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>
60 #include <net/tcp_memcontrol.h>
62 #include <asm/uaccess.h>
64 #include <trace/events/vmscan.h>
66 struct cgroup_subsys mem_cgroup_subsys __read_mostly
;
67 EXPORT_SYMBOL(mem_cgroup_subsys
);
69 #define MEM_CGROUP_RECLAIM_RETRIES 5
70 static struct mem_cgroup
*root_mem_cgroup __read_mostly
;
72 #ifdef CONFIG_MEMCG_SWAP
73 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
74 int do_swap_account __read_mostly
;
76 /* for remember boot option*/
77 #ifdef CONFIG_MEMCG_SWAP_ENABLED
78 static int really_do_swap_account __initdata
= 1;
80 static int really_do_swap_account __initdata
= 0;
84 #define do_swap_account 0
89 * Statistics for memory cgroup.
91 enum mem_cgroup_stat_index
{
93 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
95 MEM_CGROUP_STAT_CACHE
, /* # of pages charged as cache */
96 MEM_CGROUP_STAT_RSS
, /* # of pages charged as anon rss */
97 MEM_CGROUP_STAT_RSS_HUGE
, /* # of pages charged as anon huge */
98 MEM_CGROUP_STAT_FILE_MAPPED
, /* # of pages charged as file rss */
99 MEM_CGROUP_STAT_SWAP
, /* # of pages, swapped out */
100 MEM_CGROUP_STAT_NSTATS
,
103 static const char * const mem_cgroup_stat_names
[] = {
111 enum mem_cgroup_events_index
{
112 MEM_CGROUP_EVENTS_PGPGIN
, /* # of pages paged in */
113 MEM_CGROUP_EVENTS_PGPGOUT
, /* # of pages paged out */
114 MEM_CGROUP_EVENTS_PGFAULT
, /* # of page-faults */
115 MEM_CGROUP_EVENTS_PGMAJFAULT
, /* # of major page-faults */
116 MEM_CGROUP_EVENTS_NSTATS
,
119 static const char * const mem_cgroup_events_names
[] = {
126 static const char * const mem_cgroup_lru_names
[] = {
135 * Per memcg event counter is incremented at every pagein/pageout. With THP,
136 * it will be incremated by the number of pages. This counter is used for
137 * for trigger some periodic events. This is straightforward and better
138 * than using jiffies etc. to handle periodic memcg event.
140 enum mem_cgroup_events_target
{
141 MEM_CGROUP_TARGET_THRESH
,
142 MEM_CGROUP_TARGET_SOFTLIMIT
,
143 MEM_CGROUP_TARGET_NUMAINFO
,
146 #define THRESHOLDS_EVENTS_TARGET 128
147 #define SOFTLIMIT_EVENTS_TARGET 1024
148 #define NUMAINFO_EVENTS_TARGET 1024
150 struct mem_cgroup_stat_cpu
{
151 long count
[MEM_CGROUP_STAT_NSTATS
];
152 unsigned long events
[MEM_CGROUP_EVENTS_NSTATS
];
153 unsigned long nr_page_events
;
154 unsigned long targets
[MEM_CGROUP_NTARGETS
];
157 struct mem_cgroup_reclaim_iter
{
159 * last scanned hierarchy member. Valid only if last_dead_count
160 * matches memcg->dead_count of the hierarchy root group.
162 struct mem_cgroup
*last_visited
;
163 unsigned long last_dead_count
;
165 /* scan generation, increased every round-trip */
166 unsigned int generation
;
170 * per-zone information in memory controller.
172 struct mem_cgroup_per_zone
{
173 struct lruvec lruvec
;
174 unsigned long lru_size
[NR_LRU_LISTS
];
176 struct mem_cgroup_reclaim_iter reclaim_iter
[DEF_PRIORITY
+ 1];
178 struct rb_node tree_node
; /* RB tree node */
179 unsigned long long usage_in_excess
;/* Set to the value by which */
180 /* the soft limit is exceeded*/
182 struct mem_cgroup
*memcg
; /* Back pointer, we cannot */
183 /* use container_of */
186 struct mem_cgroup_per_node
{
187 struct mem_cgroup_per_zone zoneinfo
[MAX_NR_ZONES
];
191 * Cgroups above their limits are maintained in a RB-Tree, independent of
192 * their hierarchy representation
195 struct mem_cgroup_tree_per_zone
{
196 struct rb_root rb_root
;
200 struct mem_cgroup_tree_per_node
{
201 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
204 struct mem_cgroup_tree
{
205 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
208 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
210 struct mem_cgroup_threshold
{
211 struct eventfd_ctx
*eventfd
;
216 struct mem_cgroup_threshold_ary
{
217 /* An array index points to threshold just below or equal to usage. */
218 int current_threshold
;
219 /* Size of entries[] */
221 /* Array of thresholds */
222 struct mem_cgroup_threshold entries
[0];
225 struct mem_cgroup_thresholds
{
226 /* Primary thresholds array */
227 struct mem_cgroup_threshold_ary
*primary
;
229 * Spare threshold array.
230 * This is needed to make mem_cgroup_unregister_event() "never fail".
231 * It must be able to store at least primary->size - 1 entries.
233 struct mem_cgroup_threshold_ary
*spare
;
237 struct mem_cgroup_eventfd_list
{
238 struct list_head list
;
239 struct eventfd_ctx
*eventfd
;
242 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
243 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
246 * The memory controller data structure. The memory controller controls both
247 * page cache and RSS per cgroup. We would eventually like to provide
248 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
249 * to help the administrator determine what knobs to tune.
251 * TODO: Add a water mark for the memory controller. Reclaim will begin when
252 * we hit the water mark. May be even add a low water mark, such that
253 * no reclaim occurs from a cgroup at it's low water mark, this is
254 * a feature that will be implemented much later in the future.
257 struct cgroup_subsys_state css
;
259 * the counter to account for memory usage
261 struct res_counter res
;
263 /* vmpressure notifications */
264 struct vmpressure vmpressure
;
267 * the counter to account for mem+swap usage.
269 struct res_counter memsw
;
272 * the counter to account for kernel memory usage.
274 struct res_counter kmem
;
276 * Should the accounting and control be hierarchical, per subtree?
279 unsigned long kmem_account_flags
; /* See KMEM_ACCOUNTED_*, below */
285 /* OOM-Killer disable */
286 int oom_kill_disable
;
288 /* set when res.limit == memsw.limit */
289 bool memsw_is_minimum
;
291 /* protect arrays of thresholds */
292 struct mutex thresholds_lock
;
294 /* thresholds for memory usage. RCU-protected */
295 struct mem_cgroup_thresholds thresholds
;
297 /* thresholds for mem+swap usage. RCU-protected */
298 struct mem_cgroup_thresholds memsw_thresholds
;
300 /* For oom notifier event fd */
301 struct list_head oom_notify
;
304 * Should we move charges of a task when a task is moved into this
305 * mem_cgroup ? And what type of charges should we move ?
307 unsigned long move_charge_at_immigrate
;
309 * set > 0 if pages under this cgroup are moving to other cgroup.
311 atomic_t moving_account
;
312 /* taken only while moving_account > 0 */
313 spinlock_t move_lock
;
317 struct mem_cgroup_stat_cpu __percpu
*stat
;
319 * used when a cpu is offlined or other synchronizations
320 * See mem_cgroup_read_stat().
322 struct mem_cgroup_stat_cpu nocpu_base
;
323 spinlock_t pcp_counter_lock
;
326 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
327 struct tcp_memcontrol tcp_mem
;
329 #if defined(CONFIG_MEMCG_KMEM)
330 /* analogous to slab_common's slab_caches list. per-memcg */
331 struct list_head memcg_slab_caches
;
332 /* Not a spinlock, we can take a lot of time walking the list */
333 struct mutex slab_caches_mutex
;
334 /* Index in the kmem_cache->memcg_params->memcg_caches array */
338 int last_scanned_node
;
340 nodemask_t scan_nodes
;
341 atomic_t numainfo_events
;
342 atomic_t numainfo_updating
;
345 struct mem_cgroup_per_node
*nodeinfo
[0];
346 /* WARNING: nodeinfo must be the last member here */
349 static size_t memcg_size(void)
351 return sizeof(struct mem_cgroup
) +
352 nr_node_ids
* sizeof(struct mem_cgroup_per_node
);
355 /* internal only representation about the status of kmem accounting. */
357 KMEM_ACCOUNTED_ACTIVE
= 0, /* accounted by this cgroup itself */
358 KMEM_ACCOUNTED_ACTIVATED
, /* static key enabled. */
359 KMEM_ACCOUNTED_DEAD
, /* dead memcg with pending kmem charges */
362 /* We account when limit is on, but only after call sites are patched */
363 #define KMEM_ACCOUNTED_MASK \
364 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
366 #ifdef CONFIG_MEMCG_KMEM
367 static inline void memcg_kmem_set_active(struct mem_cgroup
*memcg
)
369 set_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
372 static bool memcg_kmem_is_active(struct mem_cgroup
*memcg
)
374 return test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
377 static void memcg_kmem_set_activated(struct mem_cgroup
*memcg
)
379 set_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
382 static void memcg_kmem_clear_activated(struct mem_cgroup
*memcg
)
384 clear_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
387 static void memcg_kmem_mark_dead(struct mem_cgroup
*memcg
)
390 * Our caller must use css_get() first, because memcg_uncharge_kmem()
391 * will call css_put() if it sees the memcg is dead.
394 if (test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
))
395 set_bit(KMEM_ACCOUNTED_DEAD
, &memcg
->kmem_account_flags
);
398 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup
*memcg
)
400 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD
,
401 &memcg
->kmem_account_flags
);
405 /* Stuffs for move charges at task migration. */
407 * Types of charges to be moved. "move_charge_at_immitgrate" and
408 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
411 MOVE_CHARGE_TYPE_ANON
, /* private anonymous page and swap of it */
412 MOVE_CHARGE_TYPE_FILE
, /* file page(including tmpfs) and swap of it */
416 /* "mc" and its members are protected by cgroup_mutex */
417 static struct move_charge_struct
{
418 spinlock_t lock
; /* for from, to */
419 struct mem_cgroup
*from
;
420 struct mem_cgroup
*to
;
421 unsigned long immigrate_flags
;
422 unsigned long precharge
;
423 unsigned long moved_charge
;
424 unsigned long moved_swap
;
425 struct task_struct
*moving_task
; /* a task moving charges */
426 wait_queue_head_t waitq
; /* a waitq for other context */
428 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
429 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
432 static bool move_anon(void)
434 return test_bit(MOVE_CHARGE_TYPE_ANON
, &mc
.immigrate_flags
);
437 static bool move_file(void)
439 return test_bit(MOVE_CHARGE_TYPE_FILE
, &mc
.immigrate_flags
);
443 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
444 * limit reclaim to prevent infinite loops, if they ever occur.
446 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
447 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
450 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
451 MEM_CGROUP_CHARGE_TYPE_ANON
,
452 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
453 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
457 /* for encoding cft->private value on file */
465 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
466 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
467 #define MEMFILE_ATTR(val) ((val) & 0xffff)
468 /* Used for OOM nofiier */
469 #define OOM_CONTROL (0)
472 * Reclaim flags for mem_cgroup_hierarchical_reclaim
474 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
475 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
476 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
477 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
480 * The memcg_create_mutex will be held whenever a new cgroup is created.
481 * As a consequence, any change that needs to protect against new child cgroups
482 * appearing has to hold it as well.
484 static DEFINE_MUTEX(memcg_create_mutex
);
486 struct mem_cgroup
*mem_cgroup_from_css(struct cgroup_subsys_state
*s
)
488 return s
? container_of(s
, struct mem_cgroup
, css
) : NULL
;
491 /* Some nice accessors for the vmpressure. */
492 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
495 memcg
= root_mem_cgroup
;
496 return &memcg
->vmpressure
;
499 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
501 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
504 struct vmpressure
*css_to_vmpressure(struct cgroup_subsys_state
*css
)
506 return &mem_cgroup_from_css(css
)->vmpressure
;
509 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
511 return (memcg
== root_mem_cgroup
);
514 /* Writing them here to avoid exposing memcg's inner layout */
515 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
517 void sock_update_memcg(struct sock
*sk
)
519 if (mem_cgroup_sockets_enabled
) {
520 struct mem_cgroup
*memcg
;
521 struct cg_proto
*cg_proto
;
523 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
525 /* Socket cloning can throw us here with sk_cgrp already
526 * filled. It won't however, necessarily happen from
527 * process context. So the test for root memcg given
528 * the current task's memcg won't help us in this case.
530 * Respecting the original socket's memcg is a better
531 * decision in this case.
534 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
535 css_get(&sk
->sk_cgrp
->memcg
->css
);
540 memcg
= mem_cgroup_from_task(current
);
541 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
542 if (!mem_cgroup_is_root(memcg
) &&
543 memcg_proto_active(cg_proto
) && css_tryget(&memcg
->css
)) {
544 sk
->sk_cgrp
= cg_proto
;
549 EXPORT_SYMBOL(sock_update_memcg
);
551 void sock_release_memcg(struct sock
*sk
)
553 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
554 struct mem_cgroup
*memcg
;
555 WARN_ON(!sk
->sk_cgrp
->memcg
);
556 memcg
= sk
->sk_cgrp
->memcg
;
557 css_put(&sk
->sk_cgrp
->memcg
->css
);
561 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
563 if (!memcg
|| mem_cgroup_is_root(memcg
))
566 return &memcg
->tcp_mem
.cg_proto
;
568 EXPORT_SYMBOL(tcp_proto_cgroup
);
570 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
572 if (!memcg_proto_activated(&memcg
->tcp_mem
.cg_proto
))
574 static_key_slow_dec(&memcg_socket_limit_enabled
);
577 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
582 #ifdef CONFIG_MEMCG_KMEM
584 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
585 * There are two main reasons for not using the css_id for this:
586 * 1) this works better in sparse environments, where we have a lot of memcgs,
587 * but only a few kmem-limited. Or also, if we have, for instance, 200
588 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
589 * 200 entry array for that.
