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
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
28 #include <linux/res_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
32 #include <linux/hugetlb.h>
33 #include <linux/pagemap.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/page_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
62 #include <net/tcp_memcontrol.h>
65 #include <asm/uaccess.h>
67 #include <trace/events/vmscan.h>
69 struct cgroup_subsys memory_cgrp_subsys __read_mostly
;
70 EXPORT_SYMBOL(memory_cgrp_subsys
);
72 #define MEM_CGROUP_RECLAIM_RETRIES 5
73 static struct mem_cgroup
*root_mem_cgroup __read_mostly
;
75 #ifdef CONFIG_MEMCG_SWAP
76 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
77 int do_swap_account __read_mostly
;
79 /* for remember boot option*/
80 #ifdef CONFIG_MEMCG_SWAP_ENABLED
81 static int really_do_swap_account __initdata
= 1;
83 static int really_do_swap_account __initdata
;
87 #define do_swap_account 0
91 static const char * const mem_cgroup_stat_names
[] = {
100 enum mem_cgroup_events_index
{
101 MEM_CGROUP_EVENTS_PGPGIN
, /* # of pages paged in */
102 MEM_CGROUP_EVENTS_PGPGOUT
, /* # of pages paged out */
103 MEM_CGROUP_EVENTS_PGFAULT
, /* # of page-faults */
104 MEM_CGROUP_EVENTS_PGMAJFAULT
, /* # of major page-faults */
105 MEM_CGROUP_EVENTS_NSTATS
,
108 static const char * const mem_cgroup_events_names
[] = {
115 static const char * const mem_cgroup_lru_names
[] = {
124 * Per memcg event counter is incremented at every pagein/pageout. With THP,
125 * it will be incremated by the number of pages. This counter is used for
126 * for trigger some periodic events. This is straightforward and better
127 * than using jiffies etc. to handle periodic memcg event.
129 enum mem_cgroup_events_target
{
130 MEM_CGROUP_TARGET_THRESH
,
131 MEM_CGROUP_TARGET_SOFTLIMIT
,
132 MEM_CGROUP_TARGET_NUMAINFO
,
135 #define THRESHOLDS_EVENTS_TARGET 128
136 #define SOFTLIMIT_EVENTS_TARGET 1024
137 #define NUMAINFO_EVENTS_TARGET 1024
139 struct mem_cgroup_stat_cpu
{
140 long count
[MEM_CGROUP_STAT_NSTATS
];
141 unsigned long events
[MEM_CGROUP_EVENTS_NSTATS
];
142 unsigned long nr_page_events
;
143 unsigned long targets
[MEM_CGROUP_NTARGETS
];
146 struct mem_cgroup_reclaim_iter
{
148 * last scanned hierarchy member. Valid only if last_dead_count
149 * matches memcg->dead_count of the hierarchy root group.
151 struct mem_cgroup
*last_visited
;
154 /* scan generation, increased every round-trip */
155 unsigned int generation
;
159 * per-zone information in memory controller.
161 struct mem_cgroup_per_zone
{
162 struct lruvec lruvec
;
163 unsigned long lru_size
[NR_LRU_LISTS
];
165 struct mem_cgroup_reclaim_iter reclaim_iter
[DEF_PRIORITY
+ 1];
167 struct rb_node tree_node
; /* RB tree node */
168 unsigned long long usage_in_excess
;/* Set to the value by which */
169 /* the soft limit is exceeded*/
171 struct mem_cgroup
*memcg
; /* Back pointer, we cannot */
172 /* use container_of */
175 struct mem_cgroup_per_node
{
176 struct mem_cgroup_per_zone zoneinfo
[MAX_NR_ZONES
];
180 * Cgroups above their limits are maintained in a RB-Tree, independent of
181 * their hierarchy representation
184 struct mem_cgroup_tree_per_zone
{
185 struct rb_root rb_root
;
189 struct mem_cgroup_tree_per_node
{
190 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
193 struct mem_cgroup_tree
{
194 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
197 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
199 struct mem_cgroup_threshold
{
200 struct eventfd_ctx
*eventfd
;
205 struct mem_cgroup_threshold_ary
{
206 /* An array index points to threshold just below or equal to usage. */
207 int current_threshold
;
208 /* Size of entries[] */
210 /* Array of thresholds */
211 struct mem_cgroup_threshold entries
[0];
214 struct mem_cgroup_thresholds
{
215 /* Primary thresholds array */
216 struct mem_cgroup_threshold_ary
*primary
;
218 * Spare threshold array.
219 * This is needed to make mem_cgroup_unregister_event() "never fail".
220 * It must be able to store at least primary->size - 1 entries.
222 struct mem_cgroup_threshold_ary
*spare
;
226 struct mem_cgroup_eventfd_list
{
227 struct list_head list
;
228 struct eventfd_ctx
*eventfd
;
232 * cgroup_event represents events which userspace want to receive.
234 struct mem_cgroup_event
{
236 * memcg which the event belongs to.
238 struct mem_cgroup
*memcg
;
240 * eventfd to signal userspace about the event.
242 struct eventfd_ctx
*eventfd
;
244 * Each of these stored in a list by the cgroup.
246 struct list_head list
;
248 * register_event() callback will be used to add new userspace
249 * waiter for changes related to this event. Use eventfd_signal()
250 * on eventfd to send notification to userspace.
252 int (*register_event
)(struct mem_cgroup
*memcg
,
253 struct eventfd_ctx
*eventfd
, const char *args
);
255 * unregister_event() callback will be called when userspace closes
256 * the eventfd or on cgroup removing. This callback must be set,
257 * if you want provide notification functionality.
259 void (*unregister_event
)(struct mem_cgroup
*memcg
,
260 struct eventfd_ctx
*eventfd
);
262 * All fields below needed to unregister event when
263 * userspace closes eventfd.
266 wait_queue_head_t
*wqh
;
268 struct work_struct remove
;
271 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
272 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
275 * The memory controller data structure. The memory controller controls both
276 * page cache and RSS per cgroup. We would eventually like to provide
277 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
278 * to help the administrator determine what knobs to tune.
280 * TODO: Add a water mark for the memory controller. Reclaim will begin when
281 * we hit the water mark. May be even add a low water mark, such that
282 * no reclaim occurs from a cgroup at it's low water mark, this is
283 * a feature that will be implemented much later in the future.
286 struct cgroup_subsys_state css
;
288 * the counter to account for memory usage
290 struct res_counter res
;
292 /* vmpressure notifications */
293 struct vmpressure vmpressure
;
296 * the counter to account for mem+swap usage.
298 struct res_counter memsw
;
301 * the counter to account for kernel memory usage.
303 struct res_counter kmem
;
305 * Should the accounting and control be hierarchical, per subtree?
308 unsigned long kmem_account_flags
; /* See KMEM_ACCOUNTED_*, below */
312 atomic_t oom_wakeups
;
315 /* OOM-Killer disable */
316 int oom_kill_disable
;
318 /* set when res.limit == memsw.limit */
319 bool memsw_is_minimum
;
321 /* protect arrays of thresholds */
322 struct mutex thresholds_lock
;
324 /* thresholds for memory usage. RCU-protected */
325 struct mem_cgroup_thresholds thresholds
;
327 /* thresholds for mem+swap usage. RCU-protected */
328 struct mem_cgroup_thresholds memsw_thresholds
;
330 /* For oom notifier event fd */
331 struct list_head oom_notify
;
334 * Should we move charges of a task when a task is moved into this
335 * mem_cgroup ? And what type of charges should we move ?
337 unsigned long move_charge_at_immigrate
;
339 * set > 0 if pages under this cgroup are moving to other cgroup.
341 atomic_t moving_account
;
342 /* taken only while moving_account > 0 */
343 spinlock_t move_lock
;
347 struct mem_cgroup_stat_cpu __percpu
*stat
;
349 * used when a cpu is offlined or other synchronizations
350 * See mem_cgroup_read_stat().
352 struct mem_cgroup_stat_cpu nocpu_base
;
353 spinlock_t pcp_counter_lock
;
356 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
357 struct cg_proto tcp_mem
;
359 #if defined(CONFIG_MEMCG_KMEM)
360 /* analogous to slab_common's slab_caches list, but per-memcg;
361 * protected by memcg_slab_mutex */
362 struct list_head memcg_slab_caches
;
363 /* Index in the kmem_cache->memcg_params->memcg_caches array */
367 int last_scanned_node
;
369 nodemask_t scan_nodes
;
370 atomic_t numainfo_events
;
371 atomic_t numainfo_updating
;
374 /* List of events which userspace want to receive */
375 struct list_head event_list
;
376 spinlock_t event_list_lock
;
378 struct mem_cgroup_per_node
*nodeinfo
[0];
379 /* WARNING: nodeinfo must be the last member here */
382 /* internal only representation about the status of kmem accounting. */
384 KMEM_ACCOUNTED_ACTIVE
, /* accounted by this cgroup itself */
385 KMEM_ACCOUNTED_DEAD
, /* dead memcg with pending kmem charges */
388 #ifdef CONFIG_MEMCG_KMEM
389 static inline void memcg_kmem_set_active(struct mem_cgroup
*memcg
)
391 set_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
394 static bool memcg_kmem_is_active(struct mem_cgroup
*memcg
)
396 return test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
399 static void memcg_kmem_mark_dead(struct mem_cgroup
*memcg
)
402 * Our caller must use css_get() first, because memcg_uncharge_kmem()
403 * will call css_put() if it sees the memcg is dead.
406 if (test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
))
407 set_bit(KMEM_ACCOUNTED_DEAD
, &memcg
->kmem_account_flags
);
410 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup
*memcg
)
412 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD
,
413 &memcg
->kmem_account_flags
);
417 /* Stuffs for move charges at task migration. */
419 * Types of charges to be moved. "move_charge_at_immitgrate" and
420 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
423 MOVE_CHARGE_TYPE_ANON
, /* private anonymous page and swap of it */
424 MOVE_CHARGE_TYPE_FILE
, /* file page(including tmpfs) and swap of it */
428 /* "mc" and its members are protected by cgroup_mutex */
429 static struct move_charge_struct
{
430 spinlock_t lock
; /* for from, to */
431 struct mem_cgroup
*from
;
432 struct mem_cgroup
*to
;
433 unsigned long immigrate_flags
;
434 unsigned long precharge
;
435 unsigned long moved_charge
;
436 unsigned long moved_swap
;
437 struct task_struct
*moving_task
; /* a task moving charges */
438 wait_queue_head_t waitq
; /* a waitq for other context */
440 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
441 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
444 static bool move_anon(void)
446 return test_bit(MOVE_CHARGE_TYPE_ANON
, &mc
.immigrate_flags
);
449 static bool move_file(void)
451 return test_bit(MOVE_CHARGE_TYPE_FILE
, &mc
.immigrate_flags
);
455 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
456 * limit reclaim to prevent infinite loops, if they ever occur.
458 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
459 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
462 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
463 MEM_CGROUP_CHARGE_TYPE_ANON
,
464 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
465 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
469 /* for encoding cft->private value on file */
477 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
478 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
479 #define MEMFILE_ATTR(val) ((val) & 0xffff)
480 /* Used for OOM nofiier */
481 #define OOM_CONTROL (0)
484 * Reclaim flags for mem_cgroup_hierarchical_reclaim
486 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
487 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
488 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
489 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
492 * The memcg_create_mutex will be held whenever a new cgroup is created.
493 * As a consequence, any change that needs to protect against new child cgroups
494 * appearing has to hold it as well.
496 static DEFINE_MUTEX(memcg_create_mutex
);
498 struct mem_cgroup
*mem_cgroup_from_css(struct cgroup_subsys_state
*s
)
500 return s
? container_of(s
, struct mem_cgroup
, css
) : NULL
;
503 /* Some nice accessors for the vmpressure. */
504 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
507 memcg
= root_mem_cgroup
;
508 return &memcg
->vmpressure
;
511 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
513 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
516 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
518 return (memcg
== root_mem_cgroup
);
522 * We restrict the id in the range of [1, 65535], so it can fit into
525 #define MEM_CGROUP_ID_MAX USHRT_MAX
527 static inline unsigned short mem_cgroup_id(struct mem_cgroup
*memcg
)
530 * The ID of the root cgroup is 0, but memcg treat 0 as an
531 * invalid ID, so we return (cgroup_id + 1).
533 return memcg
->css
.cgroup
->id
+ 1;
536 static inline struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
538 struct cgroup_subsys_state
*css
;
540 css
= css_from_id(id
- 1, &memory_cgrp_subsys
);
541 return mem_cgroup_from_css(css
);
544 /* Writing them here to avoid exposing memcg's inner layout */
545 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
547 void sock_update_memcg(struct sock
*sk
)
549 if (mem_cgroup_sockets_enabled
) {
550 struct mem_cgroup
*memcg
;
551 struct cg_proto
*cg_proto
;
553 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
555 /* Socket cloning can throw us here with sk_cgrp already
556 * filled. It won't however, necessarily happen from
557 * process context. So the test for root memcg given
558 * the current task's memcg won't help us in this case.
560 * Respecting the original socket's memcg is a better
561 * decision in this case.
564 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
565 css_get(&sk
->sk_cgrp
->memcg
->css
);
570 memcg
= mem_cgroup_from_task(current
);
571 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
572 if (!mem_cgroup_is_root(memcg
) &&
573 memcg_proto_active(cg_proto
) && css_tryget(&memcg
->css
)) {
574 sk
->sk_cgrp
= cg_proto
;
579 EXPORT_SYMBOL(sock_update_memcg
);
581 void sock_release_memcg(struct sock
*sk
)
583 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
584 struct mem_cgroup
*memcg
;
585 WARN_ON(!sk
->sk_cgrp
->memcg
);
586 memcg
= sk
->sk_cgrp
->memcg
;
587 css_put(&sk
->sk_cgrp
->memcg
->css
);
591 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
593 if (!memcg
|| mem_cgroup_is_root(memcg
))
596 return &memcg
->tcp_mem
;
598 EXPORT_SYMBOL(tcp_proto_cgroup
);
600 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
602 if (!memcg_proto_activated(&memcg
->tcp_mem
))
604 static_key_slow_dec(&memcg_socket_limit_enabled
);
607 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
612 #ifdef CONFIG_MEMCG_KMEM
614 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
615 * The main reason for not using cgroup id for this:
616 * this works better in sparse environments, where we have a lot of memcgs,
617 * but only a few kmem-limited. Or also, if we have, for instance, 200
618 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
619 * 200 entry array for that.
621 * The current size of the caches array is stored in
622 * memcg_limited_groups_array_size. It will double each time we have to
625 static DEFINE_IDA(kmem_limited_groups
);
626 int memcg_limited_groups_array_size
;
629 * MIN_SIZE is different than 1, because we would like to avoid going through
630 * the alloc/free process all the time. In a small machine, 4 kmem-limited
631 * cgroups is a reasonable guess. In the future, it could be a parameter or
632 * tunable, but that is strictly not necessary.
634 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
635 * this constant directly from cgroup, but it is understandable that this is
636 * better kept as an internal representation in cgroup.c. In any case, the
637 * cgrp_id space is not getting any smaller, and we don't have to necessarily
638 * increase ours as well if it increases.
640 #define MEMCG_CACHES_MIN_SIZE 4
641 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
644 * A lot of the calls to the cache allocation functions are expected to be
645 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
646 * conditional to this static branch, we'll have to allow modules that does
647 * kmem_cache_alloc and the such to see this symbol as well
649 struct static_key memcg_kmem_enabled_key
;
650 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
652 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
654 if (memcg_kmem_is_active(memcg
)) {
655 static_key_slow_dec(&memcg_kmem_enabled_key
);
656 ida_simple_remove(&kmem_limited_groups
, memcg
->kmemcg_id
);
659 * This check can't live in kmem destruction function,
660 * since the charges will outlive the cgroup
662 WARN_ON(res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0);
665 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
668 #endif /* CONFIG_MEMCG_KMEM */
670 static void disarm_static_keys(struct mem_cgroup
*memcg
)
672 disarm_sock_keys(memcg
);
673 disarm_kmem_keys(memcg
);
676 static void drain_all_stock_async(struct mem_cgroup
*memcg
);
678 static struct mem_cgroup_per_zone
*
679 mem_cgroup_zone_zoneinfo(struct mem_cgroup
*memcg
, struct zone
*zone
)
681 int nid
= zone_to_nid(zone
);
682 int zid
= zone_idx(zone
);
684 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
687 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
692 static struct mem_cgroup_per_zone
*
693 mem_cgroup_page_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
695 int nid
= page_to_nid(page
);
696 int zid
= page_zonenum(page
);
698 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
701 static struct mem_cgroup_tree_per_zone
*
702 soft_limit_tree_node_zone(int nid
, int zid
)
704 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
707 static struct mem_cgroup_tree_per_zone
*
708 soft_limit_tree_from_page(struct page
*page
)
710 int nid
= page_to_nid(page
);
711 int zid
= page_zonenum(page
);
713 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
716 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone
*mz
,
717 struct mem_cgroup_tree_per_zone
*mctz
,
718 unsigned long long new_usage_in_excess
)
720 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
721 struct rb_node
*parent
= NULL
;
722 struct mem_cgroup_per_zone
*mz_node
;
727 mz
->usage_in_excess
= new_usage_in_excess
;
728 if (!mz
->usage_in_excess
)
732 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
734 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
737 * We can't avoid mem cgroups that are over their soft
738 * limit by the same amount
740 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
743 rb_link_node(&mz
->tree_node
, parent
, p
);
744 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
748 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
749 struct mem_cgroup_tree_per_zone
*mctz
)
753 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
757 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
758 struct mem_cgroup_tree_per_zone
*mctz
)
760 spin_lock(&mctz
->lock
);
761 __mem_cgroup_remove_exceeded(mz
, mctz
);
762 spin_unlock(&mctz
->lock
);
766 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
768 unsigned long long excess
;
769 struct mem_cgroup_per_zone
*mz
;
770 struct mem_cgroup_tree_per_zone
*mctz
;
772 mctz
= soft_limit_tree_from_page(page
);
774 * Necessary to update all ancestors when hierarchy is used.
775 * because their event counter is not touched.
777 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
778 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
779 excess
= res_counter_soft_limit_excess(&memcg
->res
);
781 * We have to update the tree if mz is on RB-tree or
782 * mem is over its softlimit.
784 if (excess
|| mz
->on_tree
) {
785 spin_lock(&mctz
->lock
);
786 /* if on-tree, remove it */
788 __mem_cgroup_remove_exceeded(mz
, mctz
);
790 * Insert again. mz->usage_in_excess will be updated.
791 * If excess is 0, no tree ops.
793 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
794 spin_unlock(&mctz
->lock
);
799 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
801 struct mem_cgroup_tree_per_zone
*mctz
;
802 struct mem_cgroup_per_zone
*mz
;
806 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
807 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
808 mctz
= soft_limit_tree_node_zone(nid
, zid
);
809 mem_cgroup_remove_exceeded(mz
, mctz
);
814 static struct mem_cgroup_per_zone
*
815 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
817 struct rb_node
*rightmost
= NULL
;
818 struct mem_cgroup_per_zone
*mz
;
822 rightmost
= rb_last(&mctz
->rb_root
);
824 goto done
; /* Nothing to reclaim from */
826 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
828 * Remove the node now but someone else can add it back,
829 * we will to add it back at the end of reclaim to its correct
830 * position in the tree.
832 __mem_cgroup_remove_exceeded(mz
, mctz
);
833 if (!res_counter_soft_limit_excess(&mz
->memcg
->res
) ||
834 !css_tryget(&mz
->memcg
->css
))
840 static struct mem_cgroup_per_zone
*
841 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
843 struct mem_cgroup_per_zone
*mz
;
845 spin_lock(&mctz
->lock
);
846 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
847 spin_unlock(&mctz
->lock
);
852 * Implementation Note: reading percpu statistics for memcg.
854 * Both of vmstat[] and percpu_counter has threshold and do periodic
855 * synchronization to implement "quick" read. There are trade-off between
856 * reading cost and precision of value. Then, we may have a chance to implement
857 * a periodic synchronizion of counter in memcg's counter.
859 * But this _read() function is used for user interface now. The user accounts
860 * memory usage by memory cgroup and he _always_ requires exact value because
861 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
862 * have to visit all online cpus and make sum. So, for now, unnecessary
863 * synchronization is not implemented. (just implemented for cpu hotplug)
865 * If there are kernel internal actions which can make use of some not-exact
866 * value, and reading all cpu value can be performance bottleneck in some
867 * common workload, threashold and synchonization as vmstat[] should be
870 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
871 enum mem_cgroup_stat_index idx
)
877 for_each_online_cpu(cpu
)
878 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
879 #ifdef CONFIG_HOTPLUG_CPU
880 spin_lock(&memcg
->pcp_counter_lock
);
881 val
+= memcg
->nocpu_base
.count
[idx
];
882 spin_unlock(&memcg
->pcp_counter_lock
);
888 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
891 int val
= (charge
) ? 1 : -1;
892 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
895 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
896 enum mem_cgroup_events_index idx
)
898 unsigned long val
= 0;
902 for_each_online_cpu(cpu
)
903 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
904 #ifdef CONFIG_HOTPLUG_CPU
905 spin_lock(&memcg
->pcp_counter_lock
);
906 val
+= memcg
->nocpu_base
.events
[idx
];
907 spin_unlock(&memcg
->pcp_counter_lock
);
913 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
915 bool anon
, int nr_pages
)
918 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
919 * counted as CACHE even if it's on ANON LRU.
