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_zoneinfo(struct mem_cgroup
*memcg
, int nid
, int zid
)
681 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
682 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
685 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
690 static struct mem_cgroup_per_zone
*
691 page_cgroup_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
693 int nid
= page_to_nid(page
);
694 int zid
= page_zonenum(page
);
696 return mem_cgroup_zoneinfo(memcg
, nid
, zid
);
699 static struct mem_cgroup_tree_per_zone
*
700 soft_limit_tree_node_zone(int nid
, int zid
)
702 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
705 static struct mem_cgroup_tree_per_zone
*
706 soft_limit_tree_from_page(struct page
*page
)
708 int nid
= page_to_nid(page
);
709 int zid
= page_zonenum(page
);
711 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
715 __mem_cgroup_insert_exceeded(struct mem_cgroup
*memcg
,
716 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
);
749 __mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
750 struct mem_cgroup_per_zone
*mz
,
751 struct mem_cgroup_tree_per_zone
*mctz
)
755 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
760 mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
761 struct mem_cgroup_per_zone
*mz
,
762 struct mem_cgroup_tree_per_zone
*mctz
)
764 spin_lock(&mctz
->lock
);
765 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
766 spin_unlock(&mctz
->lock
);
770 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
772 unsigned long long excess
;
773 struct mem_cgroup_per_zone
*mz
;
774 struct mem_cgroup_tree_per_zone
*mctz
;
775 int nid
= page_to_nid(page
);
776 int zid
= page_zonenum(page
);
777 mctz
= soft_limit_tree_from_page(page
);
780 * Necessary to update all ancestors when hierarchy is used.
781 * because their event counter is not touched.
783 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
784 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
785 excess
= res_counter_soft_limit_excess(&memcg
->res
);
787 * We have to update the tree if mz is on RB-tree or
788 * mem is over its softlimit.
790 if (excess
|| mz
->on_tree
) {
791 spin_lock(&mctz
->lock
);
792 /* if on-tree, remove it */
794 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
796 * Insert again. mz->usage_in_excess will be updated.
797 * If excess is 0, no tree ops.
799 __mem_cgroup_insert_exceeded(memcg
, mz
, mctz
, excess
);
800 spin_unlock(&mctz
->lock
);
805 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
808 struct mem_cgroup_per_zone
*mz
;
809 struct mem_cgroup_tree_per_zone
*mctz
;
811 for_each_node(node
) {
812 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
813 mz
= mem_cgroup_zoneinfo(memcg
, node
, zone
);
814 mctz
= soft_limit_tree_node_zone(node
, zone
);
815 mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
820 static struct mem_cgroup_per_zone
*
821 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
823 struct rb_node
*rightmost
= NULL
;
824 struct mem_cgroup_per_zone
*mz
;
828 rightmost
= rb_last(&mctz
->rb_root
);
830 goto done
; /* Nothing to reclaim from */
832 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
834 * Remove the node now but someone else can add it back,
835 * we will to add it back at the end of reclaim to its correct
836 * position in the tree.
838 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
839 if (!res_counter_soft_limit_excess(&mz
->memcg
->res
) ||
840 !css_tryget(&mz
->memcg
->css
))
846 static struct mem_cgroup_per_zone
*
847 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
849 struct mem_cgroup_per_zone
*mz
;
851 spin_lock(&mctz
->lock
);
852 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
853 spin_unlock(&mctz
->lock
);
858 * Implementation Note: reading percpu statistics for memcg.
860 * Both of vmstat[] and percpu_counter has threshold and do periodic
861 * synchronization to implement "quick" read. There are trade-off between
862 * reading cost and precision of value. Then, we may have a chance to implement
863 * a periodic synchronizion of counter in memcg's counter.
865 * But this _read() function is used for user interface now. The user accounts
866 * memory usage by memory cgroup and he _always_ requires exact value because
867 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
868 * have to visit all online cpus and make sum. So, for now, unnecessary
869 * synchronization is not implemented. (just implemented for cpu hotplug)
871 * If there are kernel internal actions which can make use of some not-exact
872 * value, and reading all cpu value can be performance bottleneck in some
873 * common workload, threashold and synchonization as vmstat[] should be
876 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
877 enum mem_cgroup_stat_index idx
)
883 for_each_online_cpu(cpu
)
884 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
885 #ifdef CONFIG_HOTPLUG_CPU
886 spin_lock(&memcg
->pcp_counter_lock
);
887 val
+= memcg
->nocpu_base
.count
[idx
];
888 spin_unlock(&memcg
->pcp_counter_lock
);
894 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
897 int val
= (charge
) ? 1 : -1;
898 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
901 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
902 enum mem_cgroup_events_index idx
)
904 unsigned long val
= 0;
908 for_each_online_cpu(cpu
)
909 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
910 #ifdef CONFIG_HOTPLUG_CPU
911 spin_lock(&memcg
->pcp_counter_lock
);
912 val
+= memcg
->nocpu_base
.events
[idx
];
913 spin_unlock(&memcg
->pcp_counter_lock
);
919 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
921 bool anon
, int nr_pages
)
924 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
925 * counted as CACHE even if it's on ANON LRU.
928 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
931 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
934 if (PageTransHuge(page
))
935 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
938 /* pagein of a big page is an event. So, ignore page size */
940 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
942 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
943 nr_pages
= -nr_pages
; /* for event */
946 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
950 mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
952 struct mem_cgroup_per_zone
*mz
;
954 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
955 return mz
->lru_size
[lru
];
959 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup
*memcg
, int nid
, int zid
,
960 unsigned int lru_mask
)
962 struct mem_cgroup_per_zone
*mz
;
964 unsigned long ret
= 0;
966 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
969 if (BIT(lru
) & lru_mask
)
970 ret
+= mz
->lru_size
[lru
];
976 mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
977 int nid
, unsigned int lru_mask
)
982 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++)
983 total
+= mem_cgroup_zone_nr_lru_pages(memcg
,
989 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
990 unsigned int lru_mask
)
995 for_each_node_state(nid
, N_MEMORY
)
996 total
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
1000 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
1001 enum mem_cgroup_events_target target
)
1003 unsigned long val
, next
;
1005 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
1006 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
1007 /* from time_after() in jiffies.h */
1008 if ((long)next
- (long)val
< 0) {
1010 case MEM_CGROUP_TARGET_THRESH
:
1011 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
1013 case MEM_CGROUP_TARGET_SOFTLIMIT
:
1014 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
1016 case MEM_CGROUP_TARGET_NUMAINFO
:
1017 next
= val
+ NUMAINFO_EVENTS_TARGET
;
1022 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
1029 * Check events in order.
1032 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
1035 /* threshold event is triggered in finer grain than soft limit */
1036 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
1037 MEM_CGROUP_TARGET_THRESH
))) {
1039 bool do_numainfo __maybe_unused
;
1041 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
1042 MEM_CGROUP_TARGET_SOFTLIMIT
);
1043 #if MAX_NUMNODES > 1
1044 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
1045 MEM_CGROUP_TARGET_NUMAINFO
);
1049 mem_cgroup_threshold(memcg
);
1050 if (unlikely(do_softlimit
))
1051 mem_cgroup_update_tree(memcg
, page
);
1052 #if MAX_NUMNODES > 1
1053 if (unlikely(do_numainfo
))
1054 atomic_inc(&memcg
->numainfo_events
);
1060 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
1063 * mm_update_next_owner() may clear mm->owner to NULL
1064 * if it races with swapoff, page migration, etc.
1065 * So this can be called with p == NULL.
1070 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
1073 static struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
1075 struct mem_cgroup
*memcg
= NULL
;
1080 * Page cache insertions can happen withou an
1081 * actual mm context, e.g. during disk probing
1082 * on boot, loopback IO, acct() writes etc.
1085 memcg
= root_mem_cgroup
;
1087 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1088 if (unlikely(!memcg
))
1089 memcg
= root_mem_cgroup
;
1091 } while (!css_tryget(&memcg
->css
));
1097 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1098 * ref. count) or NULL if the whole root's subtree has been visited.
1100 * helper function to be used by mem_cgroup_iter
1102 static struct mem_cgroup
*__mem_cgroup_iter_next(struct mem_cgroup
*root
,
1103 struct mem_cgroup
*last_visited
)
1105 struct cgroup_subsys_state
*prev_css
, *next_css
;
1107 prev_css
= last_visited
? &last_visited
->css
: NULL
;
1109 next_css
= css_next_descendant_pre(prev_css
, &root
->css
);
1112 * Even if we found a group we have to make sure it is
1113 * alive. css && !memcg means that the groups should be
1114 * skipped and we should continue the tree walk.
1115 * last_visited css is safe to use because it is
1116 * protected by css_get and the tree walk is rcu safe.
1118 * We do not take a reference on the root of the tree walk
1119 * because we might race with the root removal when it would
1120 * be the only node in the iterated hierarchy and mem_cgroup_iter
1121 * would end up in an endless loop because it expects that at
1122 * least one valid node will be returned. Root cannot disappear
1123 * because caller of the iterator should hold it already so
1124 * skipping css reference should be safe.
1127 if ((next_css
== &root
->css
) ||
1128 ((next_css
->flags
& CSS_ONLINE
) && css_tryget(next_css
)))
1129 return mem_cgroup_from_css(next_css
);
1131 prev_css
= next_css
;
1138 static void mem_cgroup_iter_invalidate(struct mem_cgroup
*root
)
1141 * When a group in the hierarchy below root is destroyed, the
1142 * hierarchy iterator can no longer be trusted since it might
1143 * have pointed to the destroyed group. Invalidate it.
1145 atomic_inc(&root
->dead_count
);
1148 static struct mem_cgroup
*
1149 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter
*iter
,
1150 struct mem_cgroup
*root
,
1153 struct mem_cgroup
*position
= NULL
;
1155 * A cgroup destruction happens in two stages: offlining and
1156 * release. They are separated by a RCU grace period.
1158 * If the iterator is valid, we may still race with an
1159 * offlining. The RCU lock ensures the object won't be
1160 * released, tryget will fail if we lost the race.
1162 *sequence
= atomic_read(&root
->dead_count
);
1163 if (iter
->last_dead_count
== *sequence
) {
1165 position
= iter
->last_visited
;
1168 * We cannot take a reference to root because we might race
1169 * with root removal and returning NULL would end up in
1170 * an endless loop on the iterator user level when root
1171 * would be returned all the time.
1173 if (position
&& position
!= root
&&
1174 !css_tryget(&position
->css
))
1180 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter
*iter
,
1181 struct mem_cgroup
*last_visited
,
1182 struct mem_cgroup
*new_position
,
1183 struct mem_cgroup
*root
,
1186 /* root reference counting symmetric to mem_cgroup_iter_load */
1187 if (last_visited
&& last_visited
!= root
)
1188 css_put(&last_visited
->css
);
1190 * We store the sequence count from the time @last_visited was
1191 * loaded successfully instead of rereading it here so that we
1192 * don't lose destruction events in between. We could have
1193 * raced with the destruction of @new_position after all.
1195 iter
->last_visited
= new_position
;
1197 iter
->last_dead_count
= sequence
;
1201 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1202 * @root: hierarchy root
1203 * @prev: previously returned memcg, NULL on first invocation
1204 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1206 * Returns references to children of the hierarchy below @root, or
1207 * @root itself, or %NULL after a full round-trip.
1209 * Caller must pass the return value in @prev on subsequent
1210 * invocations for reference counting, or use mem_cgroup_iter_break()
1211 * to cancel a hierarchy walk before the round-trip is complete.
1213 * Reclaimers can specify a zone and a priority level in @reclaim to
1214 * divide up the memcgs in the hierarchy among all concurrent
1215 * reclaimers operating on the same zone and priority.
1217 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1218 struct mem_cgroup
*prev
,
1219 struct mem_cgroup_reclaim_cookie
*reclaim
)
1221 struct mem_cgroup
*memcg
= NULL
;
1222 struct mem_cgroup
*last_visited
= NULL
;
1224 if (mem_cgroup_disabled())
1228 root
= root_mem_cgroup
;
1230 if (prev
&& !reclaim
)
1231 last_visited
= prev
;
1233 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1241 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1242 int uninitialized_var(seq
);
1245 int nid
= zone_to_nid(reclaim
->zone
);
1246 int zid
= zone_idx(reclaim
->zone
);
1247 struct mem_cgroup_per_zone
*mz
;
1249 mz
= mem_cgroup_zoneinfo(root
, nid
, zid
);
1250 iter
= &mz
->reclaim_iter
[reclaim
->priority
];
1251 if (prev
&& reclaim
->generation
!= iter
->generation
) {
1252 iter
->last_visited
= NULL
;
1256 last_visited
= mem_cgroup_iter_load(iter
, root
, &seq
);
1259 memcg
= __mem_cgroup_iter_next(root
, last_visited
);
1262 mem_cgroup_iter_update(iter
, last_visited
, memcg
, root
,
1267 else if (!prev
&& memcg
)
1268 reclaim
->generation
= iter
->generation
;
1277 if (prev
&& prev
!= root
)
1278 css_put(&prev
->css
);
1284 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1285 * @root: hierarchy root
1286 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1288 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1289 struct mem_cgroup
*prev
)
1292 root
= root_mem_cgroup
;
1293 if (prev
&& prev
!= root
)
1294 css_put(&prev
->css
);
1298 * Iteration constructs for visiting all cgroups (under a tree). If
1299 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1300 * be used for reference counting.
1302 #define for_each_mem_cgroup_tree(iter, root) \
1303 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1305 iter = mem_cgroup_iter(root, iter, NULL))
1307 #define for_each_mem_cgroup(iter) \
1308 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1310 iter = mem_cgroup_iter(NULL, iter, NULL))
1312 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1314 struct mem_cgroup
*memcg
;
1317 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1318 if (unlikely(!memcg
))
1323 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1326 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1334 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1337 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1338 * @zone: zone of the wanted lruvec
1339 * @memcg: memcg of the wanted lruvec
1341 * Returns the lru list vector holding pages for the given @zone and
1342 * @mem. This can be the global zone lruvec, if the memory controller
1345 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1346 struct mem_cgroup
*memcg
)
1348 struct mem_cgroup_per_zone
*mz
;
1349 struct lruvec
*lruvec
;
1351 if (mem_cgroup_disabled()) {
1352 lruvec
= &zone
->lruvec
;
1356 mz
= mem_cgroup_zoneinfo(memcg
, zone_to_nid(zone
), zone_idx(zone
));
1357 lruvec
= &mz
->lruvec
;
1360 * Since a node can be onlined after the mem_cgroup was created,
1361 * we have to be prepared to initialize lruvec->zone here;
1362 * and if offlined then reonlined, we need to reinitialize it.
1364 if (unlikely(lruvec
->zone
!= zone
))
1365 lruvec
->zone
= zone
;
1370 * Following LRU functions are allowed to be used without PCG_LOCK.
1371 * Operations are called by routine of global LRU independently from memcg.
1372 * What we have to take care of here is validness of pc->mem_cgroup.
1374 * Changes to pc->mem_cgroup happens when
1377 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1378 * It is added to LRU before charge.
1379 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1380 * When moving account, the page is not on LRU. It's isolated.
1384 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1386 * @zone: zone of the page
1388 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1390 struct mem_cgroup_per_zone
*mz
;
1391 struct mem_cgroup
*memcg
;
1392 struct page_cgroup
*pc
;
1393 struct lruvec
*lruvec
;
1395 if (mem_cgroup_disabled()) {
1396 lruvec
= &zone
->lruvec
;
1400 pc
= lookup_page_cgroup(page
);
1401 memcg
= pc
->mem_cgroup
;
1404 * Surreptitiously switch any uncharged offlist page to root:
1405 * an uncharged page off lru does nothing to secure
1406 * its former mem_cgroup from sudden removal.
1408 * Our caller holds lru_lock, and PageCgroupUsed is updated
1409 * under page_cgroup lock: between them, they make all uses
1410 * of pc->mem_cgroup safe.
1412 if (!PageLRU(page
) && !PageCgroupUsed(pc
) && memcg
!= root_mem_cgroup
)
1413 pc
->mem_cgroup
= memcg
= root_mem_cgroup
;
1415 mz
= page_cgroup_zoneinfo(memcg
, page
);
1416 lruvec
= &mz
->lruvec
;
1419 * Since a node can be onlined after the mem_cgroup was created,
1420 * we have to be prepared to initialize lruvec->zone here;
1421 * and if offlined then reonlined, we need to reinitialize it.
1423 if (unlikely(lruvec
->zone
!= zone
))
1424 lruvec
->zone
= zone
;
1429 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1430 * @lruvec: mem_cgroup per zone lru vector
1431 * @lru: index of lru list the page is sitting on
1432 * @nr_pages: positive when adding or negative when removing
1434 * This function must be called when a page is added to or removed from an
1437 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1440 struct mem_cgroup_per_zone
*mz
;
1441 unsigned long *lru_size
;
1443 if (mem_cgroup_disabled())
1446 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1447 lru_size
= mz
->lru_size
+ lru
;
1448 *lru_size
+= nr_pages
;
1449 VM_BUG_ON((long)(*lru_size
) < 0);
1453 * Checks whether given mem is same or in the root_mem_cgroup's
1456 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1457 struct mem_cgroup
*memcg
)
1459 if (root_memcg
== memcg
)
1461 if (!root_memcg
->use_hierarchy
|| !memcg
)
1463 return cgroup_is_descendant(memcg
->css
.cgroup
, root_memcg
->css
.cgroup
);
1466 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1467 struct mem_cgroup
*memcg
)
1472 ret
= __mem_cgroup_same_or_subtree(root_memcg
, memcg
);
1477 bool task_in_mem_cgroup(struct task_struct
*task
,
1478 const struct mem_cgroup
*memcg
)
1480 struct mem_cgroup
*curr
= NULL
;
1481 struct task_struct
*p
;
1484 p
= find_lock_task_mm(task
);
1486 curr
= get_mem_cgroup_from_mm(p
->mm
);
1490 * All threads may have already detached their mm's, but the oom
1491 * killer still needs to detect if they have already been oom
1492 * killed to prevent needlessly killing additional tasks.
1495 curr
= mem_cgroup_from_task(task
);
1497 css_get(&curr
->css
);
1501 * We should check use_hierarchy of "memcg" not "curr". Because checking
1502 * use_hierarchy of "curr" here make this function true if hierarchy is
1503 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1504 * hierarchy(even if use_hierarchy is disabled in "memcg").
1506 ret
= mem_cgroup_same_or_subtree(memcg
, curr
);
1507 css_put(&curr
->css
);
1511 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1513 unsigned long inactive_ratio
;
1514 unsigned long inactive
;
1515 unsigned long active
;
1518 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1519 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1521 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1523 inactive_ratio
= int_sqrt(10 * gb
);
1527 return inactive
* inactive_ratio
< active
;
1530 #define mem_cgroup_from_res_counter(counter, member) \
1531 container_of(counter, struct mem_cgroup, member)
1534 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1535 * @memcg: the memory cgroup
1537 * Returns the maximum amount of memory @mem can be charged with, in
1540 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1542 unsigned long long margin
;
1544 margin
= res_counter_margin(&memcg
->res
);
1545 if (do_swap_account
)
1546 margin
= min(margin
, res_counter_margin(&memcg
->memsw
));
1547 return margin
>> PAGE_SHIFT
;
1550 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1553 if (!css_parent(&memcg
->css
))
1554 return vm_swappiness
;
1556 return memcg
->swappiness
;
1560 * memcg->moving_account is used for checking possibility that some thread is
1561 * calling move_account(). When a thread on CPU-A starts moving pages under
1562 * a memcg, other threads should check memcg->moving_account under
1563 * rcu_read_lock(), like this:
1567 * memcg->moving_account+1 if (memcg->mocing_account)
1569 * synchronize_rcu() update something.