591 * 2) In order not to violate the cgroup API, we would like to do all memory
592 * allocation in ->create(). At that point, we haven't yet allocated the
593 * css_id. Having a separate index prevents us from messing with the cgroup
596 * The current size of the caches array is stored in
597 * memcg_limited_groups_array_size. It will double each time we have to
600 static DEFINE_IDA(kmem_limited_groups
);
601 int memcg_limited_groups_array_size
;
604 * MIN_SIZE is different than 1, because we would like to avoid going through
605 * the alloc/free process all the time. In a small machine, 4 kmem-limited
606 * cgroups is a reasonable guess. In the future, it could be a parameter or
607 * tunable, but that is strictly not necessary.
609 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
610 * this constant directly from cgroup, but it is understandable that this is
611 * better kept as an internal representation in cgroup.c. In any case, the
612 * css_id space is not getting any smaller, and we don't have to necessarily
613 * increase ours as well if it increases.
615 #define MEMCG_CACHES_MIN_SIZE 4
616 #define MEMCG_CACHES_MAX_SIZE 65535
619 * A lot of the calls to the cache allocation functions are expected to be
620 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
621 * conditional to this static branch, we'll have to allow modules that does
622 * kmem_cache_alloc and the such to see this symbol as well
624 struct static_key memcg_kmem_enabled_key
;
625 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
627 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
629 if (memcg_kmem_is_active(memcg
)) {
630 static_key_slow_dec(&memcg_kmem_enabled_key
);
631 ida_simple_remove(&kmem_limited_groups
, memcg
->kmemcg_id
);
634 * This check can't live in kmem destruction function,
635 * since the charges will outlive the cgroup
637 WARN_ON(res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0);
640 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
643 #endif /* CONFIG_MEMCG_KMEM */
645 static void disarm_static_keys(struct mem_cgroup
*memcg
)
647 disarm_sock_keys(memcg
);
648 disarm_kmem_keys(memcg
);
651 static void drain_all_stock_async(struct mem_cgroup
*memcg
);
653 static struct mem_cgroup_per_zone
*
654 mem_cgroup_zoneinfo(struct mem_cgroup
*memcg
, int nid
, int zid
)
656 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
657 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
660 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
665 static struct mem_cgroup_per_zone
*
666 page_cgroup_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
668 int nid
= page_to_nid(page
);
669 int zid
= page_zonenum(page
);
671 return mem_cgroup_zoneinfo(memcg
, nid
, zid
);
674 static struct mem_cgroup_tree_per_zone
*
675 soft_limit_tree_node_zone(int nid
, int zid
)
677 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
680 static struct mem_cgroup_tree_per_zone
*
681 soft_limit_tree_from_page(struct page
*page
)
683 int nid
= page_to_nid(page
);
684 int zid
= page_zonenum(page
);
686 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
690 __mem_cgroup_insert_exceeded(struct mem_cgroup
*memcg
,
691 struct mem_cgroup_per_zone
*mz
,
692 struct mem_cgroup_tree_per_zone
*mctz
,
693 unsigned long long new_usage_in_excess
)
695 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
696 struct rb_node
*parent
= NULL
;
697 struct mem_cgroup_per_zone
*mz_node
;
702 mz
->usage_in_excess
= new_usage_in_excess
;
703 if (!mz
->usage_in_excess
)
707 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
709 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
712 * We can't avoid mem cgroups that are over their soft
713 * limit by the same amount
715 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
718 rb_link_node(&mz
->tree_node
, parent
, p
);
719 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
724 __mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
725 struct mem_cgroup_per_zone
*mz
,
726 struct mem_cgroup_tree_per_zone
*mctz
)
730 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
735 mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
736 struct mem_cgroup_per_zone
*mz
,
737 struct mem_cgroup_tree_per_zone
*mctz
)
739 spin_lock(&mctz
->lock
);
740 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
741 spin_unlock(&mctz
->lock
);
745 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
747 unsigned long long excess
;
748 struct mem_cgroup_per_zone
*mz
;
749 struct mem_cgroup_tree_per_zone
*mctz
;
750 int nid
= page_to_nid(page
);
751 int zid
= page_zonenum(page
);
752 mctz
= soft_limit_tree_from_page(page
);
755 * Necessary to update all ancestors when hierarchy is used.
756 * because their event counter is not touched.
758 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
759 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
760 excess
= res_counter_soft_limit_excess(&memcg
->res
);
762 * We have to update the tree if mz is on RB-tree or
763 * mem is over its softlimit.
765 if (excess
|| mz
->on_tree
) {
766 spin_lock(&mctz
->lock
);
767 /* if on-tree, remove it */
769 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
771 * Insert again. mz->usage_in_excess will be updated.
772 * If excess is 0, no tree ops.
774 __mem_cgroup_insert_exceeded(memcg
, mz
, mctz
, excess
);
775 spin_unlock(&mctz
->lock
);
780 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
783 struct mem_cgroup_per_zone
*mz
;
784 struct mem_cgroup_tree_per_zone
*mctz
;
786 for_each_node(node
) {
787 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
788 mz
= mem_cgroup_zoneinfo(memcg
, node
, zone
);
789 mctz
= soft_limit_tree_node_zone(node
, zone
);
790 mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
795 static struct mem_cgroup_per_zone
*
796 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
798 struct rb_node
*rightmost
= NULL
;
799 struct mem_cgroup_per_zone
*mz
;
803 rightmost
= rb_last(&mctz
->rb_root
);
805 goto done
; /* Nothing to reclaim from */
807 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
809 * Remove the node now but someone else can add it back,
810 * we will to add it back at the end of reclaim to its correct
811 * position in the tree.
813 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
814 if (!res_counter_soft_limit_excess(&mz
->memcg
->res
) ||
815 !css_tryget(&mz
->memcg
->css
))
821 static struct mem_cgroup_per_zone
*
822 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
824 struct mem_cgroup_per_zone
*mz
;
826 spin_lock(&mctz
->lock
);
827 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
828 spin_unlock(&mctz
->lock
);
833 * Implementation Note: reading percpu statistics for memcg.
835 * Both of vmstat[] and percpu_counter has threshold and do periodic
836 * synchronization to implement "quick" read. There are trade-off between
837 * reading cost and precision of value. Then, we may have a chance to implement
838 * a periodic synchronizion of counter in memcg's counter.
840 * But this _read() function is used for user interface now. The user accounts
841 * memory usage by memory cgroup and he _always_ requires exact value because
842 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
843 * have to visit all online cpus and make sum. So, for now, unnecessary
844 * synchronization is not implemented. (just implemented for cpu hotplug)
846 * If there are kernel internal actions which can make use of some not-exact
847 * value, and reading all cpu value can be performance bottleneck in some
848 * common workload, threashold and synchonization as vmstat[] should be
851 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
852 enum mem_cgroup_stat_index idx
)
858 for_each_online_cpu(cpu
)
859 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
860 #ifdef CONFIG_HOTPLUG_CPU
861 spin_lock(&memcg
->pcp_counter_lock
);
862 val
+= memcg
->nocpu_base
.count
[idx
];
863 spin_unlock(&memcg
->pcp_counter_lock
);
869 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
872 int val
= (charge
) ? 1 : -1;
873 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
876 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
877 enum mem_cgroup_events_index idx
)
879 unsigned long val
= 0;
882 for_each_online_cpu(cpu
)
883 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
884 #ifdef CONFIG_HOTPLUG_CPU
885 spin_lock(&memcg
->pcp_counter_lock
);
886 val
+= memcg
->nocpu_base
.events
[idx
];
887 spin_unlock(&memcg
->pcp_counter_lock
);
892 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
894 bool anon
, int nr_pages
)
899 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
900 * counted as CACHE even if it's on ANON LRU.
903 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
906 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
909 if (PageTransHuge(page
))
910 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
913 /* pagein of a big page is an event. So, ignore page size */
915 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
917 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
918 nr_pages
= -nr_pages
; /* for event */
921 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
927 mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
929 struct mem_cgroup_per_zone
*mz
;
931 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
932 return mz
->lru_size
[lru
];
936 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup
*memcg
, int nid
, int zid
,
937 unsigned int lru_mask
)
939 struct mem_cgroup_per_zone
*mz
;
941 unsigned long ret
= 0;
943 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
946 if (BIT(lru
) & lru_mask
)
947 ret
+= mz
->lru_size
[lru
];
953 mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
954 int nid
, unsigned int lru_mask
)
959 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++)
960 total
+= mem_cgroup_zone_nr_lru_pages(memcg
,
966 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
967 unsigned int lru_mask
)
972 for_each_node_state(nid
, N_MEMORY
)
973 total
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
977 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
978 enum mem_cgroup_events_target target
)
980 unsigned long val
, next
;
982 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
983 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
984 /* from time_after() in jiffies.h */
985 if ((long)next
- (long)val
< 0) {
987 case MEM_CGROUP_TARGET_THRESH
:
988 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
990 case MEM_CGROUP_TARGET_SOFTLIMIT
:
991 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
993 case MEM_CGROUP_TARGET_NUMAINFO
:
994 next
= val
+ NUMAINFO_EVENTS_TARGET
;
999 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
1006 * Check events in order.
1009 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
1012 /* threshold event is triggered in finer grain than soft limit */
1013 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
1014 MEM_CGROUP_TARGET_THRESH
))) {
1016 bool do_numainfo __maybe_unused
;
1018 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
1019 MEM_CGROUP_TARGET_SOFTLIMIT
);
1020 #if MAX_NUMNODES > 1
1021 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
1022 MEM_CGROUP_TARGET_NUMAINFO
);
1026 mem_cgroup_threshold(memcg
);
1027 if (unlikely(do_softlimit
))
1028 mem_cgroup_update_tree(memcg
, page
);
1029 #if MAX_NUMNODES > 1
1030 if (unlikely(do_numainfo
))
1031 atomic_inc(&memcg
->numainfo_events
);
1037 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
1040 * mm_update_next_owner() may clear mm->owner to NULL
1041 * if it races with swapoff, page migration, etc.
1042 * So this can be called with p == NULL.
1047 return mem_cgroup_from_css(task_css(p
, mem_cgroup_subsys_id
));
1050 struct mem_cgroup
*try_get_mem_cgroup_from_mm(struct mm_struct
*mm
)
1052 struct mem_cgroup
*memcg
= NULL
;
1057 * Because we have no locks, mm->owner's may be being moved to other
1058 * cgroup. We use css_tryget() here even if this looks
1059 * pessimistic (rather than adding locks here).
1063 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1064 if (unlikely(!memcg
))
1066 } while (!css_tryget(&memcg
->css
));
1072 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1073 * ref. count) or NULL if the whole root's subtree has been visited.
1075 * helper function to be used by mem_cgroup_iter
1077 static struct mem_cgroup
*__mem_cgroup_iter_next(struct mem_cgroup
*root
,
1078 struct mem_cgroup
*last_visited
)
1080 struct cgroup_subsys_state
*prev_css
, *next_css
;
1083 * Root is not visited by cgroup iterators so it needs an
1089 prev_css
= (last_visited
== root
) ? NULL
: &last_visited
->css
;
1091 next_css
= css_next_descendant_pre(prev_css
, &root
->css
);
1094 * Even if we found a group we have to make sure it is
1095 * alive. css && !memcg means that the groups should be
1096 * skipped and we should continue the tree walk.
1097 * last_visited css is safe to use because it is
1098 * protected by css_get and the tree walk is rcu safe.
1101 struct mem_cgroup
*mem
= mem_cgroup_from_css(next_css
);
1103 if (css_tryget(&mem
->css
))
1106 prev_css
= next_css
;
1114 static void mem_cgroup_iter_invalidate(struct mem_cgroup
*root
)
1117 * When a group in the hierarchy below root is destroyed, the
1118 * hierarchy iterator can no longer be trusted since it might
1119 * have pointed to the destroyed group. Invalidate it.
1121 atomic_inc(&root
->dead_count
);
1124 static struct mem_cgroup
*
1125 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter
*iter
,
1126 struct mem_cgroup
*root
,
1129 struct mem_cgroup
*position
= NULL
;
1131 * A cgroup destruction happens in two stages: offlining and
1132 * release. They are separated by a RCU grace period.
1134 * If the iterator is valid, we may still race with an
1135 * offlining. The RCU lock ensures the object won't be
1136 * released, tryget will fail if we lost the race.
1138 *sequence
= atomic_read(&root
->dead_count
);
1139 if (iter
->last_dead_count
== *sequence
) {
1141 position
= iter
->last_visited
;
1142 if (position
&& !css_tryget(&position
->css
))
1148 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter
*iter
,
1149 struct mem_cgroup
*last_visited
,
1150 struct mem_cgroup
*new_position
,
1154 css_put(&last_visited
->css
);
1156 * We store the sequence count from the time @last_visited was
1157 * loaded successfully instead of rereading it here so that we
1158 * don't lose destruction events in between. We could have
1159 * raced with the destruction of @new_position after all.
1161 iter
->last_visited
= new_position
;
1163 iter
->last_dead_count
= sequence
;
1167 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1168 * @root: hierarchy root
1169 * @prev: previously returned memcg, NULL on first invocation
1170 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1172 * Returns references to children of the hierarchy below @root, or
1173 * @root itself, or %NULL after a full round-trip.
