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/hugetlb.h>
39 #include <linux/pagemap.h>
40 #include <linux/smp.h>
41 #include <linux/page-flags.h>
42 #include <linux/backing-dev.h>
43 #include <linux/bit_spinlock.h>
44 #include <linux/rcupdate.h>
45 #include <linux/limits.h>
46 #include <linux/export.h>
47 #include <linux/mutex.h>
48 #include <linux/rbtree.h>
49 #include <linux/slab.h>
50 #include <linux/swap.h>
51 #include <linux/swapops.h>
52 #include <linux/spinlock.h>
53 #include <linux/eventfd.h>
54 #include <linux/poll.h>
55 #include <linux/sort.h>
57 #include <linux/seq_file.h>
58 #include <linux/vmpressure.h>
59 #include <linux/mm_inline.h>
60 #include <linux/swap_cgroup.h>
61 #include <linux/cpu.h>
62 #include <linux/oom.h>
63 #include <linux/lockdep.h>
64 #include <linux/file.h>
68 #include <net/tcp_memcontrol.h>
71 #include <asm/uaccess.h>
73 #include <trace/events/vmscan.h>
75 struct cgroup_subsys memory_cgrp_subsys __read_mostly
;
76 EXPORT_SYMBOL(memory_cgrp_subsys
);
78 #define MEM_CGROUP_RECLAIM_RETRIES 5
79 static struct mem_cgroup
*root_mem_cgroup __read_mostly
;
81 /* Whether the swap controller is active */
82 #ifdef CONFIG_MEMCG_SWAP
83 int do_swap_account __read_mostly
;
85 #define do_swap_account 0
88 static const char * const mem_cgroup_stat_names
[] = {
97 static const char * const mem_cgroup_events_names
[] = {
104 static const char * const mem_cgroup_lru_names
[] = {
113 * Per memcg event counter is incremented at every pagein/pageout. With THP,
114 * it will be incremated by the number of pages. This counter is used for
115 * for trigger some periodic events. This is straightforward and better
116 * than using jiffies etc. to handle periodic memcg event.
118 enum mem_cgroup_events_target
{
119 MEM_CGROUP_TARGET_THRESH
,
120 MEM_CGROUP_TARGET_SOFTLIMIT
,
121 MEM_CGROUP_TARGET_NUMAINFO
,
124 #define THRESHOLDS_EVENTS_TARGET 128
125 #define SOFTLIMIT_EVENTS_TARGET 1024
126 #define NUMAINFO_EVENTS_TARGET 1024
128 struct mem_cgroup_stat_cpu
{
129 long count
[MEM_CGROUP_STAT_NSTATS
];
130 unsigned long events
[MEMCG_NR_EVENTS
];
131 unsigned long nr_page_events
;
132 unsigned long targets
[MEM_CGROUP_NTARGETS
];
135 struct reclaim_iter
{
136 struct mem_cgroup
*position
;
137 /* scan generation, increased every round-trip */
138 unsigned int generation
;
142 * per-zone information in memory controller.
144 struct mem_cgroup_per_zone
{
145 struct lruvec lruvec
;
146 unsigned long lru_size
[NR_LRU_LISTS
];
148 struct reclaim_iter iter
[DEF_PRIORITY
+ 1];
150 struct rb_node tree_node
; /* RB tree node */
151 unsigned long usage_in_excess
;/* Set to the value by which */
152 /* the soft limit is exceeded*/
154 struct mem_cgroup
*memcg
; /* Back pointer, we cannot */
155 /* use container_of */
158 struct mem_cgroup_per_node
{
159 struct mem_cgroup_per_zone zoneinfo
[MAX_NR_ZONES
];
163 * Cgroups above their limits are maintained in a RB-Tree, independent of
164 * their hierarchy representation
167 struct mem_cgroup_tree_per_zone
{
168 struct rb_root rb_root
;
172 struct mem_cgroup_tree_per_node
{
173 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
176 struct mem_cgroup_tree
{
177 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
180 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
182 struct mem_cgroup_threshold
{
183 struct eventfd_ctx
*eventfd
;
184 unsigned long threshold
;
188 struct mem_cgroup_threshold_ary
{
189 /* An array index points to threshold just below or equal to usage. */
190 int current_threshold
;
191 /* Size of entries[] */
193 /* Array of thresholds */
194 struct mem_cgroup_threshold entries
[0];
197 struct mem_cgroup_thresholds
{
198 /* Primary thresholds array */
199 struct mem_cgroup_threshold_ary
*primary
;
201 * Spare threshold array.
202 * This is needed to make mem_cgroup_unregister_event() "never fail".
203 * It must be able to store at least primary->size - 1 entries.
205 struct mem_cgroup_threshold_ary
*spare
;
209 struct mem_cgroup_eventfd_list
{
210 struct list_head list
;
211 struct eventfd_ctx
*eventfd
;
215 * cgroup_event represents events which userspace want to receive.
217 struct mem_cgroup_event
{
219 * memcg which the event belongs to.
221 struct mem_cgroup
*memcg
;
223 * eventfd to signal userspace about the event.
225 struct eventfd_ctx
*eventfd
;
227 * Each of these stored in a list by the cgroup.
229 struct list_head list
;
231 * register_event() callback will be used to add new userspace
232 * waiter for changes related to this event. Use eventfd_signal()
233 * on eventfd to send notification to userspace.
235 int (*register_event
)(struct mem_cgroup
*memcg
,
236 struct eventfd_ctx
*eventfd
, const char *args
);
238 * unregister_event() callback will be called when userspace closes
239 * the eventfd or on cgroup removing. This callback must be set,
240 * if you want provide notification functionality.
242 void (*unregister_event
)(struct mem_cgroup
*memcg
,
243 struct eventfd_ctx
*eventfd
);
245 * All fields below needed to unregister event when
246 * userspace closes eventfd.
249 wait_queue_head_t
*wqh
;
251 struct work_struct remove
;
254 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
255 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
258 * The memory controller data structure. The memory controller controls both
259 * page cache and RSS per cgroup. We would eventually like to provide
260 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
261 * to help the administrator determine what knobs to tune.
263 * TODO: Add a water mark for the memory controller. Reclaim will begin when
264 * we hit the water mark. May be even add a low water mark, such that
265 * no reclaim occurs from a cgroup at it's low water mark, this is
266 * a feature that will be implemented much later in the future.
269 struct cgroup_subsys_state css
;
271 /* Accounted resources */
272 struct page_counter memory
;
273 struct page_counter memsw
;
274 struct page_counter kmem
;
276 /* Normal memory consumption range */
280 unsigned long soft_limit
;
282 /* vmpressure notifications */
283 struct vmpressure vmpressure
;
285 /* css_online() has been completed */
289 * Should the accounting and control be hierarchical, per subtree?
295 atomic_t oom_wakeups
;
298 /* OOM-Killer disable */
299 int oom_kill_disable
;
301 /* protect arrays of thresholds */
302 struct mutex thresholds_lock
;
304 /* thresholds for memory usage. RCU-protected */
305 struct mem_cgroup_thresholds thresholds
;
307 /* thresholds for mem+swap usage. RCU-protected */
308 struct mem_cgroup_thresholds memsw_thresholds
;
310 /* For oom notifier event fd */
311 struct list_head oom_notify
;
314 * Should we move charges of a task when a task is moved into this
315 * mem_cgroup ? And what type of charges should we move ?
317 unsigned long move_charge_at_immigrate
;
319 * set > 0 if pages under this cgroup are moving to other cgroup.
321 atomic_t moving_account
;
322 /* taken only while moving_account > 0 */
323 spinlock_t move_lock
;
324 struct task_struct
*move_lock_task
;
325 unsigned long move_lock_flags
;
329 struct mem_cgroup_stat_cpu __percpu
*stat
;
331 * used when a cpu is offlined or other synchronizations
332 * See mem_cgroup_read_stat().
334 struct mem_cgroup_stat_cpu nocpu_base
;
335 spinlock_t pcp_counter_lock
;
337 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
338 struct cg_proto tcp_mem
;
340 #if defined(CONFIG_MEMCG_KMEM)
341 /* Index in the kmem_cache->memcg_params.memcg_caches array */
343 bool kmem_acct_activated
;
344 bool kmem_acct_active
;
347 int last_scanned_node
;
349 nodemask_t scan_nodes
;
350 atomic_t numainfo_events
;
351 atomic_t numainfo_updating
;
354 /* List of events which userspace want to receive */
355 struct list_head event_list
;
356 spinlock_t event_list_lock
;
358 struct mem_cgroup_per_node
*nodeinfo
[0];
359 /* WARNING: nodeinfo must be the last member here */
362 #ifdef CONFIG_MEMCG_KMEM
363 bool memcg_kmem_is_active(struct mem_cgroup
*memcg
)
365 return memcg
->kmem_acct_active
;
369 /* Stuffs for move charges at task migration. */
371 * Types of charges to be moved.
373 #define MOVE_ANON 0x1U
374 #define MOVE_FILE 0x2U
375 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
377 /* "mc" and its members are protected by cgroup_mutex */
378 static struct move_charge_struct
{
379 spinlock_t lock
; /* for from, to */
380 struct mem_cgroup
*from
;
381 struct mem_cgroup
*to
;
383 unsigned long precharge
;
384 unsigned long moved_charge
;
385 unsigned long moved_swap
;
386 struct task_struct
*moving_task
; /* a task moving charges */
387 wait_queue_head_t waitq
; /* a waitq for other context */
389 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
390 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
394 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
395 * limit reclaim to prevent infinite loops, if they ever occur.
397 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
398 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
401 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
402 MEM_CGROUP_CHARGE_TYPE_ANON
,
403 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
404 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
408 /* for encoding cft->private value on file */
416 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
417 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
418 #define MEMFILE_ATTR(val) ((val) & 0xffff)
419 /* Used for OOM nofiier */
420 #define OOM_CONTROL (0)
423 * The memcg_create_mutex will be held whenever a new cgroup is created.
424 * As a consequence, any change that needs to protect against new child cgroups
425 * appearing has to hold it as well.
427 static DEFINE_MUTEX(memcg_create_mutex
);
429 struct mem_cgroup
*mem_cgroup_from_css(struct cgroup_subsys_state
*s
)
431 return s
? container_of(s
, struct mem_cgroup
, css
) : NULL
;
434 /* Some nice accessors for the vmpressure. */
435 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
438 memcg
= root_mem_cgroup
;
439 return &memcg
->vmpressure
;
442 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
444 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
447 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
449 return (memcg
== root_mem_cgroup
);
453 * We restrict the id in the range of [1, 65535], so it can fit into
456 #define MEM_CGROUP_ID_MAX USHRT_MAX
458 static inline unsigned short mem_cgroup_id(struct mem_cgroup
*memcg
)
460 return memcg
->css
.id
;
463 static inline struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
465 struct cgroup_subsys_state
*css
;
467 css
= css_from_id(id
, &memory_cgrp_subsys
);
468 return mem_cgroup_from_css(css
);
471 /* Writing them here to avoid exposing memcg's inner layout */
472 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
474 void sock_update_memcg(struct sock
*sk
)
476 if (mem_cgroup_sockets_enabled
) {
477 struct mem_cgroup
*memcg
;
478 struct cg_proto
*cg_proto
;
480 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
482 /* Socket cloning can throw us here with sk_cgrp already
483 * filled. It won't however, necessarily happen from
484 * process context. So the test for root memcg given
485 * the current task's memcg won't help us in this case.
487 * Respecting the original socket's memcg is a better
488 * decision in this case.
491 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
492 css_get(&sk
->sk_cgrp
->memcg
->css
);
497 memcg
= mem_cgroup_from_task(current
);
498 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
499 if (!mem_cgroup_is_root(memcg
) &&
500 memcg_proto_active(cg_proto
) &&
501 css_tryget_online(&memcg
->css
)) {
502 sk
->sk_cgrp
= cg_proto
;
507 EXPORT_SYMBOL(sock_update_memcg
);
509 void sock_release_memcg(struct sock
*sk
)
511 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
512 struct mem_cgroup
*memcg
;
513 WARN_ON(!sk
->sk_cgrp
->memcg
);
514 memcg
= sk
->sk_cgrp
->memcg
;
515 css_put(&sk
->sk_cgrp
->memcg
->css
);
519 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
521 if (!memcg
|| mem_cgroup_is_root(memcg
))
524 return &memcg
->tcp_mem
;
526 EXPORT_SYMBOL(tcp_proto_cgroup
);
530 #ifdef CONFIG_MEMCG_KMEM
532 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
533 * The main reason for not using cgroup id for this:
534 * this works better in sparse environments, where we have a lot of memcgs,
535 * but only a few kmem-limited. Or also, if we have, for instance, 200
536 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
537 * 200 entry array for that.
539 * The current size of the caches array is stored in memcg_nr_cache_ids. It
540 * will double each time we have to increase it.
542 static DEFINE_IDA(memcg_cache_ida
);
543 int memcg_nr_cache_ids
;
545 /* Protects memcg_nr_cache_ids */
546 static DECLARE_RWSEM(memcg_cache_ids_sem
);
548 void memcg_get_cache_ids(void)
550 down_read(&memcg_cache_ids_sem
);
553 void memcg_put_cache_ids(void)
555 up_read(&memcg_cache_ids_sem
);
559 * MIN_SIZE is different than 1, because we would like to avoid going through
560 * the alloc/free process all the time. In a small machine, 4 kmem-limited
561 * cgroups is a reasonable guess. In the future, it could be a parameter or
562 * tunable, but that is strictly not necessary.
564 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
565 * this constant directly from cgroup, but it is understandable that this is
566 * better kept as an internal representation in cgroup.c. In any case, the
567 * cgrp_id space is not getting any smaller, and we don't have to necessarily
568 * increase ours as well if it increases.
570 #define MEMCG_CACHES_MIN_SIZE 4
571 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
574 * A lot of the calls to the cache allocation functions are expected to be
575 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
576 * conditional to this static branch, we'll have to allow modules that does
577 * kmem_cache_alloc and the such to see this symbol as well
579 struct static_key memcg_kmem_enabled_key
;
580 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
582 #endif /* CONFIG_MEMCG_KMEM */
584 static struct mem_cgroup_per_zone
*
585 mem_cgroup_zone_zoneinfo(struct mem_cgroup
*memcg
, struct zone
*zone
)
587 int nid
= zone_to_nid(zone
);
588 int zid
= zone_idx(zone
);
590 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
593 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
598 static struct mem_cgroup_per_zone
*
599 mem_cgroup_page_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
601 int nid
= page_to_nid(page
);
602 int zid
= page_zonenum(page
);
604 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
607 static struct mem_cgroup_tree_per_zone
*
608 soft_limit_tree_node_zone(int nid
, int zid
)
610 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
613 static struct mem_cgroup_tree_per_zone
*
614 soft_limit_tree_from_page(struct page
*page
)
616 int nid
= page_to_nid(page
);
617 int zid
= page_zonenum(page
);
619 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
622 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone
*mz
,
623 struct mem_cgroup_tree_per_zone
*mctz
,
624 unsigned long new_usage_in_excess
)
626 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
627 struct rb_node
*parent
= NULL
;
628 struct mem_cgroup_per_zone
*mz_node
;
633 mz
->usage_in_excess
= new_usage_in_excess
;
634 if (!mz
->usage_in_excess
)
638 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
640 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
643 * We can't avoid mem cgroups that are over their soft
644 * limit by the same amount
646 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
649 rb_link_node(&mz
->tree_node
, parent
, p
);
650 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
654 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
655 struct mem_cgroup_tree_per_zone
*mctz
)
659 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
663 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
664 struct mem_cgroup_tree_per_zone
*mctz
)
668 spin_lock_irqsave(&mctz
->lock
, flags
);
669 __mem_cgroup_remove_exceeded(mz
, mctz
);
670 spin_unlock_irqrestore(&mctz
->lock
, flags
);
673 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
675 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
676 unsigned long soft_limit
= ACCESS_ONCE(memcg
->soft_limit
);
677 unsigned long excess
= 0;
679 if (nr_pages
> soft_limit
)
680 excess
= nr_pages
- soft_limit
;
685 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
687 unsigned long excess
;
688 struct mem_cgroup_per_zone
*mz
;
689 struct mem_cgroup_tree_per_zone
*mctz
;
691 mctz
= soft_limit_tree_from_page(page
);
693 * Necessary to update all ancestors when hierarchy is used.
694 * because their event counter is not touched.
696 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
697 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
698 excess
= soft_limit_excess(memcg
);
700 * We have to update the tree if mz is on RB-tree or
701 * mem is over its softlimit.
703 if (excess
|| mz
->on_tree
) {
706 spin_lock_irqsave(&mctz
->lock
, flags
);
707 /* if on-tree, remove it */
709 __mem_cgroup_remove_exceeded(mz
, mctz
);
711 * Insert again. mz->usage_in_excess will be updated.
712 * If excess is 0, no tree ops.
