlock(&memcg_oom_lock
);
3537 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
3539 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
3541 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
3542 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
3543 seq_printf(sf
, "oom_kill %lu\n", memcg_sum_events(memcg
, OOM_KILL
));
3547 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
3548 struct cftype
*cft
, u64 val
)
3550 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3552 /* cannot set to root cgroup and only 0 and 1 are allowed */
3553 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
3556 memcg
->oom_kill_disable
= val
;
3558 memcg_oom_recover(memcg
);
3563 #ifdef CONFIG_CGROUP_WRITEBACK
3565 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3567 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
3570 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3572 wb_domain_exit(&memcg
->cgwb_domain
);
3575 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3577 wb_domain_size_changed(&memcg
->cgwb_domain
);
3580 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
3582 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3584 if (!memcg
->css
.parent
)
3587 return &memcg
->cgwb_domain
;
3591 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3592 * @wb: bdi_writeback in question
3593 * @pfilepages: out parameter for number of file pages
3594 * @pheadroom: out parameter for number of allocatable pages according to memcg
3595 * @pdirty: out parameter for number of dirty pages
3596 * @pwriteback: out parameter for number of pages under writeback
3598 * Determine the numbers of file, headroom, dirty, and writeback pages in
3599 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3600 * is a bit more involved.
3602 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3603 * headroom is calculated as the lowest headroom of itself and the
3604 * ancestors. Note that this doesn't consider the actual amount of
3605 * available memory in the system. The caller should further cap
3606 * *@pheadroom accordingly.
3608 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
3609 unsigned long *pheadroom
, unsigned long *pdirty
,
3610 unsigned long *pwriteback
)
3612 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3613 struct mem_cgroup
*parent
;
3615 *pdirty
= memcg_page_state(memcg
, NR_FILE_DIRTY
);
3617 /* this should eventually include NR_UNSTABLE_NFS */
3618 *pwriteback
= memcg_page_state(memcg
, NR_WRITEBACK
);
3619 *pfilepages
= mem_cgroup_nr_lru_pages(memcg
, (1 << LRU_INACTIVE_FILE
) |
3620 (1 << LRU_ACTIVE_FILE
));
3621 *pheadroom
= PAGE_COUNTER_MAX
;
3623 while ((parent
= parent_mem_cgroup(memcg
))) {
3624 unsigned long ceiling
= min(memcg
->memory
.max
, memcg
->high
);
3625 unsigned long used
= page_counter_read(&memcg
->memory
);
3627 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
3632 #else /* CONFIG_CGROUP_WRITEBACK */
3634 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3639 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3643 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3647 #endif /* CONFIG_CGROUP_WRITEBACK */
3650 * DO NOT USE IN NEW FILES.
3652 * "cgroup.event_control" implementation.
3654 * This is way over-engineered. It tries to support fully configurable
3655 * events for each user. Such level of flexibility is completely
3656 * unnecessary especially in the light of the planned unified hierarchy.
3658 * Please deprecate this and replace with something simpler if at all
3663 * Unregister event and free resources.
3665 * Gets called from workqueue.
3667 static void memcg_event_remove(struct work_struct
*work
)
3669 struct mem_cgroup_event
*event
=
3670 container_of(work
, struct mem_cgroup_event
, remove
);
3671 struct mem_cgroup
*memcg
= event
->memcg
;
3673 remove_wait_queue(event
->wqh
, &event
->wait
);
3675 event
->unregister_event(memcg
, event
->eventfd
);
3677 /* Notify userspace the event is going away. */
3678 eventfd_signal(event
->eventfd
, 1);
3680 eventfd_ctx_put(event
->eventfd
);
3682 css_put(&memcg
->css
);
3686 * Gets called on EPOLLHUP on eventfd when user closes it.
3688 * Called with wqh->lock held and interrupts disabled.
3690 static int memcg_event_wake(wait_queue_entry_t
*wait
, unsigned mode
,
3691 int sync
, void *key
)
3693 struct mem_cgroup_event
*event
=
3694 container_of(wait
, struct mem_cgroup_event
, wait
);
3695 struct mem_cgroup
*memcg
= event
->memcg
;
3696 __poll_t flags
= key_to_poll(key
);
3698 if (flags
& EPOLLHUP
) {
3700 * If the event has been detached at cgroup removal, we
3701 * can simply return knowing the other side will cleanup
3704 * We can't race against event freeing since the other
3705 * side will require wqh->lock via remove_wait_queue(),
3708 spin_lock(&memcg
->event_list_lock
);
3709 if (!list_empty(&event
->list
)) {
3710 list_del_init(&event
->list
);
3712 * We are in atomic context, but cgroup_event_remove()
3713 * may sleep, so we have to call it in workqueue.
3715 schedule_work(&event
->remove
);
3717 spin_unlock(&memcg
->event_list_lock
);
3723 static void memcg_event_ptable_queue_proc(struct file
*file
,
3724 wait_queue_head_t
*wqh
, poll_table
*pt
)
3726 struct mem_cgroup_event
*event
=
3727 container_of(pt
, struct mem_cgroup_event
, pt
);
3730 add_wait_queue(wqh
, &event
->wait
);
3734 * DO NOT USE IN NEW FILES.
3736 * Parse input and register new cgroup event handler.
3738 * Input must be in format '<event_fd> <control_fd> <args>'.
3739 * Interpretation of args is defined by control file implementation.
3741 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
3742 char *buf
, size_t nbytes
, loff_t off
)
3744 struct cgroup_subsys_state
*css
= of_css(of
);
3745 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3746 struct mem_cgroup_event
*event
;
3747 struct cgroup_subsys_state
*cfile_css
;
3748 unsigned int efd
, cfd
;
3755 buf
= strstrip(buf
);
3757 efd
= simple_strtoul(buf
, &endp
, 10);
3762 cfd
= simple_strtoul(buf
, &endp
, 10);
3763 if ((*endp
!= ' ') && (*endp
!= '\0'))
3767 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
3771 event
->memcg
= memcg
;
3772 INIT_LIST_HEAD(&event
->list
);
3773 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
3774 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
3775 INIT_WORK(&event
->remove
, memcg_event_remove
);
3783 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
3784 if (IS_ERR(event
->eventfd
)) {
3785 ret
= PTR_ERR(event
->eventfd
);
3792 goto out_put_eventfd
;
3795 /* the process need read permission on control file */
3796 /* AV: shouldn't we check that it's been opened for read instead? */
3797 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
3802 * Determine the event callbacks and set them in @event. This used
3803 * to be done via struct cftype but cgroup core no longer knows
3804 * about these events. The following is crude but the whole thing
3805 * is for compatibility anyway.
3807 * DO NOT ADD NEW FILES.
3809 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
3811 if (!strcmp(name
, "memory.usage_in_bytes")) {
3812 event
->register_event
= mem_cgroup_usage_register_event
;
3813 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
3814 } else if (!strcmp(name
, "memory.oom_control")) {
3815 event
->register_event
= mem_cgroup_oom_register_event
;
3816 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
3817 } else if (!strcmp(name
, "memory.pressure_level")) {
3818 event
->register_event
= vmpressure_register_event
;
3819 event
->unregister_event
= vmpressure_unregister_event
;
3820 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
3821 event
->register_event
= memsw_cgroup_usage_register_event
;
3822 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
3829 * Verify @cfile should belong to @css. Also, remaining events are
3830 * automatically removed on cgroup destruction but the removal is
3831 * asynchronous, so take an extra ref on @css.
