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/sort.h>
50 #include <linux/seq_file.h>
51 #include <linux/vmalloc.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>
61 #include <net/tcp_memcontrol.h>
64 #include <asm/uaccess.h>
66 #include <trace/events/vmscan.h>
68 struct cgroup_subsys mem_cgroup_subsys __read_mostly
;
69 EXPORT_SYMBOL(mem_cgroup_subsys
);
71 #define MEM_CGROUP_RECLAIM_RETRIES 5
72 static struct mem_cgroup
*root_mem_cgroup __read_mostly
;
74 #ifdef CONFIG_MEMCG_SWAP
75 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
76 int do_swap_account __read_mostly
;
78 /* for remember boot option*/
79 #ifdef CONFIG_MEMCG_SWAP_ENABLED
80 static int really_do_swap_account __initdata
= 1;
82 static int really_do_swap_account __initdata
= 0;
86 #define do_swap_account 0
90 static const char * const mem_cgroup_stat_names
[] = {
99 enum mem_cgroup_events_index
{
100 MEM_CGROUP_EVENTS_PGPGIN
, /* # of pages paged in */
101 MEM_CGROUP_EVENTS_PGPGOUT
, /* # of pages paged out */
102 MEM_CGROUP_EVENTS_PGFAULT
, /* # of page-faults */
103 MEM_CGROUP_EVENTS_PGMAJFAULT
, /* # of major page-faults */
104 MEM_CGROUP_EVENTS_NSTATS
,
107 static const char * const mem_cgroup_events_names
[] = {
114 static const char * const mem_cgroup_lru_names
[] = {
123 * Per memcg event counter is incremented at every pagein/pageout. With THP,
124 * it will be incremated by the number of pages. This counter is used for
125 * for trigger some periodic events. This is straightforward and better
126 * than using jiffies etc. to handle periodic memcg event.
128 enum mem_cgroup_events_target
{
129 MEM_CGROUP_TARGET_THRESH
,
130 MEM_CGROUP_TARGET_SOFTLIMIT
,
131 MEM_CGROUP_TARGET_NUMAINFO
,
134 #define THRESHOLDS_EVENTS_TARGET 128
135 #define SOFTLIMIT_EVENTS_TARGET 1024
136 #define NUMAINFO_EVENTS_TARGET 1024
138 struct mem_cgroup_stat_cpu
{
139 long count
[MEM_CGROUP_STAT_NSTATS
];
140 unsigned long events
[MEM_CGROUP_EVENTS_NSTATS
];
141 unsigned long nr_page_events
;
142 unsigned long targets
[MEM_CGROUP_NTARGETS
];
145 struct mem_cgroup_reclaim_iter
{
147 * last scanned hierarchy member. Valid only if last_dead_count
148 * matches memcg->dead_count of the hierarchy root group.
150 struct mem_cgroup
*last_visited
;
151 unsigned long last_dead_count
;
153 /* scan generation, increased every round-trip */
154 unsigned int generation
;
158 * per-zone information in memory controller.
160 struct mem_cgroup_per_zone
{
161 struct lruvec lruvec
;
162 unsigned long lru_size
[NR_LRU_LISTS
];
164 struct mem_cgroup_reclaim_iter reclaim_iter
[DEF_PRIORITY
+ 1];
166 struct rb_node tree_node
; /* RB tree node */
167 unsigned long long usage_in_excess
;/* Set to the value by which */
168 /* the soft limit is exceeded*/
170 struct mem_cgroup
*memcg
; /* Back pointer, we cannot */
171 /* use container_of */
174 struct mem_cgroup_per_node
{
175 struct mem_cgroup_per_zone zoneinfo
[MAX_NR_ZONES
];
179 * Cgroups above their limits are maintained in a RB-Tree, independent of
180 * their hierarchy representation
183 struct mem_cgroup_tree_per_zone
{
184 struct rb_root rb_root
;
188 struct mem_cgroup_tree_per_node
{
189 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
192 struct mem_cgroup_tree
{
193 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
196 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
198 struct mem_cgroup_threshold
{
199 struct eventfd_ctx
*eventfd
;
204 struct mem_cgroup_threshold_ary
{
205 /* An array index points to threshold just below or equal to usage. */
206 int current_threshold
;
207 /* Size of entries[] */
209 /* Array of thresholds */
210 struct mem_cgroup_threshold entries
[0];
213 struct mem_cgroup_thresholds
{
214 /* Primary thresholds array */
215 struct mem_cgroup_threshold_ary
*primary
;
217 * Spare threshold array.
218 * This is needed to make mem_cgroup_unregister_event() "never fail".
219 * It must be able to store at least primary->size - 1 entries.
221 struct mem_cgroup_threshold_ary
*spare
;
225 struct mem_cgroup_eventfd_list
{
226 struct list_head list
;
227 struct eventfd_ctx
*eventfd
;
230 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
231 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
234 * The memory controller data structure. The memory controller controls both
235 * page cache and RSS per cgroup. We would eventually like to provide
236 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
237 * to help the administrator determine what knobs to tune.
239 * TODO: Add a water mark for the memory controller. Reclaim will begin when
240 * we hit the water mark. May be even add a low water mark, such that
241 * no reclaim occurs from a cgroup at it's low water mark, this is
242 * a feature that will be implemented much later in the future.
245 struct cgroup_subsys_state css
;
247 * the counter to account for memory usage
249 struct res_counter res
;
251 /* vmpressure notifications */
252 struct vmpressure vmpressure
;
255 * the counter to account for mem+swap usage.
257 struct res_counter memsw
;
260 * the counter to account for kernel memory usage.
262 struct res_counter kmem
;
264 * Should the accounting and control be hierarchical, per subtree?
267 unsigned long kmem_account_flags
; /* See KMEM_ACCOUNTED_*, below */
271 atomic_t oom_wakeups
;
274 /* OOM-Killer disable */
275 int oom_kill_disable
;
277 /* set when res.limit == memsw.limit */
278 bool memsw_is_minimum
;
280 /* protect arrays of thresholds */
281 struct mutex thresholds_lock
;
283 /* thresholds for memory usage. RCU-protected */
284 struct mem_cgroup_thresholds thresholds
;
286 /* thresholds for mem+swap usage. RCU-protected */
287 struct mem_cgroup_thresholds memsw_thresholds
;
289 /* For oom notifier event fd */
290 struct list_head oom_notify
;
293 * Should we move charges of a task when a task is moved into this
294 * mem_cgroup ? And what type of charges should we move ?
296 unsigned long move_charge_at_immigrate
;
298 * set > 0 if pages under this cgroup are moving to other cgroup.
300 atomic_t moving_account
;
301 /* taken only while moving_account > 0 */
302 spinlock_t move_lock
;
306 struct mem_cgroup_stat_cpu __percpu
*stat
;
308 * used when a cpu is offlined or other synchronizations
309 * See mem_cgroup_read_stat().
311 struct mem_cgroup_stat_cpu nocpu_base
;
312 spinlock_t pcp_counter_lock
;
315 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
316 struct cg_proto tcp_mem
;
318 #if defined(CONFIG_MEMCG_KMEM)
319 /* analogous to slab_common's slab_caches list. per-memcg */
320 struct list_head memcg_slab_caches
;
321 /* Not a spinlock, we can take a lot of time walking the list */
322 struct mutex slab_caches_mutex
;
323 /* Index in the kmem_cache->memcg_params->memcg_caches array */
327 int last_scanned_node
;
329 nodemask_t scan_nodes
;
330 atomic_t numainfo_events
;
331 atomic_t numainfo_updating
;
334 struct mem_cgroup_per_node
*nodeinfo
[0];
335 /* WARNING: nodeinfo must be the last member here */
338 static size_t memcg_size(void)
340 return sizeof(struct mem_cgroup
) +
341 nr_node_ids
* sizeof(struct mem_cgroup_per_node
);
344 /* internal only representation about the status of kmem accounting. */
346 KMEM_ACCOUNTED_ACTIVE
= 0, /* accounted by this cgroup itself */
347 KMEM_ACCOUNTED_ACTIVATED
, /* static key enabled. */
348 KMEM_ACCOUNTED_DEAD
, /* dead memcg with pending kmem charges */
351 /* We account when limit is on, but only after call sites are patched */
352 #define KMEM_ACCOUNTED_MASK \
353 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
355 #ifdef CONFIG_MEMCG_KMEM
356 static inline void memcg_kmem_set_active(struct mem_cgroup
*memcg
)
358 set_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
361 static bool memcg_kmem_is_active(struct mem_cgroup
*memcg
)
363 return test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
366 static void memcg_kmem_set_activated(struct mem_cgroup
*memcg
)
368 set_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
371 static void memcg_kmem_clear_activated(struct mem_cgroup
*memcg
)
373 clear_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
376 static void memcg_kmem_mark_dead(struct mem_cgroup
*memcg
)
379 * Our caller must use css_get() first, because memcg_uncharge_kmem()
380 * will call css_put() if it sees the memcg is dead.
383 if (test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
))
384 set_bit(KMEM_ACCOUNTED_DEAD
, &memcg
->kmem_account_flags
);
387 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup
*memcg
)
389 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD
,
390 &memcg
->kmem_account_flags
);
394 /* Stuffs for move charges at task migration. */
396 * Types of charges to be moved. "move_charge_at_immitgrate" and
397 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
400 MOVE_CHARGE_TYPE_ANON
, /* private anonymous page and swap of it */
401 MOVE_CHARGE_TYPE_FILE
, /* file page(including tmpfs) and swap of it */
405 /* "mc" and its members are protected by cgroup_mutex */
406 static struct move_charge_struct
{
407 spinlock_t lock
; /* for from, to */
408 struct mem_cgroup
*from
;
409 struct mem_cgroup
*to
;
410 unsigned long immigrate_flags
;
411 unsigned long precharge
;
412 unsigned long moved_charge
;
413 unsigned long moved_swap
;
414 struct task_struct
*moving_task
; /* a task moving charges */
415 wait_queue_head_t waitq
; /* a waitq for other context */
417 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
418 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
421 static bool move_anon(void)
423 return test_bit(MOVE_CHARGE_TYPE_ANON
, &mc
.immigrate_flags
);
426 static bool move_file(void)
428 return test_bit(MOVE_CHARGE_TYPE_FILE
, &mc
.immigrate_flags
);
432 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
433 * limit reclaim to prevent infinite loops, if they ever occur.
435 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
436 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
439 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
440 MEM_CGROUP_CHARGE_TYPE_ANON
,
441 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
442 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
446 /* for encoding cft->private value on file */
454 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
455 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
456 #define MEMFILE_ATTR(val) ((val) & 0xffff)
457 /* Used for OOM nofiier */
458 #define OOM_CONTROL (0)
461 * Reclaim flags for mem_cgroup_hierarchical_reclaim
463 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
464 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
465 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
466 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
469 * The memcg_create_mutex will be held whenever a new cgroup is created.
470 * As a consequence, any change that needs to protect against new child cgroups
471 * appearing has to hold it as well.
473 static DEFINE_MUTEX(memcg_create_mutex
);
475 struct mem_cgroup
*mem_cgroup_from_css(struct cgroup_subsys_state
*s
)
477 return s
? container_of(s
, struct mem_cgroup
, css
) : NULL
;
480 /* Some nice accessors for the vmpressure. */
481 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
484 memcg
= root_mem_cgroup
;
485 return &memcg
->vmpressure
;
488 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
490 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
493 struct vmpressure
*css_to_vmpressure(struct cgroup_subsys_state
*css
)
495 return &mem_cgroup_from_css(css
)->vmpressure
;
498 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
500 return (memcg
== root_mem_cgroup
);
504 * We restrict the id in the range of [1, 65535], so it can fit into
507 #define MEM_CGROUP_ID_MAX USHRT_MAX
509 static inline unsigned short mem_cgroup_id(struct mem_cgroup
*memcg
)
512 * The ID of the root cgroup is 0, but memcg treat 0 as an
513 * invalid ID, so we return (cgroup_id + 1).
515 return memcg
->css
.cgroup
->id
+ 1;
518 static inline struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
520 struct cgroup_subsys_state
*css
;
522 css
= css_from_id(id
- 1, &mem_cgroup_subsys
);
523 return mem_cgroup_from_css(css
);
526 /* Writing them here to avoid exposing memcg's inner layout */
527 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
529 void sock_update_memcg(struct sock
*sk
)
531 if (mem_cgroup_sockets_enabled
) {
532 struct mem_cgroup
*memcg
;
533 struct cg_proto
*cg_proto
;
535 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
537 /* Socket cloning can throw us here with sk_cgrp already
538 * filled. It won't however, necessarily happen from
539 * process context. So the test for root memcg given
540 * the current task's memcg won't help us in this case.
542 * Respecting the original socket's memcg is a better
543 * decision in this case.
546 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
547 css_get(&sk
->sk_cgrp
->memcg
->css
);
552 memcg
= mem_cgroup_from_task(current
);
553 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
554 if (!mem_cgroup_is_root(memcg
) &&
555 memcg_proto_active(cg_proto
) && css_tryget(&memcg
->css
)) {
556 sk
->sk_cgrp
= cg_proto
;
561 EXPORT_SYMBOL(sock_update_memcg
);
563 void sock_release_memcg(struct sock
*sk
)
565 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
566 struct mem_cgroup
*memcg
;
567 WARN_ON(!sk
->sk_cgrp
->memcg
);
568 memcg
= sk
->sk_cgrp
->memcg
;
569 css_put(&sk
->sk_cgrp
->memcg
->css
);
573 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
575 if (!memcg
|| mem_cgroup_is_root(memcg
))
578 return &memcg
->tcp_mem
;
580 EXPORT_SYMBOL(tcp_proto_cgroup
);
582 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
584 if (!memcg_proto_activated(&memcg
->tcp_mem
))
586 static_key_slow_dec(&memcg_socket_limit_enabled
);
589 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
594 #ifdef CONFIG_MEMCG_KMEM
596 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
597 * The main reason for not using cgroup id for this:
598 * this works better in sparse environments, where we have a lot of memcgs,
599 * but only a few kmem-limited. Or also, if we have, for instance, 200
600 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
601 * 200 entry array for that.
603 * The current size of the caches array is stored in
604 * memcg_limited_groups_array_size. It will double each time we have to
607 static DEFINE_IDA(kmem_limited_groups
);
608 int memcg_limited_groups_array_size
;
611 * MIN_SIZE is different than 1, because we would like to avoid going through
612 * the alloc/free process all the time. In a small machine, 4 kmem-limited
613 * cgroups is a reasonable guess. In the future, it could be a parameter or
614 * tunable, but that is strictly not necessary.
616 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
617 * this constant directly from cgroup, but it is understandable that this is
618 * better kept as an internal representation in cgroup.c. In any case, the
619 * cgrp_id space is not getting any smaller, and we don't have to necessarily
620 * increase ours as well if it increases.
622 #define MEMCG_CACHES_MIN_SIZE 4
623 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
626 * A lot of the calls to the cache allocation functions are expected to be
627 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
628 * conditional to this static branch, we'll have to allow modules that does
629 * kmem_cache_alloc and the such to see this symbol as well
631 struct static_key memcg_kmem_enabled_key
;
632 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
634 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
636 if (memcg_kmem_is_active(memcg
)) {
637 static_key_slow_dec(&memcg_kmem_enabled_key
);
638 ida_simple_remove(&kmem_limited_groups
, memcg
->kmemcg_id
);
641 * This check can't live in kmem destruction function,
642 * since the charges will outlive the cgroup
644 WARN_ON(res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0);
647 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
650 #endif /* CONFIG_MEMCG_KMEM */
652 static void disarm_static_keys(struct mem_cgroup
*memcg
)
654 disarm_sock_keys(memcg
);
655 disarm_kmem_keys(memcg
);
658 static void drain_all_stock_async(struct mem_cgroup
*memcg
);
660 static struct mem_cgroup_per_zone
*
661 mem_cgroup_zoneinfo(struct mem_cgroup
*memcg
, int nid
, int zid
)
663 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
664 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
667 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
672 static struct mem_cgroup_per_zone
*
673 page_cgroup_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
675 int nid
= page_to_nid(page
);
676 int zid
= page_zonenum(page
);
678 return mem_cgroup_zoneinfo(memcg
, nid
, zid
);
681 static struct mem_cgroup_tree_per_zone
*
682 soft_limit_tree_node_zone(int nid
, int zid
)
684 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
687 static struct mem_cgroup_tree_per_zone
*
688 soft_limit_tree_from_page(struct page
*page
)
690 int nid
= page_to_nid(page
);
691 int zid
= page_zonenum(page
);
693 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
697 __mem_cgroup_insert_exceeded(struct mem_cgroup
*memcg
,
698 struct mem_cgroup_per_zone
*mz
,
699 struct mem_cgroup_tree_per_zone
*mctz
,
700 unsigned long long new_usage_in_excess
)
702 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
703 struct rb_node
*parent
= NULL
;
704 struct mem_cgroup_per_zone
*mz_node
;
709 mz
->usage_in_excess
= new_usage_in_excess
;
710 if (!mz
->usage_in_excess
)
714 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
716 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
719 * We can't avoid mem cgroups that are over their soft
720 * limit by the same amount
722 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
725 rb_link_node(&mz
->tree_node
, parent
, p
);
726 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
731 __mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
732 struct mem_cgroup_per_zone
*mz
,
733 struct mem_cgroup_tree_per_zone
*mctz
)
737 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
742 mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
743 struct mem_cgroup_per_zone
*mz
,
744 struct mem_cgroup_tree_per_zone
*mctz
)
746 spin_lock(&mctz
->lock
);
747 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
748 spin_unlock(&mctz
->lock
);
752 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
754 unsigned long long excess
;
755 struct mem_cgroup_per_zone
*mz
;
756 struct mem_cgroup_tree_per_zone
*mctz
;
757 int nid
= page_to_nid(page
);
758 int zid
= page_zonenum(page
);
759 mctz
= soft_limit_tree_from_page(page
);
762 * Necessary to update all ancestors when hierarchy is used.
763 * because their event counter is not touched.
765 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
766 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
767 excess
= res_counter_soft_limit_excess(&memcg
->res
);
769 * We have to update the tree if mz is on RB-tree or
770 * mem is over its softlimit.
772 if (excess
|| mz
->on_tree
) {
773 spin_lock(&mctz
->lock
);
774 /* if on-tree, remove it */
776 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
778 * Insert again. mz->usage_in_excess will be updated.
779 * If excess is 0, no tree ops.
781 __mem_cgroup_insert_exceeded(memcg
, mz
, mctz
, excess
);
782 spin_unlock(&mctz
->lock
);
787 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
790 struct mem_cgroup_per_zone
*mz
;
791 struct mem_cgroup_tree_per_zone
*mctz
;
793 for_each_node(node
) {
794 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
795 mz
= mem_cgroup_zoneinfo(memcg
, node
, zone
);
796 mctz
= soft_limit_tree_node_zone(node
, zone
);
797 mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
802 static struct mem_cgroup_per_zone
*
803 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
805 struct rb_node
*rightmost
= NULL
;
806 struct mem_cgroup_per_zone
*mz
;
810 rightmost
= rb_last(&mctz
->rb_root
);
812 goto done
; /* Nothing to reclaim from */
814 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
816 * Remove the node now but someone else can add it back,
817 * we will to add it back at the end of reclaim to its correct
818 * position in the tree.
820 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
821 if (!res_counter_soft_limit_excess(&mz
->memcg
->res
) ||
822 !css_tryget(&mz
->memcg
->css
))
828 static struct mem_cgroup_per_zone
*
829 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
831 struct mem_cgroup_per_zone
*mz
;
833 spin_lock(&mctz
->lock
);
834 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
835 spin_unlock(&mctz
->lock
);
840 * Implementation Note: reading percpu statistics for memcg.
842 * Both of vmstat[] and percpu_counter has threshold and do periodic
843 * synchronization to implement "quick" read. There are trade-off between
844 * reading cost and precision of value. Then, we may have a chance to implement
845 * a periodic synchronizion of counter in memcg's counter.
847 * But this _read() function is used for user interface now. The user accounts
848 * memory usage by memory cgroup and he _always_ requires exact value because
849 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
850 * have to visit all online cpus and make sum. So, for now, unnecessary
851 * synchronization is not implemented. (just implemented for cpu hotplug)
853 * If there are kernel internal actions which can make use of some not-exact
854 * value, and reading all cpu value can be performance bottleneck in some
855 * common workload, threashold and synchonization as vmstat[] should be
858 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
859 enum mem_cgroup_stat_index idx
)
865 for_each_online_cpu(cpu
)
866 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
867 #ifdef CONFIG_HOTPLUG_CPU
868 spin_lock(&memcg
->pcp_counter_lock
);
869 val
+= memcg
->nocpu_base
.count
[idx
];
870 spin_unlock(&memcg
->pcp_counter_lock
);
876 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
879 int val
= (charge
) ? 1 : -1;
880 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
883 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
884 enum mem_cgroup_events_index idx
)
886 unsigned long val
= 0;
890 for_each_online_cpu(cpu
)
891 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
892 #ifdef CONFIG_HOTPLUG_CPU
893 spin_lock(&memcg
->pcp_counter_lock
);
894 val
+= memcg
->nocpu_base
.events
[idx
];
895 spin_unlock(&memcg
->pcp_counter_lock
);
901 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
903 bool anon
, int nr_pages
)
908 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
909 * counted as CACHE even if it's on ANON LRU.