714 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
715 spin_unlock_irqrestore(&mctz
->lock
, flags
);
720 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
722 struct mem_cgroup_tree_per_zone
*mctz
;
723 struct mem_cgroup_per_zone
*mz
;
727 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
728 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
729 mctz
= soft_limit_tree_node_zone(nid
, zid
);
730 mem_cgroup_remove_exceeded(mz
, mctz
);
735 static struct mem_cgroup_per_zone
*
736 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
738 struct rb_node
*rightmost
= NULL
;
739 struct mem_cgroup_per_zone
*mz
;
743 rightmost
= rb_last(&mctz
->rb_root
);
745 goto done
; /* Nothing to reclaim from */
747 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
749 * Remove the node now but someone else can add it back,
750 * we will to add it back at the end of reclaim to its correct
751 * position in the tree.
753 __mem_cgroup_remove_exceeded(mz
, mctz
);
754 if (!soft_limit_excess(mz
->memcg
) ||
755 !css_tryget_online(&mz
->memcg
->css
))
761 static struct mem_cgroup_per_zone
*
762 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
764 struct mem_cgroup_per_zone
*mz
;
766 spin_lock_irq(&mctz
->lock
);
767 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
768 spin_unlock_irq(&mctz
->lock
);
773 * Implementation Note: reading percpu statistics for memcg.
775 * Both of vmstat[] and percpu_counter has threshold and do periodic
776 * synchronization to implement "quick" read. There are trade-off between
777 * reading cost and precision of value. Then, we may have a chance to implement
778 * a periodic synchronizion of counter in memcg's counter.
780 * But this _read() function is used for user interface now. The user accounts
781 * memory usage by memory cgroup and he _always_ requires exact value because
782 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
783 * have to visit all online cpus and make sum. So, for now, unnecessary
784 * synchronization is not implemented. (just implemented for cpu hotplug)
786 * If there are kernel internal actions which can make use of some not-exact
787 * value, and reading all cpu value can be performance bottleneck in some
788 * common workload, threashold and synchonization as vmstat[] should be
791 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
792 enum mem_cgroup_stat_index idx
)
798 for_each_online_cpu(cpu
)
799 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
800 #ifdef CONFIG_HOTPLUG_CPU
801 spin_lock(&memcg
->pcp_counter_lock
);
802 val
+= memcg
->nocpu_base
.count
[idx
];
803 spin_unlock(&memcg
->pcp_counter_lock
);
809 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
810 enum mem_cgroup_events_index idx
)
812 unsigned long val
= 0;
816 for_each_online_cpu(cpu
)
817 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
818 #ifdef CONFIG_HOTPLUG_CPU
819 spin_lock(&memcg
->pcp_counter_lock
);
820 val
+= memcg
->nocpu_base
.events
[idx
];
821 spin_unlock(&memcg
->pcp_counter_lock
);
827 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
832 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
833 * counted as CACHE even if it's on ANON LRU.
836 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
839 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
842 if (PageTransHuge(page
))
843 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
846 /* pagein of a big page is an event. So, ignore page size */
848 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
850 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
851 nr_pages
= -nr_pages
; /* for event */
854 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
857 unsigned long mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
859 struct mem_cgroup_per_zone
*mz
;
861 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
862 return mz
->lru_size
[lru
];
865 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
867 unsigned int lru_mask
)
869 unsigned long nr
= 0;
872 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
874 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
875 struct mem_cgroup_per_zone
*mz
;
879 if (!(BIT(lru
) & lru_mask
))
881 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
882 nr
+= mz
->lru_size
[lru
];
888 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
889 unsigned int lru_mask
)
891 unsigned long nr
= 0;
894 for_each_node_state(nid
, N_MEMORY
)
895 nr
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
899 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
900 enum mem_cgroup_events_target target
)
902 unsigned long val
, next
;
904 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
905 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
906 /* from time_after() in jiffies.h */
907 if ((long)next
- (long)val
< 0) {
909 case MEM_CGROUP_TARGET_THRESH
:
910 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
912 case MEM_CGROUP_TARGET_SOFTLIMIT
:
913 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
915 case MEM_CGROUP_TARGET_NUMAINFO
:
916 next
= val
+ NUMAINFO_EVENTS_TARGET
;
921 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
928 * Check events in order.
931 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
933 /* threshold event is triggered in finer grain than soft limit */
934 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
935 MEM_CGROUP_TARGET_THRESH
))) {
937 bool do_numainfo __maybe_unused
;
939 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
940 MEM_CGROUP_TARGET_SOFTLIMIT
);
942 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
943 MEM_CGROUP_TARGET_NUMAINFO
);
945 mem_cgroup_threshold(memcg
);
946 if (unlikely(do_softlimit
))
947 mem_cgroup_update_tree(memcg
, page
);
949 if (unlikely(do_numainfo
))
950 atomic_inc(&memcg
->numainfo_events
);
955 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
958 * mm_update_next_owner() may clear mm->owner to NULL
959 * if it races with swapoff, page migration, etc.
960 * So this can be called with p == NULL.
965 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
968 static struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
970 struct mem_cgroup
*memcg
= NULL
;
975 * Page cache insertions can happen withou an
976 * actual mm context, e.g. during disk probing
977 * on boot, loopback IO, acct() writes etc.
980 memcg
= root_mem_cgroup
;
982 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
983 if (unlikely(!memcg
))
984 memcg
= root_mem_cgroup
;
986 } while (!css_tryget_online(&memcg
->css
));
992 * mem_cgroup_iter - iterate over memory cgroup hierarchy
993 * @root: hierarchy root
994 * @prev: previously returned memcg, NULL on first invocation
995 * @reclaim: cookie for shared reclaim walks, NULL for full walks
997 * Returns references to children of the hierarchy below @root, or
998 * @root itself, or %NULL after a full round-trip.
1000 * Caller must pass the return value in @prev on subsequent
1001 * invocations for reference counting, or use mem_cgroup_iter_break()
1002 * to cancel a hierarchy walk before the round-trip is complete.
1004 * Reclaimers can specify a zone and a priority level in @reclaim to
1005 * divide up the memcgs in the hierarchy among all concurrent
1006 * reclaimers operating on the same zone and priority.
1008 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1009 struct mem_cgroup
*prev
,
1010 struct mem_cgroup_reclaim_cookie
*reclaim
)
1012 struct reclaim_iter
*uninitialized_var(iter
);
1013 struct cgroup_subsys_state
*css
= NULL
;
1014 struct mem_cgroup
*memcg
= NULL
;
1015 struct mem_cgroup
*pos
= NULL
;
1017 if (mem_cgroup_disabled())
1021 root
= root_mem_cgroup
;
1023 if (prev
&& !reclaim
)
1026 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1035 struct mem_cgroup_per_zone
*mz
;
1037 mz
= mem_cgroup_zone_zoneinfo(root
, reclaim
->zone
);
1038 iter
= &mz
->iter
[reclaim
->priority
];
1040 if (prev
&& reclaim
->generation
!= iter
->generation
)
1044 pos
= ACCESS_ONCE(iter
->position
);
1046 * A racing update may change the position and
1047 * put the last reference, hence css_tryget(),
1048 * or retry to see the updated position.
1050 } while (pos
&& !css_tryget(&pos
->css
));
1057 css
= css_next_descendant_pre(css
, &root
->css
);
1060 * Reclaimers share the hierarchy walk, and a
1061 * new one might jump in right at the end of
1062 * the hierarchy - make sure they see at least
1063 * one group and restart from the beginning.
1071 * Verify the css and acquire a reference. The root
1072 * is provided by the caller, so we know it's alive
1073 * and kicking, and don't take an extra reference.
1075 memcg
= mem_cgroup_from_css(css
);
1077 if (css
== &root
->css
)
1080 if (css_tryget(css
)) {
1082 * Make sure the memcg is initialized:
1083 * mem_cgroup_css_online() orders the the
1084 * initialization against setting the flag.
1086 if (smp_load_acquire(&memcg
->initialized
))
1096 if (cmpxchg(&iter
->position
, pos
, memcg
) == pos
) {
1098 css_get(&memcg
->css
);
1104 * pairs with css_tryget when dereferencing iter->position
1113 reclaim
->generation
= iter
->generation
;
1119 if (prev
&& prev
!= root
)
1120 css_put(&prev
->css
);
1126 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1127 * @root: hierarchy root
1128 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1130 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1131 struct mem_cgroup
*prev
)
1134 root
= root_mem_cgroup
;
1135 if (prev
&& prev
!= root
)
1136 css_put(&prev
->css
);
1140 * Iteration constructs for visiting all cgroups (under a tree). If
1141 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1142 * be used for reference counting.
1144 #define for_each_mem_cgroup_tree(iter, root) \
1145 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1147 iter = mem_cgroup_iter(root, iter, NULL))
1149 #define for_each_mem_cgroup(iter) \
1150 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1152 iter = mem_cgroup_iter(NULL, iter, NULL))
1154 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1156 struct mem_cgroup
*memcg
;
1159 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1160 if (unlikely(!memcg
))
1165 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1168 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1176 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1179 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1180 * @zone: zone of the wanted lruvec
1181 * @memcg: memcg of the wanted lruvec
1183 * Returns the lru list vector holding pages for the given @zone and
1184 * @mem. This can be the global zone lruvec, if the memory controller
1187 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1188 struct mem_cgroup
*memcg
)
1190 struct mem_cgroup_per_zone
*mz
;
1191 struct lruvec
*lruvec
;
1193 if (mem_cgroup_disabled()) {
1194 lruvec
= &zone
->lruvec
;
1198 mz
= mem_cgroup_zone_zoneinfo(memcg
, zone
);
1199 lruvec
= &mz
->lruvec
;
1202 * Since a node can be onlined after the mem_cgroup was created,
1203 * we have to be prepared to initialize lruvec->zone here;
1204 * and if offlined then reonlined, we need to reinitialize it.
1206 if (unlikely(lruvec
->zone
!= zone
))
1207 lruvec
->zone
= zone
;
1212 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1214 * @zone: zone of the page
1216 * This function is only safe when following the LRU page isolation
1217 * and putback protocol: the LRU lock must be held, and the page must
1218 * either be PageLRU() or the caller must have isolated/allocated it.
1220 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1222 struct mem_cgroup_per_zone
*mz
;
1223 struct mem_cgroup
*memcg
;
1224 struct lruvec
*lruvec
;
1226 if (mem_cgroup_disabled()) {
1227 lruvec
= &zone
->lruvec
;
1231 memcg
= page
->mem_cgroup
;
1233 * Swapcache readahead pages are added to the LRU - and
1234 * possibly migrated - before they are charged.
1237 memcg
= root_mem_cgroup
;
1239 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
1240 lruvec
= &mz
->lruvec
;
1243 * Since a node can be onlined after the mem_cgroup was created,
1244 * we have to be prepared to initialize lruvec->zone here;
1245 * and if offlined then reonlined, we need to reinitialize it.
1247 if (unlikely(lruvec
->zone
!= zone
))
1248 lruvec
->zone
= zone
;
1253 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1254 * @lruvec: mem_cgroup per zone lru vector
1255 * @lru: index of lru list the page is sitting on
1256 * @nr_pages: positive when adding or negative when removing
1258 * This function must be called when a page is added to or removed from an
1261 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1264 struct mem_cgroup_per_zone
*mz
;
1265 unsigned long *lru_size
;
1267 if (mem_cgroup_disabled())
1270 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1271 lru_size
= mz
->lru_size
+ lru
;
1272 *lru_size
+= nr_pages
;
1273 VM_BUG_ON((long)(*lru_size
) < 0);
1276 bool mem_cgroup_is_descendant(struct mem_cgroup
*memcg
, struct mem_cgroup
*root
)
1280 if (!root
->use_hierarchy
)
1282 return cgroup_is_descendant(memcg
->css
.cgroup
, root
->css
.cgroup
);
1285 bool task_in_mem_cgroup(struct task_struct
*task
, struct mem_cgroup
*memcg
)
1287 struct mem_cgroup
*task_memcg
;
1288 struct task_struct
*p
;
1291 p
= find_lock_task_mm(task
);
1293 task_memcg
= get_mem_cgroup_from_mm(p
->mm
);
1297 * All threads may have already detached their mm's, but the oom
1298 * killer still needs to detect if they have already been oom
1299 * killed to prevent needlessly killing additional tasks.
1302 task_memcg
= mem_cgroup_from_task(task
);
1303 css_get(&task_memcg
->css
);
1306 ret
= mem_cgroup_is_descendant(task_memcg
, memcg
);
1307 css_put(&task_memcg
->css
);
1311 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1313 unsigned long inactive_ratio
;
1314 unsigned long inactive
;
1315 unsigned long active
;
1318 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1319 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1321 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1323 inactive_ratio
= int_sqrt(10 * gb
);
1327 return inactive
* inactive_ratio
< active
;
1330 bool mem_cgroup_lruvec_online(struct lruvec
*lruvec
)
1332 struct mem_cgroup_per_zone
*mz
;
1333 struct mem_cgroup
*memcg
;
1335 if (mem_cgroup_disabled())
1338 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1341 return !!(memcg
->css
.flags
& CSS_ONLINE
);
1344 #define mem_cgroup_from_counter(counter, member) \
1345 container_of(counter, struct mem_cgroup, member)
1348 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1349 * @memcg: the memory cgroup
1351 * Returns the maximum amount of memory @mem can be charged with, in
1354 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1356 unsigned long margin
= 0;
1357 unsigned long count
;
1358 unsigned long limit
;
1360 count
= page_counter_read(&memcg
->memory
);
1361 limit
= ACCESS_ONCE(memcg
->memory
.limit
);
1363 margin
= limit
- count
;
1365 if (do_swap_account
) {
1366 count
= page_counter_read(&memcg
->memsw
);
1367 limit
= ACCESS_ONCE(memcg
->memsw
.limit
);
1369 margin
= min(margin
, limit
- count
);
1375 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1378 if (mem_cgroup_disabled() || !memcg
->css
.parent
)
1379 return vm_swappiness
;
1381 return memcg
->swappiness
;
1385 * A routine for checking "mem" is under move_account() or not.
1387 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1388 * moving cgroups. This is for waiting at high-memory pressure
1391 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1393 struct mem_cgroup
*from
;
1394 struct mem_cgroup
*to
;
1397 * Unlike task_move routines, we access mc.to, mc.from not under
1398 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1400 spin_lock(&mc
.lock
);
1406 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1407 mem_cgroup_is_descendant(to
, memcg
);
1409 spin_unlock(&mc
.lock
);
1413 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1415 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1416 if (mem_cgroup_under_move(memcg
)) {
1418 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1419 /* moving charge context might have finished. */
1422 finish_wait(&mc
.waitq
, &wait
);
1429 #define K(x) ((x) << (PAGE_SHIFT-10))
1431 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1432 * @memcg: The memory cgroup that went over limit
1433 * @p: Task that is going to be killed
1435 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1438 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1440 /* oom_info_lock ensures that parallel ooms do not interleave */
1441 static DEFINE_MUTEX(oom_info_lock
);
1442 struct mem_cgroup
*iter
;
1445 mutex_lock(&oom_info_lock
);
1449 pr_info("Task in ");
1450 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1451 pr_cont(" killed as a result of limit of ");
1453 pr_info("Memory limit reached of cgroup ");
1456 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1461 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1462 K((u64
)page_counter_read(&memcg
->memory
)),
1463 K((u64
)memcg
->memory
.limit
), memcg
->memory
.failcnt
);
1464 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1465 K((u64
)page_counter_read(&memcg
->memsw
)),
1466 K((u64
)memcg
->memsw
.limit
), memcg
->memsw
.failcnt
);
1467 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1468 K((u64
)page_counter_read(&memcg
->kmem
)),
1469 K((u64
)memcg
->kmem
.limit
), memcg
->kmem
.failcnt
);
1471 for_each_mem_cgroup_tree(iter
, memcg
) {
1472 pr_info("Memory cgroup stats for ");
1473 pr_cont_cgroup_path(iter
->css
.cgroup
);
1476 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1477 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1479 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1480 K(mem_cgroup_read_stat(iter
, i
)));
1483 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1484 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1485 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1489 mutex_unlock(&oom_info_lock
);
1493 * This function returns the number of memcg under hierarchy tree. Returns
1494 * 1(self count) if no children.
1496 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1499 struct mem_cgroup
*iter
;
1501 for_each_mem_cgroup_tree(iter
, memcg
)
1507 * Return the memory (and swap, if configured) limit for a memcg.
1509 static unsigned long mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1511 unsigned long limit
;
1513 limit
= memcg
->memory
.limit
;
1514 if (mem_cgroup_swappiness(memcg
)) {
1515 unsigned long memsw_limit
;
1517 memsw_limit
= memcg
->memsw
.limit
;
1518 limit
= min(limit
+ total_swap_pages
, memsw_limit
);
1523 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1526 struct mem_cgroup
*iter
;
1527 unsigned long chosen_points
= 0;
1528 unsigned long totalpages
;
1529 unsigned int points
= 0;
1530 struct task_struct
*chosen
= NULL
;
1533 * If current has a pending SIGKILL or is exiting, then automatically
1534 * select it. The goal is to allow it to allocate so that it may
1535 * quickly exit and free its memory.