922 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
925 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
928 if (PageTransHuge(page
))
929 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
932 /* pagein of a big page is an event. So, ignore page size */
934 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
936 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
937 nr_pages
= -nr_pages
; /* for event */
940 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
943 unsigned long mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
945 struct mem_cgroup_per_zone
*mz
;
947 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
948 return mz
->lru_size
[lru
];
951 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
953 unsigned int lru_mask
)
955 unsigned long nr
= 0;
958 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
960 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
961 struct mem_cgroup_per_zone
*mz
;
965 if (!(BIT(lru
) & lru_mask
))
967 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
968 nr
+= mz
->lru_size
[lru
];
974 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
975 unsigned int lru_mask
)
977 unsigned long nr
= 0;
980 for_each_node_state(nid
, N_MEMORY
)
981 nr
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
985 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
986 enum mem_cgroup_events_target target
)
988 unsigned long val
, next
;
990 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
991 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
992 /* from time_after() in jiffies.h */
993 if ((long)next
- (long)val
< 0) {
995 case MEM_CGROUP_TARGET_THRESH
:
996 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
998 case MEM_CGROUP_TARGET_SOFTLIMIT
:
999 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
1001 case MEM_CGROUP_TARGET_NUMAINFO
:
1002 next
= val
+ NUMAINFO_EVENTS_TARGET
;
1007 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
1014 * Check events in order.
1017 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
1020 /* threshold event is triggered in finer grain than soft limit */
1021 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
1022 MEM_CGROUP_TARGET_THRESH
))) {
1024 bool do_numainfo __maybe_unused
;
1026 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
1027 MEM_CGROUP_TARGET_SOFTLIMIT
);
1028 #if MAX_NUMNODES > 1
1029 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
1030 MEM_CGROUP_TARGET_NUMAINFO
);
1034 mem_cgroup_threshold(memcg
);
1035 if (unlikely(do_softlimit
))
1036 mem_cgroup_update_tree(memcg
, page
);
1037 #if MAX_NUMNODES > 1
1038 if (unlikely(do_numainfo
))
1039 atomic_inc(&memcg
->numainfo_events
);
1045 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
1048 * mm_update_next_owner() may clear mm->owner to NULL
1049 * if it races with swapoff, page migration, etc.
1050 * So this can be called with p == NULL.
1055 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
1058 static struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
1060 struct mem_cgroup
*memcg
= NULL
;
1065 * Page cache insertions can happen withou an
1066 * actual mm context, e.g. during disk probing
1067 * on boot, loopback IO, acct() writes etc.
1070 memcg
= root_mem_cgroup
;
1072 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1073 if (unlikely(!memcg
))
1074 memcg
= root_mem_cgroup
;
1076 } while (!css_tryget(&memcg
->css
));
1082 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1083 * ref. count) or NULL if the whole root's subtree has been visited.
1085 * helper function to be used by mem_cgroup_iter
1087 static struct mem_cgroup
*__mem_cgroup_iter_next(struct mem_cgroup
*root
,
1088 struct mem_cgroup
*last_visited
)
1090 struct cgroup_subsys_state
*prev_css
, *next_css
;
1092 prev_css
= last_visited
? &last_visited
->css
: NULL
;
1094 next_css
= css_next_descendant_pre(prev_css
, &root
->css
);
1097 * Even if we found a group we have to make sure it is
1098 * alive. css && !memcg means that the groups should be
1099 * skipped and we should continue the tree walk.
1100 * last_visited css is safe to use because it is
1101 * protected by css_get and the tree walk is rcu safe.
1103 * We do not take a reference on the root of the tree walk
1104 * because we might race with the root removal when it would
1105 * be the only node in the iterated hierarchy and mem_cgroup_iter
1106 * would end up in an endless loop because it expects that at
1107 * least one valid node will be returned. Root cannot disappear
1108 * because caller of the iterator should hold it already so
1109 * skipping css reference should be safe.
1112 if ((next_css
== &root
->css
) ||
1113 ((next_css
->flags
& CSS_ONLINE
) && css_tryget(next_css
)))
1114 return mem_cgroup_from_css(next_css
);
1116 prev_css
= next_css
;
1123 static void mem_cgroup_iter_invalidate(struct mem_cgroup
*root
)
1126 * When a group in the hierarchy below root is destroyed, the
1127 * hierarchy iterator can no longer be trusted since it might
1128 * have pointed to the destroyed group. Invalidate it.
1130 atomic_inc(&root
->dead_count
);
1133 static struct mem_cgroup
*
1134 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter
*iter
,
1135 struct mem_cgroup
*root
,
1138 struct mem_cgroup
*position
= NULL
;
1140 * A cgroup destruction happens in two stages: offlining and
1141 * release. They are separated by a RCU grace period.
1143 * If the iterator is valid, we may still race with an
1144 * offlining. The RCU lock ensures the object won't be
1145 * released, tryget will fail if we lost the race.
1147 *sequence
= atomic_read(&root
->dead_count
);
1148 if (iter
->last_dead_count
== *sequence
) {
1150 position
= iter
->last_visited
;
1153 * We cannot take a reference to root because we might race
1154 * with root removal and returning NULL would end up in
1155 * an endless loop on the iterator user level when root
1156 * would be returned all the time.
1158 if (position
&& position
!= root
&&
1159 !css_tryget(&position
->css
))
1165 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter
*iter
,
1166 struct mem_cgroup
*last_visited
,
1167 struct mem_cgroup
*new_position
,
1168 struct mem_cgroup
*root
,
1171 /* root reference counting symmetric to mem_cgroup_iter_load */
1172 if (last_visited
&& last_visited
!= root
)
1173 css_put(&last_visited
->css
);
1175 * We store the sequence count from the time @last_visited was
1176 * loaded successfully instead of rereading it here so that we
1177 * don't lose destruction events in between. We could have
1178 * raced with the destruction of @new_position after all.
1180 iter
->last_visited
= new_position
;
1182 iter
->last_dead_count
= sequence
;
1186 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1187 * @root: hierarchy root
1188 * @prev: previously returned memcg, NULL on first invocation
1189 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1191 * Returns references to children of the hierarchy below @root, or
1192 * @root itself, or %NULL after a full round-trip.
1194 * Caller must pass the return value in @prev on subsequent
1195 * invocations for reference counting, or use mem_cgroup_iter_break()
1196 * to cancel a hierarchy walk before the round-trip is complete.
1198 * Reclaimers can specify a zone and a priority level in @reclaim to
1199 * divide up the memcgs in the hierarchy among all concurrent
1200 * reclaimers operating on the same zone and priority.
1202 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1203 struct mem_cgroup
*prev
,
1204 struct mem_cgroup_reclaim_cookie
*reclaim
)
1206 struct mem_cgroup
*memcg
= NULL
;
1207 struct mem_cgroup
*last_visited
= NULL
;
1209 if (mem_cgroup_disabled())
1213 root
= root_mem_cgroup
;
1215 if (prev
&& !reclaim
)
1216 last_visited
= prev
;
1218 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1226 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1227 int uninitialized_var(seq
);
1230 struct mem_cgroup_per_zone
*mz
;
1232 mz
= mem_cgroup_zone_zoneinfo(root
, reclaim
->zone
);
1233 iter
= &mz
->reclaim_iter
[reclaim
->priority
];
1234 if (prev
&& reclaim
->generation
!= iter
->generation
) {
1235 iter
->last_visited
= NULL
;
1239 last_visited
= mem_cgroup_iter_load(iter
, root
, &seq
);
1242 memcg
= __mem_cgroup_iter_next(root
, last_visited
);
1245 mem_cgroup_iter_update(iter
, last_visited
, memcg
, root
,
1250 else if (!prev
&& memcg
)
1251 reclaim
->generation
= iter
->generation
;
1260 if (prev
&& prev
!= root
)
1261 css_put(&prev
->css
);
1267 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1268 * @root: hierarchy root
1269 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1271 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1272 struct mem_cgroup
*prev
)
1275 root
= root_mem_cgroup
;
1276 if (prev
&& prev
!= root
)
1277 css_put(&prev
->css
);
1281 * Iteration constructs for visiting all cgroups (under a tree). If
1282 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1283 * be used for reference counting.
1285 #define for_each_mem_cgroup_tree(iter, root) \
1286 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1288 iter = mem_cgroup_iter(root, iter, NULL))
1290 #define for_each_mem_cgroup(iter) \
1291 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1293 iter = mem_cgroup_iter(NULL, iter, NULL))
1295 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1297 struct mem_cgroup
*memcg
;
1300 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1301 if (unlikely(!memcg
))
1306 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1309 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1317 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1320 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1321 * @zone: zone of the wanted lruvec
1322 * @memcg: memcg of the wanted lruvec
1324 * Returns the lru list vector holding pages for the given @zone and
1325 * @mem. This can be the global zone lruvec, if the memory controller
1328 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1329 struct mem_cgroup
*memcg
)
1331 struct mem_cgroup_per_zone
*mz
;
1332 struct lruvec
*lruvec
;
1334 if (mem_cgroup_disabled()) {
1335 lruvec
= &zone
->lruvec
;
1339 mz
= mem_cgroup_zone_zoneinfo(memcg
, zone
);
1340 lruvec
= &mz
->lruvec
;
1343 * Since a node can be onlined after the mem_cgroup was created,
1344 * we have to be prepared to initialize lruvec->zone here;
1345 * and if offlined then reonlined, we need to reinitialize it.
1347 if (unlikely(lruvec
->zone
!= zone
))
1348 lruvec
->zone
= zone
;
1353 * Following LRU functions are allowed to be used without PCG_LOCK.
1354 * Operations are called by routine of global LRU independently from memcg.
1355 * What we have to take care of here is validness of pc->mem_cgroup.
1357 * Changes to pc->mem_cgroup happens when
1360 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1361 * It is added to LRU before charge.
1362 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1363 * When moving account, the page is not on LRU. It's isolated.
1367 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1369 * @zone: zone of the page
1371 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1373 struct mem_cgroup_per_zone
*mz
;
1374 struct mem_cgroup
*memcg
;
1375 struct page_cgroup
*pc
;
1376 struct lruvec
*lruvec
;
1378 if (mem_cgroup_disabled()) {
1379 lruvec
= &zone
->lruvec
;
1383 pc
= lookup_page_cgroup(page
);
1384 memcg
= pc
->mem_cgroup
;
1387 * Surreptitiously switch any uncharged offlist page to root:
1388 * an uncharged page off lru does nothing to secure
1389 * its former mem_cgroup from sudden removal.
1391 * Our caller holds lru_lock, and PageCgroupUsed is updated
1392 * under page_cgroup lock: between them, they make all uses
1393 * of pc->mem_cgroup safe.
1395 if (!PageLRU(page
) && !PageCgroupUsed(pc
) && memcg
!= root_mem_cgroup
)
1396 pc
->mem_cgroup
= memcg
= root_mem_cgroup
;
1398 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
1399 lruvec
= &mz
->lruvec
;
1402 * Since a node can be onlined after the mem_cgroup was created,
1403 * we have to be prepared to initialize lruvec->zone here;
1404 * and if offlined then reonlined, we need to reinitialize it.
1406 if (unlikely(lruvec
->zone
!= zone
))
1407 lruvec
->zone
= zone
;
1412 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1413 * @lruvec: mem_cgroup per zone lru vector
1414 * @lru: index of lru list the page is sitting on
1415 * @nr_pages: positive when adding or negative when removing
1417 * This function must be called when a page is added to or removed from an
1420 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1423 struct mem_cgroup_per_zone
*mz
;
1424 unsigned long *lru_size
;
1426 if (mem_cgroup_disabled())
1429 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1430 lru_size
= mz
->lru_size
+ lru
;
1431 *lru_size
+= nr_pages
;
1432 VM_BUG_ON((long)(*lru_size
) < 0);
1436 * Checks whether given mem is same or in the root_mem_cgroup's
1439 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1440 struct mem_cgroup
*memcg
)
1442 if (root_memcg
== memcg
)
1444 if (!root_memcg
->use_hierarchy
|| !memcg
)
1446 return cgroup_is_descendant(memcg
->css
.cgroup
, root_memcg
->css
.cgroup
);
1449 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1450 struct mem_cgroup
*memcg
)
1455 ret
= __mem_cgroup_same_or_subtree(root_memcg
, memcg
);
1460 bool task_in_mem_cgroup(struct task_struct
*task
,
1461 const struct mem_cgroup
*memcg
)
1463 struct mem_cgroup
*curr
= NULL
;
1464 struct task_struct
*p
;
1467 p
= find_lock_task_mm(task
);
1469 curr
= get_mem_cgroup_from_mm(p
->mm
);
1473 * All threads may have already detached their mm's, but the oom
1474 * killer still needs to detect if they have already been oom
1475 * killed to prevent needlessly killing additional tasks.
1478 curr
= mem_cgroup_from_task(task
);
1480 css_get(&curr
->css
);
1484 * We should check use_hierarchy of "memcg" not "curr". Because checking
1485 * use_hierarchy of "curr" here make this function true if hierarchy is
1486 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1487 * hierarchy(even if use_hierarchy is disabled in "memcg").
1489 ret
= mem_cgroup_same_or_subtree(memcg
, curr
);
1490 css_put(&curr
->css
);
1494 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1496 unsigned long inactive_ratio
;
1497 unsigned long inactive
;
1498 unsigned long active
;
1501 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1502 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1504 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1506 inactive_ratio
= int_sqrt(10 * gb
);
1510 return inactive
* inactive_ratio
< active
;
1513 #define mem_cgroup_from_res_counter(counter, member) \
1514 container_of(counter, struct mem_cgroup, member)
1517 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1518 * @memcg: the memory cgroup
1520 * Returns the maximum amount of memory @mem can be charged with, in
1523 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1525 unsigned long long margin
;
1527 margin
= res_counter_margin(&memcg
->res
);
1528 if (do_swap_account
)
1529 margin
= min(margin
, res_counter_margin(&memcg
->memsw
));
1530 return margin
>> PAGE_SHIFT
;
1533 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1536 if (mem_cgroup_disabled() || !css_parent(&memcg
->css
))
1537 return vm_swappiness
;
1539 return memcg
->swappiness
;
1543 * memcg->moving_account is used for checking possibility that some thread is
1544 * calling move_account(). When a thread on CPU-A starts moving pages under
1545 * a memcg, other threads should check memcg->moving_account under
1546 * rcu_read_lock(), like this:
1550 * memcg->moving_account+1 if (memcg->mocing_account)
1552 * synchronize_rcu() update something.
1557 /* for quick checking without looking up memcg */
1558 atomic_t memcg_moving __read_mostly
;
1560 static void mem_cgroup_start_move(struct mem_cgroup
*memcg
)
1562 atomic_inc(&memcg_moving
);
1563 atomic_inc(&memcg
->moving_account
);
1567 static void mem_cgroup_end_move(struct mem_cgroup
*memcg
)
1570 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1571 * We check NULL in callee rather than caller.
1574 atomic_dec(&memcg_moving
);
1575 atomic_dec(&memcg
->moving_account
);
1580 * A routine for checking "mem" is under move_account() or not.
1582 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1583 * moving cgroups. This is for waiting at high-memory pressure
1586 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1588 struct mem_cgroup
*from
;
1589 struct mem_cgroup
*to
;
1592 * Unlike task_move routines, we access mc.to, mc.from not under
1593 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1595 spin_lock(&mc
.lock
);
1601 ret
= mem_cgroup_same_or_subtree(memcg
, from
)
1602 || mem_cgroup_same_or_subtree(memcg
, to
);
1604 spin_unlock(&mc
.lock
);
1608 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1610 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1611 if (mem_cgroup_under_move(memcg
)) {
1613 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1614 /* moving charge context might have finished. */
1617 finish_wait(&mc
.waitq
, &wait
);
1625 * Take this lock when
1626 * - a code tries to modify page's memcg while it's USED.
1627 * - a code tries to modify page state accounting in a memcg.
1629 static void move_lock_mem_cgroup(struct mem_cgroup
*memcg
,
1630 unsigned long *flags
)
1632 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
1635 static void move_unlock_mem_cgroup(struct mem_cgroup
*memcg
,
1636 unsigned long *flags
)
1638 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
1641 #define K(x) ((x) << (PAGE_SHIFT-10))
1643 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1644 * @memcg: The memory cgroup that went over limit
1645 * @p: Task that is going to be killed
1647 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1650 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1652 /* oom_info_lock ensures that parallel ooms do not interleave */
1653 static DEFINE_MUTEX(oom_info_lock
);
1654 struct mem_cgroup
*iter
;
1660 mutex_lock(&oom_info_lock
);
1663 pr_info("Task in ");
1664 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1665 pr_info(" killed as a result of limit of ");
1666 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1671 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1672 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1673 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1674 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1675 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1676 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1677 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1678 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1679 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1680 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) >> 10,
1681 res_counter_read_u64(&memcg
->kmem
, RES_LIMIT
) >> 10,
1682 res_counter_read_u64(&memcg
->kmem
, RES_FAILCNT
));
1684 for_each_mem_cgroup_tree(iter
, memcg
) {
1685 pr_info("Memory cgroup stats for ");
1686 pr_cont_cgroup_path(iter
->css
.cgroup
);
1689 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1690 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1692 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1693 K(mem_cgroup_read_stat(iter
, i
)));
1696 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1697 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1698 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1702 mutex_unlock(&oom_info_lock
);
1706 * This function returns the number of memcg under hierarchy tree. Returns
1707 * 1(self count) if no children.
1709 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1712 struct mem_cgroup
*iter
;
1714 for_each_mem_cgroup_tree(iter
, memcg
)
1720 * Return the memory (and swap, if configured) limit for a memcg.
1722 static u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1726 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1729 * Do not consider swap space if we cannot swap due to swappiness
1731 if (mem_cgroup_swappiness(memcg
)) {
1734 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1735 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1738 * If memsw is finite and limits the amount of swap space
1739 * available to this memcg, return that limit.
1741 limit
= min(limit
, memsw
);
1747 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1750 struct mem_cgroup
*iter
;
1751 unsigned long chosen_points
= 0;
1752 unsigned long totalpages
;
1753 unsigned int points
= 0;
1754 struct task_struct
*chosen
= NULL
;
1757 * If current has a pending SIGKILL or is exiting, then automatically
1758 * select it. The goal is to allow it to allocate so that it may
1759 * quickly exit and free its memory.
1761 if (fatal_signal_pending(current
) || current
->flags
& PF_EXITING
) {
1762 set_thread_flag(TIF_MEMDIE
);
1766 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1767 totalpages
= mem_cgroup_get_limit(memcg
) >> PAGE_SHIFT
? : 1;
1768 for_each_mem_cgroup_tree(iter
, memcg
) {
1769 struct css_task_iter it
;
1770 struct task_struct
*task
;
1772 css_task_iter_start(&iter
->css
, &it
);
1773 while ((task
= css_task_iter_next(&it
))) {
1774 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1776 case OOM_SCAN_SELECT
:
1778 put_task_struct(chosen
);
1780 chosen_points
= ULONG_MAX
;
1781 get_task_struct(chosen
);
1783 case OOM_SCAN_CONTINUE
:
1785 case OOM_SCAN_ABORT
:
1786 css_task_iter_end(&it
);
1787 mem_cgroup_iter_break(memcg
, iter
);
1789 put_task_struct(chosen
);
1794 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1795 if (!points
|| points
< chosen_points
)
1797 /* Prefer thread group leaders for display purposes */
1798 if (points
== chosen_points
&&
1799 thread_group_leader(chosen
))
1803 put_task_struct(chosen
);
1805 chosen_points
= points
;
1806 get_task_struct(chosen
);
1808 css_task_iter_end(&it
);
1813 points
= chosen_points
* 1000 / totalpages
;
1814 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1815 NULL
, "Memory cgroup out of memory");
1818 static unsigned long mem_cgroup_reclaim(struct mem_cgroup
*memcg
,
1820 unsigned long flags
)
1822 unsigned long total
= 0;
1823 bool noswap
= false;
1826 if (flags
& MEM_CGROUP_RECLAIM_NOSWAP
)
1828 if (!(flags
& MEM_CGROUP_RECLAIM_SHRINK
) && memcg
->memsw_is_minimum
)
1831 for (loop
= 0; loop
< MEM_CGROUP_MAX_RECLAIM_LOOPS
; loop
++) {
1833 drain_all_stock_async(memcg
);
1834 total
+= try_to_free_mem_cgroup_pages(memcg
, gfp_mask
, noswap
);
1836 * Allow limit shrinkers, which are triggered directly
1837 * by userspace, to catch signals and stop reclaim
1838 * after minimal progress, regardless of the margin.
1840 if (total
&& (flags
& MEM_CGROUP_RECLAIM_SHRINK
))
1842 if (mem_cgroup_margin(memcg
))
1845 * If nothing was reclaimed after two attempts, there
1846 * may be no reclaimable pages in this hierarchy.
1855 * test_mem_cgroup_node_reclaimable
1856 * @memcg: the target memcg
1857 * @nid: the node ID to be checked.
1858 * @noswap : specify true here if the user wants flle only information.
1860 * This function returns whether the specified memcg contains any
1861 * reclaimable pages on a node. Returns true if there are any reclaimable
1862 * pages in the node.