1574 /* for quick checking without looking up memcg */
1575 atomic_t memcg_moving __read_mostly
;
1577 static void mem_cgroup_start_move(struct mem_cgroup
*memcg
)
1579 atomic_inc(&memcg_moving
);
1580 atomic_inc(&memcg
->moving_account
);
1584 static void mem_cgroup_end_move(struct mem_cgroup
*memcg
)
1587 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1588 * We check NULL in callee rather than caller.
1591 atomic_dec(&memcg_moving
);
1592 atomic_dec(&memcg
->moving_account
);
1597 * A routine for checking "mem" is under move_account() or not.
1599 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1600 * moving cgroups. This is for waiting at high-memory pressure
1603 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1605 struct mem_cgroup
*from
;
1606 struct mem_cgroup
*to
;
1609 * Unlike task_move routines, we access mc.to, mc.from not under
1610 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1612 spin_lock(&mc
.lock
);
1618 ret
= mem_cgroup_same_or_subtree(memcg
, from
)
1619 || mem_cgroup_same_or_subtree(memcg
, to
);
1621 spin_unlock(&mc
.lock
);
1625 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1627 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1628 if (mem_cgroup_under_move(memcg
)) {
1630 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1631 /* moving charge context might have finished. */
1634 finish_wait(&mc
.waitq
, &wait
);
1642 * Take this lock when
1643 * - a code tries to modify page's memcg while it's USED.
1644 * - a code tries to modify page state accounting in a memcg.
1646 static void move_lock_mem_cgroup(struct mem_cgroup
*memcg
,
1647 unsigned long *flags
)
1649 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
1652 static void move_unlock_mem_cgroup(struct mem_cgroup
*memcg
,
1653 unsigned long *flags
)
1655 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
1658 #define K(x) ((x) << (PAGE_SHIFT-10))
1660 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1661 * @memcg: The memory cgroup that went over limit
1662 * @p: Task that is going to be killed
1664 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1667 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1669 /* oom_info_lock ensures that parallel ooms do not interleave */
1670 static DEFINE_MUTEX(oom_info_lock
);
1671 struct mem_cgroup
*iter
;
1677 mutex_lock(&oom_info_lock
);
1680 pr_info("Task in ");
1681 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1682 pr_info(" killed as a result of limit of ");
1683 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1688 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1689 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1690 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1691 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1692 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1693 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1694 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1695 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1696 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1697 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) >> 10,
1698 res_counter_read_u64(&memcg
->kmem
, RES_LIMIT
) >> 10,
1699 res_counter_read_u64(&memcg
->kmem
, RES_FAILCNT
));
1701 for_each_mem_cgroup_tree(iter
, memcg
) {
1702 pr_info("Memory cgroup stats for ");
1703 pr_cont_cgroup_path(iter
->css
.cgroup
);
1706 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1707 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1709 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1710 K(mem_cgroup_read_stat(iter
, i
)));
1713 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1714 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1715 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1719 mutex_unlock(&oom_info_lock
);
1723 * This function returns the number of memcg under hierarchy tree. Returns
1724 * 1(self count) if no children.
1726 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1729 struct mem_cgroup
*iter
;
1731 for_each_mem_cgroup_tree(iter
, memcg
)
1737 * Return the memory (and swap, if configured) limit for a memcg.
1739 static u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1743 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1746 * Do not consider swap space if we cannot swap due to swappiness
1748 if (mem_cgroup_swappiness(memcg
)) {
1751 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1752 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1755 * If memsw is finite and limits the amount of swap space
1756 * available to this memcg, return that limit.
1758 limit
= min(limit
, memsw
);
1764 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1767 struct mem_cgroup
*iter
;
1768 unsigned long chosen_points
= 0;
1769 unsigned long totalpages
;
1770 unsigned int points
= 0;
1771 struct task_struct
*chosen
= NULL
;
1774 * If current has a pending SIGKILL or is exiting, then automatically
1775 * select it. The goal is to allow it to allocate so that it may
1776 * quickly exit and free its memory.
1778 if (fatal_signal_pending(current
) || current
->flags
& PF_EXITING
) {
1779 set_thread_flag(TIF_MEMDIE
);
1783 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1784 totalpages
= mem_cgroup_get_limit(memcg
) >> PAGE_SHIFT
? : 1;
1785 for_each_mem_cgroup_tree(iter
, memcg
) {
1786 struct css_task_iter it
;
1787 struct task_struct
*task
;
1789 css_task_iter_start(&iter
->css
, &it
);
1790 while ((task
= css_task_iter_next(&it
))) {
1791 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1793 case OOM_SCAN_SELECT
:
1795 put_task_struct(chosen
);
1797 chosen_points
= ULONG_MAX
;
1798 get_task_struct(chosen
);
1800 case OOM_SCAN_CONTINUE
:
1802 case OOM_SCAN_ABORT
:
1803 css_task_iter_end(&it
);
1804 mem_cgroup_iter_break(memcg
, iter
);
1806 put_task_struct(chosen
);
1811 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1812 if (!points
|| points
< chosen_points
)
1814 /* Prefer thread group leaders for display purposes */
1815 if (points
== chosen_points
&&
1816 thread_group_leader(chosen
))
1820 put_task_struct(chosen
);
1822 chosen_points
= points
;
1823 get_task_struct(chosen
);
1825 css_task_iter_end(&it
);
1830 points
= chosen_points
* 1000 / totalpages
;
1831 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1832 NULL
, "Memory cgroup out of memory");
1835 static unsigned long mem_cgroup_reclaim(struct mem_cgroup
*memcg
,
1837 unsigned long flags
)
1839 unsigned long total
= 0;
1840 bool noswap
= false;
1843 if (flags
& MEM_CGROUP_RECLAIM_NOSWAP
)
1845 if (!(flags
& MEM_CGROUP_RECLAIM_SHRINK
) && memcg
->memsw_is_minimum
)
1848 for (loop
= 0; loop
< MEM_CGROUP_MAX_RECLAIM_LOOPS
; loop
++) {
1850 drain_all_stock_async(memcg
);
1851 total
+= try_to_free_mem_cgroup_pages(memcg
, gfp_mask
, noswap
);
1853 * Allow limit shrinkers, which are triggered directly
1854 * by userspace, to catch signals and stop reclaim
1855 * after minimal progress, regardless of the margin.
1857 if (total
&& (flags
& MEM_CGROUP_RECLAIM_SHRINK
))
1859 if (mem_cgroup_margin(memcg
))
1862 * If nothing was reclaimed after two attempts, there
1863 * may be no reclaimable pages in this hierarchy.
1872 * test_mem_cgroup_node_reclaimable
1873 * @memcg: the target memcg
1874 * @nid: the node ID to be checked.
1875 * @noswap : specify true here if the user wants flle only information.
1877 * This function returns whether the specified memcg contains any
1878 * reclaimable pages on a node. Returns true if there are any reclaimable
1879 * pages in the node.
1881 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1882 int nid
, bool noswap
)
1884 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1886 if (noswap
|| !total_swap_pages
)
1888 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1893 #if MAX_NUMNODES > 1
1896 * Always updating the nodemask is not very good - even if we have an empty
1897 * list or the wrong list here, we can start from some node and traverse all
1898 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1901 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1905 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1906 * pagein/pageout changes since the last update.
1908 if (!atomic_read(&memcg
->numainfo_events
))
1910 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1913 /* make a nodemask where this memcg uses memory from */
1914 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1916 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1918 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1919 node_clear(nid
, memcg
->scan_nodes
);
1922 atomic_set(&memcg
->numainfo_events
, 0);
1923 atomic_set(&memcg
->numainfo_updating
, 0);
1927 * Selecting a node where we start reclaim from. Because what we need is just
1928 * reducing usage counter, start from anywhere is O,K. Considering
1929 * memory reclaim from current node, there are pros. and cons.
1931 * Freeing memory from current node means freeing memory from a node which
1932 * we'll use or we've used. So, it may make LRU bad. And if several threads
1933 * hit limits, it will see a contention on a node. But freeing from remote
1934 * node means more costs for memory reclaim because of memory latency.
1936 * Now, we use round-robin. Better algorithm is welcomed.
1938 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1942 mem_cgroup_may_update_nodemask(memcg
);
1943 node
= memcg
->last_scanned_node
;
1945 node
= next_node(node
, memcg
->scan_nodes
);
1946 if (node
== MAX_NUMNODES
)
1947 node
= first_node(memcg
->scan_nodes
);
1949 * We call this when we hit limit, not when pages are added to LRU.
1950 * No LRU may hold pages because all pages are UNEVICTABLE or
1951 * memcg is too small and all pages are not on LRU. In that case,
1952 * we use curret node.
1954 if (unlikely(node
== MAX_NUMNODES
))
1955 node
= numa_node_id();
1957 memcg
->last_scanned_node
= node
;
1962 * Check all nodes whether it contains reclaimable pages or not.
1963 * For quick scan, we make use of scan_nodes. This will allow us to skip
1964 * unused nodes. But scan_nodes is lazily updated and may not cotain
1965 * enough new information. We need to do double check.
1967 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1972 * quick check...making use of scan_node.
1973 * We can skip unused nodes.
1975 if (!nodes_empty(memcg
->scan_nodes
)) {
1976 for (nid
= first_node(memcg
->scan_nodes
);
1978 nid
= next_node(nid
, memcg
->scan_nodes
)) {
1980 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1985 * Check rest of nodes.
1987 for_each_node_state(nid
, N_MEMORY
) {
1988 if (node_isset(nid
, memcg
->scan_nodes
))
1990 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1997 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
2002 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
2004 return test_mem_cgroup_node_reclaimable(memcg
, 0, noswap
);
2008 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
2011 unsigned long *total_scanned
)
2013 struct mem_cgroup
*victim
= NULL
;
2016 unsigned long excess
;
2017 unsigned long nr_scanned
;
2018 struct mem_cgroup_reclaim_cookie reclaim
= {
2023 excess
= res_counter_soft_limit_excess(&root_memcg
->res
) >> PAGE_SHIFT
;
2026 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
2031 * If we have not been able to reclaim
2032 * anything, it might because there are
2033 * no reclaimable pages under this hierarchy
2038 * We want to do more targeted reclaim.
2039 * excess >> 2 is not to excessive so as to
2040 * reclaim too much, nor too less that we keep
2041 * coming back to reclaim from this cgroup
2043 if (total
>= (excess
>> 2) ||
2044 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
2049 if (!mem_cgroup_reclaimable(victim
, false))
2051 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
2053 *total_scanned
+= nr_scanned
;
2054 if (!res_counter_soft_limit_excess(&root_memcg
->res
))
2057 mem_cgroup_iter_break(root_memcg
, victim
);
2061 #ifdef CONFIG_LOCKDEP
2062 static struct lockdep_map memcg_oom_lock_dep_map
= {
2063 .name
= "memcg_oom_lock",
2067 static DEFINE_SPINLOCK(memcg_oom_lock
);
2070 * Check OOM-Killer is already running under our hierarchy.
2071 * If someone is running, return false.
2073 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
2075 struct mem_cgroup
*iter
, *failed
= NULL
;
2077 spin_lock(&memcg_oom_lock
);
2079 for_each_mem_cgroup_tree(iter
, memcg
) {
2080 if (iter
->oom_lock
) {
2082 * this subtree of our hierarchy is already locked
2083 * so we cannot give a lock.
2086 mem_cgroup_iter_break(memcg
, iter
);
2089 iter
->oom_lock
= true;
2094 * OK, we failed to lock the whole subtree so we have
2095 * to clean up what we set up to the failing subtree
2097 for_each_mem_cgroup_tree(iter
, memcg
) {
2098 if (iter
== failed
) {
2099 mem_cgroup_iter_break(memcg
, iter
);
2102 iter
->oom_lock
= false;
2105 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
2107 spin_unlock(&memcg_oom_lock
);
2112 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
2114 struct mem_cgroup
*iter
;
2116 spin_lock(&memcg_oom_lock
);
2117 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
2118 for_each_mem_cgroup_tree(iter
, memcg
)
2119 iter
->oom_lock
= false;
2120 spin_unlock(&memcg_oom_lock
);
2123 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
2125 struct mem_cgroup
*iter
;
2127 for_each_mem_cgroup_tree(iter
, memcg
)
2128 atomic_inc(&iter
->under_oom
);
2131 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
2133 struct mem_cgroup
*iter
;
2136 * When a new child is created while the hierarchy is under oom,
2137 * mem_cgroup_oom_lock() may not be called. We have to use
2138 * atomic_add_unless() here.
2140 for_each_mem_cgroup_tree(iter
, memcg
)
2141 atomic_add_unless(&iter
->under_oom
, -1, 0);
2144 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
2146 struct oom_wait_info
{
2147 struct mem_cgroup
*memcg
;
2151 static int memcg_oom_wake_function(wait_queue_t
*wait
,
2152 unsigned mode
, int sync
, void *arg
)
2154 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
2155 struct mem_cgroup
*oom_wait_memcg
;
2156 struct oom_wait_info
*oom_wait_info
;
2158 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
2159 oom_wait_memcg
= oom_wait_info
->memcg
;
2162 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2163 * Then we can use css_is_ancestor without taking care of RCU.
2165 if (!mem_cgroup_same_or_subtree(oom_wait_memcg
, wake_memcg
)
2166 && !mem_cgroup_same_or_subtree(wake_memcg
, oom_wait_memcg
))
2168 return autoremove_wake_function(wait
, mode
, sync
, arg
);
2171 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
2173 atomic_inc(&memcg
->oom_wakeups
);
2174 /* for filtering, pass "memcg" as argument. */
2175 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
2178 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
2180 if (memcg
&& atomic_read(&memcg
->under_oom
))
2181 memcg_wakeup_oom(memcg
);
2184 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
2186 if (!current
->memcg_oom
.may_oom
)
2189 * We are in the middle of the charge context here, so we
2190 * don't want to block when potentially sitting on a callstack
2191 * that holds all kinds of filesystem and mm locks.
2193 * Also, the caller may handle a failed allocation gracefully
2194 * (like optional page cache readahead) and so an OOM killer
2195 * invocation might not even be necessary.
2197 * That's why we don't do anything here except remember the
2198 * OOM context and then deal with it at the end of the page
2199 * fault when the stack is unwound, the locks are released,
2200 * and when we know whether the fault was overall successful.
2202 css_get(&memcg
->css
);
2203 current
->memcg_oom
.memcg
= memcg
;
2204 current
->memcg_oom
.gfp_mask
= mask
;
2205 current
->memcg_oom
.order
= order
;
2209 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2210 * @handle: actually kill/wait or just clean up the OOM state
2212 * This has to be called at the end of a page fault if the memcg OOM
2213 * handler was enabled.
2215 * Memcg supports userspace OOM handling where failed allocations must
2216 * sleep on a waitqueue until the userspace task resolves the
2217 * situation. Sleeping directly in the charge context with all kinds
2218 * of locks held is not a good idea, instead we remember an OOM state
2219 * in the task and mem_cgroup_oom_synchronize() has to be called at
2220 * the end of the page fault to complete the OOM handling.
2222 * Returns %true if an ongoing memcg OOM situation was detected and
2223 * completed, %false otherwise.
2225 bool mem_cgroup_oom_synchronize(bool handle
)
2227 struct mem_cgroup
*memcg
= current
->memcg_oom
.memcg
;
2228 struct oom_wait_info owait
;
2231 /* OOM is global, do not handle */
2238 owait
.memcg
= memcg
;
2239 owait
.wait
.flags
= 0;
2240 owait
.wait
.func
= memcg_oom_wake_function
;
2241 owait
.wait
.private = current
;
2242 INIT_LIST_HEAD(&owait
.wait
.task_list
);
2244 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2245 mem_cgroup_mark_under_oom(memcg
);
2247 locked
= mem_cgroup_oom_trylock(memcg
);
2250 mem_cgroup_oom_notify(memcg
);
2252 if (locked
&& !memcg
->oom_kill_disable
) {
2253 mem_cgroup_unmark_under_oom(memcg
);
2254 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2255 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom
.gfp_mask
,
2256 current
->memcg_oom
.order
);
2259 mem_cgroup_unmark_under_oom(memcg
);
2260 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2264 mem_cgroup_oom_unlock(memcg
);
2266 * There is no guarantee that an OOM-lock contender
2267 * sees the wakeups triggered by the OOM kill
2268 * uncharges. Wake any sleepers explicitely.
2270 memcg_oom_recover(memcg
);
2273 current
->memcg_oom
.memcg
= NULL
;
2274 css_put(&memcg
->css
);
2279 * Used to update mapped file or writeback or other statistics.
2281 * Notes: Race condition
2283 * We usually use lock_page_cgroup() for accessing page_cgroup member but
2284 * it tends to be costly. But considering some conditions, we doesn't need
2285 * to do so _always_.
2287 * Considering "charge", lock_page_cgroup() is not required because all
2288 * file-stat operations happen after a page is attached to radix-tree. There
2289 * are no race with "charge".
2291 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2292 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2293 * if there are race with "uncharge". Statistics itself is properly handled
2296 * Considering "move", this is an only case we see a race. To make the race
2297 * small, we check memcg->moving_account and detect there are possibility
2298 * of race or not. If there is, we take a lock.
2301 void __mem_cgroup_begin_update_page_stat(struct page
*page
,
2302 bool *locked
, unsigned long *flags
)
2304 struct mem_cgroup
*memcg
;
2305 struct page_cgroup
*pc
;
2307 pc
= lookup_page_cgroup(page
);
2309 memcg
= pc
->mem_cgroup
;
2310 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2313 * If this memory cgroup is not under account moving, we don't
2314 * need to take move_lock_mem_cgroup(). Because we already hold
2315 * rcu_read_lock(), any calls to move_account will be delayed until
2316 * rcu_read_unlock().
2318 VM_BUG_ON(!rcu_read_lock_held());
2319 if (atomic_read(&memcg
->moving_account
) <= 0)
2322 move_lock_mem_cgroup(memcg
, flags
);
2323 if (memcg
!= pc
->mem_cgroup
|| !PageCgroupUsed(pc
)) {
2324 move_unlock_mem_cgroup(memcg
, flags
);
2330 void __mem_cgroup_end_update_page_stat(struct page
*page
, unsigned long *flags
)
2332 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2335 * It's guaranteed that pc->mem_cgroup never changes while
2336 * lock is held because a routine modifies pc->mem_cgroup
2337 * should take move_lock_mem_cgroup().
2339 move_unlock_mem_cgroup(pc
->mem_cgroup
, flags
);
2342 void mem_cgroup_update_page_stat(struct page
*page
,
2343 enum mem_cgroup_stat_index idx
, int val
)
2345 struct mem_cgroup
*memcg
;
2346 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2347 unsigned long uninitialized_var(flags
);
2349 if (mem_cgroup_disabled())
2352 VM_BUG_ON(!rcu_read_lock_held());
2353 memcg
= pc
->mem_cgroup
;
2354 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2357 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2361 * size of first charge trial. "32" comes from vmscan.c's magic value.
2362 * TODO: maybe necessary to use big numbers in big irons.