1537 if (fatal_signal_pending(current
) || task_will_free_mem(current
)) {
1538 mark_tsk_oom_victim(current
);
1542 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
, memcg
);
1543 totalpages
= mem_cgroup_get_limit(memcg
) ? : 1;
1544 for_each_mem_cgroup_tree(iter
, memcg
) {
1545 struct css_task_iter it
;
1546 struct task_struct
*task
;
1548 css_task_iter_start(&iter
->css
, &it
);
1549 while ((task
= css_task_iter_next(&it
))) {
1550 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1552 case OOM_SCAN_SELECT
:
1554 put_task_struct(chosen
);
1556 chosen_points
= ULONG_MAX
;
1557 get_task_struct(chosen
);
1559 case OOM_SCAN_CONTINUE
:
1561 case OOM_SCAN_ABORT
:
1562 css_task_iter_end(&it
);
1563 mem_cgroup_iter_break(memcg
, iter
);
1565 put_task_struct(chosen
);
1570 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1571 if (!points
|| points
< chosen_points
)
1573 /* Prefer thread group leaders for display purposes */
1574 if (points
== chosen_points
&&
1575 thread_group_leader(chosen
))
1579 put_task_struct(chosen
);
1581 chosen_points
= points
;
1582 get_task_struct(chosen
);
1584 css_task_iter_end(&it
);
1589 points
= chosen_points
* 1000 / totalpages
;
1590 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1591 NULL
, "Memory cgroup out of memory");
1594 #if MAX_NUMNODES > 1
1597 * test_mem_cgroup_node_reclaimable
1598 * @memcg: the target memcg
1599 * @nid: the node ID to be checked.
1600 * @noswap : specify true here if the user wants flle only information.
1602 * This function returns whether the specified memcg contains any
1603 * reclaimable pages on a node. Returns true if there are any reclaimable
1604 * pages in the node.
1606 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1607 int nid
, bool noswap
)
1609 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1611 if (noswap
|| !total_swap_pages
)
1613 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1620 * Always updating the nodemask is not very good - even if we have an empty
1621 * list or the wrong list here, we can start from some node and traverse all
1622 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1625 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1629 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1630 * pagein/pageout changes since the last update.
1632 if (!atomic_read(&memcg
->numainfo_events
))
1634 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1637 /* make a nodemask where this memcg uses memory from */
1638 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1640 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1642 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1643 node_clear(nid
, memcg
->scan_nodes
);
1646 atomic_set(&memcg
->numainfo_events
, 0);
1647 atomic_set(&memcg
->numainfo_updating
, 0);
1651 * Selecting a node where we start reclaim from. Because what we need is just
1652 * reducing usage counter, start from anywhere is O,K. Considering
1653 * memory reclaim from current node, there are pros. and cons.
1655 * Freeing memory from current node means freeing memory from a node which
1656 * we'll use or we've used. So, it may make LRU bad. And if several threads
1657 * hit limits, it will see a contention on a node. But freeing from remote
1658 * node means more costs for memory reclaim because of memory latency.
1660 * Now, we use round-robin. Better algorithm is welcomed.
1662 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1666 mem_cgroup_may_update_nodemask(memcg
);
1667 node
= memcg
->last_scanned_node
;
1669 node
= next_node(node
, memcg
->scan_nodes
);
1670 if (node
== MAX_NUMNODES
)
1671 node
= first_node(memcg
->scan_nodes
);
1673 * We call this when we hit limit, not when pages are added to LRU.
1674 * No LRU may hold pages because all pages are UNEVICTABLE or
1675 * memcg is too small and all pages are not on LRU. In that case,
1676 * we use curret node.
1678 if (unlikely(node
== MAX_NUMNODES
))
1679 node
= numa_node_id();
1681 memcg
->last_scanned_node
= node
;
1685 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1691 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1694 unsigned long *total_scanned
)
1696 struct mem_cgroup
*victim
= NULL
;
1699 unsigned long excess
;
1700 unsigned long nr_scanned
;
1701 struct mem_cgroup_reclaim_cookie reclaim
= {
1706 excess
= soft_limit_excess(root_memcg
);
1709 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1714 * If we have not been able to reclaim
1715 * anything, it might because there are
1716 * no reclaimable pages under this hierarchy
1721 * We want to do more targeted reclaim.
1722 * excess >> 2 is not to excessive so as to
1723 * reclaim too much, nor too less that we keep
1724 * coming back to reclaim from this cgroup
1726 if (total
>= (excess
>> 2) ||
1727 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1732 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
1734 *total_scanned
+= nr_scanned
;
1735 if (!soft_limit_excess(root_memcg
))
1738 mem_cgroup_iter_break(root_memcg
, victim
);
1742 #ifdef CONFIG_LOCKDEP
1743 static struct lockdep_map memcg_oom_lock_dep_map
= {
1744 .name
= "memcg_oom_lock",
1748 static DEFINE_SPINLOCK(memcg_oom_lock
);
1751 * Check OOM-Killer is already running under our hierarchy.
1752 * If someone is running, return false.
1754 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1756 struct mem_cgroup
*iter
, *failed
= NULL
;
1758 spin_lock(&memcg_oom_lock
);
1760 for_each_mem_cgroup_tree(iter
, memcg
) {
1761 if (iter
->oom_lock
) {
1763 * this subtree of our hierarchy is already locked
1764 * so we cannot give a lock.
1767 mem_cgroup_iter_break(memcg
, iter
);
1770 iter
->oom_lock
= true;
1775 * OK, we failed to lock the whole subtree so we have
1776 * to clean up what we set up to the failing subtree
1778 for_each_mem_cgroup_tree(iter
, memcg
) {
1779 if (iter
== failed
) {
1780 mem_cgroup_iter_break(memcg
, iter
);
1783 iter
->oom_lock
= false;
1786 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1788 spin_unlock(&memcg_oom_lock
);
1793 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1795 struct mem_cgroup
*iter
;
1797 spin_lock(&memcg_oom_lock
);
1798 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1799 for_each_mem_cgroup_tree(iter
, memcg
)
1800 iter
->oom_lock
= false;
1801 spin_unlock(&memcg_oom_lock
);
1804 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1806 struct mem_cgroup
*iter
;
1808 for_each_mem_cgroup_tree(iter
, memcg
)
1809 atomic_inc(&iter
->under_oom
);
1812 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1814 struct mem_cgroup
*iter
;
1817 * When a new child is created while the hierarchy is under oom,
1818 * mem_cgroup_oom_lock() may not be called. We have to use
1819 * atomic_add_unless() here.
1821 for_each_mem_cgroup_tree(iter
, memcg
)
1822 atomic_add_unless(&iter
->under_oom
, -1, 0);
1825 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1827 struct oom_wait_info
{
1828 struct mem_cgroup
*memcg
;
1832 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1833 unsigned mode
, int sync
, void *arg
)
1835 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1836 struct mem_cgroup
*oom_wait_memcg
;
1837 struct oom_wait_info
*oom_wait_info
;
1839 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1840 oom_wait_memcg
= oom_wait_info
->memcg
;
1842 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1843 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1845 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1848 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
1850 atomic_inc(&memcg
->oom_wakeups
);
1851 /* for filtering, pass "memcg" as argument. */
1852 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1855 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1857 if (memcg
&& atomic_read(&memcg
->under_oom
))
1858 memcg_wakeup_oom(memcg
);
1861 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1863 if (!current
->memcg_oom
.may_oom
)
1866 * We are in the middle of the charge context here, so we
1867 * don't want to block when potentially sitting on a callstack
1868 * that holds all kinds of filesystem and mm locks.
1870 * Also, the caller may handle a failed allocation gracefully
1871 * (like optional page cache readahead) and so an OOM killer
1872 * invocation might not even be necessary.
1874 * That's why we don't do anything here except remember the
1875 * OOM context and then deal with it at the end of the page
1876 * fault when the stack is unwound, the locks are released,
1877 * and when we know whether the fault was overall successful.
1879 css_get(&memcg
->css
);
1880 current
->memcg_oom
.memcg
= memcg
;
1881 current
->memcg_oom
.gfp_mask
= mask
;
1882 current
->memcg_oom
.order
= order
;
1886 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1887 * @handle: actually kill/wait or just clean up the OOM state
1889 * This has to be called at the end of a page fault if the memcg OOM
1890 * handler was enabled.
1892 * Memcg supports userspace OOM handling where failed allocations must
1893 * sleep on a waitqueue until the userspace task resolves the
1894 * situation. Sleeping directly in the charge context with all kinds
1895 * of locks held is not a good idea, instead we remember an OOM state
1896 * in the task and mem_cgroup_oom_synchronize() has to be called at
1897 * the end of the page fault to complete the OOM handling.
1899 * Returns %true if an ongoing memcg OOM situation was detected and
1900 * completed, %false otherwise.
1902 bool mem_cgroup_oom_synchronize(bool handle
)
1904 struct mem_cgroup
*memcg
= current
->memcg_oom
.memcg
;
1905 struct oom_wait_info owait
;
1908 /* OOM is global, do not handle */
1912 if (!handle
|| oom_killer_disabled
)
1915 owait
.memcg
= memcg
;
1916 owait
.wait
.flags
= 0;
1917 owait
.wait
.func
= memcg_oom_wake_function
;
1918 owait
.wait
.private = current
;
1919 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1921 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1922 mem_cgroup_mark_under_oom(memcg
);
1924 locked
= mem_cgroup_oom_trylock(memcg
);
1927 mem_cgroup_oom_notify(memcg
);
1929 if (locked
&& !memcg
->oom_kill_disable
) {
1930 mem_cgroup_unmark_under_oom(memcg
);
1931 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1932 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom
.gfp_mask
,
1933 current
->memcg_oom
.order
);
1936 mem_cgroup_unmark_under_oom(memcg
);
1937 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1941 mem_cgroup_oom_unlock(memcg
);
1943 * There is no guarantee that an OOM-lock contender
1944 * sees the wakeups triggered by the OOM kill
1945 * uncharges. Wake any sleepers explicitely.
1947 memcg_oom_recover(memcg
);
1950 current
->memcg_oom
.memcg
= NULL
;
1951 css_put(&memcg
->css
);
1956 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1957 * @page: page that is going to change accounted state
1959 * This function must mark the beginning of an accounted page state
1960 * change to prevent double accounting when the page is concurrently
1961 * being moved to another memcg:
1963 * memcg = mem_cgroup_begin_page_stat(page);
1964 * if (TestClearPageState(page))
1965 * mem_cgroup_update_page_stat(memcg, state, -1);
1966 * mem_cgroup_end_page_stat(memcg);
1968 struct mem_cgroup
*mem_cgroup_begin_page_stat(struct page
*page
)
1970 struct mem_cgroup
*memcg
;
1971 unsigned long flags
;
1974 * The RCU lock is held throughout the transaction. The fast
1975 * path can get away without acquiring the memcg->move_lock
1976 * because page moving starts with an RCU grace period.
1978 * The RCU lock also protects the memcg from being freed when
1979 * the page state that is going to change is the only thing
1980 * preventing the page from being uncharged.
1981 * E.g. end-writeback clearing PageWriteback(), which allows
1982 * migration to go ahead and uncharge the page before the
1983 * account transaction might be complete.
1987 if (mem_cgroup_disabled())
1990 memcg
= page
->mem_cgroup
;
1991 if (unlikely(!memcg
))
1994 if (atomic_read(&memcg
->moving_account
) <= 0)
1997 spin_lock_irqsave(&memcg
->move_lock
, flags
);
1998 if (memcg
!= page
->mem_cgroup
) {
1999 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2004 * When charge migration first begins, we can have locked and
2005 * unlocked page stat updates happening concurrently. Track
2006 * the task who has the lock for mem_cgroup_end_page_stat().
2008 memcg
->move_lock_task
= current
;
2009 memcg
->move_lock_flags
= flags
;
2015 * mem_cgroup_end_page_stat - finish a page state statistics transaction
2016 * @memcg: the memcg that was accounted against
2018 void mem_cgroup_end_page_stat(struct mem_cgroup
*memcg
)
2020 if (memcg
&& memcg
->move_lock_task
== current
) {
2021 unsigned long flags
= memcg
->move_lock_flags
;
2023 memcg
->move_lock_task
= NULL
;
2024 memcg
->move_lock_flags
= 0;
2026 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2033 * mem_cgroup_update_page_stat - update page state statistics
2034 * @memcg: memcg to account against
2035 * @idx: page state item to account
2036 * @val: number of pages (positive or negative)
2038 * See mem_cgroup_begin_page_stat() for locking requirements.
2040 void mem_cgroup_update_page_stat(struct mem_cgroup
*memcg
,
2041 enum mem_cgroup_stat_index idx
, int val
)
2043 VM_BUG_ON(!rcu_read_lock_held());
2046 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2050 * size of first charge trial. "32" comes from vmscan.c's magic value.
2051 * TODO: maybe necessary to use big numbers in big irons.
2053 #define CHARGE_BATCH 32U
2054 struct memcg_stock_pcp
{
2055 struct mem_cgroup
*cached
; /* this never be root cgroup */
2056 unsigned int nr_pages
;
2057 struct work_struct work
;
2058 unsigned long flags
;
2059 #define FLUSHING_CACHED_CHARGE 0
2061 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2062 static DEFINE_MUTEX(percpu_charge_mutex
);
2065 * consume_stock: Try to consume stocked charge on this cpu.
2066 * @memcg: memcg to consume from.
2067 * @nr_pages: how many pages to charge.
2069 * The charges will only happen if @memcg matches the current cpu's memcg
2070 * stock, and at least @nr_pages are available in that stock. Failure to
2071 * service an allocation will refill the stock.
2073 * returns true if successful, false otherwise.
2075 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2077 struct memcg_stock_pcp
*stock
;
2080 if (nr_pages
> CHARGE_BATCH
)
2083 stock
= &get_cpu_var(memcg_stock
);
2084 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
2085 stock
->nr_pages
-= nr_pages
;
2088 put_cpu_var(memcg_stock
);
2093 * Returns stocks cached in percpu and reset cached information.
2095 static void drain_stock(struct memcg_stock_pcp
*stock
)
2097 struct mem_cgroup
*old
= stock
->cached
;
2099 if (stock
->nr_pages
) {
2100 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
2101 if (do_swap_account
)
2102 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
2103 css_put_many(&old
->css
, stock
->nr_pages
);
2104 stock
->nr_pages
= 0;
2106 stock
->cached
= NULL
;
2110 * This must be called under preempt disabled or must be called by
2111 * a thread which is pinned to local cpu.
2113 static void drain_local_stock(struct work_struct
*dummy
)
2115 struct memcg_stock_pcp
*stock
= this_cpu_ptr(&memcg_stock
);
2117 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2121 * Cache charges(val) to local per_cpu area.
2122 * This will be consumed by consume_stock() function, later.
2124 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2126 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2128 if (stock
->cached
!= memcg
) { /* reset if necessary */
2130 stock
->cached
= memcg
;
2132 stock
->nr_pages
+= nr_pages
;
2133 put_cpu_var(memcg_stock
);
2137 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2138 * of the hierarchy under it.
2140 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
2144 /* If someone's already draining, avoid adding running more workers. */
2145 if (!mutex_trylock(&percpu_charge_mutex
))
2147 /* Notify other cpus that system-wide "drain" is running */
2150 for_each_online_cpu(cpu
) {
2151 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2152 struct mem_cgroup
*memcg
;
2154 memcg
= stock
->cached
;
2155 if (!memcg
|| !stock
->nr_pages
)
2157 if (!mem_cgroup_is_descendant(memcg
, root_memcg
))
2159 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2161 drain_local_stock(&stock
->work
);
2163 schedule_work_on(cpu
, &stock
->work
);
2168 mutex_unlock(&percpu_charge_mutex
);
2172 * This function drains percpu counter value from DEAD cpu and
2173 * move it to local cpu. Note that this function can be preempted.