912 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
915 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
918 if (PageTransHuge(page
))
919 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
922 /* pagein of a big page is an event. So, ignore page size */
924 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
926 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
927 nr_pages
= -nr_pages
; /* for event */
930 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
936 mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
938 struct mem_cgroup_per_zone
*mz
;
940 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
941 return mz
->lru_size
[lru
];
945 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup
*memcg
, int nid
, int zid
,
946 unsigned int lru_mask
)
948 struct mem_cgroup_per_zone
*mz
;
950 unsigned long ret
= 0;
952 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
955 if (BIT(lru
) & lru_mask
)
956 ret
+= mz
->lru_size
[lru
];
962 mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
963 int nid
, unsigned int lru_mask
)
968 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++)
969 total
+= mem_cgroup_zone_nr_lru_pages(memcg
,
975 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
976 unsigned int lru_mask
)
981 for_each_node_state(nid
, N_MEMORY
)
982 total
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
986 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
987 enum mem_cgroup_events_target target
)
989 unsigned long val
, next
;
991 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
992 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
993 /* from time_after() in jiffies.h */
994 if ((long)next
- (long)val
< 0) {
996 case MEM_CGROUP_TARGET_THRESH
:
997 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
999 case MEM_CGROUP_TARGET_SOFTLIMIT
:
1000 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
1002 case MEM_CGROUP_TARGET_NUMAINFO
:
1003 next
= val
+ NUMAINFO_EVENTS_TARGET
;
1008 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
1015 * Check events in order.
1018 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
1021 /* threshold event is triggered in finer grain than soft limit */
1022 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
1023 MEM_CGROUP_TARGET_THRESH
))) {
1025 bool do_numainfo __maybe_unused
;
1027 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
1028 MEM_CGROUP_TARGET_SOFTLIMIT
);
1029 #if MAX_NUMNODES > 1
1030 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
1031 MEM_CGROUP_TARGET_NUMAINFO
);
1035 mem_cgroup_threshold(memcg
);
1036 if (unlikely(do_softlimit
))
1037 mem_cgroup_update_tree(memcg
, page
);
1038 #if MAX_NUMNODES > 1
1039 if (unlikely(do_numainfo
))
1040 atomic_inc(&memcg
->numainfo_events
);
1046 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
1049 * mm_update_next_owner() may clear mm->owner to NULL
1050 * if it races with swapoff, page migration, etc.
1051 * So this can be called with p == NULL.
1056 return mem_cgroup_from_css(task_css(p
, mem_cgroup_subsys_id
));
1059 struct mem_cgroup
*try_get_mem_cgroup_from_mm(struct mm_struct
*mm
)
1061 struct mem_cgroup
*memcg
= NULL
;
1066 * Because we have no locks, mm->owner's may be being moved to other
1067 * cgroup. We use css_tryget() here even if this looks
1068 * pessimistic (rather than adding locks here).
1072 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1073 if (unlikely(!memcg
))
1075 } while (!css_tryget(&memcg
->css
));
1081 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1082 * ref. count) or NULL if the whole root's subtree has been visited.
1084 * helper function to be used by mem_cgroup_iter
1086 static struct mem_cgroup
*__mem_cgroup_iter_next(struct mem_cgroup
*root
,
1087 struct mem_cgroup
*last_visited
)
1089 struct cgroup_subsys_state
*prev_css
, *next_css
;
1091 prev_css
= last_visited
? &last_visited
->css
: NULL
;
1093 next_css
= css_next_descendant_pre(prev_css
, &root
->css
);
1096 * Even if we found a group we have to make sure it is
1097 * alive. css && !memcg means that the groups should be
1098 * skipped and we should continue the tree walk.
1099 * last_visited css is safe to use because it is
1100 * protected by css_get and the tree walk is rcu safe.
1103 struct mem_cgroup
*mem
= mem_cgroup_from_css(next_css
);
1105 if (css_tryget(&mem
->css
))
1108 prev_css
= next_css
;
1116 static void mem_cgroup_iter_invalidate(struct mem_cgroup
*root
)
1119 * When a group in the hierarchy below root is destroyed, the
1120 * hierarchy iterator can no longer be trusted since it might
1121 * have pointed to the destroyed group. Invalidate it.
1123 atomic_inc(&root
->dead_count
);
1126 static struct mem_cgroup
*
1127 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter
*iter
,
1128 struct mem_cgroup
*root
,
1131 struct mem_cgroup
*position
= NULL
;
1133 * A cgroup destruction happens in two stages: offlining and
1134 * release. They are separated by a RCU grace period.
1136 * If the iterator is valid, we may still race with an
1137 * offlining. The RCU lock ensures the object won't be
1138 * released, tryget will fail if we lost the race.
1140 *sequence
= atomic_read(&root
->dead_count
);
1141 if (iter
->last_dead_count
== *sequence
) {
1143 position
= iter
->last_visited
;
1144 if (position
&& !css_tryget(&position
->css
))
1150 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter
*iter
,
1151 struct mem_cgroup
*last_visited
,
1152 struct mem_cgroup
*new_position
,
1156 css_put(&last_visited
->css
);
1158 * We store the sequence count from the time @last_visited was
1159 * loaded successfully instead of rereading it here so that we
1160 * don't lose destruction events in between. We could have
1161 * raced with the destruction of @new_position after all.
1163 iter
->last_visited
= new_position
;
1165 iter
->last_dead_count
= sequence
;
1169 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1170 * @root: hierarchy root
1171 * @prev: previously returned memcg, NULL on first invocation
1172 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1174 * Returns references to children of the hierarchy below @root, or
1175 * @root itself, or %NULL after a full round-trip.
1177 * Caller must pass the return value in @prev on subsequent
1178 * invocations for reference counting, or use mem_cgroup_iter_break()
1179 * to cancel a hierarchy walk before the round-trip is complete.
1181 * Reclaimers can specify a zone and a priority level in @reclaim to
1182 * divide up the memcgs in the hierarchy among all concurrent
1183 * reclaimers operating on the same zone and priority.
1185 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1186 struct mem_cgroup
*prev
,
1187 struct mem_cgroup_reclaim_cookie
*reclaim
)
1189 struct mem_cgroup
*memcg
= NULL
;
1190 struct mem_cgroup
*last_visited
= NULL
;
1192 if (mem_cgroup_disabled())
1196 root
= root_mem_cgroup
;
1198 if (prev
&& !reclaim
)
1199 last_visited
= prev
;
1201 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1209 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1210 int uninitialized_var(seq
);
1213 int nid
= zone_to_nid(reclaim
->zone
);
1214 int zid
= zone_idx(reclaim
->zone
);
1215 struct mem_cgroup_per_zone
*mz
;
1217 mz
= mem_cgroup_zoneinfo(root
, nid
, zid
);
1218 iter
= &mz
->reclaim_iter
[reclaim
->priority
];
1219 if (prev
&& reclaim
->generation
!= iter
->generation
) {
1220 iter
->last_visited
= NULL
;
1224 last_visited
= mem_cgroup_iter_load(iter
, root
, &seq
);
1227 memcg
= __mem_cgroup_iter_next(root
, last_visited
);
1230 mem_cgroup_iter_update(iter
, last_visited
, memcg
, seq
);
1234 else if (!prev
&& memcg
)
1235 reclaim
->generation
= iter
->generation
;
1244 if (prev
&& prev
!= root
)
1245 css_put(&prev
->css
);
1251 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1252 * @root: hierarchy root
1253 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1255 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1256 struct mem_cgroup
*prev
)
1259 root
= root_mem_cgroup
;
1260 if (prev
&& prev
!= root
)
1261 css_put(&prev
->css
);
1265 * Iteration constructs for visiting all cgroups (under a tree). If
1266 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1267 * be used for reference counting.
1269 #define for_each_mem_cgroup_tree(iter, root) \
1270 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1272 iter = mem_cgroup_iter(root, iter, NULL))
1274 #define for_each_mem_cgroup(iter) \
1275 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1277 iter = mem_cgroup_iter(NULL, iter, NULL))
1279 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1281 struct mem_cgroup
*memcg
;
1284 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1285 if (unlikely(!memcg
))
1290 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1293 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1301 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1304 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1305 * @zone: zone of the wanted lruvec
1306 * @memcg: memcg of the wanted lruvec
1308 * Returns the lru list vector holding pages for the given @zone and
1309 * @mem. This can be the global zone lruvec, if the memory controller
1312 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1313 struct mem_cgroup
*memcg
)
1315 struct mem_cgroup_per_zone
*mz
;
1316 struct lruvec
*lruvec
;
1318 if (mem_cgroup_disabled()) {
1319 lruvec
= &zone
->lruvec
;
1323 mz
= mem_cgroup_zoneinfo(memcg
, zone_to_nid(zone
), zone_idx(zone
));
1324 lruvec
= &mz
->lruvec
;
1327 * Since a node can be onlined after the mem_cgroup was created,
1328 * we have to be prepared to initialize lruvec->zone here;
1329 * and if offlined then reonlined, we need to reinitialize it.
1331 if (unlikely(lruvec
->zone
!= zone
))
1332 lruvec
->zone
= zone
;
1337 * Following LRU functions are allowed to be used without PCG_LOCK.
1338 * Operations are called by routine of global LRU independently from memcg.
1339 * What we have to take care of here is validness of pc->mem_cgroup.
1341 * Changes to pc->mem_cgroup happens when
1344 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1345 * It is added to LRU before charge.
1346 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1347 * When moving account, the page is not on LRU. It's isolated.
1351 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1353 * @zone: zone of the page
1355 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1357 struct mem_cgroup_per_zone
*mz
;
1358 struct mem_cgroup
*memcg
;
1359 struct page_cgroup
*pc
;
1360 struct lruvec
*lruvec
;
1362 if (mem_cgroup_disabled()) {
1363 lruvec
= &zone
->lruvec
;
1367 pc
= lookup_page_cgroup(page
);
1368 memcg
= pc
->mem_cgroup
;
1371 * Surreptitiously switch any uncharged offlist page to root:
1372 * an uncharged page off lru does nothing to secure
1373 * its former mem_cgroup from sudden removal.
1375 * Our caller holds lru_lock, and PageCgroupUsed is updated
1376 * under page_cgroup lock: between them, they make all uses
1377 * of pc->mem_cgroup safe.
1379 if (!PageLRU(page
) && !PageCgroupUsed(pc
) && memcg
!= root_mem_cgroup
)
1380 pc
->mem_cgroup
= memcg
= root_mem_cgroup
;
1382 mz
= page_cgroup_zoneinfo(memcg
, page
);
1383 lruvec
= &mz
->lruvec
;
1386 * Since a node can be onlined after the mem_cgroup was created,
1387 * we have to be prepared to initialize lruvec->zone here;
1388 * and if offlined then reonlined, we need to reinitialize it.
1390 if (unlikely(lruvec
->zone
!= zone
))
1391 lruvec
->zone
= zone
;
1396 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1397 * @lruvec: mem_cgroup per zone lru vector
1398 * @lru: index of lru list the page is sitting on
1399 * @nr_pages: positive when adding or negative when removing
1401 * This function must be called when a page is added to or removed from an
1404 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1407 struct mem_cgroup_per_zone
*mz
;
1408 unsigned long *lru_size
;
1410 if (mem_cgroup_disabled())
1413 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1414 lru_size
= mz
->lru_size
+ lru
;
1415 *lru_size
+= nr_pages
;
1416 VM_BUG_ON((long)(*lru_size
) < 0);
1420 * Checks whether given mem is same or in the root_mem_cgroup's
1423 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1424 struct mem_cgroup
*memcg
)
1426 if (root_memcg
== memcg
)
1428 if (!root_memcg
->use_hierarchy
|| !memcg
)
1430 return cgroup_is_descendant(memcg
->css
.cgroup
, root_memcg
->css
.cgroup
);
1433 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1434 struct mem_cgroup
*memcg
)
1439 ret
= __mem_cgroup_same_or_subtree(root_memcg
, memcg
);
1444 bool task_in_mem_cgroup(struct task_struct
*task
,
1445 const struct mem_cgroup
*memcg
)
1447 struct mem_cgroup
*curr
= NULL
;
1448 struct task_struct
*p
;
1451 p
= find_lock_task_mm(task
);
1453 curr
= try_get_mem_cgroup_from_mm(p
->mm
);
1457 * All threads may have already detached their mm's, but the oom
1458 * killer still needs to detect if they have already been oom
1459 * killed to prevent needlessly killing additional tasks.
1462 curr
= mem_cgroup_from_task(task
);
1464 css_get(&curr
->css
);
1470 * We should check use_hierarchy of "memcg" not "curr". Because checking
1471 * use_hierarchy of "curr" here make this function true if hierarchy is
1472 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1473 * hierarchy(even if use_hierarchy is disabled in "memcg").
1475 ret
= mem_cgroup_same_or_subtree(memcg
, curr
);
1476 css_put(&curr
->css
);
1480 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1482 unsigned long inactive_ratio
;
1483 unsigned long inactive
;
1484 unsigned long active
;
1487 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1488 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1490 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1492 inactive_ratio
= int_sqrt(10 * gb
);
1496 return inactive
* inactive_ratio
< active
;
1499 #define mem_cgroup_from_res_counter(counter, member) \
1500 container_of(counter, struct mem_cgroup, member)
1503 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1504 * @memcg: the memory cgroup
1506 * Returns the maximum amount of memory @mem can be charged with, in
1509 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1511 unsigned long long margin
;
1513 margin
= res_counter_margin(&memcg
->res
);
1514 if (do_swap_account
)
1515 margin
= min(margin
, res_counter_margin(&memcg
->memsw
));
1516 return margin
>> PAGE_SHIFT
;
1519 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1522 if (!css_parent(&memcg
->css
))
1523 return vm_swappiness
;
1525 return memcg
->swappiness
;
1529 * memcg->moving_account is used for checking possibility that some thread is
1530 * calling move_account(). When a thread on CPU-A starts moving pages under
1531 * a memcg, other threads should check memcg->moving_account under
1532 * rcu_read_lock(), like this:
1536 * memcg->moving_account+1 if (memcg->mocing_account)
1538 * synchronize_rcu() update something.
1543 /* for quick checking without looking up memcg */
1544 atomic_t memcg_moving __read_mostly
;
1546 static void mem_cgroup_start_move(struct mem_cgroup
*memcg
)
1548 atomic_inc(&memcg_moving
);
1549 atomic_inc(&memcg
->moving_account
);
1553 static void mem_cgroup_end_move(struct mem_cgroup
*memcg
)
1556 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1557 * We check NULL in callee rather than caller.
1560 atomic_dec(&memcg_moving
);
1561 atomic_dec(&memcg
->moving_account
);
1566 * 2 routines for checking "mem" is under move_account() or not.
1568 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1569 * is used for avoiding races in accounting. If true,
1570 * pc->mem_cgroup may be overwritten.
1572 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1573 * under hierarchy of moving cgroups. This is for
1574 * waiting at hith-memory prressure caused by "move".
1577 static bool mem_cgroup_stolen(struct mem_cgroup
*memcg
)
1579 VM_BUG_ON(!rcu_read_lock_held());
1580 return atomic_read(&memcg
->moving_account
) > 0;
1583 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1585 struct mem_cgroup
*from
;
1586 struct mem_cgroup
*to
;
1589 * Unlike task_move routines, we access mc.to, mc.from not under
1590 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1592 spin_lock(&mc
.lock
);
1598 ret
= mem_cgroup_same_or_subtree(memcg
, from
)
1599 || mem_cgroup_same_or_subtree(memcg
, to
);
1601 spin_unlock(&mc
.lock
);
1605 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1607 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1608 if (mem_cgroup_under_move(memcg
)) {
1610 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1611 /* moving charge context might have finished. */
1614 finish_wait(&mc
.waitq
, &wait
);
1622 * Take this lock when
1623 * - a code tries to modify page's memcg while it's USED.
1624 * - a code tries to modify page state accounting in a memcg.
1625 * see mem_cgroup_stolen(), too.
1627 static void move_lock_mem_cgroup(struct mem_cgroup
*memcg
,
1628 unsigned long *flags
)
1630 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
1633 static void move_unlock_mem_cgroup(struct mem_cgroup
*memcg
,
1634 unsigned long *flags
)
1636 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
1639 #define K(x) ((x) << (PAGE_SHIFT-10))
1641 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1642 * @memcg: The memory cgroup that went over limit
1643 * @p: Task that is going to be killed
1645 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1648 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1650 struct cgroup
*task_cgrp
;
1651 struct cgroup
*mem_cgrp
;
1653 * Need a buffer in BSS, can't rely on allocations. The code relies
1654 * on the assumption that OOM is serialized for memory controller.
1655 * If this assumption is broken, revisit this code.
1657 static char memcg_name
[PATH_MAX
];
1659 struct mem_cgroup
*iter
;
1667 mem_cgrp
= memcg
->css
.cgroup
;
1668 task_cgrp
= task_cgroup(p
, mem_cgroup_subsys_id
);
1670 ret
= cgroup_path(task_cgrp
, memcg_name
, PATH_MAX
);
1673 * Unfortunately, we are unable to convert to a useful name
1674 * But we'll still print out the usage information
1681 pr_info("Task in %s killed", memcg_name
);
1684 ret
= cgroup_path(mem_cgrp
, memcg_name
, PATH_MAX
);
1692 * Continues from above, so we don't need an KERN_ level
1694 pr_cont(" as a result of limit of %s\n", memcg_name
);
1697 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1698 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1699 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1700 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1701 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1702 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1703 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1704 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1705 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1706 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) >> 10,
1707 res_counter_read_u64(&memcg
->kmem
, RES_LIMIT
) >> 10,
1708 res_counter_read_u64(&memcg
->kmem
, RES_FAILCNT
));
1710 for_each_mem_cgroup_tree(iter
, memcg
) {
1711 pr_info("Memory cgroup stats");
1714 ret
= cgroup_path(iter
->css
.cgroup
, memcg_name
, PATH_MAX
);
1716 pr_cont(" for %s", memcg_name
);
1720 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1721 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1723 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1724 K(mem_cgroup_read_stat(iter
, i
)));
1727 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1728 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1729 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1736 * This function returns the number of memcg under hierarchy tree. Returns
1737 * 1(self count) if no children.
1739 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1742 struct mem_cgroup
*iter
;
1744 for_each_mem_cgroup_tree(iter
, memcg
)
1750 * Return the memory (and swap, if configured) limit for a memcg.
1752 static u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1756 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1759 * Do not consider swap space if we cannot swap due to swappiness
1761 if (mem_cgroup_swappiness(memcg
)) {
1764 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1765 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1768 * If memsw is finite and limits the amount of swap space
1769 * available to this memcg, return that limit.
1771 limit
= min(limit
, memsw
);
1777 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1780 struct mem_cgroup
*iter
;
1781 unsigned long chosen_points
= 0;
1782 unsigned long totalpages
;
1783 unsigned int points
= 0;
1784 struct task_struct
*chosen
= NULL
;
1787 * If current has a pending SIGKILL or is exiting, then automatically
1788 * select it. The goal is to allow it to allocate so that it may
1789 * quickly exit and free its memory.
1791 if (fatal_signal_pending(current
) || current
->flags
& PF_EXITING
) {
1792 set_thread_flag(TIF_MEMDIE
);
1796 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1797 totalpages
= mem_cgroup_get_limit(memcg
) >> PAGE_SHIFT
? : 1;
1798 for_each_mem_cgroup_tree(iter
, memcg
) {
1799 struct css_task_iter it
;
1800 struct task_struct
*task
;
1802 css_task_iter_start(&iter
->css
, &it
);
1803 while ((task
= css_task_iter_next(&it
))) {
1804 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1806 case OOM_SCAN_SELECT
:
1808 put_task_struct(chosen
);
1810 chosen_points
= ULONG_MAX
;
1811 get_task_struct(chosen
);
1813 case OOM_SCAN_CONTINUE
:
1815 case OOM_SCAN_ABORT
:
1816 css_task_iter_end(&it
);
1817 mem_cgroup_iter_break(memcg
, iter
);
1819 put_task_struct(chosen
);
1824 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1825 if (points
> chosen_points
) {
1827 put_task_struct(chosen
);
1829 chosen_points
= points
;
1830 get_task_struct(chosen
);
1833 css_task_iter_end(&it
);
1838 points
= chosen_points
* 1000 / totalpages
;
1839 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1840 NULL
, "Memory cgroup out of memory");
1843 static unsigned long mem_cgroup_reclaim(struct mem_cgroup
*memcg
,
1845 unsigned long flags
)
1847 unsigned long total
= 0;
1848 bool noswap
= false;
1851 if (flags
& MEM_CGROUP_RECLAIM_NOSWAP
)
1853 if (!(flags
& MEM_CGROUP_RECLAIM_SHRINK
) && memcg
->memsw_is_minimum
)
1856 for (loop
= 0; loop
< MEM_CGROUP_MAX_RECLAIM_LOOPS
; loop
++) {
1858 drain_all_stock_async(memcg
);
1859 total
+= try_to_free_mem_cgroup_pages(memcg
, gfp_mask
, noswap
);
1861 * Allow limit shrinkers, which are triggered directly
1862 * by userspace, to catch signals and stop reclaim
1863 * after minimal progress, regardless of the margin.
1865 if (total
&& (flags
& MEM_CGROUP_RECLAIM_SHRINK
))
1867 if (mem_cgroup_margin(memcg
))
1870 * If nothing was reclaimed after two attempts, there
1871 * may be no reclaimable pages in this hierarchy.
1880 * test_mem_cgroup_node_reclaimable
1881 * @memcg: the target memcg
1882 * @nid: the node ID to be checked.
1883 * @noswap : specify true here if the user wants flle only information.