2364 #define CHARGE_BATCH 32U
2365 struct memcg_stock_pcp
{
2366 struct mem_cgroup
*cached
; /* this never be root cgroup */
2367 unsigned int nr_pages
;
2368 struct work_struct work
;
2369 unsigned long flags
;
2370 #define FLUSHING_CACHED_CHARGE 0
2372 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2373 static DEFINE_MUTEX(percpu_charge_mutex
);
2376 * consume_stock: Try to consume stocked charge on this cpu.
2377 * @memcg: memcg to consume from.
2378 * @nr_pages: how many pages to charge.
2380 * The charges will only happen if @memcg matches the current cpu's memcg
2381 * stock, and at least @nr_pages are available in that stock. Failure to
2382 * service an allocation will refill the stock.
2384 * returns true if successful, false otherwise.
2386 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2388 struct memcg_stock_pcp
*stock
;
2391 if (nr_pages
> CHARGE_BATCH
)
2394 stock
= &get_cpu_var(memcg_stock
);
2395 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
)
2396 stock
->nr_pages
-= nr_pages
;
2397 else /* need to call res_counter_charge */
2399 put_cpu_var(memcg_stock
);
2404 * Returns stocks cached in percpu to res_counter and reset cached information.
2406 static void drain_stock(struct memcg_stock_pcp
*stock
)
2408 struct mem_cgroup
*old
= stock
->cached
;
2410 if (stock
->nr_pages
) {
2411 unsigned long bytes
= stock
->nr_pages
* PAGE_SIZE
;
2413 res_counter_uncharge(&old
->res
, bytes
);
2414 if (do_swap_account
)
2415 res_counter_uncharge(&old
->memsw
, bytes
);
2416 stock
->nr_pages
= 0;
2418 stock
->cached
= NULL
;
2422 * This must be called under preempt disabled or must be called by
2423 * a thread which is pinned to local cpu.
2425 static void drain_local_stock(struct work_struct
*dummy
)
2427 struct memcg_stock_pcp
*stock
= this_cpu_ptr(&memcg_stock
);
2429 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2432 static void __init
memcg_stock_init(void)
2436 for_each_possible_cpu(cpu
) {
2437 struct memcg_stock_pcp
*stock
=
2438 &per_cpu(memcg_stock
, cpu
);
2439 INIT_WORK(&stock
->work
, drain_local_stock
);
2444 * Cache charges(val) which is from res_counter, to local per_cpu area.
2445 * This will be consumed by consume_stock() function, later.
2447 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2449 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2451 if (stock
->cached
!= memcg
) { /* reset if necessary */
2453 stock
->cached
= memcg
;
2455 stock
->nr_pages
+= nr_pages
;
2456 put_cpu_var(memcg_stock
);
2460 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2461 * of the hierarchy under it. sync flag says whether we should block
2462 * until the work is done.
2464 static void drain_all_stock(struct mem_cgroup
*root_memcg
, bool sync
)
2468 /* Notify other cpus that system-wide "drain" is running */
2471 for_each_online_cpu(cpu
) {
2472 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2473 struct mem_cgroup
*memcg
;
2475 memcg
= stock
->cached
;
2476 if (!memcg
|| !stock
->nr_pages
)
2478 if (!mem_cgroup_same_or_subtree(root_memcg
, memcg
))
2480 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2482 drain_local_stock(&stock
->work
);
2484 schedule_work_on(cpu
, &stock
->work
);
2492 for_each_online_cpu(cpu
) {
2493 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2494 if (test_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
))
2495 flush_work(&stock
->work
);
2502 * Tries to drain stocked charges in other cpus. This function is asynchronous
2503 * and just put a work per cpu for draining localy on each cpu. Caller can
2504 * expects some charges will be back to res_counter later but cannot wait for
2507 static void drain_all_stock_async(struct mem_cgroup
*root_memcg
)
2510 * If someone calls draining, avoid adding more kworker runs.
2512 if (!mutex_trylock(&percpu_charge_mutex
))
2514 drain_all_stock(root_memcg
, false);
2515 mutex_unlock(&percpu_charge_mutex
);
2518 /* This is a synchronous drain interface. */
2519 static void drain_all_stock_sync(struct mem_cgroup
*root_memcg
)
2521 /* called when force_empty is called */
2522 mutex_lock(&percpu_charge_mutex
);
2523 drain_all_stock(root_memcg
, true);
2524 mutex_unlock(&percpu_charge_mutex
);
2528 * This function drains percpu counter value from DEAD cpu and
2529 * move it to local cpu. Note that this function can be preempted.
2531 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2535 spin_lock(&memcg
->pcp_counter_lock
);
2536 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2537 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2539 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2540 memcg
->nocpu_base
.count
[i
] += x
;
2542 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2543 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2545 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2546 memcg
->nocpu_base
.events
[i
] += x
;
2548 spin_unlock(&memcg
->pcp_counter_lock
);
2551 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2552 unsigned long action
,
2555 int cpu
= (unsigned long)hcpu
;
2556 struct memcg_stock_pcp
*stock
;
2557 struct mem_cgroup
*iter
;
2559 if (action
== CPU_ONLINE
)
2562 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2565 for_each_mem_cgroup(iter
)
2566 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2568 stock
= &per_cpu(memcg_stock
, cpu
);
2574 /* See mem_cgroup_try_charge() for details */
2576 CHARGE_OK
, /* success */
2577 CHARGE_RETRY
, /* need to retry but retry is not bad */
2578 CHARGE_NOMEM
, /* we can't do more. return -ENOMEM */
2579 CHARGE_WOULDBLOCK
, /* GFP_WAIT wasn't set and no enough res. */
2582 static int mem_cgroup_do_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2583 unsigned int nr_pages
, unsigned int min_pages
,
2586 unsigned long csize
= nr_pages
* PAGE_SIZE
;
2587 struct mem_cgroup
*mem_over_limit
;
2588 struct res_counter
*fail_res
;
2589 unsigned long flags
= 0;
2592 ret
= res_counter_charge(&memcg
->res
, csize
, &fail_res
);
2595 if (!do_swap_account
)
2597 ret
= res_counter_charge(&memcg
->memsw
, csize
, &fail_res
);
2601 res_counter_uncharge(&memcg
->res
, csize
);
2602 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
2603 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
2605 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
2607 * Never reclaim on behalf of optional batching, retry with a
2608 * single page instead.
2610 if (nr_pages
> min_pages
)
2611 return CHARGE_RETRY
;
2613 if (!(gfp_mask
& __GFP_WAIT
))
2614 return CHARGE_WOULDBLOCK
;
2616 if (gfp_mask
& __GFP_NORETRY
)
2617 return CHARGE_NOMEM
;
2619 ret
= mem_cgroup_reclaim(mem_over_limit
, gfp_mask
, flags
);
2620 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2621 return CHARGE_RETRY
;
2623 * Even though the limit is exceeded at this point, reclaim
2624 * may have been able to free some pages. Retry the charge
2625 * before killing the task.
2627 * Only for regular pages, though: huge pages are rather
2628 * unlikely to succeed so close to the limit, and we fall back
2629 * to regular pages anyway in case of failure.
2631 if (nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
) && ret
)
2632 return CHARGE_RETRY
;
2635 * At task move, charge accounts can be doubly counted. So, it's
2636 * better to wait until the end of task_move if something is going on.
2638 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2639 return CHARGE_RETRY
;
2642 mem_cgroup_oom(mem_over_limit
, gfp_mask
, get_order(csize
));
2644 return CHARGE_NOMEM
;
2648 * mem_cgroup_try_charge - try charging a memcg
2649 * @memcg: memcg to charge
2650 * @nr_pages: number of pages to charge
2651 * @oom: trigger OOM if reclaim fails
2653 * Returns 0 if @memcg was charged successfully, -EINTR if the charge
2654 * was bypassed to root_mem_cgroup, and -ENOMEM if the charge failed.
2656 static int mem_cgroup_try_charge(struct mem_cgroup
*memcg
,
2658 unsigned int nr_pages
,
2661 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2662 int nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2665 if (mem_cgroup_is_root(memcg
))
2668 * Unlike in global OOM situations, memcg is not in a physical
2669 * memory shortage. Allow dying and OOM-killed tasks to
2670 * bypass the last charges so that they can exit quickly and
2671 * free their memory.
2673 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
2674 fatal_signal_pending(current
) ||
2675 current
->flags
& PF_EXITING
))
2678 if (unlikely(task_in_memcg_oom(current
)))
2681 if (gfp_mask
& __GFP_NOFAIL
)
2684 if (consume_stock(memcg
, nr_pages
))
2688 bool invoke_oom
= oom
&& !nr_oom_retries
;
2690 /* If killed, bypass charge */
2691 if (fatal_signal_pending(current
))
2694 ret
= mem_cgroup_do_charge(memcg
, gfp_mask
, batch
,
2695 nr_pages
, invoke_oom
);
2699 case CHARGE_RETRY
: /* not in OOM situation but retry */
2702 case CHARGE_WOULDBLOCK
: /* !__GFP_WAIT */
2704 case CHARGE_NOMEM
: /* OOM routine works */
2705 if (!oom
|| invoke_oom
)
2710 } while (ret
!= CHARGE_OK
);
2712 if (batch
> nr_pages
)
2713 refill_stock(memcg
, batch
- nr_pages
);
2717 if (!(gfp_mask
& __GFP_NOFAIL
))
2724 * mem_cgroup_try_charge_mm - try charging a mm
2725 * @mm: mm_struct to charge
2726 * @nr_pages: number of pages to charge
2727 * @oom: trigger OOM if reclaim fails
2729 * Returns the charged mem_cgroup associated with the given mm_struct or
2730 * NULL the charge failed.
2732 static struct mem_cgroup
*mem_cgroup_try_charge_mm(struct mm_struct
*mm
,
2734 unsigned int nr_pages
,
2738 struct mem_cgroup
*memcg
;
2741 memcg
= get_mem_cgroup_from_mm(mm
);
2742 ret
= mem_cgroup_try_charge(memcg
, gfp_mask
, nr_pages
, oom
);
2743 css_put(&memcg
->css
);
2745 memcg
= root_mem_cgroup
;
2753 * Somemtimes we have to undo a charge we got by try_charge().
2754 * This function is for that and do uncharge, put css's refcnt.
2755 * gotten by try_charge().
2757 static void __mem_cgroup_cancel_charge(struct mem_cgroup
*memcg
,
2758 unsigned int nr_pages
)
2760 if (!mem_cgroup_is_root(memcg
)) {
2761 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2763 res_counter_uncharge(&memcg
->res
, bytes
);
2764 if (do_swap_account
)
2765 res_counter_uncharge(&memcg
->memsw
, bytes
);
2770 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2771 * This is useful when moving usage to parent cgroup.
2773 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup
*memcg
,
2774 unsigned int nr_pages
)
2776 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2778 if (mem_cgroup_is_root(memcg
))
2781 res_counter_uncharge_until(&memcg
->res
, memcg
->res
.parent
, bytes
);
2782 if (do_swap_account
)
2783 res_counter_uncharge_until(&memcg
->memsw
,
2784 memcg
->memsw
.parent
, bytes
);
2788 * A helper function to get mem_cgroup from ID. must be called under
2789 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2790 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2791 * called against removed memcg.)
2793 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2795 /* ID 0 is unused ID */
2798 return mem_cgroup_from_id(id
);
2801 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2803 struct mem_cgroup
*memcg
= NULL
;
2804 struct page_cgroup
*pc
;
2808 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2810 pc
= lookup_page_cgroup(page
);
2811 lock_page_cgroup(pc
);
2812 if (PageCgroupUsed(pc
)) {
2813 memcg
= pc
->mem_cgroup
;
2814 if (memcg
&& !css_tryget(&memcg
->css
))
2816 } else if (PageSwapCache(page
)) {
2817 ent
.val
= page_private(page
);
2818 id
= lookup_swap_cgroup_id(ent
);
2820 memcg
= mem_cgroup_lookup(id
);
2821 if (memcg
&& !css_tryget(&memcg
->css
))
2825 unlock_page_cgroup(pc
);
2829 static void __mem_cgroup_commit_charge(struct mem_cgroup
*memcg
,
2831 unsigned int nr_pages
,
2832 enum charge_type ctype
,
2835 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2836 struct zone
*uninitialized_var(zone
);
2837 struct lruvec
*lruvec
;
2838 bool was_on_lru
= false;
2841 lock_page_cgroup(pc
);
2842 VM_BUG_ON_PAGE(PageCgroupUsed(pc
), page
);
2844 * we don't need page_cgroup_lock about tail pages, becase they are not
2845 * accessed by any other context at this point.
2849 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2850 * may already be on some other mem_cgroup's LRU. Take care of it.
2853 zone
= page_zone(page
);
2854 spin_lock_irq(&zone
->lru_lock
);
2855 if (PageLRU(page
)) {
2856 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2858 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2863 pc
->mem_cgroup
= memcg
;
2865 * We access a page_cgroup asynchronously without lock_page_cgroup().
2866 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2867 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2868 * before USED bit, we need memory barrier here.
2869 * See mem_cgroup_add_lru_list(), etc.
2872 SetPageCgroupUsed(pc
);
2876 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2877 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2879 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2881 spin_unlock_irq(&zone
->lru_lock
);
2884 if (ctype
== MEM_CGROUP_CHARGE_TYPE_ANON
)
2889 mem_cgroup_charge_statistics(memcg
, page
, anon
, nr_pages
);
2890 unlock_page_cgroup(pc
);
2893 * "charge_statistics" updated event counter. Then, check it.
2894 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2895 * if they exceeds softlimit.
2897 memcg_check_events(memcg
, page
);
2900 static DEFINE_MUTEX(set_limit_mutex
);
2902 #ifdef CONFIG_MEMCG_KMEM
2904 * The memcg_slab_mutex is held whenever a per memcg kmem cache is created or
2905 * destroyed. It protects memcg_caches arrays and memcg_slab_caches lists.
2907 static DEFINE_MUTEX(memcg_slab_mutex
);
2909 static DEFINE_MUTEX(activate_kmem_mutex
);
2911 static inline bool memcg_can_account_kmem(struct mem_cgroup
*memcg
)
2913 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
) &&
2914 memcg_kmem_is_active(memcg
);
2918 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2919 * in the memcg_cache_params struct.
2921 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
2923 struct kmem_cache
*cachep
;
2925 VM_BUG_ON(p
->is_root_cache
);
2926 cachep
= p
->root_cache
;
2927 return cache_from_memcg_idx(cachep
, memcg_cache_id(p
->memcg
));
2930 #ifdef CONFIG_SLABINFO
2931 static int mem_cgroup_slabinfo_read(struct seq_file
*m
, void *v
)
2933 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
2934 struct memcg_cache_params
*params
;
2936 if (!memcg_can_account_kmem(memcg
))
2939 print_slabinfo_header(m
);
2941 mutex_lock(&memcg_slab_mutex
);
2942 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
)
2943 cache_show(memcg_params_to_cache(params
), m
);
2944 mutex_unlock(&memcg_slab_mutex
);
2950 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
, u64 size
)
2952 struct res_counter
*fail_res
;
2955 ret
= res_counter_charge(&memcg
->kmem
, size
, &fail_res
);
2959 ret
= mem_cgroup_try_charge(memcg
, gfp
, size
>> PAGE_SHIFT
,
2960 oom_gfp_allowed(gfp
));
2961 if (ret
== -EINTR
) {
2963 * mem_cgroup_try_charge() chosed to bypass to root due to
2964 * OOM kill or fatal signal. Since our only options are to
2965 * either fail the allocation or charge it to this cgroup, do
2966 * it as a temporary condition. But we can't fail. From a
2967 * kmem/slab perspective, the cache has already been selected,
2968 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2971 * This condition will only trigger if the task entered
2972 * memcg_charge_kmem in a sane state, but was OOM-killed during
2973 * mem_cgroup_try_charge() above. Tasks that were already
2974 * dying when the allocation triggers should have been already
2975 * directed to the root cgroup in memcontrol.h
2977 res_counter_charge_nofail(&memcg
->res
, size
, &fail_res
);
2978 if (do_swap_account
)
2979 res_counter_charge_nofail(&memcg
->memsw
, size
,
2983 res_counter_uncharge(&memcg
->kmem
, size
);
2988 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, u64 size
)
2990 res_counter_uncharge(&memcg
->res
, size
);
2991 if (do_swap_account
)
2992 res_counter_uncharge(&memcg
->memsw
, size
);
2995 if (res_counter_uncharge(&memcg
->kmem
, size
))
2999 * Releases a reference taken in kmem_cgroup_css_offline in case
3000 * this last uncharge is racing with the offlining code or it is
3001 * outliving the memcg existence.
3003 * The memory barrier imposed by test&clear is paired with the
3004 * explicit one in memcg_kmem_mark_dead().
3006 if (memcg_kmem_test_and_clear_dead(memcg
))
3007 css_put(&memcg
->css
);
3011 * helper for acessing a memcg's index. It will be used as an index in the
3012 * child cache array in kmem_cache, and also to derive its name. This function
3013 * will return -1 when this is not a kmem-limited memcg.
3015 int memcg_cache_id(struct mem_cgroup
*memcg
)
3017 return memcg
? memcg
->kmemcg_id
: -1;
3020 static size_t memcg_caches_array_size(int num_groups
)
3023 if (num_groups
<= 0)
3026 size
= 2 * num_groups
;
3027 if (size
< MEMCG_CACHES_MIN_SIZE
)
3028 size
= MEMCG_CACHES_MIN_SIZE
;
3029 else if (size
> MEMCG_CACHES_MAX_SIZE
)
3030 size
= MEMCG_CACHES_MAX_SIZE
;
3036 * We should update the current array size iff all caches updates succeed. This
3037 * can only be done from the slab side. The slab mutex needs to be held when
3040 void memcg_update_array_size(int num
)
3042 if (num
> memcg_limited_groups_array_size
)
3043 memcg_limited_groups_array_size
= memcg_caches_array_size(num
);
3046 int memcg_update_cache_size(struct kmem_cache
*s
, int num_groups
)
3048 struct memcg_cache_params
*cur_params
= s
->memcg_params
;
3050 VM_BUG_ON(!is_root_cache(s
));
3052 if (num_groups
> memcg_limited_groups_array_size
) {
3054 struct memcg_cache_params
*new_params
;
3055 ssize_t size
= memcg_caches_array_size(num_groups
);
3057 size
*= sizeof(void *);
3058 size
+= offsetof(struct memcg_cache_params
, memcg_caches
);
3060 new_params
= kzalloc(size
, GFP_KERNEL
);
3064 new_params
->is_root_cache
= true;
3067 * There is the chance it will be bigger than
3068 * memcg_limited_groups_array_size, if we failed an allocation
3069 * in a cache, in which case all caches updated before it, will
3070 * have a bigger array.
3072 * But if that is the case, the data after
3073 * memcg_limited_groups_array_size is certainly unused
3075 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3076 if (!cur_params
->memcg_caches
[i
])
3078 new_params
->memcg_caches
[i
] =
3079 cur_params
->memcg_caches
[i
];
3083 * Ideally, we would wait until all caches succeed, and only
3084 * then free the old one. But this is not worth the extra
3085 * pointer per-cache we'd have to have for this.