1175 * Caller must pass the return value in @prev on subsequent
1176 * invocations for reference counting, or use mem_cgroup_iter_break()
1177 * to cancel a hierarchy walk before the round-trip is complete.
1179 * Reclaimers can specify a zone and a priority level in @reclaim to
1180 * divide up the memcgs in the hierarchy among all concurrent
1181 * reclaimers operating on the same zone and priority.
1183 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1184 struct mem_cgroup
*prev
,
1185 struct mem_cgroup_reclaim_cookie
*reclaim
)
1187 struct mem_cgroup
*memcg
= NULL
;
1188 struct mem_cgroup
*last_visited
= NULL
;
1190 if (mem_cgroup_disabled())
1194 root
= root_mem_cgroup
;
1196 if (prev
&& !reclaim
)
1197 last_visited
= prev
;
1199 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1207 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1208 int uninitialized_var(seq
);
1211 int nid
= zone_to_nid(reclaim
->zone
);
1212 int zid
= zone_idx(reclaim
->zone
);
1213 struct mem_cgroup_per_zone
*mz
;
1215 mz
= mem_cgroup_zoneinfo(root
, nid
, zid
);
1216 iter
= &mz
->reclaim_iter
[reclaim
->priority
];
1217 if (prev
&& reclaim
->generation
!= iter
->generation
) {
1218 iter
->last_visited
= NULL
;
1222 last_visited
= mem_cgroup_iter_load(iter
, root
, &seq
);
1225 memcg
= __mem_cgroup_iter_next(root
, last_visited
);
1228 mem_cgroup_iter_update(iter
, last_visited
, memcg
, seq
);
1232 else if (!prev
&& memcg
)
1233 reclaim
->generation
= iter
->generation
;
1242 if (prev
&& prev
!= root
)
1243 css_put(&prev
->css
);
1249 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1250 * @root: hierarchy root
1251 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1253 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1254 struct mem_cgroup
*prev
)
1257 root
= root_mem_cgroup
;
1258 if (prev
&& prev
!= root
)
1259 css_put(&prev
->css
);
1263 * Iteration constructs for visiting all cgroups (under a tree). If
1264 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1265 * be used for reference counting.
1267 #define for_each_mem_cgroup_tree(iter, root) \
1268 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1270 iter = mem_cgroup_iter(root, iter, NULL))
1272 #define for_each_mem_cgroup(iter) \
1273 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1275 iter = mem_cgroup_iter(NULL, iter, NULL))
1277 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1279 struct mem_cgroup
*memcg
;
1282 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1283 if (unlikely(!memcg
))
1288 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1291 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1299 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1302 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1303 * @zone: zone of the wanted lruvec
1304 * @memcg: memcg of the wanted lruvec
1306 * Returns the lru list vector holding pages for the given @zone and
1307 * @mem. This can be the global zone lruvec, if the memory controller
1310 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1311 struct mem_cgroup
*memcg
)
1313 struct mem_cgroup_per_zone
*mz
;
1314 struct lruvec
*lruvec
;
1316 if (mem_cgroup_disabled()) {
1317 lruvec
= &zone
->lruvec
;
1321 mz
= mem_cgroup_zoneinfo(memcg
, zone_to_nid(zone
), zone_idx(zone
));
1322 lruvec
= &mz
->lruvec
;
1325 * Since a node can be onlined after the mem_cgroup was created,
1326 * we have to be prepared to initialize lruvec->zone here;
1327 * and if offlined then reonlined, we need to reinitialize it.
1329 if (unlikely(lruvec
->zone
!= zone
))
1330 lruvec
->zone
= zone
;
1335 * Following LRU functions are allowed to be used without PCG_LOCK.
1336 * Operations are called by routine of global LRU independently from memcg.
1337 * What we have to take care of here is validness of pc->mem_cgroup.
1339 * Changes to pc->mem_cgroup happens when
1342 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1343 * It is added to LRU before charge.
1344 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1345 * When moving account, the page is not on LRU. It's isolated.
1349 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1351 * @zone: zone of the page
1353 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1355 struct mem_cgroup_per_zone
*mz
;
1356 struct mem_cgroup
*memcg
;
1357 struct page_cgroup
*pc
;
1358 struct lruvec
*lruvec
;
1360 if (mem_cgroup_disabled()) {
1361 lruvec
= &zone
->lruvec
;
1365 pc
= lookup_page_cgroup(page
);
1366 memcg
= pc
->mem_cgroup
;
1369 * Surreptitiously switch any uncharged offlist page to root:
1370 * an uncharged page off lru does nothing to secure
1371 * its former mem_cgroup from sudden removal.
1373 * Our caller holds lru_lock, and PageCgroupUsed is updated
1374 * under page_cgroup lock: between them, they make all uses
1375 * of pc->mem_cgroup safe.
1377 if (!PageLRU(page
) && !PageCgroupUsed(pc
) && memcg
!= root_mem_cgroup
)
1378 pc
->mem_cgroup
= memcg
= root_mem_cgroup
;
1380 mz
= page_cgroup_zoneinfo(memcg
, page
);
1381 lruvec
= &mz
->lruvec
;
1384 * Since a node can be onlined after the mem_cgroup was created,
1385 * we have to be prepared to initialize lruvec->zone here;
1386 * and if offlined then reonlined, we need to reinitialize it.
1388 if (unlikely(lruvec
->zone
!= zone
))
1389 lruvec
->zone
= zone
;
1394 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1395 * @lruvec: mem_cgroup per zone lru vector
1396 * @lru: index of lru list the page is sitting on
1397 * @nr_pages: positive when adding or negative when removing
1399 * This function must be called when a page is added to or removed from an
1402 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1405 struct mem_cgroup_per_zone
*mz
;
1406 unsigned long *lru_size
;
1408 if (mem_cgroup_disabled())
1411 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1412 lru_size
= mz
->lru_size
+ lru
;
1413 *lru_size
+= nr_pages
;
1414 VM_BUG_ON((long)(*lru_size
) < 0);
1418 * Checks whether given mem is same or in the root_mem_cgroup's
1421 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1422 struct mem_cgroup
*memcg
)
1424 if (root_memcg
== memcg
)
1426 if (!root_memcg
->use_hierarchy
|| !memcg
)
1428 return css_is_ancestor(&memcg
->css
, &root_memcg
->css
);
1431 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1432 struct mem_cgroup
*memcg
)
1437 ret
= __mem_cgroup_same_or_subtree(root_memcg
, memcg
);
1442 bool task_in_mem_cgroup(struct task_struct
*task
,
1443 const struct mem_cgroup
*memcg
)
1445 struct mem_cgroup
*curr
= NULL
;
1446 struct task_struct
*p
;
1449 p
= find_lock_task_mm(task
);
1451 curr
= try_get_mem_cgroup_from_mm(p
->mm
);
1455 * All threads may have already detached their mm's, but the oom
1456 * killer still needs to detect if they have already been oom
1457 * killed to prevent needlessly killing additional tasks.
1460 curr
= mem_cgroup_from_task(task
);
1462 css_get(&curr
->css
);
1468 * We should check use_hierarchy of "memcg" not "curr". Because checking
1469 * use_hierarchy of "curr" here make this function true if hierarchy is
1470 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1471 * hierarchy(even if use_hierarchy is disabled in "memcg").
1473 ret
= mem_cgroup_same_or_subtree(memcg
, curr
);
1474 css_put(&curr
->css
);
1478 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1480 unsigned long inactive_ratio
;
1481 unsigned long inactive
;
1482 unsigned long active
;
1485 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1486 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1488 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1490 inactive_ratio
= int_sqrt(10 * gb
);
1494 return inactive
* inactive_ratio
< active
;
1497 #define mem_cgroup_from_res_counter(counter, member) \
1498 container_of(counter, struct mem_cgroup, member)
1501 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1502 * @memcg: the memory cgroup
1504 * Returns the maximum amount of memory @mem can be charged with, in
1507 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1509 unsigned long long margin
;
1511 margin
= res_counter_margin(&memcg
->res
);
1512 if (do_swap_account
)
1513 margin
= min(margin
, res_counter_margin(&memcg
->memsw
));
1514 return margin
>> PAGE_SHIFT
;
1517 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1520 if (!css_parent(&memcg
->css
))
1521 return vm_swappiness
;
1523 return memcg
->swappiness
;
1527 * memcg->moving_account is used for checking possibility that some thread is
1528 * calling move_account(). When a thread on CPU-A starts moving pages under
1529 * a memcg, other threads should check memcg->moving_account under
1530 * rcu_read_lock(), like this:
1534 * memcg->moving_account+1 if (memcg->mocing_account)
1536 * synchronize_rcu() update something.
1541 /* for quick checking without looking up memcg */
1542 atomic_t memcg_moving __read_mostly
;
1544 static void mem_cgroup_start_move(struct mem_cgroup
*memcg
)
1546 atomic_inc(&memcg_moving
);
1547 atomic_inc(&memcg
->moving_account
);
1551 static void mem_cgroup_end_move(struct mem_cgroup
*memcg
)
1554 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1555 * We check NULL in callee rather than caller.
1558 atomic_dec(&memcg_moving
);
1559 atomic_dec(&memcg
->moving_account
);
1564 * 2 routines for checking "mem" is under move_account() or not.
1566 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1567 * is used for avoiding races in accounting. If true,
1568 * pc->mem_cgroup may be overwritten.
1570 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1571 * under hierarchy of moving cgroups. This is for
1572 * waiting at hith-memory prressure caused by "move".
1575 static bool mem_cgroup_stolen(struct mem_cgroup
*memcg
)
1577 VM_BUG_ON(!rcu_read_lock_held());
1578 return atomic_read(&memcg
->moving_account
) > 0;
1581 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1583 struct mem_cgroup
*from
;
1584 struct mem_cgroup
*to
;
1587 * Unlike task_move routines, we access mc.to, mc.from not under
1588 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1590 spin_lock(&mc
.lock
);
1596 ret
= mem_cgroup_same_or_subtree(memcg
, from
)
1597 || mem_cgroup_same_or_subtree(memcg
, to
);
1599 spin_unlock(&mc
.lock
);
1603 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1605 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1606 if (mem_cgroup_under_move(memcg
)) {
1608 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1609 /* moving charge context might have finished. */
1612 finish_wait(&mc
.waitq
, &wait
);
1620 * Take this lock when
1621 * - a code tries to modify page's memcg while it's USED.
1622 * - a code tries to modify page state accounting in a memcg.
1623 * see mem_cgroup_stolen(), too.
1625 static void move_lock_mem_cgroup(struct mem_cgroup
*memcg
,
1626 unsigned long *flags
)
1628 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
1631 static void move_unlock_mem_cgroup(struct mem_cgroup
*memcg
,
1632 unsigned long *flags
)
1634 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
1637 #define K(x) ((x) << (PAGE_SHIFT-10))
1639 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1640 * @memcg: The memory cgroup that went over limit
1641 * @p: Task that is going to be killed
1643 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1646 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1648 struct cgroup
*task_cgrp
;
1649 struct cgroup
*mem_cgrp
;
1651 * Need a buffer in BSS, can't rely on allocations. The code relies
1652 * on the assumption that OOM is serialized for memory controller.
1653 * If this assumption is broken, revisit this code.
1655 static char memcg_name
[PATH_MAX
];
1657 struct mem_cgroup
*iter
;
1665 mem_cgrp
= memcg
->css
.cgroup
;
1666 task_cgrp
= task_cgroup(p
, mem_cgroup_subsys_id
);
1668 ret
= cgroup_path(task_cgrp
, memcg_name
, PATH_MAX
);
1671 * Unfortunately, we are unable to convert to a useful name
1672 * But we'll still print out the usage information
1679 pr_info("Task in %s killed", memcg_name
);
1682 ret
= cgroup_path(mem_cgrp
, memcg_name
, PATH_MAX
);
1690 * Continues from above, so we don't need an KERN_ level
1692 pr_cont(" as a result of limit of %s\n", memcg_name
);
1695 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1696 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1697 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1698 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1699 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1700 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1701 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1702 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1703 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1704 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) >> 10,
1705 res_counter_read_u64(&memcg
->kmem
, RES_LIMIT
) >> 10,
1706 res_counter_read_u64(&memcg
->kmem
, RES_FAILCNT
));
1708 for_each_mem_cgroup_tree(iter
, memcg
) {
1709 pr_info("Memory cgroup stats");
1712 ret
= cgroup_path(iter
->css
.cgroup
, memcg_name
, PATH_MAX
);
1714 pr_cont(" for %s", memcg_name
);
1718 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1719 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1721 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1722 K(mem_cgroup_read_stat(iter
, i
)));
1725 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1726 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1727 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1734 * This function returns the number of memcg under hierarchy tree. Returns
1735 * 1(self count) if no children.
1737 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1740 struct mem_cgroup
*iter
;
1742 for_each_mem_cgroup_tree(iter
, memcg
)
1748 * Return the memory (and swap, if configured) limit for a memcg.
1750 static u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1754 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1757 * Do not consider swap space if we cannot swap due to swappiness
1759 if (mem_cgroup_swappiness(memcg
)) {
1762 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1763 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1766 * If memsw is finite and limits the amount of swap space
1767 * available to this memcg, return that limit.
1769 limit
= min(limit
, memsw
);
1775 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1778 struct mem_cgroup
*iter
;
1779 unsigned long chosen_points
= 0;
1780 unsigned long totalpages
;
1781 unsigned int points
= 0;
1782 struct task_struct
*chosen
= NULL
;
1785 * If current has a pending SIGKILL or is exiting, then automatically
1786 * select it. The goal is to allow it to allocate so that it may
1787 * quickly exit and free its memory.