1885 * This function returns whether the specified memcg contains any
1886 * reclaimable pages on a node. Returns true if there are any reclaimable
1887 * pages in the node.
1889 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1890 int nid
, bool noswap
)
1892 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1894 if (noswap
|| !total_swap_pages
)
1896 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1901 #if MAX_NUMNODES > 1
1904 * Always updating the nodemask is not very good - even if we have an empty
1905 * list or the wrong list here, we can start from some node and traverse all
1906 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1909 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1913 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1914 * pagein/pageout changes since the last update.
1916 if (!atomic_read(&memcg
->numainfo_events
))
1918 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1921 /* make a nodemask where this memcg uses memory from */
1922 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1924 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1926 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1927 node_clear(nid
, memcg
->scan_nodes
);
1930 atomic_set(&memcg
->numainfo_events
, 0);
1931 atomic_set(&memcg
->numainfo_updating
, 0);
1935 * Selecting a node where we start reclaim from. Because what we need is just
1936 * reducing usage counter, start from anywhere is O,K. Considering
1937 * memory reclaim from current node, there are pros. and cons.
1939 * Freeing memory from current node means freeing memory from a node which
1940 * we'll use or we've used. So, it may make LRU bad. And if several threads
1941 * hit limits, it will see a contention on a node. But freeing from remote
1942 * node means more costs for memory reclaim because of memory latency.
1944 * Now, we use round-robin. Better algorithm is welcomed.
1946 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1950 mem_cgroup_may_update_nodemask(memcg
);
1951 node
= memcg
->last_scanned_node
;
1953 node
= next_node(node
, memcg
->scan_nodes
);
1954 if (node
== MAX_NUMNODES
)
1955 node
= first_node(memcg
->scan_nodes
);
1957 * We call this when we hit limit, not when pages are added to LRU.
1958 * No LRU may hold pages because all pages are UNEVICTABLE or
1959 * memcg is too small and all pages are not on LRU. In that case,
1960 * we use curret node.
1962 if (unlikely(node
== MAX_NUMNODES
))
1963 node
= numa_node_id();
1965 memcg
->last_scanned_node
= node
;
1970 * Check all nodes whether it contains reclaimable pages or not.
1971 * For quick scan, we make use of scan_nodes. This will allow us to skip
1972 * unused nodes. But scan_nodes is lazily updated and may not cotain
1973 * enough new information. We need to do double check.
1975 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1980 * quick check...making use of scan_node.
1981 * We can skip unused nodes.
1983 if (!nodes_empty(memcg
->scan_nodes
)) {
1984 for (nid
= first_node(memcg
->scan_nodes
);
1986 nid
= next_node(nid
, memcg
->scan_nodes
)) {
1988 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1993 * Check rest of nodes.
1995 for_each_node_state(nid
, N_MEMORY
) {
1996 if (node_isset(nid
, memcg
->scan_nodes
))
1998 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
2005 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
2010 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
2012 return test_mem_cgroup_node_reclaimable(memcg
, 0, noswap
);
2016 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
2019 unsigned long *total_scanned
)
2021 struct mem_cgroup
*victim
= NULL
;
2024 unsigned long excess
;
2025 unsigned long nr_scanned
;
2026 struct mem_cgroup_reclaim_cookie reclaim
= {
2031 excess
= res_counter_soft_limit_excess(&root_memcg
->res
) >> PAGE_SHIFT
;
2034 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
2039 * If we have not been able to reclaim
2040 * anything, it might because there are
2041 * no reclaimable pages under this hierarchy
2046 * We want to do more targeted reclaim.
2047 * excess >> 2 is not to excessive so as to
2048 * reclaim too much, nor too less that we keep
2049 * coming back to reclaim from this cgroup
2051 if (total
>= (excess
>> 2) ||
2052 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
2057 if (!mem_cgroup_reclaimable(victim
, false))
2059 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
2061 *total_scanned
+= nr_scanned
;
2062 if (!res_counter_soft_limit_excess(&root_memcg
->res
))
2065 mem_cgroup_iter_break(root_memcg
, victim
);
2069 #ifdef CONFIG_LOCKDEP
2070 static struct lockdep_map memcg_oom_lock_dep_map
= {
2071 .name
= "memcg_oom_lock",
2075 static DEFINE_SPINLOCK(memcg_oom_lock
);
2078 * Check OOM-Killer is already running under our hierarchy.
2079 * If someone is running, return false.
2081 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
2083 struct mem_cgroup
*iter
, *failed
= NULL
;
2085 spin_lock(&memcg_oom_lock
);
2087 for_each_mem_cgroup_tree(iter
, memcg
) {
2088 if (iter
->oom_lock
) {
2090 * this subtree of our hierarchy is already locked
2091 * so we cannot give a lock.
2094 mem_cgroup_iter_break(memcg
, iter
);
2097 iter
->oom_lock
= true;
2102 * OK, we failed to lock the whole subtree so we have
2103 * to clean up what we set up to the failing subtree
2105 for_each_mem_cgroup_tree(iter
, memcg
) {
2106 if (iter
== failed
) {
2107 mem_cgroup_iter_break(memcg
, iter
);
2110 iter
->oom_lock
= false;
2113 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
2115 spin_unlock(&memcg_oom_lock
);
2120 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
2122 struct mem_cgroup
*iter
;
2124 spin_lock(&memcg_oom_lock
);
2125 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
2126 for_each_mem_cgroup_tree(iter
, memcg
)
2127 iter
->oom_lock
= false;
2128 spin_unlock(&memcg_oom_lock
);
2131 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
2133 struct mem_cgroup
*iter
;
2135 for_each_mem_cgroup_tree(iter
, memcg
)
2136 atomic_inc(&iter
->under_oom
);
2139 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
2141 struct mem_cgroup
*iter
;
2144 * When a new child is created while the hierarchy is under oom,
2145 * mem_cgroup_oom_lock() may not be called. We have to use
2146 * atomic_add_unless() here.
2148 for_each_mem_cgroup_tree(iter
, memcg
)
2149 atomic_add_unless(&iter
->under_oom
, -1, 0);
2152 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
2154 struct oom_wait_info
{
2155 struct mem_cgroup
*memcg
;
2159 static int memcg_oom_wake_function(wait_queue_t
*wait
,
2160 unsigned mode
, int sync
, void *arg
)
2162 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
2163 struct mem_cgroup
*oom_wait_memcg
;
2164 struct oom_wait_info
*oom_wait_info
;
2166 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
2167 oom_wait_memcg
= oom_wait_info
->memcg
;
2170 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2171 * Then we can use css_is_ancestor without taking care of RCU.
2173 if (!mem_cgroup_same_or_subtree(oom_wait_memcg
, wake_memcg
)
2174 && !mem_cgroup_same_or_subtree(wake_memcg
, oom_wait_memcg
))
2176 return autoremove_wake_function(wait
, mode
, sync
, arg
);
2179 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
2181 atomic_inc(&memcg
->oom_wakeups
);
2182 /* for filtering, pass "memcg" as argument. */
2183 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
2186 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
2188 if (memcg
&& atomic_read(&memcg
->under_oom
))
2189 memcg_wakeup_oom(memcg
);
2192 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
2194 if (!current
->memcg_oom
.may_oom
)
2197 * We are in the middle of the charge context here, so we
2198 * don't want to block when potentially sitting on a callstack
2199 * that holds all kinds of filesystem and mm locks.
2201 * Also, the caller may handle a failed allocation gracefully
2202 * (like optional page cache readahead) and so an OOM killer
2203 * invocation might not even be necessary.
2205 * That's why we don't do anything here except remember the
2206 * OOM context and then deal with it at the end of the page
2207 * fault when the stack is unwound, the locks are released,
2208 * and when we know whether the fault was overall successful.
2210 css_get(&memcg
->css
);
2211 current
->memcg_oom
.memcg
= memcg
;
2212 current
->memcg_oom
.gfp_mask
= mask
;
2213 current
->memcg_oom
.order
= order
;
2217 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2218 * @handle: actually kill/wait or just clean up the OOM state
2220 * This has to be called at the end of a page fault if the memcg OOM
2221 * handler was enabled.
2223 * Memcg supports userspace OOM handling where failed allocations must
2224 * sleep on a waitqueue until the userspace task resolves the
2225 * situation. Sleeping directly in the charge context with all kinds
2226 * of locks held is not a good idea, instead we remember an OOM state
2227 * in the task and mem_cgroup_oom_synchronize() has to be called at
2228 * the end of the page fault to complete the OOM handling.
2230 * Returns %true if an ongoing memcg OOM situation was detected and
2231 * completed, %false otherwise.
2233 bool mem_cgroup_oom_synchronize(bool handle
)
2235 struct mem_cgroup
*memcg
= current
->memcg_oom
.memcg
;
2236 struct oom_wait_info owait
;
2239 /* OOM is global, do not handle */
2246 owait
.memcg
= memcg
;
2247 owait
.wait
.flags
= 0;
2248 owait
.wait
.func
= memcg_oom_wake_function
;
2249 owait
.wait
.private = current
;
2250 INIT_LIST_HEAD(&owait
.wait
.task_list
);
2252 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2253 mem_cgroup_mark_under_oom(memcg
);
2255 locked
= mem_cgroup_oom_trylock(memcg
);
2258 mem_cgroup_oom_notify(memcg
);
2260 if (locked
&& !memcg
->oom_kill_disable
) {
2261 mem_cgroup_unmark_under_oom(memcg
);
2262 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2263 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom
.gfp_mask
,
2264 current
->memcg_oom
.order
);
2267 mem_cgroup_unmark_under_oom(memcg
);
2268 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2272 mem_cgroup_oom_unlock(memcg
);
2274 * There is no guarantee that an OOM-lock contender
2275 * sees the wakeups triggered by the OOM kill
2276 * uncharges. Wake any sleepers explicitely.
2278 memcg_oom_recover(memcg
);
2281 current
->memcg_oom
.memcg
= NULL
;
2282 css_put(&memcg
->css
);
2287 * Currently used to update mapped file statistics, but the routine can be
2288 * generalized to update other statistics as well.
2290 * Notes: Race condition
2292 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2293 * it tends to be costly. But considering some conditions, we doesn't need
2294 * to do so _always_.
2296 * Considering "charge", lock_page_cgroup() is not required because all
2297 * file-stat operations happen after a page is attached to radix-tree. There
2298 * are no race with "charge".
2300 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2301 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2302 * if there are race with "uncharge". Statistics itself is properly handled
2305 * Considering "move", this is an only case we see a race. To make the race
2306 * small, we check mm->moving_account and detect there are possibility of race
2307 * If there is, we take a lock.
2310 void __mem_cgroup_begin_update_page_stat(struct page
*page
,
2311 bool *locked
, unsigned long *flags
)
2313 struct mem_cgroup
*memcg
;
2314 struct page_cgroup
*pc
;
2316 pc
= lookup_page_cgroup(page
);
2318 memcg
= pc
->mem_cgroup
;
2319 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2322 * If this memory cgroup is not under account moving, we don't
2323 * need to take move_lock_mem_cgroup(). Because we already hold
2324 * rcu_read_lock(), any calls to move_account will be delayed until
2325 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2327 if (!mem_cgroup_stolen(memcg
))
2330 move_lock_mem_cgroup(memcg
, flags
);
2331 if (memcg
!= pc
->mem_cgroup
|| !PageCgroupUsed(pc
)) {
2332 move_unlock_mem_cgroup(memcg
, flags
);
2338 void __mem_cgroup_end_update_page_stat(struct page
*page
, unsigned long *flags
)
2340 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2343 * It's guaranteed that pc->mem_cgroup never changes while
2344 * lock is held because a routine modifies pc->mem_cgroup
2345 * should take move_lock_mem_cgroup().
2347 move_unlock_mem_cgroup(pc
->mem_cgroup
, flags
);
2350 void mem_cgroup_update_page_stat(struct page
*page
,
2351 enum mem_cgroup_stat_index idx
, int val
)
2353 struct mem_cgroup
*memcg
;
2354 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2355 unsigned long uninitialized_var(flags
);
2357 if (mem_cgroup_disabled())
2360 VM_BUG_ON(!rcu_read_lock_held());
2361 memcg
= pc
->mem_cgroup
;
2362 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2365 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2369 * size of first charge trial. "32" comes from vmscan.c's magic value.
2370 * TODO: maybe necessary to use big numbers in big irons.
2372 #define CHARGE_BATCH 32U
2373 struct memcg_stock_pcp
{
2374 struct mem_cgroup
*cached
; /* this never be root cgroup */
2375 unsigned int nr_pages
;
2376 struct work_struct work
;
2377 unsigned long flags
;
2378 #define FLUSHING_CACHED_CHARGE 0
2380 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2381 static DEFINE_MUTEX(percpu_charge_mutex
);
2384 * consume_stock: Try to consume stocked charge on this cpu.
2385 * @memcg: memcg to consume from.
2386 * @nr_pages: how many pages to charge.
2388 * The charges will only happen if @memcg matches the current cpu's memcg
2389 * stock, and at least @nr_pages are available in that stock. Failure to
2390 * service an allocation will refill the stock.
2392 * returns true if successful, false otherwise.
2394 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2396 struct memcg_stock_pcp
*stock
;
2399 if (nr_pages
> CHARGE_BATCH
)
2402 stock
= &get_cpu_var(memcg_stock
);
2403 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
)
2404 stock
->nr_pages
-= nr_pages
;
2405 else /* need to call res_counter_charge */
2407 put_cpu_var(memcg_stock
);
2412 * Returns stocks cached in percpu to res_counter and reset cached information.
2414 static void drain_stock(struct memcg_stock_pcp
*stock
)
2416 struct mem_cgroup
*old
= stock
->cached
;
2418 if (stock
->nr_pages
) {
2419 unsigned long bytes
= stock
->nr_pages
* PAGE_SIZE
;
2421 res_counter_uncharge(&old
->res
, bytes
);
2422 if (do_swap_account
)
2423 res_counter_uncharge(&old
->memsw
, bytes
);
2424 stock
->nr_pages
= 0;
2426 stock
->cached
= NULL
;
2430 * This must be called under preempt disabled or must be called by
2431 * a thread which is pinned to local cpu.
2433 static void drain_local_stock(struct work_struct
*dummy
)
2435 struct memcg_stock_pcp
*stock
= &__get_cpu_var(memcg_stock
);
2437 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2440 static void __init
memcg_stock_init(void)
2444 for_each_possible_cpu(cpu
) {
2445 struct memcg_stock_pcp
*stock
=
2446 &per_cpu(memcg_stock
, cpu
);
2447 INIT_WORK(&stock
->work
, drain_local_stock
);
2452 * Cache charges(val) which is from res_counter, to local per_cpu area.
2453 * This will be consumed by consume_stock() function, later.
2455 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2457 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2459 if (stock
->cached
!= memcg
) { /* reset if necessary */
2461 stock
->cached
= memcg
;
2463 stock
->nr_pages
+= nr_pages
;
2464 put_cpu_var(memcg_stock
);
2468 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2469 * of the hierarchy under it. sync flag says whether we should block
2470 * until the work is done.
2472 static void drain_all_stock(struct mem_cgroup
*root_memcg
, bool sync
)
2476 /* Notify other cpus that system-wide "drain" is running */
2479 for_each_online_cpu(cpu
) {
2480 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2481 struct mem_cgroup
*memcg
;
2483 memcg
= stock
->cached
;
2484 if (!memcg
|| !stock
->nr_pages
)
2486 if (!mem_cgroup_same_or_subtree(root_memcg
, memcg
))
2488 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2490 drain_local_stock(&stock
->work
);
2492 schedule_work_on(cpu
, &stock
->work
);
2500 for_each_online_cpu(cpu
) {
2501 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2502 if (test_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
))
2503 flush_work(&stock
->work
);
2510 * Tries to drain stocked charges in other cpus. This function is asynchronous
2511 * and just put a work per cpu for draining localy on each cpu. Caller can
2512 * expects some charges will be back to res_counter later but cannot wait for
2515 static void drain_all_stock_async(struct mem_cgroup
*root_memcg
)
2518 * If someone calls draining, avoid adding more kworker runs.
2520 if (!mutex_trylock(&percpu_charge_mutex
))
2522 drain_all_stock(root_memcg
, false);
2523 mutex_unlock(&percpu_charge_mutex
);
2526 /* This is a synchronous drain interface. */
2527 static void drain_all_stock_sync(struct mem_cgroup
*root_memcg
)
2529 /* called when force_empty is called */
2530 mutex_lock(&percpu_charge_mutex
);
2531 drain_all_stock(root_memcg
, true);
2532 mutex_unlock(&percpu_charge_mutex
);
2536 * This function drains percpu counter value from DEAD cpu and
2537 * move it to local cpu. Note that this function can be preempted.
2539 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2543 spin_lock(&memcg
->pcp_counter_lock
);
2544 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2545 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2547 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2548 memcg
->nocpu_base
.count
[i
] += x
;
2550 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2551 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2553 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2554 memcg
->nocpu_base
.events
[i
] += x
;
2556 spin_unlock(&memcg
->pcp_counter_lock
);
2559 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2560 unsigned long action
,
2563 int cpu
= (unsigned long)hcpu
;
2564 struct memcg_stock_pcp
*stock
;
2565 struct mem_cgroup
*iter
;
2567 if (action
== CPU_ONLINE
)
2570 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2573 for_each_mem_cgroup(iter
)
2574 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2576 stock
= &per_cpu(memcg_stock
, cpu
);
2582 /* See __mem_cgroup_try_charge() for details */
2584 CHARGE_OK
, /* success */
2585 CHARGE_RETRY
, /* need to retry but retry is not bad */
2586 CHARGE_NOMEM
, /* we can't do more. return -ENOMEM */
2587 CHARGE_WOULDBLOCK
, /* GFP_WAIT wasn't set and no enough res. */
2590 static int mem_cgroup_do_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2591 unsigned int nr_pages
, unsigned int min_pages
,
2594 unsigned long csize
= nr_pages
* PAGE_SIZE
;
2595 struct mem_cgroup
*mem_over_limit
;
2596 struct res_counter
*fail_res
;
2597 unsigned long flags
= 0;
2600 ret
= res_counter_charge(&memcg
->res
, csize
, &fail_res
);
2603 if (!do_swap_account
)
2605 ret
= res_counter_charge(&memcg
->memsw
, csize
, &fail_res
);
2609 res_counter_uncharge(&memcg
->res
, csize
);
2610 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
2611 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
2613 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
2615 * Never reclaim on behalf of optional batching, retry with a
2616 * single page instead.
2618 if (nr_pages
> min_pages
)
2619 return CHARGE_RETRY
;
2621 if (!(gfp_mask
& __GFP_WAIT
))
2622 return CHARGE_WOULDBLOCK
;
2624 if (gfp_mask
& __GFP_NORETRY
)
2625 return CHARGE_NOMEM
;
2627 ret
= mem_cgroup_reclaim(mem_over_limit
, gfp_mask
, flags
);
2628 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2629 return CHARGE_RETRY
;
2631 * Even though the limit is exceeded at this point, reclaim
2632 * may have been able to free some pages. Retry the charge
2633 * before killing the task.
2635 * Only for regular pages, though: huge pages are rather
2636 * unlikely to succeed so close to the limit, and we fall back
2637 * to regular pages anyway in case of failure.
2639 if (nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
) && ret
)
2640 return CHARGE_RETRY
;
2643 * At task move, charge accounts can be doubly counted. So, it's
2644 * better to wait until the end of task_move if something is going on.
2646 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2647 return CHARGE_RETRY
;
2650 mem_cgroup_oom(mem_over_limit
, gfp_mask
, get_order(csize
));
2652 return CHARGE_NOMEM
;
2656 * __mem_cgroup_try_charge() does
2657 * 1. detect memcg to be charged against from passed *mm and *ptr,
2658 * 2. update res_counter
2659 * 3. call memory reclaim if necessary.
2661 * In some special case, if the task is fatal, fatal_signal_pending() or
2662 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2663 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2664 * as possible without any hazards. 2: all pages should have a valid
2665 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2666 * pointer, that is treated as a charge to root_mem_cgroup.
2668 * So __mem_cgroup_try_charge() will return
2669 * 0 ... on success, filling *ptr with a valid memcg pointer.
2670 * -ENOMEM ... charge failure because of resource limits.
2671 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2673 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2674 * the oom-killer can be invoked.
2676 static int __mem_cgroup_try_charge(struct mm_struct
*mm
,
2678 unsigned int nr_pages
,
2679 struct mem_cgroup
**ptr
,
2682 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2683 int nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2684 struct mem_cgroup
*memcg
= NULL
;
2688 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2689 * in system level. So, allow to go ahead dying process in addition to
2692 if (unlikely(test_thread_flag(TIF_MEMDIE
)
2693 || fatal_signal_pending(current
)))
2696 if (unlikely(task_in_memcg_oom(current
)))
2700 * We always charge the cgroup the mm_struct belongs to.
2701 * The mm_struct's mem_cgroup changes on task migration if the
2702 * thread group leader migrates. It's possible that mm is not
2703 * set, if so charge the root memcg (happens for pagecache usage).
2706 *ptr
= root_mem_cgroup
;
2708 if (*ptr
) { /* css should be a valid one */
2710 if (mem_cgroup_is_root(memcg
))
2712 if (consume_stock(memcg
, nr_pages
))
2714 css_get(&memcg
->css
);
2716 struct task_struct
*p
;
2719 p
= rcu_dereference(mm
->owner
);
2721 * Because we don't have task_lock(), "p" can exit.
2722 * In that case, "memcg" can point to root or p can be NULL with
2723 * race with swapoff. Then, we have small risk of mis-accouning.