1864 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1865 int nid
, bool noswap
)
1867 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1869 if (noswap
|| !total_swap_pages
)
1871 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1876 #if MAX_NUMNODES > 1
1879 * Always updating the nodemask is not very good - even if we have an empty
1880 * list or the wrong list here, we can start from some node and traverse all
1881 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1884 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1888 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1889 * pagein/pageout changes since the last update.
1891 if (!atomic_read(&memcg
->numainfo_events
))
1893 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1896 /* make a nodemask where this memcg uses memory from */
1897 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1899 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1901 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1902 node_clear(nid
, memcg
->scan_nodes
);
1905 atomic_set(&memcg
->numainfo_events
, 0);
1906 atomic_set(&memcg
->numainfo_updating
, 0);
1910 * Selecting a node where we start reclaim from. Because what we need is just
1911 * reducing usage counter, start from anywhere is O,K. Considering
1912 * memory reclaim from current node, there are pros. and cons.
1914 * Freeing memory from current node means freeing memory from a node which
1915 * we'll use or we've used. So, it may make LRU bad. And if several threads
1916 * hit limits, it will see a contention on a node. But freeing from remote
1917 * node means more costs for memory reclaim because of memory latency.
1919 * Now, we use round-robin. Better algorithm is welcomed.
1921 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1925 mem_cgroup_may_update_nodemask(memcg
);
1926 node
= memcg
->last_scanned_node
;
1928 node
= next_node(node
, memcg
->scan_nodes
);
1929 if (node
== MAX_NUMNODES
)
1930 node
= first_node(memcg
->scan_nodes
);
1932 * We call this when we hit limit, not when pages are added to LRU.
1933 * No LRU may hold pages because all pages are UNEVICTABLE or
1934 * memcg is too small and all pages are not on LRU. In that case,
1935 * we use curret node.
1937 if (unlikely(node
== MAX_NUMNODES
))
1938 node
= numa_node_id();
1940 memcg
->last_scanned_node
= node
;
1945 * Check all nodes whether it contains reclaimable pages or not.
1946 * For quick scan, we make use of scan_nodes. This will allow us to skip
1947 * unused nodes. But scan_nodes is lazily updated and may not cotain
1948 * enough new information. We need to do double check.
1950 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1955 * quick check...making use of scan_node.
1956 * We can skip unused nodes.
1958 if (!nodes_empty(memcg
->scan_nodes
)) {
1959 for (nid
= first_node(memcg
->scan_nodes
);
1961 nid
= next_node(nid
, memcg
->scan_nodes
)) {
1963 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1968 * Check rest of nodes.
1970 for_each_node_state(nid
, N_MEMORY
) {
1971 if (node_isset(nid
, memcg
->scan_nodes
))
1973 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1980 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1985 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1987 return test_mem_cgroup_node_reclaimable(memcg
, 0, noswap
);
1991 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1994 unsigned long *total_scanned
)
1996 struct mem_cgroup
*victim
= NULL
;
1999 unsigned long excess
;
2000 unsigned long nr_scanned
;
2001 struct mem_cgroup_reclaim_cookie reclaim
= {
2006 excess
= res_counter_soft_limit_excess(&root_memcg
->res
) >> PAGE_SHIFT
;
2009 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
2014 * If we have not been able to reclaim
2015 * anything, it might because there are
2016 * no reclaimable pages under this hierarchy
2021 * We want to do more targeted reclaim.
2022 * excess >> 2 is not to excessive so as to
2023 * reclaim too much, nor too less that we keep
2024 * coming back to reclaim from this cgroup
2026 if (total
>= (excess
>> 2) ||
2027 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
2032 if (!mem_cgroup_reclaimable(victim
, false))
2034 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
2036 *total_scanned
+= nr_scanned
;
2037 if (!res_counter_soft_limit_excess(&root_memcg
->res
))
2040 mem_cgroup_iter_break(root_memcg
, victim
);
2044 #ifdef CONFIG_LOCKDEP
2045 static struct lockdep_map memcg_oom_lock_dep_map
= {
2046 .name
= "memcg_oom_lock",
2050 static DEFINE_SPINLOCK(memcg_oom_lock
);
2053 * Check OOM-Killer is already running under our hierarchy.
2054 * If someone is running, return false.
2056 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
2058 struct mem_cgroup
*iter
, *failed
= NULL
;
2060 spin_lock(&memcg_oom_lock
);
2062 for_each_mem_cgroup_tree(iter
, memcg
) {
2063 if (iter
->oom_lock
) {
2065 * this subtree of our hierarchy is already locked
2066 * so we cannot give a lock.
2069 mem_cgroup_iter_break(memcg
, iter
);
2072 iter
->oom_lock
= true;
2077 * OK, we failed to lock the whole subtree so we have
2078 * to clean up what we set up to the failing subtree
2080 for_each_mem_cgroup_tree(iter
, memcg
) {
2081 if (iter
== failed
) {
2082 mem_cgroup_iter_break(memcg
, iter
);
2085 iter
->oom_lock
= false;
2088 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
2090 spin_unlock(&memcg_oom_lock
);
2095 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
2097 struct mem_cgroup
*iter
;
2099 spin_lock(&memcg_oom_lock
);
2100 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
2101 for_each_mem_cgroup_tree(iter
, memcg
)
2102 iter
->oom_lock
= false;
2103 spin_unlock(&memcg_oom_lock
);
2106 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
2108 struct mem_cgroup
*iter
;
2110 for_each_mem_cgroup_tree(iter
, memcg
)
2111 atomic_inc(&iter
->under_oom
);
2114 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
2116 struct mem_cgroup
*iter
;
2119 * When a new child is created while the hierarchy is under oom,
2120 * mem_cgroup_oom_lock() may not be called. We have to use
2121 * atomic_add_unless() here.
2123 for_each_mem_cgroup_tree(iter
, memcg
)
2124 atomic_add_unless(&iter
->under_oom
, -1, 0);
2127 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
2129 struct oom_wait_info
{
2130 struct mem_cgroup
*memcg
;
2134 static int memcg_oom_wake_function(wait_queue_t
*wait
,
2135 unsigned mode
, int sync
, void *arg
)
2137 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
2138 struct mem_cgroup
*oom_wait_memcg
;
2139 struct oom_wait_info
*oom_wait_info
;
2141 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
2142 oom_wait_memcg
= oom_wait_info
->memcg
;
2145 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2146 * Then we can use css_is_ancestor without taking care of RCU.
2148 if (!mem_cgroup_same_or_subtree(oom_wait_memcg
, wake_memcg
)
2149 && !mem_cgroup_same_or_subtree(wake_memcg
, oom_wait_memcg
))
2151 return autoremove_wake_function(wait
, mode
, sync
, arg
);
2154 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
2156 atomic_inc(&memcg
->oom_wakeups
);
2157 /* for filtering, pass "memcg" as argument. */
2158 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
2161 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
2163 if (memcg
&& atomic_read(&memcg
->under_oom
))
2164 memcg_wakeup_oom(memcg
);
2167 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
2169 if (!current
->memcg_oom
.may_oom
)
2172 * We are in the middle of the charge context here, so we
2173 * don't want to block when potentially sitting on a callstack
2174 * that holds all kinds of filesystem and mm locks.
2176 * Also, the caller may handle a failed allocation gracefully
2177 * (like optional page cache readahead) and so an OOM killer
2178 * invocation might not even be necessary.
2180 * That's why we don't do anything here except remember the
2181 * OOM context and then deal with it at the end of the page
2182 * fault when the stack is unwound, the locks are released,
2183 * and when we know whether the fault was overall successful.
2185 css_get(&memcg
->css
);
2186 current
->memcg_oom
.memcg
= memcg
;
2187 current
->memcg_oom
.gfp_mask
= mask
;
2188 current
->memcg_oom
.order
= order
;
2192 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2193 * @handle: actually kill/wait or just clean up the OOM state
2195 * This has to be called at the end of a page fault if the memcg OOM
2196 * handler was enabled.
2198 * Memcg supports userspace OOM handling where failed allocations must
2199 * sleep on a waitqueue until the userspace task resolves the
2200 * situation. Sleeping directly in the charge context with all kinds
2201 * of locks held is not a good idea, instead we remember an OOM state
2202 * in the task and mem_cgroup_oom_synchronize() has to be called at
2203 * the end of the page fault to complete the OOM handling.
2205 * Returns %true if an ongoing memcg OOM situation was detected and
2206 * completed, %false otherwise.
2208 bool mem_cgroup_oom_synchronize(bool handle
)
2210 struct mem_cgroup
*memcg
= current
->memcg_oom
.memcg
;
2211 struct oom_wait_info owait
;
2214 /* OOM is global, do not handle */
2221 owait
.memcg
= memcg
;
2222 owait
.wait
.flags
= 0;
2223 owait
.wait
.func
= memcg_oom_wake_function
;
2224 owait
.wait
.private = current
;
2225 INIT_LIST_HEAD(&owait
.wait
.task_list
);
2227 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2228 mem_cgroup_mark_under_oom(memcg
);
2230 locked
= mem_cgroup_oom_trylock(memcg
);
2233 mem_cgroup_oom_notify(memcg
);
2235 if (locked
&& !memcg
->oom_kill_disable
) {
2236 mem_cgroup_unmark_under_oom(memcg
);
2237 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2238 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom
.gfp_mask
,
2239 current
->memcg_oom
.order
);
2242 mem_cgroup_unmark_under_oom(memcg
);
2243 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2247 mem_cgroup_oom_unlock(memcg
);
2249 * There is no guarantee that an OOM-lock contender
2250 * sees the wakeups triggered by the OOM kill
2251 * uncharges. Wake any sleepers explicitely.
2253 memcg_oom_recover(memcg
);
2256 current
->memcg_oom
.memcg
= NULL
;
2257 css_put(&memcg
->css
);
2262 * Used to update mapped file or writeback or other statistics.
2264 * Notes: Race condition
2266 * We usually use lock_page_cgroup() for accessing page_cgroup member but
2267 * it tends to be costly. But considering some conditions, we doesn't need
2268 * to do so _always_.
2270 * Considering "charge", lock_page_cgroup() is not required because all
2271 * file-stat operations happen after a page is attached to radix-tree. There
2272 * are no race with "charge".
2274 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2275 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2276 * if there are race with "uncharge". Statistics itself is properly handled
2279 * Considering "move", this is an only case we see a race. To make the race
2280 * small, we check memcg->moving_account and detect there are possibility
2281 * of race or not. If there is, we take a lock.
2284 void __mem_cgroup_begin_update_page_stat(struct page
*page
,
2285 bool *locked
, unsigned long *flags
)
2287 struct mem_cgroup
*memcg
;
2288 struct page_cgroup
*pc
;
2290 pc
= lookup_page_cgroup(page
);
2292 memcg
= pc
->mem_cgroup
;
2293 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2296 * If this memory cgroup is not under account moving, we don't
2297 * need to take move_lock_mem_cgroup(). Because we already hold
2298 * rcu_read_lock(), any calls to move_account will be delayed until
2299 * rcu_read_unlock().
2301 VM_BUG_ON(!rcu_read_lock_held());
2302 if (atomic_read(&memcg
->moving_account
) <= 0)
2305 move_lock_mem_cgroup(memcg
, flags
);
2306 if (memcg
!= pc
->mem_cgroup
|| !PageCgroupUsed(pc
)) {
2307 move_unlock_mem_cgroup(memcg
, flags
);
2313 void __mem_cgroup_end_update_page_stat(struct page
*page
, unsigned long *flags
)
2315 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2318 * It's guaranteed that pc->mem_cgroup never changes while
2319 * lock is held because a routine modifies pc->mem_cgroup
2320 * should take move_lock_mem_cgroup().
2322 move_unlock_mem_cgroup(pc
->mem_cgroup
, flags
);
2325 void mem_cgroup_update_page_stat(struct page
*page
,
2326 enum mem_cgroup_stat_index idx
, int val
)
2328 struct mem_cgroup
*memcg
;
2329 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2330 unsigned long uninitialized_var(flags
);
2332 if (mem_cgroup_disabled())
2335 VM_BUG_ON(!rcu_read_lock_held());
2336 memcg
= pc
->mem_cgroup
;
2337 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2340 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2344 * size of first charge trial. "32" comes from vmscan.c's magic value.
2345 * TODO: maybe necessary to use big numbers in big irons.
2347 #define CHARGE_BATCH 32U
2348 struct memcg_stock_pcp
{
2349 struct mem_cgroup
*cached
; /* this never be root cgroup */
2350 unsigned int nr_pages
;
2351 struct work_struct work
;
2352 unsigned long flags
;
2353 #define FLUSHING_CACHED_CHARGE 0
2355 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2356 static DEFINE_MUTEX(percpu_charge_mutex
);
2359 * consume_stock: Try to consume stocked charge on this cpu.
2360 * @memcg: memcg to consume from.
2361 * @nr_pages: how many pages to charge.
2363 * The charges will only happen if @memcg matches the current cpu's memcg
2364 * stock, and at least @nr_pages are available in that stock. Failure to
2365 * service an allocation will refill the stock.
2367 * returns true if successful, false otherwise.
2369 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2371 struct memcg_stock_pcp
*stock
;
2374 if (nr_pages
> CHARGE_BATCH
)
2377 stock
= &get_cpu_var(memcg_stock
);
2378 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
)
2379 stock
->nr_pages
-= nr_pages
;
2380 else /* need to call res_counter_charge */
2382 put_cpu_var(memcg_stock
);
2387 * Returns stocks cached in percpu to res_counter and reset cached information.
2389 static void drain_stock(struct memcg_stock_pcp
*stock
)
2391 struct mem_cgroup
*old
= stock
->cached
;
2393 if (stock
->nr_pages
) {
2394 unsigned long bytes
= stock
->nr_pages
* PAGE_SIZE
;
2396 res_counter_uncharge(&old
->res
, bytes
);
2397 if (do_swap_account
)
2398 res_counter_uncharge(&old
->memsw
, bytes
);
2399 stock
->nr_pages
= 0;
2401 stock
->cached
= NULL
;
2405 * This must be called under preempt disabled or must be called by
2406 * a thread which is pinned to local cpu.
2408 static void drain_local_stock(struct work_struct
*dummy
)
2410 struct memcg_stock_pcp
*stock
= this_cpu_ptr(&memcg_stock
);
2412 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2415 static void __init
memcg_stock_init(void)
2419 for_each_possible_cpu(cpu
) {
2420 struct memcg_stock_pcp
*stock
=
2421 &per_cpu(memcg_stock
, cpu
);
2422 INIT_WORK(&stock
->work
, drain_local_stock
);
2427 * Cache charges(val) which is from res_counter, to local per_cpu area.
2428 * This will be consumed by consume_stock() function, later.
2430 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2432 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2434 if (stock
->cached
!= memcg
) { /* reset if necessary */
2436 stock
->cached
= memcg
;
2438 stock
->nr_pages
+= nr_pages
;
2439 put_cpu_var(memcg_stock
);
2443 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2444 * of the hierarchy under it. sync flag says whether we should block
2445 * until the work is done.
2447 static void drain_all_stock(struct mem_cgroup
*root_memcg
, bool sync
)
2451 /* Notify other cpus that system-wide "drain" is running */
2454 for_each_online_cpu(cpu
) {
2455 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2456 struct mem_cgroup
*memcg
;
2458 memcg
= stock
->cached
;
2459 if (!memcg
|| !stock
->nr_pages
)
2461 if (!mem_cgroup_same_or_subtree(root_memcg
, memcg
))
2463 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2465 drain_local_stock(&stock
->work
);
2467 schedule_work_on(cpu
, &stock
->work
);
2475 for_each_online_cpu(cpu
) {
2476 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2477 if (test_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
))
2478 flush_work(&stock
->work
);
2485 * Tries to drain stocked charges in other cpus. This function is asynchronous
2486 * and just put a work per cpu for draining localy on each cpu. Caller can
2487 * expects some charges will be back to res_counter later but cannot wait for
2490 static void drain_all_stock_async(struct mem_cgroup
*root_memcg
)
2493 * If someone calls draining, avoid adding more kworker runs.
2495 if (!mutex_trylock(&percpu_charge_mutex
))
2497 drain_all_stock(root_memcg
, false);
2498 mutex_unlock(&percpu_charge_mutex
);
2501 /* This is a synchronous drain interface. */
2502 static void drain_all_stock_sync(struct mem_cgroup
*root_memcg
)
2504 /* called when force_empty is called */
2505 mutex_lock(&percpu_charge_mutex
);
2506 drain_all_stock(root_memcg
, true);
2507 mutex_unlock(&percpu_charge_mutex
);
2511 * This function drains percpu counter value from DEAD cpu and
2512 * move it to local cpu. Note that this function can be preempted.
2514 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2518 spin_lock(&memcg
->pcp_counter_lock
);
2519 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2520 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2522 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2523 memcg
->nocpu_base
.count
[i
] += x
;
2525 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2526 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2528 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2529 memcg
->nocpu_base
.events
[i
] += x
;
2531 spin_unlock(&memcg
->pcp_counter_lock
);
2534 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2535 unsigned long action
,
2538 int cpu
= (unsigned long)hcpu
;
2539 struct memcg_stock_pcp
*stock
;
2540 struct mem_cgroup
*iter
;
2542 if (action
== CPU_ONLINE
)
2545 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2548 for_each_mem_cgroup(iter
)
2549 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2551 stock
= &per_cpu(memcg_stock
, cpu
);
2557 /* See mem_cgroup_try_charge() for details */
2559 CHARGE_OK
, /* success */
2560 CHARGE_RETRY
, /* need to retry but retry is not bad */
2561 CHARGE_NOMEM
, /* we can't do more. return -ENOMEM */
2562 CHARGE_WOULDBLOCK
, /* GFP_WAIT wasn't set and no enough res. */
2565 static int mem_cgroup_do_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2566 unsigned int nr_pages
, unsigned int min_pages
,
2569 unsigned long csize
= nr_pages
* PAGE_SIZE
;
2570 struct mem_cgroup
*mem_over_limit
;
2571 struct res_counter
*fail_res
;
2572 unsigned long flags
= 0;
2575 ret
= res_counter_charge(&memcg
->res
, csize
, &fail_res
);
2578 if (!do_swap_account
)
2580 ret
= res_counter_charge(&memcg
->memsw
, csize
, &fail_res
);
2584 res_counter_uncharge(&memcg
->res
, csize
);
2585 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
2586 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
2588 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
2590 * Never reclaim on behalf of optional batching, retry with a
2591 * single page instead.
2593 if (nr_pages
> min_pages
)
2594 return CHARGE_RETRY
;
2596 if (!(gfp_mask
& __GFP_WAIT
))
2597 return CHARGE_WOULDBLOCK
;
2599 if (gfp_mask
& __GFP_NORETRY
)
2600 return CHARGE_NOMEM
;
2602 ret
= mem_cgroup_reclaim(mem_over_limit
, gfp_mask
, flags
);
2603 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2604 return CHARGE_RETRY
;
2606 * Even though the limit is exceeded at this point, reclaim
2607 * may have been able to free some pages. Retry the charge
2608 * before killing the task.
2610 * Only for regular pages, though: huge pages are rather
2611 * unlikely to succeed so close to the limit, and we fall back
2612 * to regular pages anyway in case of failure.
2614 if (nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
) && ret
)
2615 return CHARGE_RETRY
;
2618 * At task move, charge accounts can be doubly counted. So, it's
2619 * better to wait until the end of task_move if something is going on.
2621 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2622 return CHARGE_RETRY
;
2625 mem_cgroup_oom(mem_over_limit
, gfp_mask
, get_order(csize
));
2627 return CHARGE_NOMEM
;
2631 * mem_cgroup_try_charge - try charging a memcg
2632 * @memcg: memcg to charge
2633 * @nr_pages: number of pages to charge
2634 * @oom: trigger OOM if reclaim fails
2636 * Returns 0 if @memcg was charged successfully, -EINTR if the charge
2637 * was bypassed to root_mem_cgroup, and -ENOMEM if the charge failed.
2639 static int mem_cgroup_try_charge(struct mem_cgroup
*memcg
,
2641 unsigned int nr_pages
,
2644 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2645 int nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2648 if (mem_cgroup_is_root(memcg
))
2651 * Unlike in global OOM situations, memcg is not in a physical
2652 * memory shortage. Allow dying and OOM-killed tasks to
2653 * bypass the last charges so that they can exit quickly and
2654 * free their memory.
2656 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
2657 fatal_signal_pending(current
) ||
2658 current
->flags
& PF_EXITING
))
2661 if (unlikely(task_in_memcg_oom(current
)))
2664 if (gfp_mask
& __GFP_NOFAIL
)
2667 if (consume_stock(memcg
, nr_pages
))
2671 bool invoke_oom
= oom
&& !nr_oom_retries
;
2673 /* If killed, bypass charge */
2674 if (fatal_signal_pending(current
))
2677 ret
= mem_cgroup_do_charge(memcg
, gfp_mask
, batch
,
2678 nr_pages
, invoke_oom
);
2682 case CHARGE_RETRY
: /* not in OOM situation but retry */
2685 case CHARGE_WOULDBLOCK
: /* !__GFP_WAIT */
2687 case CHARGE_NOMEM
: /* OOM routine works */
2688 if (!oom
|| invoke_oom
)
2693 } while (ret
!= CHARGE_OK
);
2695 if (batch
> nr_pages
)
2696 refill_stock(memcg
, batch
- nr_pages
);
2700 if (!(gfp_mask
& __GFP_NOFAIL
))
2707 * mem_cgroup_try_charge_mm - try charging a mm
2708 * @mm: mm_struct to charge
2709 * @nr_pages: number of pages to charge
2710 * @oom: trigger OOM if reclaim fails
2712 * Returns the charged mem_cgroup associated with the given mm_struct or
2713 * NULL the charge failed.