1789 if (fatal_signal_pending(current
) || current
->flags
& PF_EXITING
) {
1790 set_thread_flag(TIF_MEMDIE
);
1794 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1795 totalpages
= mem_cgroup_get_limit(memcg
) >> PAGE_SHIFT
? : 1;
1796 for_each_mem_cgroup_tree(iter
, memcg
) {
1797 struct css_task_iter it
;
1798 struct task_struct
*task
;
1800 css_task_iter_start(&iter
->css
, &it
);
1801 while ((task
= css_task_iter_next(&it
))) {
1802 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1804 case OOM_SCAN_SELECT
:
1806 put_task_struct(chosen
);
1808 chosen_points
= ULONG_MAX
;
1809 get_task_struct(chosen
);
1811 case OOM_SCAN_CONTINUE
:
1813 case OOM_SCAN_ABORT
:
1814 css_task_iter_end(&it
);
1815 mem_cgroup_iter_break(memcg
, iter
);
1817 put_task_struct(chosen
);
1822 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1823 if (points
> chosen_points
) {
1825 put_task_struct(chosen
);
1827 chosen_points
= points
;
1828 get_task_struct(chosen
);
1831 css_task_iter_end(&it
);
1836 points
= chosen_points
* 1000 / totalpages
;
1837 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1838 NULL
, "Memory cgroup out of memory");
1841 static unsigned long mem_cgroup_reclaim(struct mem_cgroup
*memcg
,
1843 unsigned long flags
)
1845 unsigned long total
= 0;
1846 bool noswap
= false;
1849 if (flags
& MEM_CGROUP_RECLAIM_NOSWAP
)
1851 if (!(flags
& MEM_CGROUP_RECLAIM_SHRINK
) && memcg
->memsw_is_minimum
)
1854 for (loop
= 0; loop
< MEM_CGROUP_MAX_RECLAIM_LOOPS
; loop
++) {
1856 drain_all_stock_async(memcg
);
1857 total
+= try_to_free_mem_cgroup_pages(memcg
, gfp_mask
, noswap
);
1859 * Allow limit shrinkers, which are triggered directly
1860 * by userspace, to catch signals and stop reclaim
1861 * after minimal progress, regardless of the margin.
1863 if (total
&& (flags
& MEM_CGROUP_RECLAIM_SHRINK
))
1865 if (mem_cgroup_margin(memcg
))
1868 * If nothing was reclaimed after two attempts, there
1869 * may be no reclaimable pages in this hierarchy.
1878 * test_mem_cgroup_node_reclaimable
1879 * @memcg: the target memcg
1880 * @nid: the node ID to be checked.
1881 * @noswap : specify true here if the user wants flle only information.
1883 * This function returns whether the specified memcg contains any
1884 * reclaimable pages on a node. Returns true if there are any reclaimable
1885 * pages in the node.
1887 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1888 int nid
, bool noswap
)
1890 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1892 if (noswap
|| !total_swap_pages
)
1894 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1899 #if MAX_NUMNODES > 1
1902 * Always updating the nodemask is not very good - even if we have an empty
1903 * list or the wrong list here, we can start from some node and traverse all
1904 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1907 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1911 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1912 * pagein/pageout changes since the last update.
1914 if (!atomic_read(&memcg
->numainfo_events
))
1916 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1919 /* make a nodemask where this memcg uses memory from */
1920 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1922 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1924 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1925 node_clear(nid
, memcg
->scan_nodes
);
1928 atomic_set(&memcg
->numainfo_events
, 0);
1929 atomic_set(&memcg
->numainfo_updating
, 0);
1933 * Selecting a node where we start reclaim from. Because what we need is just
1934 * reducing usage counter, start from anywhere is O,K. Considering
1935 * memory reclaim from current node, there are pros. and cons.
1937 * Freeing memory from current node means freeing memory from a node which
1938 * we'll use or we've used. So, it may make LRU bad. And if several threads
1939 * hit limits, it will see a contention on a node. But freeing from remote
1940 * node means more costs for memory reclaim because of memory latency.
1942 * Now, we use round-robin. Better algorithm is welcomed.
1944 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1948 mem_cgroup_may_update_nodemask(memcg
);
1949 node
= memcg
->last_scanned_node
;
1951 node
= next_node(node
, memcg
->scan_nodes
);
1952 if (node
== MAX_NUMNODES
)
1953 node
= first_node(memcg
->scan_nodes
);
1955 * We call this when we hit limit, not when pages are added to LRU.
1956 * No LRU may hold pages because all pages are UNEVICTABLE or
1957 * memcg is too small and all pages are not on LRU. In that case,
1958 * we use curret node.
1960 if (unlikely(node
== MAX_NUMNODES
))
1961 node
= numa_node_id();
1963 memcg
->last_scanned_node
= node
;
1968 * Check all nodes whether it contains reclaimable pages or not.
1969 * For quick scan, we make use of scan_nodes. This will allow us to skip
1970 * unused nodes. But scan_nodes is lazily updated and may not cotain
1971 * enough new information. We need to do double check.
1973 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1978 * quick check...making use of scan_node.
1979 * We can skip unused nodes.
1981 if (!nodes_empty(memcg
->scan_nodes
)) {
1982 for (nid
= first_node(memcg
->scan_nodes
);
1984 nid
= next_node(nid
, memcg
->scan_nodes
)) {
1986 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1991 * Check rest of nodes.
1993 for_each_node_state(nid
, N_MEMORY
) {
1994 if (node_isset(nid
, memcg
->scan_nodes
))
1996 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
2003 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
2008 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
2010 return test_mem_cgroup_node_reclaimable(memcg
, 0, noswap
);
2014 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
2017 unsigned long *total_scanned
)
2019 struct mem_cgroup
*victim
= NULL
;
2022 unsigned long excess
;
2023 unsigned long nr_scanned
;
2024 struct mem_cgroup_reclaim_cookie reclaim
= {
2029 excess
= res_counter_soft_limit_excess(&root_memcg
->res
) >> PAGE_SHIFT
;
2032 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
2037 * If we have not been able to reclaim
2038 * anything, it might because there are
2039 * no reclaimable pages under this hierarchy
2044 * We want to do more targeted reclaim.
2045 * excess >> 2 is not to excessive so as to
2046 * reclaim too much, nor too less that we keep
2047 * coming back to reclaim from this cgroup
2049 if (total
>= (excess
>> 2) ||
2050 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
2055 if (!mem_cgroup_reclaimable(victim
, false))
2057 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
2059 *total_scanned
+= nr_scanned
;
2060 if (!res_counter_soft_limit_excess(&root_memcg
->res
))
2063 mem_cgroup_iter_break(root_memcg
, victim
);
2068 * Check OOM-Killer is already running under our hierarchy.
2069 * If someone is running, return false.
2070 * Has to be called with memcg_oom_lock
2072 static bool mem_cgroup_oom_lock(struct mem_cgroup
*memcg
)
2074 struct mem_cgroup
*iter
, *failed
= NULL
;
2076 for_each_mem_cgroup_tree(iter
, memcg
) {
2077 if (iter
->oom_lock
) {
2079 * this subtree of our hierarchy is already locked
2080 * so we cannot give a lock.
2083 mem_cgroup_iter_break(memcg
, iter
);
2086 iter
->oom_lock
= true;
2093 * OK, we failed to lock the whole subtree so we have to clean up
2094 * what we set up to the failing subtree
2096 for_each_mem_cgroup_tree(iter
, memcg
) {
2097 if (iter
== failed
) {
2098 mem_cgroup_iter_break(memcg
, iter
);
2101 iter
->oom_lock
= false;
2107 * Has to be called with memcg_oom_lock
2109 static int mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
2111 struct mem_cgroup
*iter
;
2113 for_each_mem_cgroup_tree(iter
, memcg
)
2114 iter
->oom_lock
= false;
2118 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
2120 struct mem_cgroup
*iter
;
2122 for_each_mem_cgroup_tree(iter
, memcg
)
2123 atomic_inc(&iter
->under_oom
);
2126 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
2128 struct mem_cgroup
*iter
;
2131 * When a new child is created while the hierarchy is under oom,
2132 * mem_cgroup_oom_lock() may not be called. We have to use
2133 * atomic_add_unless() here.
2135 for_each_mem_cgroup_tree(iter
, memcg
)
2136 atomic_add_unless(&iter
->under_oom
, -1, 0);
2139 static DEFINE_SPINLOCK(memcg_oom_lock
);
2140 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
2142 struct oom_wait_info
{
2143 struct mem_cgroup
*memcg
;
2147 static int memcg_oom_wake_function(wait_queue_t
*wait
,
2148 unsigned mode
, int sync
, void *arg
)
2150 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
2151 struct mem_cgroup
*oom_wait_memcg
;
2152 struct oom_wait_info
*oom_wait_info
;
2154 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
2155 oom_wait_memcg
= oom_wait_info
->memcg
;
2158 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2159 * Then we can use css_is_ancestor without taking care of RCU.
2161 if (!mem_cgroup_same_or_subtree(oom_wait_memcg
, wake_memcg
)
2162 && !mem_cgroup_same_or_subtree(wake_memcg
, oom_wait_memcg
))
2164 return autoremove_wake_function(wait
, mode
, sync
, arg
);
2167 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
2169 /* for filtering, pass "memcg" as argument. */
2170 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
2173 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
2175 if (memcg
&& atomic_read(&memcg
->under_oom
))
2176 memcg_wakeup_oom(memcg
);
2180 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2182 static bool mem_cgroup_handle_oom(struct mem_cgroup
*memcg
, gfp_t mask
,
2185 struct oom_wait_info owait
;
2186 bool locked
, need_to_kill
;
2188 owait
.memcg
= memcg
;
2189 owait
.wait
.flags
= 0;
2190 owait
.wait
.func
= memcg_oom_wake_function
;
2191 owait
.wait
.private = current
;
2192 INIT_LIST_HEAD(&owait
.wait
.task_list
);
2193 need_to_kill
= true;
2194 mem_cgroup_mark_under_oom(memcg
);
2196 /* At first, try to OOM lock hierarchy under memcg.*/
2197 spin_lock(&memcg_oom_lock
);
2198 locked
= mem_cgroup_oom_lock(memcg
);
2200 * Even if signal_pending(), we can't quit charge() loop without
2201 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2202 * under OOM is always welcomed, use TASK_KILLABLE here.
2204 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2205 if (!locked
|| memcg
->oom_kill_disable
)
2206 need_to_kill
= false;
2208 mem_cgroup_oom_notify(memcg
);
2209 spin_unlock(&memcg_oom_lock
);
2212 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2213 mem_cgroup_out_of_memory(memcg
, mask
, order
);
2216 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2218 spin_lock(&memcg_oom_lock
);
2220 mem_cgroup_oom_unlock(memcg
);
2221 memcg_wakeup_oom(memcg
);
2222 spin_unlock(&memcg_oom_lock
);
2224 mem_cgroup_unmark_under_oom(memcg
);
2226 if (test_thread_flag(TIF_MEMDIE
) || fatal_signal_pending(current
))
2228 /* Give chance to dying process */
2229 schedule_timeout_uninterruptible(1);
2234 * Currently used to update mapped file statistics, but the routine can be
2235 * generalized to update other statistics as well.
2237 * Notes: Race condition
2239 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2240 * it tends to be costly. But considering some conditions, we doesn't need
2241 * to do so _always_.
2243 * Considering "charge", lock_page_cgroup() is not required because all
2244 * file-stat operations happen after a page is attached to radix-tree. There
2245 * are no race with "charge".
2247 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2248 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2249 * if there are race with "uncharge". Statistics itself is properly handled
2252 * Considering "move", this is an only case we see a race. To make the race
2253 * small, we check mm->moving_account and detect there are possibility of race
2254 * If there is, we take a lock.
2257 void __mem_cgroup_begin_update_page_stat(struct page
*page
,
2258 bool *locked
, unsigned long *flags
)
2260 struct mem_cgroup
*memcg
;
2261 struct page_cgroup
*pc
;
2263 pc
= lookup_page_cgroup(page
);
2265 memcg
= pc
->mem_cgroup
;
2266 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2269 * If this memory cgroup is not under account moving, we don't
2270 * need to take move_lock_mem_cgroup(). Because we already hold
2271 * rcu_read_lock(), any calls to move_account will be delayed until
2272 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2274 if (!mem_cgroup_stolen(memcg
))
2277 move_lock_mem_cgroup(memcg
, flags
);
2278 if (memcg
!= pc
->mem_cgroup
|| !PageCgroupUsed(pc
)) {
2279 move_unlock_mem_cgroup(memcg
, flags
);
2285 void __mem_cgroup_end_update_page_stat(struct page
*page
, unsigned long *flags
)
2287 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2290 * It's guaranteed that pc->mem_cgroup never changes while
2291 * lock is held because a routine modifies pc->mem_cgroup
2292 * should take move_lock_mem_cgroup().
2294 move_unlock_mem_cgroup(pc
->mem_cgroup
, flags
);
2297 void mem_cgroup_update_page_stat(struct page
*page
,
2298 enum mem_cgroup_page_stat_item idx
, int val
)
2300 struct mem_cgroup
*memcg
;
2301 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2302 unsigned long uninitialized_var(flags
);
2304 if (mem_cgroup_disabled())
2307 memcg
= pc
->mem_cgroup
;
2308 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2312 case MEMCG_NR_FILE_MAPPED
:
2313 idx
= MEM_CGROUP_STAT_FILE_MAPPED
;
2319 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2323 * size of first charge trial. "32" comes from vmscan.c's magic value.