2724 * But such kind of mis-account by race always happens because
2725 * we don't have cgroup_mutex(). It's overkill and we allo that
2727 * (*) swapoff at el will charge against mm-struct not against
2728 * task-struct. So, mm->owner can be NULL.
2730 memcg
= mem_cgroup_from_task(p
);
2732 memcg
= root_mem_cgroup
;
2733 if (mem_cgroup_is_root(memcg
)) {
2737 if (consume_stock(memcg
, nr_pages
)) {
2739 * It seems dagerous to access memcg without css_get().
2740 * But considering how consume_stok works, it's not
2741 * necessary. If consume_stock success, some charges
2742 * from this memcg are cached on this cpu. So, we
2743 * don't need to call css_get()/css_tryget() before
2744 * calling consume_stock().
2749 /* after here, we may be blocked. we need to get refcnt */
2750 if (!css_tryget(&memcg
->css
)) {
2758 bool invoke_oom
= oom
&& !nr_oom_retries
;
2760 /* If killed, bypass charge */
2761 if (fatal_signal_pending(current
)) {
2762 css_put(&memcg
->css
);
2766 ret
= mem_cgroup_do_charge(memcg
, gfp_mask
, batch
,
2767 nr_pages
, invoke_oom
);
2771 case CHARGE_RETRY
: /* not in OOM situation but retry */
2773 css_put(&memcg
->css
);
2776 case CHARGE_WOULDBLOCK
: /* !__GFP_WAIT */
2777 css_put(&memcg
->css
);
2779 case CHARGE_NOMEM
: /* OOM routine works */
2780 if (!oom
|| invoke_oom
) {
2781 css_put(&memcg
->css
);
2787 } while (ret
!= CHARGE_OK
);
2789 if (batch
> nr_pages
)
2790 refill_stock(memcg
, batch
- nr_pages
);
2791 css_put(&memcg
->css
);
2796 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2801 *ptr
= root_mem_cgroup
;
2806 * Somemtimes we have to undo a charge we got by try_charge().
2807 * This function is for that and do uncharge, put css's refcnt.
2808 * gotten by try_charge().
2810 static void __mem_cgroup_cancel_charge(struct mem_cgroup
*memcg
,
2811 unsigned int nr_pages
)
2813 if (!mem_cgroup_is_root(memcg
)) {
2814 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2816 res_counter_uncharge(&memcg
->res
, bytes
);
2817 if (do_swap_account
)
2818 res_counter_uncharge(&memcg
->memsw
, bytes
);
2823 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2824 * This is useful when moving usage to parent cgroup.
2826 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup
*memcg
,
2827 unsigned int nr_pages
)
2829 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2831 if (mem_cgroup_is_root(memcg
))
2834 res_counter_uncharge_until(&memcg
->res
, memcg
->res
.parent
, bytes
);
2835 if (do_swap_account
)
2836 res_counter_uncharge_until(&memcg
->memsw
,
2837 memcg
->memsw
.parent
, bytes
);
2841 * A helper function to get mem_cgroup from ID. must be called under
2842 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2843 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2844 * called against removed memcg.)
2846 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2848 /* ID 0 is unused ID */
2851 return mem_cgroup_from_id(id
);
2854 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2856 struct mem_cgroup
*memcg
= NULL
;
2857 struct page_cgroup
*pc
;
2861 VM_BUG_ON(!PageLocked(page
));
2863 pc
= lookup_page_cgroup(page
);
2864 lock_page_cgroup(pc
);
2865 if (PageCgroupUsed(pc
)) {
2866 memcg
= pc
->mem_cgroup
;
2867 if (memcg
&& !css_tryget(&memcg
->css
))
2869 } else if (PageSwapCache(page
)) {
2870 ent
.val
= page_private(page
);
2871 id
= lookup_swap_cgroup_id(ent
);
2873 memcg
= mem_cgroup_lookup(id
);
2874 if (memcg
&& !css_tryget(&memcg
->css
))
2878 unlock_page_cgroup(pc
);
2882 static void __mem_cgroup_commit_charge(struct mem_cgroup
*memcg
,
2884 unsigned int nr_pages
,
2885 enum charge_type ctype
,
2888 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2889 struct zone
*uninitialized_var(zone
);
2890 struct lruvec
*lruvec
;
2891 bool was_on_lru
= false;
2894 lock_page_cgroup(pc
);
2895 VM_BUG_ON(PageCgroupUsed(pc
));
2897 * we don't need page_cgroup_lock about tail pages, becase they are not
2898 * accessed by any other context at this point.
2902 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2903 * may already be on some other mem_cgroup's LRU. Take care of it.
2906 zone
= page_zone(page
);
2907 spin_lock_irq(&zone
->lru_lock
);
2908 if (PageLRU(page
)) {
2909 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2911 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2916 pc
->mem_cgroup
= memcg
;
2918 * We access a page_cgroup asynchronously without lock_page_cgroup().
2919 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2920 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2921 * before USED bit, we need memory barrier here.
2922 * See mem_cgroup_add_lru_list(), etc.
2925 SetPageCgroupUsed(pc
);
2929 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2930 VM_BUG_ON(PageLRU(page
));
2932 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2934 spin_unlock_irq(&zone
->lru_lock
);
2937 if (ctype
== MEM_CGROUP_CHARGE_TYPE_ANON
)
2942 mem_cgroup_charge_statistics(memcg
, page
, anon
, nr_pages
);
2943 unlock_page_cgroup(pc
);
2946 * "charge_statistics" updated event counter. Then, check it.
2947 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2948 * if they exceeds softlimit.
2950 memcg_check_events(memcg
, page
);
2953 static DEFINE_MUTEX(set_limit_mutex
);
2955 #ifdef CONFIG_MEMCG_KMEM
2956 static inline bool memcg_can_account_kmem(struct mem_cgroup
*memcg
)
2958 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
) &&
2959 (memcg
->kmem_account_flags
& KMEM_ACCOUNTED_MASK
);
2963 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2964 * in the memcg_cache_params struct.
2966 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
2968 struct kmem_cache
*cachep
;
2970 VM_BUG_ON(p
->is_root_cache
);
2971 cachep
= p
->root_cache
;
2972 return cache_from_memcg_idx(cachep
, memcg_cache_id(p
->memcg
));
2975 #ifdef CONFIG_SLABINFO
2976 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state
*css
,
2977 struct cftype
*cft
, struct seq_file
*m
)
2979 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2980 struct memcg_cache_params
*params
;
2982 if (!memcg_can_account_kmem(memcg
))
2985 print_slabinfo_header(m
);
2987 mutex_lock(&memcg
->slab_caches_mutex
);
2988 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
)
2989 cache_show(memcg_params_to_cache(params
), m
);
2990 mutex_unlock(&memcg
->slab_caches_mutex
);
2996 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
, u64 size
)
2998 struct res_counter
*fail_res
;
2999 struct mem_cgroup
*_memcg
;
3002 ret
= res_counter_charge(&memcg
->kmem
, size
, &fail_res
);
3007 ret
= __mem_cgroup_try_charge(NULL
, gfp
, size
>> PAGE_SHIFT
,
3008 &_memcg
, oom_gfp_allowed(gfp
));
3010 if (ret
== -EINTR
) {
3012 * __mem_cgroup_try_charge() chosed to bypass to root due to
3013 * OOM kill or fatal signal. Since our only options are to
3014 * either fail the allocation or charge it to this cgroup, do
3015 * it as a temporary condition. But we can't fail. From a
3016 * kmem/slab perspective, the cache has already been selected,
3017 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3020 * This condition will only trigger if the task entered
3021 * memcg_charge_kmem in a sane state, but was OOM-killed during
3022 * __mem_cgroup_try_charge() above. Tasks that were already
3023 * dying when the allocation triggers should have been already
3024 * directed to the root cgroup in memcontrol.h
3026 res_counter_charge_nofail(&memcg
->res
, size
, &fail_res
);
3027 if (do_swap_account
)
3028 res_counter_charge_nofail(&memcg
->memsw
, size
,
3032 res_counter_uncharge(&memcg
->kmem
, size
);
3037 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, u64 size
)
3039 res_counter_uncharge(&memcg
->res
, size
);
3040 if (do_swap_account
)
3041 res_counter_uncharge(&memcg
->memsw
, size
);
3044 if (res_counter_uncharge(&memcg
->kmem
, size
))
3048 * Releases a reference taken in kmem_cgroup_css_offline in case
3049 * this last uncharge is racing with the offlining code or it is
3050 * outliving the memcg existence.
3052 * The memory barrier imposed by test&clear is paired with the
3053 * explicit one in memcg_kmem_mark_dead().
3055 if (memcg_kmem_test_and_clear_dead(memcg
))
3056 css_put(&memcg
->css
);
3059 void memcg_cache_list_add(struct mem_cgroup
*memcg
, struct kmem_cache
*cachep
)
3064 mutex_lock(&memcg
->slab_caches_mutex
);
3065 list_add(&cachep
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
3066 mutex_unlock(&memcg
->slab_caches_mutex
);
3070 * helper for acessing a memcg's index. It will be used as an index in the
3071 * child cache array in kmem_cache, and also to derive its name. This function
3072 * will return -1 when this is not a kmem-limited memcg.
3074 int memcg_cache_id(struct mem_cgroup
*memcg
)
3076 return memcg
? memcg
->kmemcg_id
: -1;
3080 * This ends up being protected by the set_limit mutex, during normal
3081 * operation, because that is its main call site.
3083 * But when we create a new cache, we can call this as well if its parent
3084 * is kmem-limited. That will have to hold set_limit_mutex as well.
3086 int memcg_update_cache_sizes(struct mem_cgroup
*memcg
)
3090 num
= ida_simple_get(&kmem_limited_groups
,
3091 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
3095 * After this point, kmem_accounted (that we test atomically in
3096 * the beginning of this conditional), is no longer 0. This
3097 * guarantees only one process will set the following boolean
3098 * to true. We don't need test_and_set because we're protected
3099 * by the set_limit_mutex anyway.
3101 memcg_kmem_set_activated(memcg
);
3103 ret
= memcg_update_all_caches(num
+1);
3105 ida_simple_remove(&kmem_limited_groups
, num
);
3106 memcg_kmem_clear_activated(memcg
);
3110 memcg
->kmemcg_id
= num
;
3111 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
3112 mutex_init(&memcg
->slab_caches_mutex
);
3116 static size_t memcg_caches_array_size(int num_groups
)
3119 if (num_groups
<= 0)
3122 size
= 2 * num_groups
;
3123 if (size
< MEMCG_CACHES_MIN_SIZE
)
3124 size
= MEMCG_CACHES_MIN_SIZE
;
3125 else if (size
> MEMCG_CACHES_MAX_SIZE
)
3126 size
= MEMCG_CACHES_MAX_SIZE
;
3132 * We should update the current array size iff all caches updates succeed. This
3133 * can only be done from the slab side. The slab mutex needs to be held when
3136 void memcg_update_array_size(int num
)
3138 if (num
> memcg_limited_groups_array_size
)
3139 memcg_limited_groups_array_size
= memcg_caches_array_size(num
);
3142 static void kmem_cache_destroy_work_func(struct work_struct
*w
);
3144 int memcg_update_cache_size(struct kmem_cache
*s
, int num_groups
)
3146 struct memcg_cache_params
*cur_params
= s
->memcg_params
;
3148 VM_BUG_ON(!is_root_cache(s
));
3150 if (num_groups
> memcg_limited_groups_array_size
) {
3152 ssize_t size
= memcg_caches_array_size(num_groups
);
3154 size
*= sizeof(void *);
3155 size
+= offsetof(struct memcg_cache_params
, memcg_caches
);
3157 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3158 if (!s
->memcg_params
) {
3159 s
->memcg_params
= cur_params
;
3163 s
->memcg_params
->is_root_cache
= true;
3166 * There is the chance it will be bigger than
3167 * memcg_limited_groups_array_size, if we failed an allocation
3168 * in a cache, in which case all caches updated before it, will
3169 * have a bigger array.
3171 * But if that is the case, the data after
3172 * memcg_limited_groups_array_size is certainly unused
3174 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3175 if (!cur_params
->memcg_caches
[i
])
3177 s
->memcg_params
->memcg_caches
[i
] =
3178 cur_params
->memcg_caches
[i
];
3182 * Ideally, we would wait until all caches succeed, and only
3183 * then free the old one. But this is not worth the extra
3184 * pointer per-cache we'd have to have for this.
3186 * It is not a big deal if some caches are left with a size
3187 * bigger than the others. And all updates will reset this
3195 int memcg_register_cache(struct mem_cgroup
*memcg
, struct kmem_cache
*s
,
3196 struct kmem_cache
*root_cache
)
3200 if (!memcg_kmem_enabled())
3204 size
= offsetof(struct memcg_cache_params
, memcg_caches
);
3205 size
+= memcg_limited_groups_array_size
* sizeof(void *);
3207 size
= sizeof(struct memcg_cache_params
);
3209 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3210 if (!s
->memcg_params
)
3214 s
->memcg_params
->memcg
= memcg
;
3215 s
->memcg_params
->root_cache
= root_cache
;
3216 INIT_WORK(&s
->memcg_params
->destroy
,
3217 kmem_cache_destroy_work_func
);
3219 s
->memcg_params
->is_root_cache
= true;
3224 void memcg_release_cache(struct kmem_cache
*s
)
3226 struct kmem_cache
*root
;
3227 struct mem_cgroup
*memcg
;
3231 * This happens, for instance, when a root cache goes away before we
3234 if (!s
->memcg_params
)
3237 if (s
->memcg_params
->is_root_cache
)
3240 memcg
= s
->memcg_params
->memcg
;
3241 id
= memcg_cache_id(memcg
);
3243 root
= s
->memcg_params
->root_cache
;
3244 root
->memcg_params
->memcg_caches
[id
] = NULL
;
3246 mutex_lock(&memcg
->slab_caches_mutex
);
3247 list_del(&s
->memcg_params
->list
);
3248 mutex_unlock(&memcg
->slab_caches_mutex
);
3250 css_put(&memcg
->css
);
3252 kfree(s
->memcg_params
);
3256 * During the creation a new cache, we need to disable our accounting mechanism
3257 * altogether. This is true even if we are not creating, but rather just
3258 * enqueing new caches to be created.
3260 * This is because that process will trigger allocations; some visible, like
3261 * explicit kmallocs to auxiliary data structures, name strings and internal
3262 * cache structures; some well concealed, like INIT_WORK() that can allocate
3263 * objects during debug.
3265 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3266 * to it. This may not be a bounded recursion: since the first cache creation
3267 * failed to complete (waiting on the allocation), we'll just try to create the
3268 * cache again, failing at the same point.
3270 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3271 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3272 * inside the following two functions.
3274 static inline void memcg_stop_kmem_account(void)
3276 VM_BUG_ON(!current
->mm
);
3277 current
->memcg_kmem_skip_account
++;
3280 static inline void memcg_resume_kmem_account(void)
3282 VM_BUG_ON(!current
->mm
);
3283 current
->memcg_kmem_skip_account
--;
3286 static void kmem_cache_destroy_work_func(struct work_struct
*w
)
3288 struct kmem_cache
*cachep
;
3289 struct memcg_cache_params
*p
;
3291 p
= container_of(w
, struct memcg_cache_params
, destroy
);
3293 cachep
= memcg_params_to_cache(p
);
3296 * If we get down to 0 after shrink, we could delete right away.
3297 * However, memcg_release_pages() already puts us back in the workqueue
3298 * in that case. If we proceed deleting, we'll get a dangling
3299 * reference, and removing the object from the workqueue in that case
3300 * is unnecessary complication. We are not a fast path.
3302 * Note that this case is fundamentally different from racing with
3303 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3304 * kmem_cache_shrink, not only we would be reinserting a dead cache
3305 * into the queue, but doing so from inside the worker racing to
3308 * So if we aren't down to zero, we'll just schedule a worker and try
3311 if (atomic_read(&cachep
->memcg_params
->nr_pages
) != 0) {
3312 kmem_cache_shrink(cachep
);
3313 if (atomic_read(&cachep
->memcg_params
->nr_pages
) == 0)
3316 kmem_cache_destroy(cachep
);
3319 void mem_cgroup_destroy_cache(struct kmem_cache
*cachep
)
3321 if (!cachep
->memcg_params
->dead
)
3325 * There are many ways in which we can get here.
3327 * We can get to a memory-pressure situation while the delayed work is
3328 * still pending to run. The vmscan shrinkers can then release all
3329 * cache memory and get us to destruction. If this is the case, we'll
3330 * be executed twice, which is a bug (the second time will execute over
3331 * bogus data). In this case, cancelling the work should be fine.
3333 * But we can also get here from the worker itself, if
3334 * kmem_cache_shrink is enough to shake all the remaining objects and
3335 * get the page count to 0. In this case, we'll deadlock if we try to
3336 * cancel the work (the worker runs with an internal lock held, which
3337 * is the same lock we would hold for cancel_work_sync().)
3339 * Since we can't possibly know who got us here, just refrain from
3340 * running if there is already work pending
3342 if (work_pending(&cachep
->memcg_params
->destroy
))
3345 * We have to defer the actual destroying to a workqueue, because
3346 * we might currently be in a context that cannot sleep.
3348 schedule_work(&cachep
->memcg_params
->destroy
);
3352 * This lock protects updaters, not readers. We want readers to be as fast as
3353 * they can, and they will either see NULL or a valid cache value. Our model
3354 * allow them to see NULL, in which case the root memcg will be selected.
3356 * We need this lock because multiple allocations to the same cache from a non
3357 * will span more than one worker. Only one of them can create the cache.
3359 static DEFINE_MUTEX(memcg_cache_mutex
);
3362 * Called with memcg_cache_mutex held
3364 static struct kmem_cache
*kmem_cache_dup(struct mem_cgroup
*memcg
,
3365 struct kmem_cache
*s
)
3367 struct kmem_cache
*new;
3368 static char *tmp_name
= NULL
;
3370 lockdep_assert_held(&memcg_cache_mutex
);
3373 * kmem_cache_create_memcg duplicates the given name and
3374 * cgroup_name for this name requires RCU context.
3375 * This static temporary buffer is used to prevent from
3376 * pointless shortliving allocation.
3379 tmp_name
= kmalloc(PATH_MAX
, GFP_KERNEL
);
3385 snprintf(tmp_name
, PATH_MAX
, "%s(%d:%s)", s
->name
,
3386 memcg_cache_id(memcg
), cgroup_name(memcg
->css
.cgroup
));
3389 new = kmem_cache_create_memcg(memcg
, tmp_name
, s
->object_size
, s
->align
,
3390 (s
->flags
& ~SLAB_PANIC
), s
->ctor
, s
);
3393 new->allocflags
|= __GFP_KMEMCG
;
3398 static struct kmem_cache
*memcg_create_kmem_cache(struct mem_cgroup
*memcg
,
3399 struct kmem_cache
*cachep
)
3401 struct kmem_cache
*new_cachep
;
3404 BUG_ON(!memcg_can_account_kmem(memcg
));
3406 idx
= memcg_cache_id(memcg
);
3408 mutex_lock(&memcg_cache_mutex
);
3409 new_cachep
= cache_from_memcg_idx(cachep
, idx
);
3411 css_put(&memcg
->css
);
3415 new_cachep
= kmem_cache_dup(memcg
, cachep
);
3416 if (new_cachep
== NULL
) {
3417 new_cachep
= cachep
;
3418 css_put(&memcg
->css
);
3422 atomic_set(&new_cachep
->memcg_params
->nr_pages
, 0);
3424 cachep
->memcg_params
->memcg_caches
[idx
] = new_cachep
;
3426 * the readers won't lock, make sure everybody sees the updated value,
3427 * so they won't put stuff in the queue again for no reason
3431 mutex_unlock(&memcg_cache_mutex
);
3435 void kmem_cache_destroy_memcg_children(struct kmem_cache
*s
)
3437 struct kmem_cache
*c
;
3440 if (!s
->memcg_params
)
3442 if (!s
->memcg_params
->is_root_cache
)
3446 * If the cache is being destroyed, we trust that there is no one else
3447 * requesting objects from it. Even if there are, the sanity checks in
3448 * kmem_cache_destroy should caught this ill-case.
3450 * Still, we don't want anyone else freeing memcg_caches under our
3451 * noses, which can happen if a new memcg comes to life. As usual,
3452 * we'll take the set_limit_mutex to protect ourselves against this.
3454 mutex_lock(&set_limit_mutex
);
3455 for_each_memcg_cache_index(i
) {
3456 c
= cache_from_memcg_idx(s
, i
);
3461 * We will now manually delete the caches, so to avoid races
3462 * we need to cancel all pending destruction workers and
3463 * proceed with destruction ourselves.
3465 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3466 * and that could spawn the workers again: it is likely that
3467 * the cache still have active pages until this very moment.
3468 * This would lead us back to mem_cgroup_destroy_cache.
3470 * But that will not execute at all if the "dead" flag is not
3471 * set, so flip it down to guarantee we are in control.