2175 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2179 spin_lock(&memcg
->pcp_counter_lock
);
2180 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2181 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2183 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2184 memcg
->nocpu_base
.count
[i
] += x
;
2186 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2187 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2189 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2190 memcg
->nocpu_base
.events
[i
] += x
;
2192 spin_unlock(&memcg
->pcp_counter_lock
);
2195 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2196 unsigned long action
,
2199 int cpu
= (unsigned long)hcpu
;
2200 struct memcg_stock_pcp
*stock
;
2201 struct mem_cgroup
*iter
;
2203 if (action
== CPU_ONLINE
)
2206 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2209 for_each_mem_cgroup(iter
)
2210 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2212 stock
= &per_cpu(memcg_stock
, cpu
);
2217 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2218 unsigned int nr_pages
)
2220 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2221 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2222 struct mem_cgroup
*mem_over_limit
;
2223 struct page_counter
*counter
;
2224 unsigned long nr_reclaimed
;
2225 bool may_swap
= true;
2226 bool drained
= false;
2229 if (mem_cgroup_is_root(memcg
))
2232 if (consume_stock(memcg
, nr_pages
))
2235 if (!do_swap_account
||
2236 !page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
2237 if (!page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
2239 if (do_swap_account
)
2240 page_counter_uncharge(&memcg
->memsw
, batch
);
2241 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
2243 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
2247 if (batch
> nr_pages
) {
2253 * Unlike in global OOM situations, memcg is not in a physical
2254 * memory shortage. Allow dying and OOM-killed tasks to
2255 * bypass the last charges so that they can exit quickly and
2256 * free their memory.
2258 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
2259 fatal_signal_pending(current
) ||
2260 current
->flags
& PF_EXITING
))
2263 if (unlikely(task_in_memcg_oom(current
)))
2266 if (!(gfp_mask
& __GFP_WAIT
))
2269 mem_cgroup_events(mem_over_limit
, MEMCG_MAX
, 1);
2271 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
2272 gfp_mask
, may_swap
);
2274 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2278 drain_all_stock(mem_over_limit
);
2283 if (gfp_mask
& __GFP_NORETRY
)
2286 * Even though the limit is exceeded at this point, reclaim
2287 * may have been able to free some pages. Retry the charge
2288 * before killing the task.
2290 * Only for regular pages, though: huge pages are rather
2291 * unlikely to succeed so close to the limit, and we fall back
2292 * to regular pages anyway in case of failure.
2294 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2297 * At task move, charge accounts can be doubly counted. So, it's
2298 * better to wait until the end of task_move if something is going on.
2300 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2306 if (gfp_mask
& __GFP_NOFAIL
)
2309 if (fatal_signal_pending(current
))
2312 mem_cgroup_events(mem_over_limit
, MEMCG_OOM
, 1);
2314 mem_cgroup_oom(mem_over_limit
, gfp_mask
, get_order(nr_pages
));
2316 if (!(gfp_mask
& __GFP_NOFAIL
))
2322 css_get_many(&memcg
->css
, batch
);
2323 if (batch
> nr_pages
)
2324 refill_stock(memcg
, batch
- nr_pages
);
2326 * If the hierarchy is above the normal consumption range,
2327 * make the charging task trim their excess contribution.
2330 if (page_counter_read(&memcg
->memory
) <= memcg
->high
)
2332 mem_cgroup_events(memcg
, MEMCG_HIGH
, 1);
2333 try_to_free_mem_cgroup_pages(memcg
, nr_pages
, gfp_mask
, true);
2334 } while ((memcg
= parent_mem_cgroup(memcg
)));
2339 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2341 if (mem_cgroup_is_root(memcg
))
2344 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2345 if (do_swap_account
)
2346 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2348 css_put_many(&memcg
->css
, nr_pages
);
2352 * A helper function to get mem_cgroup from ID. must be called under
2353 * rcu_read_lock(). The caller is responsible for calling
2354 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2355 * refcnt from swap can be called against removed memcg.)
2357 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2359 /* ID 0 is unused ID */
2362 return mem_cgroup_from_id(id
);
2366 * try_get_mem_cgroup_from_page - look up page's memcg association
2369 * Look up, get a css reference, and return the memcg that owns @page.
2371 * The page must be locked to prevent racing with swap-in and page
2372 * cache charges. If coming from an unlocked page table, the caller
2373 * must ensure the page is on the LRU or this can race with charging.
2375 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2377 struct mem_cgroup
*memcg
;
2381 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2383 memcg
= page
->mem_cgroup
;
2385 if (!css_tryget_online(&memcg
->css
))
2387 } else if (PageSwapCache(page
)) {
2388 ent
.val
= page_private(page
);
2389 id
= lookup_swap_cgroup_id(ent
);
2391 memcg
= mem_cgroup_lookup(id
);
2392 if (memcg
&& !css_tryget_online(&memcg
->css
))
2399 static void lock_page_lru(struct page
*page
, int *isolated
)
2401 struct zone
*zone
= page_zone(page
);
2403 spin_lock_irq(&zone
->lru_lock
);
2404 if (PageLRU(page
)) {
2405 struct lruvec
*lruvec
;
2407 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2409 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2415 static void unlock_page_lru(struct page
*page
, int isolated
)
2417 struct zone
*zone
= page_zone(page
);
2420 struct lruvec
*lruvec
;
2422 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2423 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2425 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2427 spin_unlock_irq(&zone
->lru_lock
);
2430 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2435 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2438 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2439 * may already be on some other mem_cgroup's LRU. Take care of it.
2442 lock_page_lru(page
, &isolated
);
2445 * Nobody should be changing or seriously looking at
2446 * page->mem_cgroup at this point:
2448 * - the page is uncharged
2450 * - the page is off-LRU
2452 * - an anonymous fault has exclusive page access, except for
2453 * a locked page table
2455 * - a page cache insertion, a swapin fault, or a migration
2456 * have the page locked
2458 page
->mem_cgroup
= memcg
;
2461 unlock_page_lru(page
, isolated
);
2464 #ifdef CONFIG_MEMCG_KMEM
2465 int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
,
2466 unsigned long nr_pages
)
2468 struct page_counter
*counter
;
2471 ret
= page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
);
2475 ret
= try_charge(memcg
, gfp
, nr_pages
);
2476 if (ret
== -EINTR
) {
2478 * try_charge() chose to bypass to root due to OOM kill or
2479 * fatal signal. Since our only options are to either fail
2480 * the allocation or charge it to this cgroup, do it as a
2481 * temporary condition. But we can't fail. From a kmem/slab
2482 * perspective, the cache has already been selected, by
2483 * mem_cgroup_kmem_get_cache(), so it is too late to change
2486 * This condition will only trigger if the task entered
2487 * memcg_charge_kmem in a sane state, but was OOM-killed
2488 * during try_charge() above. Tasks that were already dying
2489 * when the allocation triggers should have been already
2490 * directed to the root cgroup in memcontrol.h
2492 page_counter_charge(&memcg
->memory
, nr_pages
);
2493 if (do_swap_account
)
2494 page_counter_charge(&memcg
->memsw
, nr_pages
);
2495 css_get_many(&memcg
->css
, nr_pages
);
2498 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2503 void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, unsigned long nr_pages
)
2505 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2506 if (do_swap_account
)
2507 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2509 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2511 css_put_many(&memcg
->css
, nr_pages
);
2515 * helper for acessing a memcg's index. It will be used as an index in the
2516 * child cache array in kmem_cache, and also to derive its name. This function
2517 * will return -1 when this is not a kmem-limited memcg.
2519 int memcg_cache_id(struct mem_cgroup
*memcg
)
2521 return memcg
? memcg
->kmemcg_id
: -1;
2524 static int memcg_alloc_cache_id(void)
2529 id
= ida_simple_get(&memcg_cache_ida
,
2530 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2534 if (id
< memcg_nr_cache_ids
)
2538 * There's no space for the new id in memcg_caches arrays,
2539 * so we have to grow them.
2541 down_write(&memcg_cache_ids_sem
);
2543 size
= 2 * (id
+ 1);
2544 if (size
< MEMCG_CACHES_MIN_SIZE
)
2545 size
= MEMCG_CACHES_MIN_SIZE
;
2546 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2547 size
= MEMCG_CACHES_MAX_SIZE
;
2549 err
= memcg_update_all_caches(size
);
2551 err
= memcg_update_all_list_lrus(size
);
2553 memcg_nr_cache_ids
= size
;
2555 up_write(&memcg_cache_ids_sem
);
2558 ida_simple_remove(&memcg_cache_ida
, id
);
2564 static void memcg_free_cache_id(int id
)
2566 ida_simple_remove(&memcg_cache_ida
, id
);
2569 struct memcg_kmem_cache_create_work
{
2570 struct mem_cgroup
*memcg
;
2571 struct kmem_cache
*cachep
;
2572 struct work_struct work
;
2575 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2577 struct memcg_kmem_cache_create_work
*cw
=
2578 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2579 struct mem_cgroup
*memcg
= cw
->memcg
;
2580 struct kmem_cache
*cachep
= cw
->cachep
;
2582 memcg_create_kmem_cache(memcg
, cachep
);
2584 css_put(&memcg
->css
);
2589 * Enqueue the creation of a per-memcg kmem_cache.
2591 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2592 struct kmem_cache
*cachep
)
2594 struct memcg_kmem_cache_create_work
*cw
;
2596 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
2600 css_get(&memcg
->css
);
2603 cw
->cachep
= cachep
;
2604 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2606 schedule_work(&cw
->work
);
2609 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2610 struct kmem_cache
*cachep
)
2613 * We need to stop accounting when we kmalloc, because if the
2614 * corresponding kmalloc cache is not yet created, the first allocation
2615 * in __memcg_schedule_kmem_cache_create will recurse.
2617 * However, it is better to enclose the whole function. Depending on
2618 * the debugging options enabled, INIT_WORK(), for instance, can
2619 * trigger an allocation. This too, will make us recurse. Because at
2620 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2621 * the safest choice is to do it like this, wrapping the whole function.
2623 current
->memcg_kmem_skip_account
= 1;
2624 __memcg_schedule_kmem_cache_create(memcg
, cachep
);
2625 current
->memcg_kmem_skip_account
= 0;
2629 * Return the kmem_cache we're supposed to use for a slab allocation.
2630 * We try to use the current memcg's version of the cache.
2632 * If the cache does not exist yet, if we are the first user of it,
2633 * we either create it immediately, if possible, or create it asynchronously
2635 * In the latter case, we will let the current allocation go through with
2636 * the original cache.
2638 * Can't be called in interrupt context or from kernel threads.
2639 * This function needs to be called with rcu_read_lock() held.
2641 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
)
2643 struct mem_cgroup
*memcg
;
2644 struct kmem_cache
*memcg_cachep
;
2647 VM_BUG_ON(!is_root_cache(cachep
));
2649 if (current
->memcg_kmem_skip_account
)
2652 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2653 kmemcg_id
= ACCESS_ONCE(memcg
->kmemcg_id
);
2657 memcg_cachep
= cache_from_memcg_idx(cachep
, kmemcg_id
);
2658 if (likely(memcg_cachep
))
2659 return memcg_cachep
;
2662 * If we are in a safe context (can wait, and not in interrupt
2663 * context), we could be be predictable and return right away.
2664 * This would guarantee that the allocation being performed
2665 * already belongs in the new cache.
2667 * However, there are some clashes that can arrive from locking.
2668 * For instance, because we acquire the slab_mutex while doing
2669 * memcg_create_kmem_cache, this means no further allocation
2670 * could happen with the slab_mutex held. So it's better to
2673 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2675 css_put(&memcg
->css
);
2679 void __memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2681 if (!is_root_cache(cachep
))
2682 css_put(&cachep
->memcg_params
.memcg
->css
);
2686 * We need to verify if the allocation against current->mm->owner's memcg is
2687 * possible for the given order. But the page is not allocated yet, so we'll
2688 * need a further commit step to do the final arrangements.
2690 * It is possible for the task to switch cgroups in this mean time, so at
2691 * commit time, we can't rely on task conversion any longer. We'll then use
2692 * the handle argument to return to the caller which cgroup we should commit
2693 * against. We could also return the memcg directly and avoid the pointer
2694 * passing, but a boolean return value gives better semantics considering
2695 * the compiled-out case as well.
2697 * Returning true means the allocation is possible.
2700 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
2702 struct mem_cgroup
*memcg
;
2707 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2709 if (!memcg_kmem_is_active(memcg
)) {
2710 css_put(&memcg
->css
);
2714 ret
= memcg_charge_kmem(memcg
, gfp
, 1 << order
);
2718 css_put(&memcg
->css
);
2722 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2725 VM_BUG_ON(mem_cgroup_is_root(memcg
));
2727 /* The page allocation failed. Revert */
2729 memcg_uncharge_kmem(memcg
, 1 << order
);
2732 page
->mem_cgroup
= memcg
;
2735 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
2737 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2742 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2744 memcg_uncharge_kmem(memcg
, 1 << order
);
2745 page
->mem_cgroup
= NULL
;
2748 struct mem_cgroup
*__mem_cgroup_from_kmem(void *ptr
)
2750 struct mem_cgroup
*memcg
= NULL
;
2751 struct kmem_cache
*cachep
;
2754 page
= virt_to_head_page(ptr
);
2755 if (PageSlab(page
)) {
2756 cachep
= page
->slab_cache
;
2757 if (!is_root_cache(cachep
))
2758 memcg
= cachep
->memcg_params
.memcg
;
2760 /* page allocated by alloc_kmem_pages */
2761 memcg
= page
->mem_cgroup
;
2765 #endif /* CONFIG_MEMCG_KMEM */
2767 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2770 * Because tail pages are not marked as "used", set it. We're under
2771 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2772 * charge/uncharge will be never happen and move_account() is done under
2773 * compound_lock(), so we don't have to take care of races.
2775 void mem_cgroup_split_huge_fixup(struct page
*head
)
2779 if (mem_cgroup_disabled())
2782 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
2783 head
[i
].mem_cgroup
= head
->mem_cgroup
;
2785 __this_cpu_sub(head
->mem_cgroup
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
2788 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2790 #ifdef CONFIG_MEMCG_SWAP
2791 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
2794 int val
= (charge
) ? 1 : -1;
2795 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
2799 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2800 * @entry: swap entry to be moved
2801 * @from: mem_cgroup which the entry is moved from
2802 * @to: mem_cgroup which the entry is moved to
2804 * It succeeds only when the swap_cgroup's record for this entry is the same
2805 * as the mem_cgroup's id of @from.
2807 * Returns 0 on success, -EINVAL on failure.
2809 * The caller must have charged to @to, IOW, called page_counter_charge() about
2810 * both res and memsw, and called css_get().
2812 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2813 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2815 unsigned short old_id
, new_id
;
2817 old_id
= mem_cgroup_id(from
);
2818 new_id
= mem_cgroup_id(to
);
2820 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2821 mem_cgroup_swap_statistics(from
, false);
2822 mem_cgroup_swap_statistics(to
, true);
2828 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2829 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2835 static DEFINE_MUTEX(memcg_limit_mutex
);
2837 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
2838 unsigned long limit
)
2840 unsigned long curusage
;
2841 unsigned long oldusage
;
2842 bool enlarge
= false;
2847 * For keeping hierarchical_reclaim simple, how long we should retry
2848 * is depends on callers. We set our retry-count to be function
2849 * of # of children which we should visit in this loop.
2851 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2852 mem_cgroup_count_children(memcg
);
2854 oldusage
= page_counter_read(&memcg
->memory
);
2857 if (signal_pending(current
)) {
2862 mutex_lock(&memcg_limit_mutex
);
2863 if (limit
> memcg
->memsw
.limit
) {
2864 mutex_unlock(&memcg_limit_mutex
);
2868 if (limit
> memcg
->memory
.limit
)
2870 ret
= page_counter_limit(&memcg
->memory
, limit
);
2871 mutex_unlock(&memcg_limit_mutex
);
2876 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, true);
2878 curusage
= page_counter_read(&memcg
->memory
);
2879 /* Usage is reduced ? */
2880 if (curusage
>= oldusage
)
2883 oldusage
= curusage
;
2884 } while (retry_count
);
2886 if (!ret
&& enlarge
)
2887 memcg_oom_recover(memcg
);
2892 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
2893 unsigned long limit
)
2895 unsigned long curusage
;
2896 unsigned long oldusage
;
2897 bool enlarge
= false;
2901 /* see mem_cgroup_resize_res_limit */
2902 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2903 mem_cgroup_count_children(memcg
);
2905 oldusage
= page_counter_read(&memcg
->memsw
);
2908 if (signal_pending(current
)) {
2913 mutex_lock(&memcg_limit_mutex
);
2914 if (limit
< memcg
->memory
.limit
) {
2915 mutex_unlock(&memcg_limit_mutex
);
2919 if (limit
> memcg
->memsw
.limit
)
2921 ret
= page_counter_limit(&memcg
->memsw
, limit
);
2922 mutex_unlock(&memcg_limit_mutex
);
2927 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, false);
2929 curusage
= page_counter_read(&memcg
->memsw
);
2930 /* Usage is reduced ? */
2931 if (curusage
>= oldusage
)
2934 oldusage
= curusage
;
2935 } while (retry_count
);
2937 if (!ret
&& enlarge
)
2938 memcg_oom_recover(memcg
);
2943 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
2945 unsigned long *total_scanned
)
2947 unsigned long nr_reclaimed
= 0;
2948 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
2949 unsigned long reclaimed
;
2951 struct mem_cgroup_tree_per_zone
*mctz
;
2952 unsigned long excess
;
2953 unsigned long nr_scanned
;
2958 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
2960 * This loop can run a while, specially if mem_cgroup's continuously
2961 * keep exceeding their soft limit and putting the system under
2968 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
2973 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
2974 gfp_mask
, &nr_scanned
);
2975 nr_reclaimed
+= reclaimed
;
2976 *total_scanned
+= nr_scanned
;
2977 spin_lock_irq(&mctz
->lock
);
2978 __mem_cgroup_remove_exceeded(mz
, mctz
);
2981 * If we failed to reclaim anything from this memory cgroup
2982 * it is time to move on to the next cgroup
2986 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
2988 excess
= soft_limit_excess(mz
->memcg
);
2990 * One school of thought says that we should not add
2991 * back the node to the tree if reclaim returns 0.