2324 * TODO: maybe necessary to use big numbers in big irons.
2326 #define CHARGE_BATCH 32U
2327 struct memcg_stock_pcp
{
2328 struct mem_cgroup
*cached
; /* this never be root cgroup */
2329 unsigned int nr_pages
;
2330 struct work_struct work
;
2331 unsigned long flags
;
2332 #define FLUSHING_CACHED_CHARGE 0
2334 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2335 static DEFINE_MUTEX(percpu_charge_mutex
);
2338 * consume_stock: Try to consume stocked charge on this cpu.
2339 * @memcg: memcg to consume from.
2340 * @nr_pages: how many pages to charge.
2342 * The charges will only happen if @memcg matches the current cpu's memcg
2343 * stock, and at least @nr_pages are available in that stock. Failure to
2344 * service an allocation will refill the stock.
2346 * returns true if successful, false otherwise.
2348 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2350 struct memcg_stock_pcp
*stock
;
2353 if (nr_pages
> CHARGE_BATCH
)
2356 stock
= &get_cpu_var(memcg_stock
);
2357 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
)
2358 stock
->nr_pages
-= nr_pages
;
2359 else /* need to call res_counter_charge */
2361 put_cpu_var(memcg_stock
);
2366 * Returns stocks cached in percpu to res_counter and reset cached information.
2368 static void drain_stock(struct memcg_stock_pcp
*stock
)
2370 struct mem_cgroup
*old
= stock
->cached
;
2372 if (stock
->nr_pages
) {
2373 unsigned long bytes
= stock
->nr_pages
* PAGE_SIZE
;
2375 res_counter_uncharge(&old
->res
, bytes
);
2376 if (do_swap_account
)
2377 res_counter_uncharge(&old
->memsw
, bytes
);
2378 stock
->nr_pages
= 0;
2380 stock
->cached
= NULL
;
2384 * This must be called under preempt disabled or must be called by
2385 * a thread which is pinned to local cpu.
2387 static void drain_local_stock(struct work_struct
*dummy
)
2389 struct memcg_stock_pcp
*stock
= &__get_cpu_var(memcg_stock
);
2391 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2394 static void __init
memcg_stock_init(void)
2398 for_each_possible_cpu(cpu
) {
2399 struct memcg_stock_pcp
*stock
=
2400 &per_cpu(memcg_stock
, cpu
);
2401 INIT_WORK(&stock
->work
, drain_local_stock
);
2406 * Cache charges(val) which is from res_counter, to local per_cpu area.
2407 * This will be consumed by consume_stock() function, later.
2409 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2411 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2413 if (stock
->cached
!= memcg
) { /* reset if necessary */
2415 stock
->cached
= memcg
;
2417 stock
->nr_pages
+= nr_pages
;
2418 put_cpu_var(memcg_stock
);
2422 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2423 * of the hierarchy under it. sync flag says whether we should block
2424 * until the work is done.
2426 static void drain_all_stock(struct mem_cgroup
*root_memcg
, bool sync
)
2430 /* Notify other cpus that system-wide "drain" is running */
2433 for_each_online_cpu(cpu
) {
2434 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2435 struct mem_cgroup
*memcg
;
2437 memcg
= stock
->cached
;
2438 if (!memcg
|| !stock
->nr_pages
)
2440 if (!mem_cgroup_same_or_subtree(root_memcg
, memcg
))
2442 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2444 drain_local_stock(&stock
->work
);
2446 schedule_work_on(cpu
, &stock
->work
);
2454 for_each_online_cpu(cpu
) {
2455 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2456 if (test_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
))
2457 flush_work(&stock
->work
);
2464 * Tries to drain stocked charges in other cpus. This function is asynchronous
2465 * and just put a work per cpu for draining localy on each cpu. Caller can
2466 * expects some charges will be back to res_counter later but cannot wait for
2469 static void drain_all_stock_async(struct mem_cgroup
*root_memcg
)
2472 * If someone calls draining, avoid adding more kworker runs.
2474 if (!mutex_trylock(&percpu_charge_mutex
))
2476 drain_all_stock(root_memcg
, false);
2477 mutex_unlock(&percpu_charge_mutex
);
2480 /* This is a synchronous drain interface. */
2481 static void drain_all_stock_sync(struct mem_cgroup
*root_memcg
)
2483 /* called when force_empty is called */
2484 mutex_lock(&percpu_charge_mutex
);
2485 drain_all_stock(root_memcg
, true);
2486 mutex_unlock(&percpu_charge_mutex
);
2490 * This function drains percpu counter value from DEAD cpu and
2491 * move it to local cpu. Note that this function can be preempted.
2493 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2497 spin_lock(&memcg
->pcp_counter_lock
);
2498 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2499 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2501 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2502 memcg
->nocpu_base
.count
[i
] += x
;
2504 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2505 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2507 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2508 memcg
->nocpu_base
.events
[i
] += x
;
2510 spin_unlock(&memcg
->pcp_counter_lock
);
2513 static int __cpuinit
memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2514 unsigned long action
,
2517 int cpu
= (unsigned long)hcpu
;
2518 struct memcg_stock_pcp
*stock
;
2519 struct mem_cgroup
*iter
;
2521 if (action
== CPU_ONLINE
)
2524 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2527 for_each_mem_cgroup(iter
)
2528 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2530 stock
= &per_cpu(memcg_stock
, cpu
);
2536 /* See __mem_cgroup_try_charge() for details */
2538 CHARGE_OK
, /* success */
2539 CHARGE_RETRY
, /* need to retry but retry is not bad */
2540 CHARGE_NOMEM
, /* we can't do more. return -ENOMEM */
2541 CHARGE_WOULDBLOCK
, /* GFP_WAIT wasn't set and no enough res. */
2542 CHARGE_OOM_DIE
, /* the current is killed because of OOM */
2545 static int mem_cgroup_do_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2546 unsigned int nr_pages
, unsigned int min_pages
,
2549 unsigned long csize
= nr_pages
* PAGE_SIZE
;
2550 struct mem_cgroup
*mem_over_limit
;
2551 struct res_counter
*fail_res
;
2552 unsigned long flags
= 0;
2555 ret
= res_counter_charge(&memcg
->res
, csize
, &fail_res
);
2558 if (!do_swap_account
)
2560 ret
= res_counter_charge(&memcg
->memsw
, csize
, &fail_res
);
2564 res_counter_uncharge(&memcg
->res
, csize
);
2565 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
2566 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
2568 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
2570 * Never reclaim on behalf of optional batching, retry with a
2571 * single page instead.
2573 if (nr_pages
> min_pages
)
2574 return CHARGE_RETRY
;
2576 if (!(gfp_mask
& __GFP_WAIT
))
2577 return CHARGE_WOULDBLOCK
;
2579 if (gfp_mask
& __GFP_NORETRY
)
2580 return CHARGE_NOMEM
;
2582 ret
= mem_cgroup_reclaim(mem_over_limit
, gfp_mask
, flags
);
2583 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2584 return CHARGE_RETRY
;
2586 * Even though the limit is exceeded at this point, reclaim
2587 * may have been able to free some pages. Retry the charge
2588 * before killing the task.
2590 * Only for regular pages, though: huge pages are rather
2591 * unlikely to succeed so close to the limit, and we fall back
2592 * to regular pages anyway in case of failure.
2594 if (nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
) && ret
)
2595 return CHARGE_RETRY
;
2598 * At task move, charge accounts can be doubly counted. So, it's
2599 * better to wait until the end of task_move if something is going on.
2601 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2602 return CHARGE_RETRY
;
2604 /* If we don't need to call oom-killer at el, return immediately */
2606 return CHARGE_NOMEM
;
2608 if (!mem_cgroup_handle_oom(mem_over_limit
, gfp_mask
, get_order(csize
)))
2609 return CHARGE_OOM_DIE
;
2611 return CHARGE_RETRY
;
2615 * __mem_cgroup_try_charge() does
2616 * 1. detect memcg to be charged against from passed *mm and *ptr,
2617 * 2. update res_counter
2618 * 3. call memory reclaim if necessary.
2620 * In some special case, if the task is fatal, fatal_signal_pending() or
2621 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2622 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2623 * as possible without any hazards. 2: all pages should have a valid
2624 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2625 * pointer, that is treated as a charge to root_mem_cgroup.
2627 * So __mem_cgroup_try_charge() will return
2628 * 0 ... on success, filling *ptr with a valid memcg pointer.
2629 * -ENOMEM ... charge failure because of resource limits.
2630 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2632 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2633 * the oom-killer can be invoked.
2635 static int __mem_cgroup_try_charge(struct mm_struct
*mm
,
2637 unsigned int nr_pages
,
2638 struct mem_cgroup
**ptr
,
2641 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2642 int nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2643 struct mem_cgroup
*memcg
= NULL
;
2647 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2648 * in system level. So, allow to go ahead dying process in addition to
2651 if (unlikely(test_thread_flag(TIF_MEMDIE
)
2652 || fatal_signal_pending(current
)))
2656 * We always charge the cgroup the mm_struct belongs to.
2657 * The mm_struct's mem_cgroup changes on task migration if the
2658 * thread group leader migrates. It's possible that mm is not
2659 * set, if so charge the root memcg (happens for pagecache usage).
2662 *ptr
= root_mem_cgroup
;
2664 if (*ptr
) { /* css should be a valid one */
2666 if (mem_cgroup_is_root(memcg
))
2668 if (consume_stock(memcg
, nr_pages
))
2670 css_get(&memcg
->css
);
2672 struct task_struct
*p
;
2675 p
= rcu_dereference(mm
->owner
);
2677 * Because we don't have task_lock(), "p" can exit.
2678 * In that case, "memcg" can point to root or p can be NULL with
2679 * race with swapoff. Then, we have small risk of mis-accouning.
2680 * But such kind of mis-account by race always happens because
2681 * we don't have cgroup_mutex(). It's overkill and we allo that
2683 * (*) swapoff at el will charge against mm-struct not against
2684 * task-struct. So, mm->owner can be NULL.
2686 memcg
= mem_cgroup_from_task(p
);
2688 memcg
= root_mem_cgroup
;
2689 if (mem_cgroup_is_root(memcg
)) {
2693 if (consume_stock(memcg
, nr_pages
)) {
2695 * It seems dagerous to access memcg without css_get().
2696 * But considering how consume_stok works, it's not
2697 * necessary. If consume_stock success, some charges
2698 * from this memcg are cached on this cpu. So, we
2699 * don't need to call css_get()/css_tryget() before
2700 * calling consume_stock().
2705 /* after here, we may be blocked. we need to get refcnt */
2706 if (!css_tryget(&memcg
->css
)) {
2716 /* If killed, bypass charge */
2717 if (fatal_signal_pending(current
)) {
2718 css_put(&memcg
->css
);
2723 if (oom
&& !nr_oom_retries
) {
2725 nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2728 ret
= mem_cgroup_do_charge(memcg
, gfp_mask
, batch
, nr_pages
,
2733 case CHARGE_RETRY
: /* not in OOM situation but retry */
2735 css_put(&memcg
->css
);
2738 case CHARGE_WOULDBLOCK
: /* !__GFP_WAIT */
2739 css_put(&memcg
->css
);
2741 case CHARGE_NOMEM
: /* OOM routine works */
2743 css_put(&memcg
->css
);
2746 /* If oom, we never return -ENOMEM */
2749 case CHARGE_OOM_DIE
: /* Killed by OOM Killer */
2750 css_put(&memcg
->css
);
2753 } while (ret
!= CHARGE_OK
);
2755 if (batch
> nr_pages
)
2756 refill_stock(memcg
, batch
- nr_pages
);
2757 css_put(&memcg
->css
);
2765 *ptr
= root_mem_cgroup
;
2770 * Somemtimes we have to undo a charge we got by try_charge().
2771 * This function is for that and do uncharge, put css's refcnt.
2772 * gotten by try_charge().
2774 static void __mem_cgroup_cancel_charge(struct mem_cgroup
*memcg
,
2775 unsigned int nr_pages
)
2777 if (!mem_cgroup_is_root(memcg
)) {
2778 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2780 res_counter_uncharge(&memcg
->res
, bytes
);
2781 if (do_swap_account
)
2782 res_counter_uncharge(&memcg
->memsw
, bytes
);
2787 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2788 * This is useful when moving usage to parent cgroup.
2790 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup
*memcg
,
2791 unsigned int nr_pages
)
2793 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2795 if (mem_cgroup_is_root(memcg
))
2798 res_counter_uncharge_until(&memcg
->res
, memcg
->res
.parent
, bytes
);
2799 if (do_swap_account
)
2800 res_counter_uncharge_until(&memcg
->memsw
,
2801 memcg
->memsw
.parent
, bytes
);
2805 * A helper function to get mem_cgroup from ID. must be called under
2806 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2807 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2808 * called against removed memcg.)