3833 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
3834 &memory_cgrp_subsys
);
3836 if (IS_ERR(cfile_css
))
3838 if (cfile_css
!= css
) {
3843 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
3847 vfs_poll(efile
.file
, &event
->pt
);
3849 spin_lock(&memcg
->event_list_lock
);
3850 list_add(&event
->list
, &memcg
->event_list
);
3851 spin_unlock(&memcg
->event_list_lock
);
3863 eventfd_ctx_put(event
->eventfd
);
3872 static struct cftype mem_cgroup_legacy_files
[] = {
3874 .name
= "usage_in_bytes",
3875 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
3876 .read_u64
= mem_cgroup_read_u64
,
3879 .name
= "max_usage_in_bytes",
3880 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
3881 .write
= mem_cgroup_reset
,
3882 .read_u64
= mem_cgroup_read_u64
,
3885 .name
= "limit_in_bytes",
3886 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
3887 .write
= mem_cgroup_write
,
3888 .read_u64
= mem_cgroup_read_u64
,
3891 .name
= "soft_limit_in_bytes",
3892 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
3893 .write
= mem_cgroup_write
,
3894 .read_u64
= mem_cgroup_read_u64
,
3898 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
3899 .write
= mem_cgroup_reset
,
3900 .read_u64
= mem_cgroup_read_u64
,
3904 .seq_show
= memcg_stat_show
,
3907 .name
= "force_empty",
3908 .write
= mem_cgroup_force_empty_write
,
3911 .name
= "use_hierarchy",
3912 .write_u64
= mem_cgroup_hierarchy_write
,
3913 .read_u64
= mem_cgroup_hierarchy_read
,
3916 .name
= "cgroup.event_control", /* XXX: for compat */
3917 .write
= memcg_write_event_control
,
3918 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
3921 .name
= "swappiness",
3922 .read_u64
= mem_cgroup_swappiness_read
,
3923 .write_u64
= mem_cgroup_swappiness_write
,
3926 .name
= "move_charge_at_immigrate",
3927 .read_u64
= mem_cgroup_move_charge_read
,
3928 .write_u64
= mem_cgroup_move_charge_write
,
3931 .name
= "oom_control",
3932 .seq_show
= mem_cgroup_oom_control_read
,
3933 .write_u64
= mem_cgroup_oom_control_write
,
3934 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
3937 .name
= "pressure_level",
3941 .name
= "numa_stat",
3942 .seq_show
= memcg_numa_stat_show
,
3946 .name
= "kmem.limit_in_bytes",
3947 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
3948 .write
= mem_cgroup_write
,
3949 .read_u64
= mem_cgroup_read_u64
,
3952 .name
= "kmem.usage_in_bytes",
3953 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
3954 .read_u64
= mem_cgroup_read_u64
,
3957 .name
= "kmem.failcnt",
3958 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
3959 .write
= mem_cgroup_reset
,
3960 .read_u64
= mem_cgroup_read_u64
,
3963 .name
= "kmem.max_usage_in_bytes",
3964 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
3965 .write
= mem_cgroup_reset
,
3966 .read_u64
= mem_cgroup_read_u64
,
3968 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
3970 .name
= "kmem.slabinfo",
3971 .seq_start
= memcg_slab_start
,
3972 .seq_next
= memcg_slab_next
,
3973 .seq_stop
= memcg_slab_stop
,
3974 .seq_show
= memcg_slab_show
,
3978 .name
= "kmem.tcp.limit_in_bytes",
3979 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
3980 .write
= mem_cgroup_write
,
3981 .read_u64
= mem_cgroup_read_u64
,
3984 .name
= "kmem.tcp.usage_in_bytes",
3985 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
3986 .read_u64
= mem_cgroup_read_u64
,
3989 .name
= "kmem.tcp.failcnt",
3990 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
3991 .write
= mem_cgroup_reset
,
3992 .read_u64
= mem_cgroup_read_u64
,
3995 .name
= "kmem.tcp.max_usage_in_bytes",
3996 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
3997 .write
= mem_cgroup_reset
,
3998 .read_u64
= mem_cgroup_read_u64
,
4000 { }, /* terminate */
4004 * Private memory cgroup IDR
4006 * Swap-out records and page cache shadow entries need to store memcg
4007 * references in constrained space, so we maintain an ID space that is
4008 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4009 * memory-controlled cgroups to 64k.
4011 * However, there usually are many references to the oflline CSS after
4012 * the cgroup has been destroyed, such as page cache or reclaimable
4013 * slab objects, that don't need to hang on to the ID. We want to keep
4014 * those dead CSS from occupying IDs, or we might quickly exhaust the
4015 * relatively small ID space and prevent the creation of new cgroups
4016 * even when there are much fewer than 64k cgroups - possibly none.
4018 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4019 * be freed and recycled when it's no longer needed, which is usually
4020 * when the CSS is offlined.
4022 * The only exception to that are records of swapped out tmpfs/shmem
4023 * pages that need to be attributed to live ancestors on swapin. But
4024 * those references are manageable from userspace.
4027 static DEFINE_IDR(mem_cgroup_idr
);
4029 static void mem_cgroup_id_get_many(struct mem_cgroup
*memcg
, unsigned int n
)
4031 VM_BUG_ON(atomic_read(&memcg
->id
.ref
) <= 0);
4032 atomic_add(n
, &memcg
->id
.ref
);
4035 static void mem_cgroup_id_put_many(struct mem_cgroup
*memcg
, unsigned int n
)
4037 VM_BUG_ON(atomic_read(&memcg
->id
.ref
) < n
);
4038 if (atomic_sub_and_test(n
, &memcg
->id
.ref
)) {
4039 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
4042 /* Memcg ID pins CSS */
4043 css_put(&memcg
->css
);
4047 static inline void mem_cgroup_id_get(struct mem_cgroup
*memcg
)
4049 mem_cgroup_id_get_many(memcg
, 1);
4052 static inline void mem_cgroup_id_put(struct mem_cgroup
*memcg
)
4054 mem_cgroup_id_put_many(memcg
, 1);
4058 * mem_cgroup_from_id - look up a memcg from a memcg id
4059 * @id: the memcg id to look up
4061 * Caller must hold rcu_read_lock().
4063 struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
4065 WARN_ON_ONCE(!rcu_read_lock_held());
4066 return idr_find(&mem_cgroup_idr
, id
);
4069 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4071 struct mem_cgroup_per_node
*pn
;
4074 * This routine is called against possible nodes.
4075 * But it's BUG to call kmalloc() against offline node.
4077 * TODO: this routine can waste much memory for nodes which will
4078 * never be onlined. It's better to use memory hotplug callback
4081 if (!node_state(node
, N_NORMAL_MEMORY
))
4083 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4087 pn
->lruvec_stat_cpu
= alloc_percpu(struct lruvec_stat
);
4088 if (!pn
->lruvec_stat_cpu
) {
4093 lruvec_init(&pn
->lruvec
);
4094 pn
->usage_in_excess
= 0;
4095 pn
->on_tree
= false;
4098 memcg
->nodeinfo
[node
] = pn
;
4102 static void free_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4104 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[node
];
4109 free_percpu(pn
->lruvec_stat_cpu
);
4113 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
4118 free_mem_cgroup_per_node_info(memcg
, node
);
4119 free_percpu(memcg
->stat_cpu
);
4123 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
4125 memcg_wb_domain_exit(memcg
);
4126 __mem_cgroup_free(memcg
);
4129 static struct mem_cgroup
*mem_cgroup_alloc(void)
4131 struct mem_cgroup
*memcg
;
4135 size
= sizeof(struct mem_cgroup
);
4136 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4138 memcg
= kzalloc(size
, GFP_KERNEL
);
4142 memcg
->id
.id
= idr_alloc(&mem_cgroup_idr
, NULL
,
4143 1, MEM_CGROUP_ID_MAX
,
4145 if (memcg
->id
.id
< 0)
4148 memcg
->stat_cpu
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4149 if (!memcg
->stat_cpu
)
4153 if (alloc_mem_cgroup_per_node_info(memcg
, node
))
4156 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
4159 INIT_WORK(&memcg
->high_work
, high_work_func
);
4160 memcg
->last_scanned_node
= MAX_NUMNODES
;
4161 INIT_LIST_HEAD(&memcg
->oom_notify
);
4162 mutex_init(&memcg
->thresholds_lock
);
4163 spin_lock_init(&memcg
->move_lock
);
4164 vmpressure_init(&memcg
->vmpressure
);
4165 INIT_LIST_HEAD(&memcg
->event_list
);
4166 spin_lock_init(&memcg
->event_list_lock
);
4167 memcg
->socket_pressure
= jiffies
;
4169 memcg
->kmemcg_id
= -1;
4171 #ifdef CONFIG_CGROUP_WRITEBACK
4172 INIT_LIST_HEAD(&memcg
->cgwb_list
);
4174 idr_replace(&mem_cgroup_idr
, memcg
, memcg
->id
.id
);
4177 if (memcg
->id
.id
> 0)
4178 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
4179 __mem_cgroup_free(memcg
);
4183 static struct cgroup_subsys_state
* __ref
4184 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4186 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
4187 struct mem_cgroup
*memcg
;
4188 long error
= -ENOMEM
;
4190 memcg
= mem_cgroup_alloc();
4192 return ERR_PTR(error
);
4194 memcg
->high
= PAGE_COUNTER_MAX
;
4195 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4197 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4198 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4200 if (parent
&& parent
->use_hierarchy
) {
4201 memcg
->use_hierarchy
= true;
4202 page_counter_init(&memcg
->memory
, &parent
->memory
);
4203 page_counter_init(&memcg
->swap
, &parent
->swap
);
4204 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4205 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4206 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
4208 page_counter_init(&memcg
->memory
, NULL
);
4209 page_counter_init(&memcg
->swap
, NULL
);
4210 page_counter_init(&memcg
->memsw
, NULL
);
4211 page_counter_init(&memcg
->kmem
, NULL
);
4212 page_counter_init(&memcg
->tcpmem
, NULL
);
4214 * Deeper hierachy with use_hierarchy == false doesn't make
4215 * much sense so let cgroup subsystem know about this
4216 * unfortunate state in our controller.
4218 if (parent
!= root_mem_cgroup
)
4219 memory_cgrp_subsys
.broken_hierarchy
= true;
4222 /* The following stuff does not apply to the root */
4224 root_mem_cgroup
= memcg
;
4228 error
= memcg_online_kmem(memcg
);
4232 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4233 static_branch_inc(&memcg_sockets_enabled_key
);
4237 mem_cgroup_free(memcg
);
4238 return ERR_PTR(-ENOMEM
);
4241 static int mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
4243 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4245 /* Online state pins memcg ID, memcg ID pins CSS */
4246 atomic_set(&memcg
->id
.ref
, 1);
4251 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
4253 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4254 struct mem_cgroup_event
*event
, *tmp
;
4257 * Unregister events and notify userspace.