3473 c
->memcg_params
->dead
= false;
3474 cancel_work_sync(&c
->memcg_params
->destroy
);
3475 kmem_cache_destroy(c
);
3477 mutex_unlock(&set_limit_mutex
);
3480 struct create_work
{
3481 struct mem_cgroup
*memcg
;
3482 struct kmem_cache
*cachep
;
3483 struct work_struct work
;
3486 static void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3488 struct kmem_cache
*cachep
;
3489 struct memcg_cache_params
*params
;
3491 if (!memcg_kmem_is_active(memcg
))
3494 mutex_lock(&memcg
->slab_caches_mutex
);
3495 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
) {
3496 cachep
= memcg_params_to_cache(params
);
3497 cachep
->memcg_params
->dead
= true;
3498 schedule_work(&cachep
->memcg_params
->destroy
);
3500 mutex_unlock(&memcg
->slab_caches_mutex
);
3503 static void memcg_create_cache_work_func(struct work_struct
*w
)
3505 struct create_work
*cw
;
3507 cw
= container_of(w
, struct create_work
, work
);
3508 memcg_create_kmem_cache(cw
->memcg
, cw
->cachep
);
3513 * Enqueue the creation of a per-memcg kmem_cache.
3515 static void __memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3516 struct kmem_cache
*cachep
)
3518 struct create_work
*cw
;
3520 cw
= kmalloc(sizeof(struct create_work
), GFP_NOWAIT
);
3522 css_put(&memcg
->css
);
3527 cw
->cachep
= cachep
;
3529 INIT_WORK(&cw
->work
, memcg_create_cache_work_func
);
3530 schedule_work(&cw
->work
);
3533 static void memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3534 struct kmem_cache
*cachep
)
3537 * We need to stop accounting when we kmalloc, because if the
3538 * corresponding kmalloc cache is not yet created, the first allocation
3539 * in __memcg_create_cache_enqueue will recurse.
3541 * However, it is better to enclose the whole function. Depending on
3542 * the debugging options enabled, INIT_WORK(), for instance, can
3543 * trigger an allocation. This too, will make us recurse. Because at
3544 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3545 * the safest choice is to do it like this, wrapping the whole function.
3547 memcg_stop_kmem_account();
3548 __memcg_create_cache_enqueue(memcg
, cachep
);
3549 memcg_resume_kmem_account();
3552 * Return the kmem_cache we're supposed to use for a slab allocation.
3553 * We try to use the current memcg's version of the cache.
3555 * If the cache does not exist yet, if we are the first user of it,
3556 * we either create it immediately, if possible, or create it asynchronously
3558 * In the latter case, we will let the current allocation go through with
3559 * the original cache.
3561 * Can't be called in interrupt context or from kernel threads.
3562 * This function needs to be called with rcu_read_lock() held.
3564 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
3567 struct mem_cgroup
*memcg
;
3570 VM_BUG_ON(!cachep
->memcg_params
);
3571 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
3573 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3577 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
3579 if (!memcg_can_account_kmem(memcg
))
3582 idx
= memcg_cache_id(memcg
);
3585 * barrier to mare sure we're always seeing the up to date value. The
3586 * code updating memcg_caches will issue a write barrier to match this.
3588 read_barrier_depends();
3589 if (likely(cache_from_memcg_idx(cachep
, idx
))) {
3590 cachep
= cache_from_memcg_idx(cachep
, idx
);
3594 /* The corresponding put will be done in the workqueue. */
3595 if (!css_tryget(&memcg
->css
))
3600 * If we are in a safe context (can wait, and not in interrupt
3601 * context), we could be be predictable and return right away.
3602 * This would guarantee that the allocation being performed
3603 * already belongs in the new cache.
3605 * However, there are some clashes that can arrive from locking.
3606 * For instance, because we acquire the slab_mutex while doing
3607 * kmem_cache_dup, this means no further allocation could happen
3608 * with the slab_mutex held.
3610 * Also, because cache creation issue get_online_cpus(), this
3611 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3612 * that ends up reversed during cpu hotplug. (cpuset allocates
3613 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3614 * better to defer everything.
3616 memcg_create_cache_enqueue(memcg
, cachep
);
3622 EXPORT_SYMBOL(__memcg_kmem_get_cache
);
3625 * We need to verify if the allocation against current->mm->owner's memcg is
3626 * possible for the given order. But the page is not allocated yet, so we'll
3627 * need a further commit step to do the final arrangements.
3629 * It is possible for the task to switch cgroups in this mean time, so at
3630 * commit time, we can't rely on task conversion any longer. We'll then use
3631 * the handle argument to return to the caller which cgroup we should commit
3632 * against. We could also return the memcg directly and avoid the pointer
3633 * passing, but a boolean return value gives better semantics considering
3634 * the compiled-out case as well.
3636 * Returning true means the allocation is possible.
3639 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
3641 struct mem_cgroup
*memcg
;
3647 * Disabling accounting is only relevant for some specific memcg
3648 * internal allocations. Therefore we would initially not have such
3649 * check here, since direct calls to the page allocator that are marked
3650 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3651 * concerned with cache allocations, and by having this test at
3652 * memcg_kmem_get_cache, we are already able to relay the allocation to
3653 * the root cache and bypass the memcg cache altogether.
3655 * There is one exception, though: the SLUB allocator does not create
3656 * large order caches, but rather service large kmallocs directly from
3657 * the page allocator. Therefore, the following sequence when backed by
3658 * the SLUB allocator:
3660 * memcg_stop_kmem_account();
3661 * kmalloc(<large_number>)
3662 * memcg_resume_kmem_account();
3664 * would effectively ignore the fact that we should skip accounting,
3665 * since it will drive us directly to this function without passing
3666 * through the cache selector memcg_kmem_get_cache. Such large
3667 * allocations are extremely rare but can happen, for instance, for the
3668 * cache arrays. We bring this test here.
3670 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3673 memcg
= try_get_mem_cgroup_from_mm(current
->mm
);
3676 * very rare case described in mem_cgroup_from_task. Unfortunately there
3677 * isn't much we can do without complicating this too much, and it would
3678 * be gfp-dependent anyway. Just let it go
3680 if (unlikely(!memcg
))
3683 if (!memcg_can_account_kmem(memcg
)) {
3684 css_put(&memcg
->css
);
3688 ret
= memcg_charge_kmem(memcg
, gfp
, PAGE_SIZE
<< order
);
3692 css_put(&memcg
->css
);
3696 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
3699 struct page_cgroup
*pc
;
3701 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3703 /* The page allocation failed. Revert */
3705 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3709 pc
= lookup_page_cgroup(page
);
3710 lock_page_cgroup(pc
);
3711 pc
->mem_cgroup
= memcg
;
3712 SetPageCgroupUsed(pc
);
3713 unlock_page_cgroup(pc
);
3716 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
3718 struct mem_cgroup
*memcg
= NULL
;
3719 struct page_cgroup
*pc
;
3722 pc
= lookup_page_cgroup(page
);
3724 * Fast unlocked return. Theoretically might have changed, have to
3725 * check again after locking.
3727 if (!PageCgroupUsed(pc
))
3730 lock_page_cgroup(pc
);
3731 if (PageCgroupUsed(pc
)) {
3732 memcg
= pc
->mem_cgroup
;
3733 ClearPageCgroupUsed(pc
);
3735 unlock_page_cgroup(pc
);
3738 * We trust that only if there is a memcg associated with the page, it
3739 * is a valid allocation
3744 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3745 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3748 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3751 #endif /* CONFIG_MEMCG_KMEM */
3753 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3755 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3757 * Because tail pages are not marked as "used", set it. We're under
3758 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3759 * charge/uncharge will be never happen and move_account() is done under
3760 * compound_lock(), so we don't have to take care of races.
3762 void mem_cgroup_split_huge_fixup(struct page
*head
)
3764 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
3765 struct page_cgroup
*pc
;
3766 struct mem_cgroup
*memcg
;
3769 if (mem_cgroup_disabled())
3772 memcg
= head_pc
->mem_cgroup
;
3773 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3775 pc
->mem_cgroup
= memcg
;
3776 smp_wmb();/* see __commit_charge() */
3777 pc
->flags
= head_pc
->flags
& ~PCGF_NOCOPY_AT_SPLIT
;
3779 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
3782 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3785 void mem_cgroup_move_account_page_stat(struct mem_cgroup
*from
,
3786 struct mem_cgroup
*to
,
3787 unsigned int nr_pages
,
3788 enum mem_cgroup_stat_index idx
)
3790 /* Update stat data for mem_cgroup */
3792 __this_cpu_sub(from
->stat
->count
[idx
], nr_pages
);
3793 __this_cpu_add(to
->stat
->count
[idx
], nr_pages
);
3798 * mem_cgroup_move_account - move account of the page
3800 * @nr_pages: number of regular pages (>1 for huge pages)
3801 * @pc: page_cgroup of the page.
3802 * @from: mem_cgroup which the page is moved from.
3803 * @to: mem_cgroup which the page is moved to. @from != @to.
3805 * The caller must confirm following.
3806 * - page is not on LRU (isolate_page() is useful.)
3807 * - compound_lock is held when nr_pages > 1
3809 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3812 static int mem_cgroup_move_account(struct page
*page
,
3813 unsigned int nr_pages
,
3814 struct page_cgroup
*pc
,
3815 struct mem_cgroup
*from
,
3816 struct mem_cgroup
*to
)
3818 unsigned long flags
;
3820 bool anon
= PageAnon(page
);
3822 VM_BUG_ON(from
== to
);
3823 VM_BUG_ON(PageLRU(page
));
3825 * The page is isolated from LRU. So, collapse function
3826 * will not handle this page. But page splitting can happen.
3827 * Do this check under compound_page_lock(). The caller should
3831 if (nr_pages
> 1 && !PageTransHuge(page
))
3834 lock_page_cgroup(pc
);
3837 if (!PageCgroupUsed(pc
) || pc
->mem_cgroup
!= from
)
3840 move_lock_mem_cgroup(from
, &flags
);
3842 if (!anon
&& page_mapped(page
))
3843 mem_cgroup_move_account_page_stat(from
, to
, nr_pages
,
3844 MEM_CGROUP_STAT_FILE_MAPPED
);
3846 if (PageWriteback(page
))
3847 mem_cgroup_move_account_page_stat(from
, to
, nr_pages
,
3848 MEM_CGROUP_STAT_WRITEBACK
);
3850 mem_cgroup_charge_statistics(from
, page
, anon
, -nr_pages
);
3852 /* caller should have done css_get */
3853 pc
->mem_cgroup
= to
;
3854 mem_cgroup_charge_statistics(to
, page
, anon
, nr_pages
);
3855 move_unlock_mem_cgroup(from
, &flags
);
3858 unlock_page_cgroup(pc
);
3862 memcg_check_events(to
, page
);
3863 memcg_check_events(from
, page
);
3869 * mem_cgroup_move_parent - moves page to the parent group
3870 * @page: the page to move
3871 * @pc: page_cgroup of the page
3872 * @child: page's cgroup
3874 * move charges to its parent or the root cgroup if the group has no
3875 * parent (aka use_hierarchy==0).
3876 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3877 * mem_cgroup_move_account fails) the failure is always temporary and
3878 * it signals a race with a page removal/uncharge or migration. In the
3879 * first case the page is on the way out and it will vanish from the LRU
3880 * on the next attempt and the call should be retried later.
3881 * Isolation from the LRU fails only if page has been isolated from
3882 * the LRU since we looked at it and that usually means either global
3883 * reclaim or migration going on. The page will either get back to the
3885 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3886 * (!PageCgroupUsed) or moved to a different group. The page will
3887 * disappear in the next attempt.
3889 static int mem_cgroup_move_parent(struct page
*page
,
3890 struct page_cgroup
*pc
,
3891 struct mem_cgroup
*child
)
3893 struct mem_cgroup
*parent
;
3894 unsigned int nr_pages
;
3895 unsigned long uninitialized_var(flags
);
3898 VM_BUG_ON(mem_cgroup_is_root(child
));
3901 if (!get_page_unless_zero(page
))
3903 if (isolate_lru_page(page
))
3906 nr_pages
= hpage_nr_pages(page
);
3908 parent
= parent_mem_cgroup(child
);
3910 * If no parent, move charges to root cgroup.
3913 parent
= root_mem_cgroup
;
3916 VM_BUG_ON(!PageTransHuge(page
));
3917 flags
= compound_lock_irqsave(page
);
3920 ret
= mem_cgroup_move_account(page
, nr_pages
,
3923 __mem_cgroup_cancel_local_charge(child
, nr_pages
);
3926 compound_unlock_irqrestore(page
, flags
);
3927 putback_lru_page(page
);
3935 * Charge the memory controller for page usage.
3937 * 0 if the charge was successful
3938 * < 0 if the cgroup is over its limit
3940 static int mem_cgroup_charge_common(struct page
*page
, struct mm_struct
*mm
,
3941 gfp_t gfp_mask
, enum charge_type ctype
)
3943 struct mem_cgroup
*memcg
= NULL
;
3944 unsigned int nr_pages
= 1;
3948 if (PageTransHuge(page
)) {
3949 nr_pages
<<= compound_order(page
);
3950 VM_BUG_ON(!PageTransHuge(page
));
3952 * Never OOM-kill a process for a huge page. The
3953 * fault handler will fall back to regular pages.
3958 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, nr_pages
, &memcg
, oom
);
3961 __mem_cgroup_commit_charge(memcg
, page
, nr_pages
, ctype
, false);
3965 int mem_cgroup_newpage_charge(struct page
*page
,
3966 struct mm_struct
*mm
, gfp_t gfp_mask
)
3968 if (mem_cgroup_disabled())
3970 VM_BUG_ON(page_mapped(page
));
3971 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
3973 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
3974 MEM_CGROUP_CHARGE_TYPE_ANON
);
3978 * While swap-in, try_charge -> commit or cancel, the page is locked.
3979 * And when try_charge() successfully returns, one refcnt to memcg without
3980 * struct page_cgroup is acquired. This refcnt will be consumed by
3981 * "commit()" or removed by "cancel()"
3983 static int __mem_cgroup_try_charge_swapin(struct mm_struct
*mm
,
3986 struct mem_cgroup
**memcgp
)
3988 struct mem_cgroup
*memcg
;
3989 struct page_cgroup
*pc
;
3992 pc
= lookup_page_cgroup(page
);
3994 * Every swap fault against a single page tries to charge the
3995 * page, bail as early as possible. shmem_unuse() encounters
3996 * already charged pages, too. The USED bit is protected by
3997 * the page lock, which serializes swap cache removal, which
3998 * in turn serializes uncharging.
4000 if (PageCgroupUsed(pc
))
4002 if (!do_swap_account
)
4004 memcg
= try_get_mem_cgroup_from_page(page
);
4008 ret
= __mem_cgroup_try_charge(NULL
, mask
, 1, memcgp
, true);
4009 css_put(&memcg
->css
);
4014 ret
= __mem_cgroup_try_charge(mm
, mask
, 1, memcgp
, true);
4020 int mem_cgroup_try_charge_swapin(struct mm_struct
*mm
, struct page
*page
,
4021 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
4024 if (mem_cgroup_disabled())
4027 * A racing thread's fault, or swapoff, may have already
4028 * updated the pte, and even removed page from swap cache: in
4029 * those cases unuse_pte()'s pte_same() test will fail; but
4030 * there's also a KSM case which does need to charge the page.
4032 if (!PageSwapCache(page
)) {
4035 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, 1, memcgp
, true);
4040 return __mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, memcgp
);
4043 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup
*memcg
)
4045 if (mem_cgroup_disabled())
4049 __mem_cgroup_cancel_charge(memcg
, 1);
4053 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*memcg
,
4054 enum charge_type ctype
)
4056 if (mem_cgroup_disabled())
4061 __mem_cgroup_commit_charge(memcg
, page
, 1, ctype
, true);
4063 * Now swap is on-memory. This means this page may be
4064 * counted both as mem and swap....double count.
4065 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4066 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4067 * may call delete_from_swap_cache() before reach here.
4069 if (do_swap_account
&& PageSwapCache(page
)) {
4070 swp_entry_t ent
= {.val
= page_private(page
)};
4071 mem_cgroup_uncharge_swap(ent
);
4075 void mem_cgroup_commit_charge_swapin(struct page
*page
,
4076 struct mem_cgroup
*memcg
)
4078 __mem_cgroup_commit_charge_swapin(page
, memcg
,
4079 MEM_CGROUP_CHARGE_TYPE_ANON
);
4082 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
4085 struct mem_cgroup
*memcg
= NULL
;
4086 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4089 if (mem_cgroup_disabled())
4091 if (PageCompound(page
))
4094 if (!PageSwapCache(page
))
4095 ret
= mem_cgroup_charge_common(page
, mm
, gfp_mask
, type
);
4096 else { /* page is swapcache/shmem */
4097 ret
= __mem_cgroup_try_charge_swapin(mm
, page
,
4100 __mem_cgroup_commit_charge_swapin(page
, memcg
, type
);
4105 static void mem_cgroup_do_uncharge(struct mem_cgroup
*memcg
,
4106 unsigned int nr_pages
,
4107 const enum charge_type ctype
)
4109 struct memcg_batch_info
*batch
= NULL
;
4110 bool uncharge_memsw
= true;
4112 /* If swapout, usage of swap doesn't decrease */
4113 if (!do_swap_account
|| ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
)
4114 uncharge_memsw
= false;
4116 batch
= ¤t
->memcg_batch
;
4118 * In usual, we do css_get() when we remember memcg pointer.
4119 * But in this case, we keep res->usage until end of a series of
4120 * uncharges. Then, it's ok to ignore memcg's refcnt.
4123 batch
->memcg
= memcg
;
4125 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4126 * In those cases, all pages freed continuously can be expected to be in
4127 * the same cgroup and we have chance to coalesce uncharges.
4128 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4129 * because we want to do uncharge as soon as possible.
4132 if (!batch
->do_batch
|| test_thread_flag(TIF_MEMDIE
))
4133 goto direct_uncharge
;
4136 goto direct_uncharge
;
4139 * In typical case, batch->memcg == mem. This means we can
4140 * merge a series of uncharges to an uncharge of res_counter.
4141 * If not, we uncharge res_counter ony by one.
4143 if (batch
->memcg
!= memcg
)
4144 goto direct_uncharge
;
4145 /* remember freed charge and uncharge it later */
4148 batch
->memsw_nr_pages
++;
4151 res_counter_uncharge(&memcg
->res
, nr_pages
* PAGE_SIZE
);
4153 res_counter_uncharge(&memcg
->memsw
, nr_pages
* PAGE_SIZE
);
4154 if (unlikely(batch
->memcg
!= memcg
))
4155 memcg_oom_recover(memcg
);
4159 * uncharge if !page_mapped(page)
4161 static struct mem_cgroup
*
4162 __mem_cgroup_uncharge_common(struct page
*page
, enum charge_type ctype
,
4165 struct mem_cgroup
*memcg
= NULL
;
4166 unsigned int nr_pages
= 1;
4167 struct page_cgroup
*pc
;
4170 if (mem_cgroup_disabled())
4173 if (PageTransHuge(page
)) {
4174 nr_pages
<<= compound_order(page
);
4175 VM_BUG_ON(!PageTransHuge(page
));
4178 * Check if our page_cgroup is valid
4180 pc
= lookup_page_cgroup(page
);
4181 if (unlikely(!PageCgroupUsed(pc
)))
4184 lock_page_cgroup(pc
);
4186 memcg
= pc
->mem_cgroup
;
4188 if (!PageCgroupUsed(pc
))
4191 anon
= PageAnon(page
);
4194 case MEM_CGROUP_CHARGE_TYPE_ANON
:
4196 * Generally PageAnon tells if it's the anon statistics to be
4197 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4198 * used before page reached the stage of being marked PageAnon.
4202 case MEM_CGROUP_CHARGE_TYPE_DROP
:
4203 /* See mem_cgroup_prepare_migration() */
4204 if (page_mapped(page
))
4207 * Pages under migration may not be uncharged. But
4208 * end_migration() /must/ be the one uncharging the
4209 * unused post-migration page and so it has to call
4210 * here with the migration bit still set. See the
4211 * res_counter handling below.
4213 if (!end_migration
&& PageCgroupMigration(pc
))
4216 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT
:
4217 if (!PageAnon(page
)) { /* Shared memory */
4218 if (page
->mapping
&& !page_is_file_cache(page
))
4220 } else if (page_mapped(page
)) /* Anon */
4227 mem_cgroup_charge_statistics(memcg
, page
, anon
, -nr_pages
);
4229 ClearPageCgroupUsed(pc
);
4231 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4232 * freed from LRU. This is safe because uncharged page is expected not
4233 * to be reused (freed soon). Exception is SwapCache, it's handled by
4234 * special functions.
4237 unlock_page_cgroup(pc
);
4239 * even after unlock, we have memcg->res.usage here and this memcg
4240 * will never be freed, so it's safe to call css_get().
4242 memcg_check_events(memcg
, page
);
4243 if (do_swap_account
&& ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
) {
4244 mem_cgroup_swap_statistics(memcg
, true);
4245 css_get(&memcg
->css
);
4248 * Migration does not charge the res_counter for the
4249 * replacement page, so leave it alone when phasing out the
4250 * page that is unused after the migration.
4252 if (!end_migration
&& !mem_cgroup_is_root(memcg
))
4253 mem_cgroup_do_uncharge(memcg
, nr_pages
, ctype
);
4258 unlock_page_cgroup(pc
);
4262 void mem_cgroup_uncharge_page(struct page
*page
)
4265 if (page_mapped(page
))
4267 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
4269 * If the page is in swap cache, uncharge should be deferred
4270 * to the swap path, which also properly accounts swap usage
4271 * and handles memcg lifetime.