3087 * It is not a big deal if some caches are left with a size
3088 * bigger than the others. And all updates will reset this
3091 rcu_assign_pointer(s
->memcg_params
, new_params
);
3093 kfree_rcu(cur_params
, rcu_head
);
3098 int memcg_alloc_cache_params(struct mem_cgroup
*memcg
, struct kmem_cache
*s
,
3099 struct kmem_cache
*root_cache
)
3103 if (!memcg_kmem_enabled())
3107 size
= offsetof(struct memcg_cache_params
, memcg_caches
);
3108 size
+= memcg_limited_groups_array_size
* sizeof(void *);
3110 size
= sizeof(struct memcg_cache_params
);
3112 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3113 if (!s
->memcg_params
)
3117 s
->memcg_params
->memcg
= memcg
;
3118 s
->memcg_params
->root_cache
= root_cache
;
3119 css_get(&memcg
->css
);
3121 s
->memcg_params
->is_root_cache
= true;
3126 void memcg_free_cache_params(struct kmem_cache
*s
)
3128 if (!s
->memcg_params
)
3130 if (!s
->memcg_params
->is_root_cache
)
3131 css_put(&s
->memcg_params
->memcg
->css
);
3132 kfree(s
->memcg_params
);
3135 static void memcg_kmem_create_cache(struct mem_cgroup
*memcg
,
3136 struct kmem_cache
*root_cache
)
3138 static char *memcg_name_buf
; /* protected by memcg_slab_mutex */
3139 struct kmem_cache
*cachep
;
3142 lockdep_assert_held(&memcg_slab_mutex
);
3144 id
= memcg_cache_id(memcg
);
3147 * Since per-memcg caches are created asynchronously on first
3148 * allocation (see memcg_kmem_get_cache()), several threads can try to
3149 * create the same cache, but only one of them may succeed.
3151 if (cache_from_memcg_idx(root_cache
, id
))
3154 if (!memcg_name_buf
) {
3155 memcg_name_buf
= kmalloc(NAME_MAX
+ 1, GFP_KERNEL
);
3156 if (!memcg_name_buf
)
3160 cgroup_name(memcg
->css
.cgroup
, memcg_name_buf
, NAME_MAX
+ 1);
3161 cachep
= kmem_cache_create_memcg(memcg
, root_cache
, memcg_name_buf
);
3163 * If we could not create a memcg cache, do not complain, because
3164 * that's not critical at all as we can always proceed with the root
3170 list_add(&cachep
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
3173 * Since readers won't lock (see cache_from_memcg_idx()), we need a
3174 * barrier here to ensure nobody will see the kmem_cache partially
3179 BUG_ON(root_cache
->memcg_params
->memcg_caches
[id
]);
3180 root_cache
->memcg_params
->memcg_caches
[id
] = cachep
;
3183 static void memcg_kmem_destroy_cache(struct kmem_cache
*cachep
)
3185 struct kmem_cache
*root_cache
;
3186 struct mem_cgroup
*memcg
;
3189 lockdep_assert_held(&memcg_slab_mutex
);
3191 BUG_ON(is_root_cache(cachep
));
3193 root_cache
= cachep
->memcg_params
->root_cache
;
3194 memcg
= cachep
->memcg_params
->memcg
;
3195 id
= memcg_cache_id(memcg
);
3197 BUG_ON(root_cache
->memcg_params
->memcg_caches
[id
] != cachep
);
3198 root_cache
->memcg_params
->memcg_caches
[id
] = NULL
;
3200 list_del(&cachep
->memcg_params
->list
);
3202 kmem_cache_destroy(cachep
);
3206 * During the creation a new cache, we need to disable our accounting mechanism
3207 * altogether. This is true even if we are not creating, but rather just
3208 * enqueing new caches to be created.
3210 * This is because that process will trigger allocations; some visible, like
3211 * explicit kmallocs to auxiliary data structures, name strings and internal
3212 * cache structures; some well concealed, like INIT_WORK() that can allocate
3213 * objects during debug.
3215 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3216 * to it. This may not be a bounded recursion: since the first cache creation
3217 * failed to complete (waiting on the allocation), we'll just try to create the
3218 * cache again, failing at the same point.
3220 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3221 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3222 * inside the following two functions.
3224 static inline void memcg_stop_kmem_account(void)
3226 VM_BUG_ON(!current
->mm
);
3227 current
->memcg_kmem_skip_account
++;
3230 static inline void memcg_resume_kmem_account(void)
3232 VM_BUG_ON(!current
->mm
);
3233 current
->memcg_kmem_skip_account
--;
3236 int __kmem_cache_destroy_memcg_children(struct kmem_cache
*s
)
3238 struct kmem_cache
*c
;
3241 mutex_lock(&memcg_slab_mutex
);
3242 for_each_memcg_cache_index(i
) {
3243 c
= cache_from_memcg_idx(s
, i
);
3247 memcg_kmem_destroy_cache(c
);
3249 if (cache_from_memcg_idx(s
, i
))
3252 mutex_unlock(&memcg_slab_mutex
);
3256 static void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3258 struct kmem_cache
*cachep
;
3259 struct memcg_cache_params
*params
, *tmp
;
3261 if (!memcg_kmem_is_active(memcg
))
3264 mutex_lock(&memcg_slab_mutex
);
3265 list_for_each_entry_safe(params
, tmp
, &memcg
->memcg_slab_caches
, list
) {
3266 cachep
= memcg_params_to_cache(params
);
3267 kmem_cache_shrink(cachep
);
3268 if (atomic_read(&cachep
->memcg_params
->nr_pages
) == 0)
3269 memcg_kmem_destroy_cache(cachep
);
3271 mutex_unlock(&memcg_slab_mutex
);
3274 struct create_work
{
3275 struct mem_cgroup
*memcg
;
3276 struct kmem_cache
*cachep
;
3277 struct work_struct work
;
3280 static void memcg_create_cache_work_func(struct work_struct
*w
)
3282 struct create_work
*cw
= container_of(w
, struct create_work
, work
);
3283 struct mem_cgroup
*memcg
= cw
->memcg
;
3284 struct kmem_cache
*cachep
= cw
->cachep
;
3286 mutex_lock(&memcg_slab_mutex
);
3287 memcg_kmem_create_cache(memcg
, cachep
);
3288 mutex_unlock(&memcg_slab_mutex
);
3290 css_put(&memcg
->css
);
3295 * Enqueue the creation of a per-memcg kmem_cache.
3297 static void __memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3298 struct kmem_cache
*cachep
)
3300 struct create_work
*cw
;
3302 cw
= kmalloc(sizeof(struct create_work
), GFP_NOWAIT
);
3304 css_put(&memcg
->css
);
3309 cw
->cachep
= cachep
;
3311 INIT_WORK(&cw
->work
, memcg_create_cache_work_func
);
3312 schedule_work(&cw
->work
);
3315 static void memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3316 struct kmem_cache
*cachep
)
3319 * We need to stop accounting when we kmalloc, because if the
3320 * corresponding kmalloc cache is not yet created, the first allocation
3321 * in __memcg_create_cache_enqueue will recurse.
3323 * However, it is better to enclose the whole function. Depending on
3324 * the debugging options enabled, INIT_WORK(), for instance, can
3325 * trigger an allocation. This too, will make us recurse. Because at
3326 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3327 * the safest choice is to do it like this, wrapping the whole function.
3329 memcg_stop_kmem_account();
3330 __memcg_create_cache_enqueue(memcg
, cachep
);
3331 memcg_resume_kmem_account();
3334 int __memcg_charge_slab(struct kmem_cache
*cachep
, gfp_t gfp
, int order
)
3338 res
= memcg_charge_kmem(cachep
->memcg_params
->memcg
, gfp
,
3339 PAGE_SIZE
<< order
);
3341 atomic_add(1 << order
, &cachep
->memcg_params
->nr_pages
);
3345 void __memcg_uncharge_slab(struct kmem_cache
*cachep
, int order
)
3347 memcg_uncharge_kmem(cachep
->memcg_params
->memcg
, PAGE_SIZE
<< order
);
3348 atomic_sub(1 << order
, &cachep
->memcg_params
->nr_pages
);
3352 * Return the kmem_cache we're supposed to use for a slab allocation.
3353 * We try to use the current memcg's version of the cache.
3355 * If the cache does not exist yet, if we are the first user of it,
3356 * we either create it immediately, if possible, or create it asynchronously
3358 * In the latter case, we will let the current allocation go through with
3359 * the original cache.
3361 * Can't be called in interrupt context or from kernel threads.
3362 * This function needs to be called with rcu_read_lock() held.
3364 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
3367 struct mem_cgroup
*memcg
;
3368 struct kmem_cache
*memcg_cachep
;
3370 VM_BUG_ON(!cachep
->memcg_params
);
3371 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
3373 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3377 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
3379 if (!memcg_can_account_kmem(memcg
))
3382 memcg_cachep
= cache_from_memcg_idx(cachep
, memcg_cache_id(memcg
));
3383 if (likely(memcg_cachep
)) {
3384 cachep
= memcg_cachep
;
3388 /* The corresponding put will be done in the workqueue. */
3389 if (!css_tryget(&memcg
->css
))
3394 * If we are in a safe context (can wait, and not in interrupt
3395 * context), we could be be predictable and return right away.
3396 * This would guarantee that the allocation being performed
3397 * already belongs in the new cache.
3399 * However, there are some clashes that can arrive from locking.
3400 * For instance, because we acquire the slab_mutex while doing
3401 * kmem_cache_dup, this means no further allocation could happen
3402 * with the slab_mutex held.
3404 * Also, because cache creation issue get_online_cpus(), this
3405 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3406 * that ends up reversed during cpu hotplug. (cpuset allocates
3407 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3408 * better to defer everything.
3410 memcg_create_cache_enqueue(memcg
, cachep
);
3418 * We need to verify if the allocation against current->mm->owner's memcg is
3419 * possible for the given order. But the page is not allocated yet, so we'll
3420 * need a further commit step to do the final arrangements.
3422 * It is possible for the task to switch cgroups in this mean time, so at
3423 * commit time, we can't rely on task conversion any longer. We'll then use
3424 * the handle argument to return to the caller which cgroup we should commit
3425 * against. We could also return the memcg directly and avoid the pointer
3426 * passing, but a boolean return value gives better semantics considering
3427 * the compiled-out case as well.
3429 * Returning true means the allocation is possible.
3432 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
3434 struct mem_cgroup
*memcg
;
3440 * Disabling accounting is only relevant for some specific memcg
3441 * internal allocations. Therefore we would initially not have such
3442 * check here, since direct calls to the page allocator that are
3443 * accounted to kmemcg (alloc_kmem_pages and friends) only happen
3444 * outside memcg core. We are mostly concerned with cache allocations,
3445 * and by having this test at memcg_kmem_get_cache, we are already able
3446 * to relay the allocation to the root cache and bypass the memcg cache
3449 * There is one exception, though: the SLUB allocator does not create
3450 * large order caches, but rather service large kmallocs directly from
3451 * the page allocator. Therefore, the following sequence when backed by
3452 * the SLUB allocator:
3454 * memcg_stop_kmem_account();
3455 * kmalloc(<large_number>)
3456 * memcg_resume_kmem_account();
3458 * would effectively ignore the fact that we should skip accounting,
3459 * since it will drive us directly to this function without passing
3460 * through the cache selector memcg_kmem_get_cache. Such large
3461 * allocations are extremely rare but can happen, for instance, for the
3462 * cache arrays. We bring this test here.
3464 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3467 memcg
= get_mem_cgroup_from_mm(current
->mm
);
3469 if (!memcg_can_account_kmem(memcg
)) {
3470 css_put(&memcg
->css
);
3474 ret
= memcg_charge_kmem(memcg
, gfp
, PAGE_SIZE
<< order
);
3478 css_put(&memcg
->css
);
3482 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
3485 struct page_cgroup
*pc
;
3487 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3489 /* The page allocation failed. Revert */
3491 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3495 pc
= lookup_page_cgroup(page
);
3496 lock_page_cgroup(pc
);
3497 pc
->mem_cgroup
= memcg
;
3498 SetPageCgroupUsed(pc
);
3499 unlock_page_cgroup(pc
);
3502 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
3504 struct mem_cgroup
*memcg
= NULL
;
3505 struct page_cgroup
*pc
;
3508 pc
= lookup_page_cgroup(page
);
3510 * Fast unlocked return. Theoretically might have changed, have to
3511 * check again after locking.
3513 if (!PageCgroupUsed(pc
))
3516 lock_page_cgroup(pc
);
3517 if (PageCgroupUsed(pc
)) {
3518 memcg
= pc
->mem_cgroup
;
3519 ClearPageCgroupUsed(pc
);
3521 unlock_page_cgroup(pc
);
3524 * We trust that only if there is a memcg associated with the page, it
3525 * is a valid allocation
3530 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
3531 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3534 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3537 #endif /* CONFIG_MEMCG_KMEM */
3539 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3541 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3543 * Because tail pages are not marked as "used", set it. We're under
3544 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3545 * charge/uncharge will be never happen and move_account() is done under
3546 * compound_lock(), so we don't have to take care of races.
3548 void mem_cgroup_split_huge_fixup(struct page
*head
)
3550 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
3551 struct page_cgroup
*pc
;
3552 struct mem_cgroup
*memcg
;
3555 if (mem_cgroup_disabled())
3558 memcg
= head_pc
->mem_cgroup
;
3559 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3561 pc
->mem_cgroup
= memcg
;
3562 smp_wmb();/* see __commit_charge() */
3563 pc
->flags
= head_pc
->flags
& ~PCGF_NOCOPY_AT_SPLIT
;
3565 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
3568 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3571 * mem_cgroup_move_account - move account of the page
3573 * @nr_pages: number of regular pages (>1 for huge pages)
3574 * @pc: page_cgroup of the page.
3575 * @from: mem_cgroup which the page is moved from.
3576 * @to: mem_cgroup which the page is moved to. @from != @to.
3578 * The caller must confirm following.
3579 * - page is not on LRU (isolate_page() is useful.)
3580 * - compound_lock is held when nr_pages > 1
3582 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3585 static int mem_cgroup_move_account(struct page
*page
,
3586 unsigned int nr_pages
,
3587 struct page_cgroup
*pc
,
3588 struct mem_cgroup
*from
,
3589 struct mem_cgroup
*to
)
3591 unsigned long flags
;
3593 bool anon
= PageAnon(page
);
3595 VM_BUG_ON(from
== to
);
3596 VM_BUG_ON_PAGE(PageLRU(page
), page
);
3598 * The page is isolated from LRU. So, collapse function
3599 * will not handle this page. But page splitting can happen.
3600 * Do this check under compound_page_lock(). The caller should
3604 if (nr_pages
> 1 && !PageTransHuge(page
))
3607 lock_page_cgroup(pc
);
3610 if (!PageCgroupUsed(pc
) || pc
->mem_cgroup
!= from
)
3613 move_lock_mem_cgroup(from
, &flags
);
3615 if (!anon
&& page_mapped(page
)) {
3616 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
3618 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
3622 if (PageWriteback(page
)) {
3623 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
3625 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
3629 mem_cgroup_charge_statistics(from
, page
, anon
, -nr_pages
);
3631 /* caller should have done css_get */
3632 pc
->mem_cgroup
= to
;
3633 mem_cgroup_charge_statistics(to
, page
, anon
, nr_pages
);
3634 move_unlock_mem_cgroup(from
, &flags
);
3637 unlock_page_cgroup(pc
);
3641 memcg_check_events(to
, page
);
3642 memcg_check_events(from
, page
);
3648 * mem_cgroup_move_parent - moves page to the parent group
3649 * @page: the page to move
3650 * @pc: page_cgroup of the page
3651 * @child: page's cgroup
3653 * move charges to its parent or the root cgroup if the group has no
3654 * parent (aka use_hierarchy==0).
3655 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3656 * mem_cgroup_move_account fails) the failure is always temporary and
3657 * it signals a race with a page removal/uncharge or migration. In the
3658 * first case the page is on the way out and it will vanish from the LRU
3659 * on the next attempt and the call should be retried later.
3660 * Isolation from the LRU fails only if page has been isolated from
3661 * the LRU since we looked at it and that usually means either global
3662 * reclaim or migration going on. The page will either get back to the
3664 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3665 * (!PageCgroupUsed) or moved to a different group. The page will
3666 * disappear in the next attempt.
3668 static int mem_cgroup_move_parent(struct page
*page
,
3669 struct page_cgroup
*pc
,
3670 struct mem_cgroup
*child
)
3672 struct mem_cgroup
*parent
;
3673 unsigned int nr_pages
;
3674 unsigned long uninitialized_var(flags
);
3677 VM_BUG_ON(mem_cgroup_is_root(child
));
3680 if (!get_page_unless_zero(page
))
3682 if (isolate_lru_page(page
))
3685 nr_pages
= hpage_nr_pages(page
);
3687 parent
= parent_mem_cgroup(child
);
3689 * If no parent, move charges to root cgroup.
3692 parent
= root_mem_cgroup
;
3695 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
3696 flags
= compound_lock_irqsave(page
);
3699 ret
= mem_cgroup_move_account(page
, nr_pages
,
3702 __mem_cgroup_cancel_local_charge(child
, nr_pages
);
3705 compound_unlock_irqrestore(page
, flags
);
3706 putback_lru_page(page
);
3713 int mem_cgroup_charge_anon(struct page
*page
,
3714 struct mm_struct
*mm
, gfp_t gfp_mask
)
3716 unsigned int nr_pages
= 1;
3717 struct mem_cgroup
*memcg
;
3720 if (mem_cgroup_disabled())
3723 VM_BUG_ON_PAGE(page_mapped(page
), page
);
3724 VM_BUG_ON_PAGE(page
->mapping
&& !PageAnon(page
), page
);
3727 if (PageTransHuge(page
)) {
3728 nr_pages
<<= compound_order(page
);
3729 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
3731 * Never OOM-kill a process for a huge page. The
3732 * fault handler will fall back to regular pages.
3737 memcg
= mem_cgroup_try_charge_mm(mm
, gfp_mask
, nr_pages
, oom
);
3740 __mem_cgroup_commit_charge(memcg
, page
, nr_pages
,
3741 MEM_CGROUP_CHARGE_TYPE_ANON
, false);
3746 * While swap-in, try_charge -> commit or cancel, the page is locked.
3747 * And when try_charge() successfully returns, one refcnt to memcg without
3748 * struct page_cgroup is acquired. This refcnt will be consumed by
3749 * "commit()" or removed by "cancel()"
3751 static int __mem_cgroup_try_charge_swapin(struct mm_struct
*mm
,
3754 struct mem_cgroup
**memcgp
)
3756 struct mem_cgroup
*memcg
= NULL
;
3757 struct page_cgroup
*pc
;
3760 pc
= lookup_page_cgroup(page
);
3762 * Every swap fault against a single page tries to charge the
3763 * page, bail as early as possible. shmem_unuse() encounters
3764 * already charged pages, too. The USED bit is protected by
3765 * the page lock, which serializes swap cache removal, which
3766 * in turn serializes uncharging.
3768 if (PageCgroupUsed(pc
))
3770 if (do_swap_account
)
3771 memcg
= try_get_mem_cgroup_from_page(page
);
3773 memcg
= get_mem_cgroup_from_mm(mm
);
3774 ret
= mem_cgroup_try_charge(memcg
, mask
, 1, true);
3775 css_put(&memcg
->css
);
3777 memcg
= root_mem_cgroup
;
3785 int mem_cgroup_try_charge_swapin(struct mm_struct
*mm
, struct page
*page
,
3786 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
3788 if (mem_cgroup_disabled()) {
3793 * A racing thread's fault, or swapoff, may have already
3794 * updated the pte, and even removed page from swap cache: in
3795 * those cases unuse_pte()'s pte_same() test will fail; but
3796 * there's also a KSM case which does need to charge the page.