4258 * Notify userspace about cgroup removing only after rmdir of cgroup
4259 * directory to avoid race between userspace and kernelspace.
4261 spin_lock(&memcg
->event_list_lock
);
4262 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
4263 list_del_init(&event
->list
);
4264 schedule_work(&event
->remove
);
4266 spin_unlock(&memcg
->event_list_lock
);
4268 page_counter_set_low(&memcg
->memory
, 0);
4270 memcg_offline_kmem(memcg
);
4271 wb_memcg_offline(memcg
);
4273 mem_cgroup_id_put(memcg
);
4276 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
4278 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4280 invalidate_reclaim_iterators(memcg
);
4283 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
4285 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4287 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4288 static_branch_dec(&memcg_sockets_enabled_key
);
4290 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
4291 static_branch_dec(&memcg_sockets_enabled_key
);
4293 vmpressure_cleanup(&memcg
->vmpressure
);
4294 cancel_work_sync(&memcg
->high_work
);
4295 mem_cgroup_remove_from_trees(memcg
);
4296 memcg_free_kmem(memcg
);
4297 mem_cgroup_free(memcg
);
4301 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4302 * @css: the target css
4304 * Reset the states of the mem_cgroup associated with @css. This is
4305 * invoked when the userland requests disabling on the default hierarchy
4306 * but the memcg is pinned through dependency. The memcg should stop
4307 * applying policies and should revert to the vanilla state as it may be
4308 * made visible again.
4310 * The current implementation only resets the essential configurations.
4311 * This needs to be expanded to cover all the visible parts.
4313 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
4315 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4317 page_counter_set_max(&memcg
->memory
, PAGE_COUNTER_MAX
);
4318 page_counter_set_max(&memcg
->swap
, PAGE_COUNTER_MAX
);
4319 page_counter_set_max(&memcg
->memsw
, PAGE_COUNTER_MAX
);
4320 page_counter_set_max(&memcg
->kmem
, PAGE_COUNTER_MAX
);
4321 page_counter_set_max(&memcg
->tcpmem
, PAGE_COUNTER_MAX
);
4322 page_counter_set_low(&memcg
->memory
, 0);
4323 memcg
->high
= PAGE_COUNTER_MAX
;
4324 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4325 memcg_wb_domain_size_changed(memcg
);
4329 /* Handlers for move charge at task migration. */
4330 static int mem_cgroup_do_precharge(unsigned long count
)
4334 /* Try a single bulk charge without reclaim first, kswapd may wake */
4335 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
4337 mc
.precharge
+= count
;
4341 /* Try charges one by one with reclaim, but do not retry */
4343 ret
= try_charge(mc
.to
, GFP_KERNEL
| __GFP_NORETRY
, 1);
4357 enum mc_target_type
{
4364 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4365 unsigned long addr
, pte_t ptent
)
4367 struct page
*page
= _vm_normal_page(vma
, addr
, ptent
, true);
4369 if (!page
|| !page_mapped(page
))
4371 if (PageAnon(page
)) {
4372 if (!(mc
.flags
& MOVE_ANON
))
4375 if (!(mc
.flags
& MOVE_FILE
))
4378 if (!get_page_unless_zero(page
))
4384 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
4385 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4386 pte_t ptent
, swp_entry_t
*entry
)
4388 struct page
*page
= NULL
;
4389 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4391 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
4395 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
4396 * a device and because they are not accessible by CPU they are store
4397 * as special swap entry in the CPU page table.
4399 if (is_device_private_entry(ent
)) {
4400 page
= device_private_entry_to_page(ent
);
4402 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
4403 * a refcount of 1 when free (unlike normal page)
4405 if (!page_ref_add_unless(page
, 1, 1))
4411 * Because lookup_swap_cache() updates some statistics counter,
4412 * we call find_get_page() with swapper_space directly.
4414 page
= find_get_page(swap_address_space(ent
), swp_offset(ent
));
4415 if (do_memsw_account())
4416 entry
->val
= ent
.val
;
4421 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4422 pte_t ptent
, swp_entry_t
*entry
)
4428 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4429 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4431 struct page
*page
= NULL
;
4432 struct address_space
*mapping
;
4435 if (!vma
->vm_file
) /* anonymous vma */
4437 if (!(mc
.flags
& MOVE_FILE
))
4440 mapping
= vma
->vm_file
->f_mapping
;
4441 pgoff
= linear_page_index(vma
, addr
);
4443 /* page is moved even if it's not RSS of this task(page-faulted). */
4445 /* shmem/tmpfs may report page out on swap: account for that too. */
4446 if (shmem_mapping(mapping
)) {
4447 page
= find_get_entry(mapping
, pgoff
);
4448 if (radix_tree_exceptional_entry(page
)) {
4449 swp_entry_t swp
= radix_to_swp_entry(page
);
4450 if (do_memsw_account())
4452 page
= find_get_page(swap_address_space(swp
),
4456 page
= find_get_page(mapping
, pgoff
);
4458 page
= find_get_page(mapping
, pgoff
);
4464 * mem_cgroup_move_account - move account of the page
4466 * @compound: charge the page as compound or small page
4467 * @from: mem_cgroup which the page is moved from.
4468 * @to: mem_cgroup which the page is moved to. @from != @to.
4470 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4472 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4475 static int mem_cgroup_move_account(struct page
*page
,
4477 struct mem_cgroup
*from
,
4478 struct mem_cgroup
*to
)
4480 unsigned long flags
;
4481 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
4485 VM_BUG_ON(from
== to
);
4486 VM_BUG_ON_PAGE(PageLRU(page
), page
);
4487 VM_BUG_ON(compound
&& !PageTransHuge(page
));
4490 * Prevent mem_cgroup_migrate() from looking at
4491 * page->mem_cgroup of its source page while we change it.
4494 if (!trylock_page(page
))
4498 if (page
->mem_cgroup
!= from
)
4501 anon
= PageAnon(page
);
4503 spin_lock_irqsave(&from
->move_lock
, flags
);
4505 if (!anon
&& page_mapped(page
)) {
4506 __mod_memcg_state(from
, NR_FILE_MAPPED
, -nr_pages
);
4507 __mod_memcg_state(to
, NR_FILE_MAPPED
, nr_pages
);
4511 * move_lock grabbed above and caller set from->moving_account, so
4512 * mod_memcg_page_state will serialize updates to PageDirty.
4513 * So mapping should be stable for dirty pages.
4515 if (!anon
&& PageDirty(page
)) {
4516 struct address_space
*mapping
= page_mapping(page
);
4518 if (mapping_cap_account_dirty(mapping
)) {
4519 __mod_memcg_state(from
, NR_FILE_DIRTY
, -nr_pages
);
4520 __mod_memcg_state(to
, NR_FILE_DIRTY
, nr_pages
);
4524 if (PageWriteback(page
)) {
4525 __mod_memcg_state(from
, NR_WRITEBACK
, -nr_pages
);
4526 __mod_memcg_state(to
, NR_WRITEBACK
, nr_pages
);
4530 * It is safe to change page->mem_cgroup here because the page
4531 * is referenced, charged, and isolated - we can't race with
4532 * uncharging, charging, migration, or LRU putback.
4535 /* caller should have done css_get */
4536 page
->mem_cgroup
= to
;
4537 spin_unlock_irqrestore(&from
->move_lock
, flags
);
4541 local_irq_disable();
4542 mem_cgroup_charge_statistics(to
, page
, compound
, nr_pages
);
4543 memcg_check_events(to
, page
);
4544 mem_cgroup_charge_statistics(from
, page
, compound
, -nr_pages
);
4545 memcg_check_events(from
, page
);
4554 * get_mctgt_type - get target type of moving charge
4555 * @vma: the vma the pte to be checked belongs
4556 * @addr: the address corresponding to the pte to be checked
4557 * @ptent: the pte to be checked
4558 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4561 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4562 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4563 * move charge. if @target is not NULL, the page is stored in target->page
4564 * with extra refcnt got(Callers should handle it).
4565 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4566 * target for charge migration. if @target is not NULL, the entry is stored
4568 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PUBLIC
4569 * or MEMORY_DEVICE_PRIVATE (so ZONE_DEVICE page and thus not on the lru).
4570 * For now we such page is charge like a regular page would be as for all
4571 * intent and purposes it is just special memory taking the place of a
4574 * See Documentations/vm/hmm.txt and include/linux/hmm.h
4576 * Called with pte lock held.
4579 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
4580 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4582 struct page
*page
= NULL
;
4583 enum mc_target_type ret
= MC_TARGET_NONE
;
4584 swp_entry_t ent
= { .val
= 0 };
4586 if (pte_present(ptent
))
4587 page
= mc_handle_present_pte(vma
, addr
, ptent
);
4588 else if (is_swap_pte(ptent
))
4589 page
= mc_handle_swap_pte(vma
, ptent
, &ent
);
4590 else if (pte_none(ptent
))
4591 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
4593 if (!page
&& !ent
.val
)
4597 * Do only loose check w/o serialization.