4273 * Note that this check is not stable and reclaim may add the
4274 * page to swap cache at any time after this. However, if the
4275 * page is not in swap cache by the time page->mapcount hits
4276 * 0, there won't be any page table references to the swap
4277 * slot, and reclaim will free it and not actually write the
4280 if (PageSwapCache(page
))
4282 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_ANON
, false);
4285 void mem_cgroup_uncharge_cache_page(struct page
*page
)
4287 VM_BUG_ON(page_mapped(page
));
4288 VM_BUG_ON(page
->mapping
);
4289 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_CACHE
, false);
4293 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4294 * In that cases, pages are freed continuously and we can expect pages
4295 * are in the same memcg. All these calls itself limits the number of
4296 * pages freed at once, then uncharge_start/end() is called properly.
4297 * This may be called prural(2) times in a context,
4300 void mem_cgroup_uncharge_start(void)
4302 current
->memcg_batch
.do_batch
++;
4303 /* We can do nest. */
4304 if (current
->memcg_batch
.do_batch
== 1) {
4305 current
->memcg_batch
.memcg
= NULL
;
4306 current
->memcg_batch
.nr_pages
= 0;
4307 current
->memcg_batch
.memsw_nr_pages
= 0;
4311 void mem_cgroup_uncharge_end(void)
4313 struct memcg_batch_info
*batch
= ¤t
->memcg_batch
;
4315 if (!batch
->do_batch
)
4319 if (batch
->do_batch
) /* If stacked, do nothing. */
4325 * This "batch->memcg" is valid without any css_get/put etc...
4326 * bacause we hide charges behind us.
4328 if (batch
->nr_pages
)
4329 res_counter_uncharge(&batch
->memcg
->res
,
4330 batch
->nr_pages
* PAGE_SIZE
);
4331 if (batch
->memsw_nr_pages
)
4332 res_counter_uncharge(&batch
->memcg
->memsw
,
4333 batch
->memsw_nr_pages
* PAGE_SIZE
);
4334 memcg_oom_recover(batch
->memcg
);
4335 /* forget this pointer (for sanity check) */
4336 batch
->memcg
= NULL
;
4341 * called after __delete_from_swap_cache() and drop "page" account.
4342 * memcg information is recorded to swap_cgroup of "ent"
4345 mem_cgroup_uncharge_swapcache(struct page
*page
, swp_entry_t ent
, bool swapout
)
4347 struct mem_cgroup
*memcg
;
4348 int ctype
= MEM_CGROUP_CHARGE_TYPE_SWAPOUT
;
4350 if (!swapout
) /* this was a swap cache but the swap is unused ! */
4351 ctype
= MEM_CGROUP_CHARGE_TYPE_DROP
;
4353 memcg
= __mem_cgroup_uncharge_common(page
, ctype
, false);
4356 * record memcg information, if swapout && memcg != NULL,
4357 * css_get() was called in uncharge().
4359 if (do_swap_account
&& swapout
&& memcg
)
4360 swap_cgroup_record(ent
, mem_cgroup_id(memcg
));
4364 #ifdef CONFIG_MEMCG_SWAP
4366 * called from swap_entry_free(). remove record in swap_cgroup and
4367 * uncharge "memsw" account.
4369 void mem_cgroup_uncharge_swap(swp_entry_t ent
)
4371 struct mem_cgroup
*memcg
;
4374 if (!do_swap_account
)
4377 id
= swap_cgroup_record(ent
, 0);
4379 memcg
= mem_cgroup_lookup(id
);
4382 * We uncharge this because swap is freed.
4383 * This memcg can be obsolete one. We avoid calling css_tryget
4385 if (!mem_cgroup_is_root(memcg
))
4386 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
4387 mem_cgroup_swap_statistics(memcg
, false);
4388 css_put(&memcg
->css
);
4394 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4395 * @entry: swap entry to be moved
4396 * @from: mem_cgroup which the entry is moved from
4397 * @to: mem_cgroup which the entry is moved to
4399 * It succeeds only when the swap_cgroup's record for this entry is the same
4400 * as the mem_cgroup's id of @from.
4402 * Returns 0 on success, -EINVAL on failure.
4404 * The caller must have charged to @to, IOW, called res_counter_charge() about
4405 * both res and memsw, and called css_get().
4407 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
4408 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4410 unsigned short old_id
, new_id
;
4412 old_id
= mem_cgroup_id(from
);
4413 new_id
= mem_cgroup_id(to
);
4415 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
4416 mem_cgroup_swap_statistics(from
, false);
4417 mem_cgroup_swap_statistics(to
, true);
4419 * This function is only called from task migration context now.
4420 * It postpones res_counter and refcount handling till the end
4421 * of task migration(mem_cgroup_clear_mc()) for performance
4422 * improvement. But we cannot postpone css_get(to) because if
4423 * the process that has been moved to @to does swap-in, the
4424 * refcount of @to might be decreased to 0.
4426 * We are in attach() phase, so the cgroup is guaranteed to be
4427 * alive, so we can just call css_get().
4435 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
4436 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4443 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4446 void mem_cgroup_prepare_migration(struct page
*page
, struct page
*newpage
,
4447 struct mem_cgroup
**memcgp
)
4449 struct mem_cgroup
*memcg
= NULL
;
4450 unsigned int nr_pages
= 1;
4451 struct page_cgroup
*pc
;
4452 enum charge_type ctype
;
4456 if (mem_cgroup_disabled())
4459 if (PageTransHuge(page
))
4460 nr_pages
<<= compound_order(page
);
4462 pc
= lookup_page_cgroup(page
);
4463 lock_page_cgroup(pc
);
4464 if (PageCgroupUsed(pc
)) {
4465 memcg
= pc
->mem_cgroup
;
4466 css_get(&memcg
->css
);
4468 * At migrating an anonymous page, its mapcount goes down
4469 * to 0 and uncharge() will be called. But, even if it's fully
4470 * unmapped, migration may fail and this page has to be
4471 * charged again. We set MIGRATION flag here and delay uncharge
4472 * until end_migration() is called
4474 * Corner Case Thinking
4476 * When the old page was mapped as Anon and it's unmap-and-freed
4477 * while migration was ongoing.
4478 * If unmap finds the old page, uncharge() of it will be delayed
4479 * until end_migration(). If unmap finds a new page, it's
4480 * uncharged when it make mapcount to be 1->0. If unmap code
4481 * finds swap_migration_entry, the new page will not be mapped
4482 * and end_migration() will find it(mapcount==0).
4485 * When the old page was mapped but migraion fails, the kernel
4486 * remaps it. A charge for it is kept by MIGRATION flag even
4487 * if mapcount goes down to 0. We can do remap successfully
4488 * without charging it again.
4491 * The "old" page is under lock_page() until the end of
4492 * migration, so, the old page itself will not be swapped-out.
4493 * If the new page is swapped out before end_migraton, our
4494 * hook to usual swap-out path will catch the event.
4497 SetPageCgroupMigration(pc
);
4499 unlock_page_cgroup(pc
);
4501 * If the page is not charged at this point,
4509 * We charge new page before it's used/mapped. So, even if unlock_page()
4510 * is called before end_migration, we can catch all events on this new
4511 * page. In the case new page is migrated but not remapped, new page's
4512 * mapcount will be finally 0 and we call uncharge in end_migration().
4515 ctype
= MEM_CGROUP_CHARGE_TYPE_ANON
;
4517 ctype
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4519 * The page is committed to the memcg, but it's not actually
4520 * charged to the res_counter since we plan on replacing the
4521 * old one and only one page is going to be left afterwards.
4523 __mem_cgroup_commit_charge(memcg
, newpage
, nr_pages
, ctype
, false);
4526 /* remove redundant charge if migration failed*/
4527 void mem_cgroup_end_migration(struct mem_cgroup
*memcg
,
4528 struct page
*oldpage
, struct page
*newpage
, bool migration_ok
)
4530 struct page
*used
, *unused
;
4531 struct page_cgroup
*pc
;
4537 if (!migration_ok
) {
4544 anon
= PageAnon(used
);
4545 __mem_cgroup_uncharge_common(unused
,
4546 anon
? MEM_CGROUP_CHARGE_TYPE_ANON
4547 : MEM_CGROUP_CHARGE_TYPE_CACHE
,
4549 css_put(&memcg
->css
);
4551 * We disallowed uncharge of pages under migration because mapcount
4552 * of the page goes down to zero, temporarly.
4553 * Clear the flag and check the page should be charged.
4555 pc
= lookup_page_cgroup(oldpage
);
4556 lock_page_cgroup(pc
);
4557 ClearPageCgroupMigration(pc
);
4558 unlock_page_cgroup(pc
);
4561 * If a page is a file cache, radix-tree replacement is very atomic
4562 * and we can skip this check. When it was an Anon page, its mapcount
4563 * goes down to 0. But because we added MIGRATION flage, it's not
4564 * uncharged yet. There are several case but page->mapcount check
4565 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4566 * check. (see prepare_charge() also)
4569 mem_cgroup_uncharge_page(used
);
4573 * At replace page cache, newpage is not under any memcg but it's on
4574 * LRU. So, this function doesn't touch res_counter but handles LRU
4575 * in correct way. Both pages are locked so we cannot race with uncharge.
4577 void mem_cgroup_replace_page_cache(struct page
*oldpage
,
4578 struct page
*newpage
)
4580 struct mem_cgroup
*memcg
= NULL
;
4581 struct page_cgroup
*pc
;
4582 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4584 if (mem_cgroup_disabled())
4587 pc
= lookup_page_cgroup(oldpage
);
4588 /* fix accounting on old pages */
4589 lock_page_cgroup(pc
);
4590 if (PageCgroupUsed(pc
)) {
4591 memcg
= pc
->mem_cgroup
;
4592 mem_cgroup_charge_statistics(memcg
, oldpage
, false, -1);
4593 ClearPageCgroupUsed(pc
);
4595 unlock_page_cgroup(pc
);
4598 * When called from shmem_replace_page(), in some cases the
4599 * oldpage has already been charged, and in some cases not.
4604 * Even if newpage->mapping was NULL before starting replacement,
4605 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4606 * LRU while we overwrite pc->mem_cgroup.
4608 __mem_cgroup_commit_charge(memcg
, newpage
, 1, type
, true);
4611 #ifdef CONFIG_DEBUG_VM
4612 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
4614 struct page_cgroup
*pc
;
4616 pc
= lookup_page_cgroup(page
);
4618 * Can be NULL while feeding pages into the page allocator for
4619 * the first time, i.e. during boot or memory hotplug;
4620 * or when mem_cgroup_disabled().
4622 if (likely(pc
) && PageCgroupUsed(pc
))
4627 bool mem_cgroup_bad_page_check(struct page
*page
)
4629 if (mem_cgroup_disabled())
4632 return lookup_page_cgroup_used(page
) != NULL
;
4635 void mem_cgroup_print_bad_page(struct page
*page
)
4637 struct page_cgroup
*pc
;
4639 pc
= lookup_page_cgroup_used(page
);
4641 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4642 pc
, pc
->flags
, pc
->mem_cgroup
);
4647 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
4648 unsigned long long val
)
4651 u64 memswlimit
, memlimit
;
4653 int children
= mem_cgroup_count_children(memcg
);
4654 u64 curusage
, oldusage
;
4658 * For keeping hierarchical_reclaim simple, how long we should retry
4659 * is depends on callers. We set our retry-count to be function
4660 * of # of children which we should visit in this loop.
4662 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
4664 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4667 while (retry_count
) {
4668 if (signal_pending(current
)) {
4673 * Rather than hide all in some function, I do this in
4674 * open coded manner. You see what this really does.
4675 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4677 mutex_lock(&set_limit_mutex
);
4678 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4679 if (memswlimit
< val
) {
4681 mutex_unlock(&set_limit_mutex
);
4685 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4689 ret
= res_counter_set_limit(&memcg
->res
, val
);
4691 if (memswlimit
== val
)
4692 memcg
->memsw_is_minimum
= true;
4694 memcg
->memsw_is_minimum
= false;
4696 mutex_unlock(&set_limit_mutex
);
4701 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4702 MEM_CGROUP_RECLAIM_SHRINK
);
4703 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4704 /* Usage is reduced ? */
4705 if (curusage
>= oldusage
)
4708 oldusage
= curusage
;
4710 if (!ret
&& enlarge
)
4711 memcg_oom_recover(memcg
);
4716 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
4717 unsigned long long val
)
4720 u64 memlimit
, memswlimit
, oldusage
, curusage
;
4721 int children
= mem_cgroup_count_children(memcg
);
4725 /* see mem_cgroup_resize_res_limit */
4726 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
4727 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4728 while (retry_count
) {
4729 if (signal_pending(current
)) {
4734 * Rather than hide all in some function, I do this in
4735 * open coded manner. You see what this really does.
4736 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4738 mutex_lock(&set_limit_mutex
);
4739 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4740 if (memlimit
> val
) {
4742 mutex_unlock(&set_limit_mutex
);
4745 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4746 if (memswlimit
< val
)
4748 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
4750 if (memlimit
== val
)
4751 memcg
->memsw_is_minimum
= true;
4753 memcg
->memsw_is_minimum
= false;
4755 mutex_unlock(&set_limit_mutex
);
4760 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4761 MEM_CGROUP_RECLAIM_NOSWAP
|
4762 MEM_CGROUP_RECLAIM_SHRINK
);
4763 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4764 /* Usage is reduced ? */
4765 if (curusage
>= oldusage
)
4768 oldusage
= curusage
;
4770 if (!ret
&& enlarge
)
4771 memcg_oom_recover(memcg
);
4775 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
4777 unsigned long *total_scanned
)
4779 unsigned long nr_reclaimed
= 0;
4780 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
4781 unsigned long reclaimed
;
4783 struct mem_cgroup_tree_per_zone
*mctz
;
4784 unsigned long long excess
;
4785 unsigned long nr_scanned
;
4790 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
4792 * This loop can run a while, specially if mem_cgroup's continuously
4793 * keep exceeding their soft limit and putting the system under
4800 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
4805 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
4806 gfp_mask
, &nr_scanned
);
4807 nr_reclaimed
+= reclaimed
;
4808 *total_scanned
+= nr_scanned
;
4809 spin_lock(&mctz
->lock
);
4812 * If we failed to reclaim anything from this memory cgroup
4813 * it is time to move on to the next cgroup
4819 * Loop until we find yet another one.
4821 * By the time we get the soft_limit lock
4822 * again, someone might have aded the
4823 * group back on the RB tree. Iterate to
4824 * make sure we get a different mem.
4825 * mem_cgroup_largest_soft_limit_node returns
4826 * NULL if no other cgroup is present on
4830 __mem_cgroup_largest_soft_limit_node(mctz
);
4832 css_put(&next_mz
->memcg
->css
);
4833 else /* next_mz == NULL or other memcg */
4837 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
4838 excess
= res_counter_soft_limit_excess(&mz
->memcg
->res
);
4840 * One school of thought says that we should not add
4841 * back the node to the tree if reclaim returns 0.
4842 * But our reclaim could return 0, simply because due
4843 * to priority we are exposing a smaller subset of
4844 * memory to reclaim from. Consider this as a longer
4847 /* If excess == 0, no tree ops */
4848 __mem_cgroup_insert_exceeded(mz
->memcg
, mz
, mctz
, excess
);
4849 spin_unlock(&mctz
->lock
);
4850 css_put(&mz
->memcg
->css
);
4853 * Could not reclaim anything and there are no more
4854 * mem cgroups to try or we seem to be looping without
4855 * reclaiming anything.
4857 if (!nr_reclaimed
&&
4859 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
4861 } while (!nr_reclaimed
);
4863 css_put(&next_mz
->memcg
->css
);
4864 return nr_reclaimed
;
4868 * mem_cgroup_force_empty_list - clears LRU of a group
4869 * @memcg: group to clear
4872 * @lru: lru to to clear
4874 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4875 * reclaim the pages page themselves - pages are moved to the parent (or root)
4878 static void mem_cgroup_force_empty_list(struct mem_cgroup
*memcg
,
4879 int node
, int zid
, enum lru_list lru
)
4881 struct lruvec
*lruvec
;
4882 unsigned long flags
;
4883 struct list_head
*list
;
4887 zone
= &NODE_DATA(node
)->node_zones
[zid
];
4888 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
4889 list
= &lruvec
->lists
[lru
];
4893 struct page_cgroup
*pc
;
4896 spin_lock_irqsave(&zone
->lru_lock
, flags
);
4897 if (list_empty(list
)) {
4898 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4901 page
= list_entry(list
->prev
, struct page
, lru
);
4903 list_move(&page
->lru
, list
);
4905 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4908 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4910 pc
= lookup_page_cgroup(page
);
4912 if (mem_cgroup_move_parent(page
, pc
, memcg
)) {
4913 /* found lock contention or "pc" is obsolete. */
4918 } while (!list_empty(list
));
4922 * make mem_cgroup's charge to be 0 if there is no task by moving
4923 * all the charges and pages to the parent.
4924 * This enables deleting this mem_cgroup.
4926 * Caller is responsible for holding css reference on the memcg.
4928 static void mem_cgroup_reparent_charges(struct mem_cgroup
*memcg
)
4934 /* This is for making all *used* pages to be on LRU. */
4935 lru_add_drain_all();
4936 drain_all_stock_sync(memcg
);
4937 mem_cgroup_start_move(memcg
);
4938 for_each_node_state(node
, N_MEMORY
) {
4939 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
4942 mem_cgroup_force_empty_list(memcg
,
4947 mem_cgroup_end_move(memcg
);
4948 memcg_oom_recover(memcg
);
4952 * Kernel memory may not necessarily be trackable to a specific
4953 * process. So they are not migrated, and therefore we can't
4954 * expect their value to drop to 0 here.
4955 * Having res filled up with kmem only is enough.
4957 * This is a safety check because mem_cgroup_force_empty_list
4958 * could have raced with mem_cgroup_replace_page_cache callers
4959 * so the lru seemed empty but the page could have been added
4960 * right after the check. RES_USAGE should be safe as we always
4961 * charge before adding to the LRU.
4963 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
) -
4964 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
);
4965 } while (usage
> 0);
4968 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
4970 lockdep_assert_held(&memcg_create_mutex
);
4972 * The lock does not prevent addition or deletion to the list
4973 * of children, but it prevents a new child from being
4974 * initialized based on this parent in css_online(), so it's
4975 * enough to decide whether hierarchically inherited
4976 * attributes can still be changed or not.
4978 return memcg
->use_hierarchy
&&
4979 !list_empty(&memcg
->css
.cgroup
->children
);
4983 * Reclaims as many pages from the given memcg as possible and moves
4984 * the rest to the parent.
4986 * Caller is responsible for holding css reference for memcg.
4988 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
4990 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
4991 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
4993 /* returns EBUSY if there is a task or if we come here twice. */
4994 if (cgroup_task_count(cgrp
) || !list_empty(&cgrp
->children
))
4997 /* we call try-to-free pages for make this cgroup empty */
4998 lru_add_drain_all();
4999 /* try to free all pages in this cgroup */
5000 while (nr_retries
&& res_counter_read_u64(&memcg
->res
, RES_USAGE
) > 0) {
5003 if (signal_pending(current
))
5006 progress
= try_to_free_mem_cgroup_pages(memcg
, GFP_KERNEL
,
5010 /* maybe some writeback is necessary */
5011 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
5016 mem_cgroup_reparent_charges(memcg
);
5021 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state
*css
,
5024 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5026 if (mem_cgroup_is_root(memcg
))
5028 return mem_cgroup_force_empty(memcg
);
5031 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
5034 return mem_cgroup_from_css(css
)->use_hierarchy
;
5037 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
5038 struct cftype
*cft
, u64 val
)
5041 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5042 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5044 mutex_lock(&memcg_create_mutex
);
5046 if (memcg
->use_hierarchy
== val
)
5050 * If parent's use_hierarchy is set, we can't make any modifications
5051 * in the child subtrees. If it is unset, then the change can
5052 * occur, provided the current cgroup has no children.
5054 * For the root cgroup, parent_mem is NULL, we allow value to be
5055 * set if there are no children.
5057 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
5058 (val
== 1 || val
== 0)) {
5059 if (list_empty(&memcg
->css
.cgroup
->children
))
5060 memcg
->use_hierarchy
= val
;
5067 mutex_unlock(&memcg_create_mutex
);
5073 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup
*memcg
,
5074 enum mem_cgroup_stat_index idx
)
5076 struct mem_cgroup
*iter
;
5079 /* Per-cpu values can be negative, use a signed accumulator */
5080 for_each_mem_cgroup_tree(iter
, memcg
)
5081 val
+= mem_cgroup_read_stat(iter
, idx
);
5083 if (val
< 0) /* race ? */
5088 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
5092 if (!mem_cgroup_is_root(memcg
)) {
5094 return res_counter_read_u64(&memcg
->res
, RES_USAGE
);
5096 return res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
5100 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5101 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5103 val
= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
5104 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_RSS
);
5107 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
5109 return val
<< PAGE_SHIFT
;
5112 static ssize_t
mem_cgroup_read(struct cgroup_subsys_state
*css
,
5113 struct cftype
*cft
, struct file
*file
,
5114 char __user
*buf
, size_t nbytes
, loff_t
*ppos
)
5116 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5122 type
= MEMFILE_TYPE(cft
->private);
5123 name
= MEMFILE_ATTR(cft
->private);
5127 if (name
== RES_USAGE
)
5128 val
= mem_cgroup_usage(memcg
, false);
5130 val
= res_counter_read_u64(&memcg
->res
, name
);
5133 if (name
== RES_USAGE
)
5134 val
= mem_cgroup_usage(memcg
, true);
5136 val
= res_counter_read_u64(&memcg
->memsw
, name
);
5139 val
= res_counter_read_u64(&memcg
->kmem
, name
);
5145 len
= scnprintf(str
, sizeof(str
), "%llu\n", (unsigned long long)val
);
5146 return simple_read_from_buffer(buf
, nbytes
, ppos
, str
, len
);
5149 static int memcg_update_kmem_limit(struct cgroup_subsys_state
*css
, u64 val
)
5152 #ifdef CONFIG_MEMCG_KMEM
5153 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5155 * For simplicity, we won't allow this to be disabled. It also can't
5156 * be changed if the cgroup has children already, or if tasks had
5159 * If tasks join before we set the limit, a person looking at
5160 * kmem.usage_in_bytes will have no way to determine when it took
5161 * place, which makes the value quite meaningless.