3798 if (!PageSwapCache(page
)) {
3799 struct mem_cgroup
*memcg
;
3801 memcg
= mem_cgroup_try_charge_mm(mm
, gfp_mask
, 1, true);
3807 return __mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, memcgp
);
3810 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup
*memcg
)
3812 if (mem_cgroup_disabled())
3816 __mem_cgroup_cancel_charge(memcg
, 1);
3820 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*memcg
,
3821 enum charge_type ctype
)
3823 if (mem_cgroup_disabled())
3828 __mem_cgroup_commit_charge(memcg
, page
, 1, ctype
, true);
3830 * Now swap is on-memory. This means this page may be
3831 * counted both as mem and swap....double count.
3832 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3833 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3834 * may call delete_from_swap_cache() before reach here.
3836 if (do_swap_account
&& PageSwapCache(page
)) {
3837 swp_entry_t ent
= {.val
= page_private(page
)};
3838 mem_cgroup_uncharge_swap(ent
);
3842 void mem_cgroup_commit_charge_swapin(struct page
*page
,
3843 struct mem_cgroup
*memcg
)
3845 __mem_cgroup_commit_charge_swapin(page
, memcg
,
3846 MEM_CGROUP_CHARGE_TYPE_ANON
);
3849 int mem_cgroup_charge_file(struct page
*page
, struct mm_struct
*mm
,
3852 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
3853 struct mem_cgroup
*memcg
;
3856 if (mem_cgroup_disabled())
3858 if (PageCompound(page
))
3861 if (PageSwapCache(page
)) { /* shmem */
3862 ret
= __mem_cgroup_try_charge_swapin(mm
, page
,
3866 __mem_cgroup_commit_charge_swapin(page
, memcg
, type
);
3870 memcg
= mem_cgroup_try_charge_mm(mm
, gfp_mask
, 1, true);
3873 __mem_cgroup_commit_charge(memcg
, page
, 1, type
, false);
3877 static void mem_cgroup_do_uncharge(struct mem_cgroup
*memcg
,
3878 unsigned int nr_pages
,
3879 const enum charge_type ctype
)
3881 struct memcg_batch_info
*batch
= NULL
;
3882 bool uncharge_memsw
= true;
3884 /* If swapout, usage of swap doesn't decrease */
3885 if (!do_swap_account
|| ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
)
3886 uncharge_memsw
= false;
3888 batch
= ¤t
->memcg_batch
;
3890 * In usual, we do css_get() when we remember memcg pointer.
3891 * But in this case, we keep res->usage until end of a series of
3892 * uncharges. Then, it's ok to ignore memcg's refcnt.
3895 batch
->memcg
= memcg
;
3897 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3898 * In those cases, all pages freed continuously can be expected to be in
3899 * the same cgroup and we have chance to coalesce uncharges.
3900 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3901 * because we want to do uncharge as soon as possible.
3904 if (!batch
->do_batch
|| test_thread_flag(TIF_MEMDIE
))
3905 goto direct_uncharge
;
3908 goto direct_uncharge
;
3911 * In typical case, batch->memcg == mem. This means we can
3912 * merge a series of uncharges to an uncharge of res_counter.
3913 * If not, we uncharge res_counter ony by one.
3915 if (batch
->memcg
!= memcg
)
3916 goto direct_uncharge
;
3917 /* remember freed charge and uncharge it later */
3920 batch
->memsw_nr_pages
++;
3923 res_counter_uncharge(&memcg
->res
, nr_pages
* PAGE_SIZE
);
3925 res_counter_uncharge(&memcg
->memsw
, nr_pages
* PAGE_SIZE
);
3926 if (unlikely(batch
->memcg
!= memcg
))
3927 memcg_oom_recover(memcg
);
3931 * uncharge if !page_mapped(page)
3933 static struct mem_cgroup
*
3934 __mem_cgroup_uncharge_common(struct page
*page
, enum charge_type ctype
,
3937 struct mem_cgroup
*memcg
= NULL
;
3938 unsigned int nr_pages
= 1;
3939 struct page_cgroup
*pc
;
3942 if (mem_cgroup_disabled())
3945 if (PageTransHuge(page
)) {
3946 nr_pages
<<= compound_order(page
);
3947 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
3950 * Check if our page_cgroup is valid
3952 pc
= lookup_page_cgroup(page
);
3953 if (unlikely(!PageCgroupUsed(pc
)))
3956 lock_page_cgroup(pc
);
3958 memcg
= pc
->mem_cgroup
;
3960 if (!PageCgroupUsed(pc
))
3963 anon
= PageAnon(page
);
3966 case MEM_CGROUP_CHARGE_TYPE_ANON
:
3968 * Generally PageAnon tells if it's the anon statistics to be
3969 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
3970 * used before page reached the stage of being marked PageAnon.
3974 case MEM_CGROUP_CHARGE_TYPE_DROP
:
3975 /* See mem_cgroup_prepare_migration() */
3976 if (page_mapped(page
))
3979 * Pages under migration may not be uncharged. But
3980 * end_migration() /must/ be the one uncharging the
3981 * unused post-migration page and so it has to call
3982 * here with the migration bit still set. See the
3983 * res_counter handling below.
3985 if (!end_migration
&& PageCgroupMigration(pc
))
3988 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT
:
3989 if (!PageAnon(page
)) { /* Shared memory */
3990 if (page
->mapping
&& !page_is_file_cache(page
))
3992 } else if (page_mapped(page
)) /* Anon */
3999 mem_cgroup_charge_statistics(memcg
, page
, anon
, -nr_pages
);
4001 ClearPageCgroupUsed(pc
);
4003 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4004 * freed from LRU. This is safe because uncharged page is expected not
4005 * to be reused (freed soon). Exception is SwapCache, it's handled by
4006 * special functions.
4009 unlock_page_cgroup(pc
);
4011 * even after unlock, we have memcg->res.usage here and this memcg
4012 * will never be freed, so it's safe to call css_get().
4014 memcg_check_events(memcg
, page
);
4015 if (do_swap_account
&& ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
) {
4016 mem_cgroup_swap_statistics(memcg
, true);
4017 css_get(&memcg
->css
);
4020 * Migration does not charge the res_counter for the
4021 * replacement page, so leave it alone when phasing out the
4022 * page that is unused after the migration.
4024 if (!end_migration
&& !mem_cgroup_is_root(memcg
))
4025 mem_cgroup_do_uncharge(memcg
, nr_pages
, ctype
);
4030 unlock_page_cgroup(pc
);
4034 void mem_cgroup_uncharge_page(struct page
*page
)
4037 if (page_mapped(page
))
4039 VM_BUG_ON_PAGE(page
->mapping
&& !PageAnon(page
), page
);
4041 * If the page is in swap cache, uncharge should be deferred
4042 * to the swap path, which also properly accounts swap usage
4043 * and handles memcg lifetime.
4045 * Note that this check is not stable and reclaim may add the
4046 * page to swap cache at any time after this. However, if the
4047 * page is not in swap cache by the time page->mapcount hits
4048 * 0, there won't be any page table references to the swap
4049 * slot, and reclaim will free it and not actually write the
4052 if (PageSwapCache(page
))
4054 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_ANON
, false);
4057 void mem_cgroup_uncharge_cache_page(struct page
*page
)
4059 VM_BUG_ON_PAGE(page_mapped(page
), page
);
4060 VM_BUG_ON_PAGE(page
->mapping
, page
);
4061 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_CACHE
, false);
4065 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4066 * In that cases, pages are freed continuously and we can expect pages
4067 * are in the same memcg. All these calls itself limits the number of
4068 * pages freed at once, then uncharge_start/end() is called properly.
4069 * This may be called prural(2) times in a context,
4072 void mem_cgroup_uncharge_start(void)
4074 current
->memcg_batch
.do_batch
++;
4075 /* We can do nest. */
4076 if (current
->memcg_batch
.do_batch
== 1) {
4077 current
->memcg_batch
.memcg
= NULL
;
4078 current
->memcg_batch
.nr_pages
= 0;
4079 current
->memcg_batch
.memsw_nr_pages
= 0;
4083 void mem_cgroup_uncharge_end(void)
4085 struct memcg_batch_info
*batch
= ¤t
->memcg_batch
;
4087 if (!batch
->do_batch
)
4091 if (batch
->do_batch
) /* If stacked, do nothing. */
4097 * This "batch->memcg" is valid without any css_get/put etc...
4098 * bacause we hide charges behind us.
4100 if (batch
->nr_pages
)
4101 res_counter_uncharge(&batch
->memcg
->res
,
4102 batch
->nr_pages
* PAGE_SIZE
);
4103 if (batch
->memsw_nr_pages
)
4104 res_counter_uncharge(&batch
->memcg
->memsw
,
4105 batch
->memsw_nr_pages
* PAGE_SIZE
);
4106 memcg_oom_recover(batch
->memcg
);
4107 /* forget this pointer (for sanity check) */
4108 batch
->memcg
= NULL
;
4113 * called after __delete_from_swap_cache() and drop "page" account.
4114 * memcg information is recorded to swap_cgroup of "ent"
4117 mem_cgroup_uncharge_swapcache(struct page
*page
, swp_entry_t ent
, bool swapout
)
4119 struct mem_cgroup
*memcg
;
4120 int ctype
= MEM_CGROUP_CHARGE_TYPE_SWAPOUT
;
4122 if (!swapout
) /* this was a swap cache but the swap is unused ! */
4123 ctype
= MEM_CGROUP_CHARGE_TYPE_DROP
;
4125 memcg
= __mem_cgroup_uncharge_common(page
, ctype
, false);
4128 * record memcg information, if swapout && memcg != NULL,
4129 * css_get() was called in uncharge().
4131 if (do_swap_account
&& swapout
&& memcg
)
4132 swap_cgroup_record(ent
, mem_cgroup_id(memcg
));
4136 #ifdef CONFIG_MEMCG_SWAP
4138 * called from swap_entry_free(). remove record in swap_cgroup and
4139 * uncharge "memsw" account.
4141 void mem_cgroup_uncharge_swap(swp_entry_t ent
)
4143 struct mem_cgroup
*memcg
;
4146 if (!do_swap_account
)
4149 id
= swap_cgroup_record(ent
, 0);
4151 memcg
= mem_cgroup_lookup(id
);
4154 * We uncharge this because swap is freed.
4155 * This memcg can be obsolete one. We avoid calling css_tryget
4157 if (!mem_cgroup_is_root(memcg
))
4158 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
4159 mem_cgroup_swap_statistics(memcg
, false);
4160 css_put(&memcg
->css
);
4166 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4167 * @entry: swap entry to be moved
4168 * @from: mem_cgroup which the entry is moved from
4169 * @to: mem_cgroup which the entry is moved to
4171 * It succeeds only when the swap_cgroup's record for this entry is the same
4172 * as the mem_cgroup's id of @from.
4174 * Returns 0 on success, -EINVAL on failure.
4176 * The caller must have charged to @to, IOW, called res_counter_charge() about
4177 * both res and memsw, and called css_get().
4179 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
4180 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4182 unsigned short old_id
, new_id
;
4184 old_id
= mem_cgroup_id(from
);
4185 new_id
= mem_cgroup_id(to
);
4187 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
4188 mem_cgroup_swap_statistics(from
, false);
4189 mem_cgroup_swap_statistics(to
, true);
4191 * This function is only called from task migration context now.
4192 * It postpones res_counter and refcount handling till the end
4193 * of task migration(mem_cgroup_clear_mc()) for performance
4194 * improvement. But we cannot postpone css_get(to) because if
4195 * the process that has been moved to @to does swap-in, the
4196 * refcount of @to might be decreased to 0.
4198 * We are in attach() phase, so the cgroup is guaranteed to be
4199 * alive, so we can just call css_get().
4207 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
4208 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4215 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4218 void mem_cgroup_prepare_migration(struct page
*page
, struct page
*newpage
,
4219 struct mem_cgroup
**memcgp
)
4221 struct mem_cgroup
*memcg
= NULL
;
4222 unsigned int nr_pages
= 1;
4223 struct page_cgroup
*pc
;
4224 enum charge_type ctype
;
4228 if (mem_cgroup_disabled())
4231 if (PageTransHuge(page
))
4232 nr_pages
<<= compound_order(page
);
4234 pc
= lookup_page_cgroup(page
);
4235 lock_page_cgroup(pc
);
4236 if (PageCgroupUsed(pc
)) {
4237 memcg
= pc
->mem_cgroup
;
4238 css_get(&memcg
->css
);
4240 * At migrating an anonymous page, its mapcount goes down
4241 * to 0 and uncharge() will be called. But, even if it's fully
4242 * unmapped, migration may fail and this page has to be
4243 * charged again. We set MIGRATION flag here and delay uncharge
4244 * until end_migration() is called
4246 * Corner Case Thinking
4248 * When the old page was mapped as Anon and it's unmap-and-freed
4249 * while migration was ongoing.
4250 * If unmap finds the old page, uncharge() of it will be delayed
4251 * until end_migration(). If unmap finds a new page, it's
4252 * uncharged when it make mapcount to be 1->0. If unmap code
4253 * finds swap_migration_entry, the new page will not be mapped
4254 * and end_migration() will find it(mapcount==0).
4257 * When the old page was mapped but migraion fails, the kernel
4258 * remaps it. A charge for it is kept by MIGRATION flag even
4259 * if mapcount goes down to 0. We can do remap successfully
4260 * without charging it again.
4263 * The "old" page is under lock_page() until the end of
4264 * migration, so, the old page itself will not be swapped-out.
4265 * If the new page is swapped out before end_migraton, our
4266 * hook to usual swap-out path will catch the event.
4269 SetPageCgroupMigration(pc
);
4271 unlock_page_cgroup(pc
);
4273 * If the page is not charged at this point,
4281 * We charge new page before it's used/mapped. So, even if unlock_page()
4282 * is called before end_migration, we can catch all events on this new
4283 * page. In the case new page is migrated but not remapped, new page's
4284 * mapcount will be finally 0 and we call uncharge in end_migration().
4287 ctype
= MEM_CGROUP_CHARGE_TYPE_ANON
;
4289 ctype
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4291 * The page is committed to the memcg, but it's not actually
4292 * charged to the res_counter since we plan on replacing the
4293 * old one and only one page is going to be left afterwards.
4295 __mem_cgroup_commit_charge(memcg
, newpage
, nr_pages
, ctype
, false);
4298 /* remove redundant charge if migration failed*/
4299 void mem_cgroup_end_migration(struct mem_cgroup
*memcg
,
4300 struct page
*oldpage
, struct page
*newpage
, bool migration_ok
)
4302 struct page
*used
, *unused
;
4303 struct page_cgroup
*pc
;
4309 if (!migration_ok
) {
4316 anon
= PageAnon(used
);
4317 __mem_cgroup_uncharge_common(unused
,
4318 anon
? MEM_CGROUP_CHARGE_TYPE_ANON
4319 : MEM_CGROUP_CHARGE_TYPE_CACHE
,
4321 css_put(&memcg
->css
);
4323 * We disallowed uncharge of pages under migration because mapcount
4324 * of the page goes down to zero, temporarly.
4325 * Clear the flag and check the page should be charged.
4327 pc
= lookup_page_cgroup(oldpage
);
4328 lock_page_cgroup(pc
);
4329 ClearPageCgroupMigration(pc
);
4330 unlock_page_cgroup(pc
);
4333 * If a page is a file cache, radix-tree replacement is very atomic
4334 * and we can skip this check. When it was an Anon page, its mapcount
4335 * goes down to 0. But because we added MIGRATION flage, it's not
4336 * uncharged yet. There are several case but page->mapcount check
4337 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4338 * check. (see prepare_charge() also)
4341 mem_cgroup_uncharge_page(used
);
4345 * At replace page cache, newpage is not under any memcg but it's on
4346 * LRU. So, this function doesn't touch res_counter but handles LRU
4347 * in correct way. Both pages are locked so we cannot race with uncharge.
4349 void mem_cgroup_replace_page_cache(struct page
*oldpage
,
4350 struct page
*newpage
)
4352 struct mem_cgroup
*memcg
= NULL
;
4353 struct page_cgroup
*pc
;
4354 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4356 if (mem_cgroup_disabled())
4359 pc
= lookup_page_cgroup(oldpage
);
4360 /* fix accounting on old pages */
4361 lock_page_cgroup(pc
);
4362 if (PageCgroupUsed(pc
)) {
4363 memcg
= pc
->mem_cgroup
;
4364 mem_cgroup_charge_statistics(memcg
, oldpage
, false, -1);
4365 ClearPageCgroupUsed(pc
);
4367 unlock_page_cgroup(pc
);
4370 * When called from shmem_replace_page(), in some cases the
4371 * oldpage has already been charged, and in some cases not.
4376 * Even if newpage->mapping was NULL before starting replacement,
4377 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4378 * LRU while we overwrite pc->mem_cgroup.
4380 __mem_cgroup_commit_charge(memcg
, newpage
, 1, type
, true);
4383 #ifdef CONFIG_DEBUG_VM
4384 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
4386 struct page_cgroup
*pc
;
4388 pc
= lookup_page_cgroup(page
);
4390 * Can be NULL while feeding pages into the page allocator for
4391 * the first time, i.e. during boot or memory hotplug;
4392 * or when mem_cgroup_disabled().
4394 if (likely(pc
) && PageCgroupUsed(pc
))
4399 bool mem_cgroup_bad_page_check(struct page
*page
)
4401 if (mem_cgroup_disabled())
4404 return lookup_page_cgroup_used(page
) != NULL
;
4407 void mem_cgroup_print_bad_page(struct page
*page
)
4409 struct page_cgroup
*pc
;
4411 pc
= lookup_page_cgroup_used(page
);
4413 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4414 pc
, pc
->flags
, pc
->mem_cgroup
);
4419 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
4420 unsigned long long val
)
4423 u64 memswlimit
, memlimit
;
4425 int children
= mem_cgroup_count_children(memcg
);
4426 u64 curusage
, oldusage
;
4430 * For keeping hierarchical_reclaim simple, how long we should retry
4431 * is depends on callers. We set our retry-count to be function
4432 * of # of children which we should visit in this loop.
4434 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
4436 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4439 while (retry_count
) {
4440 if (signal_pending(current
)) {
4445 * Rather than hide all in some function, I do this in
4446 * open coded manner. You see what this really does.
4447 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4449 mutex_lock(&set_limit_mutex
);
4450 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4451 if (memswlimit
< val
) {
4453 mutex_unlock(&set_limit_mutex
);
4457 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4461 ret
= res_counter_set_limit(&memcg
->res
, val
);
4463 if (memswlimit
== val
)
4464 memcg
->memsw_is_minimum
= true;
4466 memcg
->memsw_is_minimum
= false;
4468 mutex_unlock(&set_limit_mutex
);
4473 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4474 MEM_CGROUP_RECLAIM_SHRINK
);
4475 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4476 /* Usage is reduced ? */
4477 if (curusage
>= oldusage
)
4480 oldusage
= curusage
;
4482 if (!ret
&& enlarge
)
4483 memcg_oom_recover(memcg
);
4488 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
4489 unsigned long long val
)
4492 u64 memlimit
, memswlimit
, oldusage
, curusage
;
4493 int children
= mem_cgroup_count_children(memcg
);
4497 /* see mem_cgroup_resize_res_limit */
4498 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
4499 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4500 while (retry_count
) {
4501 if (signal_pending(current
)) {
4506 * Rather than hide all in some function, I do this in
4507 * open coded manner. You see what this really does.