2715 static struct mem_cgroup
*mem_cgroup_try_charge_mm(struct mm_struct
*mm
,
2717 unsigned int nr_pages
,
2721 struct mem_cgroup
*memcg
;
2724 memcg
= get_mem_cgroup_from_mm(mm
);
2725 ret
= mem_cgroup_try_charge(memcg
, gfp_mask
, nr_pages
, oom
);
2726 css_put(&memcg
->css
);
2728 memcg
= root_mem_cgroup
;
2736 * Somemtimes we have to undo a charge we got by try_charge().
2737 * This function is for that and do uncharge, put css's refcnt.
2738 * gotten by try_charge().
2740 static void __mem_cgroup_cancel_charge(struct mem_cgroup
*memcg
,
2741 unsigned int nr_pages
)
2743 if (!mem_cgroup_is_root(memcg
)) {
2744 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2746 res_counter_uncharge(&memcg
->res
, bytes
);
2747 if (do_swap_account
)
2748 res_counter_uncharge(&memcg
->memsw
, bytes
);
2753 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2754 * This is useful when moving usage to parent cgroup.
2756 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup
*memcg
,
2757 unsigned int nr_pages
)
2759 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2761 if (mem_cgroup_is_root(memcg
))
2764 res_counter_uncharge_until(&memcg
->res
, memcg
->res
.parent
, bytes
);
2765 if (do_swap_account
)
2766 res_counter_uncharge_until(&memcg
->memsw
,
2767 memcg
->memsw
.parent
, bytes
);
2771 * A helper function to get mem_cgroup from ID. must be called under
2772 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2773 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2774 * called against removed memcg.)
2776 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2778 /* ID 0 is unused ID */
2781 return mem_cgroup_from_id(id
);
2784 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2786 struct mem_cgroup
*memcg
= NULL
;
2787 struct page_cgroup
*pc
;
2791 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2793 pc
= lookup_page_cgroup(page
);
2794 lock_page_cgroup(pc
);
2795 if (PageCgroupUsed(pc
)) {
2796 memcg
= pc
->mem_cgroup
;
2797 if (memcg
&& !css_tryget(&memcg
->css
))
2799 } else if (PageSwapCache(page
)) {
2800 ent
.val
= page_private(page
);
2801 id
= lookup_swap_cgroup_id(ent
);
2803 memcg
= mem_cgroup_lookup(id
);
2804 if (memcg
&& !css_tryget(&memcg
->css
))
2808 unlock_page_cgroup(pc
);
2812 static void __mem_cgroup_commit_charge(struct mem_cgroup
*memcg
,
2814 unsigned int nr_pages
,
2815 enum charge_type ctype
,
2818 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2819 struct zone
*uninitialized_var(zone
);
2820 struct lruvec
*lruvec
;
2821 bool was_on_lru
= false;
2824 lock_page_cgroup(pc
);
2825 VM_BUG_ON_PAGE(PageCgroupUsed(pc
), page
);
2827 * we don't need page_cgroup_lock about tail pages, becase they are not
2828 * accessed by any other context at this point.
2832 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2833 * may already be on some other mem_cgroup's LRU. Take care of it.
2836 zone
= page_zone(page
);
2837 spin_lock_irq(&zone
->lru_lock
);
2838 if (PageLRU(page
)) {
2839 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2841 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2846 pc
->mem_cgroup
= memcg
;
2848 * We access a page_cgroup asynchronously without lock_page_cgroup().
2849 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2850 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2851 * before USED bit, we need memory barrier here.
2852 * See mem_cgroup_add_lru_list(), etc.
2855 SetPageCgroupUsed(pc
);
2859 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2860 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2862 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2864 spin_unlock_irq(&zone
->lru_lock
);
2867 if (ctype
== MEM_CGROUP_CHARGE_TYPE_ANON
)
2872 mem_cgroup_charge_statistics(memcg
, page
, anon
, nr_pages
);
2873 unlock_page_cgroup(pc
);
2876 * "charge_statistics" updated event counter. Then, check it.
2877 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2878 * if they exceeds softlimit.
2880 memcg_check_events(memcg
, page
);
2883 static DEFINE_MUTEX(set_limit_mutex
);
2885 #ifdef CONFIG_MEMCG_KMEM
2887 * The memcg_slab_mutex is held whenever a per memcg kmem cache is created or
2888 * destroyed. It protects memcg_caches arrays and memcg_slab_caches lists.
2890 static DEFINE_MUTEX(memcg_slab_mutex
);
2892 static DEFINE_MUTEX(activate_kmem_mutex
);
2894 static inline bool memcg_can_account_kmem(struct mem_cgroup
*memcg
)
2896 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
) &&
2897 memcg_kmem_is_active(memcg
);
2901 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2902 * in the memcg_cache_params struct.
2904 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
2906 struct kmem_cache
*cachep
;
2908 VM_BUG_ON(p
->is_root_cache
);
2909 cachep
= p
->root_cache
;
2910 return cache_from_memcg_idx(cachep
, memcg_cache_id(p
->memcg
));
2913 #ifdef CONFIG_SLABINFO
2914 static int mem_cgroup_slabinfo_read(struct seq_file
*m
, void *v
)
2916 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
2917 struct memcg_cache_params
*params
;
2919 if (!memcg_can_account_kmem(memcg
))
2922 print_slabinfo_header(m
);
2924 mutex_lock(&memcg_slab_mutex
);
2925 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
)
2926 cache_show(memcg_params_to_cache(params
), m
);
2927 mutex_unlock(&memcg_slab_mutex
);
2933 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
, u64 size
)
2935 struct res_counter
*fail_res
;
2938 ret
= res_counter_charge(&memcg
->kmem
, size
, &fail_res
);
2942 ret
= mem_cgroup_try_charge(memcg
, gfp
, size
>> PAGE_SHIFT
,
2943 oom_gfp_allowed(gfp
));
2944 if (ret
== -EINTR
) {
2946 * mem_cgroup_try_charge() chosed to bypass to root due to
2947 * OOM kill or fatal signal. Since our only options are to
2948 * either fail the allocation or charge it to this cgroup, do
2949 * it as a temporary condition. But we can't fail. From a
2950 * kmem/slab perspective, the cache has already been selected,
2951 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2954 * This condition will only trigger if the task entered
2955 * memcg_charge_kmem in a sane state, but was OOM-killed during
2956 * mem_cgroup_try_charge() above. Tasks that were already
2957 * dying when the allocation triggers should have been already
2958 * directed to the root cgroup in memcontrol.h
2960 res_counter_charge_nofail(&memcg
->res
, size
, &fail_res
);
2961 if (do_swap_account
)
2962 res_counter_charge_nofail(&memcg
->memsw
, size
,
2966 res_counter_uncharge(&memcg
->kmem
, size
);
2971 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, u64 size
)
2973 res_counter_uncharge(&memcg
->res
, size
);
2974 if (do_swap_account
)
2975 res_counter_uncharge(&memcg
->memsw
, size
);
2978 if (res_counter_uncharge(&memcg
->kmem
, size
))
2982 * Releases a reference taken in kmem_cgroup_css_offline in case
2983 * this last uncharge is racing with the offlining code or it is
2984 * outliving the memcg existence.
2986 * The memory barrier imposed by test&clear is paired with the
2987 * explicit one in memcg_kmem_mark_dead().
2989 if (memcg_kmem_test_and_clear_dead(memcg
))
2990 css_put(&memcg
->css
);
2994 * helper for acessing a memcg's index. It will be used as an index in the
2995 * child cache array in kmem_cache, and also to derive its name. This function
2996 * will return -1 when this is not a kmem-limited memcg.
2998 int memcg_cache_id(struct mem_cgroup
*memcg
)
3000 return memcg
? memcg
->kmemcg_id
: -1;
3003 static size_t memcg_caches_array_size(int num_groups
)
3006 if (num_groups
<= 0)
3009 size
= 2 * num_groups
;
3010 if (size
< MEMCG_CACHES_MIN_SIZE
)
3011 size
= MEMCG_CACHES_MIN_SIZE
;
3012 else if (size
> MEMCG_CACHES_MAX_SIZE
)
3013 size
= MEMCG_CACHES_MAX_SIZE
;
3019 * We should update the current array size iff all caches updates succeed. This
3020 * can only be done from the slab side. The slab mutex needs to be held when
3023 void memcg_update_array_size(int num
)
3025 if (num
> memcg_limited_groups_array_size
)
3026 memcg_limited_groups_array_size
= memcg_caches_array_size(num
);
3029 int memcg_update_cache_size(struct kmem_cache
*s
, int num_groups
)
3031 struct memcg_cache_params
*cur_params
= s
->memcg_params
;
3033 VM_BUG_ON(!is_root_cache(s
));
3035 if (num_groups
> memcg_limited_groups_array_size
) {
3037 struct memcg_cache_params
*new_params
;
3038 ssize_t size
= memcg_caches_array_size(num_groups
);
3040 size
*= sizeof(void *);
3041 size
+= offsetof(struct memcg_cache_params
, memcg_caches
);
3043 new_params
= kzalloc(size
, GFP_KERNEL
);
3047 new_params
->is_root_cache
= true;
3050 * There is the chance it will be bigger than
3051 * memcg_limited_groups_array_size, if we failed an allocation
3052 * in a cache, in which case all caches updated before it, will
3053 * have a bigger array.
3055 * But if that is the case, the data after
3056 * memcg_limited_groups_array_size is certainly unused
3058 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3059 if (!cur_params
->memcg_caches
[i
])
3061 new_params
->memcg_caches
[i
] =
3062 cur_params
->memcg_caches
[i
];
3066 * Ideally, we would wait until all caches succeed, and only
3067 * then free the old one. But this is not worth the extra
3068 * pointer per-cache we'd have to have for this.
3070 * It is not a big deal if some caches are left with a size
3071 * bigger than the others. And all updates will reset this
3074 rcu_assign_pointer(s
->memcg_params
, new_params
);
3076 kfree_rcu(cur_params
, rcu_head
);
3081 int memcg_alloc_cache_params(struct mem_cgroup
*memcg
, struct kmem_cache
*s
,
3082 struct kmem_cache
*root_cache
)
3086 if (!memcg_kmem_enabled())
3090 size
= offsetof(struct memcg_cache_params
, memcg_caches
);
3091 size
+= memcg_limited_groups_array_size
* sizeof(void *);
3093 size
= sizeof(struct memcg_cache_params
);
3095 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3096 if (!s
->memcg_params
)
3100 s
->memcg_params
->memcg
= memcg
;
3101 s
->memcg_params
->root_cache
= root_cache
;
3102 css_get(&memcg
->css
);
3104 s
->memcg_params
->is_root_cache
= true;
3109 void memcg_free_cache_params(struct kmem_cache
*s
)
3111 if (!s
->memcg_params
)
3113 if (!s
->memcg_params
->is_root_cache
)
3114 css_put(&s
->memcg_params
->memcg
->css
);
3115 kfree(s
->memcg_params
);
3118 static void memcg_register_cache(struct mem_cgroup
*memcg
,
3119 struct kmem_cache
*root_cache
)
3121 static char memcg_name_buf
[NAME_MAX
+ 1]; /* protected by
3123 struct kmem_cache
*cachep
;
3126 lockdep_assert_held(&memcg_slab_mutex
);
3128 id
= memcg_cache_id(memcg
);
3131 * Since per-memcg caches are created asynchronously on first
3132 * allocation (see memcg_kmem_get_cache()), several threads can try to
3133 * create the same cache, but only one of them may succeed.
3135 if (cache_from_memcg_idx(root_cache
, id
))
3138 cgroup_name(memcg
->css
.cgroup
, memcg_name_buf
, NAME_MAX
+ 1);
3139 cachep
= memcg_create_kmem_cache(memcg
, root_cache
, memcg_name_buf
);
3141 * If we could not create a memcg cache, do not complain, because
3142 * that's not critical at all as we can always proceed with the root
3148 list_add(&cachep
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
3151 * Since readers won't lock (see cache_from_memcg_idx()), we need a
3152 * barrier here to ensure nobody will see the kmem_cache partially
3157 BUG_ON(root_cache
->memcg_params
->memcg_caches
[id
]);
3158 root_cache
->memcg_params
->memcg_caches
[id
] = cachep
;
3161 static void memcg_unregister_cache(struct kmem_cache
*cachep
)
3163 struct kmem_cache
*root_cache
;
3164 struct mem_cgroup
*memcg
;
3167 lockdep_assert_held(&memcg_slab_mutex
);
3169 BUG_ON(is_root_cache(cachep
));
3171 root_cache
= cachep
->memcg_params
->root_cache
;
3172 memcg
= cachep
->memcg_params
->memcg
;
3173 id
= memcg_cache_id(memcg
);
3175 BUG_ON(root_cache
->memcg_params
->memcg_caches
[id
] != cachep
);
3176 root_cache
->memcg_params
->memcg_caches
[id
] = NULL
;
3178 list_del(&cachep
->memcg_params
->list
);
3180 kmem_cache_destroy(cachep
);
3184 * During the creation a new cache, we need to disable our accounting mechanism
3185 * altogether. This is true even if we are not creating, but rather just
3186 * enqueing new caches to be created.
3188 * This is because that process will trigger allocations; some visible, like
3189 * explicit kmallocs to auxiliary data structures, name strings and internal
3190 * cache structures; some well concealed, like INIT_WORK() that can allocate
3191 * objects during debug.
3193 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3194 * to it. This may not be a bounded recursion: since the first cache creation
3195 * failed to complete (waiting on the allocation), we'll just try to create the
3196 * cache again, failing at the same point.
3198 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3199 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3200 * inside the following two functions.
3202 static inline void memcg_stop_kmem_account(void)
3204 VM_BUG_ON(!current
->mm
);
3205 current
->memcg_kmem_skip_account
++;
3208 static inline void memcg_resume_kmem_account(void)
3210 VM_BUG_ON(!current
->mm
);
3211 current
->memcg_kmem_skip_account
--;
3214 int __memcg_cleanup_cache_params(struct kmem_cache
*s
)
3216 struct kmem_cache
*c
;
3219 mutex_lock(&memcg_slab_mutex
);
3220 for_each_memcg_cache_index(i
) {
3221 c
= cache_from_memcg_idx(s
, i
);
3225 memcg_unregister_cache(c
);
3227 if (cache_from_memcg_idx(s
, i
))
3230 mutex_unlock(&memcg_slab_mutex
);
3234 static void memcg_unregister_all_caches(struct mem_cgroup
*memcg
)
3236 struct kmem_cache
*cachep
;
3237 struct memcg_cache_params
*params
, *tmp
;
3239 if (!memcg_kmem_is_active(memcg
))
3242 mutex_lock(&memcg_slab_mutex
);
3243 list_for_each_entry_safe(params
, tmp
, &memcg
->memcg_slab_caches
, list
) {
3244 cachep
= memcg_params_to_cache(params
);
3245 kmem_cache_shrink(cachep
);
3246 if (atomic_read(&cachep
->memcg_params
->nr_pages
) == 0)
3247 memcg_unregister_cache(cachep
);
3249 mutex_unlock(&memcg_slab_mutex
);
3252 struct memcg_register_cache_work
{
3253 struct mem_cgroup
*memcg
;
3254 struct kmem_cache
*cachep
;
3255 struct work_struct work
;
3258 static void memcg_register_cache_func(struct work_struct
*w
)
3260 struct memcg_register_cache_work
*cw
=
3261 container_of(w
, struct memcg_register_cache_work
, work
);
3262 struct mem_cgroup
*memcg
= cw
->memcg
;
3263 struct kmem_cache
*cachep
= cw
->cachep
;
3265 mutex_lock(&memcg_slab_mutex
);
3266 memcg_register_cache(memcg
, cachep
);
3267 mutex_unlock(&memcg_slab_mutex
);
3269 css_put(&memcg
->css
);
3274 * Enqueue the creation of a per-memcg kmem_cache.
3276 static void __memcg_schedule_register_cache(struct mem_cgroup
*memcg
,
3277 struct kmem_cache
*cachep
)
3279 struct memcg_register_cache_work
*cw
;
3281 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
3283 css_put(&memcg
->css
);
3288 cw
->cachep
= cachep
;
3290 INIT_WORK(&cw
->work
, memcg_register_cache_func
);
3291 schedule_work(&cw
->work
);
3294 static void memcg_schedule_register_cache(struct mem_cgroup
*memcg
,
3295 struct kmem_cache
*cachep
)
3298 * We need to stop accounting when we kmalloc, because if the
3299 * corresponding kmalloc cache is not yet created, the first allocation
3300 * in __memcg_schedule_register_cache will recurse.
3302 * However, it is better to enclose the whole function. Depending on
3303 * the debugging options enabled, INIT_WORK(), for instance, can
3304 * trigger an allocation. This too, will make us recurse. Because at
3305 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3306 * the safest choice is to do it like this, wrapping the whole function.
3308 memcg_stop_kmem_account();
3309 __memcg_schedule_register_cache(memcg
, cachep
);
3310 memcg_resume_kmem_account();
3313 int __memcg_charge_slab(struct kmem_cache
*cachep
, gfp_t gfp
, int order
)
3317 res
= memcg_charge_kmem(cachep
->memcg_params
->memcg
, gfp
,
3318 PAGE_SIZE
<< order
);
3320 atomic_add(1 << order
, &cachep
->memcg_params
->nr_pages
);
3324 void __memcg_uncharge_slab(struct kmem_cache
*cachep
, int order
)
3326 memcg_uncharge_kmem(cachep
->memcg_params
->memcg
, PAGE_SIZE
<< order
);
3327 atomic_sub(1 << order
, &cachep
->memcg_params
->nr_pages
);
3331 * Return the kmem_cache we're supposed to use for a slab allocation.
3332 * We try to use the current memcg's version of the cache.
3334 * If the cache does not exist yet, if we are the first user of it,
3335 * we either create it immediately, if possible, or create it asynchronously
3337 * In the latter case, we will let the current allocation go through with
3338 * the original cache.
3340 * Can't be called in interrupt context or from kernel threads.
3341 * This function needs to be called with rcu_read_lock() held.
3343 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
3346 struct mem_cgroup
*memcg
;
3347 struct kmem_cache
*memcg_cachep
;
3349 VM_BUG_ON(!cachep
->memcg_params
);
3350 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
3352 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3356 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
3358 if (!memcg_can_account_kmem(memcg
))
3361 memcg_cachep
= cache_from_memcg_idx(cachep
, memcg_cache_id(memcg
));
3362 if (likely(memcg_cachep
)) {
3363 cachep
= memcg_cachep
;
3367 /* The corresponding put will be done in the workqueue. */
3368 if (!css_tryget(&memcg
->css
))
3373 * If we are in a safe context (can wait, and not in interrupt
3374 * context), we could be be predictable and return right away.
3375 * This would guarantee that the allocation being performed
3376 * already belongs in the new cache.
3378 * However, there are some clashes that can arrive from locking.
3379 * For instance, because we acquire the slab_mutex while doing
3380 * memcg_create_kmem_cache, this means no further allocation
3381 * could happen with the slab_mutex held. So it's better to
3384 memcg_schedule_register_cache(memcg
, cachep
);
3392 * We need to verify if the allocation against current->mm->owner's memcg is
3393 * possible for the given order. But the page is not allocated yet, so we'll
3394 * need a further commit step to do the final arrangements.
3396 * It is possible for the task to switch cgroups in this mean time, so at
3397 * commit time, we can't rely on task conversion any longer. We'll then use
3398 * the handle argument to return to the caller which cgroup we should commit
3399 * against. We could also return the memcg directly and avoid the pointer
3400 * passing, but a boolean return value gives better semantics considering
3401 * the compiled-out case as well.
3403 * Returning true means the allocation is possible.
3406 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
3408 struct mem_cgroup
*memcg
;
3414 * Disabling accounting is only relevant for some specific memcg
3415 * internal allocations. Therefore we would initially not have such
3416 * check here, since direct calls to the page allocator that are
3417 * accounted to kmemcg (alloc_kmem_pages and friends) only happen
3418 * outside memcg core. We are mostly concerned with cache allocations,
3419 * and by having this test at memcg_kmem_get_cache, we are already able
3420 * to relay the allocation to the root cache and bypass the memcg cache
3423 * There is one exception, though: the SLUB allocator does not create
3424 * large order caches, but rather service large kmallocs directly from
3425 * the page allocator. Therefore, the following sequence when backed by
3426 * the SLUB allocator:
3428 * memcg_stop_kmem_account();
3429 * kmalloc(<large_number>)
3430 * memcg_resume_kmem_account();
3432 * would effectively ignore the fact that we should skip accounting,
3433 * since it will drive us directly to this function without passing
3434 * through the cache selector memcg_kmem_get_cache. Such large
3435 * allocations are extremely rare but can happen, for instance, for the
3436 * cache arrays. We bring this test here.