2810 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2812 struct cgroup_subsys_state
*css
;
2814 /* ID 0 is unused ID */
2817 css
= css_lookup(&mem_cgroup_subsys
, id
);
2820 return mem_cgroup_from_css(css
);
2823 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2825 struct mem_cgroup
*memcg
= NULL
;
2826 struct page_cgroup
*pc
;
2830 VM_BUG_ON(!PageLocked(page
));
2832 pc
= lookup_page_cgroup(page
);
2833 lock_page_cgroup(pc
);
2834 if (PageCgroupUsed(pc
)) {
2835 memcg
= pc
->mem_cgroup
;
2836 if (memcg
&& !css_tryget(&memcg
->css
))
2838 } else if (PageSwapCache(page
)) {
2839 ent
.val
= page_private(page
);
2840 id
= lookup_swap_cgroup_id(ent
);
2842 memcg
= mem_cgroup_lookup(id
);
2843 if (memcg
&& !css_tryget(&memcg
->css
))
2847 unlock_page_cgroup(pc
);
2851 static void __mem_cgroup_commit_charge(struct mem_cgroup
*memcg
,
2853 unsigned int nr_pages
,
2854 enum charge_type ctype
,
2857 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2858 struct zone
*uninitialized_var(zone
);
2859 struct lruvec
*lruvec
;
2860 bool was_on_lru
= false;
2863 lock_page_cgroup(pc
);
2864 VM_BUG_ON(PageCgroupUsed(pc
));
2866 * we don't need page_cgroup_lock about tail pages, becase they are not
2867 * accessed by any other context at this point.
2871 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2872 * may already be on some other mem_cgroup's LRU. Take care of it.
2875 zone
= page_zone(page
);
2876 spin_lock_irq(&zone
->lru_lock
);
2877 if (PageLRU(page
)) {
2878 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2880 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2885 pc
->mem_cgroup
= memcg
;
2887 * We access a page_cgroup asynchronously without lock_page_cgroup().
2888 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2889 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2890 * before USED bit, we need memory barrier here.
2891 * See mem_cgroup_add_lru_list(), etc.
2894 SetPageCgroupUsed(pc
);
2898 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2899 VM_BUG_ON(PageLRU(page
));
2901 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2903 spin_unlock_irq(&zone
->lru_lock
);
2906 if (ctype
== MEM_CGROUP_CHARGE_TYPE_ANON
)
2911 mem_cgroup_charge_statistics(memcg
, page
, anon
, nr_pages
);
2912 unlock_page_cgroup(pc
);
2915 * "charge_statistics" updated event counter. Then, check it.
2916 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2917 * if they exceeds softlimit.
2919 memcg_check_events(memcg
, page
);
2922 static DEFINE_MUTEX(set_limit_mutex
);
2924 #ifdef CONFIG_MEMCG_KMEM
2925 static inline bool memcg_can_account_kmem(struct mem_cgroup
*memcg
)
2927 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
) &&
2928 (memcg
->kmem_account_flags
& KMEM_ACCOUNTED_MASK
);
2932 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2933 * in the memcg_cache_params struct.
2935 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
2937 struct kmem_cache
*cachep
;
2939 VM_BUG_ON(p
->is_root_cache
);
2940 cachep
= p
->root_cache
;
2941 return cachep
->memcg_params
->memcg_caches
[memcg_cache_id(p
->memcg
)];
2944 #ifdef CONFIG_SLABINFO
2945 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state
*css
,
2946 struct cftype
*cft
, struct seq_file
*m
)
2948 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2949 struct memcg_cache_params
*params
;
2951 if (!memcg_can_account_kmem(memcg
))
2954 print_slabinfo_header(m
);
2956 mutex_lock(&memcg
->slab_caches_mutex
);
2957 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
)
2958 cache_show(memcg_params_to_cache(params
), m
);
2959 mutex_unlock(&memcg
->slab_caches_mutex
);
2965 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
, u64 size
)
2967 struct res_counter
*fail_res
;
2968 struct mem_cgroup
*_memcg
;
2972 ret
= res_counter_charge(&memcg
->kmem
, size
, &fail_res
);
2977 * Conditions under which we can wait for the oom_killer. Those are
2978 * the same conditions tested by the core page allocator
2980 may_oom
= (gfp
& __GFP_FS
) && !(gfp
& __GFP_NORETRY
);
2983 ret
= __mem_cgroup_try_charge(NULL
, gfp
, size
>> PAGE_SHIFT
,
2986 if (ret
== -EINTR
) {
2988 * __mem_cgroup_try_charge() chosed to bypass to root due to
2989 * OOM kill or fatal signal. Since our only options are to
2990 * either fail the allocation or charge it to this cgroup, do
2991 * it as a temporary condition. But we can't fail. From a
2992 * kmem/slab perspective, the cache has already been selected,
2993 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2996 * This condition will only trigger if the task entered
2997 * memcg_charge_kmem in a sane state, but was OOM-killed during
2998 * __mem_cgroup_try_charge() above. Tasks that were already
2999 * dying when the allocation triggers should have been already
3000 * directed to the root cgroup in memcontrol.h
3002 res_counter_charge_nofail(&memcg
->res
, size
, &fail_res
);
3003 if (do_swap_account
)
3004 res_counter_charge_nofail(&memcg
->memsw
, size
,
3008 res_counter_uncharge(&memcg
->kmem
, size
);
3013 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, u64 size
)
3015 res_counter_uncharge(&memcg
->res
, size
);
3016 if (do_swap_account
)
3017 res_counter_uncharge(&memcg
->memsw
, size
);
3020 if (res_counter_uncharge(&memcg
->kmem
, size
))
3024 * Releases a reference taken in kmem_cgroup_css_offline in case
3025 * this last uncharge is racing with the offlining code or it is
3026 * outliving the memcg existence.
3028 * The memory barrier imposed by test&clear is paired with the
3029 * explicit one in memcg_kmem_mark_dead().
3031 if (memcg_kmem_test_and_clear_dead(memcg
))
3032 css_put(&memcg
->css
);
3035 void memcg_cache_list_add(struct mem_cgroup
*memcg
, struct kmem_cache
*cachep
)
3040 mutex_lock(&memcg
->slab_caches_mutex
);
3041 list_add(&cachep
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
3042 mutex_unlock(&memcg
->slab_caches_mutex
);
3046 * helper for acessing a memcg's index. It will be used as an index in the
3047 * child cache array in kmem_cache, and also to derive its name. This function
3048 * will return -1 when this is not a kmem-limited memcg.
3050 int memcg_cache_id(struct mem_cgroup
*memcg
)
3052 return memcg
? memcg
->kmemcg_id
: -1;
3056 * This ends up being protected by the set_limit mutex, during normal
3057 * operation, because that is its main call site.
3059 * But when we create a new cache, we can call this as well if its parent
3060 * is kmem-limited. That will have to hold set_limit_mutex as well.
3062 int memcg_update_cache_sizes(struct mem_cgroup
*memcg
)
3066 num
= ida_simple_get(&kmem_limited_groups
,
3067 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
3071 * After this point, kmem_accounted (that we test atomically in
3072 * the beginning of this conditional), is no longer 0. This
3073 * guarantees only one process will set the following boolean
3074 * to true. We don't need test_and_set because we're protected
3075 * by the set_limit_mutex anyway.
3077 memcg_kmem_set_activated(memcg
);
3079 ret
= memcg_update_all_caches(num
+1);
3081 ida_simple_remove(&kmem_limited_groups
, num
);
3082 memcg_kmem_clear_activated(memcg
);
3086 memcg
->kmemcg_id
= num
;
3087 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
3088 mutex_init(&memcg
->slab_caches_mutex
);
3092 static size_t memcg_caches_array_size(int num_groups
)
3095 if (num_groups
<= 0)
3098 size
= 2 * num_groups
;
3099 if (size
< MEMCG_CACHES_MIN_SIZE
)
3100 size
= MEMCG_CACHES_MIN_SIZE
;
3101 else if (size
> MEMCG_CACHES_MAX_SIZE
)
3102 size
= MEMCG_CACHES_MAX_SIZE
;
3108 * We should update the current array size iff all caches updates succeed. This
3109 * can only be done from the slab side. The slab mutex needs to be held when
3112 void memcg_update_array_size(int num
)
3114 if (num
> memcg_limited_groups_array_size
)
3115 memcg_limited_groups_array_size
= memcg_caches_array_size(num
);
3118 static void kmem_cache_destroy_work_func(struct work_struct
*w
);
3120 int memcg_update_cache_size(struct kmem_cache
*s
, int num_groups
)
3122 struct memcg_cache_params
*cur_params
= s
->memcg_params
;
3124 VM_BUG_ON(s
->memcg_params
&& !s
->memcg_params
->is_root_cache
);
3126 if (num_groups
> memcg_limited_groups_array_size
) {
3128 ssize_t size
= memcg_caches_array_size(num_groups
);
3130 size
*= sizeof(void *);
3131 size
+= sizeof(struct memcg_cache_params
);
3133 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3134 if (!s
->memcg_params
) {
3135 s
->memcg_params
= cur_params
;
3139 s
->memcg_params
->is_root_cache
= true;
3142 * There is the chance it will be bigger than
3143 * memcg_limited_groups_array_size, if we failed an allocation
3144 * in a cache, in which case all caches updated before it, will
3145 * have a bigger array.
3147 * But if that is the case, the data after
3148 * memcg_limited_groups_array_size is certainly unused
3150 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3151 if (!cur_params
->memcg_caches
[i
])
3153 s
->memcg_params
->memcg_caches
[i
] =
3154 cur_params
->memcg_caches
[i
];
3158 * Ideally, we would wait until all caches succeed, and only
3159 * then free the old one. But this is not worth the extra
3160 * pointer per-cache we'd have to have for this.
3162 * It is not a big deal if some caches are left with a size
3163 * bigger than the others. And all updates will reset this
3171 int memcg_register_cache(struct mem_cgroup
*memcg
, struct kmem_cache
*s
,
3172 struct kmem_cache
*root_cache
)
3174 size_t size
= sizeof(struct memcg_cache_params
);
3176 if (!memcg_kmem_enabled())
3180 size
+= memcg_limited_groups_array_size
* sizeof(void *);
3182 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3183 if (!s
->memcg_params
)
3186 INIT_WORK(&s
->memcg_params
->destroy
,
3187 kmem_cache_destroy_work_func
);
3189 s
->memcg_params
->memcg
= memcg
;
3190 s
->memcg_params
->root_cache
= root_cache
;
3192 s
->memcg_params
->is_root_cache
= true;
3197 void memcg_release_cache(struct kmem_cache
*s
)
3199 struct kmem_cache
*root
;
3200 struct mem_cgroup
*memcg
;
3204 * This happens, for instance, when a root cache goes away before we
3207 if (!s
->memcg_params
)
3210 if (s
->memcg_params
->is_root_cache
)
3213 memcg
= s
->memcg_params
->memcg
;
3214 id
= memcg_cache_id(memcg
);
3216 root
= s
->memcg_params
->root_cache
;
3217 root
->memcg_params
->memcg_caches
[id
] = NULL
;
3219 mutex_lock(&memcg
->slab_caches_mutex
);
3220 list_del(&s
->memcg_params
->list
);
3221 mutex_unlock(&memcg
->slab_caches_mutex
);
3223 css_put(&memcg
->css
);
3225 kfree(s
->memcg_params
);
3229 * During the creation a new cache, we need to disable our accounting mechanism
3230 * altogether. This is true even if we are not creating, but rather just
3231 * enqueing new caches to be created.
3233 * This is because that process will trigger allocations; some visible, like
3234 * explicit kmallocs to auxiliary data structures, name strings and internal
3235 * cache structures; some well concealed, like INIT_WORK() that can allocate
3236 * objects during debug.
3238 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3239 * to it. This may not be a bounded recursion: since the first cache creation
3240 * failed to complete (waiting on the allocation), we'll just try to create the
3241 * cache again, failing at the same point.
3243 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3244 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3245 * inside the following two functions.
3247 static inline void memcg_stop_kmem_account(void)
3249 VM_BUG_ON(!current
->mm
);
3250 current
->memcg_kmem_skip_account
++;
3253 static inline void memcg_resume_kmem_account(void)
3255 VM_BUG_ON(!current
->mm
);
3256 current
->memcg_kmem_skip_account
--;
3259 static void kmem_cache_destroy_work_func(struct work_struct
*w
)
3261 struct kmem_cache
*cachep
;
3262 struct memcg_cache_params
*p
;
3264 p
= container_of(w
, struct memcg_cache_params
, destroy
);
3266 cachep
= memcg_params_to_cache(p
);
3269 * If we get down to 0 after shrink, we could delete right away.
3270 * However, memcg_release_pages() already puts us back in the workqueue
3271 * in that case. If we proceed deleting, we'll get a dangling
3272 * reference, and removing the object from the workqueue in that case
3273 * is unnecessary complication. We are not a fast path.
3275 * Note that this case is fundamentally different from racing with
3276 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3277 * kmem_cache_shrink, not only we would be reinserting a dead cache
3278 * into the queue, but doing so from inside the worker racing to
3281 * So if we aren't down to zero, we'll just schedule a worker and try
3284 if (atomic_read(&cachep
->memcg_params
->nr_pages
) != 0) {
3285 kmem_cache_shrink(cachep
);
3286 if (atomic_read(&cachep
->memcg_params
->nr_pages
) == 0)
3289 kmem_cache_destroy(cachep
);
3292 void mem_cgroup_destroy_cache(struct kmem_cache
*cachep
)
3294 if (!cachep
->memcg_params
->dead
)
3298 * There are many ways in which we can get here.
3300 * We can get to a memory-pressure situation while the delayed work is
3301 * still pending to run. The vmscan shrinkers can then release all
3302 * cache memory and get us to destruction. If this is the case, we'll
3303 * be executed twice, which is a bug (the second time will execute over
3304 * bogus data). In this case, cancelling the work should be fine.
3306 * But we can also get here from the worker itself, if
3307 * kmem_cache_shrink is enough to shake all the remaining objects and
3308 * get the page count to 0. In this case, we'll deadlock if we try to
3309 * cancel the work (the worker runs with an internal lock held, which
3310 * is the same lock we would hold for cancel_work_sync().)