4508 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4510 mutex_lock(&set_limit_mutex
);
4511 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4512 if (memlimit
> val
) {
4514 mutex_unlock(&set_limit_mutex
);
4517 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4518 if (memswlimit
< val
)
4520 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
4522 if (memlimit
== val
)
4523 memcg
->memsw_is_minimum
= true;
4525 memcg
->memsw_is_minimum
= false;
4527 mutex_unlock(&set_limit_mutex
);
4532 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4533 MEM_CGROUP_RECLAIM_NOSWAP
|
4534 MEM_CGROUP_RECLAIM_SHRINK
);
4535 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4536 /* Usage is reduced ? */
4537 if (curusage
>= oldusage
)
4540 oldusage
= curusage
;
4542 if (!ret
&& enlarge
)
4543 memcg_oom_recover(memcg
);
4547 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
4549 unsigned long *total_scanned
)
4551 unsigned long nr_reclaimed
= 0;
4552 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
4553 unsigned long reclaimed
;
4555 struct mem_cgroup_tree_per_zone
*mctz
;
4556 unsigned long long excess
;
4557 unsigned long nr_scanned
;
4562 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
4564 * This loop can run a while, specially if mem_cgroup's continuously
4565 * keep exceeding their soft limit and putting the system under
4572 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
4577 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
4578 gfp_mask
, &nr_scanned
);
4579 nr_reclaimed
+= reclaimed
;
4580 *total_scanned
+= nr_scanned
;
4581 spin_lock(&mctz
->lock
);
4584 * If we failed to reclaim anything from this memory cgroup
4585 * it is time to move on to the next cgroup
4591 * Loop until we find yet another one.
4593 * By the time we get the soft_limit lock
4594 * again, someone might have aded the
4595 * group back on the RB tree. Iterate to
4596 * make sure we get a different mem.
4597 * mem_cgroup_largest_soft_limit_node returns
4598 * NULL if no other cgroup is present on
4602 __mem_cgroup_largest_soft_limit_node(mctz
);
4604 css_put(&next_mz
->memcg
->css
);
4605 else /* next_mz == NULL or other memcg */
4609 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
4610 excess
= res_counter_soft_limit_excess(&mz
->memcg
->res
);
4612 * One school of thought says that we should not add
4613 * back the node to the tree if reclaim returns 0.
4614 * But our reclaim could return 0, simply because due
4615 * to priority we are exposing a smaller subset of
4616 * memory to reclaim from. Consider this as a longer
4619 /* If excess == 0, no tree ops */
4620 __mem_cgroup_insert_exceeded(mz
->memcg
, mz
, mctz
, excess
);
4621 spin_unlock(&mctz
->lock
);
4622 css_put(&mz
->memcg
->css
);
4625 * Could not reclaim anything and there are no more
4626 * mem cgroups to try or we seem to be looping without
4627 * reclaiming anything.
4629 if (!nr_reclaimed
&&
4631 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
4633 } while (!nr_reclaimed
);
4635 css_put(&next_mz
->memcg
->css
);
4636 return nr_reclaimed
;
4640 * mem_cgroup_force_empty_list - clears LRU of a group
4641 * @memcg: group to clear
4644 * @lru: lru to to clear
4646 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4647 * reclaim the pages page themselves - pages are moved to the parent (or root)
4650 static void mem_cgroup_force_empty_list(struct mem_cgroup
*memcg
,
4651 int node
, int zid
, enum lru_list lru
)
4653 struct lruvec
*lruvec
;
4654 unsigned long flags
;
4655 struct list_head
*list
;
4659 zone
= &NODE_DATA(node
)->node_zones
[zid
];
4660 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
4661 list
= &lruvec
->lists
[lru
];
4665 struct page_cgroup
*pc
;
4668 spin_lock_irqsave(&zone
->lru_lock
, flags
);
4669 if (list_empty(list
)) {
4670 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4673 page
= list_entry(list
->prev
, struct page
, lru
);
4675 list_move(&page
->lru
, list
);
4677 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4680 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4682 pc
= lookup_page_cgroup(page
);
4684 if (mem_cgroup_move_parent(page
, pc
, memcg
)) {
4685 /* found lock contention or "pc" is obsolete. */
4690 } while (!list_empty(list
));
4694 * make mem_cgroup's charge to be 0 if there is no task by moving
4695 * all the charges and pages to the parent.
4696 * This enables deleting this mem_cgroup.
4698 * Caller is responsible for holding css reference on the memcg.
4700 static void mem_cgroup_reparent_charges(struct mem_cgroup
*memcg
)
4706 /* This is for making all *used* pages to be on LRU. */
4707 lru_add_drain_all();
4708 drain_all_stock_sync(memcg
);
4709 mem_cgroup_start_move(memcg
);
4710 for_each_node_state(node
, N_MEMORY
) {
4711 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
4714 mem_cgroup_force_empty_list(memcg
,
4719 mem_cgroup_end_move(memcg
);
4720 memcg_oom_recover(memcg
);
4724 * Kernel memory may not necessarily be trackable to a specific
4725 * process. So they are not migrated, and therefore we can't
4726 * expect their value to drop to 0 here.
4727 * Having res filled up with kmem only is enough.
4729 * This is a safety check because mem_cgroup_force_empty_list
4730 * could have raced with mem_cgroup_replace_page_cache callers
4731 * so the lru seemed empty but the page could have been added
4732 * right after the check. RES_USAGE should be safe as we always
4733 * charge before adding to the LRU.
4735 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
) -
4736 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
);
4737 } while (usage
> 0);
4740 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
4742 lockdep_assert_held(&memcg_create_mutex
);
4744 * The lock does not prevent addition or deletion to the list
4745 * of children, but it prevents a new child from being
4746 * initialized based on this parent in css_online(), so it's
4747 * enough to decide whether hierarchically inherited
4748 * attributes can still be changed or not.
4750 return memcg
->use_hierarchy
&&
4751 !list_empty(&memcg
->css
.cgroup
->children
);
4755 * Reclaims as many pages from the given memcg as possible and moves
4756 * the rest to the parent.
4758 * Caller is responsible for holding css reference for memcg.
4760 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
4762 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
4763 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
4765 /* returns EBUSY if there is a task or if we come here twice. */
4766 if (cgroup_has_tasks(cgrp
) || !list_empty(&cgrp
->children
))
4769 /* we call try-to-free pages for make this cgroup empty */
4770 lru_add_drain_all();
4771 /* try to free all pages in this cgroup */
4772 while (nr_retries
&& res_counter_read_u64(&memcg
->res
, RES_USAGE
) > 0) {
4775 if (signal_pending(current
))
4778 progress
= try_to_free_mem_cgroup_pages(memcg
, GFP_KERNEL
,
4782 /* maybe some writeback is necessary */
4783 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4788 mem_cgroup_reparent_charges(memcg
);
4793 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state
*css
,
4796 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4798 if (mem_cgroup_is_root(memcg
))
4800 return mem_cgroup_force_empty(memcg
);
4803 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
4806 return mem_cgroup_from_css(css
)->use_hierarchy
;
4809 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
4810 struct cftype
*cft
, u64 val
)
4813 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4814 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
4816 mutex_lock(&memcg_create_mutex
);
4818 if (memcg
->use_hierarchy
== val
)
4822 * If parent's use_hierarchy is set, we can't make any modifications
4823 * in the child subtrees. If it is unset, then the change can
4824 * occur, provided the current cgroup has no children.
4826 * For the root cgroup, parent_mem is NULL, we allow value to be
4827 * set if there are no children.
4829 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
4830 (val
== 1 || val
== 0)) {
4831 if (list_empty(&memcg
->css
.cgroup
->children
))
4832 memcg
->use_hierarchy
= val
;
4839 mutex_unlock(&memcg_create_mutex
);
4845 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup
*memcg
,
4846 enum mem_cgroup_stat_index idx
)
4848 struct mem_cgroup
*iter
;
4851 /* Per-cpu values can be negative, use a signed accumulator */
4852 for_each_mem_cgroup_tree(iter
, memcg
)
4853 val
+= mem_cgroup_read_stat(iter
, idx
);
4855 if (val
< 0) /* race ? */
4860 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
4864 if (!mem_cgroup_is_root(memcg
)) {
4866 return res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4868 return res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4872 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
4873 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
4875 val
= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
4876 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_RSS
);
4879 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
4881 return val
<< PAGE_SHIFT
;
4884 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
4887 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4892 type
= MEMFILE_TYPE(cft
->private);
4893 name
= MEMFILE_ATTR(cft
->private);
4897 if (name
== RES_USAGE
)
4898 val
= mem_cgroup_usage(memcg
, false);
4900 val
= res_counter_read_u64(&memcg
->res
, name
);
4903 if (name
== RES_USAGE
)
4904 val
= mem_cgroup_usage(memcg
, true);
4906 val
= res_counter_read_u64(&memcg
->memsw
, name
);
4909 val
= res_counter_read_u64(&memcg
->kmem
, name
);
4918 #ifdef CONFIG_MEMCG_KMEM
4919 /* should be called with activate_kmem_mutex held */
4920 static int __memcg_activate_kmem(struct mem_cgroup
*memcg
,
4921 unsigned long long limit
)
4926 if (memcg_kmem_is_active(memcg
))
4930 * We are going to allocate memory for data shared by all memory
4931 * cgroups so let's stop accounting here.
4933 memcg_stop_kmem_account();
4936 * For simplicity, we won't allow this to be disabled. It also can't
4937 * be changed if the cgroup has children already, or if tasks had
4940 * If tasks join before we set the limit, a person looking at
4941 * kmem.usage_in_bytes will have no way to determine when it took
4942 * place, which makes the value quite meaningless.
4944 * After it first became limited, changes in the value of the limit are
4945 * of course permitted.
4947 mutex_lock(&memcg_create_mutex
);
4948 if (cgroup_has_tasks(memcg
->css
.cgroup
) || memcg_has_children(memcg
))
4950 mutex_unlock(&memcg_create_mutex
);
4954 memcg_id
= ida_simple_get(&kmem_limited_groups
,
4955 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
4962 * Make sure we have enough space for this cgroup in each root cache's
4965 mutex_lock(&memcg_slab_mutex
);
4966 err
= memcg_update_all_caches(memcg_id
+ 1);
4967 mutex_unlock(&memcg_slab_mutex
);
4971 memcg
->kmemcg_id
= memcg_id
;
4972 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
4975 * We couldn't have accounted to this cgroup, because it hasn't got the
4976 * active bit set yet, so this should succeed.
4978 err
= res_counter_set_limit(&memcg
->kmem
, limit
);
4981 static_key_slow_inc(&memcg_kmem_enabled_key
);
4983 * Setting the active bit after enabling static branching will
4984 * guarantee no one starts accounting before all call sites are
4987 memcg_kmem_set_active(memcg
);
4989 memcg_resume_kmem_account();
4993 ida_simple_remove(&kmem_limited_groups
, memcg_id
);
4997 static int memcg_activate_kmem(struct mem_cgroup
*memcg
,
4998 unsigned long long limit
)
5002 mutex_lock(&activate_kmem_mutex
);
5003 ret
= __memcg_activate_kmem(memcg
, limit
);
5004 mutex_unlock(&activate_kmem_mutex
);
5008 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
5009 unsigned long long val
)
5013 if (!memcg_kmem_is_active(memcg
))
5014 ret
= memcg_activate_kmem(memcg
, val
);
5016 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
5020 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
5023 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
5028 mutex_lock(&activate_kmem_mutex
);
5030 * If the parent cgroup is not kmem-active now, it cannot be activated
5031 * after this point, because it has at least one child already.
5033 if (memcg_kmem_is_active(parent
))
5034 ret
= __memcg_activate_kmem(memcg
, RES_COUNTER_MAX
);
5035 mutex_unlock(&activate_kmem_mutex
);
5039 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
5040 unsigned long long val
)
5044 #endif /* CONFIG_MEMCG_KMEM */
5047 * The user of this function is...
5050 static int mem_cgroup_write(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
5053 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5056 unsigned long long val
;
5059 type
= MEMFILE_TYPE(cft
->private);
5060 name
= MEMFILE_ATTR(cft
->private);
5064 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
5068 /* This function does all necessary parse...reuse it */
5069 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5073 ret
= mem_cgroup_resize_limit(memcg
, val
);
5074 else if (type
== _MEMSWAP
)
5075 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
5076 else if (type
== _KMEM
)
5077 ret
= memcg_update_kmem_limit(memcg
, val
);
5081 case RES_SOFT_LIMIT
:
5082 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5086 * For memsw, soft limits are hard to implement in terms
5087 * of semantics, for now, we support soft limits for
5088 * control without swap
5091 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
5096 ret
= -EINVAL
; /* should be BUG() ? */
5102 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
5103 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
5105 unsigned long long min_limit
, min_memsw_limit
, tmp
;
5107 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5108 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5109 if (!memcg
->use_hierarchy
)
5112 while (css_parent(&memcg
->css
)) {
5113 memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5114 if (!memcg
->use_hierarchy
)
5116 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5117 min_limit
= min(min_limit
, tmp
);
5118 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5119 min_memsw_limit
= min(min_memsw_limit
, tmp
);
5122 *mem_limit
= min_limit
;
5123 *memsw_limit
= min_memsw_limit
;
5126 static int mem_cgroup_reset(struct cgroup_subsys_state
*css
, unsigned int event
)
5128 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5132 type
= MEMFILE_TYPE(event
);
5133 name
= MEMFILE_ATTR(event
);
5138 res_counter_reset_max(&memcg
->res
);
5139 else if (type
== _MEMSWAP
)
5140 res_counter_reset_max(&memcg
->memsw
);
5141 else if (type
== _KMEM
)
5142 res_counter_reset_max(&memcg
->kmem
);
5148 res_counter_reset_failcnt(&memcg
->res
);
5149 else if (type
== _MEMSWAP
)
5150 res_counter_reset_failcnt(&memcg
->memsw
);
5151 else if (type
== _KMEM
)
5152 res_counter_reset_failcnt(&memcg
->kmem
);
5161 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
5164 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
5168 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5169 struct cftype
*cft
, u64 val
)
5171 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5173 if (val
>= (1 << NR_MOVE_TYPE
))
5177 * No kind of locking is needed in here, because ->can_attach() will
5178 * check this value once in the beginning of the process, and then carry
5179 * on with stale data. This means that changes to this value will only
5180 * affect task migrations starting after the change.
5182 memcg
->move_charge_at_immigrate
= val
;
5186 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5187 struct cftype
*cft
, u64 val
)
5194 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
5198 unsigned int lru_mask
;
5201 static const struct numa_stat stats
[] = {
5202 { "total", LRU_ALL
},
5203 { "file", LRU_ALL_FILE
},
5204 { "anon", LRU_ALL_ANON
},
5205 { "unevictable", BIT(LRU_UNEVICTABLE
) },
5207 const struct numa_stat
*stat
;
5210 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5212 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
5213 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
5214 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
5215 for_each_node_state(nid
, N_MEMORY
) {
5216 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5218 seq_printf(m
, " N%d=%lu", nid
, nr
);
5223 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
5224 struct mem_cgroup
*iter
;
5227 for_each_mem_cgroup_tree(iter
, memcg
)
5228 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
5229 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
5230 for_each_node_state(nid
, N_MEMORY
) {
5232 for_each_mem_cgroup_tree(iter
, memcg
)
5233 nr
+= mem_cgroup_node_nr_lru_pages(
5234 iter
, nid
, stat
->lru_mask
);
5235 seq_printf(m
, " N%d=%lu", nid
, nr
);
5242 #endif /* CONFIG_NUMA */
5244 static inline void mem_cgroup_lru_names_not_uptodate(void)
5246 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
5249 static int memcg_stat_show(struct seq_file
*m
, void *v
)
5251 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5252 struct mem_cgroup
*mi
;
5255 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5256 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5258 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
5259 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
5262 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
5263 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
5264 mem_cgroup_read_events(memcg
, i
));
5266 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
5267 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
5268 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
5270 /* Hierarchical information */
5272 unsigned long long limit
, memsw_limit
;
5273 memcg_get_hierarchical_limit(memcg
, &limit
, &memsw_limit
);
5274 seq_printf(m
, "hierarchical_memory_limit %llu\n", limit
);
5275 if (do_swap_account
)
5276 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
5280 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5283 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5285 for_each_mem_cgroup_tree(mi
, memcg
)
5286 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
5287 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
5290 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
5291 unsigned long long val
= 0;
5293 for_each_mem_cgroup_tree(mi
, memcg
)
5294 val
+= mem_cgroup_read_events(mi
, i
);
5295 seq_printf(m
, "total_%s %llu\n",
5296 mem_cgroup_events_names
[i
], val
);
5299 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5300 unsigned long long val
= 0;
5302 for_each_mem_cgroup_tree(mi
, memcg
)
5303 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
5304 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
5307 #ifdef CONFIG_DEBUG_VM
5310 struct mem_cgroup_per_zone
*mz
;
5311 struct zone_reclaim_stat
*rstat
;
5312 unsigned long recent_rotated
[2] = {0, 0};
5313 unsigned long recent_scanned
[2] = {0, 0};
5315 for_each_online_node(nid
)
5316 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
5317 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
5318 rstat
= &mz
->lruvec
.reclaim_stat
;
5320 recent_rotated
[0] += rstat
->recent_rotated
[0];
5321 recent_rotated
[1] += rstat
->recent_rotated
[1];
5322 recent_scanned
[0] += rstat
->recent_scanned
[0];
5323 recent_scanned
[1] += rstat
->recent_scanned
[1];
5325 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
5326 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
5327 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
5328 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
5335 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
5338 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5340 return mem_cgroup_swappiness(memcg
);
5343 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
5344 struct cftype
*cft
, u64 val
)
5346 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5351 if (css_parent(css
))
5352 memcg
->swappiness
= val
;
5354 vm_swappiness
= val
;
5359 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
5361 struct mem_cgroup_threshold_ary
*t
;
5367 t
= rcu_dereference(memcg
->thresholds
.primary
);
5369 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
5374 usage
= mem_cgroup_usage(memcg
, swap
);
5377 * current_threshold points to threshold just below or equal to usage.
5378 * If it's not true, a threshold was crossed after last
5379 * call of __mem_cgroup_threshold().
5381 i
= t
->current_threshold
;
5384 * Iterate backward over array of thresholds starting from
5385 * current_threshold and check if a threshold is crossed.
5386 * If none of thresholds below usage is crossed, we read
5387 * only one element of the array here.
5389 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
5390 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5392 /* i = current_threshold + 1 */
5396 * Iterate forward over array of thresholds starting from
5397 * current_threshold+1 and check if a threshold is crossed.
5398 * If none of thresholds above usage is crossed, we read
5399 * only one element of the array here.