2992 * But our reclaim could return 0, simply because due
2993 * to priority we are exposing a smaller subset of
2994 * memory to reclaim from. Consider this as a longer
2997 /* If excess == 0, no tree ops */
2998 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
2999 spin_unlock_irq(&mctz
->lock
);
3000 css_put(&mz
->memcg
->css
);
3003 * Could not reclaim anything and there are no more
3004 * mem cgroups to try or we seem to be looping without
3005 * reclaiming anything.
3007 if (!nr_reclaimed
&&
3009 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
3011 } while (!nr_reclaimed
);
3013 css_put(&next_mz
->memcg
->css
);
3014 return nr_reclaimed
;
3018 * Test whether @memcg has children, dead or alive. Note that this
3019 * function doesn't care whether @memcg has use_hierarchy enabled and
3020 * returns %true if there are child csses according to the cgroup
3021 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3023 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
3028 * The lock does not prevent addition or deletion of children, but
3029 * it prevents a new child from being initialized based on this
3030 * parent in css_online(), so it's enough to decide whether
3031 * hierarchically inherited attributes can still be changed or not.
3033 lockdep_assert_held(&memcg_create_mutex
);
3036 ret
= css_next_child(NULL
, &memcg
->css
);
3042 * Reclaims as many pages from the given memcg as possible and moves
3043 * the rest to the parent.
3045 * Caller is responsible for holding css reference for memcg.
3047 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
3049 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
3051 /* we call try-to-free pages for make this cgroup empty */
3052 lru_add_drain_all();
3053 /* try to free all pages in this cgroup */
3054 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
3057 if (signal_pending(current
))
3060 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
3064 /* maybe some writeback is necessary */
3065 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3073 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
3074 char *buf
, size_t nbytes
,
3077 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3079 if (mem_cgroup_is_root(memcg
))
3081 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
3084 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
3087 return mem_cgroup_from_css(css
)->use_hierarchy
;
3090 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
3091 struct cftype
*cft
, u64 val
)
3094 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3095 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
3097 mutex_lock(&memcg_create_mutex
);
3099 if (memcg
->use_hierarchy
== val
)
3103 * If parent's use_hierarchy is set, we can't make any modifications
3104 * in the child subtrees. If it is unset, then the change can
3105 * occur, provided the current cgroup has no children.
3107 * For the root cgroup, parent_mem is NULL, we allow value to be
3108 * set if there are no children.
3110 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
3111 (val
== 1 || val
== 0)) {
3112 if (!memcg_has_children(memcg
))
3113 memcg
->use_hierarchy
= val
;
3120 mutex_unlock(&memcg_create_mutex
);
3125 static unsigned long tree_stat(struct mem_cgroup
*memcg
,
3126 enum mem_cgroup_stat_index idx
)
3128 struct mem_cgroup
*iter
;
3131 /* Per-cpu values can be negative, use a signed accumulator */
3132 for_each_mem_cgroup_tree(iter
, memcg
)
3133 val
+= mem_cgroup_read_stat(iter
, idx
);
3135 if (val
< 0) /* race ? */
3140 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
3144 if (mem_cgroup_is_root(memcg
)) {
3145 val
= tree_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
3146 val
+= tree_stat(memcg
, MEM_CGROUP_STAT_RSS
);
3148 val
+= tree_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
3151 val
= page_counter_read(&memcg
->memory
);
3153 val
= page_counter_read(&memcg
->memsw
);
3155 return val
<< PAGE_SHIFT
;
3166 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
3169 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3170 struct page_counter
*counter
;
3172 switch (MEMFILE_TYPE(cft
->private)) {
3174 counter
= &memcg
->memory
;
3177 counter
= &memcg
->memsw
;
3180 counter
= &memcg
->kmem
;
3186 switch (MEMFILE_ATTR(cft
->private)) {
3188 if (counter
== &memcg
->memory
)
3189 return mem_cgroup_usage(memcg
, false);
3190 if (counter
== &memcg
->memsw
)
3191 return mem_cgroup_usage(memcg
, true);
3192 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
3194 return (u64
)counter
->limit
* PAGE_SIZE
;
3196 return (u64
)counter
->watermark
* PAGE_SIZE
;
3198 return counter
->failcnt
;
3199 case RES_SOFT_LIMIT
:
3200 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
3206 #ifdef CONFIG_MEMCG_KMEM
3207 static int memcg_activate_kmem(struct mem_cgroup
*memcg
,
3208 unsigned long nr_pages
)
3213 BUG_ON(memcg
->kmemcg_id
>= 0);
3214 BUG_ON(memcg
->kmem_acct_activated
);
3215 BUG_ON(memcg
->kmem_acct_active
);
3218 * For simplicity, we won't allow this to be disabled. It also can't
3219 * be changed if the cgroup has children already, or if tasks had
3222 * If tasks join before we set the limit, a person looking at
3223 * kmem.usage_in_bytes will have no way to determine when it took
3224 * place, which makes the value quite meaningless.
3226 * After it first became limited, changes in the value of the limit are
3227 * of course permitted.
3229 mutex_lock(&memcg_create_mutex
);
3230 if (cgroup_has_tasks(memcg
->css
.cgroup
) ||
3231 (memcg
->use_hierarchy
&& memcg_has_children(memcg
)))
3233 mutex_unlock(&memcg_create_mutex
);
3237 memcg_id
= memcg_alloc_cache_id();
3244 * We couldn't have accounted to this cgroup, because it hasn't got
3245 * activated yet, so this should succeed.
3247 err
= page_counter_limit(&memcg
->kmem
, nr_pages
);
3250 static_key_slow_inc(&memcg_kmem_enabled_key
);
3252 * A memory cgroup is considered kmem-active as soon as it gets
3253 * kmemcg_id. Setting the id after enabling static branching will
3254 * guarantee no one starts accounting before all call sites are
3257 memcg
->kmemcg_id
= memcg_id
;
3258 memcg
->kmem_acct_activated
= true;
3259 memcg
->kmem_acct_active
= true;
3264 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
3265 unsigned long limit
)
3269 mutex_lock(&memcg_limit_mutex
);
3270 if (!memcg_kmem_is_active(memcg
))
3271 ret
= memcg_activate_kmem(memcg
, limit
);
3273 ret
= page_counter_limit(&memcg
->kmem
, limit
);
3274 mutex_unlock(&memcg_limit_mutex
);
3278 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
3281 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
3286 mutex_lock(&memcg_limit_mutex
);
3288 * If the parent cgroup is not kmem-active now, it cannot be activated
3289 * after this point, because it has at least one child already.
3291 if (memcg_kmem_is_active(parent
))
3292 ret
= memcg_activate_kmem(memcg
, PAGE_COUNTER_MAX
);
3293 mutex_unlock(&memcg_limit_mutex
);
3297 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
3298 unsigned long limit
)
3302 #endif /* CONFIG_MEMCG_KMEM */
3305 * The user of this function is...
3308 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
3309 char *buf
, size_t nbytes
, loff_t off
)
3311 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3312 unsigned long nr_pages
;
3315 buf
= strstrip(buf
);
3316 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
3320 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3322 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3326 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3328 ret
= mem_cgroup_resize_limit(memcg
, nr_pages
);
3331 ret
= mem_cgroup_resize_memsw_limit(memcg
, nr_pages
);
3334 ret
= memcg_update_kmem_limit(memcg
, nr_pages
);
3338 case RES_SOFT_LIMIT
:
3339 memcg
->soft_limit
= nr_pages
;
3343 return ret
?: nbytes
;
3346 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3347 size_t nbytes
, loff_t off
)
3349 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3350 struct page_counter
*counter
;
3352 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3354 counter
= &memcg
->memory
;
3357 counter
= &memcg
->memsw
;
3360 counter
= &memcg
->kmem
;
3366 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3368 page_counter_reset_watermark(counter
);
3371 counter
->failcnt
= 0;
3380 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3383 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3387 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3388 struct cftype
*cft
, u64 val
)
3390 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3392 if (val
& ~MOVE_MASK
)
3396 * No kind of locking is needed in here, because ->can_attach() will
3397 * check this value once in the beginning of the process, and then carry
3398 * on with stale data. This means that changes to this value will only
3399 * affect task migrations starting after the change.
3401 memcg
->move_charge_at_immigrate
= val
;
3405 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3406 struct cftype
*cft
, u64 val
)
3413 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3417 unsigned int lru_mask
;
3420 static const struct numa_stat stats
[] = {
3421 { "total", LRU_ALL
},
3422 { "file", LRU_ALL_FILE
},
3423 { "anon", LRU_ALL_ANON
},
3424 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3426 const struct numa_stat
*stat
;
3429 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3431 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3432 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3433 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3434 for_each_node_state(nid
, N_MEMORY
) {
3435 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3437 seq_printf(m
, " N%d=%lu", nid
, nr
);
3442 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3443 struct mem_cgroup
*iter
;
3446 for_each_mem_cgroup_tree(iter
, memcg
)
3447 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3448 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3449 for_each_node_state(nid
, N_MEMORY
) {
3451 for_each_mem_cgroup_tree(iter
, memcg
)
3452 nr
+= mem_cgroup_node_nr_lru_pages(
3453 iter
, nid
, stat
->lru_mask
);
3454 seq_printf(m
, " N%d=%lu", nid
, nr
);
3461 #endif /* CONFIG_NUMA */
3463 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3465 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3466 unsigned long memory
, memsw
;
3467 struct mem_cgroup
*mi
;
3470 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names
) !=
3471 MEM_CGROUP_STAT_NSTATS
);
3472 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names
) !=
3473 MEM_CGROUP_EVENTS_NSTATS
);
3474 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3476 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3477 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
3479 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
3480 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
3483 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
3484 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
3485 mem_cgroup_read_events(memcg
, i
));
3487 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3488 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3489 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
3491 /* Hierarchical information */
3492 memory
= memsw
= PAGE_COUNTER_MAX
;
3493 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3494 memory
= min(memory
, mi
->memory
.limit
);
3495 memsw
= min(memsw
, mi
->memsw
.limit
);
3497 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3498 (u64
)memory
* PAGE_SIZE
);
3499 if (do_swap_account
)
3500 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3501 (u64
)memsw
* PAGE_SIZE
);
3503 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3506 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
3508 for_each_mem_cgroup_tree(mi
, memcg
)
3509 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
3510 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
3513 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
3514 unsigned long long val
= 0;
3516 for_each_mem_cgroup_tree(mi
, memcg
)
3517 val
+= mem_cgroup_read_events(mi
, i
);
3518 seq_printf(m
, "total_%s %llu\n",
3519 mem_cgroup_events_names
[i
], val
);
3522 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
3523 unsigned long long val
= 0;
3525 for_each_mem_cgroup_tree(mi
, memcg
)
3526 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
3527 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
3530 #ifdef CONFIG_DEBUG_VM
3533 struct mem_cgroup_per_zone
*mz
;
3534 struct zone_reclaim_stat
*rstat
;
3535 unsigned long recent_rotated
[2] = {0, 0};
3536 unsigned long recent_scanned
[2] = {0, 0};
3538 for_each_online_node(nid
)
3539 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
3540 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
3541 rstat
= &mz
->lruvec
.reclaim_stat
;
3543 recent_rotated
[0] += rstat
->recent_rotated
[0];
3544 recent_rotated
[1] += rstat
->recent_rotated
[1];
3545 recent_scanned
[0] += rstat
->recent_scanned
[0];
3546 recent_scanned
[1] += rstat
->recent_scanned
[1];
3548 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3549 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3550 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3551 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3558 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3561 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3563 return mem_cgroup_swappiness(memcg
);
3566 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3567 struct cftype
*cft
, u64 val
)
3569 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3575 memcg
->swappiness
= val
;
3577 vm_swappiness
= val
;
3582 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3584 struct mem_cgroup_threshold_ary
*t
;
3585 unsigned long usage
;
3590 t
= rcu_dereference(memcg
->thresholds
.primary
);
3592 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3597 usage
= mem_cgroup_usage(memcg
, swap
);
3600 * current_threshold points to threshold just below or equal to usage.
3601 * If it's not true, a threshold was crossed after last
3602 * call of __mem_cgroup_threshold().
3604 i
= t
->current_threshold
;
3607 * Iterate backward over array of thresholds starting from
3608 * current_threshold and check if a threshold is crossed.
3609 * If none of thresholds below usage is crossed, we read
3610 * only one element of the array here.
3612 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3613 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3615 /* i = current_threshold + 1 */
3619 * Iterate forward over array of thresholds starting from
3620 * current_threshold+1 and check if a threshold is crossed.
3621 * If none of thresholds above usage is crossed, we read
3622 * only one element of the array here.
3624 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3625 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3627 /* Update current_threshold */
3628 t
->current_threshold
= i
- 1;
3633 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3636 __mem_cgroup_threshold(memcg
, false);
3637 if (do_swap_account
)
3638 __mem_cgroup_threshold(memcg
, true);
3640 memcg
= parent_mem_cgroup(memcg
);
3644 static int compare_thresholds(const void *a
, const void *b
)
3646 const struct mem_cgroup_threshold
*_a
= a
;
3647 const struct mem_cgroup_threshold
*_b
= b
;
3649 if (_a
->threshold
> _b
->threshold
)
3652 if (_a
->threshold
< _b
->threshold
)
3658 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3660 struct mem_cgroup_eventfd_list
*ev
;
3662 spin_lock(&memcg_oom_lock
);
3664 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3665 eventfd_signal(ev
->eventfd
, 1);
3667 spin_unlock(&memcg_oom_lock
);
3671 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3673 struct mem_cgroup
*iter
;
3675 for_each_mem_cgroup_tree(iter
, memcg
)
3676 mem_cgroup_oom_notify_cb(iter
);
3679 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3680 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
3682 struct mem_cgroup_thresholds
*thresholds
;
3683 struct mem_cgroup_threshold_ary
*new;
3684 unsigned long threshold
;
3685 unsigned long usage
;
3688 ret
= page_counter_memparse(args
, "-1", &threshold
);
3692 mutex_lock(&memcg
->thresholds_lock
);
3695 thresholds
= &memcg
->thresholds
;
3696 usage
= mem_cgroup_usage(memcg
, false);
3697 } else if (type
== _MEMSWAP
) {
3698 thresholds
= &memcg
->memsw_thresholds
;
3699 usage
= mem_cgroup_usage(memcg
, true);
3703 /* Check if a threshold crossed before adding a new one */
3704 if (thresholds
->primary
)
3705 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3707 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3709 /* Allocate memory for new array of thresholds */
3710 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
3718 /* Copy thresholds (if any) to new array */
3719 if (thresholds
->primary
) {
3720 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3721 sizeof(struct mem_cgroup_threshold
));
3724 /* Add new threshold */
3725 new->entries
[size
- 1].eventfd
= eventfd
;
3726 new->entries
[size
- 1].threshold
= threshold
;
3728 /* Sort thresholds. Registering of new threshold isn't time-critical */
3729 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3730 compare_thresholds
, NULL
);
3732 /* Find current threshold */
3733 new->current_threshold
= -1;
3734 for (i
= 0; i
< size
; i
++) {
3735 if (new->entries
[i
].threshold
<= usage
) {
3737 * new->current_threshold will not be used until
3738 * rcu_assign_pointer(), so it's safe to increment
3741 ++new->current_threshold
;
3746 /* Free old spare buffer and save old primary buffer as spare */
3747 kfree(thresholds
->spare
);
3748 thresholds
->spare
= thresholds
->primary
;
3750 rcu_assign_pointer(thresholds
->primary
, new);
3752 /* To be sure that nobody uses thresholds */
3756 mutex_unlock(&memcg
->thresholds_lock
);
3761 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3762 struct eventfd_ctx
*eventfd
, const char *args
)
3764 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
3767 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3768 struct eventfd_ctx
*eventfd
, const char *args
)
3770 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
3773 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3774 struct eventfd_ctx
*eventfd
, enum res_type type
)
3776 struct mem_cgroup_thresholds
*thresholds
;
3777 struct mem_cgroup_threshold_ary
*new;
3778 unsigned long usage
;
3781 mutex_lock(&memcg
->thresholds_lock
);
3784 thresholds
= &memcg
->thresholds
;
3785 usage
= mem_cgroup_usage(memcg
, false);
3786 } else if (type
== _MEMSWAP
) {
3787 thresholds
= &memcg
->memsw_thresholds
;
3788 usage
= mem_cgroup_usage(memcg
, true);
3792 if (!thresholds
->primary
)
3795 /* Check if a threshold crossed before removing */
3796 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3798 /* Calculate new number of threshold */
3800 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
3801 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
3805 new = thresholds
->spare
;
3807 /* Set thresholds array to NULL if we don't have thresholds */
3816 /* Copy thresholds and find current threshold */
3817 new->current_threshold
= -1;
3818 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
3819 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
3822 new->entries
[j
] = thresholds
->primary
->entries
[i
];
3823 if (new->entries
[j
].threshold
<= usage
) {
3825 * new->current_threshold will not be used
3826 * until rcu_assign_pointer(), so it's safe to increment
3829 ++new->current_threshold
;
3835 /* Swap primary and spare array */
3836 thresholds
->spare
= thresholds
->primary
;
3837 /* If all events are unregistered, free the spare array */
3839 kfree(thresholds
->spare
);
3840 thresholds
->spare
= NULL
;
3843 rcu_assign_pointer(thresholds
->primary
, new);
3845 /* To be sure that nobody uses thresholds */
3848 mutex_unlock(&memcg
->thresholds_lock
);
3851 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3852 struct eventfd_ctx
*eventfd
)
3854 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
3857 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3858 struct eventfd_ctx
*eventfd
)
3860 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
3863 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
3864 struct eventfd_ctx
*eventfd
, const char *args
)
3866 struct mem_cgroup_eventfd_list
*event
;
3868 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
3872 spin_lock(&memcg_oom_lock
);
3874 event
->eventfd
= eventfd
;
3875 list_add(&event
->list
, &memcg
->oom_notify
);
3877 /* already in OOM ? */
3878 if (atomic_read(&memcg
->under_oom
))
3879 eventfd_signal(eventfd
, 1);
3880 spin_unlock(&memcg_oom_lock
);
3885 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
3886 struct eventfd_ctx
*eventfd
)
3888 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
3890 spin_lock(&memcg_oom_lock
);
3892 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
3893 if (ev
->eventfd
== eventfd
) {
3894 list_del(&ev
->list
);
3899 spin_unlock(&memcg_oom_lock
);
3902 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
3904 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
3906 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
3907 seq_printf(sf
, "under_oom %d\n", (bool)atomic_read(&memcg
->under_oom
));
3911 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
3912 struct cftype
*cft
, u64 val
)
3914 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3916 /* cannot set to root cgroup and only 0 and 1 are allowed */
3917 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
3920 memcg
->oom_kill_disable
= val
;
3922 memcg_oom_recover(memcg
);
3927 #ifdef CONFIG_MEMCG_KMEM
3928 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
3932 ret
= memcg_propagate_kmem(memcg
);
3936 return mem_cgroup_sockets_init(memcg
, ss
);
3939 static void memcg_deactivate_kmem(struct mem_cgroup
*memcg
)
3941 struct cgroup_subsys_state
*css
;
3942 struct mem_cgroup
*parent
, *child
;
3945 if (!memcg
->kmem_acct_active
)
3949 * Clear the 'active' flag before clearing memcg_caches arrays entries.