4598 * mem_cgroup_move_account() checks the page is valid or
4599 * not under LRU exclusion.
4601 if (page
->mem_cgroup
== mc
.from
) {
4602 ret
= MC_TARGET_PAGE
;
4603 if (is_device_private_page(page
) ||
4604 is_device_public_page(page
))
4605 ret
= MC_TARGET_DEVICE
;
4607 target
->page
= page
;
4609 if (!ret
|| !target
)
4613 * There is a swap entry and a page doesn't exist or isn't charged.
4614 * But we cannot move a tail-page in a THP.
4616 if (ent
.val
&& !ret
&& (!page
|| !PageTransCompound(page
)) &&
4617 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
4618 ret
= MC_TARGET_SWAP
;
4625 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4627 * We don't consider PMD mapped swapping or file mapped pages because THP does
4628 * not support them for now.
4629 * Caller should make sure that pmd_trans_huge(pmd) is true.
4631 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4632 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4634 struct page
*page
= NULL
;
4635 enum mc_target_type ret
= MC_TARGET_NONE
;
4637 if (unlikely(is_swap_pmd(pmd
))) {
4638 VM_BUG_ON(thp_migration_supported() &&
4639 !is_pmd_migration_entry(pmd
));
4642 page
= pmd_page(pmd
);
4643 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
4644 if (!(mc
.flags
& MOVE_ANON
))
4646 if (page
->mem_cgroup
== mc
.from
) {
4647 ret
= MC_TARGET_PAGE
;
4650 target
->page
= page
;
4656 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4657 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4659 return MC_TARGET_NONE
;
4663 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
4664 unsigned long addr
, unsigned long end
,
4665 struct mm_walk
*walk
)
4667 struct vm_area_struct
*vma
= walk
->vma
;
4671 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4674 * Note their can not be MC_TARGET_DEVICE for now as we do not
4675 * support transparent huge page with MEMORY_DEVICE_PUBLIC or
4676 * MEMORY_DEVICE_PRIVATE but this might change.
4678 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
4679 mc
.precharge
+= HPAGE_PMD_NR
;
4684 if (pmd_trans_unstable(pmd
))
4686 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4687 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
4688 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
4689 mc
.precharge
++; /* increment precharge temporarily */
4690 pte_unmap_unlock(pte
- 1, ptl
);
4696 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
4698 unsigned long precharge
;
4700 struct mm_walk mem_cgroup_count_precharge_walk
= {
4701 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
4704 down_read(&mm
->mmap_sem
);
4705 walk_page_range(0, mm
->highest_vm_end
,
4706 &mem_cgroup_count_precharge_walk
);
4707 up_read(&mm
->mmap_sem
);
4709 precharge
= mc
.precharge
;
4715 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
4717 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
4719 VM_BUG_ON(mc
.moving_task
);
4720 mc
.moving_task
= current
;
4721 return mem_cgroup_do_precharge(precharge
);
4724 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4725 static void __mem_cgroup_clear_mc(void)
4727 struct mem_cgroup
*from
= mc
.from
;
4728 struct mem_cgroup
*to
= mc
.to
;
4730 /* we must uncharge all the leftover precharges from mc.to */
4732 cancel_charge(mc
.to
, mc
.precharge
);
4736 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4737 * we must uncharge here.
4739 if (mc
.moved_charge
) {
4740 cancel_charge(mc
.from
, mc
.moved_charge
);
4741 mc
.moved_charge
= 0;
4743 /* we must fixup refcnts and charges */
4744 if (mc
.moved_swap
) {
4745 /* uncharge swap account from the old cgroup */
4746 if (!mem_cgroup_is_root(mc
.from
))
4747 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
4749 mem_cgroup_id_put_many(mc
.from
, mc
.moved_swap
);
4752 * we charged both to->memory and to->memsw, so we
4753 * should uncharge to->memory.
4755 if (!mem_cgroup_is_root(mc
.to
))
4756 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
4758 mem_cgroup_id_get_many(mc
.to
, mc
.moved_swap
);
4759 css_put_many(&mc
.to
->css
, mc
.moved_swap
);
4763 memcg_oom_recover(from
);
4764 memcg_oom_recover(to
);
4765 wake_up_all(&mc
.waitq
);
4768 static void mem_cgroup_clear_mc(void)
4770 struct mm_struct
*mm
= mc
.mm
;
4773 * we must clear moving_task before waking up waiters at the end of
4776 mc
.moving_task
= NULL
;
4777 __mem_cgroup_clear_mc();
4778 spin_lock(&mc
.lock
);
4782 spin_unlock(&mc
.lock
);
4787 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
4789 struct cgroup_subsys_state
*css
;
4790 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
4791 struct mem_cgroup
*from
;
4792 struct task_struct
*leader
, *p
;
4793 struct mm_struct
*mm
;
4794 unsigned long move_flags
;
4797 /* charge immigration isn't supported on the default hierarchy */
4798 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
4802 * Multi-process migrations only happen on the default hierarchy
4803 * where charge immigration is not used. Perform charge
4804 * immigration if @tset contains a leader and whine if there are
4808 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
4811 memcg
= mem_cgroup_from_css(css
);
4817 * We are now commited to this value whatever it is. Changes in this
4818 * tunable will only affect upcoming migrations, not the current one.
4819 * So we need to save it, and keep it going.
4821 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
4825 from
= mem_cgroup_from_task(p
);
4827 VM_BUG_ON(from
== memcg
);
4829 mm
= get_task_mm(p
);
4832 /* We move charges only when we move a owner of the mm */
4833 if (mm
->owner
== p
) {
4836 VM_BUG_ON(mc
.precharge
);
4837 VM_BUG_ON(mc
.moved_charge
);
4838 VM_BUG_ON(mc
.moved_swap
);
4840 spin_lock(&mc
.lock
);
4844 mc
.flags
= move_flags
;
4845 spin_unlock(&mc
.lock
);
4846 /* We set mc.moving_task later */
4848 ret
= mem_cgroup_precharge_mc(mm
);
4850 mem_cgroup_clear_mc();
4857 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
4860 mem_cgroup_clear_mc();
4863 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
4864 unsigned long addr
, unsigned long end
,
4865 struct mm_walk
*walk
)
4868 struct vm_area_struct
*vma
= walk
->vma
;
4871 enum mc_target_type target_type
;
4872 union mc_target target
;
4875 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4877 if (mc
.precharge
< HPAGE_PMD_NR
) {
4881 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
4882 if (target_type
== MC_TARGET_PAGE
) {
4884 if (!isolate_lru_page(page
)) {
4885 if (!mem_cgroup_move_account(page
, true,
4887 mc
.precharge
-= HPAGE_PMD_NR
;
4888 mc
.moved_charge
+= HPAGE_PMD_NR
;
4890 putback_lru_page(page
);
4893 } else if (target_type
== MC_TARGET_DEVICE
) {
4895 if (!mem_cgroup_move_account(page
, true,
4897 mc
.precharge
-= HPAGE_PMD_NR
;
4898 mc
.moved_charge
+= HPAGE_PMD_NR
;
4906 if (pmd_trans_unstable(pmd
))
4909 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4910 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
4911 pte_t ptent
= *(pte
++);
4912 bool device
= false;
4918 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
4919 case MC_TARGET_DEVICE
:
4922 case MC_TARGET_PAGE
:
4925 * We can have a part of the split pmd here. Moving it
4926 * can be done but it would be too convoluted so simply
4927 * ignore such a partial THP and keep it in original
4928 * memcg. There should be somebody mapping the head.
4930 if (PageTransCompound(page
))
4932 if (!device
&& isolate_lru_page(page
))
4934 if (!mem_cgroup_move_account(page
, false,
4937 /* we uncharge from mc.from later. */
4941 putback_lru_page(page
);
4942 put
: /* get_mctgt_type() gets the page */
4945 case MC_TARGET_SWAP
:
4947 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
4949 /* we fixup refcnts and charges later. */
4957 pte_unmap_unlock(pte
- 1, ptl
);
4962 * We have consumed all precharges we got in can_attach().
4963 * We try charge one by one, but don't do any additional
4964 * charges to mc.to if we have failed in charge once in attach()
4967 ret
= mem_cgroup_do_precharge(1);
4975 static void mem_cgroup_move_charge(void)
4977 struct mm_walk mem_cgroup_move_charge_walk
= {
4978 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
4982 lru_add_drain_all();
4984 * Signal lock_page_memcg() to take the memcg's move_lock
4985 * while we're moving its pages to another memcg. Then wait
4986 * for already started RCU-only updates to finish.
4988 atomic_inc(&mc
.from
->moving_account
);
4991 if (unlikely(!down_read_trylock(&mc
.mm
->mmap_sem
))) {
4993 * Someone who are holding the mmap_sem might be waiting in
4994 * waitq. So we cancel all extra charges, wake up all waiters,
4995 * and retry. Because we cancel precharges, we might not be able
4996 * to move enough charges, but moving charge is a best-effort
4997 * feature anyway, so it wouldn't be a big problem.