5401 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
5402 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5404 /* Update current_threshold */
5405 t
->current_threshold
= i
- 1;
5410 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
5413 __mem_cgroup_threshold(memcg
, false);
5414 if (do_swap_account
)
5415 __mem_cgroup_threshold(memcg
, true);
5417 memcg
= parent_mem_cgroup(memcg
);
5421 static int compare_thresholds(const void *a
, const void *b
)
5423 const struct mem_cgroup_threshold
*_a
= a
;
5424 const struct mem_cgroup_threshold
*_b
= b
;
5426 if (_a
->threshold
> _b
->threshold
)
5429 if (_a
->threshold
< _b
->threshold
)
5435 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
5437 struct mem_cgroup_eventfd_list
*ev
;
5439 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
5440 eventfd_signal(ev
->eventfd
, 1);
5444 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
5446 struct mem_cgroup
*iter
;
5448 for_each_mem_cgroup_tree(iter
, memcg
)
5449 mem_cgroup_oom_notify_cb(iter
);
5452 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
5453 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
5455 struct mem_cgroup_thresholds
*thresholds
;
5456 struct mem_cgroup_threshold_ary
*new;
5457 u64 threshold
, usage
;
5460 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
5464 mutex_lock(&memcg
->thresholds_lock
);
5467 thresholds
= &memcg
->thresholds
;
5468 else if (type
== _MEMSWAP
)
5469 thresholds
= &memcg
->memsw_thresholds
;
5473 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5475 /* Check if a threshold crossed before adding a new one */
5476 if (thresholds
->primary
)
5477 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5479 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
5481 /* Allocate memory for new array of thresholds */
5482 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
5490 /* Copy thresholds (if any) to new array */
5491 if (thresholds
->primary
) {
5492 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
5493 sizeof(struct mem_cgroup_threshold
));
5496 /* Add new threshold */
5497 new->entries
[size
- 1].eventfd
= eventfd
;
5498 new->entries
[size
- 1].threshold
= threshold
;
5500 /* Sort thresholds. Registering of new threshold isn't time-critical */
5501 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
5502 compare_thresholds
, NULL
);
5504 /* Find current threshold */
5505 new->current_threshold
= -1;
5506 for (i
= 0; i
< size
; i
++) {
5507 if (new->entries
[i
].threshold
<= usage
) {
5509 * new->current_threshold will not be used until
5510 * rcu_assign_pointer(), so it's safe to increment
5513 ++new->current_threshold
;
5518 /* Free old spare buffer and save old primary buffer as spare */
5519 kfree(thresholds
->spare
);
5520 thresholds
->spare
= thresholds
->primary
;
5522 rcu_assign_pointer(thresholds
->primary
, new);
5524 /* To be sure that nobody uses thresholds */
5528 mutex_unlock(&memcg
->thresholds_lock
);
5533 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
5534 struct eventfd_ctx
*eventfd
, const char *args
)
5536 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
5539 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
5540 struct eventfd_ctx
*eventfd
, const char *args
)
5542 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
5545 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
5546 struct eventfd_ctx
*eventfd
, enum res_type type
)
5548 struct mem_cgroup_thresholds
*thresholds
;
5549 struct mem_cgroup_threshold_ary
*new;
5553 mutex_lock(&memcg
->thresholds_lock
);
5555 thresholds
= &memcg
->thresholds
;
5556 else if (type
== _MEMSWAP
)
5557 thresholds
= &memcg
->memsw_thresholds
;
5561 if (!thresholds
->primary
)
5564 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5566 /* Check if a threshold crossed before removing */
5567 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5569 /* Calculate new number of threshold */
5571 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
5572 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
5576 new = thresholds
->spare
;
5578 /* Set thresholds array to NULL if we don't have thresholds */
5587 /* Copy thresholds and find current threshold */
5588 new->current_threshold
= -1;
5589 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
5590 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
5593 new->entries
[j
] = thresholds
->primary
->entries
[i
];
5594 if (new->entries
[j
].threshold
<= usage
) {
5596 * new->current_threshold will not be used
5597 * until rcu_assign_pointer(), so it's safe to increment
5600 ++new->current_threshold
;
5606 /* Swap primary and spare array */
5607 thresholds
->spare
= thresholds
->primary
;
5608 /* If all events are unregistered, free the spare array */
5610 kfree(thresholds
->spare
);
5611 thresholds
->spare
= NULL
;
5614 rcu_assign_pointer(thresholds
->primary
, new);
5616 /* To be sure that nobody uses thresholds */
5619 mutex_unlock(&memcg
->thresholds_lock
);
5622 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
5623 struct eventfd_ctx
*eventfd
)
5625 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
5628 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
5629 struct eventfd_ctx
*eventfd
)
5631 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
5634 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
5635 struct eventfd_ctx
*eventfd
, const char *args
)
5637 struct mem_cgroup_eventfd_list
*event
;
5639 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
5643 spin_lock(&memcg_oom_lock
);
5645 event
->eventfd
= eventfd
;
5646 list_add(&event
->list
, &memcg
->oom_notify
);
5648 /* already in OOM ? */
5649 if (atomic_read(&memcg
->under_oom
))
5650 eventfd_signal(eventfd
, 1);
5651 spin_unlock(&memcg_oom_lock
);
5656 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
5657 struct eventfd_ctx
*eventfd
)
5659 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
5661 spin_lock(&memcg_oom_lock
);
5663 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
5664 if (ev
->eventfd
== eventfd
) {
5665 list_del(&ev
->list
);
5670 spin_unlock(&memcg_oom_lock
);
5673 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
5675 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
5677 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
5678 seq_printf(sf
, "under_oom %d\n", (bool)atomic_read(&memcg
->under_oom
));
5682 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
5683 struct cftype
*cft
, u64 val
)
5685 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5687 /* cannot set to root cgroup and only 0 and 1 are allowed */
5688 if (!css_parent(css
) || !((val
== 0) || (val
== 1)))
5691 memcg
->oom_kill_disable
= val
;
5693 memcg_oom_recover(memcg
);
5698 #ifdef CONFIG_MEMCG_KMEM
5699 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5703 memcg
->kmemcg_id
= -1;
5704 ret
= memcg_propagate_kmem(memcg
);
5708 return mem_cgroup_sockets_init(memcg
, ss
);
5711 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5713 mem_cgroup_sockets_destroy(memcg
);
5716 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5718 if (!memcg_kmem_is_active(memcg
))
5722 * kmem charges can outlive the cgroup. In the case of slab
5723 * pages, for instance, a page contain objects from various
5724 * processes. As we prevent from taking a reference for every
5725 * such allocation we have to be careful when doing uncharge
5726 * (see memcg_uncharge_kmem) and here during offlining.
5728 * The idea is that that only the _last_ uncharge which sees
5729 * the dead memcg will drop the last reference. An additional
5730 * reference is taken here before the group is marked dead
5731 * which is then paired with css_put during uncharge resp. here.
5733 * Although this might sound strange as this path is called from
5734 * css_offline() when the referencemight have dropped down to 0
5735 * and shouldn't be incremented anymore (css_tryget would fail)
5736 * we do not have other options because of the kmem allocations
5739 css_get(&memcg
->css
);
5741 memcg_kmem_mark_dead(memcg
);
5743 if (res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0)
5746 if (memcg_kmem_test_and_clear_dead(memcg
))
5747 css_put(&memcg
->css
);
5750 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5755 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5759 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5765 * DO NOT USE IN NEW FILES.
5767 * "cgroup.event_control" implementation.
5769 * This is way over-engineered. It tries to support fully configurable
5770 * events for each user. Such level of flexibility is completely
5771 * unnecessary especially in the light of the planned unified hierarchy.
5773 * Please deprecate this and replace with something simpler if at all
5778 * Unregister event and free resources.
5780 * Gets called from workqueue.
5782 static void memcg_event_remove(struct work_struct
*work
)
5784 struct mem_cgroup_event
*event
=
5785 container_of(work
, struct mem_cgroup_event
, remove
);
5786 struct mem_cgroup
*memcg
= event
->memcg
;
5788 remove_wait_queue(event
->wqh
, &event
->wait
);
5790 event
->unregister_event(memcg
, event
->eventfd
);
5792 /* Notify userspace the event is going away. */
5793 eventfd_signal(event
->eventfd
, 1);
5795 eventfd_ctx_put(event
->eventfd
);
5797 css_put(&memcg
->css
);
5801 * Gets called on POLLHUP on eventfd when user closes it.
5803 * Called with wqh->lock held and interrupts disabled.
5805 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
5806 int sync
, void *key
)
5808 struct mem_cgroup_event
*event
=
5809 container_of(wait
, struct mem_cgroup_event
, wait
);
5810 struct mem_cgroup
*memcg
= event
->memcg
;
5811 unsigned long flags
= (unsigned long)key
;
5813 if (flags
& POLLHUP
) {
5815 * If the event has been detached at cgroup removal, we
5816 * can simply return knowing the other side will cleanup
5819 * We can't race against event freeing since the other
5820 * side will require wqh->lock via remove_wait_queue(),
5823 spin_lock(&memcg
->event_list_lock
);
5824 if (!list_empty(&event
->list
)) {
5825 list_del_init(&event
->list
);
5827 * We are in atomic context, but cgroup_event_remove()
5828 * may sleep, so we have to call it in workqueue.
5830 schedule_work(&event
->remove
);
5832 spin_unlock(&memcg
->event_list_lock
);
5838 static void memcg_event_ptable_queue_proc(struct file
*file
,
5839 wait_queue_head_t
*wqh
, poll_table
*pt
)
5841 struct mem_cgroup_event
*event
=
5842 container_of(pt
, struct mem_cgroup_event
, pt
);
5845 add_wait_queue(wqh
, &event
->wait
);
5849 * DO NOT USE IN NEW FILES.
5851 * Parse input and register new cgroup event handler.
5853 * Input must be in format '<event_fd> <control_fd> <args>'.
5854 * Interpretation of args is defined by control file implementation.
5856 static int memcg_write_event_control(struct cgroup_subsys_state
*css
,
5857 struct cftype
*cft
, char *buffer
)
5859 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5860 struct mem_cgroup_event
*event
;
5861 struct cgroup_subsys_state
*cfile_css
;
5862 unsigned int efd
, cfd
;
5869 efd
= simple_strtoul(buffer
, &endp
, 10);
5874 cfd
= simple_strtoul(buffer
, &endp
, 10);
5875 if ((*endp
!= ' ') && (*endp
!= '\0'))
5879 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
5883 event
->memcg
= memcg
;
5884 INIT_LIST_HEAD(&event
->list
);
5885 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
5886 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
5887 INIT_WORK(&event
->remove
, memcg_event_remove
);
5895 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
5896 if (IS_ERR(event
->eventfd
)) {
5897 ret
= PTR_ERR(event
->eventfd
);
5904 goto out_put_eventfd
;
5907 /* the process need read permission on control file */
5908 /* AV: shouldn't we check that it's been opened for read instead? */
5909 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
5914 * Determine the event callbacks and set them in @event. This used
5915 * to be done via struct cftype but cgroup core no longer knows
5916 * about these events. The following is crude but the whole thing
5917 * is for compatibility anyway.
5919 * DO NOT ADD NEW FILES.
5921 name
= cfile
.file
->f_dentry
->d_name
.name
;
5923 if (!strcmp(name
, "memory.usage_in_bytes")) {
5924 event
->register_event
= mem_cgroup_usage_register_event
;
5925 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
5926 } else if (!strcmp(name
, "memory.oom_control")) {
5927 event
->register_event
= mem_cgroup_oom_register_event
;
5928 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
5929 } else if (!strcmp(name
, "memory.pressure_level")) {
5930 event
->register_event
= vmpressure_register_event
;
5931 event
->unregister_event
= vmpressure_unregister_event
;
5932 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
5933 event
->register_event
= memsw_cgroup_usage_register_event
;
5934 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
5941 * Verify @cfile should belong to @css. Also, remaining events are
5942 * automatically removed on cgroup destruction but the removal is
5943 * asynchronous, so take an extra ref on @css.
5945 cfile_css
= css_tryget_from_dir(cfile
.file
->f_dentry
->d_parent
,
5946 &memory_cgrp_subsys
);
5948 if (IS_ERR(cfile_css
))
5950 if (cfile_css
!= css
) {
5955 ret
= event
->register_event(memcg
, event
->eventfd
, buffer
);
5959 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
5961 spin_lock(&memcg
->event_list_lock
);
5962 list_add(&event
->list
, &memcg
->event_list
);
5963 spin_unlock(&memcg
->event_list_lock
);
5975 eventfd_ctx_put(event
->eventfd
);
5984 static struct cftype mem_cgroup_files
[] = {
5986 .name
= "usage_in_bytes",
5987 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
5988 .read_u64
= mem_cgroup_read_u64
,
5991 .name
= "max_usage_in_bytes",
5992 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
5993 .trigger
= mem_cgroup_reset
,
5994 .read_u64
= mem_cgroup_read_u64
,
5997 .name
= "limit_in_bytes",
5998 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
5999 .write_string
= mem_cgroup_write
,
6000 .read_u64
= mem_cgroup_read_u64
,
6003 .name
= "soft_limit_in_bytes",
6004 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
6005 .write_string
= mem_cgroup_write
,
6006 .read_u64
= mem_cgroup_read_u64
,
6010 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
6011 .trigger
= mem_cgroup_reset
,
6012 .read_u64
= mem_cgroup_read_u64
,
6016 .seq_show
= memcg_stat_show
,
6019 .name
= "force_empty",
6020 .trigger
= mem_cgroup_force_empty_write
,
6023 .name
= "use_hierarchy",
6024 .flags
= CFTYPE_INSANE
,
6025 .write_u64
= mem_cgroup_hierarchy_write
,
6026 .read_u64
= mem_cgroup_hierarchy_read
,
6029 .name
= "cgroup.event_control", /* XXX: for compat */
6030 .write_string
= memcg_write_event_control
,
6031 .flags
= CFTYPE_NO_PREFIX
,
6035 .name
= "swappiness",
6036 .read_u64
= mem_cgroup_swappiness_read
,
6037 .write_u64
= mem_cgroup_swappiness_write
,
6040 .name
= "move_charge_at_immigrate",
6041 .read_u64
= mem_cgroup_move_charge_read
,
6042 .write_u64
= mem_cgroup_move_charge_write
,
6045 .name
= "oom_control",
6046 .seq_show
= mem_cgroup_oom_control_read
,
6047 .write_u64
= mem_cgroup_oom_control_write
,
6048 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
6051 .name
= "pressure_level",
6055 .name
= "numa_stat",
6056 .seq_show
= memcg_numa_stat_show
,
6059 #ifdef CONFIG_MEMCG_KMEM
6061 .name
= "kmem.limit_in_bytes",
6062 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
6063 .write_string
= mem_cgroup_write
,
6064 .read_u64
= mem_cgroup_read_u64
,
6067 .name
= "kmem.usage_in_bytes",
6068 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
6069 .read_u64
= mem_cgroup_read_u64
,
6072 .name
= "kmem.failcnt",
6073 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
6074 .trigger
= mem_cgroup_reset
,
6075 .read_u64
= mem_cgroup_read_u64
,
6078 .name
= "kmem.max_usage_in_bytes",
6079 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
6080 .trigger
= mem_cgroup_reset
,
6081 .read_u64
= mem_cgroup_read_u64
,
6083 #ifdef CONFIG_SLABINFO
6085 .name
= "kmem.slabinfo",
6086 .seq_show
= mem_cgroup_slabinfo_read
,
6090 { }, /* terminate */
6093 #ifdef CONFIG_MEMCG_SWAP
6094 static struct cftype memsw_cgroup_files
[] = {
6096 .name
= "memsw.usage_in_bytes",
6097 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
6098 .read_u64
= mem_cgroup_read_u64
,
6101 .name
= "memsw.max_usage_in_bytes",
6102 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
6103 .trigger
= mem_cgroup_reset
,
6104 .read_u64
= mem_cgroup_read_u64
,
6107 .name
= "memsw.limit_in_bytes",
6108 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
6109 .write_string
= mem_cgroup_write
,
6110 .read_u64
= mem_cgroup_read_u64
,
6113 .name
= "memsw.failcnt",
6114 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
6115 .trigger
= mem_cgroup_reset
,
6116 .read_u64
= mem_cgroup_read_u64
,
6118 { }, /* terminate */
6121 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6123 struct mem_cgroup_per_node
*pn
;
6124 struct mem_cgroup_per_zone
*mz
;
6125 int zone
, tmp
= node
;
6127 * This routine is called against possible nodes.
6128 * But it's BUG to call kmalloc() against offline node.
6130 * TODO: this routine can waste much memory for nodes which will
6131 * never be onlined. It's better to use memory hotplug callback
6134 if (!node_state(node
, N_NORMAL_MEMORY
))
6136 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
6140 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6141 mz
= &pn
->zoneinfo
[zone
];
6142 lruvec_init(&mz
->lruvec
);
6143 mz
->usage_in_excess
= 0;
6144 mz
->on_tree
= false;
6147 memcg
->nodeinfo
[node
] = pn
;
6151 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6153 kfree(memcg
->nodeinfo
[node
]);
6156 static struct mem_cgroup
*mem_cgroup_alloc(void)
6158 struct mem_cgroup
*memcg
;
6161 size
= sizeof(struct mem_cgroup
);
6162 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
6164 memcg
= kzalloc(size
, GFP_KERNEL
);
6168 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
6171 spin_lock_init(&memcg
->pcp_counter_lock
);
6180 * At destroying mem_cgroup, references from swap_cgroup can remain.
6181 * (scanning all at force_empty is too costly...)
6183 * Instead of clearing all references at force_empty, we remember
6184 * the number of reference from swap_cgroup and free mem_cgroup when
6185 * it goes down to 0.
6187 * Removal of cgroup itself succeeds regardless of refs from swap.
6190 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
6194 mem_cgroup_remove_from_trees(memcg
);
6197 free_mem_cgroup_per_zone_info(memcg
, node
);
6199 free_percpu(memcg
->stat
);
6202 * We need to make sure that (at least for now), the jump label
6203 * destruction code runs outside of the cgroup lock. This is because
6204 * get_online_cpus(), which is called from the static_branch update,
6205 * can't be called inside the cgroup_lock. cpusets are the ones
6206 * enforcing this dependency, so if they ever change, we might as well.
6208 * schedule_work() will guarantee this happens. Be careful if you need
6209 * to move this code around, and make sure it is outside
6212 disarm_static_keys(memcg
);
6217 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6219 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
6221 if (!memcg
->res
.parent
)
6223 return mem_cgroup_from_res_counter(memcg
->res
.parent
, res
);
6225 EXPORT_SYMBOL(parent_mem_cgroup
);
6227 static void __init
mem_cgroup_soft_limit_tree_init(void)
6229 struct mem_cgroup_tree_per_node
*rtpn
;
6230 struct mem_cgroup_tree_per_zone
*rtpz
;
6231 int tmp
, node
, zone
;
6233 for_each_node(node
) {
6235 if (!node_state(node
, N_NORMAL_MEMORY
))
6237 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
6240 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
6242 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6243 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
6244 rtpz
->rb_root
= RB_ROOT
;
6245 spin_lock_init(&rtpz
->lock
);
6250 static struct cgroup_subsys_state
* __ref
6251 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
6253 struct mem_cgroup
*memcg
;
6254 long error
= -ENOMEM
;
6257 memcg
= mem_cgroup_alloc();
6259 return ERR_PTR(error
);
6262 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
6266 if (parent_css
== NULL
) {
6267 root_mem_cgroup
= memcg
;
6268 res_counter_init(&memcg
->res
, NULL
);
6269 res_counter_init(&memcg
->memsw
, NULL
);
6270 res_counter_init(&memcg
->kmem
, NULL
);
6273 memcg
->last_scanned_node
= MAX_NUMNODES
;
6274 INIT_LIST_HEAD(&memcg
->oom_notify
);
6275 memcg
->move_charge_at_immigrate
= 0;
6276 mutex_init(&memcg
->thresholds_lock
);
6277 spin_lock_init(&memcg
->move_lock
);
6278 vmpressure_init(&memcg
->vmpressure
);
6279 INIT_LIST_HEAD(&memcg
->event_list
);
6280 spin_lock_init(&memcg
->event_list_lock
);
6285 __mem_cgroup_free(memcg
);
6286 return ERR_PTR(error
);
6290 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
6292 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6293 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(css
));
6295 if (css
->cgroup
->id
> MEM_CGROUP_ID_MAX
)
6301 mutex_lock(&memcg_create_mutex
);
6303 memcg
->use_hierarchy
= parent
->use_hierarchy
;
6304 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
6305 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
6307 if (parent
->use_hierarchy
) {
6308 res_counter_init(&memcg
->res
, &parent
->res
);
6309 res_counter_init(&memcg
->memsw
, &parent
->memsw
);
6310 res_counter_init(&memcg
->kmem
, &parent
->kmem
);
6313 * No need to take a reference to the parent because cgroup
6314 * core guarantees its existence.