3950 * Since we take the slab_mutex in memcg_deactivate_kmem_caches(), it
3951 * guarantees no cache will be created for this cgroup after we are
3952 * done (see memcg_create_kmem_cache()).
3954 memcg
->kmem_acct_active
= false;
3956 memcg_deactivate_kmem_caches(memcg
);
3958 kmemcg_id
= memcg
->kmemcg_id
;
3959 BUG_ON(kmemcg_id
< 0);
3961 parent
= parent_mem_cgroup(memcg
);
3963 parent
= root_mem_cgroup
;
3966 * Change kmemcg_id of this cgroup and all its descendants to the
3967 * parent's id, and then move all entries from this cgroup's list_lrus
3968 * to ones of the parent. After we have finished, all list_lrus
3969 * corresponding to this cgroup are guaranteed to remain empty. The
3970 * ordering is imposed by list_lru_node->lock taken by
3971 * memcg_drain_all_list_lrus().
3973 css_for_each_descendant_pre(css
, &memcg
->css
) {
3974 child
= mem_cgroup_from_css(css
);
3975 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
3976 child
->kmemcg_id
= parent
->kmemcg_id
;
3977 if (!memcg
->use_hierarchy
)
3980 memcg_drain_all_list_lrus(kmemcg_id
, parent
->kmemcg_id
);
3982 memcg_free_cache_id(kmemcg_id
);
3985 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
3987 if (memcg
->kmem_acct_activated
) {
3988 memcg_destroy_kmem_caches(memcg
);
3989 static_key_slow_dec(&memcg_kmem_enabled_key
);
3990 WARN_ON(page_counter_read(&memcg
->kmem
));
3992 mem_cgroup_sockets_destroy(memcg
);
3995 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
4000 static void memcg_deactivate_kmem(struct mem_cgroup
*memcg
)
4004 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
4010 * DO NOT USE IN NEW FILES.
4012 * "cgroup.event_control" implementation.
4014 * This is way over-engineered. It tries to support fully configurable
4015 * events for each user. Such level of flexibility is completely
4016 * unnecessary especially in the light of the planned unified hierarchy.
4018 * Please deprecate this and replace with something simpler if at all
4023 * Unregister event and free resources.
4025 * Gets called from workqueue.
4027 static void memcg_event_remove(struct work_struct
*work
)
4029 struct mem_cgroup_event
*event
=
4030 container_of(work
, struct mem_cgroup_event
, remove
);
4031 struct mem_cgroup
*memcg
= event
->memcg
;
4033 remove_wait_queue(event
->wqh
, &event
->wait
);
4035 event
->unregister_event(memcg
, event
->eventfd
);
4037 /* Notify userspace the event is going away. */
4038 eventfd_signal(event
->eventfd
, 1);
4040 eventfd_ctx_put(event
->eventfd
);
4042 css_put(&memcg
->css
);
4046 * Gets called on POLLHUP on eventfd when user closes it.
4048 * Called with wqh->lock held and interrupts disabled.
4050 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
4051 int sync
, void *key
)
4053 struct mem_cgroup_event
*event
=
4054 container_of(wait
, struct mem_cgroup_event
, wait
);
4055 struct mem_cgroup
*memcg
= event
->memcg
;
4056 unsigned long flags
= (unsigned long)key
;
4058 if (flags
& POLLHUP
) {
4060 * If the event has been detached at cgroup removal, we
4061 * can simply return knowing the other side will cleanup
4064 * We can't race against event freeing since the other
4065 * side will require wqh->lock via remove_wait_queue(),
4068 spin_lock(&memcg
->event_list_lock
);
4069 if (!list_empty(&event
->list
)) {
4070 list_del_init(&event
->list
);
4072 * We are in atomic context, but cgroup_event_remove()
4073 * may sleep, so we have to call it in workqueue.
4075 schedule_work(&event
->remove
);
4077 spin_unlock(&memcg
->event_list_lock
);
4083 static void memcg_event_ptable_queue_proc(struct file
*file
,
4084 wait_queue_head_t
*wqh
, poll_table
*pt
)
4086 struct mem_cgroup_event
*event
=
4087 container_of(pt
, struct mem_cgroup_event
, pt
);
4090 add_wait_queue(wqh
, &event
->wait
);
4094 * DO NOT USE IN NEW FILES.
4096 * Parse input and register new cgroup event handler.
4098 * Input must be in format '<event_fd> <control_fd> <args>'.
4099 * Interpretation of args is defined by control file implementation.
4101 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
4102 char *buf
, size_t nbytes
, loff_t off
)
4104 struct cgroup_subsys_state
*css
= of_css(of
);
4105 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4106 struct mem_cgroup_event
*event
;
4107 struct cgroup_subsys_state
*cfile_css
;
4108 unsigned int efd
, cfd
;
4115 buf
= strstrip(buf
);
4117 efd
= simple_strtoul(buf
, &endp
, 10);
4122 cfd
= simple_strtoul(buf
, &endp
, 10);
4123 if ((*endp
!= ' ') && (*endp
!= '\0'))
4127 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
4131 event
->memcg
= memcg
;
4132 INIT_LIST_HEAD(&event
->list
);
4133 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
4134 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
4135 INIT_WORK(&event
->remove
, memcg_event_remove
);
4143 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
4144 if (IS_ERR(event
->eventfd
)) {
4145 ret
= PTR_ERR(event
->eventfd
);
4152 goto out_put_eventfd
;
4155 /* the process need read permission on control file */
4156 /* AV: shouldn't we check that it's been opened for read instead? */
4157 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
4162 * Determine the event callbacks and set them in @event. This used
4163 * to be done via struct cftype but cgroup core no longer knows
4164 * about these events. The following is crude but the whole thing
4165 * is for compatibility anyway.
4167 * DO NOT ADD NEW FILES.
4169 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
4171 if (!strcmp(name
, "memory.usage_in_bytes")) {
4172 event
->register_event
= mem_cgroup_usage_register_event
;
4173 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
4174 } else if (!strcmp(name
, "memory.oom_control")) {
4175 event
->register_event
= mem_cgroup_oom_register_event
;
4176 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
4177 } else if (!strcmp(name
, "memory.pressure_level")) {
4178 event
->register_event
= vmpressure_register_event
;
4179 event
->unregister_event
= vmpressure_unregister_event
;
4180 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
4181 event
->register_event
= memsw_cgroup_usage_register_event
;
4182 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
4189 * Verify @cfile should belong to @css. Also, remaining events are
4190 * automatically removed on cgroup destruction but the removal is
4191 * asynchronous, so take an extra ref on @css.
4193 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
4194 &memory_cgrp_subsys
);
4196 if (IS_ERR(cfile_css
))
4198 if (cfile_css
!= css
) {
4203 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
4207 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
4209 spin_lock(&memcg
->event_list_lock
);
4210 list_add(&event
->list
, &memcg
->event_list
);
4211 spin_unlock(&memcg
->event_list_lock
);
4223 eventfd_ctx_put(event
->eventfd
);
4232 static struct cftype mem_cgroup_legacy_files
[] = {
4234 .name
= "usage_in_bytes",
4235 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
4236 .read_u64
= mem_cgroup_read_u64
,
4239 .name
= "max_usage_in_bytes",
4240 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
4241 .write
= mem_cgroup_reset
,
4242 .read_u64
= mem_cgroup_read_u64
,
4245 .name
= "limit_in_bytes",
4246 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
4247 .write
= mem_cgroup_write
,
4248 .read_u64
= mem_cgroup_read_u64
,
4251 .name
= "soft_limit_in_bytes",
4252 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
4253 .write
= mem_cgroup_write
,
4254 .read_u64
= mem_cgroup_read_u64
,
4258 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
4259 .write
= mem_cgroup_reset
,
4260 .read_u64
= mem_cgroup_read_u64
,
4264 .seq_show
= memcg_stat_show
,
4267 .name
= "force_empty",
4268 .write
= mem_cgroup_force_empty_write
,
4271 .name
= "use_hierarchy",
4272 .write_u64
= mem_cgroup_hierarchy_write
,
4273 .read_u64
= mem_cgroup_hierarchy_read
,
4276 .name
= "cgroup.event_control", /* XXX: for compat */
4277 .write
= memcg_write_event_control
,
4278 .flags
= CFTYPE_NO_PREFIX
,
4282 .name
= "swappiness",
4283 .read_u64
= mem_cgroup_swappiness_read
,
4284 .write_u64
= mem_cgroup_swappiness_write
,
4287 .name
= "move_charge_at_immigrate",
4288 .read_u64
= mem_cgroup_move_charge_read
,
4289 .write_u64
= mem_cgroup_move_charge_write
,
4292 .name
= "oom_control",
4293 .seq_show
= mem_cgroup_oom_control_read
,
4294 .write_u64
= mem_cgroup_oom_control_write
,
4295 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4298 .name
= "pressure_level",
4302 .name
= "numa_stat",
4303 .seq_show
= memcg_numa_stat_show
,
4306 #ifdef CONFIG_MEMCG_KMEM
4308 .name
= "kmem.limit_in_bytes",
4309 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
4310 .write
= mem_cgroup_write
,
4311 .read_u64
= mem_cgroup_read_u64
,
4314 .name
= "kmem.usage_in_bytes",
4315 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
4316 .read_u64
= mem_cgroup_read_u64
,
4319 .name
= "kmem.failcnt",
4320 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
4321 .write
= mem_cgroup_reset
,
4322 .read_u64
= mem_cgroup_read_u64
,
4325 .name
= "kmem.max_usage_in_bytes",
4326 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4327 .write
= mem_cgroup_reset
,
4328 .read_u64
= mem_cgroup_read_u64
,
4330 #ifdef CONFIG_SLABINFO
4332 .name
= "kmem.slabinfo",
4333 .seq_start
= slab_start
,
4334 .seq_next
= slab_next
,
4335 .seq_stop
= slab_stop
,
4336 .seq_show
= memcg_slab_show
,
4340 { }, /* terminate */
4343 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4345 struct mem_cgroup_per_node
*pn
;
4346 struct mem_cgroup_per_zone
*mz
;
4347 int zone
, tmp
= node
;
4349 * This routine is called against possible nodes.
4350 * But it's BUG to call kmalloc() against offline node.
4352 * TODO: this routine can waste much memory for nodes which will
4353 * never be onlined. It's better to use memory hotplug callback
4356 if (!node_state(node
, N_NORMAL_MEMORY
))
4358 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4362 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
4363 mz
= &pn
->zoneinfo
[zone
];
4364 lruvec_init(&mz
->lruvec
);
4365 mz
->usage_in_excess
= 0;
4366 mz
->on_tree
= false;
4369 memcg
->nodeinfo
[node
] = pn
;
4373 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4375 kfree(memcg
->nodeinfo
[node
]);
4378 static struct mem_cgroup
*mem_cgroup_alloc(void)
4380 struct mem_cgroup
*memcg
;
4383 size
= sizeof(struct mem_cgroup
);
4384 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4386 memcg
= kzalloc(size
, GFP_KERNEL
);
4390 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4393 spin_lock_init(&memcg
->pcp_counter_lock
);
4402 * At destroying mem_cgroup, references from swap_cgroup can remain.
4403 * (scanning all at force_empty is too costly...)
4405 * Instead of clearing all references at force_empty, we remember
4406 * the number of reference from swap_cgroup and free mem_cgroup when
4407 * it goes down to 0.
4409 * Removal of cgroup itself succeeds regardless of refs from swap.
4412 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
4416 mem_cgroup_remove_from_trees(memcg
);
4419 free_mem_cgroup_per_zone_info(memcg
, node
);
4421 free_percpu(memcg
->stat
);
4426 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4428 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
4430 if (!memcg
->memory
.parent
)
4432 return mem_cgroup_from_counter(memcg
->memory
.parent
, memory
);
4434 EXPORT_SYMBOL(parent_mem_cgroup
);
4436 static struct cgroup_subsys_state
* __ref
4437 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4439 struct mem_cgroup
*memcg
;
4440 long error
= -ENOMEM
;
4443 memcg
= mem_cgroup_alloc();
4445 return ERR_PTR(error
);
4448 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
4452 if (parent_css
== NULL
) {
4453 root_mem_cgroup
= memcg
;
4454 page_counter_init(&memcg
->memory
, NULL
);
4455 memcg
->high
= PAGE_COUNTER_MAX
;
4456 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4457 page_counter_init(&memcg
->memsw
, NULL
);
4458 page_counter_init(&memcg
->kmem
, NULL
);
4461 memcg
->last_scanned_node
= MAX_NUMNODES
;
4462 INIT_LIST_HEAD(&memcg
->oom_notify
);
4463 memcg
->move_charge_at_immigrate
= 0;
4464 mutex_init(&memcg
->thresholds_lock
);
4465 spin_lock_init(&memcg
->move_lock
);
4466 vmpressure_init(&memcg
->vmpressure
);
4467 INIT_LIST_HEAD(&memcg
->event_list
);
4468 spin_lock_init(&memcg
->event_list_lock
);
4469 #ifdef CONFIG_MEMCG_KMEM
4470 memcg
->kmemcg_id
= -1;
4476 __mem_cgroup_free(memcg
);
4477 return ERR_PTR(error
);
4481 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
4483 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4484 struct mem_cgroup
*parent
= mem_cgroup_from_css(css
->parent
);
4487 if (css
->id
> MEM_CGROUP_ID_MAX
)
4493 mutex_lock(&memcg_create_mutex
);
4495 memcg
->use_hierarchy
= parent
->use_hierarchy
;
4496 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4497 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4499 if (parent
->use_hierarchy
) {
4500 page_counter_init(&memcg
->memory
, &parent
->memory
);
4501 memcg
->high
= PAGE_COUNTER_MAX
;
4502 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4503 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4504 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4507 * No need to take a reference to the parent because cgroup
4508 * core guarantees its existence.