3438 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3441 memcg
= get_mem_cgroup_from_mm(current
->mm
);
3443 if (!memcg_can_account_kmem(memcg
)) {
3444 css_put(&memcg
->css
);
3448 ret
= memcg_charge_kmem(memcg
, gfp
, PAGE_SIZE
<< order
);
3452 css_put(&memcg
->css
);
3456 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
3459 struct page_cgroup
*pc
;
3461 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3463 /* The page allocation failed. Revert */
3465 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3469 pc
= lookup_page_cgroup(page
);
3470 lock_page_cgroup(pc
);
3471 pc
->mem_cgroup
= memcg
;
3472 SetPageCgroupUsed(pc
);
3473 unlock_page_cgroup(pc
);
3476 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
3478 struct mem_cgroup
*memcg
= NULL
;
3479 struct page_cgroup
*pc
;
3482 pc
= lookup_page_cgroup(page
);
3484 * Fast unlocked return. Theoretically might have changed, have to
3485 * check again after locking.
3487 if (!PageCgroupUsed(pc
))
3490 lock_page_cgroup(pc
);
3491 if (PageCgroupUsed(pc
)) {
3492 memcg
= pc
->mem_cgroup
;
3493 ClearPageCgroupUsed(pc
);
3495 unlock_page_cgroup(pc
);
3498 * We trust that only if there is a memcg associated with the page, it
3499 * is a valid allocation
3504 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
3505 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3508 static inline void memcg_unregister_all_caches(struct mem_cgroup
*memcg
)
3511 #endif /* CONFIG_MEMCG_KMEM */
3513 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3515 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3517 * Because tail pages are not marked as "used", set it. We're under
3518 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3519 * charge/uncharge will be never happen and move_account() is done under
3520 * compound_lock(), so we don't have to take care of races.
3522 void mem_cgroup_split_huge_fixup(struct page
*head
)
3524 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
3525 struct page_cgroup
*pc
;
3526 struct mem_cgroup
*memcg
;
3529 if (mem_cgroup_disabled())
3532 memcg
= head_pc
->mem_cgroup
;
3533 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3535 pc
->mem_cgroup
= memcg
;
3536 smp_wmb();/* see __commit_charge() */
3537 pc
->flags
= head_pc
->flags
& ~PCGF_NOCOPY_AT_SPLIT
;
3539 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
3542 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3545 * mem_cgroup_move_account - move account of the page
3547 * @nr_pages: number of regular pages (>1 for huge pages)
3548 * @pc: page_cgroup of the page.
3549 * @from: mem_cgroup which the page is moved from.
3550 * @to: mem_cgroup which the page is moved to. @from != @to.
3552 * The caller must confirm following.
3553 * - page is not on LRU (isolate_page() is useful.)
3554 * - compound_lock is held when nr_pages > 1
3556 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3559 static int mem_cgroup_move_account(struct page
*page
,
3560 unsigned int nr_pages
,
3561 struct page_cgroup
*pc
,
3562 struct mem_cgroup
*from
,
3563 struct mem_cgroup
*to
)
3565 unsigned long flags
;
3567 bool anon
= PageAnon(page
);
3569 VM_BUG_ON(from
== to
);
3570 VM_BUG_ON_PAGE(PageLRU(page
), page
);
3572 * The page is isolated from LRU. So, collapse function
3573 * will not handle this page. But page splitting can happen.
3574 * Do this check under compound_page_lock(). The caller should
3578 if (nr_pages
> 1 && !PageTransHuge(page
))
3581 lock_page_cgroup(pc
);
3584 if (!PageCgroupUsed(pc
) || pc
->mem_cgroup
!= from
)
3587 move_lock_mem_cgroup(from
, &flags
);
3589 if (!anon
&& page_mapped(page
)) {
3590 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
3592 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
3596 if (PageWriteback(page
)) {
3597 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
3599 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
3603 mem_cgroup_charge_statistics(from
, page
, anon
, -nr_pages
);
3605 /* caller should have done css_get */
3606 pc
->mem_cgroup
= to
;
3607 mem_cgroup_charge_statistics(to
, page
, anon
, nr_pages
);
3608 move_unlock_mem_cgroup(from
, &flags
);
3611 unlock_page_cgroup(pc
);
3615 memcg_check_events(to
, page
);
3616 memcg_check_events(from
, page
);
3622 * mem_cgroup_move_parent - moves page to the parent group
3623 * @page: the page to move
3624 * @pc: page_cgroup of the page
3625 * @child: page's cgroup
3627 * move charges to its parent or the root cgroup if the group has no
3628 * parent (aka use_hierarchy==0).
3629 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3630 * mem_cgroup_move_account fails) the failure is always temporary and
3631 * it signals a race with a page removal/uncharge or migration. In the
3632 * first case the page is on the way out and it will vanish from the LRU
3633 * on the next attempt and the call should be retried later.
3634 * Isolation from the LRU fails only if page has been isolated from
3635 * the LRU since we looked at it and that usually means either global
3636 * reclaim or migration going on. The page will either get back to the
3638 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3639 * (!PageCgroupUsed) or moved to a different group. The page will
3640 * disappear in the next attempt.
3642 static int mem_cgroup_move_parent(struct page
*page
,
3643 struct page_cgroup
*pc
,
3644 struct mem_cgroup
*child
)
3646 struct mem_cgroup
*parent
;
3647 unsigned int nr_pages
;
3648 unsigned long uninitialized_var(flags
);
3651 VM_BUG_ON(mem_cgroup_is_root(child
));
3654 if (!get_page_unless_zero(page
))
3656 if (isolate_lru_page(page
))
3659 nr_pages
= hpage_nr_pages(page
);
3661 parent
= parent_mem_cgroup(child
);
3663 * If no parent, move charges to root cgroup.
3666 parent
= root_mem_cgroup
;
3669 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
3670 flags
= compound_lock_irqsave(page
);
3673 ret
= mem_cgroup_move_account(page
, nr_pages
,
3676 __mem_cgroup_cancel_local_charge(child
, nr_pages
);
3679 compound_unlock_irqrestore(page
, flags
);
3680 putback_lru_page(page
);
3687 int mem_cgroup_charge_anon(struct page
*page
,
3688 struct mm_struct
*mm
, gfp_t gfp_mask
)
3690 unsigned int nr_pages
= 1;
3691 struct mem_cgroup
*memcg
;
3694 if (mem_cgroup_disabled())
3697 VM_BUG_ON_PAGE(page_mapped(page
), page
);
3698 VM_BUG_ON_PAGE(page
->mapping
&& !PageAnon(page
), page
);
3701 if (PageTransHuge(page
)) {
3702 nr_pages
<<= compound_order(page
);
3703 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
3705 * Never OOM-kill a process for a huge page. The
3706 * fault handler will fall back to regular pages.
3711 memcg
= mem_cgroup_try_charge_mm(mm
, gfp_mask
, nr_pages
, oom
);
3714 __mem_cgroup_commit_charge(memcg
, page
, nr_pages
,
3715 MEM_CGROUP_CHARGE_TYPE_ANON
, false);
3720 * While swap-in, try_charge -> commit or cancel, the page is locked.
3721 * And when try_charge() successfully returns, one refcnt to memcg without
3722 * struct page_cgroup is acquired. This refcnt will be consumed by
3723 * "commit()" or removed by "cancel()"
3725 static int __mem_cgroup_try_charge_swapin(struct mm_struct
*mm
,
3728 struct mem_cgroup
**memcgp
)
3730 struct mem_cgroup
*memcg
= NULL
;
3731 struct page_cgroup
*pc
;
3734 pc
= lookup_page_cgroup(page
);
3736 * Every swap fault against a single page tries to charge the
3737 * page, bail as early as possible. shmem_unuse() encounters
3738 * already charged pages, too. The USED bit is protected by
3739 * the page lock, which serializes swap cache removal, which
3740 * in turn serializes uncharging.
3742 if (PageCgroupUsed(pc
))
3744 if (do_swap_account
)
3745 memcg
= try_get_mem_cgroup_from_page(page
);
3747 memcg
= get_mem_cgroup_from_mm(mm
);
3748 ret
= mem_cgroup_try_charge(memcg
, mask
, 1, true);
3749 css_put(&memcg
->css
);
3751 memcg
= root_mem_cgroup
;
3759 int mem_cgroup_try_charge_swapin(struct mm_struct
*mm
, struct page
*page
,
3760 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
3762 if (mem_cgroup_disabled()) {
3767 * A racing thread's fault, or swapoff, may have already
3768 * updated the pte, and even removed page from swap cache: in
3769 * those cases unuse_pte()'s pte_same() test will fail; but
3770 * there's also a KSM case which does need to charge the page.
3772 if (!PageSwapCache(page
)) {
3773 struct mem_cgroup
*memcg
;
3775 memcg
= mem_cgroup_try_charge_mm(mm
, gfp_mask
, 1, true);
3781 return __mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, memcgp
);
3784 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup
*memcg
)
3786 if (mem_cgroup_disabled())
3790 __mem_cgroup_cancel_charge(memcg
, 1);
3794 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*memcg
,
3795 enum charge_type ctype
)
3797 if (mem_cgroup_disabled())
3802 __mem_cgroup_commit_charge(memcg
, page
, 1, ctype
, true);
3804 * Now swap is on-memory. This means this page may be
3805 * counted both as mem and swap....double count.
3806 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3807 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3808 * may call delete_from_swap_cache() before reach here.
3810 if (do_swap_account
&& PageSwapCache(page
)) {
3811 swp_entry_t ent
= {.val
= page_private(page
)};
3812 mem_cgroup_uncharge_swap(ent
);
3816 void mem_cgroup_commit_charge_swapin(struct page
*page
,
3817 struct mem_cgroup
*memcg
)
3819 __mem_cgroup_commit_charge_swapin(page
, memcg
,
3820 MEM_CGROUP_CHARGE_TYPE_ANON
);
3823 int mem_cgroup_charge_file(struct page
*page
, struct mm_struct
*mm
,
3826 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
3827 struct mem_cgroup
*memcg
;
3830 if (mem_cgroup_disabled())
3832 if (PageCompound(page
))
3835 if (PageSwapCache(page
)) { /* shmem */
3836 ret
= __mem_cgroup_try_charge_swapin(mm
, page
,
3840 __mem_cgroup_commit_charge_swapin(page
, memcg
, type
);
3844 memcg
= mem_cgroup_try_charge_mm(mm
, gfp_mask
, 1, true);
3847 __mem_cgroup_commit_charge(memcg
, page
, 1, type
, false);
3851 static void mem_cgroup_do_uncharge(struct mem_cgroup
*memcg
,
3852 unsigned int nr_pages
,
3853 const enum charge_type ctype
)
3855 struct memcg_batch_info
*batch
= NULL
;
3856 bool uncharge_memsw
= true;
3858 /* If swapout, usage of swap doesn't decrease */
3859 if (!do_swap_account
|| ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
)
3860 uncharge_memsw
= false;
3862 batch
= ¤t
->memcg_batch
;
3864 * In usual, we do css_get() when we remember memcg pointer.
3865 * But in this case, we keep res->usage until end of a series of
3866 * uncharges. Then, it's ok to ignore memcg's refcnt.
3869 batch
->memcg
= memcg
;
3871 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3872 * In those cases, all pages freed continuously can be expected to be in
3873 * the same cgroup and we have chance to coalesce uncharges.
3874 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3875 * because we want to do uncharge as soon as possible.
3878 if (!batch
->do_batch
|| test_thread_flag(TIF_MEMDIE
))
3879 goto direct_uncharge
;
3882 goto direct_uncharge
;
3885 * In typical case, batch->memcg == mem. This means we can
3886 * merge a series of uncharges to an uncharge of res_counter.
3887 * If not, we uncharge res_counter ony by one.
3889 if (batch
->memcg
!= memcg
)
3890 goto direct_uncharge
;
3891 /* remember freed charge and uncharge it later */
3894 batch
->memsw_nr_pages
++;
3897 res_counter_uncharge(&memcg
->res
, nr_pages
* PAGE_SIZE
);
3899 res_counter_uncharge(&memcg
->memsw
, nr_pages
* PAGE_SIZE
);
3900 if (unlikely(batch
->memcg
!= memcg
))
3901 memcg_oom_recover(memcg
);
3905 * uncharge if !page_mapped(page)
3907 static struct mem_cgroup
*
3908 __mem_cgroup_uncharge_common(struct page
*page
, enum charge_type ctype
,
3911 struct mem_cgroup
*memcg
= NULL
;
3912 unsigned int nr_pages
= 1;
3913 struct page_cgroup
*pc
;
3916 if (mem_cgroup_disabled())
3919 if (PageTransHuge(page
)) {
3920 nr_pages
<<= compound_order(page
);
3921 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
3924 * Check if our page_cgroup is valid
3926 pc
= lookup_page_cgroup(page
);
3927 if (unlikely(!PageCgroupUsed(pc
)))
3930 lock_page_cgroup(pc
);
3932 memcg
= pc
->mem_cgroup
;
3934 if (!PageCgroupUsed(pc
))
3937 anon
= PageAnon(page
);
3940 case MEM_CGROUP_CHARGE_TYPE_ANON
:
3942 * Generally PageAnon tells if it's the anon statistics to be
3943 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
3944 * used before page reached the stage of being marked PageAnon.
3948 case MEM_CGROUP_CHARGE_TYPE_DROP
:
3949 /* See mem_cgroup_prepare_migration() */
3950 if (page_mapped(page
))
3953 * Pages under migration may not be uncharged. But
3954 * end_migration() /must/ be the one uncharging the
3955 * unused post-migration page and so it has to call
3956 * here with the migration bit still set. See the
3957 * res_counter handling below.
3959 if (!end_migration
&& PageCgroupMigration(pc
))
3962 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT
:
3963 if (!PageAnon(page
)) { /* Shared memory */
3964 if (page
->mapping
&& !page_is_file_cache(page
))
3966 } else if (page_mapped(page
)) /* Anon */
3973 mem_cgroup_charge_statistics(memcg
, page
, anon
, -nr_pages
);
3975 ClearPageCgroupUsed(pc
);
3977 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3978 * freed from LRU. This is safe because uncharged page is expected not
3979 * to be reused (freed soon). Exception is SwapCache, it's handled by
3980 * special functions.
3983 unlock_page_cgroup(pc
);
3985 * even after unlock, we have memcg->res.usage here and this memcg
3986 * will never be freed, so it's safe to call css_get().
3988 memcg_check_events(memcg
, page
);
3989 if (do_swap_account
&& ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
) {
3990 mem_cgroup_swap_statistics(memcg
, true);
3991 css_get(&memcg
->css
);
3994 * Migration does not charge the res_counter for the
3995 * replacement page, so leave it alone when phasing out the
3996 * page that is unused after the migration.
3998 if (!end_migration
&& !mem_cgroup_is_root(memcg
))
3999 mem_cgroup_do_uncharge(memcg
, nr_pages
, ctype
);
4004 unlock_page_cgroup(pc
);
4008 void mem_cgroup_uncharge_page(struct page
*page
)
4011 if (page_mapped(page
))
4013 VM_BUG_ON_PAGE(page
->mapping
&& !PageAnon(page
), page
);
4015 * If the page is in swap cache, uncharge should be deferred
4016 * to the swap path, which also properly accounts swap usage
4017 * and handles memcg lifetime.
4019 * Note that this check is not stable and reclaim may add the
4020 * page to swap cache at any time after this. However, if the
4021 * page is not in swap cache by the time page->mapcount hits
4022 * 0, there won't be any page table references to the swap
4023 * slot, and reclaim will free it and not actually write the
4026 if (PageSwapCache(page
))
4028 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_ANON
, false);
4031 void mem_cgroup_uncharge_cache_page(struct page
*page
)
4033 VM_BUG_ON_PAGE(page_mapped(page
), page
);
4034 VM_BUG_ON_PAGE(page
->mapping
, page
);
4035 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_CACHE
, false);
4039 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4040 * In that cases, pages are freed continuously and we can expect pages
4041 * are in the same memcg. All these calls itself limits the number of
4042 * pages freed at once, then uncharge_start/end() is called properly.
4043 * This may be called prural(2) times in a context,
4046 void mem_cgroup_uncharge_start(void)
4048 current
->memcg_batch
.do_batch
++;
4049 /* We can do nest. */
4050 if (current
->memcg_batch
.do_batch
== 1) {
4051 current
->memcg_batch
.memcg
= NULL
;
4052 current
->memcg_batch
.nr_pages
= 0;
4053 current
->memcg_batch
.memsw_nr_pages
= 0;
4057 void mem_cgroup_uncharge_end(void)
4059 struct memcg_batch_info
*batch
= ¤t
->memcg_batch
;
4061 if (!batch
->do_batch
)
4065 if (batch
->do_batch
) /* If stacked, do nothing. */
4071 * This "batch->memcg" is valid without any css_get/put etc...
4072 * bacause we hide charges behind us.
4074 if (batch
->nr_pages
)
4075 res_counter_uncharge(&batch
->memcg
->res
,
4076 batch
->nr_pages
* PAGE_SIZE
);
4077 if (batch
->memsw_nr_pages
)
4078 res_counter_uncharge(&batch
->memcg
->memsw
,
4079 batch
->memsw_nr_pages
* PAGE_SIZE
);
4080 memcg_oom_recover(batch
->memcg
);
4081 /* forget this pointer (for sanity check) */
4082 batch
->memcg
= NULL
;
4087 * called after __delete_from_swap_cache() and drop "page" account.
4088 * memcg information is recorded to swap_cgroup of "ent"
4091 mem_cgroup_uncharge_swapcache(struct page
*page
, swp_entry_t ent
, bool swapout
)
4093 struct mem_cgroup
*memcg
;
4094 int ctype
= MEM_CGROUP_CHARGE_TYPE_SWAPOUT
;
4096 if (!swapout
) /* this was a swap cache but the swap is unused ! */
4097 ctype
= MEM_CGROUP_CHARGE_TYPE_DROP
;
4099 memcg
= __mem_cgroup_uncharge_common(page
, ctype
, false);
4102 * record memcg information, if swapout && memcg != NULL,
4103 * css_get() was called in uncharge().
4105 if (do_swap_account
&& swapout
&& memcg
)
4106 swap_cgroup_record(ent
, mem_cgroup_id(memcg
));
4110 #ifdef CONFIG_MEMCG_SWAP
4112 * called from swap_entry_free(). remove record in swap_cgroup and
4113 * uncharge "memsw" account.
4115 void mem_cgroup_uncharge_swap(swp_entry_t ent
)
4117 struct mem_cgroup
*memcg
;
4120 if (!do_swap_account
)
4123 id
= swap_cgroup_record(ent
, 0);
4125 memcg
= mem_cgroup_lookup(id
);
4128 * We uncharge this because swap is freed.
4129 * This memcg can be obsolete one. We avoid calling css_tryget
4131 if (!mem_cgroup_is_root(memcg
))
4132 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
4133 mem_cgroup_swap_statistics(memcg
, false);
4134 css_put(&memcg
->css
);
4140 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4141 * @entry: swap entry to be moved
4142 * @from: mem_cgroup which the entry is moved from
4143 * @to: mem_cgroup which the entry is moved to
4145 * It succeeds only when the swap_cgroup's record for this entry is the same
4146 * as the mem_cgroup's id of @from.
4148 * Returns 0 on success, -EINVAL on failure.
4150 * The caller must have charged to @to, IOW, called res_counter_charge() about
4151 * both res and memsw, and called css_get().
4153 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
4154 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4156 unsigned short old_id
, new_id
;
4158 old_id
= mem_cgroup_id(from
);
4159 new_id
= mem_cgroup_id(to
);
4161 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
4162 mem_cgroup_swap_statistics(from
, false);
4163 mem_cgroup_swap_statistics(to
, true);
4165 * This function is only called from task migration context now.
4166 * It postpones res_counter and refcount handling till the end
4167 * of task migration(mem_cgroup_clear_mc()) for performance
4168 * improvement. But we cannot postpone css_get(to) because if
4169 * the process that has been moved to @to does swap-in, the
4170 * refcount of @to might be decreased to 0.
4172 * We are in attach() phase, so the cgroup is guaranteed to be
4173 * alive, so we can just call css_get().
4181 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
4182 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4189 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4192 void mem_cgroup_prepare_migration(struct page
*page
, struct page
*newpage
,
4193 struct mem_cgroup
**memcgp
)
4195 struct mem_cgroup
*memcg
= NULL
;
4196 unsigned int nr_pages
= 1;
4197 struct page_cgroup
*pc
;
4198 enum charge_type ctype
;
4202 if (mem_cgroup_disabled())
4205 if (PageTransHuge(page
))
4206 nr_pages
<<= compound_order(page
);
4208 pc
= lookup_page_cgroup(page
);
4209 lock_page_cgroup(pc
);
4210 if (PageCgroupUsed(pc
)) {
4211 memcg
= pc
->mem_cgroup
;
4212 css_get(&memcg
->css
);
4214 * At migrating an anonymous page, its mapcount goes down
4215 * to 0 and uncharge() will be called. But, even if it's fully
4216 * unmapped, migration may fail and this page has to be
4217 * charged again. We set MIGRATION flag here and delay uncharge
4218 * until end_migration() is called
4220 * Corner Case Thinking
4222 * When the old page was mapped as Anon and it's unmap-and-freed
4223 * while migration was ongoing.
4224 * If unmap finds the old page, uncharge() of it will be delayed
4225 * until end_migration(). If unmap finds a new page, it's
4226 * uncharged when it make mapcount to be 1->0. If unmap code
4227 * finds swap_migration_entry, the new page will not be mapped
4228 * and end_migration() will find it(mapcount==0).
4231 * When the old page was mapped but migraion fails, the kernel
4232 * remaps it. A charge for it is kept by MIGRATION flag even
4233 * if mapcount goes down to 0. We can do remap successfully
4234 * without charging it again.