6317 res_counter_init(&memcg
->res
, NULL
);
6318 res_counter_init(&memcg
->memsw
, NULL
);
6319 res_counter_init(&memcg
->kmem
, NULL
);
6321 * Deeper hierachy with use_hierarchy == false doesn't make
6322 * much sense so let cgroup subsystem know about this
6323 * unfortunate state in our controller.
6325 if (parent
!= root_mem_cgroup
)
6326 memory_cgrp_subsys
.broken_hierarchy
= true;
6328 mutex_unlock(&memcg_create_mutex
);
6330 return memcg_init_kmem(memcg
, &memory_cgrp_subsys
);
6334 * Announce all parents that a group from their hierarchy is gone.
6336 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup
*memcg
)
6338 struct mem_cgroup
*parent
= memcg
;
6340 while ((parent
= parent_mem_cgroup(parent
)))
6341 mem_cgroup_iter_invalidate(parent
);
6344 * if the root memcg is not hierarchical we have to check it
6347 if (!root_mem_cgroup
->use_hierarchy
)
6348 mem_cgroup_iter_invalidate(root_mem_cgroup
);
6351 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
6353 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6354 struct mem_cgroup_event
*event
, *tmp
;
6355 struct cgroup_subsys_state
*iter
;
6358 * Unregister events and notify userspace.
6359 * Notify userspace about cgroup removing only after rmdir of cgroup
6360 * directory to avoid race between userspace and kernelspace.
6362 spin_lock(&memcg
->event_list_lock
);
6363 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
6364 list_del_init(&event
->list
);
6365 schedule_work(&event
->remove
);
6367 spin_unlock(&memcg
->event_list_lock
);
6369 kmem_cgroup_css_offline(memcg
);
6371 mem_cgroup_invalidate_reclaim_iterators(memcg
);
6374 * This requires that offlining is serialized. Right now that is
6375 * guaranteed because css_killed_work_fn() holds the cgroup_mutex.
6377 css_for_each_descendant_post(iter
, css
)
6378 mem_cgroup_reparent_charges(mem_cgroup_from_css(iter
));
6380 mem_cgroup_destroy_all_caches(memcg
);
6381 vmpressure_cleanup(&memcg
->vmpressure
);
6384 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
6386 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6388 * XXX: css_offline() would be where we should reparent all
6389 * memory to prepare the cgroup for destruction. However,
6390 * memcg does not do css_tryget() and res_counter charging
6391 * under the same RCU lock region, which means that charging
6392 * could race with offlining. Offlining only happens to
6393 * cgroups with no tasks in them but charges can show up
6394 * without any tasks from the swapin path when the target
6395 * memcg is looked up from the swapout record and not from the
6396 * current task as it usually is. A race like this can leak
6397 * charges and put pages with stale cgroup pointers into
6401 * lookup_swap_cgroup_id()
6403 * mem_cgroup_lookup()
6406 * disable css_tryget()
6409 * reparent_charges()
6410 * res_counter_charge()
6413 * pc->mem_cgroup = dead memcg
6416 * The bulk of the charges are still moved in offline_css() to
6417 * avoid pinning a lot of pages in case a long-term reference
6418 * like a swapout record is deferring the css_free() to long
6419 * after offlining. But this makes sure we catch any charges
6420 * made after offlining:
6422 mem_cgroup_reparent_charges(memcg
);
6424 memcg_destroy_kmem(memcg
);
6425 __mem_cgroup_free(memcg
);
6429 /* Handlers for move charge at task migration. */
6430 #define PRECHARGE_COUNT_AT_ONCE 256
6431 static int mem_cgroup_do_precharge(unsigned long count
)
6434 int batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6435 struct mem_cgroup
*memcg
= mc
.to
;
6437 if (mem_cgroup_is_root(memcg
)) {
6438 mc
.precharge
+= count
;
6439 /* we don't need css_get for root */
6442 /* try to charge at once */
6444 struct res_counter
*dummy
;
6446 * "memcg" cannot be under rmdir() because we've already checked
6447 * by cgroup_lock_live_cgroup() that it is not removed and we
6448 * are still under the same cgroup_mutex. So we can postpone
6451 if (res_counter_charge(&memcg
->res
, PAGE_SIZE
* count
, &dummy
))
6453 if (do_swap_account
&& res_counter_charge(&memcg
->memsw
,
6454 PAGE_SIZE
* count
, &dummy
)) {
6455 res_counter_uncharge(&memcg
->res
, PAGE_SIZE
* count
);
6458 mc
.precharge
+= count
;
6462 /* fall back to one by one charge */
6464 if (signal_pending(current
)) {
6468 if (!batch_count
--) {
6469 batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6472 ret
= mem_cgroup_try_charge(memcg
, GFP_KERNEL
, 1, false);
6474 /* mem_cgroup_clear_mc() will do uncharge later */
6482 * get_mctgt_type - get target type of moving charge
6483 * @vma: the vma the pte to be checked belongs
6484 * @addr: the address corresponding to the pte to be checked
6485 * @ptent: the pte to be checked
6486 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6489 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6490 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6491 * move charge. if @target is not NULL, the page is stored in target->page
6492 * with extra refcnt got(Callers should handle it).
6493 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6494 * target for charge migration. if @target is not NULL, the entry is stored
6497 * Called with pte lock held.
6504 enum mc_target_type
{
6510 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
6511 unsigned long addr
, pte_t ptent
)
6513 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
6515 if (!page
|| !page_mapped(page
))
6517 if (PageAnon(page
)) {
6518 /* we don't move shared anon */
6521 } else if (!move_file())
6522 /* we ignore mapcount for file pages */
6524 if (!get_page_unless_zero(page
))
6531 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6532 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6534 struct page
*page
= NULL
;
6535 swp_entry_t ent
= pte_to_swp_entry(ptent
);
6537 if (!move_anon() || non_swap_entry(ent
))
6540 * Because lookup_swap_cache() updates some statistics counter,
6541 * we call find_get_page() with swapper_space directly.
6543 page
= find_get_page(swap_address_space(ent
), ent
.val
);
6544 if (do_swap_account
)
6545 entry
->val
= ent
.val
;
6550 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6551 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6557 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
6558 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6560 struct page
*page
= NULL
;
6561 struct address_space
*mapping
;
6564 if (!vma
->vm_file
) /* anonymous vma */
6569 mapping
= vma
->vm_file
->f_mapping
;
6570 if (pte_none(ptent
))
6571 pgoff
= linear_page_index(vma
, addr
);
6572 else /* pte_file(ptent) is true */
6573 pgoff
= pte_to_pgoff(ptent
);
6575 /* page is moved even if it's not RSS of this task(page-faulted). */
6577 /* shmem/tmpfs may report page out on swap: account for that too. */
6578 if (shmem_mapping(mapping
)) {
6579 page
= find_get_entry(mapping
, pgoff
);
6580 if (radix_tree_exceptional_entry(page
)) {
6581 swp_entry_t swp
= radix_to_swp_entry(page
);
6582 if (do_swap_account
)
6584 page
= find_get_page(swap_address_space(swp
), swp
.val
);
6587 page
= find_get_page(mapping
, pgoff
);
6589 page
= find_get_page(mapping
, pgoff
);
6594 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
6595 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
6597 struct page
*page
= NULL
;
6598 struct page_cgroup
*pc
;
6599 enum mc_target_type ret
= MC_TARGET_NONE
;
6600 swp_entry_t ent
= { .val
= 0 };
6602 if (pte_present(ptent
))
6603 page
= mc_handle_present_pte(vma
, addr
, ptent
);
6604 else if (is_swap_pte(ptent
))
6605 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
6606 else if (pte_none(ptent
) || pte_file(ptent
))
6607 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
6609 if (!page
&& !ent
.val
)
6612 pc
= lookup_page_cgroup(page
);
6614 * Do only loose check w/o page_cgroup lock.
6615 * mem_cgroup_move_account() checks the pc is valid or not under
6618 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6619 ret
= MC_TARGET_PAGE
;
6621 target
->page
= page
;
6623 if (!ret
|| !target
)
6626 /* There is a swap entry and a page doesn't exist or isn't charged */
6627 if (ent
.val
&& !ret
&&
6628 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
6629 ret
= MC_TARGET_SWAP
;
6636 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6638 * We don't consider swapping or file mapped pages because THP does not
6639 * support them for now.
6640 * Caller should make sure that pmd_trans_huge(pmd) is true.
6642 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6643 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6645 struct page
*page
= NULL
;
6646 struct page_cgroup
*pc
;
6647 enum mc_target_type ret
= MC_TARGET_NONE
;
6649 page
= pmd_page(pmd
);
6650 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
6653 pc
= lookup_page_cgroup(page
);
6654 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6655 ret
= MC_TARGET_PAGE
;
6658 target
->page
= page
;
6664 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6665 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6667 return MC_TARGET_NONE
;
6671 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
6672 unsigned long addr
, unsigned long end
,
6673 struct mm_walk
*walk
)
6675 struct vm_area_struct
*vma
= walk
->private;
6679 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
6680 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
6681 mc
.precharge
+= HPAGE_PMD_NR
;
6686 if (pmd_trans_unstable(pmd
))
6688 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6689 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
6690 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
6691 mc
.precharge
++; /* increment precharge temporarily */
6692 pte_unmap_unlock(pte
- 1, ptl
);
6698 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
6700 unsigned long precharge
;
6701 struct vm_area_struct
*vma
;
6703 down_read(&mm
->mmap_sem
);
6704 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6705 struct mm_walk mem_cgroup_count_precharge_walk
= {
6706 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
6710 if (is_vm_hugetlb_page(vma
))
6712 walk_page_range(vma
->vm_start
, vma
->vm_end
,
6713 &mem_cgroup_count_precharge_walk
);
6715 up_read(&mm
->mmap_sem
);
6717 precharge
= mc
.precharge
;
6723 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
6725 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
6727 VM_BUG_ON(mc
.moving_task
);
6728 mc
.moving_task
= current
;
6729 return mem_cgroup_do_precharge(precharge
);
6732 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6733 static void __mem_cgroup_clear_mc(void)
6735 struct mem_cgroup
*from
= mc
.from
;
6736 struct mem_cgroup
*to
= mc
.to
;
6739 /* we must uncharge all the leftover precharges from mc.to */
6741 __mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
6745 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6746 * we must uncharge here.
6748 if (mc
.moved_charge
) {
6749 __mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
6750 mc
.moved_charge
= 0;
6752 /* we must fixup refcnts and charges */
6753 if (mc
.moved_swap
) {
6754 /* uncharge swap account from the old cgroup */
6755 if (!mem_cgroup_is_root(mc
.from
))
6756 res_counter_uncharge(&mc
.from
->memsw
,
6757 PAGE_SIZE
* mc
.moved_swap
);
6759 for (i
= 0; i
< mc
.moved_swap
; i
++)
6760 css_put(&mc
.from
->css
);
6762 if (!mem_cgroup_is_root(mc
.to
)) {
6764 * we charged both to->res and to->memsw, so we should
6767 res_counter_uncharge(&mc
.to
->res
,
6768 PAGE_SIZE
* mc
.moved_swap
);
6770 /* we've already done css_get(mc.to) */
6773 memcg_oom_recover(from
);
6774 memcg_oom_recover(to
);
6775 wake_up_all(&mc
.waitq
);
6778 static void mem_cgroup_clear_mc(void)
6780 struct mem_cgroup
*from
= mc
.from
;
6783 * we must clear moving_task before waking up waiters at the end of
6786 mc
.moving_task
= NULL
;
6787 __mem_cgroup_clear_mc();
6788 spin_lock(&mc
.lock
);
6791 spin_unlock(&mc
.lock
);
6792 mem_cgroup_end_move(from
);
6795 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6796 struct cgroup_taskset
*tset
)
6798 struct task_struct
*p
= cgroup_taskset_first(tset
);
6800 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6801 unsigned long move_charge_at_immigrate
;
6804 * We are now commited to this value whatever it is. Changes in this
6805 * tunable will only affect upcoming migrations, not the current one.
6806 * So we need to save it, and keep it going.
6808 move_charge_at_immigrate
= memcg
->move_charge_at_immigrate
;
6809 if (move_charge_at_immigrate
) {
6810 struct mm_struct
*mm
;
6811 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
6813 VM_BUG_ON(from
== memcg
);
6815 mm
= get_task_mm(p
);
6818 /* We move charges only when we move a owner of the mm */
6819 if (mm
->owner
== p
) {
6822 VM_BUG_ON(mc
.precharge
);
6823 VM_BUG_ON(mc
.moved_charge
);
6824 VM_BUG_ON(mc
.moved_swap
);
6825 mem_cgroup_start_move(from
);
6826 spin_lock(&mc
.lock
);
6829 mc
.immigrate_flags
= move_charge_at_immigrate
;
6830 spin_unlock(&mc
.lock
);
6831 /* We set mc.moving_task later */
6833 ret
= mem_cgroup_precharge_mc(mm
);
6835 mem_cgroup_clear_mc();
6842 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6843 struct cgroup_taskset
*tset
)
6845 mem_cgroup_clear_mc();
6848 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6849 unsigned long addr
, unsigned long end
,
6850 struct mm_walk
*walk
)
6853 struct vm_area_struct
*vma
= walk
->private;
6856 enum mc_target_type target_type
;
6857 union mc_target target
;
6859 struct page_cgroup
*pc
;
6862 * We don't take compound_lock() here but no race with splitting thp
6864 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6865 * under splitting, which means there's no concurrent thp split,
6866 * - if another thread runs into split_huge_page() just after we
6867 * entered this if-block, the thread must wait for page table lock
6868 * to be unlocked in __split_huge_page_splitting(), where the main
6869 * part of thp split is not executed yet.
6871 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
6872 if (mc
.precharge
< HPAGE_PMD_NR
) {
6876 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6877 if (target_type
== MC_TARGET_PAGE
) {
6879 if (!isolate_lru_page(page
)) {
6880 pc
= lookup_page_cgroup(page
);
6881 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
6882 pc
, mc
.from
, mc
.to
)) {
6883 mc
.precharge
-= HPAGE_PMD_NR
;
6884 mc
.moved_charge
+= HPAGE_PMD_NR
;
6886 putback_lru_page(page
);
6894 if (pmd_trans_unstable(pmd
))
6897 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6898 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6899 pte_t ptent
= *(pte
++);
6905 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6906 case MC_TARGET_PAGE
:
6908 if (isolate_lru_page(page
))
6910 pc
= lookup_page_cgroup(page
);
6911 if (!mem_cgroup_move_account(page
, 1, pc
,
6914 /* we uncharge from mc.from later. */
6917 putback_lru_page(page
);
6918 put
: /* get_mctgt_type() gets the page */
6921 case MC_TARGET_SWAP
:
6923 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6925 /* we fixup refcnts and charges later. */
6933 pte_unmap_unlock(pte
- 1, ptl
);
6938 * We have consumed all precharges we got in can_attach().
6939 * We try charge one by one, but don't do any additional
6940 * charges to mc.to if we have failed in charge once in attach()
6943 ret
= mem_cgroup_do_precharge(1);
6951 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
6953 struct vm_area_struct
*vma
;
6955 lru_add_drain_all();
6957 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
6959 * Someone who are holding the mmap_sem might be waiting in
6960 * waitq. So we cancel all extra charges, wake up all waiters,
6961 * and retry. Because we cancel precharges, we might not be able
6962 * to move enough charges, but moving charge is a best-effort
6963 * feature anyway, so it wouldn't be a big problem.
6965 __mem_cgroup_clear_mc();
6969 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6971 struct mm_walk mem_cgroup_move_charge_walk
= {
6972 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6976 if (is_vm_hugetlb_page(vma
))
6978 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
6979 &mem_cgroup_move_charge_walk
);
6982 * means we have consumed all precharges and failed in
6983 * doing additional charge. Just abandon here.
6987 up_read(&mm
->mmap_sem
);
6990 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6991 struct cgroup_taskset
*tset
)
6993 struct task_struct
*p
= cgroup_taskset_first(tset
);
6994 struct mm_struct
*mm
= get_task_mm(p
);
6998 mem_cgroup_move_charge(mm
);
7002 mem_cgroup_clear_mc();
7004 #else /* !CONFIG_MMU */
7005 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
7006 struct cgroup_taskset
*tset
)
7010 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
7011 struct cgroup_taskset
*tset
)
7014 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
7015 struct cgroup_taskset
*tset
)
7021 * Cgroup retains root cgroups across [un]mount cycles making it necessary
7022 * to verify sane_behavior flag on each mount attempt.
7024 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
7027 * use_hierarchy is forced with sane_behavior. cgroup core
7028 * guarantees that @root doesn't have any children, so turning it
7029 * on for the root memcg is enough.
7031 if (cgroup_sane_behavior(root_css
->cgroup
))
7032 mem_cgroup_from_css(root_css
)->use_hierarchy
= true;
7035 struct cgroup_subsys memory_cgrp_subsys
= {
7036 .css_alloc
= mem_cgroup_css_alloc
,
7037 .css_online
= mem_cgroup_css_online
,
7038 .css_offline
= mem_cgroup_css_offline
,
7039 .css_free
= mem_cgroup_css_free
,
7040 .can_attach
= mem_cgroup_can_attach
,
7041 .cancel_attach
= mem_cgroup_cancel_attach
,
7042 .attach
= mem_cgroup_move_task
,
7043 .bind
= mem_cgroup_bind
,
7044 .base_cftypes
= mem_cgroup_files
,
7048 #ifdef CONFIG_MEMCG_SWAP
7049 static int __init
enable_swap_account(char *s
)
7051 if (!strcmp(s
, "1"))
7052 really_do_swap_account
= 1;
7053 else if (!strcmp(s
, "0"))
7054 really_do_swap_account
= 0;
7057 __setup("swapaccount=", enable_swap_account
);
7059 static void __init
memsw_file_init(void)
7061 WARN_ON(cgroup_add_cftypes(&memory_cgrp_subsys
, memsw_cgroup_files
));
7064 static void __init
enable_swap_cgroup(void)
7066 if (!mem_cgroup_disabled() && really_do_swap_account
) {
7067 do_swap_account
= 1;
7073 static void __init
enable_swap_cgroup(void)
7079 * subsys_initcall() for memory controller.
7081 * Some parts like hotcpu_notifier() have to be initialized from this context
7082 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7083 * everything that doesn't depend on a specific mem_cgroup structure should
7084 * be initialized from here.
7086 static int __init
mem_cgroup_init(void)
7088 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
7089 enable_swap_cgroup();
7090 mem_cgroup_soft_limit_tree_init();
7094 subsys_initcall(mem_cgroup_init
);