4999 __mem_cgroup_clear_mc();
5004 * When we have consumed all precharges and failed in doing
5005 * additional charge, the page walk just aborts.
5007 walk_page_range(0, mc
.mm
->highest_vm_end
, &mem_cgroup_move_charge_walk
);
5009 up_read(&mc
.mm
->mmap_sem
);
5010 atomic_dec(&mc
.from
->moving_account
);
5013 static void mem_cgroup_move_task(void)
5016 mem_cgroup_move_charge();
5017 mem_cgroup_clear_mc();
5020 #else /* !CONFIG_MMU */
5021 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
5025 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
5028 static void mem_cgroup_move_task(void)
5034 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5035 * to verify whether we're attached to the default hierarchy on each mount
5038 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
5041 * use_hierarchy is forced on the default hierarchy. cgroup core
5042 * guarantees that @root doesn't have any children, so turning it
5043 * on for the root memcg is enough.
5045 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5046 root_mem_cgroup
->use_hierarchy
= true;
5048 root_mem_cgroup
->use_hierarchy
= false;
5051 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
5054 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5056 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
5059 static int memory_low_show(struct seq_file
*m
, void *v
)
5061 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5062 unsigned long low
= READ_ONCE(memcg
->memory
.low
);
5064 if (low
== PAGE_COUNTER_MAX
)
5065 seq_puts(m
, "max\n");
5067 seq_printf(m
, "%llu\n", (u64
)low
* PAGE_SIZE
);
5072 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
5073 char *buf
, size_t nbytes
, loff_t off
)
5075 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5079 buf
= strstrip(buf
);
5080 err
= page_counter_memparse(buf
, "max", &low
);
5084 page_counter_set_low(&memcg
->memory
, low
);
5089 static int memory_high_show(struct seq_file
*m
, void *v
)
5091 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5092 unsigned long high
= READ_ONCE(memcg
->high
);
5094 if (high
== PAGE_COUNTER_MAX
)
5095 seq_puts(m
, "max\n");
5097 seq_printf(m
, "%llu\n", (u64
)high
* PAGE_SIZE
);
5102 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
5103 char *buf
, size_t nbytes
, loff_t off
)
5105 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5106 unsigned long nr_pages
;
5110 buf
= strstrip(buf
);
5111 err
= page_counter_memparse(buf
, "max", &high
);
5117 nr_pages
= page_counter_read(&memcg
->memory
);
5118 if (nr_pages
> high
)
5119 try_to_free_mem_cgroup_pages(memcg
, nr_pages
- high
,
5122 memcg_wb_domain_size_changed(memcg
);
5126 static int memory_max_show(struct seq_file
*m
, void *v
)
5128 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5129 unsigned long max
= READ_ONCE(memcg
->memory
.max
);
5131 if (max
== PAGE_COUNTER_MAX
)
5132 seq_puts(m
, "max\n");
5134 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5139 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
5140 char *buf
, size_t nbytes
, loff_t off
)
5142 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5143 unsigned int nr_reclaims
= MEM_CGROUP_RECLAIM_RETRIES
;
5144 bool drained
= false;
5148 buf
= strstrip(buf
);
5149 err
= page_counter_memparse(buf
, "max", &max
);
5153 xchg(&memcg
->memory
.max
, max
);
5156 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
5158 if (nr_pages
<= max
)
5161 if (signal_pending(current
)) {
5167 drain_all_stock(memcg
);
5173 if (!try_to_free_mem_cgroup_pages(memcg
, nr_pages
- max
,
5179 memcg_memory_event(memcg
, MEMCG_OOM
);
5180 if (!mem_cgroup_out_of_memory(memcg
, GFP_KERNEL
, 0))
5184 memcg_wb_domain_size_changed(memcg
);
5188 static int memory_events_show(struct seq_file
*m
, void *v
)
5190 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5192 seq_printf(m
, "low %lu\n",
5193 atomic_long_read(&memcg
->memory_events
[MEMCG_LOW
]));
5194 seq_printf(m
, "high %lu\n",
5195 atomic_long_read(&memcg
->memory_events
[MEMCG_HIGH
]));
5196 seq_printf(m
, "max %lu\n",
5197 atomic_long_read(&memcg
->memory_events
[MEMCG_MAX
]));
5198 seq_printf(m
, "oom %lu\n",
5199 atomic_long_read(&memcg
->memory_events
[MEMCG_OOM
]));
5200 seq_printf(m
, "oom_kill %lu\n", memcg_sum_events(memcg
, OOM_KILL
));
5205 static int memory_stat_show(struct seq_file
*m
, void *v
)
5207 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5208 unsigned long stat
[MEMCG_NR_STAT
];
5209 unsigned long events
[NR_VM_EVENT_ITEMS
];
5213 * Provide statistics on the state of the memory subsystem as
5214 * well as cumulative event counters that show past behavior.
5216 * This list is ordered following a combination of these gradients:
5217 * 1) generic big picture -> specifics and details
5218 * 2) reflecting userspace activity -> reflecting kernel heuristics
5220 * Current memory state:
5223 tree_stat(memcg
, stat
);
5224 tree_events(memcg
, events
);
5226 seq_printf(m
, "anon %llu\n",
5227 (u64
)stat
[MEMCG_RSS
] * PAGE_SIZE
);
5228 seq_printf(m
, "file %llu\n",
5229 (u64
)stat
[MEMCG_CACHE
] * PAGE_SIZE
);
5230 seq_printf(m
, "kernel_stack %llu\n",
5231 (u64
)stat
[MEMCG_KERNEL_STACK_KB
] * 1024);
5232 seq_printf(m
, "slab %llu\n",
5233 (u64
)(stat
[NR_SLAB_RECLAIMABLE
] +
5234 stat
[NR_SLAB_UNRECLAIMABLE
]) * PAGE_SIZE
);
5235 seq_printf(m
, "sock %llu\n",
5236 (u64
)stat
[MEMCG_SOCK
] * PAGE_SIZE
);
5238 seq_printf(m
, "shmem %llu\n",
5239 (u64
)stat
[NR_SHMEM
] * PAGE_SIZE
);
5240 seq_printf(m
, "file_mapped %llu\n",
5241 (u64
)stat
[NR_FILE_MAPPED
] * PAGE_SIZE
);
5242 seq_printf(m
, "file_dirty %llu\n",
5243 (u64
)stat
[NR_FILE_DIRTY
] * PAGE_SIZE
);
5244 seq_printf(m
, "file_writeback %llu\n",
5245 (u64
)stat
[NR_WRITEBACK
] * PAGE_SIZE
);
5247 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5248 struct mem_cgroup
*mi
;
5249 unsigned long val
= 0;
5251 for_each_mem_cgroup_tree(mi
, memcg
)
5252 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
));
5253 seq_printf(m
, "%s %llu\n",
5254 mem_cgroup_lru_names
[i
], (u64
)val
* PAGE_SIZE
);
5257 seq_printf(m
, "slab_reclaimable %llu\n",
5258 (u64
)stat
[NR_SLAB_RECLAIMABLE
] * PAGE_SIZE
);
5259 seq_printf(m
, "slab_unreclaimable %llu\n",
5260 (u64
)stat
[NR_SLAB_UNRECLAIMABLE
] * PAGE_SIZE
);
5262 /* Accumulated memory events */
5264 seq_printf(m
, "pgfault %lu\n", events
[PGFAULT
]);
5265 seq_printf(m
, "pgmajfault %lu\n", events
[PGMAJFAULT
]);
5267 seq_printf(m
, "pgrefill %lu\n", events
[PGREFILL
]);
5268 seq_printf(m
, "pgscan %lu\n", events
[PGSCAN_KSWAPD
] +
5269 events
[PGSCAN_DIRECT
]);
5270 seq_printf(m
, "pgsteal %lu\n", events
[PGSTEAL_KSWAPD
] +
5271 events
[PGSTEAL_DIRECT
]);
5272 seq_printf(m
, "pgactivate %lu\n", events
[PGACTIVATE
]);
5273 seq_printf(m
, "pgdeactivate %lu\n", events
[PGDEACTIVATE
]);
5274 seq_printf(m
, "pglazyfree %lu\n", events
[PGLAZYFREE
]);
5275 seq_printf(m
, "pglazyfreed %lu\n", events
[PGLAZYFREED
]);
5277 seq_printf(m
, "workingset_refault %lu\n",
5278 stat
[WORKINGSET_REFAULT
]);
5279 seq_printf(m
, "workingset_activate %lu\n",
5280 stat
[WORKINGSET_ACTIVATE
]);
5281 seq_printf(m
, "workingset_nodereclaim %lu\n",
5282 stat
[WORKINGSET_NODERECLAIM
]);
5287 static struct cftype memory_files
[] = {
5290 .flags
= CFTYPE_NOT_ON_ROOT
,
5291 .read_u64
= memory_current_read
,
5295 .flags
= CFTYPE_NOT_ON_ROOT
,
5296 .seq_show
= memory_low_show
,
5297 .write
= memory_low_write
,
5301 .flags
= CFTYPE_NOT_ON_ROOT
,
5302 .seq_show
= memory_high_show
,
5303 .write
= memory_high_write
,
5307 .flags
= CFTYPE_NOT_ON_ROOT
,
5308 .seq_show
= memory_max_show
,
5309 .write
= memory_max_write
,
5313 .flags
= CFTYPE_NOT_ON_ROOT
,
5314 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
5315 .seq_show
= memory_events_show
,
5319 .flags
= CFTYPE_NOT_ON_ROOT
,
5320 .seq_show
= memory_stat_show
,
5325 struct cgroup_subsys memory_cgrp_subsys
= {
5326 .css_alloc
= mem_cgroup_css_alloc
,
5327 .css_online
= mem_cgroup_css_online
,
5328 .css_offline
= mem_cgroup_css_offline
,
5329 .css_released
= mem_cgroup_css_released
,
5330 .css_free
= mem_cgroup_css_free
,
5331 .css_reset
= mem_cgroup_css_reset
,
5332 .can_attach
= mem_cgroup_can_attach
,
5333 .cancel_attach
= mem_cgroup_cancel_attach
,
5334 .post_attach
= mem_cgroup_move_task
,
5335 .bind
= mem_cgroup_bind
,
5336 .dfl_cftypes
= memory_files
,
5337 .legacy_cftypes
= mem_cgroup_legacy_files
,
5342 * mem_cgroup_low - check if memory consumption is in the normal range
5343 * @root: the top ancestor of the sub-tree being checked
5344 * @memcg: the memory cgroup to check
5346 * WARNING: This function is not stateless! It can only be used as part
5347 * of a top-down tree iteration, not for isolated queries.