5163 * After it first became limited, changes in the value of the limit are
5164 * of course permitted.
5166 mutex_lock(&memcg_create_mutex
);
5167 mutex_lock(&set_limit_mutex
);
5168 if (!memcg
->kmem_account_flags
&& val
!= RES_COUNTER_MAX
) {
5169 if (cgroup_task_count(css
->cgroup
) || memcg_has_children(memcg
)) {
5173 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
5176 ret
= memcg_update_cache_sizes(memcg
);
5178 res_counter_set_limit(&memcg
->kmem
, RES_COUNTER_MAX
);
5181 static_key_slow_inc(&memcg_kmem_enabled_key
);
5183 * setting the active bit after the inc will guarantee no one
5184 * starts accounting before all call sites are patched
5186 memcg_kmem_set_active(memcg
);
5188 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
5190 mutex_unlock(&set_limit_mutex
);
5191 mutex_unlock(&memcg_create_mutex
);
5196 #ifdef CONFIG_MEMCG_KMEM
5197 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
5200 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
5204 memcg
->kmem_account_flags
= parent
->kmem_account_flags
;
5206 * When that happen, we need to disable the static branch only on those
5207 * memcgs that enabled it. To achieve this, we would be forced to
5208 * complicate the code by keeping track of which memcgs were the ones
5209 * that actually enabled limits, and which ones got it from its
5212 * It is a lot simpler just to do static_key_slow_inc() on every child
5213 * that is accounted.
5215 if (!memcg_kmem_is_active(memcg
))
5219 * __mem_cgroup_free() will issue static_key_slow_dec() because this
5220 * memcg is active already. If the later initialization fails then the
5221 * cgroup core triggers the cleanup so we do not have to do it here.
5223 static_key_slow_inc(&memcg_kmem_enabled_key
);
5225 mutex_lock(&set_limit_mutex
);
5226 memcg_stop_kmem_account();
5227 ret
= memcg_update_cache_sizes(memcg
);
5228 memcg_resume_kmem_account();
5229 mutex_unlock(&set_limit_mutex
);
5233 #endif /* CONFIG_MEMCG_KMEM */
5236 * The user of this function is...
5239 static int mem_cgroup_write(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
5242 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5245 unsigned long long val
;
5248 type
= MEMFILE_TYPE(cft
->private);
5249 name
= MEMFILE_ATTR(cft
->private);
5253 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
5257 /* This function does all necessary parse...reuse it */
5258 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5262 ret
= mem_cgroup_resize_limit(memcg
, val
);
5263 else if (type
== _MEMSWAP
)
5264 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
5265 else if (type
== _KMEM
)
5266 ret
= memcg_update_kmem_limit(css
, val
);
5270 case RES_SOFT_LIMIT
:
5271 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5275 * For memsw, soft limits are hard to implement in terms
5276 * of semantics, for now, we support soft limits for
5277 * control without swap
5280 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
5285 ret
= -EINVAL
; /* should be BUG() ? */
5291 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
5292 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
5294 unsigned long long min_limit
, min_memsw_limit
, tmp
;
5296 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5297 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5298 if (!memcg
->use_hierarchy
)
5301 while (css_parent(&memcg
->css
)) {
5302 memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5303 if (!memcg
->use_hierarchy
)
5305 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5306 min_limit
= min(min_limit
, tmp
);
5307 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5308 min_memsw_limit
= min(min_memsw_limit
, tmp
);
5311 *mem_limit
= min_limit
;
5312 *memsw_limit
= min_memsw_limit
;
5315 static int mem_cgroup_reset(struct cgroup_subsys_state
*css
, unsigned int event
)
5317 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5321 type
= MEMFILE_TYPE(event
);
5322 name
= MEMFILE_ATTR(event
);
5327 res_counter_reset_max(&memcg
->res
);
5328 else if (type
== _MEMSWAP
)
5329 res_counter_reset_max(&memcg
->memsw
);
5330 else if (type
== _KMEM
)
5331 res_counter_reset_max(&memcg
->kmem
);
5337 res_counter_reset_failcnt(&memcg
->res
);
5338 else if (type
== _MEMSWAP
)
5339 res_counter_reset_failcnt(&memcg
->memsw
);
5340 else if (type
== _KMEM
)
5341 res_counter_reset_failcnt(&memcg
->kmem
);
5350 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
5353 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
5357 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5358 struct cftype
*cft
, u64 val
)
5360 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5362 if (val
>= (1 << NR_MOVE_TYPE
))
5366 * No kind of locking is needed in here, because ->can_attach() will
5367 * check this value once in the beginning of the process, and then carry
5368 * on with stale data. This means that changes to this value will only
5369 * affect task migrations starting after the change.
5371 memcg
->move_charge_at_immigrate
= val
;
5375 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5376 struct cftype
*cft
, u64 val
)
5383 static int memcg_numa_stat_show(struct cgroup_subsys_state
*css
,
5384 struct cftype
*cft
, struct seq_file
*m
)
5388 unsigned int lru_mask
;
5391 static const struct numa_stat stats
[] = {
5392 { "total", LRU_ALL
},
5393 { "file", LRU_ALL_FILE
},
5394 { "anon", LRU_ALL_ANON
},
5395 { "unevictable", BIT(LRU_UNEVICTABLE
) },
5397 const struct numa_stat
*stat
;
5400 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5402 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
5403 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
5404 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
5405 for_each_node_state(nid
, N_MEMORY
) {
5406 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5408 seq_printf(m
, " N%d=%lu", nid
, nr
);
5413 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
5414 struct mem_cgroup
*iter
;
5417 for_each_mem_cgroup_tree(iter
, memcg
)
5418 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
5419 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
5420 for_each_node_state(nid
, N_MEMORY
) {
5422 for_each_mem_cgroup_tree(iter
, memcg
)
5423 nr
+= mem_cgroup_node_nr_lru_pages(
5424 iter
, nid
, stat
->lru_mask
);
5425 seq_printf(m
, " N%d=%lu", nid
, nr
);
5432 #endif /* CONFIG_NUMA */
5434 static inline void mem_cgroup_lru_names_not_uptodate(void)
5436 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
5439 static int memcg_stat_show(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
5442 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5443 struct mem_cgroup
*mi
;
5446 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5447 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5449 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
5450 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
5453 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
5454 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
5455 mem_cgroup_read_events(memcg
, i
));
5457 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
5458 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
5459 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
5461 /* Hierarchical information */
5463 unsigned long long limit
, memsw_limit
;
5464 memcg_get_hierarchical_limit(memcg
, &limit
, &memsw_limit
);
5465 seq_printf(m
, "hierarchical_memory_limit %llu\n", limit
);
5466 if (do_swap_account
)
5467 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
5471 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5474 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5476 for_each_mem_cgroup_tree(mi
, memcg
)
5477 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
5478 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
5481 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
5482 unsigned long long val
= 0;
5484 for_each_mem_cgroup_tree(mi
, memcg
)
5485 val
+= mem_cgroup_read_events(mi
, i
);
5486 seq_printf(m
, "total_%s %llu\n",
5487 mem_cgroup_events_names
[i
], val
);
5490 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5491 unsigned long long val
= 0;
5493 for_each_mem_cgroup_tree(mi
, memcg
)
5494 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
5495 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
5498 #ifdef CONFIG_DEBUG_VM
5501 struct mem_cgroup_per_zone
*mz
;
5502 struct zone_reclaim_stat
*rstat
;
5503 unsigned long recent_rotated
[2] = {0, 0};
5504 unsigned long recent_scanned
[2] = {0, 0};
5506 for_each_online_node(nid
)
5507 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
5508 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
5509 rstat
= &mz
->lruvec
.reclaim_stat
;
5511 recent_rotated
[0] += rstat
->recent_rotated
[0];
5512 recent_rotated
[1] += rstat
->recent_rotated
[1];
5513 recent_scanned
[0] += rstat
->recent_scanned
[0];
5514 recent_scanned
[1] += rstat
->recent_scanned
[1];
5516 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
5517 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
5518 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
5519 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
5526 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
5529 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5531 return mem_cgroup_swappiness(memcg
);
5534 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
5535 struct cftype
*cft
, u64 val
)
5537 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5538 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5540 if (val
> 100 || !parent
)
5543 mutex_lock(&memcg_create_mutex
);
5545 /* If under hierarchy, only empty-root can set this value */
5546 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5547 mutex_unlock(&memcg_create_mutex
);
5551 memcg
->swappiness
= val
;
5553 mutex_unlock(&memcg_create_mutex
);
5558 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
5560 struct mem_cgroup_threshold_ary
*t
;
5566 t
= rcu_dereference(memcg
->thresholds
.primary
);
5568 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
5573 usage
= mem_cgroup_usage(memcg
, swap
);
5576 * current_threshold points to threshold just below or equal to usage.
5577 * If it's not true, a threshold was crossed after last
5578 * call of __mem_cgroup_threshold().
5580 i
= t
->current_threshold
;
5583 * Iterate backward over array of thresholds starting from
5584 * current_threshold and check if a threshold is crossed.
5585 * If none of thresholds below usage is crossed, we read
5586 * only one element of the array here.
5588 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
5589 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5591 /* i = current_threshold + 1 */
5595 * Iterate forward over array of thresholds starting from
5596 * current_threshold+1 and check if a threshold is crossed.
5597 * If none of thresholds above usage is crossed, we read
5598 * only one element of the array here.
5600 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
5601 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5603 /* Update current_threshold */
5604 t
->current_threshold
= i
- 1;
5609 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
5612 __mem_cgroup_threshold(memcg
, false);
5613 if (do_swap_account
)
5614 __mem_cgroup_threshold(memcg
, true);
5616 memcg
= parent_mem_cgroup(memcg
);
5620 static int compare_thresholds(const void *a
, const void *b
)
5622 const struct mem_cgroup_threshold
*_a
= a
;
5623 const struct mem_cgroup_threshold
*_b
= b
;
5625 if (_a
->threshold
> _b
->threshold
)
5628 if (_a
->threshold
< _b
->threshold
)
5634 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
5636 struct mem_cgroup_eventfd_list
*ev
;
5638 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
5639 eventfd_signal(ev
->eventfd
, 1);
5643 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
5645 struct mem_cgroup
*iter
;
5647 for_each_mem_cgroup_tree(iter
, memcg
)
5648 mem_cgroup_oom_notify_cb(iter
);
5651 static int mem_cgroup_usage_register_event(struct cgroup_subsys_state
*css
,
5652 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5654 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5655 struct mem_cgroup_thresholds
*thresholds
;
5656 struct mem_cgroup_threshold_ary
*new;
5657 enum res_type type
= MEMFILE_TYPE(cft
->private);
5658 u64 threshold
, usage
;
5661 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
5665 mutex_lock(&memcg
->thresholds_lock
);
5668 thresholds
= &memcg
->thresholds
;
5669 else if (type
== _MEMSWAP
)
5670 thresholds
= &memcg
->memsw_thresholds
;
5674 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5676 /* Check if a threshold crossed before adding a new one */
5677 if (thresholds
->primary
)
5678 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5680 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
5682 /* Allocate memory for new array of thresholds */
5683 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
5691 /* Copy thresholds (if any) to new array */
5692 if (thresholds
->primary
) {
5693 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
5694 sizeof(struct mem_cgroup_threshold
));
5697 /* Add new threshold */
5698 new->entries
[size
- 1].eventfd
= eventfd
;
5699 new->entries
[size
- 1].threshold
= threshold
;
5701 /* Sort thresholds. Registering of new threshold isn't time-critical */
5702 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
5703 compare_thresholds
, NULL
);
5705 /* Find current threshold */
5706 new->current_threshold
= -1;
5707 for (i
= 0; i
< size
; i
++) {
5708 if (new->entries
[i
].threshold
<= usage
) {
5710 * new->current_threshold will not be used until
5711 * rcu_assign_pointer(), so it's safe to increment
5714 ++new->current_threshold
;
5719 /* Free old spare buffer and save old primary buffer as spare */
5720 kfree(thresholds
->spare
);
5721 thresholds
->spare
= thresholds
->primary
;
5723 rcu_assign_pointer(thresholds
->primary
, new);
5725 /* To be sure that nobody uses thresholds */
5729 mutex_unlock(&memcg
->thresholds_lock
);
5734 static void mem_cgroup_usage_unregister_event(struct cgroup_subsys_state
*css
,
5735 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5737 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5738 struct mem_cgroup_thresholds
*thresholds
;
5739 struct mem_cgroup_threshold_ary
*new;
5740 enum res_type type
= MEMFILE_TYPE(cft
->private);
5744 mutex_lock(&memcg
->thresholds_lock
);
5746 thresholds
= &memcg
->thresholds
;
5747 else if (type
== _MEMSWAP
)
5748 thresholds
= &memcg
->memsw_thresholds
;
5752 if (!thresholds
->primary
)
5755 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5757 /* Check if a threshold crossed before removing */
5758 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5760 /* Calculate new number of threshold */
5762 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
5763 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
5767 new = thresholds
->spare
;
5769 /* Set thresholds array to NULL if we don't have thresholds */
5778 /* Copy thresholds and find current threshold */
5779 new->current_threshold
= -1;
5780 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
5781 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
5784 new->entries
[j
] = thresholds
->primary
->entries
[i
];
5785 if (new->entries
[j
].threshold
<= usage
) {
5787 * new->current_threshold will not be used
5788 * until rcu_assign_pointer(), so it's safe to increment
5791 ++new->current_threshold
;
5797 /* Swap primary and spare array */
5798 thresholds
->spare
= thresholds
->primary
;
5799 /* If all events are unregistered, free the spare array */
5801 kfree(thresholds
->spare
);
5802 thresholds
->spare
= NULL
;
5805 rcu_assign_pointer(thresholds
->primary
, new);
5807 /* To be sure that nobody uses thresholds */
5810 mutex_unlock(&memcg
->thresholds_lock
);
5813 static int mem_cgroup_oom_register_event(struct cgroup_subsys_state
*css
,
5814 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5816 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5817 struct mem_cgroup_eventfd_list
*event
;
5818 enum res_type type
= MEMFILE_TYPE(cft
->private);
5820 BUG_ON(type
!= _OOM_TYPE
);
5821 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
5825 spin_lock(&memcg_oom_lock
);
5827 event
->eventfd
= eventfd
;
5828 list_add(&event
->list
, &memcg
->oom_notify
);
5830 /* already in OOM ? */
5831 if (atomic_read(&memcg
->under_oom
))
5832 eventfd_signal(eventfd
, 1);
5833 spin_unlock(&memcg_oom_lock
);
5838 static void mem_cgroup_oom_unregister_event(struct cgroup_subsys_state
*css
,
5839 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5841 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5842 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
5843 enum res_type type
= MEMFILE_TYPE(cft
->private);
5845 BUG_ON(type
!= _OOM_TYPE
);
5847 spin_lock(&memcg_oom_lock
);
5849 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
5850 if (ev
->eventfd
== eventfd
) {
5851 list_del(&ev
->list
);
5856 spin_unlock(&memcg_oom_lock
);
5859 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state
*css
,
5860 struct cftype
*cft
, struct cgroup_map_cb
*cb
)
5862 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5864 cb
->fill(cb
, "oom_kill_disable", memcg
->oom_kill_disable
);
5866 if (atomic_read(&memcg
->under_oom
))
5867 cb
->fill(cb
, "under_oom", 1);
5869 cb
->fill(cb
, "under_oom", 0);
5873 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
5874 struct cftype
*cft
, u64 val
)
5876 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5877 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5879 /* cannot set to root cgroup and only 0 and 1 are allowed */
5880 if (!parent
|| !((val
== 0) || (val
== 1)))
5883 mutex_lock(&memcg_create_mutex
);
5884 /* oom-kill-disable is a flag for subhierarchy. */
5885 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5886 mutex_unlock(&memcg_create_mutex
);
5889 memcg
->oom_kill_disable
= val
;
5891 memcg_oom_recover(memcg
);
5892 mutex_unlock(&memcg_create_mutex
);
5896 #ifdef CONFIG_MEMCG_KMEM
5897 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5901 memcg
->kmemcg_id
= -1;
5902 ret
= memcg_propagate_kmem(memcg
);
5906 return mem_cgroup_sockets_init(memcg
, ss
);
5909 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5911 mem_cgroup_sockets_destroy(memcg
);
5914 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5916 if (!memcg_kmem_is_active(memcg
))
5920 * kmem charges can outlive the cgroup. In the case of slab
5921 * pages, for instance, a page contain objects from various
5922 * processes. As we prevent from taking a reference for every
5923 * such allocation we have to be careful when doing uncharge
5924 * (see memcg_uncharge_kmem) and here during offlining.
5926 * The idea is that that only the _last_ uncharge which sees
5927 * the dead memcg will drop the last reference. An additional
5928 * reference is taken here before the group is marked dead
5929 * which is then paired with css_put during uncharge resp. here.
5931 * Although this might sound strange as this path is called from
5932 * css_offline() when the referencemight have dropped down to 0
5933 * and shouldn't be incremented anymore (css_tryget would fail)
5934 * we do not have other options because of the kmem allocations
5937 css_get(&memcg
->css
);
5939 memcg_kmem_mark_dead(memcg
);
5941 if (res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0)
5944 if (memcg_kmem_test_and_clear_dead(memcg
))
5945 css_put(&memcg
->css
);
5948 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5953 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5957 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5962 static struct cftype mem_cgroup_files
[] = {
5964 .name
= "usage_in_bytes",
5965 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
5966 .read
= mem_cgroup_read
,
5967 .register_event
= mem_cgroup_usage_register_event
,
5968 .unregister_event
= mem_cgroup_usage_unregister_event
,
5971 .name
= "max_usage_in_bytes",
5972 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
5973 .trigger
= mem_cgroup_reset
,
5974 .read
= mem_cgroup_read
,
5977 .name
= "limit_in_bytes",
5978 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
5979 .write_string
= mem_cgroup_write
,
5980 .read
= mem_cgroup_read
,
5983 .name
= "soft_limit_in_bytes",
5984 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
5985 .write_string
= mem_cgroup_write
,
5986 .read
= mem_cgroup_read
,
5990 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
5991 .trigger
= mem_cgroup_reset
,
5992 .read
= mem_cgroup_read
,
5996 .read_seq_string
= memcg_stat_show
,
5999 .name
= "force_empty",
6000 .trigger
= mem_cgroup_force_empty_write
,
6003 .name
= "use_hierarchy",
6004 .flags
= CFTYPE_INSANE
,
6005 .write_u64
= mem_cgroup_hierarchy_write
,
6006 .read_u64
= mem_cgroup_hierarchy_read
,
6009 .name
= "swappiness",
6010 .read_u64
= mem_cgroup_swappiness_read
,
6011 .write_u64
= mem_cgroup_swappiness_write
,
6014 .name
= "move_charge_at_immigrate",
6015 .read_u64
= mem_cgroup_move_charge_read
,
6016 .write_u64
= mem_cgroup_move_charge_write
,
6019 .name
= "oom_control",
6020 .read_map
= mem_cgroup_oom_control_read
,
6021 .write_u64
= mem_cgroup_oom_control_write
,
6022 .register_event
= mem_cgroup_oom_register_event
,
6023 .unregister_event
= mem_cgroup_oom_unregister_event
,
6024 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
6027 .name
= "pressure_level",
6028 .register_event
= vmpressure_register_event
,
6029 .unregister_event
= vmpressure_unregister_event
,
6033 .name
= "numa_stat",
6034 .read_seq_string
= memcg_numa_stat_show
,
6037 #ifdef CONFIG_MEMCG_KMEM
6039 .name
= "kmem.limit_in_bytes",
6040 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
6041 .write_string
= mem_cgroup_write
,
6042 .read
= mem_cgroup_read
,
6045 .name
= "kmem.usage_in_bytes",
6046 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
6047 .read
= mem_cgroup_read
,
6050 .name
= "kmem.failcnt",
6051 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
6052 .trigger
= mem_cgroup_reset
,
6053 .read
= mem_cgroup_read
,
6056 .name
= "kmem.max_usage_in_bytes",
6057 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
6058 .trigger
= mem_cgroup_reset
,
6059 .read
= mem_cgroup_read
,
6061 #ifdef CONFIG_SLABINFO
6063 .name
= "kmem.slabinfo",
6064 .read_seq_string
= mem_cgroup_slabinfo_read
,
6068 { }, /* terminate */
6071 #ifdef CONFIG_MEMCG_SWAP
6072 static struct cftype memsw_cgroup_files
[] = {
6074 .name
= "memsw.usage_in_bytes",
6075 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
6076 .read
= mem_cgroup_read
,
6077 .register_event
= mem_cgroup_usage_register_event
,
6078 .unregister_event
= mem_cgroup_usage_unregister_event
,
6081 .name
= "memsw.max_usage_in_bytes",
6082 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
6083 .trigger
= mem_cgroup_reset
,
6084 .read
= mem_cgroup_read
,
6087 .name
= "memsw.limit_in_bytes",
6088 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
6089 .write_string
= mem_cgroup_write
,
6090 .read
= mem_cgroup_read
,
6093 .name
= "memsw.failcnt",
6094 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
6095 .trigger
= mem_cgroup_reset
,
6096 .read
= mem_cgroup_read
,
6098 { }, /* terminate */
6101 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6103 struct mem_cgroup_per_node
*pn
;
6104 struct mem_cgroup_per_zone
*mz
;
6105 int zone
, tmp
= node
;
6107 * This routine is called against possible nodes.