3312 * Since we can't possibly know who got us here, just refrain from
3313 * running if there is already work pending
3315 if (work_pending(&cachep
->memcg_params
->destroy
))
3318 * We have to defer the actual destroying to a workqueue, because
3319 * we might currently be in a context that cannot sleep.
3321 schedule_work(&cachep
->memcg_params
->destroy
);
3325 * This lock protects updaters, not readers. We want readers to be as fast as
3326 * they can, and they will either see NULL or a valid cache value. Our model
3327 * allow them to see NULL, in which case the root memcg will be selected.
3329 * We need this lock because multiple allocations to the same cache from a non
3330 * will span more than one worker. Only one of them can create the cache.
3332 static DEFINE_MUTEX(memcg_cache_mutex
);
3335 * Called with memcg_cache_mutex held
3337 static struct kmem_cache
*kmem_cache_dup(struct mem_cgroup
*memcg
,
3338 struct kmem_cache
*s
)
3340 struct kmem_cache
*new;
3341 static char *tmp_name
= NULL
;
3343 lockdep_assert_held(&memcg_cache_mutex
);
3346 * kmem_cache_create_memcg duplicates the given name and
3347 * cgroup_name for this name requires RCU context.
3348 * This static temporary buffer is used to prevent from
3349 * pointless shortliving allocation.
3352 tmp_name
= kmalloc(PATH_MAX
, GFP_KERNEL
);
3358 snprintf(tmp_name
, PATH_MAX
, "%s(%d:%s)", s
->name
,
3359 memcg_cache_id(memcg
), cgroup_name(memcg
->css
.cgroup
));
3362 new = kmem_cache_create_memcg(memcg
, tmp_name
, s
->object_size
, s
->align
,
3363 (s
->flags
& ~SLAB_PANIC
), s
->ctor
, s
);
3366 new->allocflags
|= __GFP_KMEMCG
;
3371 static struct kmem_cache
*memcg_create_kmem_cache(struct mem_cgroup
*memcg
,
3372 struct kmem_cache
*cachep
)
3374 struct kmem_cache
*new_cachep
;
3377 BUG_ON(!memcg_can_account_kmem(memcg
));
3379 idx
= memcg_cache_id(memcg
);
3381 mutex_lock(&memcg_cache_mutex
);
3382 new_cachep
= cachep
->memcg_params
->memcg_caches
[idx
];
3384 css_put(&memcg
->css
);
3388 new_cachep
= kmem_cache_dup(memcg
, cachep
);
3389 if (new_cachep
== NULL
) {
3390 new_cachep
= cachep
;
3391 css_put(&memcg
->css
);
3395 atomic_set(&new_cachep
->memcg_params
->nr_pages
, 0);
3397 cachep
->memcg_params
->memcg_caches
[idx
] = new_cachep
;
3399 * the readers won't lock, make sure everybody sees the updated value,
3400 * so they won't put stuff in the queue again for no reason
3404 mutex_unlock(&memcg_cache_mutex
);
3408 void kmem_cache_destroy_memcg_children(struct kmem_cache
*s
)
3410 struct kmem_cache
*c
;
3413 if (!s
->memcg_params
)
3415 if (!s
->memcg_params
->is_root_cache
)
3419 * If the cache is being destroyed, we trust that there is no one else
3420 * requesting objects from it. Even if there are, the sanity checks in
3421 * kmem_cache_destroy should caught this ill-case.
3423 * Still, we don't want anyone else freeing memcg_caches under our
3424 * noses, which can happen if a new memcg comes to life. As usual,
3425 * we'll take the set_limit_mutex to protect ourselves against this.
3427 mutex_lock(&set_limit_mutex
);
3428 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3429 c
= s
->memcg_params
->memcg_caches
[i
];
3434 * We will now manually delete the caches, so to avoid races
3435 * we need to cancel all pending destruction workers and
3436 * proceed with destruction ourselves.
3438 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3439 * and that could spawn the workers again: it is likely that
3440 * the cache still have active pages until this very moment.
3441 * This would lead us back to mem_cgroup_destroy_cache.
3443 * But that will not execute at all if the "dead" flag is not
3444 * set, so flip it down to guarantee we are in control.
3446 c
->memcg_params
->dead
= false;
3447 cancel_work_sync(&c
->memcg_params
->destroy
);
3448 kmem_cache_destroy(c
);
3450 mutex_unlock(&set_limit_mutex
);
3453 struct create_work
{
3454 struct mem_cgroup
*memcg
;
3455 struct kmem_cache
*cachep
;
3456 struct work_struct work
;
3459 static void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3461 struct kmem_cache
*cachep
;
3462 struct memcg_cache_params
*params
;
3464 if (!memcg_kmem_is_active(memcg
))
3467 mutex_lock(&memcg
->slab_caches_mutex
);
3468 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
) {
3469 cachep
= memcg_params_to_cache(params
);
3470 cachep
->memcg_params
->dead
= true;
3471 schedule_work(&cachep
->memcg_params
->destroy
);
3473 mutex_unlock(&memcg
->slab_caches_mutex
);
3476 static void memcg_create_cache_work_func(struct work_struct
*w
)
3478 struct create_work
*cw
;
3480 cw
= container_of(w
, struct create_work
, work
);
3481 memcg_create_kmem_cache(cw
->memcg
, cw
->cachep
);
3486 * Enqueue the creation of a per-memcg kmem_cache.
3488 static void __memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3489 struct kmem_cache
*cachep
)
3491 struct create_work
*cw
;
3493 cw
= kmalloc(sizeof(struct create_work
), GFP_NOWAIT
);
3495 css_put(&memcg
->css
);
3500 cw
->cachep
= cachep
;
3502 INIT_WORK(&cw
->work
, memcg_create_cache_work_func
);
3503 schedule_work(&cw
->work
);
3506 static void memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3507 struct kmem_cache
*cachep
)
3510 * We need to stop accounting when we kmalloc, because if the
3511 * corresponding kmalloc cache is not yet created, the first allocation
3512 * in __memcg_create_cache_enqueue will recurse.
3514 * However, it is better to enclose the whole function. Depending on
3515 * the debugging options enabled, INIT_WORK(), for instance, can
3516 * trigger an allocation. This too, will make us recurse. Because at
3517 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3518 * the safest choice is to do it like this, wrapping the whole function.
3520 memcg_stop_kmem_account();
3521 __memcg_create_cache_enqueue(memcg
, cachep
);
3522 memcg_resume_kmem_account();
3525 * Return the kmem_cache we're supposed to use for a slab allocation.
3526 * We try to use the current memcg's version of the cache.
3528 * If the cache does not exist yet, if we are the first user of it,
3529 * we either create it immediately, if possible, or create it asynchronously
3531 * In the latter case, we will let the current allocation go through with
3532 * the original cache.
3534 * Can't be called in interrupt context or from kernel threads.
3535 * This function needs to be called with rcu_read_lock() held.
3537 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
3540 struct mem_cgroup
*memcg
;
3543 VM_BUG_ON(!cachep
->memcg_params
);
3544 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
3546 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3550 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
3552 if (!memcg_can_account_kmem(memcg
))
3555 idx
= memcg_cache_id(memcg
);
3558 * barrier to mare sure we're always seeing the up to date value. The
3559 * code updating memcg_caches will issue a write barrier to match this.
3561 read_barrier_depends();
3562 if (likely(cachep
->memcg_params
->memcg_caches
[idx
])) {
3563 cachep
= cachep
->memcg_params
->memcg_caches
[idx
];
3567 /* The corresponding put will be done in the workqueue. */
3568 if (!css_tryget(&memcg
->css
))
3573 * If we are in a safe context (can wait, and not in interrupt
3574 * context), we could be be predictable and return right away.
3575 * This would guarantee that the allocation being performed
3576 * already belongs in the new cache.
3578 * However, there are some clashes that can arrive from locking.
3579 * For instance, because we acquire the slab_mutex while doing
3580 * kmem_cache_dup, this means no further allocation could happen
3581 * with the slab_mutex held.
3583 * Also, because cache creation issue get_online_cpus(), this
3584 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3585 * that ends up reversed during cpu hotplug. (cpuset allocates
3586 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3587 * better to defer everything.
3589 memcg_create_cache_enqueue(memcg
, cachep
);
3595 EXPORT_SYMBOL(__memcg_kmem_get_cache
);
3598 * We need to verify if the allocation against current->mm->owner's memcg is
3599 * possible for the given order. But the page is not allocated yet, so we'll
3600 * need a further commit step to do the final arrangements.
3602 * It is possible for the task to switch cgroups in this mean time, so at
3603 * commit time, we can't rely on task conversion any longer. We'll then use
3604 * the handle argument to return to the caller which cgroup we should commit
3605 * against. We could also return the memcg directly and avoid the pointer
3606 * passing, but a boolean return value gives better semantics considering
3607 * the compiled-out case as well.
3609 * Returning true means the allocation is possible.
3612 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
3614 struct mem_cgroup
*memcg
;
3620 * Disabling accounting is only relevant for some specific memcg
3621 * internal allocations. Therefore we would initially not have such
3622 * check here, since direct calls to the page allocator that are marked
3623 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3624 * concerned with cache allocations, and by having this test at
3625 * memcg_kmem_get_cache, we are already able to relay the allocation to
3626 * the root cache and bypass the memcg cache altogether.
3628 * There is one exception, though: the SLUB allocator does not create
3629 * large order caches, but rather service large kmallocs directly from
3630 * the page allocator. Therefore, the following sequence when backed by
3631 * the SLUB allocator:
3633 * memcg_stop_kmem_account();
3634 * kmalloc(<large_number>)
3635 * memcg_resume_kmem_account();
3637 * would effectively ignore the fact that we should skip accounting,
3638 * since it will drive us directly to this function without passing
3639 * through the cache selector memcg_kmem_get_cache. Such large
3640 * allocations are extremely rare but can happen, for instance, for the
3641 * cache arrays. We bring this test here.
3643 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3646 memcg
= try_get_mem_cgroup_from_mm(current
->mm
);
3649 * very rare case described in mem_cgroup_from_task. Unfortunately there
3650 * isn't much we can do without complicating this too much, and it would
3651 * be gfp-dependent anyway. Just let it go
3653 if (unlikely(!memcg
))
3656 if (!memcg_can_account_kmem(memcg
)) {
3657 css_put(&memcg
->css
);
3661 ret
= memcg_charge_kmem(memcg
, gfp
, PAGE_SIZE
<< order
);
3665 css_put(&memcg
->css
);
3669 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
3672 struct page_cgroup
*pc
;
3674 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3676 /* The page allocation failed. Revert */
3678 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3682 pc
= lookup_page_cgroup(page
);
3683 lock_page_cgroup(pc
);
3684 pc
->mem_cgroup
= memcg
;
3685 SetPageCgroupUsed(pc
);
3686 unlock_page_cgroup(pc
);
3689 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
3691 struct mem_cgroup
*memcg
= NULL
;
3692 struct page_cgroup
*pc
;
3695 pc
= lookup_page_cgroup(page
);
3697 * Fast unlocked return. Theoretically might have changed, have to
3698 * check again after locking.
3700 if (!PageCgroupUsed(pc
))
3703 lock_page_cgroup(pc
);
3704 if (PageCgroupUsed(pc
)) {
3705 memcg
= pc
->mem_cgroup
;
3706 ClearPageCgroupUsed(pc
);
3708 unlock_page_cgroup(pc
);
3711 * We trust that only if there is a memcg associated with the page, it
3712 * is a valid allocation
3717 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3718 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3721 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3724 #endif /* CONFIG_MEMCG_KMEM */
3726 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3728 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3730 * Because tail pages are not marked as "used", set it. We're under
3731 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3732 * charge/uncharge will be never happen and move_account() is done under
3733 * compound_lock(), so we don't have to take care of races.
3735 void mem_cgroup_split_huge_fixup(struct page
*head
)
3737 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
3738 struct page_cgroup
*pc
;
3739 struct mem_cgroup
*memcg
;
3742 if (mem_cgroup_disabled())
3745 memcg
= head_pc
->mem_cgroup
;
3746 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3748 pc
->mem_cgroup
= memcg
;
3749 smp_wmb();/* see __commit_charge() */
3750 pc
->flags
= head_pc
->flags
& ~PCGF_NOCOPY_AT_SPLIT
;
3752 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
3755 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3758 * mem_cgroup_move_account - move account of the page
3760 * @nr_pages: number of regular pages (>1 for huge pages)
3761 * @pc: page_cgroup of the page.
3762 * @from: mem_cgroup which the page is moved from.
3763 * @to: mem_cgroup which the page is moved to. @from != @to.
3765 * The caller must confirm following.
3766 * - page is not on LRU (isolate_page() is useful.)
3767 * - compound_lock is held when nr_pages > 1
3769 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3772 static int mem_cgroup_move_account(struct page
*page
,
3773 unsigned int nr_pages
,
3774 struct page_cgroup
*pc
,
3775 struct mem_cgroup
*from
,
3776 struct mem_cgroup
*to
)
3778 unsigned long flags
;
3780 bool anon
= PageAnon(page
);
3782 VM_BUG_ON(from
== to
);
3783 VM_BUG_ON(PageLRU(page
));
3785 * The page is isolated from LRU. So, collapse function
3786 * will not handle this page. But page splitting can happen.