5349 * Returns %true if memory consumption of @memcg is in the normal range.
5351 * @root is exclusive; it is never low when looked at directly
5353 * To provide a proper hierarchical behavior, effective memory.low value
5356 * Effective memory.low is always equal or less than the original memory.low.
5357 * If there is no memory.low overcommittment (which is always true for
5358 * top-level memory cgroups), these two values are equal.
5359 * Otherwise, it's a part of parent's effective memory.low,
5360 * calculated as a cgroup's memory.low usage divided by sum of sibling's
5361 * memory.low usages, where memory.low usage is the size of actually
5365 * elow = min( memory.low, parent->elow * ------------------ ),
5366 * siblings_low_usage
5368 * | memory.current, if memory.current < memory.low
5373 * Such definition of the effective memory.low provides the expected
5374 * hierarchical behavior: parent's memory.low value is limiting
5375 * children, unprotected memory is reclaimed first and cgroups,
5376 * which are not using their guarantee do not affect actual memory
5379 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
5381 * A A/memory.low = 2G, A/memory.current = 6G
5383 * BC DE B/memory.low = 3G B/memory.current = 2G
5384 * C/memory.low = 1G C/memory.current = 2G
5385 * D/memory.low = 0 D/memory.current = 2G
5386 * E/memory.low = 10G E/memory.current = 0
5388 * and the memory pressure is applied, the following memory distribution
5389 * is expected (approximately):
5391 * A/memory.current = 2G
5393 * B/memory.current = 1.3G
5394 * C/memory.current = 0.6G
5395 * D/memory.current = 0
5396 * E/memory.current = 0
5398 * These calculations require constant tracking of the actual low usages
5399 * (see propagate_low_usage()), as well as recursive calculation of
5400 * effective memory.low values. But as we do call mem_cgroup_low()
5401 * path for each memory cgroup top-down from the reclaim,
5402 * it's possible to optimize this part, and save calculated elow
5403 * for next usage. This part is intentionally racy, but it's ok,
5404 * as memory.low is a best-effort mechanism.
5406 bool mem_cgroup_low(struct mem_cgroup
*root
, struct mem_cgroup
*memcg
)
5408 unsigned long usage
, low_usage
, siblings_low_usage
;
5409 unsigned long elow
, parent_elow
;
5410 struct mem_cgroup
*parent
;
5412 if (mem_cgroup_disabled())
5416 root
= root_mem_cgroup
;
5420 elow
= memcg
->memory
.low
;
5421 usage
= page_counter_read(&memcg
->memory
);
5422 parent
= parent_mem_cgroup(memcg
);
5427 parent_elow
= READ_ONCE(parent
->memory
.elow
);
5428 elow
= min(elow
, parent_elow
);
5430 if (!elow
|| !parent_elow
)
5433 low_usage
= min(usage
, memcg
->memory
.low
);
5434 siblings_low_usage
= atomic_long_read(
5435 &parent
->memory
.children_low_usage
);
5437 if (!low_usage
|| !siblings_low_usage
)
5440 elow
= min(elow
, parent_elow
* low_usage
/ siblings_low_usage
);
5442 memcg
->memory
.elow
= elow
;
5443 return usage
&& usage
<= elow
;
5447 * mem_cgroup_try_charge - try charging a page
5448 * @page: page to charge
5449 * @mm: mm context of the victim
5450 * @gfp_mask: reclaim mode
5451 * @memcgp: charged memcg return
5452 * @compound: charge the page as compound or small page
5454 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5455 * pages according to @gfp_mask if necessary.
5457 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5458 * Otherwise, an error code is returned.
5460 * After page->mapping has been set up, the caller must finalize the
5461 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5462 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5464 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
5465 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
,
5468 struct mem_cgroup
*memcg
= NULL
;
5469 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5472 if (mem_cgroup_disabled())
5475 if (PageSwapCache(page
)) {
5477 * Every swap fault against a single page tries to charge the
5478 * page, bail as early as possible. shmem_unuse() encounters
5479 * already charged pages, too. The USED bit is protected by
5480 * the page lock, which serializes swap cache removal, which
5481 * in turn serializes uncharging.
5483 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5484 if (compound_head(page
)->mem_cgroup
)
5487 if (do_swap_account
) {
5488 swp_entry_t ent
= { .val
= page_private(page
), };
5489 unsigned short id
= lookup_swap_cgroup_id(ent
);
5492 memcg
= mem_cgroup_from_id(id
);
5493 if (memcg
&& !css_tryget_online(&memcg
->css
))
5500 memcg
= get_mem_cgroup_from_mm(mm
);
5502 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
5504 css_put(&memcg
->css
);
5511 * mem_cgroup_commit_charge - commit a page charge
5512 * @page: page to charge
5513 * @memcg: memcg to charge the page to
5514 * @lrucare: page might be on LRU already
5515 * @compound: charge the page as compound or small page
5517 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5518 * after page->mapping has been set up. This must happen atomically
5519 * as part of the page instantiation, i.e. under the page table lock
5520 * for anonymous pages, under the page lock for page and swap cache.
5522 * In addition, the page must not be on the LRU during the commit, to
5523 * prevent racing with task migration. If it might be, use @lrucare.
5525 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5527 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5528 bool lrucare
, bool compound
)
5530 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5532 VM_BUG_ON_PAGE(!page
->mapping
, page
);
5533 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
5535 if (mem_cgroup_disabled())
5538 * Swap faults will attempt to charge the same page multiple
5539 * times. But reuse_swap_page() might have removed the page
5540 * from swapcache already, so we can't check PageSwapCache().
5545 commit_charge(page
, memcg
, lrucare
);
5547 local_irq_disable();
5548 mem_cgroup_charge_statistics(memcg
, page
, compound
, nr_pages
);
5549 memcg_check_events(memcg
, page
);
5552 if (do_memsw_account() && PageSwapCache(page
)) {
5553 swp_entry_t entry
= { .val
= page_private(page
) };
5555 * The swap entry might not get freed for a long time,
5556 * let's not wait for it. The page already received a
5557 * memory+swap charge, drop the swap entry duplicate.
5559 mem_cgroup_uncharge_swap(entry
, nr_pages
);
5564 * mem_cgroup_cancel_charge - cancel a page charge
5565 * @page: page to charge
5566 * @memcg: memcg to charge the page to
5567 * @compound: charge the page as compound or small page
5569 * Cancel a charge transaction started by mem_cgroup_try_charge().
5571 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5574 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5576 if (mem_cgroup_disabled())
5579 * Swap faults will attempt to charge the same page multiple
5580 * times. But reuse_swap_page() might have removed the page
5581 * from swapcache already, so we can't check PageSwapCache().