4511 page_counter_init(&memcg
->memory
, NULL
);
4512 memcg
->high
= PAGE_COUNTER_MAX
;
4513 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4514 page_counter_init(&memcg
->memsw
, NULL
);
4515 page_counter_init(&memcg
->kmem
, NULL
);
4517 * Deeper hierachy with use_hierarchy == false doesn't make
4518 * much sense so let cgroup subsystem know about this
4519 * unfortunate state in our controller.
4521 if (parent
!= root_mem_cgroup
)
4522 memory_cgrp_subsys
.broken_hierarchy
= true;
4524 mutex_unlock(&memcg_create_mutex
);
4526 ret
= memcg_init_kmem(memcg
, &memory_cgrp_subsys
);
4531 * Make sure the memcg is initialized: mem_cgroup_iter()
4532 * orders reading memcg->initialized against its callers
4533 * reading the memcg members.
4535 smp_store_release(&memcg
->initialized
, 1);
4540 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
4542 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4543 struct mem_cgroup_event
*event
, *tmp
;
4546 * Unregister events and notify userspace.
4547 * Notify userspace about cgroup removing only after rmdir of cgroup
4548 * directory to avoid race between userspace and kernelspace.
4550 spin_lock(&memcg
->event_list_lock
);
4551 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
4552 list_del_init(&event
->list
);
4553 schedule_work(&event
->remove
);
4555 spin_unlock(&memcg
->event_list_lock
);
4557 vmpressure_cleanup(&memcg
->vmpressure
);
4559 memcg_deactivate_kmem(memcg
);
4562 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
4564 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4566 memcg_destroy_kmem(memcg
);
4567 __mem_cgroup_free(memcg
);
4571 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4572 * @css: the target css
4574 * Reset the states of the mem_cgroup associated with @css. This is
4575 * invoked when the userland requests disabling on the default hierarchy
4576 * but the memcg is pinned through dependency. The memcg should stop
4577 * applying policies and should revert to the vanilla state as it may be
4578 * made visible again.
4580 * The current implementation only resets the essential configurations.
4581 * This needs to be expanded to cover all the visible parts.
4583 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
4585 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4587 mem_cgroup_resize_limit(memcg
, PAGE_COUNTER_MAX
);
4588 mem_cgroup_resize_memsw_limit(memcg
, PAGE_COUNTER_MAX
);
4589 memcg_update_kmem_limit(memcg
, PAGE_COUNTER_MAX
);
4591 memcg
->high
= PAGE_COUNTER_MAX
;
4592 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4596 /* Handlers for move charge at task migration. */
4597 static int mem_cgroup_do_precharge(unsigned long count
)
4601 /* Try a single bulk charge without reclaim first */
4602 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_WAIT
, count
);
4604 mc
.precharge
+= count
;
4607 if (ret
== -EINTR
) {
4608 cancel_charge(root_mem_cgroup
, count
);
4612 /* Try charges one by one with reclaim */
4614 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_NORETRY
, 1);
4616 * In case of failure, any residual charges against
4617 * mc.to will be dropped by mem_cgroup_clear_mc()
4618 * later on. However, cancel any charges that are
4619 * bypassed to root right away or they'll be lost.
4622 cancel_charge(root_mem_cgroup
, 1);
4632 * get_mctgt_type - get target type of moving charge
4633 * @vma: the vma the pte to be checked belongs
4634 * @addr: the address corresponding to the pte to be checked
4635 * @ptent: the pte to be checked
4636 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4639 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4640 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4641 * move charge. if @target is not NULL, the page is stored in target->page
4642 * with extra refcnt got(Callers should handle it).
4643 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4644 * target for charge migration. if @target is not NULL, the entry is stored
4647 * Called with pte lock held.
4654 enum mc_target_type
{
4660 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4661 unsigned long addr
, pte_t ptent
)
4663 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
4665 if (!page
|| !page_mapped(page
))
4667 if (PageAnon(page
)) {
4668 if (!(mc
.flags
& MOVE_ANON
))
4671 if (!(mc
.flags
& MOVE_FILE
))
4674 if (!get_page_unless_zero(page
))
4681 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4682 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4684 struct page
*page
= NULL
;
4685 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4687 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
4690 * Because lookup_swap_cache() updates some statistics counter,
4691 * we call find_get_page() with swapper_space directly.
4693 page
= find_get_page(swap_address_space(ent
), ent
.val
);
4694 if (do_swap_account
)
4695 entry
->val
= ent
.val
;
4700 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4701 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4707 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4708 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4710 struct page
*page
= NULL
;
4711 struct address_space
*mapping
;
4714 if (!vma
->vm_file
) /* anonymous vma */
4716 if (!(mc
.flags
& MOVE_FILE
))
4719 mapping
= vma
->vm_file
->f_mapping
;
4720 pgoff
= linear_page_index(vma
, addr
);
4722 /* page is moved even if it's not RSS of this task(page-faulted). */
4724 /* shmem/tmpfs may report page out on swap: account for that too. */
4725 if (shmem_mapping(mapping
)) {
4726 page
= find_get_entry(mapping
, pgoff
);
4727 if (radix_tree_exceptional_entry(page
)) {
4728 swp_entry_t swp
= radix_to_swp_entry(page
);
4729 if (do_swap_account
)
4731 page
= find_get_page(swap_address_space(swp
), swp
.val
);
4734 page
= find_get_page(mapping
, pgoff
);
4736 page
= find_get_page(mapping
, pgoff
);
4742 * mem_cgroup_move_account - move account of the page
4744 * @nr_pages: number of regular pages (>1 for huge pages)
4745 * @from: mem_cgroup which the page is moved from.
4746 * @to: mem_cgroup which the page is moved to. @from != @to.
4748 * The caller must confirm following.
4749 * - page is not on LRU (isolate_page() is useful.)
4750 * - compound_lock is held when nr_pages > 1
4752 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4755 static int mem_cgroup_move_account(struct page
*page
,
4756 unsigned int nr_pages
,
4757 struct mem_cgroup
*from
,
4758 struct mem_cgroup
*to
)
4760 unsigned long flags
;
4763 VM_BUG_ON(from
== to
);
4764 VM_BUG_ON_PAGE(PageLRU(page
), page
);
4766 * The page is isolated from LRU. So, collapse function
4767 * will not handle this page. But page splitting can happen.
4768 * Do this check under compound_page_lock(). The caller should
4772 if (nr_pages
> 1 && !PageTransHuge(page
))
4776 * Prevent mem_cgroup_migrate() from looking at page->mem_cgroup
4777 * of its source page while we change it: page migration takes
4778 * both pages off the LRU, but page cache replacement doesn't.
4780 if (!trylock_page(page
))
4784 if (page
->mem_cgroup
!= from
)
4787 spin_lock_irqsave(&from
->move_lock
, flags
);
4789 if (!PageAnon(page
) && page_mapped(page
)) {
4790 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4792 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4796 if (PageWriteback(page
)) {
4797 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4799 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4804 * It is safe to change page->mem_cgroup here because the page
4805 * is referenced, charged, and isolated - we can't race with
4806 * uncharging, charging, migration, or LRU putback.
4809 /* caller should have done css_get */
4810 page
->mem_cgroup
= to
;
4811 spin_unlock_irqrestore(&from
->move_lock
, flags
);
4815 local_irq_disable();
4816 mem_cgroup_charge_statistics(to
, page
, nr_pages
);
4817 memcg_check_events(to
, page
);
4818 mem_cgroup_charge_statistics(from
, page
, -nr_pages
);
4819 memcg_check_events(from
, page
);
4827 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
4828 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4830 struct page
*page
= NULL
;
4831 enum mc_target_type ret
= MC_TARGET_NONE
;
4832 swp_entry_t ent
= { .val
= 0 };
4834 if (pte_present(ptent
))
4835 page
= mc_handle_present_pte(vma
, addr
, ptent
);
4836 else if (is_swap_pte(ptent
))
4837 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
4838 else if (pte_none(ptent
))
4839 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
4841 if (!page
&& !ent
.val
)
4845 * Do only loose check w/o serialization.
4846 * mem_cgroup_move_account() checks the page is valid or
4847 * not under LRU exclusion.
4849 if (page
->mem_cgroup
== mc
.from
) {
4850 ret
= MC_TARGET_PAGE
;
4852 target
->page
= page
;
4854 if (!ret
|| !target
)
4857 /* There is a swap entry and a page doesn't exist or isn't charged */
4858 if (ent
.val
&& !ret
&&
4859 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
4860 ret
= MC_TARGET_SWAP
;
4867 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4869 * We don't consider swapping or file mapped pages because THP does not
4870 * support them for now.
4871 * Caller should make sure that pmd_trans_huge(pmd) is true.
4873 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4874 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4876 struct page
*page
= NULL
;
4877 enum mc_target_type ret
= MC_TARGET_NONE
;
4879 page
= pmd_page(pmd
);
4880 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
4881 if (!(mc
.flags
& MOVE_ANON
))
4883 if (page
->mem_cgroup
== mc
.from
) {
4884 ret
= MC_TARGET_PAGE
;
4887 target
->page
= page
;
4893 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4894 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4896 return MC_TARGET_NONE
;
4900 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
4901 unsigned long addr
, unsigned long end
,
4902 struct mm_walk
*walk
)
4904 struct vm_area_struct
*vma
= walk
->vma
;
4908 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
4909 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
4910 mc
.precharge
+= HPAGE_PMD_NR
;
4915 if (pmd_trans_unstable(pmd
))
4917 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4918 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
4919 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
4920 mc
.precharge
++; /* increment precharge temporarily */
4921 pte_unmap_unlock(pte
- 1, ptl
);
4927 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
4929 unsigned long precharge
;
4931 struct mm_walk mem_cgroup_count_precharge_walk
= {
4932 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
4935 down_read(&mm
->mmap_sem
);
4936 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk
);
4937 up_read(&mm
->mmap_sem
);
4939 precharge
= mc
.precharge
;
4945 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
4947 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
4949 VM_BUG_ON(mc
.moving_task
);
4950 mc
.moving_task
= current
;
4951 return mem_cgroup_do_precharge(precharge
);
4954 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4955 static void __mem_cgroup_clear_mc(void)
4957 struct mem_cgroup
*from
= mc
.from
;
4958 struct mem_cgroup
*to
= mc
.to
;
4960 /* we must uncharge all the leftover precharges from mc.to */
4962 cancel_charge(mc
.to
, mc
.precharge
);
4966 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4967 * we must uncharge here.
4969 if (mc
.moved_charge
) {
4970 cancel_charge(mc
.from
, mc
.moved_charge
);
4971 mc
.moved_charge
= 0;
4973 /* we must fixup refcnts and charges */
4974 if (mc
.moved_swap
) {
4975 /* uncharge swap account from the old cgroup */
4976 if (!mem_cgroup_is_root(mc
.from
))
4977 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
4980 * we charged both to->memory and to->memsw, so we
4981 * should uncharge to->memory.
4983 if (!mem_cgroup_is_root(mc
.to
))
4984 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
4986 css_put_many(&mc
.from
->css
, mc
.moved_swap
);
4988 /* we've already done css_get(mc.to) */
4991 memcg_oom_recover(from
);
4992 memcg_oom_recover(to
);
4993 wake_up_all(&mc
.waitq
);
4996 static void mem_cgroup_clear_mc(void)
4999 * we must clear moving_task before waking up waiters at the end of
5002 mc
.moving_task
= NULL
;
5003 __mem_cgroup_clear_mc();
5004 spin_lock(&mc
.lock
);
5007 spin_unlock(&mc
.lock
);
5010 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
5011 struct cgroup_taskset
*tset
)
5013 struct task_struct
*p
= cgroup_taskset_first(tset
);
5015 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5016 unsigned long move_flags
;
5019 * We are now commited to this value whatever it is. Changes in this
5020 * tunable will only affect upcoming migrations, not the current one.
5021 * So we need to save it, and keep it going.
5023 move_flags
= ACCESS_ONCE(memcg
->move_charge_at_immigrate
);
5025 struct mm_struct
*mm
;
5026 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
5028 VM_BUG_ON(from
== memcg
);
5030 mm
= get_task_mm(p
);
5033 /* We move charges only when we move a owner of the mm */
5034 if (mm
->owner
== p
) {
5037 VM_BUG_ON(mc
.precharge
);
5038 VM_BUG_ON(mc
.moved_charge
);
5039 VM_BUG_ON(mc
.moved_swap
);
5041 spin_lock(&mc
.lock
);
5044 mc
.flags
= move_flags
;
5045 spin_unlock(&mc
.lock
);
5046 /* We set mc.moving_task later */
5048 ret
= mem_cgroup_precharge_mc(mm
);
5050 mem_cgroup_clear_mc();
5057 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
5058 struct cgroup_taskset
*tset
)
5061 mem_cgroup_clear_mc();
5064 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
5065 unsigned long addr
, unsigned long end
,
5066 struct mm_walk
*walk
)
5069 struct vm_area_struct
*vma
= walk
->vma
;
5072 enum mc_target_type target_type
;
5073 union mc_target target
;
5077 * We don't take compound_lock() here but no race with splitting thp
5079 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5080 * under splitting, which means there's no concurrent thp split,
5081 * - if another thread runs into split_huge_page() just after we
5082 * entered this if-block, the thread must wait for page table lock
5083 * to be unlocked in __split_huge_page_splitting(), where the main
5084 * part of thp split is not executed yet.
5086 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
5087 if (mc
.precharge
< HPAGE_PMD_NR
) {
5091 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
5092 if (target_type
== MC_TARGET_PAGE
) {
5094 if (!isolate_lru_page(page
)) {
5095 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
5097 mc
.precharge
-= HPAGE_PMD_NR
;
5098 mc
.moved_charge
+= HPAGE_PMD_NR
;
5100 putback_lru_page(page
);
5108 if (pmd_trans_unstable(pmd
))
5111 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5112 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
5113 pte_t ptent
= *(pte
++);
5119 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
5120 case MC_TARGET_PAGE
:
5122 if (isolate_lru_page(page
))
5124 if (!mem_cgroup_move_account(page
, 1, mc
.from
, mc
.to
)) {
5126 /* we uncharge from mc.from later. */
5129 putback_lru_page(page
);
5130 put
: /* get_mctgt_type() gets the page */
5133 case MC_TARGET_SWAP
:
5135 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
5137 /* we fixup refcnts and charges later. */
5145 pte_unmap_unlock(pte
- 1, ptl
);
5150 * We have consumed all precharges we got in can_attach().
5151 * We try charge one by one, but don't do any additional
5152 * charges to mc.to if we have failed in charge once in attach()
5155 ret
= mem_cgroup_do_precharge(1);
5163 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
5165 struct mm_walk mem_cgroup_move_charge_walk
= {
5166 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
5170 lru_add_drain_all();
5172 * Signal mem_cgroup_begin_page_stat() to take the memcg's
5173 * move_lock while we're moving its pages to another memcg.
5174 * Then wait for already started RCU-only updates to finish.
5176 atomic_inc(&mc
.from
->moving_account
);
5179 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
5181 * Someone who are holding the mmap_sem might be waiting in
5182 * waitq. So we cancel all extra charges, wake up all waiters,
5183 * and retry. Because we cancel precharges, we might not be able
5184 * to move enough charges, but moving charge is a best-effort
5185 * feature anyway, so it wouldn't be a big problem.
5187 __mem_cgroup_clear_mc();
5192 * When we have consumed all precharges and failed in doing
5193 * additional charge, the page walk just aborts.
5195 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk
);
5196 up_read(&mm
->mmap_sem
);
5197 atomic_dec(&mc
.from
->moving_account
);
5200 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
5201 struct cgroup_taskset
*tset
)
5203 struct task_struct
*p
= cgroup_taskset_first(tset
);
5204 struct mm_struct
*mm
= get_task_mm(p
);
5208 mem_cgroup_move_charge(mm
);
5212 mem_cgroup_clear_mc();
5214 #else /* !CONFIG_MMU */
5215 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
5216 struct cgroup_taskset
*tset
)
5220 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
5221 struct cgroup_taskset
*tset
)
5224 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
5225 struct cgroup_taskset
*tset
)
5231 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5232 * to verify whether we're attached to the default hierarchy on each mount
5235 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
5238 * use_hierarchy is forced on the default hierarchy. cgroup core
5239 * guarantees that @root doesn't have any children, so turning it
5240 * on for the root memcg is enough.