4237 * The "old" page is under lock_page() until the end of
4238 * migration, so, the old page itself will not be swapped-out.
4239 * If the new page is swapped out before end_migraton, our
4240 * hook to usual swap-out path will catch the event.
4243 SetPageCgroupMigration(pc
);
4245 unlock_page_cgroup(pc
);
4247 * If the page is not charged at this point,
4255 * We charge new page before it's used/mapped. So, even if unlock_page()
4256 * is called before end_migration, we can catch all events on this new
4257 * page. In the case new page is migrated but not remapped, new page's
4258 * mapcount will be finally 0 and we call uncharge in end_migration().
4261 ctype
= MEM_CGROUP_CHARGE_TYPE_ANON
;
4263 ctype
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4265 * The page is committed to the memcg, but it's not actually
4266 * charged to the res_counter since we plan on replacing the
4267 * old one and only one page is going to be left afterwards.
4269 __mem_cgroup_commit_charge(memcg
, newpage
, nr_pages
, ctype
, false);
4272 /* remove redundant charge if migration failed*/
4273 void mem_cgroup_end_migration(struct mem_cgroup
*memcg
,
4274 struct page
*oldpage
, struct page
*newpage
, bool migration_ok
)
4276 struct page
*used
, *unused
;
4277 struct page_cgroup
*pc
;
4283 if (!migration_ok
) {
4290 anon
= PageAnon(used
);
4291 __mem_cgroup_uncharge_common(unused
,
4292 anon
? MEM_CGROUP_CHARGE_TYPE_ANON
4293 : MEM_CGROUP_CHARGE_TYPE_CACHE
,
4295 css_put(&memcg
->css
);
4297 * We disallowed uncharge of pages under migration because mapcount
4298 * of the page goes down to zero, temporarly.
4299 * Clear the flag and check the page should be charged.
4301 pc
= lookup_page_cgroup(oldpage
);
4302 lock_page_cgroup(pc
);
4303 ClearPageCgroupMigration(pc
);
4304 unlock_page_cgroup(pc
);
4307 * If a page is a file cache, radix-tree replacement is very atomic
4308 * and we can skip this check. When it was an Anon page, its mapcount
4309 * goes down to 0. But because we added MIGRATION flage, it's not
4310 * uncharged yet. There are several case but page->mapcount check
4311 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4312 * check. (see prepare_charge() also)
4315 mem_cgroup_uncharge_page(used
);
4319 * At replace page cache, newpage is not under any memcg but it's on
4320 * LRU. So, this function doesn't touch res_counter but handles LRU
4321 * in correct way. Both pages are locked so we cannot race with uncharge.
4323 void mem_cgroup_replace_page_cache(struct page
*oldpage
,
4324 struct page
*newpage
)
4326 struct mem_cgroup
*memcg
= NULL
;
4327 struct page_cgroup
*pc
;
4328 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4330 if (mem_cgroup_disabled())
4333 pc
= lookup_page_cgroup(oldpage
);
4334 /* fix accounting on old pages */
4335 lock_page_cgroup(pc
);
4336 if (PageCgroupUsed(pc
)) {
4337 memcg
= pc
->mem_cgroup
;
4338 mem_cgroup_charge_statistics(memcg
, oldpage
, false, -1);
4339 ClearPageCgroupUsed(pc
);
4341 unlock_page_cgroup(pc
);
4344 * When called from shmem_replace_page(), in some cases the
4345 * oldpage has already been charged, and in some cases not.
4350 * Even if newpage->mapping was NULL before starting replacement,
4351 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4352 * LRU while we overwrite pc->mem_cgroup.
4354 __mem_cgroup_commit_charge(memcg
, newpage
, 1, type
, true);
4357 #ifdef CONFIG_DEBUG_VM
4358 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
4360 struct page_cgroup
*pc
;
4362 pc
= lookup_page_cgroup(page
);
4364 * Can be NULL while feeding pages into the page allocator for
4365 * the first time, i.e. during boot or memory hotplug;
4366 * or when mem_cgroup_disabled().
4368 if (likely(pc
) && PageCgroupUsed(pc
))
4373 bool mem_cgroup_bad_page_check(struct page
*page
)
4375 if (mem_cgroup_disabled())
4378 return lookup_page_cgroup_used(page
) != NULL
;
4381 void mem_cgroup_print_bad_page(struct page
*page
)
4383 struct page_cgroup
*pc
;
4385 pc
= lookup_page_cgroup_used(page
);
4387 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4388 pc
, pc
->flags
, pc
->mem_cgroup
);
4393 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
4394 unsigned long long val
)
4397 u64 memswlimit
, memlimit
;
4399 int children
= mem_cgroup_count_children(memcg
);
4400 u64 curusage
, oldusage
;
4404 * For keeping hierarchical_reclaim simple, how long we should retry
4405 * is depends on callers. We set our retry-count to be function
4406 * of # of children which we should visit in this loop.
4408 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
4410 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4413 while (retry_count
) {
4414 if (signal_pending(current
)) {
4419 * Rather than hide all in some function, I do this in
4420 * open coded manner. You see what this really does.
4421 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4423 mutex_lock(&set_limit_mutex
);
4424 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4425 if (memswlimit
< val
) {
4427 mutex_unlock(&set_limit_mutex
);
4431 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4435 ret
= res_counter_set_limit(&memcg
->res
, val
);
4437 if (memswlimit
== val
)
4438 memcg
->memsw_is_minimum
= true;
4440 memcg
->memsw_is_minimum
= false;
4442 mutex_unlock(&set_limit_mutex
);
4447 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4448 MEM_CGROUP_RECLAIM_SHRINK
);
4449 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4450 /* Usage is reduced ? */
4451 if (curusage
>= oldusage
)
4454 oldusage
= curusage
;
4456 if (!ret
&& enlarge
)
4457 memcg_oom_recover(memcg
);
4462 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
4463 unsigned long long val
)
4466 u64 memlimit
, memswlimit
, oldusage
, curusage
;
4467 int children
= mem_cgroup_count_children(memcg
);
4471 /* see mem_cgroup_resize_res_limit */
4472 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
4473 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4474 while (retry_count
) {
4475 if (signal_pending(current
)) {
4480 * Rather than hide all in some function, I do this in
4481 * open coded manner. You see what this really does.
4482 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4484 mutex_lock(&set_limit_mutex
);
4485 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4486 if (memlimit
> val
) {
4488 mutex_unlock(&set_limit_mutex
);
4491 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4492 if (memswlimit
< val
)
4494 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
4496 if (memlimit
== val
)
4497 memcg
->memsw_is_minimum
= true;
4499 memcg
->memsw_is_minimum
= false;
4501 mutex_unlock(&set_limit_mutex
);
4506 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4507 MEM_CGROUP_RECLAIM_NOSWAP
|
4508 MEM_CGROUP_RECLAIM_SHRINK
);
4509 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4510 /* Usage is reduced ? */
4511 if (curusage
>= oldusage
)
4514 oldusage
= curusage
;
4516 if (!ret
&& enlarge
)
4517 memcg_oom_recover(memcg
);
4521 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
4523 unsigned long *total_scanned
)
4525 unsigned long nr_reclaimed
= 0;
4526 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
4527 unsigned long reclaimed
;
4529 struct mem_cgroup_tree_per_zone
*mctz
;
4530 unsigned long long excess
;
4531 unsigned long nr_scanned
;
4536 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
4538 * This loop can run a while, specially if mem_cgroup's continuously
4539 * keep exceeding their soft limit and putting the system under
4546 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
4551 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
4552 gfp_mask
, &nr_scanned
);
4553 nr_reclaimed
+= reclaimed
;
4554 *total_scanned
+= nr_scanned
;
4555 spin_lock(&mctz
->lock
);
4558 * If we failed to reclaim anything from this memory cgroup
4559 * it is time to move on to the next cgroup
4565 * Loop until we find yet another one.
4567 * By the time we get the soft_limit lock
4568 * again, someone might have aded the
4569 * group back on the RB tree. Iterate to
4570 * make sure we get a different mem.
4571 * mem_cgroup_largest_soft_limit_node returns
4572 * NULL if no other cgroup is present on
4576 __mem_cgroup_largest_soft_limit_node(mctz
);
4578 css_put(&next_mz
->memcg
->css
);
4579 else /* next_mz == NULL or other memcg */
4583 __mem_cgroup_remove_exceeded(mz
, mctz
);
4584 excess
= res_counter_soft_limit_excess(&mz
->memcg
->res
);
4586 * One school of thought says that we should not add
4587 * back the node to the tree if reclaim returns 0.
4588 * But our reclaim could return 0, simply because due
4589 * to priority we are exposing a smaller subset of
4590 * memory to reclaim from. Consider this as a longer
4593 /* If excess == 0, no tree ops */
4594 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
4595 spin_unlock(&mctz
->lock
);
4596 css_put(&mz
->memcg
->css
);
4599 * Could not reclaim anything and there are no more
4600 * mem cgroups to try or we seem to be looping without
4601 * reclaiming anything.
4603 if (!nr_reclaimed
&&
4605 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
4607 } while (!nr_reclaimed
);
4609 css_put(&next_mz
->memcg
->css
);
4610 return nr_reclaimed
;
4614 * mem_cgroup_force_empty_list - clears LRU of a group
4615 * @memcg: group to clear
4618 * @lru: lru to to clear
4620 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4621 * reclaim the pages page themselves - pages are moved to the parent (or root)
4624 static void mem_cgroup_force_empty_list(struct mem_cgroup
*memcg
,
4625 int node
, int zid
, enum lru_list lru
)
4627 struct lruvec
*lruvec
;
4628 unsigned long flags
;
4629 struct list_head
*list
;
4633 zone
= &NODE_DATA(node
)->node_zones
[zid
];
4634 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
4635 list
= &lruvec
->lists
[lru
];
4639 struct page_cgroup
*pc
;
4642 spin_lock_irqsave(&zone
->lru_lock
, flags
);
4643 if (list_empty(list
)) {
4644 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4647 page
= list_entry(list
->prev
, struct page
, lru
);
4649 list_move(&page
->lru
, list
);
4651 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4654 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4656 pc
= lookup_page_cgroup(page
);
4658 if (mem_cgroup_move_parent(page
, pc
, memcg
)) {
4659 /* found lock contention or "pc" is obsolete. */
4664 } while (!list_empty(list
));
4668 * make mem_cgroup's charge to be 0 if there is no task by moving
4669 * all the charges and pages to the parent.
4670 * This enables deleting this mem_cgroup.
4672 * Caller is responsible for holding css reference on the memcg.
4674 static void mem_cgroup_reparent_charges(struct mem_cgroup
*memcg
)
4680 /* This is for making all *used* pages to be on LRU. */
4681 lru_add_drain_all();
4682 drain_all_stock_sync(memcg
);
4683 mem_cgroup_start_move(memcg
);
4684 for_each_node_state(node
, N_MEMORY
) {
4685 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
4688 mem_cgroup_force_empty_list(memcg
,
4693 mem_cgroup_end_move(memcg
);
4694 memcg_oom_recover(memcg
);
4698 * Kernel memory may not necessarily be trackable to a specific
4699 * process. So they are not migrated, and therefore we can't
4700 * expect their value to drop to 0 here.
4701 * Having res filled up with kmem only is enough.
4703 * This is a safety check because mem_cgroup_force_empty_list
4704 * could have raced with mem_cgroup_replace_page_cache callers
4705 * so the lru seemed empty but the page could have been added
4706 * right after the check. RES_USAGE should be safe as we always
4707 * charge before adding to the LRU.
4709 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
) -
4710 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
);
4711 } while (usage
> 0);
4714 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
4716 lockdep_assert_held(&memcg_create_mutex
);
4718 * The lock does not prevent addition or deletion to the list
4719 * of children, but it prevents a new child from being
4720 * initialized based on this parent in css_online(), so it's
4721 * enough to decide whether hierarchically inherited
4722 * attributes can still be changed or not.
4724 return memcg
->use_hierarchy
&&
4725 !list_empty(&memcg
->css
.cgroup
->children
);
4729 * Reclaims as many pages from the given memcg as possible and moves
4730 * the rest to the parent.
4732 * Caller is responsible for holding css reference for memcg.
4734 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
4736 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
4737 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
4739 /* returns EBUSY if there is a task or if we come here twice. */
4740 if (cgroup_has_tasks(cgrp
) || !list_empty(&cgrp
->children
))
4743 /* we call try-to-free pages for make this cgroup empty */
4744 lru_add_drain_all();
4745 /* try to free all pages in this cgroup */
4746 while (nr_retries
&& res_counter_read_u64(&memcg
->res
, RES_USAGE
) > 0) {
4749 if (signal_pending(current
))
4752 progress
= try_to_free_mem_cgroup_pages(memcg
, GFP_KERNEL
,
4756 /* maybe some writeback is necessary */
4757 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4762 mem_cgroup_reparent_charges(memcg
);
4767 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state
*css
,
4770 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4772 if (mem_cgroup_is_root(memcg
))
4774 return mem_cgroup_force_empty(memcg
);
4777 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
4780 return mem_cgroup_from_css(css
)->use_hierarchy
;
4783 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
4784 struct cftype
*cft
, u64 val
)
4787 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4788 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
4790 mutex_lock(&memcg_create_mutex
);
4792 if (memcg
->use_hierarchy
== val
)
4796 * If parent's use_hierarchy is set, we can't make any modifications
4797 * in the child subtrees. If it is unset, then the change can
4798 * occur, provided the current cgroup has no children.
4800 * For the root cgroup, parent_mem is NULL, we allow value to be
4801 * set if there are no children.
4803 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
4804 (val
== 1 || val
== 0)) {
4805 if (list_empty(&memcg
->css
.cgroup
->children
))
4806 memcg
->use_hierarchy
= val
;
4813 mutex_unlock(&memcg_create_mutex
);
4819 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup
*memcg
,
4820 enum mem_cgroup_stat_index idx
)
4822 struct mem_cgroup
*iter
;
4825 /* Per-cpu values can be negative, use a signed accumulator */
4826 for_each_mem_cgroup_tree(iter
, memcg
)
4827 val
+= mem_cgroup_read_stat(iter
, idx
);
4829 if (val
< 0) /* race ? */
4834 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
4838 if (!mem_cgroup_is_root(memcg
)) {
4840 return res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4842 return res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4846 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
4847 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
4849 val
= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
4850 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_RSS
);
4853 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
4855 return val
<< PAGE_SHIFT
;
4858 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
4861 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4866 type
= MEMFILE_TYPE(cft
->private);
4867 name
= MEMFILE_ATTR(cft
->private);
4871 if (name
== RES_USAGE
)
4872 val
= mem_cgroup_usage(memcg
, false);
4874 val
= res_counter_read_u64(&memcg
->res
, name
);
4877 if (name
== RES_USAGE
)
4878 val
= mem_cgroup_usage(memcg
, true);
4880 val
= res_counter_read_u64(&memcg
->memsw
, name
);
4883 val
= res_counter_read_u64(&memcg
->kmem
, name
);
4892 #ifdef CONFIG_MEMCG_KMEM
4893 /* should be called with activate_kmem_mutex held */
4894 static int __memcg_activate_kmem(struct mem_cgroup
*memcg
,
4895 unsigned long long limit
)
4900 if (memcg_kmem_is_active(memcg
))
4904 * We are going to allocate memory for data shared by all memory
4905 * cgroups so let's stop accounting here.
4907 memcg_stop_kmem_account();
4910 * For simplicity, we won't allow this to be disabled. It also can't
4911 * be changed if the cgroup has children already, or if tasks had
4914 * If tasks join before we set the limit, a person looking at
4915 * kmem.usage_in_bytes will have no way to determine when it took
4916 * place, which makes the value quite meaningless.
4918 * After it first became limited, changes in the value of the limit are
4919 * of course permitted.
4921 mutex_lock(&memcg_create_mutex
);
4922 if (cgroup_has_tasks(memcg
->css
.cgroup
) || memcg_has_children(memcg
))
4924 mutex_unlock(&memcg_create_mutex
);
4928 memcg_id
= ida_simple_get(&kmem_limited_groups
,
4929 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
4936 * Make sure we have enough space for this cgroup in each root cache's
4939 mutex_lock(&memcg_slab_mutex
);
4940 err
= memcg_update_all_caches(memcg_id
+ 1);
4941 mutex_unlock(&memcg_slab_mutex
);
4945 memcg
->kmemcg_id
= memcg_id
;
4946 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
4949 * We couldn't have accounted to this cgroup, because it hasn't got the
4950 * active bit set yet, so this should succeed.
4952 err
= res_counter_set_limit(&memcg
->kmem
, limit
);
4955 static_key_slow_inc(&memcg_kmem_enabled_key
);
4957 * Setting the active bit after enabling static branching will
4958 * guarantee no one starts accounting before all call sites are
4961 memcg_kmem_set_active(memcg
);
4963 memcg_resume_kmem_account();
4967 ida_simple_remove(&kmem_limited_groups
, memcg_id
);
4971 static int memcg_activate_kmem(struct mem_cgroup
*memcg
,
4972 unsigned long long limit
)
4976 mutex_lock(&activate_kmem_mutex
);
4977 ret
= __memcg_activate_kmem(memcg
, limit
);
4978 mutex_unlock(&activate_kmem_mutex
);
4982 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
4983 unsigned long long val
)
4987 if (!memcg_kmem_is_active(memcg
))
4988 ret
= memcg_activate_kmem(memcg
, val
);
4990 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
4994 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
4997 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
5002 mutex_lock(&activate_kmem_mutex
);
5004 * If the parent cgroup is not kmem-active now, it cannot be activated
5005 * after this point, because it has at least one child already.
5007 if (memcg_kmem_is_active(parent
))
5008 ret
= __memcg_activate_kmem(memcg
, RES_COUNTER_MAX
);
5009 mutex_unlock(&activate_kmem_mutex
);
5013 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
5014 unsigned long long val
)
5018 #endif /* CONFIG_MEMCG_KMEM */
5021 * The user of this function is...
5024 static int mem_cgroup_write(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
5027 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5030 unsigned long long val
;
5033 type
= MEMFILE_TYPE(cft
->private);
5034 name
= MEMFILE_ATTR(cft
->private);
5038 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
5042 /* This function does all necessary parse...reuse it */
5043 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5047 ret
= mem_cgroup_resize_limit(memcg
, val
);
5048 else if (type
== _MEMSWAP
)
5049 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
5050 else if (type
== _KMEM
)
5051 ret
= memcg_update_kmem_limit(memcg
, val
);
5055 case RES_SOFT_LIMIT
:
5056 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5060 * For memsw, soft limits are hard to implement in terms
5061 * of semantics, for now, we support soft limits for
5062 * control without swap
5065 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
5070 ret
= -EINVAL
; /* should be BUG() ? */
5076 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
5077 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
5079 unsigned long long min_limit
, min_memsw_limit
, tmp
;
5081 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5082 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5083 if (!memcg
->use_hierarchy
)
5086 while (css_parent(&memcg
->css
)) {
5087 memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5088 if (!memcg
->use_hierarchy
)
5090 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5091 min_limit
= min(min_limit
, tmp
);
5092 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5093 min_memsw_limit
= min(min_memsw_limit
, tmp
);
5096 *mem_limit
= min_limit
;
5097 *memsw_limit
= min_memsw_limit
;
5100 static int mem_cgroup_reset(struct cgroup_subsys_state
*css
, unsigned int event
)
5102 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5106 type
= MEMFILE_TYPE(event
);
5107 name
= MEMFILE_ATTR(event
);
5112 res_counter_reset_max(&memcg
->res
);
5113 else if (type
== _MEMSWAP
)
5114 res_counter_reset_max(&memcg
->memsw
);
5115 else if (type
== _KMEM
)
5116 res_counter_reset_max(&memcg
->kmem
);
5122 res_counter_reset_failcnt(&memcg
->res
);
5123 else if (type
== _MEMSWAP
)
5124 res_counter_reset_failcnt(&memcg
->memsw
);
5125 else if (type
== _KMEM
)
5126 res_counter_reset_failcnt(&memcg
->kmem
);
5135 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
5138 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
5142 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5143 struct cftype
*cft
, u64 val
)
5145 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5147 if (val
>= (1 << NR_MOVE_TYPE
))
5151 * No kind of locking is needed in here, because ->can_attach() will
5152 * check this value once in the beginning of the process, and then carry
5153 * on with stale data. This means that changes to this value will only
5154 * affect task migrations starting after the change.