5586 cancel_charge(memcg
, nr_pages
);
5589 struct uncharge_gather
{
5590 struct mem_cgroup
*memcg
;
5591 unsigned long pgpgout
;
5592 unsigned long nr_anon
;
5593 unsigned long nr_file
;
5594 unsigned long nr_kmem
;
5595 unsigned long nr_huge
;
5596 unsigned long nr_shmem
;
5597 struct page
*dummy_page
;
5600 static inline void uncharge_gather_clear(struct uncharge_gather
*ug
)
5602 memset(ug
, 0, sizeof(*ug
));
5605 static void uncharge_batch(const struct uncharge_gather
*ug
)
5607 unsigned long nr_pages
= ug
->nr_anon
+ ug
->nr_file
+ ug
->nr_kmem
;
5608 unsigned long flags
;
5610 if (!mem_cgroup_is_root(ug
->memcg
)) {
5611 page_counter_uncharge(&ug
->memcg
->memory
, nr_pages
);
5612 if (do_memsw_account())
5613 page_counter_uncharge(&ug
->memcg
->memsw
, nr_pages
);
5614 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && ug
->nr_kmem
)
5615 page_counter_uncharge(&ug
->memcg
->kmem
, ug
->nr_kmem
);
5616 memcg_oom_recover(ug
->memcg
);
5619 local_irq_save(flags
);
5620 __mod_memcg_state(ug
->memcg
, MEMCG_RSS
, -ug
->nr_anon
);
5621 __mod_memcg_state(ug
->memcg
, MEMCG_CACHE
, -ug
->nr_file
);
5622 __mod_memcg_state(ug
->memcg
, MEMCG_RSS_HUGE
, -ug
->nr_huge
);
5623 __mod_memcg_state(ug
->memcg
, NR_SHMEM
, -ug
->nr_shmem
);
5624 __count_memcg_events(ug
->memcg
, PGPGOUT
, ug
->pgpgout
);
5625 __this_cpu_add(ug
->memcg
->stat_cpu
->nr_page_events
, nr_pages
);
5626 memcg_check_events(ug
->memcg
, ug
->dummy_page
);
5627 local_irq_restore(flags
);
5629 if (!mem_cgroup_is_root(ug
->memcg
))
5630 css_put_many(&ug
->memcg
->css
, nr_pages
);
5633 static void uncharge_page(struct page
*page
, struct uncharge_gather
*ug
)
5635 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5636 VM_BUG_ON_PAGE(page_count(page
) && !is_zone_device_page(page
) &&
5637 !PageHWPoison(page
) , page
);
5639 if (!page
->mem_cgroup
)
5643 * Nobody should be changing or seriously looking at
5644 * page->mem_cgroup at this point, we have fully
5645 * exclusive access to the page.
5648 if (ug
->memcg
!= page
->mem_cgroup
) {
5651 uncharge_gather_clear(ug
);
5653 ug
->memcg
= page
->mem_cgroup
;
5656 if (!PageKmemcg(page
)) {
5657 unsigned int nr_pages
= 1;
5659 if (PageTransHuge(page
)) {
5660 nr_pages
<<= compound_order(page
);
5661 ug
->nr_huge
+= nr_pages
;
5664 ug
->nr_anon
+= nr_pages
;
5666 ug
->nr_file
+= nr_pages
;
5667 if (PageSwapBacked(page
))
5668 ug
->nr_shmem
+= nr_pages
;
5672 ug
->nr_kmem
+= 1 << compound_order(page
);
5673 __ClearPageKmemcg(page
);
5676 ug
->dummy_page
= page
;
5677 page
->mem_cgroup
= NULL
;
5680 static void uncharge_list(struct list_head
*page_list
)
5682 struct uncharge_gather ug
;
5683 struct list_head
*next
;
5685 uncharge_gather_clear(&ug
);
5688 * Note that the list can be a single page->lru; hence the
5689 * do-while loop instead of a simple list_for_each_entry().
5691 next
= page_list
->next
;
5695 page
= list_entry(next
, struct page
, lru
);
5696 next
= page
->lru
.next
;
5698 uncharge_page(page
, &ug
);
5699 } while (next
!= page_list
);
5702 uncharge_batch(&ug
);
5706 * mem_cgroup_uncharge - uncharge a page
5707 * @page: page to uncharge
5709 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5710 * mem_cgroup_commit_charge().
5712 void mem_cgroup_uncharge(struct page
*page
)
5714 struct uncharge_gather ug
;
5716 if (mem_cgroup_disabled())
5719 /* Don't touch page->lru of any random page, pre-check: */
5720 if (!page
->mem_cgroup
)
5723 uncharge_gather_clear(&ug
);
5724 uncharge_page(page
, &ug
);
5725 uncharge_batch(&ug
);
5729 * mem_cgroup_uncharge_list - uncharge a list of page
5730 * @page_list: list of pages to uncharge
5732 * Uncharge a list of pages previously charged with
5733 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5735 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
5737 if (mem_cgroup_disabled())
5740 if (!list_empty(page_list
))
5741 uncharge_list(page_list
);
5745 * mem_cgroup_migrate - charge a page's replacement
5746 * @oldpage: currently circulating page
5747 * @newpage: replacement page
5749 * Charge @newpage as a replacement page for @oldpage. @oldpage will
5750 * be uncharged upon free.
5752 * Both pages must be locked, @newpage->mapping must be set up.
5754 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
)
5756 struct mem_cgroup
*memcg
;
5757 unsigned int nr_pages
;
5759 unsigned long flags
;
5761 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
5762 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
5763 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
5764 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
5767 if (mem_cgroup_disabled())
5770 /* Page cache replacement: new page already charged? */
5771 if (newpage
->mem_cgroup
)
5774 /* Swapcache readahead pages can get replaced before being charged */
5775 memcg
= oldpage
->mem_cgroup
;
5779 /* Force-charge the new page. The old one will be freed soon */
5780 compound
= PageTransHuge(newpage
);
5781 nr_pages
= compound
? hpage_nr_pages(newpage
) : 1;
5783 page_counter_charge(&memcg
->memory
, nr_pages
);
5784 if (do_memsw_account())
5785 page_counter_charge(&memcg
->memsw
, nr_pages
);
5786 css_get_many(&memcg
->css
, nr_pages
);
5788 commit_charge(newpage
, memcg
, false);
5790 local_irq_save(flags
);
5791 mem_cgroup_charge_statistics(memcg
, newpage
, compound
, nr_pages
);
5792 memcg_check_events(memcg
, newpage
);
5793 local_irq_restore(flags
);
5796 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
5797 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
5799 void mem_cgroup_sk_alloc(struct sock
*sk
)
5801 struct mem_cgroup
*memcg
;
5803 if (!mem_cgroup_sockets_enabled
)
5807 * Socket cloning can throw us here with sk_memcg already
5808 * filled. It won't however, necessarily happen from
5809 * process context. So the test for root memcg given
5810 * the current task's memcg won't help us in this case.
5812 * Respecting the original socket's memcg is a better
5813 * decision in this case.
5816 css_get(&sk
->sk_memcg
->css
);
5821 memcg
= mem_cgroup_from_task(current
);
5822 if (memcg
== root_mem_cgroup
)
5824 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
5826 if (css_tryget_online(&memcg
->css
))
5827 sk
->sk_memcg
= memcg
;
5832 void mem_cgroup_sk_free(struct sock
*sk
)
5835 css_put(&sk
->sk_memcg
->css
);
5839 * mem_cgroup_charge_skmem - charge socket memory
5840 * @memcg: memcg to charge
5841 * @nr_pages: number of pages to charge
5843 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5844 * @memcg's configured limit, %false if the charge had to be forced.
5846 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5848 gfp_t gfp_mask
= GFP_KERNEL
;
5850 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5851 struct page_counter
*fail
;
5853 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
5854 memcg
->tcpmem_pressure
= 0;
5857 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
5858 memcg
->tcpmem_pressure
= 1;
5862 /* Don't block in the packet receive path */
5864 gfp_mask
= GFP_NOWAIT
;
5866 mod_memcg_state(memcg
, MEMCG_SOCK
, nr_pages
);
5868 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0)
5871 try_charge(memcg
, gfp_mask
|__GFP_NOFAIL
, nr_pages
);
5876 * mem_cgroup_uncharge_skmem - uncharge socket memory
5877 * @memcg: memcg to uncharge
5878 * @nr_pages: number of pages to uncharge
5880 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5882 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5883 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
5887 mod_memcg_state(memcg
, MEMCG_SOCK
, -nr_pages
);
5889 refill_stock(memcg
, nr_pages
);
5892 static int __init
cgroup_memory(char *s
)
5896 while ((token
= strsep(&s
, ",")) != NULL
) {
5899 if (!strcmp(token
, "nosocket"))
5900 cgroup_memory_nosocket
= true;
5901 if (!strcmp(token
, "nokmem"))
5902 cgroup_memory_nokmem
= true;
5906 __setup("cgroup.memory=", cgroup_memory
);
5909 * subsys_initcall() for memory controller.
5911 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
5912 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
5913 * basically everything that doesn't depend on a specific mem_cgroup structure
5914 * should be initialized from here.
5916 static int __init
mem_cgroup_init(void)
5922 * Kmem cache creation is mostly done with the slab_mutex held,
5923 * so use a workqueue with limited concurrency to avoid stalling
5924 * all worker threads in case lots of cgroups are created and
5925 * destroyed simultaneously.