5242 if (cgroup_on_dfl(root_css
->cgroup
))
5243 root_mem_cgroup
->use_hierarchy
= true;
5245 root_mem_cgroup
->use_hierarchy
= false;
5248 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
5251 return mem_cgroup_usage(mem_cgroup_from_css(css
), false);
5254 static int memory_low_show(struct seq_file
*m
, void *v
)
5256 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5257 unsigned long low
= ACCESS_ONCE(memcg
->low
);
5259 if (low
== PAGE_COUNTER_MAX
)
5260 seq_puts(m
, "max\n");
5262 seq_printf(m
, "%llu\n", (u64
)low
* PAGE_SIZE
);
5267 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
5268 char *buf
, size_t nbytes
, loff_t off
)
5270 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5274 buf
= strstrip(buf
);
5275 err
= page_counter_memparse(buf
, "max", &low
);
5284 static int memory_high_show(struct seq_file
*m
, void *v
)
5286 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5287 unsigned long high
= ACCESS_ONCE(memcg
->high
);
5289 if (high
== PAGE_COUNTER_MAX
)
5290 seq_puts(m
, "max\n");
5292 seq_printf(m
, "%llu\n", (u64
)high
* PAGE_SIZE
);
5297 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
5298 char *buf
, size_t nbytes
, loff_t off
)
5300 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5304 buf
= strstrip(buf
);
5305 err
= page_counter_memparse(buf
, "max", &high
);
5314 static int memory_max_show(struct seq_file
*m
, void *v
)
5316 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5317 unsigned long max
= ACCESS_ONCE(memcg
->memory
.limit
);
5319 if (max
== PAGE_COUNTER_MAX
)
5320 seq_puts(m
, "max\n");
5322 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5327 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
5328 char *buf
, size_t nbytes
, loff_t off
)
5330 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5334 buf
= strstrip(buf
);
5335 err
= page_counter_memparse(buf
, "max", &max
);
5339 err
= mem_cgroup_resize_limit(memcg
, max
);
5346 static int memory_events_show(struct seq_file
*m
, void *v
)
5348 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5350 seq_printf(m
, "low %lu\n", mem_cgroup_read_events(memcg
, MEMCG_LOW
));
5351 seq_printf(m
, "high %lu\n", mem_cgroup_read_events(memcg
, MEMCG_HIGH
));
5352 seq_printf(m
, "max %lu\n", mem_cgroup_read_events(memcg
, MEMCG_MAX
));
5353 seq_printf(m
, "oom %lu\n", mem_cgroup_read_events(memcg
, MEMCG_OOM
));
5358 static struct cftype memory_files
[] = {
5361 .read_u64
= memory_current_read
,
5365 .flags
= CFTYPE_NOT_ON_ROOT
,
5366 .seq_show
= memory_low_show
,
5367 .write
= memory_low_write
,
5371 .flags
= CFTYPE_NOT_ON_ROOT
,
5372 .seq_show
= memory_high_show
,
5373 .write
= memory_high_write
,
5377 .flags
= CFTYPE_NOT_ON_ROOT
,
5378 .seq_show
= memory_max_show
,
5379 .write
= memory_max_write
,
5383 .flags
= CFTYPE_NOT_ON_ROOT
,
5384 .seq_show
= memory_events_show
,
5389 struct cgroup_subsys memory_cgrp_subsys
= {
5390 .css_alloc
= mem_cgroup_css_alloc
,
5391 .css_online
= mem_cgroup_css_online
,
5392 .css_offline
= mem_cgroup_css_offline
,
5393 .css_free
= mem_cgroup_css_free
,
5394 .css_reset
= mem_cgroup_css_reset
,
5395 .can_attach
= mem_cgroup_can_attach
,
5396 .cancel_attach
= mem_cgroup_cancel_attach
,
5397 .attach
= mem_cgroup_move_task
,
5398 .bind
= mem_cgroup_bind
,
5399 .dfl_cftypes
= memory_files
,
5400 .legacy_cftypes
= mem_cgroup_legacy_files
,
5405 * mem_cgroup_events - count memory events against a cgroup
5406 * @memcg: the memory cgroup
5407 * @idx: the event index
5408 * @nr: the number of events to account for
5410 void mem_cgroup_events(struct mem_cgroup
*memcg
,
5411 enum mem_cgroup_events_index idx
,
5414 this_cpu_add(memcg
->stat
->events
[idx
], nr
);
5418 * mem_cgroup_low - check if memory consumption is below the normal range
5419 * @root: the highest ancestor to consider
5420 * @memcg: the memory cgroup to check
5422 * Returns %true if memory consumption of @memcg, and that of all
5423 * configurable ancestors up to @root, is below the normal range.
5425 bool mem_cgroup_low(struct mem_cgroup
*root
, struct mem_cgroup
*memcg
)
5427 if (mem_cgroup_disabled())
5431 * The toplevel group doesn't have a configurable range, so
5432 * it's never low when looked at directly, and it is not
5433 * considered an ancestor when assessing the hierarchy.
5436 if (memcg
== root_mem_cgroup
)
5439 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5442 while (memcg
!= root
) {
5443 memcg
= parent_mem_cgroup(memcg
);
5445 if (memcg
== root_mem_cgroup
)
5448 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5455 * mem_cgroup_try_charge - try charging a page
5456 * @page: page to charge
5457 * @mm: mm context of the victim
5458 * @gfp_mask: reclaim mode
5459 * @memcgp: charged memcg return
5461 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5462 * pages according to @gfp_mask if necessary.
5464 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5465 * Otherwise, an error code is returned.
5467 * After page->mapping has been set up, the caller must finalize the
5468 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5469 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5471 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
5472 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
5474 struct mem_cgroup
*memcg
= NULL
;
5475 unsigned int nr_pages
= 1;
5478 if (mem_cgroup_disabled())
5481 if (PageSwapCache(page
)) {
5483 * Every swap fault against a single page tries to charge the
5484 * page, bail as early as possible. shmem_unuse() encounters
5485 * already charged pages, too. The USED bit is protected by
5486 * the page lock, which serializes swap cache removal, which
5487 * in turn serializes uncharging.
5489 if (page
->mem_cgroup
)
5493 if (PageTransHuge(page
)) {
5494 nr_pages
<<= compound_order(page
);
5495 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5498 if (do_swap_account
&& PageSwapCache(page
))
5499 memcg
= try_get_mem_cgroup_from_page(page
);
5501 memcg
= get_mem_cgroup_from_mm(mm
);
5503 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
5505 css_put(&memcg
->css
);
5507 if (ret
== -EINTR
) {
5508 memcg
= root_mem_cgroup
;
5517 * mem_cgroup_commit_charge - commit a page charge
5518 * @page: page to charge
5519 * @memcg: memcg to charge the page to
5520 * @lrucare: page might be on LRU already
5522 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5523 * after page->mapping has been set up. This must happen atomically
5524 * as part of the page instantiation, i.e. under the page table lock
5525 * for anonymous pages, under the page lock for page and swap cache.
5527 * In addition, the page must not be on the LRU during the commit, to
5528 * prevent racing with task migration. If it might be, use @lrucare.
5530 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5532 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5535 unsigned int nr_pages
= 1;
5537 VM_BUG_ON_PAGE(!page
->mapping
, page
);
5538 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
5540 if (mem_cgroup_disabled())
5543 * Swap faults will attempt to charge the same page multiple
5544 * times. But reuse_swap_page() might have removed the page
5545 * from swapcache already, so we can't check PageSwapCache().
5550 commit_charge(page
, memcg
, lrucare
);
5552 if (PageTransHuge(page
)) {
5553 nr_pages
<<= compound_order(page
);
5554 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5557 local_irq_disable();
5558 mem_cgroup_charge_statistics(memcg
, page
, nr_pages
);
5559 memcg_check_events(memcg
, page
);
5562 if (do_swap_account
&& PageSwapCache(page
)) {
5563 swp_entry_t entry
= { .val
= page_private(page
) };
5565 * The swap entry might not get freed for a long time,
5566 * let's not wait for it. The page already received a
5567 * memory+swap charge, drop the swap entry duplicate.
5569 mem_cgroup_uncharge_swap(entry
);
5574 * mem_cgroup_cancel_charge - cancel a page charge
5575 * @page: page to charge
5576 * @memcg: memcg to charge the page to
5578 * Cancel a charge transaction started by mem_cgroup_try_charge().
5580 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
)
5582 unsigned int nr_pages
= 1;
5584 if (mem_cgroup_disabled())
5587 * Swap faults will attempt to charge the same page multiple
5588 * times. But reuse_swap_page() might have removed the page
5589 * from swapcache already, so we can't check PageSwapCache().
5594 if (PageTransHuge(page
)) {
5595 nr_pages
<<= compound_order(page
);
5596 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5599 cancel_charge(memcg
, nr_pages
);
5602 static void uncharge_batch(struct mem_cgroup
*memcg
, unsigned long pgpgout
,
5603 unsigned long nr_anon
, unsigned long nr_file
,
5604 unsigned long nr_huge
, struct page
*dummy_page
)
5606 unsigned long nr_pages
= nr_anon
+ nr_file
;
5607 unsigned long flags
;
5609 if (!mem_cgroup_is_root(memcg
)) {
5610 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5611 if (do_swap_account
)
5612 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
5613 memcg_oom_recover(memcg
);
5616 local_irq_save(flags
);
5617 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_anon
);
5618 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_file
);
5619 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
], nr_huge
);
5620 __this_cpu_add(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
], pgpgout
);
5621 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
5622 memcg_check_events(memcg
, dummy_page
);
5623 local_irq_restore(flags
);
5625 if (!mem_cgroup_is_root(memcg
))
5626 css_put_many(&memcg
->css
, nr_pages
);
5629 static void uncharge_list(struct list_head
*page_list
)
5631 struct mem_cgroup
*memcg
= NULL
;
5632 unsigned long nr_anon
= 0;
5633 unsigned long nr_file
= 0;
5634 unsigned long nr_huge
= 0;
5635 unsigned long pgpgout
= 0;
5636 struct list_head
*next
;
5639 next
= page_list
->next
;
5641 unsigned int nr_pages
= 1;
5643 page
= list_entry(next
, struct page
, lru
);
5644 next
= page
->lru
.next
;
5646 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5647 VM_BUG_ON_PAGE(page_count(page
), page
);
5649 if (!page
->mem_cgroup
)
5653 * Nobody should be changing or seriously looking at
5654 * page->mem_cgroup at this point, we have fully
5655 * exclusive access to the page.
5658 if (memcg
!= page
->mem_cgroup
) {
5660 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5662 pgpgout
= nr_anon
= nr_file
= nr_huge
= 0;
5664 memcg
= page
->mem_cgroup
;
5667 if (PageTransHuge(page
)) {
5668 nr_pages
<<= compound_order(page
);
5669 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5670 nr_huge
+= nr_pages
;
5674 nr_anon
+= nr_pages
;
5676 nr_file
+= nr_pages
;
5678 page
->mem_cgroup
= NULL
;
5681 } while (next
!= page_list
);
5684 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5689 * mem_cgroup_uncharge - uncharge a page
5690 * @page: page to uncharge
5692 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5693 * mem_cgroup_commit_charge().
5695 void mem_cgroup_uncharge(struct page
*page
)
5697 if (mem_cgroup_disabled())
5700 /* Don't touch page->lru of any random page, pre-check: */
5701 if (!page
->mem_cgroup
)
5704 INIT_LIST_HEAD(&page
->lru
);
5705 uncharge_list(&page
->lru
);
5709 * mem_cgroup_uncharge_list - uncharge a list of page
5710 * @page_list: list of pages to uncharge
5712 * Uncharge a list of pages previously charged with
5713 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5715 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
5717 if (mem_cgroup_disabled())
5720 if (!list_empty(page_list
))
5721 uncharge_list(page_list
);
5725 * mem_cgroup_migrate - migrate a charge to another page
5726 * @oldpage: currently charged page
5727 * @newpage: page to transfer the charge to
5728 * @lrucare: either or both pages might be on the LRU already
5730 * Migrate the charge from @oldpage to @newpage.
5732 * Both pages must be locked, @newpage->mapping must be set up.
5734 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
,
5737 struct mem_cgroup
*memcg
;
5740 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
5741 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
5742 VM_BUG_ON_PAGE(!lrucare
&& PageLRU(oldpage
), oldpage
);
5743 VM_BUG_ON_PAGE(!lrucare
&& PageLRU(newpage
), newpage
);
5744 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
5745 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
5748 if (mem_cgroup_disabled())
5751 /* Page cache replacement: new page already charged? */
5752 if (newpage
->mem_cgroup
)
5756 * Swapcache readahead pages can get migrated before being
5757 * charged, and migration from compaction can happen to an
5758 * uncharged page when the PFN walker finds a page that
5759 * reclaim just put back on the LRU but has not released yet.
5761 memcg
= oldpage
->mem_cgroup
;
5766 lock_page_lru(oldpage
, &isolated
);
5768 oldpage
->mem_cgroup
= NULL
;
5771 unlock_page_lru(oldpage
, isolated
);
5773 commit_charge(newpage
, memcg
, lrucare
);
5777 * subsys_initcall() for memory controller.
5779 * Some parts like hotcpu_notifier() have to be initialized from this context
5780 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5781 * everything that doesn't depend on a specific mem_cgroup structure should
5782 * be initialized from here.
5784 static int __init
mem_cgroup_init(void)
5788 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
5790 for_each_possible_cpu(cpu
)
5791 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
5794 for_each_node(node
) {
5795 struct mem_cgroup_tree_per_node
*rtpn
;
5798 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
5799 node_online(node
) ? node
: NUMA_NO_NODE
);
5801 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5802 struct mem_cgroup_tree_per_zone
*rtpz
;
5804 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
5805 rtpz
->rb_root
= RB_ROOT
;
5806 spin_lock_init(&rtpz
->lock
);
5808 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
5813 subsys_initcall(mem_cgroup_init
);
5815 #ifdef CONFIG_MEMCG_SWAP
5817 * mem_cgroup_swapout - transfer a memsw charge to swap
5818 * @page: page whose memsw charge to transfer
5819 * @entry: swap entry to move the charge to
5821 * Transfer the memsw charge of @page to @entry.
5823 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
5825 struct mem_cgroup
*memcg
;
5826 unsigned short oldid
;
5828 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5829 VM_BUG_ON_PAGE(page_count(page
), page
);
5831 if (!do_swap_account
)
5834 memcg
= page
->mem_cgroup
;
5836 /* Readahead page, never charged */
5840 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5841 VM_BUG_ON_PAGE(oldid
, page
);
5842 mem_cgroup_swap_statistics(memcg
, true);
5844 page
->mem_cgroup
= NULL
;
5846 if (!mem_cgroup_is_root(memcg
))
5847 page_counter_uncharge(&memcg
->memory
, 1);
5849 /* XXX: caller holds IRQ-safe mapping->tree_lock */
5850 VM_BUG_ON(!irqs_disabled());
5852 mem_cgroup_charge_statistics(memcg
, page
, -1);
5853 memcg_check_events(memcg
, page
);
5857 * mem_cgroup_uncharge_swap - uncharge a swap entry
5858 * @entry: swap entry to uncharge
5860 * Drop the memsw charge associated with @entry.
5862 void mem_cgroup_uncharge_swap(swp_entry_t entry
)
5864 struct mem_cgroup
*memcg
;
5867 if (!do_swap_account
)
5870 id
= swap_cgroup_record(entry
, 0);
5872 memcg
= mem_cgroup_lookup(id
);
5874 if (!mem_cgroup_is_root(memcg
))
5875 page_counter_uncharge(&memcg
->memsw
, 1);
5876 mem_cgroup_swap_statistics(memcg
, false);
5877 css_put(&memcg
->css
);
5882 /* for remember boot option*/
5883 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5884 static int really_do_swap_account __initdata
= 1;
5886 static int really_do_swap_account __initdata
;
5889 static int __init
enable_swap_account(char *s
)
5891 if (!strcmp(s
, "1"))
5892 really_do_swap_account
= 1;
5893 else if (!strcmp(s
, "0"))
5894 really_do_swap_account
= 0;
5897 __setup("swapaccount=", enable_swap_account
);
5899 static struct cftype memsw_cgroup_files
[] = {
5901 .name
= "memsw.usage_in_bytes",
5902 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
5903 .read_u64
= mem_cgroup_read_u64
,
5906 .name
= "memsw.max_usage_in_bytes",
5907 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
5908 .write
= mem_cgroup_reset
,
5909 .read_u64
= mem_cgroup_read_u64
,
5912 .name
= "memsw.limit_in_bytes",
5913 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
5914 .write
= mem_cgroup_write
,
5915 .read_u64
= mem_cgroup_read_u64
,
5918 .name
= "memsw.failcnt",
5919 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
5920 .write
= mem_cgroup_reset
,
5921 .read_u64
= mem_cgroup_read_u64
,
5923 { }, /* terminate */
5926 static int __init
mem_cgroup_swap_init(void)
5928 if (!mem_cgroup_disabled() && really_do_swap_account
) {
5929 do_swap_account
= 1;
5930 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
5931 memsw_cgroup_files
));
5935 subsys_initcall(mem_cgroup_swap_init
);
5937 #endif /* CONFIG_MEMCG_SWAP */