5156 memcg
->move_charge_at_immigrate
= val
;
5160 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5161 struct cftype
*cft
, u64 val
)
5168 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
5172 unsigned int lru_mask
;
5175 static const struct numa_stat stats
[] = {
5176 { "total", LRU_ALL
},
5177 { "file", LRU_ALL_FILE
},
5178 { "anon", LRU_ALL_ANON
},
5179 { "unevictable", BIT(LRU_UNEVICTABLE
) },
5181 const struct numa_stat
*stat
;
5184 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5186 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
5187 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
5188 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
5189 for_each_node_state(nid
, N_MEMORY
) {
5190 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5192 seq_printf(m
, " N%d=%lu", nid
, nr
);
5197 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
5198 struct mem_cgroup
*iter
;
5201 for_each_mem_cgroup_tree(iter
, memcg
)
5202 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
5203 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
5204 for_each_node_state(nid
, N_MEMORY
) {
5206 for_each_mem_cgroup_tree(iter
, memcg
)
5207 nr
+= mem_cgroup_node_nr_lru_pages(
5208 iter
, nid
, stat
->lru_mask
);
5209 seq_printf(m
, " N%d=%lu", nid
, nr
);
5216 #endif /* CONFIG_NUMA */
5218 static inline void mem_cgroup_lru_names_not_uptodate(void)
5220 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
5223 static int memcg_stat_show(struct seq_file
*m
, void *v
)
5225 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5226 struct mem_cgroup
*mi
;
5229 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5230 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5232 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
5233 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
5236 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
5237 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
5238 mem_cgroup_read_events(memcg
, i
));
5240 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
5241 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
5242 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
5244 /* Hierarchical information */
5246 unsigned long long limit
, memsw_limit
;
5247 memcg_get_hierarchical_limit(memcg
, &limit
, &memsw_limit
);
5248 seq_printf(m
, "hierarchical_memory_limit %llu\n", limit
);
5249 if (do_swap_account
)
5250 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
5254 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5257 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5259 for_each_mem_cgroup_tree(mi
, memcg
)
5260 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
5261 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
5264 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
5265 unsigned long long val
= 0;
5267 for_each_mem_cgroup_tree(mi
, memcg
)
5268 val
+= mem_cgroup_read_events(mi
, i
);
5269 seq_printf(m
, "total_%s %llu\n",
5270 mem_cgroup_events_names
[i
], val
);
5273 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5274 unsigned long long val
= 0;
5276 for_each_mem_cgroup_tree(mi
, memcg
)
5277 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
5278 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
5281 #ifdef CONFIG_DEBUG_VM
5284 struct mem_cgroup_per_zone
*mz
;
5285 struct zone_reclaim_stat
*rstat
;
5286 unsigned long recent_rotated
[2] = {0, 0};
5287 unsigned long recent_scanned
[2] = {0, 0};
5289 for_each_online_node(nid
)
5290 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
5291 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
5292 rstat
= &mz
->lruvec
.reclaim_stat
;
5294 recent_rotated
[0] += rstat
->recent_rotated
[0];
5295 recent_rotated
[1] += rstat
->recent_rotated
[1];
5296 recent_scanned
[0] += rstat
->recent_scanned
[0];
5297 recent_scanned
[1] += rstat
->recent_scanned
[1];
5299 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
5300 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
5301 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
5302 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
5309 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
5312 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5314 return mem_cgroup_swappiness(memcg
);
5317 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
5318 struct cftype
*cft
, u64 val
)
5320 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5325 if (css_parent(css
))
5326 memcg
->swappiness
= val
;
5328 vm_swappiness
= val
;
5333 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
5335 struct mem_cgroup_threshold_ary
*t
;
5341 t
= rcu_dereference(memcg
->thresholds
.primary
);
5343 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
5348 usage
= mem_cgroup_usage(memcg
, swap
);
5351 * current_threshold points to threshold just below or equal to usage.
5352 * If it's not true, a threshold was crossed after last
5353 * call of __mem_cgroup_threshold().
5355 i
= t
->current_threshold
;
5358 * Iterate backward over array of thresholds starting from
5359 * current_threshold and check if a threshold is crossed.
5360 * If none of thresholds below usage is crossed, we read
5361 * only one element of the array here.
5363 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
5364 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5366 /* i = current_threshold + 1 */
5370 * Iterate forward over array of thresholds starting from
5371 * current_threshold+1 and check if a threshold is crossed.
5372 * If none of thresholds above usage is crossed, we read
5373 * only one element of the array here.
5375 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
5376 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5378 /* Update current_threshold */
5379 t
->current_threshold
= i
- 1;
5384 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
5387 __mem_cgroup_threshold(memcg
, false);
5388 if (do_swap_account
)
5389 __mem_cgroup_threshold(memcg
, true);
5391 memcg
= parent_mem_cgroup(memcg
);
5395 static int compare_thresholds(const void *a
, const void *b
)
5397 const struct mem_cgroup_threshold
*_a
= a
;
5398 const struct mem_cgroup_threshold
*_b
= b
;
5400 if (_a
->threshold
> _b
->threshold
)
5403 if (_a
->threshold
< _b
->threshold
)
5409 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
5411 struct mem_cgroup_eventfd_list
*ev
;
5413 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
5414 eventfd_signal(ev
->eventfd
, 1);
5418 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
5420 struct mem_cgroup
*iter
;
5422 for_each_mem_cgroup_tree(iter
, memcg
)
5423 mem_cgroup_oom_notify_cb(iter
);
5426 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
5427 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
5429 struct mem_cgroup_thresholds
*thresholds
;
5430 struct mem_cgroup_threshold_ary
*new;
5431 u64 threshold
, usage
;
5434 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
5438 mutex_lock(&memcg
->thresholds_lock
);
5441 thresholds
= &memcg
->thresholds
;
5442 else if (type
== _MEMSWAP
)
5443 thresholds
= &memcg
->memsw_thresholds
;
5447 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5449 /* Check if a threshold crossed before adding a new one */
5450 if (thresholds
->primary
)
5451 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5453 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
5455 /* Allocate memory for new array of thresholds */
5456 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
5464 /* Copy thresholds (if any) to new array */
5465 if (thresholds
->primary
) {
5466 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
5467 sizeof(struct mem_cgroup_threshold
));
5470 /* Add new threshold */
5471 new->entries
[size
- 1].eventfd
= eventfd
;
5472 new->entries
[size
- 1].threshold
= threshold
;
5474 /* Sort thresholds. Registering of new threshold isn't time-critical */
5475 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
5476 compare_thresholds
, NULL
);
5478 /* Find current threshold */
5479 new->current_threshold
= -1;
5480 for (i
= 0; i
< size
; i
++) {
5481 if (new->entries
[i
].threshold
<= usage
) {
5483 * new->current_threshold will not be used until
5484 * rcu_assign_pointer(), so it's safe to increment
5487 ++new->current_threshold
;
5492 /* Free old spare buffer and save old primary buffer as spare */
5493 kfree(thresholds
->spare
);
5494 thresholds
->spare
= thresholds
->primary
;
5496 rcu_assign_pointer(thresholds
->primary
, new);
5498 /* To be sure that nobody uses thresholds */
5502 mutex_unlock(&memcg
->thresholds_lock
);
5507 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
5508 struct eventfd_ctx
*eventfd
, const char *args
)
5510 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
5513 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
5514 struct eventfd_ctx
*eventfd
, const char *args
)
5516 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
5519 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
5520 struct eventfd_ctx
*eventfd
, enum res_type type
)
5522 struct mem_cgroup_thresholds
*thresholds
;
5523 struct mem_cgroup_threshold_ary
*new;
5527 mutex_lock(&memcg
->thresholds_lock
);
5529 thresholds
= &memcg
->thresholds
;
5530 else if (type
== _MEMSWAP
)
5531 thresholds
= &memcg
->memsw_thresholds
;
5535 if (!thresholds
->primary
)
5538 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5540 /* Check if a threshold crossed before removing */
5541 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5543 /* Calculate new number of threshold */
5545 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
5546 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
5550 new = thresholds
->spare
;
5552 /* Set thresholds array to NULL if we don't have thresholds */
5561 /* Copy thresholds and find current threshold */
5562 new->current_threshold
= -1;
5563 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
5564 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
5567 new->entries
[j
] = thresholds
->primary
->entries
[i
];
5568 if (new->entries
[j
].threshold
<= usage
) {
5570 * new->current_threshold will not be used
5571 * until rcu_assign_pointer(), so it's safe to increment
5574 ++new->current_threshold
;
5580 /* Swap primary and spare array */
5581 thresholds
->spare
= thresholds
->primary
;
5582 /* If all events are unregistered, free the spare array */
5584 kfree(thresholds
->spare
);
5585 thresholds
->spare
= NULL
;
5588 rcu_assign_pointer(thresholds
->primary
, new);
5590 /* To be sure that nobody uses thresholds */
5593 mutex_unlock(&memcg
->thresholds_lock
);
5596 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
5597 struct eventfd_ctx
*eventfd
)
5599 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
5602 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
5603 struct eventfd_ctx
*eventfd
)
5605 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
5608 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
5609 struct eventfd_ctx
*eventfd
, const char *args
)
5611 struct mem_cgroup_eventfd_list
*event
;
5613 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
5617 spin_lock(&memcg_oom_lock
);
5619 event
->eventfd
= eventfd
;
5620 list_add(&event
->list
, &memcg
->oom_notify
);
5622 /* already in OOM ? */
5623 if (atomic_read(&memcg
->under_oom
))
5624 eventfd_signal(eventfd
, 1);
5625 spin_unlock(&memcg_oom_lock
);
5630 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
5631 struct eventfd_ctx
*eventfd
)
5633 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
5635 spin_lock(&memcg_oom_lock
);
5637 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
5638 if (ev
->eventfd
== eventfd
) {
5639 list_del(&ev
->list
);
5644 spin_unlock(&memcg_oom_lock
);
5647 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
5649 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
5651 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
5652 seq_printf(sf
, "under_oom %d\n", (bool)atomic_read(&memcg
->under_oom
));
5656 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
5657 struct cftype
*cft
, u64 val
)
5659 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5661 /* cannot set to root cgroup and only 0 and 1 are allowed */
5662 if (!css_parent(css
) || !((val
== 0) || (val
== 1)))
5665 memcg
->oom_kill_disable
= val
;
5667 memcg_oom_recover(memcg
);
5672 #ifdef CONFIG_MEMCG_KMEM
5673 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5677 memcg
->kmemcg_id
= -1;
5678 ret
= memcg_propagate_kmem(memcg
);
5682 return mem_cgroup_sockets_init(memcg
, ss
);
5685 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5687 mem_cgroup_sockets_destroy(memcg
);
5690 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5692 if (!memcg_kmem_is_active(memcg
))
5696 * kmem charges can outlive the cgroup. In the case of slab
5697 * pages, for instance, a page contain objects from various
5698 * processes. As we prevent from taking a reference for every
5699 * such allocation we have to be careful when doing uncharge
5700 * (see memcg_uncharge_kmem) and here during offlining.
5702 * The idea is that that only the _last_ uncharge which sees
5703 * the dead memcg will drop the last reference. An additional
5704 * reference is taken here before the group is marked dead
5705 * which is then paired with css_put during uncharge resp. here.
5707 * Although this might sound strange as this path is called from
5708 * css_offline() when the referencemight have dropped down to 0
5709 * and shouldn't be incremented anymore (css_tryget would fail)
5710 * we do not have other options because of the kmem allocations
5713 css_get(&memcg
->css
);
5715 memcg_kmem_mark_dead(memcg
);
5717 if (res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0)
5720 if (memcg_kmem_test_and_clear_dead(memcg
))
5721 css_put(&memcg
->css
);
5724 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5729 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5733 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5739 * DO NOT USE IN NEW FILES.
5741 * "cgroup.event_control" implementation.
5743 * This is way over-engineered. It tries to support fully configurable
5744 * events for each user. Such level of flexibility is completely
5745 * unnecessary especially in the light of the planned unified hierarchy.
5747 * Please deprecate this and replace with something simpler if at all
5752 * Unregister event and free resources.
5754 * Gets called from workqueue.
5756 static void memcg_event_remove(struct work_struct
*work
)
5758 struct mem_cgroup_event
*event
=
5759 container_of(work
, struct mem_cgroup_event
, remove
);
5760 struct mem_cgroup
*memcg
= event
->memcg
;
5762 remove_wait_queue(event
->wqh
, &event
->wait
);
5764 event
->unregister_event(memcg
, event
->eventfd
);
5766 /* Notify userspace the event is going away. */
5767 eventfd_signal(event
->eventfd
, 1);
5769 eventfd_ctx_put(event
->eventfd
);
5771 css_put(&memcg
->css
);
5775 * Gets called on POLLHUP on eventfd when user closes it.
5777 * Called with wqh->lock held and interrupts disabled.
5779 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
5780 int sync
, void *key
)
5782 struct mem_cgroup_event
*event
=
5783 container_of(wait
, struct mem_cgroup_event
, wait
);
5784 struct mem_cgroup
*memcg
= event
->memcg
;
5785 unsigned long flags
= (unsigned long)key
;
5787 if (flags
& POLLHUP
) {
5789 * If the event has been detached at cgroup removal, we
5790 * can simply return knowing the other side will cleanup
5793 * We can't race against event freeing since the other
5794 * side will require wqh->lock via remove_wait_queue(),
5797 spin_lock(&memcg
->event_list_lock
);
5798 if (!list_empty(&event
->list
)) {
5799 list_del_init(&event
->list
);
5801 * We are in atomic context, but cgroup_event_remove()
5802 * may sleep, so we have to call it in workqueue.
5804 schedule_work(&event
->remove
);
5806 spin_unlock(&memcg
->event_list_lock
);
5812 static void memcg_event_ptable_queue_proc(struct file
*file
,
5813 wait_queue_head_t
*wqh
, poll_table
*pt
)
5815 struct mem_cgroup_event
*event
=
5816 container_of(pt
, struct mem_cgroup_event
, pt
);
5819 add_wait_queue(wqh
, &event
->wait
);
5823 * DO NOT USE IN NEW FILES.
5825 * Parse input and register new cgroup event handler.
5827 * Input must be in format '<event_fd> <control_fd> <args>'.
5828 * Interpretation of args is defined by control file implementation.
5830 static int memcg_write_event_control(struct cgroup_subsys_state
*css
,
5831 struct cftype
*cft
, char *buffer
)
5833 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5834 struct mem_cgroup_event
*event
;
5835 struct cgroup_subsys_state
*cfile_css
;
5836 unsigned int efd
, cfd
;
5843 efd
= simple_strtoul(buffer
, &endp
, 10);
5848 cfd
= simple_strtoul(buffer
, &endp
, 10);
5849 if ((*endp
!= ' ') && (*endp
!= '\0'))
5853 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
5857 event
->memcg
= memcg
;
5858 INIT_LIST_HEAD(&event
->list
);
5859 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
5860 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
5861 INIT_WORK(&event
->remove
, memcg_event_remove
);
5869 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
5870 if (IS_ERR(event
->eventfd
)) {
5871 ret
= PTR_ERR(event
->eventfd
);
5878 goto out_put_eventfd
;
5881 /* the process need read permission on control file */
5882 /* AV: shouldn't we check that it's been opened for read instead? */
5883 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
5888 * Determine the event callbacks and set them in @event. This used
5889 * to be done via struct cftype but cgroup core no longer knows
5890 * about these events. The following is crude but the whole thing
5891 * is for compatibility anyway.
5893 * DO NOT ADD NEW FILES.
5895 name
= cfile
.file
->f_dentry
->d_name
.name
;
5897 if (!strcmp(name
, "memory.usage_in_bytes")) {
5898 event
->register_event
= mem_cgroup_usage_register_event
;
5899 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
5900 } else if (!strcmp(name
, "memory.oom_control")) {
5901 event
->register_event
= mem_cgroup_oom_register_event
;
5902 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
5903 } else if (!strcmp(name
, "memory.pressure_level")) {
5904 event
->register_event
= vmpressure_register_event
;
5905 event
->unregister_event
= vmpressure_unregister_event
;
5906 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
5907 event
->register_event
= memsw_cgroup_usage_register_event
;
5908 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
5915 * Verify @cfile should belong to @css. Also, remaining events are
5916 * automatically removed on cgroup destruction but the removal is
5917 * asynchronous, so take an extra ref on @css.
5919 cfile_css
= css_tryget_from_dir(cfile
.file
->f_dentry
->d_parent
,
5920 &memory_cgrp_subsys
);
5922 if (IS_ERR(cfile_css
))
5924 if (cfile_css
!= css
) {
5929 ret
= event
->register_event(memcg
, event
->eventfd
, buffer
);
5933 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
5935 spin_lock(&memcg
->event_list_lock
);
5936 list_add(&event
->list
, &memcg
->event_list
);
5937 spin_unlock(&memcg
->event_list_lock
);
5949 eventfd_ctx_put(event
->eventfd
);
5958 static struct cftype mem_cgroup_files
[] = {
5960 .name
= "usage_in_bytes",
5961 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
5962 .read_u64
= mem_cgroup_read_u64
,
5965 .name
= "max_usage_in_bytes",
5966 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
5967 .trigger
= mem_cgroup_reset
,
5968 .read_u64
= mem_cgroup_read_u64
,
5971 .name
= "limit_in_bytes",
5972 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
5973 .write_string
= mem_cgroup_write
,
5974 .read_u64
= mem_cgroup_read_u64
,
5977 .name
= "soft_limit_in_bytes",
5978 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
5979 .write_string
= mem_cgroup_write
,
5980 .read_u64
= mem_cgroup_read_u64
,
5984 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
5985 .trigger
= mem_cgroup_reset
,
5986 .read_u64
= mem_cgroup_read_u64
,
5990 .seq_show
= memcg_stat_show
,
5993 .name
= "force_empty",
5994 .trigger
= mem_cgroup_force_empty_write
,
5997 .name
= "use_hierarchy",
5998 .flags
= CFTYPE_INSANE
,
5999 .write_u64
= mem_cgroup_hierarchy_write
,
6000 .read_u64
= mem_cgroup_hierarchy_read
,
6003 .name
= "cgroup.event_control", /* XXX: for compat */
6004 .write_string
= memcg_write_event_control
,
6005 .flags
= CFTYPE_NO_PREFIX
,
6009 .name
= "swappiness",
6010 .read_u64
= mem_cgroup_swappiness_read
,
6011 .write_u64
= mem_cgroup_swappiness_write
,
6014 .name
= "move_charge_at_immigrate",
6015 .read_u64
= mem_cgroup_move_charge_read
,
6016 .write_u64
= mem_cgroup_move_charge_write
,
6019 .name
= "oom_control",
6020 .seq_show
= mem_cgroup_oom_control_read
,
6021 .write_u64
= mem_cgroup_oom_control_write
,
6022 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
6025 .name
= "pressure_level",
6029 .name
= "numa_stat",
6030 .seq_show
= memcg_numa_stat_show
,
6033 #ifdef CONFIG_MEMCG_KMEM
6035 .name
= "kmem.limit_in_bytes",
6036 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
6037 .write_string
= mem_cgroup_write
,
6038 .read_u64
= mem_cgroup_read_u64
,
6041 .name
= "kmem.usage_in_bytes",
6042 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
6043 .read_u64
= mem_cgroup_read_u64
,
6046 .name
= "kmem.failcnt",
6047 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
6048 .trigger
= mem_cgroup_reset
,
6049 .read_u64
= mem_cgroup_read_u64
,
6052 .name
= "kmem.max_usage_in_bytes",
6053 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
6054 .trigger
= mem_cgroup_reset
,
6055 .read_u64
= mem_cgroup_read_u64
,
6057 #ifdef CONFIG_SLABINFO
6059 .name
= "kmem.slabinfo",
6060 .seq_show
= mem_cgroup_slabinfo_read
,
6064 { }, /* terminate */
6067 #ifdef CONFIG_MEMCG_SWAP
6068 static struct cftype memsw_cgroup_files
[] = {
6070 .name
= "memsw.usage_in_bytes",
6071 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
6072 .read_u64
= mem_cgroup_read_u64
,
6075 .name
= "memsw.max_usage_in_bytes",
6076 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
6077 .trigger
= mem_cgroup_reset
,
6078 .read_u64
= mem_cgroup_read_u64
,
6081 .name
= "memsw.limit_in_bytes",
6082 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
6083 .write_string
= mem_cgroup_write
,
6084 .read_u64
= mem_cgroup_read_u64
,
6087 .name
= "memsw.failcnt",
6088 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
6089 .trigger
= mem_cgroup_reset
,
6090 .read_u64
= mem_cgroup_read_u64
,
6092 { }, /* terminate */
6095 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6097 struct mem_cgroup_per_node
*pn
;
6098 struct mem_cgroup_per_zone
*mz
;
6099 int zone
, tmp
= node
;
6101 * This routine is called against possible nodes.
6102 * But it's BUG to call kmalloc() against offline node.
6104 * TODO: this routine can waste much memory for nodes which will
6105 * never be onlined. It's better to use memory hotplug callback
6108 if (!node_state(node
, N_NORMAL_MEMORY
))
6110 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
6114 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6115 mz
= &pn
->zoneinfo
[zone
];
6116 lruvec_init(&mz
->lruvec
);
6117 mz
->usage_in_excess
= 0;
6118 mz
->on_tree
= false;
6121 memcg
->nodeinfo
[node
] = pn
;
6125 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6127 kfree(memcg
->nodeinfo
[node
]);
6130 static struct mem_cgroup
*mem_cgroup_alloc(void)
6132 struct mem_cgroup
*memcg
;
6135 size
= sizeof(struct mem_cgroup
);
6136 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
6138 memcg
= kzalloc(size
, GFP_KERNEL
);
6142 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
6145 spin_lock_init(&memcg
->pcp_counter_lock
);
6154 * At destroying mem_cgroup, references from swap_cgroup can remain.
6155 * (scanning all at force_empty is too costly...)
6157 * Instead of clearing all references at force_empty, we remember
6158 * the number of reference from swap_cgroup and free mem_cgroup when
6159 * it goes down to 0.
6161 * Removal of cgroup itself succeeds regardless of refs from swap.
6164 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
6168 mem_cgroup_remove_from_trees(memcg
);
6171 free_mem_cgroup_per_zone_info(memcg
, node
);
6173 free_percpu(memcg
->stat
);
6176 * We need to make sure that (at least for now), the jump label
6177 * destruction code runs outside of the cgroup lock. This is because
6178 * get_online_cpus(), which is called from the static_branch update,
6179 * can't be called inside the cgroup_lock. cpusets are the ones
6180 * enforcing this dependency, so if they ever change, we might as well.