5927 memcg_kmem_cache_wq
= alloc_workqueue("memcg_kmem_cache", 0, 1);
5928 BUG_ON(!memcg_kmem_cache_wq
);
5931 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD
, "mm/memctrl:dead", NULL
,
5932 memcg_hotplug_cpu_dead
);
5934 for_each_possible_cpu(cpu
)
5935 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
5938 for_each_node(node
) {
5939 struct mem_cgroup_tree_per_node
*rtpn
;
5941 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
5942 node_online(node
) ? node
: NUMA_NO_NODE
);
5944 rtpn
->rb_root
= RB_ROOT
;
5945 rtpn
->rb_rightmost
= NULL
;
5946 spin_lock_init(&rtpn
->lock
);
5947 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
5952 subsys_initcall(mem_cgroup_init
);
5954 #ifdef CONFIG_MEMCG_SWAP
5955 static struct mem_cgroup
*mem_cgroup_id_get_online(struct mem_cgroup
*memcg
)
5957 while (!atomic_inc_not_zero(&memcg
->id
.ref
)) {
5959 * The root cgroup cannot be destroyed, so it's refcount must
5962 if (WARN_ON_ONCE(memcg
== root_mem_cgroup
)) {
5966 memcg
= parent_mem_cgroup(memcg
);
5968 memcg
= root_mem_cgroup
;
5974 * mem_cgroup_swapout - transfer a memsw charge to swap
5975 * @page: page whose memsw charge to transfer
5976 * @entry: swap entry to move the charge to
5978 * Transfer the memsw charge of @page to @entry.
5980 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
5982 struct mem_cgroup
*memcg
, *swap_memcg
;
5983 unsigned int nr_entries
;
5984 unsigned short oldid
;
5986 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5987 VM_BUG_ON_PAGE(page_count(page
), page
);
5989 if (!do_memsw_account())
5992 memcg
= page
->mem_cgroup
;
5994 /* Readahead page, never charged */
5999 * In case the memcg owning these pages has been offlined and doesn't
6000 * have an ID allocated to it anymore, charge the closest online
6001 * ancestor for the swap instead and transfer the memory+swap charge.
6003 swap_memcg
= mem_cgroup_id_get_online(memcg
);
6004 nr_entries
= hpage_nr_pages(page
);
6005 /* Get references for the tail pages, too */
6007 mem_cgroup_id_get_many(swap_memcg
, nr_entries
- 1);
6008 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(swap_memcg
),
6010 VM_BUG_ON_PAGE(oldid
, page
);
6011 mod_memcg_state(swap_memcg
, MEMCG_SWAP
, nr_entries
);
6013 page
->mem_cgroup
= NULL
;
6015 if (!mem_cgroup_is_root(memcg
))
6016 page_counter_uncharge(&memcg
->memory
, nr_entries
);
6018 if (memcg
!= swap_memcg
) {
6019 if (!mem_cgroup_is_root(swap_memcg
))
6020 page_counter_charge(&swap_memcg
->memsw
, nr_entries
);
6021 page_counter_uncharge(&memcg
->memsw
, nr_entries
);
6025 * Interrupts should be disabled here because the caller holds the
6026 * i_pages lock which is taken with interrupts-off. It is
6027 * important here to have the interrupts disabled because it is the
6028 * only synchronisation we have for updating the per-CPU variables.
6030 VM_BUG_ON(!irqs_disabled());
6031 mem_cgroup_charge_statistics(memcg
, page
, PageTransHuge(page
),
6033 memcg_check_events(memcg
, page
);
6035 if (!mem_cgroup_is_root(memcg
))
6036 css_put_many(&memcg
->css
, nr_entries
);
6040 * mem_cgroup_try_charge_swap - try charging swap space for a page
6041 * @page: page being added to swap
6042 * @entry: swap entry to charge
6044 * Try to charge @page's memcg for the swap space at @entry.
6046 * Returns 0 on success, -ENOMEM on failure.
6048 int mem_cgroup_try_charge_swap(struct page
*page
, swp_entry_t entry
)
6050 unsigned int nr_pages
= hpage_nr_pages(page
);
6051 struct page_counter
*counter
;
6052 struct mem_cgroup
*memcg
;
6053 unsigned short oldid
;
6055 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) || !do_swap_account
)
6058 memcg
= page
->mem_cgroup
;
6060 /* Readahead page, never charged */
6065 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
6069 memcg
= mem_cgroup_id_get_online(memcg
);
6071 if (!mem_cgroup_is_root(memcg
) &&
6072 !page_counter_try_charge(&memcg
->swap
, nr_pages
, &counter
)) {
6073 memcg_memory_event(memcg
, MEMCG_SWAP_MAX
);
6074 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
6075 mem_cgroup_id_put(memcg
);
6079 /* Get references for the tail pages, too */
6081 mem_cgroup_id_get_many(memcg
, nr_pages
- 1);
6082 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
), nr_pages
);
6083 VM_BUG_ON_PAGE(oldid
, page
);
6084 mod_memcg_state(memcg
, MEMCG_SWAP
, nr_pages
);
6090 * mem_cgroup_uncharge_swap - uncharge swap space
6091 * @entry: swap entry to uncharge
6092 * @nr_pages: the amount of swap space to uncharge
6094 void mem_cgroup_uncharge_swap(swp_entry_t entry
, unsigned int nr_pages
)
6096 struct mem_cgroup
*memcg
;
6099 if (!do_swap_account
)
6102 id
= swap_cgroup_record(entry
, 0, nr_pages
);
6104 memcg
= mem_cgroup_from_id(id
);
6106 if (!mem_cgroup_is_root(memcg
)) {
6107 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
6108 page_counter_uncharge(&memcg
->swap
, nr_pages
);
6110 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
6112 mod_memcg_state(memcg
, MEMCG_SWAP
, -nr_pages
);
6113 mem_cgroup_id_put_many(memcg
, nr_pages
);
6118 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
6120 long nr_swap_pages
= get_nr_swap_pages();
6122 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
6123 return nr_swap_pages
;
6124 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
6125 nr_swap_pages
= min_t(long, nr_swap_pages
,
6126 READ_ONCE(memcg
->swap
.max
) -
6127 page_counter_read(&memcg
->swap
));
6128 return nr_swap_pages
;
6131 bool mem_cgroup_swap_full(struct page
*page
)
6133 struct mem_cgroup
*memcg
;
6135 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
6139 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
6142 memcg
= page
->mem_cgroup
;
6146 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
6147 if (page_counter_read(&memcg
->swap
) * 2 >= memcg
->swap
.max
)
6153 /* for remember boot option*/
6154 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6155 static int really_do_swap_account __initdata
= 1;
6157 static int really_do_swap_account __initdata
;
6160 static int __init
enable_swap_account(char *s
)
6162 if (!strcmp(s
, "1"))
6163 really_do_swap_account
= 1;
6164 else if (!strcmp(s
, "0"))
6165 really_do_swap_account
= 0;
6168 __setup("swapaccount=", enable_swap_account
);
6170 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
6173 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6175 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
6178 static int swap_max_show(struct seq_file
*m
, void *v
)
6180 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
6181 unsigned long max
= READ_ONCE(memcg
->swap
.max
);
6183 if (max
== PAGE_COUNTER_MAX
)
6184 seq_puts(m
, "max\n");
6186 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
6191 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
6192 char *buf
, size_t nbytes
, loff_t off
)
6194 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6198 buf
= strstrip(buf
);
6199 err
= page_counter_memparse(buf
, "max", &max
);
6203 mutex_lock(&memcg_max_mutex
);
6204 err
= page_counter_set_max(&memcg
->swap
, max
);
6205 mutex_unlock(&memcg_max_mutex
);
6212 static int swap_events_show(struct seq_file
*m
, void *v
)
6214 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
6216 seq_printf(m
, "max %lu\n",
6217 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_MAX
]));
6218 seq_printf(m
, "fail %lu\n",
6219 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_FAIL
]));
6224 static struct cftype swap_files
[] = {
6226 .name
= "swap.current",
6227 .flags
= CFTYPE_NOT_ON_ROOT
,
6228 .read_u64
= swap_current_read
,
6232 .flags
= CFTYPE_NOT_ON_ROOT
,
6233 .seq_show
= swap_max_show
,
6234 .write
= swap_max_write
,
6237 .name
= "swap.events",
6238 .flags
= CFTYPE_NOT_ON_ROOT
,
6239 .file_offset
= offsetof(struct mem_cgroup
, swap_events_file
),
6240 .seq_show
= swap_events_show
,
6245 static struct cftype memsw_cgroup_files
[] = {
6247 .name
= "memsw.usage_in_bytes",
6248 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
6249 .read_u64
= mem_cgroup_read_u64
,
6252 .name
= "memsw.max_usage_in_bytes",
6253 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
6254 .write
= mem_cgroup_reset
,
6255 .read_u64
= mem_cgroup_read_u64
,
6258 .name
= "memsw.limit_in_bytes",
6259 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
6260 .write
= mem_cgroup_write
,
6261 .read_u64
= mem_cgroup_read_u64
,
6264 .name
= "memsw.failcnt",
6265 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
6266 .write
= mem_cgroup_reset
,
6267 .read_u64
= mem_cgroup_read_u64
,
6269 { }, /* terminate */
6272 static int __init
mem_cgroup_swap_init(void)
6274 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6275 do_swap_account
= 1;
6276 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
,
6278 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
6279 memsw_cgroup_files
));
6283 subsys_initcall(mem_cgroup_swap_init
);
6285 #endif /* CONFIG_MEMCG_SWAP */