6108 * But it's BUG to call kmalloc() against offline node.
6110 * TODO: this routine can waste much memory for nodes which will
6111 * never be onlined. It's better to use memory hotplug callback
6114 if (!node_state(node
, N_NORMAL_MEMORY
))
6116 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
6120 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6121 mz
= &pn
->zoneinfo
[zone
];
6122 lruvec_init(&mz
->lruvec
);
6123 mz
->usage_in_excess
= 0;
6124 mz
->on_tree
= false;
6127 memcg
->nodeinfo
[node
] = pn
;
6131 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6133 kfree(memcg
->nodeinfo
[node
]);
6136 static struct mem_cgroup
*mem_cgroup_alloc(void)
6138 struct mem_cgroup
*memcg
;
6139 size_t size
= memcg_size();
6141 /* Can be very big if nr_node_ids is very big */
6142 if (size
< PAGE_SIZE
)
6143 memcg
= kzalloc(size
, GFP_KERNEL
);
6145 memcg
= vzalloc(size
);
6150 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
6153 spin_lock_init(&memcg
->pcp_counter_lock
);
6157 if (size
< PAGE_SIZE
)
6165 * At destroying mem_cgroup, references from swap_cgroup can remain.
6166 * (scanning all at force_empty is too costly...)
6168 * Instead of clearing all references at force_empty, we remember
6169 * the number of reference from swap_cgroup and free mem_cgroup when
6170 * it goes down to 0.
6172 * Removal of cgroup itself succeeds regardless of refs from swap.
6175 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
6178 size_t size
= memcg_size();
6180 mem_cgroup_remove_from_trees(memcg
);
6183 free_mem_cgroup_per_zone_info(memcg
, node
);
6185 free_percpu(memcg
->stat
);
6188 * We need to make sure that (at least for now), the jump label
6189 * destruction code runs outside of the cgroup lock. This is because
6190 * get_online_cpus(), which is called from the static_branch update,
6191 * can't be called inside the cgroup_lock. cpusets are the ones
6192 * enforcing this dependency, so if they ever change, we might as well.
6194 * schedule_work() will guarantee this happens. Be careful if you need
6195 * to move this code around, and make sure it is outside
6198 disarm_static_keys(memcg
);
6199 if (size
< PAGE_SIZE
)
6206 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6208 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
6210 if (!memcg
->res
.parent
)
6212 return mem_cgroup_from_res_counter(memcg
->res
.parent
, res
);
6214 EXPORT_SYMBOL(parent_mem_cgroup
);
6216 static void __init
mem_cgroup_soft_limit_tree_init(void)
6218 struct mem_cgroup_tree_per_node
*rtpn
;
6219 struct mem_cgroup_tree_per_zone
*rtpz
;
6220 int tmp
, node
, zone
;
6222 for_each_node(node
) {
6224 if (!node_state(node
, N_NORMAL_MEMORY
))
6226 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
6229 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
6231 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6232 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
6233 rtpz
->rb_root
= RB_ROOT
;
6234 spin_lock_init(&rtpz
->lock
);
6239 static struct cgroup_subsys_state
* __ref
6240 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
6242 struct mem_cgroup
*memcg
;
6243 long error
= -ENOMEM
;
6246 memcg
= mem_cgroup_alloc();
6248 return ERR_PTR(error
);
6251 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
6255 if (parent_css
== NULL
) {
6256 root_mem_cgroup
= memcg
;
6257 res_counter_init(&memcg
->res
, NULL
);
6258 res_counter_init(&memcg
->memsw
, NULL
);
6259 res_counter_init(&memcg
->kmem
, NULL
);
6262 memcg
->last_scanned_node
= MAX_NUMNODES
;
6263 INIT_LIST_HEAD(&memcg
->oom_notify
);
6264 memcg
->move_charge_at_immigrate
= 0;
6265 mutex_init(&memcg
->thresholds_lock
);
6266 spin_lock_init(&memcg
->move_lock
);
6267 vmpressure_init(&memcg
->vmpressure
);
6272 __mem_cgroup_free(memcg
);
6273 return ERR_PTR(error
);
6277 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
6279 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6280 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(css
));
6283 if (css
->cgroup
->id
> MEM_CGROUP_ID_MAX
)
6289 mutex_lock(&memcg_create_mutex
);
6291 memcg
->use_hierarchy
= parent
->use_hierarchy
;
6292 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
6293 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
6295 if (parent
->use_hierarchy
) {
6296 res_counter_init(&memcg
->res
, &parent
->res
);
6297 res_counter_init(&memcg
->memsw
, &parent
->memsw
);
6298 res_counter_init(&memcg
->kmem
, &parent
->kmem
);
6301 * No need to take a reference to the parent because cgroup
6302 * core guarantees its existence.
6305 res_counter_init(&memcg
->res
, NULL
);
6306 res_counter_init(&memcg
->memsw
, NULL
);
6307 res_counter_init(&memcg
->kmem
, NULL
);
6309 * Deeper hierachy with use_hierarchy == false doesn't make
6310 * much sense so let cgroup subsystem know about this
6311 * unfortunate state in our controller.
6313 if (parent
!= root_mem_cgroup
)
6314 mem_cgroup_subsys
.broken_hierarchy
= true;
6317 error
= memcg_init_kmem(memcg
, &mem_cgroup_subsys
);
6318 mutex_unlock(&memcg_create_mutex
);
6323 * Announce all parents that a group from their hierarchy is gone.
6325 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup
*memcg
)
6327 struct mem_cgroup
*parent
= memcg
;
6329 while ((parent
= parent_mem_cgroup(parent
)))
6330 mem_cgroup_iter_invalidate(parent
);
6333 * if the root memcg is not hierarchical we have to check it
6336 if (!root_mem_cgroup
->use_hierarchy
)
6337 mem_cgroup_iter_invalidate(root_mem_cgroup
);
6340 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
6342 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6344 kmem_cgroup_css_offline(memcg
);
6346 mem_cgroup_invalidate_reclaim_iterators(memcg
);
6347 mem_cgroup_reparent_charges(memcg
);
6348 mem_cgroup_destroy_all_caches(memcg
);
6349 vmpressure_cleanup(&memcg
->vmpressure
);
6352 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
6354 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6356 memcg_destroy_kmem(memcg
);
6357 __mem_cgroup_free(memcg
);
6361 /* Handlers for move charge at task migration. */
6362 #define PRECHARGE_COUNT_AT_ONCE 256
6363 static int mem_cgroup_do_precharge(unsigned long count
)
6366 int batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6367 struct mem_cgroup
*memcg
= mc
.to
;
6369 if (mem_cgroup_is_root(memcg
)) {
6370 mc
.precharge
+= count
;
6371 /* we don't need css_get for root */
6374 /* try to charge at once */
6376 struct res_counter
*dummy
;
6378 * "memcg" cannot be under rmdir() because we've already checked
6379 * by cgroup_lock_live_cgroup() that it is not removed and we
6380 * are still under the same cgroup_mutex. So we can postpone
6383 if (res_counter_charge(&memcg
->res
, PAGE_SIZE
* count
, &dummy
))
6385 if (do_swap_account
&& res_counter_charge(&memcg
->memsw
,
6386 PAGE_SIZE
* count
, &dummy
)) {
6387 res_counter_uncharge(&memcg
->res
, PAGE_SIZE
* count
);
6390 mc
.precharge
+= count
;
6394 /* fall back to one by one charge */
6396 if (signal_pending(current
)) {
6400 if (!batch_count
--) {
6401 batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6404 ret
= __mem_cgroup_try_charge(NULL
,
6405 GFP_KERNEL
, 1, &memcg
, false);
6407 /* mem_cgroup_clear_mc() will do uncharge later */
6415 * get_mctgt_type - get target type of moving charge
6416 * @vma: the vma the pte to be checked belongs
6417 * @addr: the address corresponding to the pte to be checked
6418 * @ptent: the pte to be checked
6419 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6422 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6423 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6424 * move charge. if @target is not NULL, the page is stored in target->page
6425 * with extra refcnt got(Callers should handle it).
6426 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6427 * target for charge migration. if @target is not NULL, the entry is stored
6430 * Called with pte lock held.
6437 enum mc_target_type
{
6443 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
6444 unsigned long addr
, pte_t ptent
)
6446 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
6448 if (!page
|| !page_mapped(page
))
6450 if (PageAnon(page
)) {
6451 /* we don't move shared anon */
6454 } else if (!move_file())
6455 /* we ignore mapcount for file pages */
6457 if (!get_page_unless_zero(page
))
6464 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6465 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6467 struct page
*page
= NULL
;
6468 swp_entry_t ent
= pte_to_swp_entry(ptent
);
6470 if (!move_anon() || non_swap_entry(ent
))
6473 * Because lookup_swap_cache() updates some statistics counter,
6474 * we call find_get_page() with swapper_space directly.
6476 page
= find_get_page(swap_address_space(ent
), ent
.val
);
6477 if (do_swap_account
)
6478 entry
->val
= ent
.val
;
6483 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6484 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6490 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
6491 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6493 struct page
*page
= NULL
;
6494 struct address_space
*mapping
;
6497 if (!vma
->vm_file
) /* anonymous vma */
6502 mapping
= vma
->vm_file
->f_mapping
;
6503 if (pte_none(ptent
))
6504 pgoff
= linear_page_index(vma
, addr
);
6505 else /* pte_file(ptent) is true */
6506 pgoff
= pte_to_pgoff(ptent
);
6508 /* page is moved even if it's not RSS of this task(page-faulted). */
6509 page
= find_get_page(mapping
, pgoff
);
6512 /* shmem/tmpfs may report page out on swap: account for that too. */
6513 if (radix_tree_exceptional_entry(page
)) {
6514 swp_entry_t swap
= radix_to_swp_entry(page
);
6515 if (do_swap_account
)
6517 page
= find_get_page(swap_address_space(swap
), swap
.val
);
6523 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
6524 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
6526 struct page
*page
= NULL
;
6527 struct page_cgroup
*pc
;
6528 enum mc_target_type ret
= MC_TARGET_NONE
;
6529 swp_entry_t ent
= { .val
= 0 };
6531 if (pte_present(ptent
))
6532 page
= mc_handle_present_pte(vma
, addr
, ptent
);
6533 else if (is_swap_pte(ptent
))
6534 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
6535 else if (pte_none(ptent
) || pte_file(ptent
))
6536 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
6538 if (!page
&& !ent
.val
)
6541 pc
= lookup_page_cgroup(page
);
6543 * Do only loose check w/o page_cgroup lock.
6544 * mem_cgroup_move_account() checks the pc is valid or not under
6547 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6548 ret
= MC_TARGET_PAGE
;
6550 target
->page
= page
;
6552 if (!ret
|| !target
)
6555 /* There is a swap entry and a page doesn't exist or isn't charged */
6556 if (ent
.val
&& !ret
&&
6557 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
6558 ret
= MC_TARGET_SWAP
;
6565 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6567 * We don't consider swapping or file mapped pages because THP does not
6568 * support them for now.
6569 * Caller should make sure that pmd_trans_huge(pmd) is true.
6571 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6572 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6574 struct page
*page
= NULL
;
6575 struct page_cgroup
*pc
;
6576 enum mc_target_type ret
= MC_TARGET_NONE
;
6578 page
= pmd_page(pmd
);
6579 VM_BUG_ON(!page
|| !PageHead(page
));
6582 pc
= lookup_page_cgroup(page
);
6583 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6584 ret
= MC_TARGET_PAGE
;
6587 target
->page
= page
;
6593 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6594 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6596 return MC_TARGET_NONE
;
6600 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
6601 unsigned long addr
, unsigned long end
,
6602 struct mm_walk
*walk
)
6604 struct vm_area_struct
*vma
= walk
->private;
6608 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
6609 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
6610 mc
.precharge
+= HPAGE_PMD_NR
;
6615 if (pmd_trans_unstable(pmd
))
6617 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6618 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
6619 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
6620 mc
.precharge
++; /* increment precharge temporarily */
6621 pte_unmap_unlock(pte
- 1, ptl
);
6627 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
6629 unsigned long precharge
;
6630 struct vm_area_struct
*vma
;
6632 down_read(&mm
->mmap_sem
);
6633 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6634 struct mm_walk mem_cgroup_count_precharge_walk
= {
6635 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
6639 if (is_vm_hugetlb_page(vma
))
6641 walk_page_range(vma
->vm_start
, vma
->vm_end
,
6642 &mem_cgroup_count_precharge_walk
);
6644 up_read(&mm
->mmap_sem
);
6646 precharge
= mc
.precharge
;
6652 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
6654 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
6656 VM_BUG_ON(mc
.moving_task
);
6657 mc
.moving_task
= current
;
6658 return mem_cgroup_do_precharge(precharge
);
6661 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6662 static void __mem_cgroup_clear_mc(void)
6664 struct mem_cgroup
*from
= mc
.from
;
6665 struct mem_cgroup
*to
= mc
.to
;
6668 /* we must uncharge all the leftover precharges from mc.to */
6670 __mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
6674 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6675 * we must uncharge here.
6677 if (mc
.moved_charge
) {
6678 __mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
6679 mc
.moved_charge
= 0;
6681 /* we must fixup refcnts and charges */
6682 if (mc
.moved_swap
) {
6683 /* uncharge swap account from the old cgroup */
6684 if (!mem_cgroup_is_root(mc
.from
))
6685 res_counter_uncharge(&mc
.from
->memsw
,
6686 PAGE_SIZE
* mc
.moved_swap
);
6688 for (i
= 0; i
< mc
.moved_swap
; i
++)
6689 css_put(&mc
.from
->css
);
6691 if (!mem_cgroup_is_root(mc
.to
)) {
6693 * we charged both to->res and to->memsw, so we should
6696 res_counter_uncharge(&mc
.to
->res
,
6697 PAGE_SIZE
* mc
.moved_swap
);
6699 /* we've already done css_get(mc.to) */
6702 memcg_oom_recover(from
);
6703 memcg_oom_recover(to
);
6704 wake_up_all(&mc
.waitq
);
6707 static void mem_cgroup_clear_mc(void)
6709 struct mem_cgroup
*from
= mc
.from
;
6712 * we must clear moving_task before waking up waiters at the end of
6715 mc
.moving_task
= NULL
;
6716 __mem_cgroup_clear_mc();
6717 spin_lock(&mc
.lock
);
6720 spin_unlock(&mc
.lock
);
6721 mem_cgroup_end_move(from
);
6724 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6725 struct cgroup_taskset
*tset
)
6727 struct task_struct
*p
= cgroup_taskset_first(tset
);
6729 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6730 unsigned long move_charge_at_immigrate
;
6733 * We are now commited to this value whatever it is. Changes in this
6734 * tunable will only affect upcoming migrations, not the current one.
6735 * So we need to save it, and keep it going.
6737 move_charge_at_immigrate
= memcg
->move_charge_at_immigrate
;
6738 if (move_charge_at_immigrate
) {
6739 struct mm_struct
*mm
;
6740 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
6742 VM_BUG_ON(from
== memcg
);
6744 mm
= get_task_mm(p
);
6747 /* We move charges only when we move a owner of the mm */
6748 if (mm
->owner
== p
) {
6751 VM_BUG_ON(mc
.precharge
);
6752 VM_BUG_ON(mc
.moved_charge
);
6753 VM_BUG_ON(mc
.moved_swap
);
6754 mem_cgroup_start_move(from
);
6755 spin_lock(&mc
.lock
);
6758 mc
.immigrate_flags
= move_charge_at_immigrate
;
6759 spin_unlock(&mc
.lock
);
6760 /* We set mc.moving_task later */
6762 ret
= mem_cgroup_precharge_mc(mm
);
6764 mem_cgroup_clear_mc();
6771 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6772 struct cgroup_taskset
*tset
)
6774 mem_cgroup_clear_mc();
6777 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6778 unsigned long addr
, unsigned long end
,
6779 struct mm_walk
*walk
)
6782 struct vm_area_struct
*vma
= walk
->private;
6785 enum mc_target_type target_type
;
6786 union mc_target target
;
6788 struct page_cgroup
*pc
;
6791 * We don't take compound_lock() here but no race with splitting thp
6793 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6794 * under splitting, which means there's no concurrent thp split,
6795 * - if another thread runs into split_huge_page() just after we
6796 * entered this if-block, the thread must wait for page table lock
6797 * to be unlocked in __split_huge_page_splitting(), where the main
6798 * part of thp split is not executed yet.
6800 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
6801 if (mc
.precharge
< HPAGE_PMD_NR
) {
6805 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6806 if (target_type
== MC_TARGET_PAGE
) {
6808 if (!isolate_lru_page(page
)) {
6809 pc
= lookup_page_cgroup(page
);
6810 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
6811 pc
, mc
.from
, mc
.to
)) {
6812 mc
.precharge
-= HPAGE_PMD_NR
;
6813 mc
.moved_charge
+= HPAGE_PMD_NR
;
6815 putback_lru_page(page
);
6823 if (pmd_trans_unstable(pmd
))
6826 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6827 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6828 pte_t ptent
= *(pte
++);
6834 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6835 case MC_TARGET_PAGE
:
6837 if (isolate_lru_page(page
))
6839 pc
= lookup_page_cgroup(page
);
6840 if (!mem_cgroup_move_account(page
, 1, pc
,
6843 /* we uncharge from mc.from later. */
6846 putback_lru_page(page
);
6847 put
: /* get_mctgt_type() gets the page */
6850 case MC_TARGET_SWAP
:
6852 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6854 /* we fixup refcnts and charges later. */
6862 pte_unmap_unlock(pte
- 1, ptl
);
6867 * We have consumed all precharges we got in can_attach().
6868 * We try charge one by one, but don't do any additional
6869 * charges to mc.to if we have failed in charge once in attach()
6872 ret
= mem_cgroup_do_precharge(1);
6880 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
6882 struct vm_area_struct
*vma
;
6884 lru_add_drain_all();
6886 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
6888 * Someone who are holding the mmap_sem might be waiting in
6889 * waitq. So we cancel all extra charges, wake up all waiters,
6890 * and retry. Because we cancel precharges, we might not be able
6891 * to move enough charges, but moving charge is a best-effort
6892 * feature anyway, so it wouldn't be a big problem.
6894 __mem_cgroup_clear_mc();
6898 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6900 struct mm_walk mem_cgroup_move_charge_walk
= {
6901 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6905 if (is_vm_hugetlb_page(vma
))
6907 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
6908 &mem_cgroup_move_charge_walk
);
6911 * means we have consumed all precharges and failed in
6912 * doing additional charge. Just abandon here.
6916 up_read(&mm
->mmap_sem
);
6919 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6920 struct cgroup_taskset
*tset
)
6922 struct task_struct
*p
= cgroup_taskset_first(tset
);
6923 struct mm_struct
*mm
= get_task_mm(p
);
6927 mem_cgroup_move_charge(mm
);
6931 mem_cgroup_clear_mc();
6933 #else /* !CONFIG_MMU */
6934 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6935 struct cgroup_taskset
*tset
)
6939 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6940 struct cgroup_taskset
*tset
)
6943 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6944 struct cgroup_taskset
*tset
)
6950 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6951 * to verify sane_behavior flag on each mount attempt.
6953 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
6956 * use_hierarchy is forced with sane_behavior. cgroup core
6957 * guarantees that @root doesn't have any children, so turning it
6958 * on for the root memcg is enough.
6960 if (cgroup_sane_behavior(root_css
->cgroup
))
6961 mem_cgroup_from_css(root_css
)->use_hierarchy
= true;
6964 struct cgroup_subsys mem_cgroup_subsys
= {
6966 .subsys_id
= mem_cgroup_subsys_id
,
6967 .css_alloc
= mem_cgroup_css_alloc
,
6968 .css_online
= mem_cgroup_css_online
,
6969 .css_offline
= mem_cgroup_css_offline
,
6970 .css_free
= mem_cgroup_css_free
,
6971 .can_attach
= mem_cgroup_can_attach
,
6972 .cancel_attach
= mem_cgroup_cancel_attach
,
6973 .attach
= mem_cgroup_move_task
,
6974 .bind
= mem_cgroup_bind
,
6975 .base_cftypes
= mem_cgroup_files
,
6979 #ifdef CONFIG_MEMCG_SWAP
6980 static int __init
enable_swap_account(char *s
)
6982 if (!strcmp(s
, "1"))
6983 really_do_swap_account
= 1;
6984 else if (!strcmp(s
, "0"))
6985 really_do_swap_account
= 0;
6988 __setup("swapaccount=", enable_swap_account
);
6990 static void __init
memsw_file_init(void)
6992 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys
, memsw_cgroup_files
));
6995 static void __init
enable_swap_cgroup(void)
6997 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6998 do_swap_account
= 1;
7004 static void __init
enable_swap_cgroup(void)
7010 * subsys_initcall() for memory controller.
7012 * Some parts like hotcpu_notifier() have to be initialized from this context
7013 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7014 * everything that doesn't depend on a specific mem_cgroup structure should
7015 * be initialized from here.
7017 static int __init
mem_cgroup_init(void)
7019 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
7020 enable_swap_cgroup();
7021 mem_cgroup_soft_limit_tree_init();
7025 subsys_initcall(mem_cgroup_init
);