6182 * schedule_work() will guarantee this happens. Be careful if you need
6183 * to move this code around, and make sure it is outside
6186 disarm_static_keys(memcg
);
6191 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6193 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
6195 if (!memcg
->res
.parent
)
6197 return mem_cgroup_from_res_counter(memcg
->res
.parent
, res
);
6199 EXPORT_SYMBOL(parent_mem_cgroup
);
6201 static void __init
mem_cgroup_soft_limit_tree_init(void)
6203 struct mem_cgroup_tree_per_node
*rtpn
;
6204 struct mem_cgroup_tree_per_zone
*rtpz
;
6205 int tmp
, node
, zone
;
6207 for_each_node(node
) {
6209 if (!node_state(node
, N_NORMAL_MEMORY
))
6211 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
6214 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
6216 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6217 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
6218 rtpz
->rb_root
= RB_ROOT
;
6219 spin_lock_init(&rtpz
->lock
);
6224 static struct cgroup_subsys_state
* __ref
6225 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
6227 struct mem_cgroup
*memcg
;
6228 long error
= -ENOMEM
;
6231 memcg
= mem_cgroup_alloc();
6233 return ERR_PTR(error
);
6236 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
6240 if (parent_css
== NULL
) {
6241 root_mem_cgroup
= memcg
;
6242 res_counter_init(&memcg
->res
, NULL
);
6243 res_counter_init(&memcg
->memsw
, NULL
);
6244 res_counter_init(&memcg
->kmem
, NULL
);
6247 memcg
->last_scanned_node
= MAX_NUMNODES
;
6248 INIT_LIST_HEAD(&memcg
->oom_notify
);
6249 memcg
->move_charge_at_immigrate
= 0;
6250 mutex_init(&memcg
->thresholds_lock
);
6251 spin_lock_init(&memcg
->move_lock
);
6252 vmpressure_init(&memcg
->vmpressure
);
6253 INIT_LIST_HEAD(&memcg
->event_list
);
6254 spin_lock_init(&memcg
->event_list_lock
);
6259 __mem_cgroup_free(memcg
);
6260 return ERR_PTR(error
);
6264 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
6266 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6267 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(css
));
6269 if (css
->cgroup
->id
> MEM_CGROUP_ID_MAX
)
6275 mutex_lock(&memcg_create_mutex
);
6277 memcg
->use_hierarchy
= parent
->use_hierarchy
;
6278 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
6279 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
6281 if (parent
->use_hierarchy
) {
6282 res_counter_init(&memcg
->res
, &parent
->res
);
6283 res_counter_init(&memcg
->memsw
, &parent
->memsw
);
6284 res_counter_init(&memcg
->kmem
, &parent
->kmem
);
6287 * No need to take a reference to the parent because cgroup
6288 * core guarantees its existence.
6291 res_counter_init(&memcg
->res
, NULL
);
6292 res_counter_init(&memcg
->memsw
, NULL
);
6293 res_counter_init(&memcg
->kmem
, NULL
);
6295 * Deeper hierachy with use_hierarchy == false doesn't make
6296 * much sense so let cgroup subsystem know about this
6297 * unfortunate state in our controller.
6299 if (parent
!= root_mem_cgroup
)
6300 memory_cgrp_subsys
.broken_hierarchy
= true;
6302 mutex_unlock(&memcg_create_mutex
);
6304 return memcg_init_kmem(memcg
, &memory_cgrp_subsys
);
6308 * Announce all parents that a group from their hierarchy is gone.
6310 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup
*memcg
)
6312 struct mem_cgroup
*parent
= memcg
;
6314 while ((parent
= parent_mem_cgroup(parent
)))
6315 mem_cgroup_iter_invalidate(parent
);
6318 * if the root memcg is not hierarchical we have to check it
6321 if (!root_mem_cgroup
->use_hierarchy
)
6322 mem_cgroup_iter_invalidate(root_mem_cgroup
);
6325 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
6327 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6328 struct mem_cgroup_event
*event
, *tmp
;
6329 struct cgroup_subsys_state
*iter
;
6332 * Unregister events and notify userspace.
6333 * Notify userspace about cgroup removing only after rmdir of cgroup
6334 * directory to avoid race between userspace and kernelspace.
6336 spin_lock(&memcg
->event_list_lock
);
6337 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
6338 list_del_init(&event
->list
);
6339 schedule_work(&event
->remove
);
6341 spin_unlock(&memcg
->event_list_lock
);
6343 kmem_cgroup_css_offline(memcg
);
6345 mem_cgroup_invalidate_reclaim_iterators(memcg
);
6348 * This requires that offlining is serialized. Right now that is
6349 * guaranteed because css_killed_work_fn() holds the cgroup_mutex.
6351 css_for_each_descendant_post(iter
, css
)
6352 mem_cgroup_reparent_charges(mem_cgroup_from_css(iter
));
6354 memcg_unregister_all_caches(memcg
);
6355 vmpressure_cleanup(&memcg
->vmpressure
);
6358 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
6360 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6362 * XXX: css_offline() would be where we should reparent all
6363 * memory to prepare the cgroup for destruction. However,
6364 * memcg does not do css_tryget() and res_counter charging
6365 * under the same RCU lock region, which means that charging
6366 * could race with offlining. Offlining only happens to
6367 * cgroups with no tasks in them but charges can show up
6368 * without any tasks from the swapin path when the target
6369 * memcg is looked up from the swapout record and not from the
6370 * current task as it usually is. A race like this can leak
6371 * charges and put pages with stale cgroup pointers into
6375 * lookup_swap_cgroup_id()
6377 * mem_cgroup_lookup()
6380 * disable css_tryget()
6383 * reparent_charges()
6384 * res_counter_charge()
6387 * pc->mem_cgroup = dead memcg
6390 * The bulk of the charges are still moved in offline_css() to
6391 * avoid pinning a lot of pages in case a long-term reference
6392 * like a swapout record is deferring the css_free() to long
6393 * after offlining. But this makes sure we catch any charges
6394 * made after offlining:
6396 mem_cgroup_reparent_charges(memcg
);
6398 memcg_destroy_kmem(memcg
);
6399 __mem_cgroup_free(memcg
);
6403 /* Handlers for move charge at task migration. */
6404 #define PRECHARGE_COUNT_AT_ONCE 256
6405 static int mem_cgroup_do_precharge(unsigned long count
)
6408 int batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6409 struct mem_cgroup
*memcg
= mc
.to
;
6411 if (mem_cgroup_is_root(memcg
)) {
6412 mc
.precharge
+= count
;
6413 /* we don't need css_get for root */
6416 /* try to charge at once */
6418 struct res_counter
*dummy
;
6420 * "memcg" cannot be under rmdir() because we've already checked
6421 * by cgroup_lock_live_cgroup() that it is not removed and we
6422 * are still under the same cgroup_mutex. So we can postpone
6425 if (res_counter_charge(&memcg
->res
, PAGE_SIZE
* count
, &dummy
))
6427 if (do_swap_account
&& res_counter_charge(&memcg
->memsw
,
6428 PAGE_SIZE
* count
, &dummy
)) {
6429 res_counter_uncharge(&memcg
->res
, PAGE_SIZE
* count
);
6432 mc
.precharge
+= count
;
6436 /* fall back to one by one charge */
6438 if (signal_pending(current
)) {
6442 if (!batch_count
--) {
6443 batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6446 ret
= mem_cgroup_try_charge(memcg
, GFP_KERNEL
, 1, false);
6448 /* mem_cgroup_clear_mc() will do uncharge later */
6456 * get_mctgt_type - get target type of moving charge
6457 * @vma: the vma the pte to be checked belongs
6458 * @addr: the address corresponding to the pte to be checked
6459 * @ptent: the pte to be checked
6460 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6463 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6464 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6465 * move charge. if @target is not NULL, the page is stored in target->page
6466 * with extra refcnt got(Callers should handle it).
6467 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6468 * target for charge migration. if @target is not NULL, the entry is stored
6471 * Called with pte lock held.
6478 enum mc_target_type
{
6484 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
6485 unsigned long addr
, pte_t ptent
)
6487 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
6489 if (!page
|| !page_mapped(page
))
6491 if (PageAnon(page
)) {
6492 /* we don't move shared anon */
6495 } else if (!move_file())
6496 /* we ignore mapcount for file pages */
6498 if (!get_page_unless_zero(page
))
6505 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6506 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6508 struct page
*page
= NULL
;
6509 swp_entry_t ent
= pte_to_swp_entry(ptent
);
6511 if (!move_anon() || non_swap_entry(ent
))
6514 * Because lookup_swap_cache() updates some statistics counter,
6515 * we call find_get_page() with swapper_space directly.
6517 page
= find_get_page(swap_address_space(ent
), ent
.val
);
6518 if (do_swap_account
)
6519 entry
->val
= ent
.val
;
6524 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6525 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6531 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
6532 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6534 struct page
*page
= NULL
;
6535 struct address_space
*mapping
;
6538 if (!vma
->vm_file
) /* anonymous vma */
6543 mapping
= vma
->vm_file
->f_mapping
;
6544 if (pte_none(ptent
))
6545 pgoff
= linear_page_index(vma
, addr
);
6546 else /* pte_file(ptent) is true */
6547 pgoff
= pte_to_pgoff(ptent
);
6549 /* page is moved even if it's not RSS of this task(page-faulted). */
6551 /* shmem/tmpfs may report page out on swap: account for that too. */
6552 if (shmem_mapping(mapping
)) {
6553 page
= find_get_entry(mapping
, pgoff
);
6554 if (radix_tree_exceptional_entry(page
)) {
6555 swp_entry_t swp
= radix_to_swp_entry(page
);
6556 if (do_swap_account
)
6558 page
= find_get_page(swap_address_space(swp
), swp
.val
);
6561 page
= find_get_page(mapping
, pgoff
);
6563 page
= find_get_page(mapping
, pgoff
);
6568 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
6569 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
6571 struct page
*page
= NULL
;
6572 struct page_cgroup
*pc
;
6573 enum mc_target_type ret
= MC_TARGET_NONE
;
6574 swp_entry_t ent
= { .val
= 0 };
6576 if (pte_present(ptent
))
6577 page
= mc_handle_present_pte(vma
, addr
, ptent
);
6578 else if (is_swap_pte(ptent
))
6579 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
6580 else if (pte_none(ptent
) || pte_file(ptent
))
6581 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
6583 if (!page
&& !ent
.val
)
6586 pc
= lookup_page_cgroup(page
);
6588 * Do only loose check w/o page_cgroup lock.
6589 * mem_cgroup_move_account() checks the pc is valid or not under
6592 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6593 ret
= MC_TARGET_PAGE
;
6595 target
->page
= page
;
6597 if (!ret
|| !target
)
6600 /* There is a swap entry and a page doesn't exist or isn't charged */
6601 if (ent
.val
&& !ret
&&
6602 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
6603 ret
= MC_TARGET_SWAP
;
6610 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6612 * We don't consider swapping or file mapped pages because THP does not
6613 * support them for now.
6614 * Caller should make sure that pmd_trans_huge(pmd) is true.
6616 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6617 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6619 struct page
*page
= NULL
;
6620 struct page_cgroup
*pc
;
6621 enum mc_target_type ret
= MC_TARGET_NONE
;
6623 page
= pmd_page(pmd
);
6624 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
6627 pc
= lookup_page_cgroup(page
);
6628 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6629 ret
= MC_TARGET_PAGE
;
6632 target
->page
= page
;
6638 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6639 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6641 return MC_TARGET_NONE
;
6645 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
6646 unsigned long addr
, unsigned long end
,
6647 struct mm_walk
*walk
)
6649 struct vm_area_struct
*vma
= walk
->private;
6653 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
6654 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
6655 mc
.precharge
+= HPAGE_PMD_NR
;
6660 if (pmd_trans_unstable(pmd
))
6662 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6663 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
6664 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
6665 mc
.precharge
++; /* increment precharge temporarily */
6666 pte_unmap_unlock(pte
- 1, ptl
);
6672 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
6674 unsigned long precharge
;
6675 struct vm_area_struct
*vma
;
6677 down_read(&mm
->mmap_sem
);
6678 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6679 struct mm_walk mem_cgroup_count_precharge_walk
= {
6680 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
6684 if (is_vm_hugetlb_page(vma
))
6686 walk_page_range(vma
->vm_start
, vma
->vm_end
,
6687 &mem_cgroup_count_precharge_walk
);
6689 up_read(&mm
->mmap_sem
);
6691 precharge
= mc
.precharge
;
6697 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
6699 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
6701 VM_BUG_ON(mc
.moving_task
);
6702 mc
.moving_task
= current
;
6703 return mem_cgroup_do_precharge(precharge
);
6706 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6707 static void __mem_cgroup_clear_mc(void)
6709 struct mem_cgroup
*from
= mc
.from
;
6710 struct mem_cgroup
*to
= mc
.to
;
6713 /* we must uncharge all the leftover precharges from mc.to */
6715 __mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
6719 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6720 * we must uncharge here.
6722 if (mc
.moved_charge
) {
6723 __mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
6724 mc
.moved_charge
= 0;
6726 /* we must fixup refcnts and charges */
6727 if (mc
.moved_swap
) {
6728 /* uncharge swap account from the old cgroup */
6729 if (!mem_cgroup_is_root(mc
.from
))
6730 res_counter_uncharge(&mc
.from
->memsw
,
6731 PAGE_SIZE
* mc
.moved_swap
);
6733 for (i
= 0; i
< mc
.moved_swap
; i
++)
6734 css_put(&mc
.from
->css
);
6736 if (!mem_cgroup_is_root(mc
.to
)) {
6738 * we charged both to->res and to->memsw, so we should
6741 res_counter_uncharge(&mc
.to
->res
,
6742 PAGE_SIZE
* mc
.moved_swap
);
6744 /* we've already done css_get(mc.to) */
6747 memcg_oom_recover(from
);
6748 memcg_oom_recover(to
);
6749 wake_up_all(&mc
.waitq
);
6752 static void mem_cgroup_clear_mc(void)
6754 struct mem_cgroup
*from
= mc
.from
;
6757 * we must clear moving_task before waking up waiters at the end of
6760 mc
.moving_task
= NULL
;
6761 __mem_cgroup_clear_mc();
6762 spin_lock(&mc
.lock
);
6765 spin_unlock(&mc
.lock
);
6766 mem_cgroup_end_move(from
);
6769 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6770 struct cgroup_taskset
*tset
)
6772 struct task_struct
*p
= cgroup_taskset_first(tset
);
6774 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6775 unsigned long move_charge_at_immigrate
;
6778 * We are now commited to this value whatever it is. Changes in this
6779 * tunable will only affect upcoming migrations, not the current one.
6780 * So we need to save it, and keep it going.
6782 move_charge_at_immigrate
= memcg
->move_charge_at_immigrate
;
6783 if (move_charge_at_immigrate
) {
6784 struct mm_struct
*mm
;
6785 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
6787 VM_BUG_ON(from
== memcg
);
6789 mm
= get_task_mm(p
);
6792 /* We move charges only when we move a owner of the mm */
6793 if (mm
->owner
== p
) {
6796 VM_BUG_ON(mc
.precharge
);
6797 VM_BUG_ON(mc
.moved_charge
);
6798 VM_BUG_ON(mc
.moved_swap
);
6799 mem_cgroup_start_move(from
);
6800 spin_lock(&mc
.lock
);
6803 mc
.immigrate_flags
= move_charge_at_immigrate
;
6804 spin_unlock(&mc
.lock
);
6805 /* We set mc.moving_task later */
6807 ret
= mem_cgroup_precharge_mc(mm
);
6809 mem_cgroup_clear_mc();
6816 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6817 struct cgroup_taskset
*tset
)
6819 mem_cgroup_clear_mc();
6822 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6823 unsigned long addr
, unsigned long end
,
6824 struct mm_walk
*walk
)
6827 struct vm_area_struct
*vma
= walk
->private;
6830 enum mc_target_type target_type
;
6831 union mc_target target
;
6833 struct page_cgroup
*pc
;
6836 * We don't take compound_lock() here but no race with splitting thp
6838 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6839 * under splitting, which means there's no concurrent thp split,
6840 * - if another thread runs into split_huge_page() just after we
6841 * entered this if-block, the thread must wait for page table lock
6842 * to be unlocked in __split_huge_page_splitting(), where the main
6843 * part of thp split is not executed yet.
6845 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
6846 if (mc
.precharge
< HPAGE_PMD_NR
) {
6850 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6851 if (target_type
== MC_TARGET_PAGE
) {
6853 if (!isolate_lru_page(page
)) {
6854 pc
= lookup_page_cgroup(page
);
6855 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
6856 pc
, mc
.from
, mc
.to
)) {
6857 mc
.precharge
-= HPAGE_PMD_NR
;
6858 mc
.moved_charge
+= HPAGE_PMD_NR
;
6860 putback_lru_page(page
);
6868 if (pmd_trans_unstable(pmd
))
6871 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6872 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6873 pte_t ptent
= *(pte
++);
6879 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6880 case MC_TARGET_PAGE
:
6882 if (isolate_lru_page(page
))
6884 pc
= lookup_page_cgroup(page
);
6885 if (!mem_cgroup_move_account(page
, 1, pc
,
6888 /* we uncharge from mc.from later. */
6891 putback_lru_page(page
);
6892 put
: /* get_mctgt_type() gets the page */
6895 case MC_TARGET_SWAP
:
6897 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6899 /* we fixup refcnts and charges later. */
6907 pte_unmap_unlock(pte
- 1, ptl
);
6912 * We have consumed all precharges we got in can_attach().
6913 * We try charge one by one, but don't do any additional
6914 * charges to mc.to if we have failed in charge once in attach()
6917 ret
= mem_cgroup_do_precharge(1);
6925 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
6927 struct vm_area_struct
*vma
;
6929 lru_add_drain_all();
6931 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
6933 * Someone who are holding the mmap_sem might be waiting in
6934 * waitq. So we cancel all extra charges, wake up all waiters,
6935 * and retry. Because we cancel precharges, we might not be able
6936 * to move enough charges, but moving charge is a best-effort
6937 * feature anyway, so it wouldn't be a big problem.
6939 __mem_cgroup_clear_mc();
6943 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6945 struct mm_walk mem_cgroup_move_charge_walk
= {
6946 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6950 if (is_vm_hugetlb_page(vma
))
6952 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
6953 &mem_cgroup_move_charge_walk
);
6956 * means we have consumed all precharges and failed in
6957 * doing additional charge. Just abandon here.
6961 up_read(&mm
->mmap_sem
);
6964 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6965 struct cgroup_taskset
*tset
)
6967 struct task_struct
*p
= cgroup_taskset_first(tset
);
6968 struct mm_struct
*mm
= get_task_mm(p
);
6972 mem_cgroup_move_charge(mm
);
6976 mem_cgroup_clear_mc();
6978 #else /* !CONFIG_MMU */
6979 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6980 struct cgroup_taskset
*tset
)
6984 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6985 struct cgroup_taskset
*tset
)
6988 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6989 struct cgroup_taskset
*tset
)
6995 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6996 * to verify sane_behavior flag on each mount attempt.
6998 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
7001 * use_hierarchy is forced with sane_behavior. cgroup core
7002 * guarantees that @root doesn't have any children, so turning it
7003 * on for the root memcg is enough.
7005 if (cgroup_sane_behavior(root_css
->cgroup
))
7006 mem_cgroup_from_css(root_css
)->use_hierarchy
= true;
7009 struct cgroup_subsys memory_cgrp_subsys
= {
7010 .css_alloc
= mem_cgroup_css_alloc
,
7011 .css_online
= mem_cgroup_css_online
,
7012 .css_offline
= mem_cgroup_css_offline
,
7013 .css_free
= mem_cgroup_css_free
,
7014 .can_attach
= mem_cgroup_can_attach
,
7015 .cancel_attach
= mem_cgroup_cancel_attach
,
7016 .attach
= mem_cgroup_move_task
,
7017 .bind
= mem_cgroup_bind
,
7018 .base_cftypes
= mem_cgroup_files
,
7022 #ifdef CONFIG_MEMCG_SWAP
7023 static int __init
enable_swap_account(char *s
)
7025 if (!strcmp(s
, "1"))
7026 really_do_swap_account
= 1;
7027 else if (!strcmp(s
, "0"))
7028 really_do_swap_account
= 0;
7031 __setup("swapaccount=", enable_swap_account
);
7033 static void __init
memsw_file_init(void)
7035 WARN_ON(cgroup_add_cftypes(&memory_cgrp_subsys
, memsw_cgroup_files
));
7038 static void __init
enable_swap_cgroup(void)
7040 if (!mem_cgroup_disabled() && really_do_swap_account
) {
7041 do_swap_account
= 1;
7047 static void __init
enable_swap_cgroup(void)
7053 * subsys_initcall() for memory controller.
7055 * Some parts like hotcpu_notifier() have to be initialized from this context
7056 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7057 * everything that doesn't depend on a specific mem_cgroup structure should
7058 * be initialized from here.
7060 static int __init
mem_cgroup_init(void)
7062 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
7063 enable_swap_cgroup();
7064 mem_cgroup_soft_limit_tree_init();
7068 subsys_initcall(mem_cgroup_init
);