3787 * Do this check under compound_page_lock(). The caller should
3791 if (nr_pages
> 1 && !PageTransHuge(page
))
3794 lock_page_cgroup(pc
);
3797 if (!PageCgroupUsed(pc
) || pc
->mem_cgroup
!= from
)
3800 move_lock_mem_cgroup(from
, &flags
);
3802 if (!anon
&& page_mapped(page
)) {
3803 /* Update mapped_file data for mem_cgroup */
3805 __this_cpu_dec(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
3806 __this_cpu_inc(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
3809 mem_cgroup_charge_statistics(from
, page
, anon
, -nr_pages
);
3811 /* caller should have done css_get */
3812 pc
->mem_cgroup
= to
;
3813 mem_cgroup_charge_statistics(to
, page
, anon
, nr_pages
);
3814 move_unlock_mem_cgroup(from
, &flags
);
3817 unlock_page_cgroup(pc
);
3821 memcg_check_events(to
, page
);
3822 memcg_check_events(from
, page
);
3828 * mem_cgroup_move_parent - moves page to the parent group
3829 * @page: the page to move
3830 * @pc: page_cgroup of the page
3831 * @child: page's cgroup
3833 * move charges to its parent or the root cgroup if the group has no
3834 * parent (aka use_hierarchy==0).
3835 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3836 * mem_cgroup_move_account fails) the failure is always temporary and
3837 * it signals a race with a page removal/uncharge or migration. In the
3838 * first case the page is on the way out and it will vanish from the LRU
3839 * on the next attempt and the call should be retried later.
3840 * Isolation from the LRU fails only if page has been isolated from
3841 * the LRU since we looked at it and that usually means either global
3842 * reclaim or migration going on. The page will either get back to the
3844 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3845 * (!PageCgroupUsed) or moved to a different group. The page will
3846 * disappear in the next attempt.
3848 static int mem_cgroup_move_parent(struct page
*page
,
3849 struct page_cgroup
*pc
,
3850 struct mem_cgroup
*child
)
3852 struct mem_cgroup
*parent
;
3853 unsigned int nr_pages
;
3854 unsigned long uninitialized_var(flags
);
3857 VM_BUG_ON(mem_cgroup_is_root(child
));
3860 if (!get_page_unless_zero(page
))
3862 if (isolate_lru_page(page
))
3865 nr_pages
= hpage_nr_pages(page
);
3867 parent
= parent_mem_cgroup(child
);
3869 * If no parent, move charges to root cgroup.
3872 parent
= root_mem_cgroup
;
3875 VM_BUG_ON(!PageTransHuge(page
));
3876 flags
= compound_lock_irqsave(page
);
3879 ret
= mem_cgroup_move_account(page
, nr_pages
,
3882 __mem_cgroup_cancel_local_charge(child
, nr_pages
);
3885 compound_unlock_irqrestore(page
, flags
);
3886 putback_lru_page(page
);
3894 * Charge the memory controller for page usage.
3896 * 0 if the charge was successful
3897 * < 0 if the cgroup is over its limit
3899 static int mem_cgroup_charge_common(struct page
*page
, struct mm_struct
*mm
,
3900 gfp_t gfp_mask
, enum charge_type ctype
)
3902 struct mem_cgroup
*memcg
= NULL
;
3903 unsigned int nr_pages
= 1;
3907 if (PageTransHuge(page
)) {
3908 nr_pages
<<= compound_order(page
);
3909 VM_BUG_ON(!PageTransHuge(page
));
3911 * Never OOM-kill a process for a huge page. The
3912 * fault handler will fall back to regular pages.
3917 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, nr_pages
, &memcg
, oom
);
3920 __mem_cgroup_commit_charge(memcg
, page
, nr_pages
, ctype
, false);
3924 int mem_cgroup_newpage_charge(struct page
*page
,
3925 struct mm_struct
*mm
, gfp_t gfp_mask
)
3927 if (mem_cgroup_disabled())
3929 VM_BUG_ON(page_mapped(page
));
3930 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
3932 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
3933 MEM_CGROUP_CHARGE_TYPE_ANON
);
3937 * While swap-in, try_charge -> commit or cancel, the page is locked.
3938 * And when try_charge() successfully returns, one refcnt to memcg without
3939 * struct page_cgroup is acquired. This refcnt will be consumed by
3940 * "commit()" or removed by "cancel()"
3942 static int __mem_cgroup_try_charge_swapin(struct mm_struct
*mm
,
3945 struct mem_cgroup
**memcgp
)
3947 struct mem_cgroup
*memcg
;
3948 struct page_cgroup
*pc
;
3951 pc
= lookup_page_cgroup(page
);
3953 * Every swap fault against a single page tries to charge the
3954 * page, bail as early as possible. shmem_unuse() encounters
3955 * already charged pages, too. The USED bit is protected by
3956 * the page lock, which serializes swap cache removal, which
3957 * in turn serializes uncharging.
3959 if (PageCgroupUsed(pc
))
3961 if (!do_swap_account
)
3963 memcg
= try_get_mem_cgroup_from_page(page
);
3967 ret
= __mem_cgroup_try_charge(NULL
, mask
, 1, memcgp
, true);
3968 css_put(&memcg
->css
);
3973 ret
= __mem_cgroup_try_charge(mm
, mask
, 1, memcgp
, true);
3979 int mem_cgroup_try_charge_swapin(struct mm_struct
*mm
, struct page
*page
,
3980 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
3983 if (mem_cgroup_disabled())
3986 * A racing thread's fault, or swapoff, may have already
3987 * updated the pte, and even removed page from swap cache: in
3988 * those cases unuse_pte()'s pte_same() test will fail; but
3989 * there's also a KSM case which does need to charge the page.
3991 if (!PageSwapCache(page
)) {
3994 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, 1, memcgp
, true);
3999 return __mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, memcgp
);
4002 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup
*memcg
)
4004 if (mem_cgroup_disabled())
4008 __mem_cgroup_cancel_charge(memcg
, 1);
4012 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*memcg
,
4013 enum charge_type ctype
)
4015 if (mem_cgroup_disabled())
4020 __mem_cgroup_commit_charge(memcg
, page
, 1, ctype
, true);
4022 * Now swap is on-memory. This means this page may be
4023 * counted both as mem and swap....double count.
4024 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4025 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4026 * may call delete_from_swap_cache() before reach here.
4028 if (do_swap_account
&& PageSwapCache(page
)) {
4029 swp_entry_t ent
= {.val
= page_private(page
)};
4030 mem_cgroup_uncharge_swap(ent
);
4034 void mem_cgroup_commit_charge_swapin(struct page
*page
,
4035 struct mem_cgroup
*memcg
)
4037 __mem_cgroup_commit_charge_swapin(page
, memcg
,
4038 MEM_CGROUP_CHARGE_TYPE_ANON
);
4041 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
4044 struct mem_cgroup
*memcg
= NULL
;
4045 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4048 if (mem_cgroup_disabled())
4050 if (PageCompound(page
))
4053 if (!PageSwapCache(page
))
4054 ret
= mem_cgroup_charge_common(page
, mm
, gfp_mask
, type
);
4055 else { /* page is swapcache/shmem */
4056 ret
= __mem_cgroup_try_charge_swapin(mm
, page
,
4059 __mem_cgroup_commit_charge_swapin(page
, memcg
, type
);
4064 static void mem_cgroup_do_uncharge(struct mem_cgroup
*memcg
,
4065 unsigned int nr_pages
,
4066 const enum charge_type ctype
)
4068 struct memcg_batch_info
*batch
= NULL
;
4069 bool uncharge_memsw
= true;
4071 /* If swapout, usage of swap doesn't decrease */
4072 if (!do_swap_account
|| ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
)
4073 uncharge_memsw
= false;
4075 batch
= ¤t
->memcg_batch
;
4077 * In usual, we do css_get() when we remember memcg pointer.
4078 * But in this case, we keep res->usage until end of a series of
4079 * uncharges. Then, it's ok to ignore memcg's refcnt.
4082 batch
->memcg
= memcg
;
4084 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4085 * In those cases, all pages freed continuously can be expected to be in
4086 * the same cgroup and we have chance to coalesce uncharges.
4087 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4088 * because we want to do uncharge as soon as possible.
4091 if (!batch
->do_batch
|| test_thread_flag(TIF_MEMDIE
))
4092 goto direct_uncharge
;
4095 goto direct_uncharge
;
4098 * In typical case, batch->memcg == mem. This means we can
4099 * merge a series of uncharges to an uncharge of res_counter.
4100 * If not, we uncharge res_counter ony by one.
4102 if (batch
->memcg
!= memcg
)
4103 goto direct_uncharge
;
4104 /* remember freed charge and uncharge it later */
4107 batch
->memsw_nr_pages
++;
4110 res_counter_uncharge(&memcg
->res
, nr_pages
* PAGE_SIZE
);
4112 res_counter_uncharge(&memcg
->memsw
, nr_pages
* PAGE_SIZE
);
4113 if (unlikely(batch
->memcg
!= memcg
))
4114 memcg_oom_recover(memcg
);
4118 * uncharge if !page_mapped(page)
4120 static struct mem_cgroup
*
4121 __mem_cgroup_uncharge_common(struct page
*page
, enum charge_type ctype
,
4124 struct mem_cgroup
*memcg
= NULL
;
4125 unsigned int nr_pages
= 1;
4126 struct page_cgroup
*pc
;
4129 if (mem_cgroup_disabled())
4132 if (PageTransHuge(page
)) {
4133 nr_pages
<<= compound_order(page
);
4134 VM_BUG_ON(!PageTransHuge(page
));
4137 * Check if our page_cgroup is valid
4139 pc
= lookup_page_cgroup(page
);
4140 if (unlikely(!PageCgroupUsed(pc
)))
4143 lock_page_cgroup(pc
);
4145 memcg
= pc
->mem_cgroup
;
4147 if (!PageCgroupUsed(pc
))
4150 anon
= PageAnon(page
);
4153 case MEM_CGROUP_CHARGE_TYPE_ANON
:
4155 * Generally PageAnon tells if it's the anon statistics to be
4156 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4157 * used before page reached the stage of being marked PageAnon.
4161 case MEM_CGROUP_CHARGE_TYPE_DROP
:
4162 /* See mem_cgroup_prepare_migration() */
4163 if (page_mapped(page
))
4166 * Pages under migration may not be uncharged. But
4167 * end_migration() /must/ be the one uncharging the
4168 * unused post-migration page and so it has to call
4169 * here with the migration bit still set. See the
4170 * res_counter handling below.
4172 if (!end_migration
&& PageCgroupMigration(pc
))
4175 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT
:
4176 if (!PageAnon(page
)) { /* Shared memory */
4177 if (page
->mapping
&& !page_is_file_cache(page
))
4179 } else if (page_mapped(page
)) /* Anon */
4186 mem_cgroup_charge_statistics(memcg
, page
, anon
, -nr_pages
);
4188 ClearPageCgroupUsed(pc
);
4190 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4191 * freed from LRU. This is safe because uncharged page is expected not
4192 * to be reused (freed soon). Exception is SwapCache, it's handled by
4193 * special functions.
4196 unlock_page_cgroup(pc
);
4198 * even after unlock, we have memcg->res.usage here and this memcg
4199 * will never be freed, so it's safe to call css_get().
4201 memcg_check_events(memcg
, page
);
4202 if (do_swap_account
&& ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
) {
4203 mem_cgroup_swap_statistics(memcg
, true);
4204 css_get(&memcg
->css
);
4207 * Migration does not charge the res_counter for the
4208 * replacement page, so leave it alone when phasing out the
4209 * page that is unused after the migration.
4211 if (!end_migration
&& !mem_cgroup_is_root(memcg
))
4212 mem_cgroup_do_uncharge(memcg
, nr_pages
, ctype
);
4217 unlock_page_cgroup(pc
);
4221 void mem_cgroup_uncharge_page(struct page
*page
)
4224 if (page_mapped(page
))
4226 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
4228 * If the page is in swap cache, uncharge should be deferred
4229 * to the swap path, which also properly accounts swap usage
4230 * and handles memcg lifetime.
4232 * Note that this check is not stable and reclaim may add the
4233 * page to swap cache at any time after this. However, if the
4234 * page is not in swap cache by the time page->mapcount hits
4235 * 0, there won't be any page table references to the swap
4236 * slot, and reclaim will free it and not actually write the
4239 if (PageSwapCache(page
))
4241 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_ANON
, false);
4244 void mem_cgroup_uncharge_cache_page(struct page
*page
)
4246 VM_BUG_ON(page_mapped(page
));
4247 VM_BUG_ON(page
->mapping
);
4248 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_CACHE
, false);
4252 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4253 * In that cases, pages are freed continuously and we can expect pages
4254 * are in the same memcg. All these calls itself limits the number of
4255 * pages freed at once, then uncharge_start/end() is called properly.
4256 * This may be called prural(2) times in a context,
4259 void mem_cgroup_uncharge_start(void)
4261 current
->memcg_batch
.do_batch
++;
4262 /* We can do nest. */
4263 if (current
->memcg_batch
.do_batch
== 1) {
4264 current
->memcg_batch
.memcg
= NULL
;
4265 current
->memcg_batch
.nr_pages
= 0;
4266 current
->memcg_batch
.memsw_nr_pages
= 0;
4270 void mem_cgroup_uncharge_end(void)
4272 struct memcg_batch_info
*batch
= ¤t
->memcg_batch
;
4274 if (!batch
->do_batch
)
4278 if (batch
->do_batch
) /* If stacked, do nothing. */
4284 * This "batch->memcg" is valid without any css_get/put etc...
4285 * bacause we hide charges behind us.
4287 if (batch
->nr_pages
)
4288 res_counter_uncharge(&batch
->memcg
->res
,
4289 batch
->nr_pages
* PAGE_SIZE
);
4290 if (batch
->memsw_nr_pages
)
4291 res_counter_uncharge(&batch
->memcg
->memsw
,
4292 batch
->memsw_nr_pages
* PAGE_SIZE
);
4293 memcg_oom_recover(batch
->memcg
);
4294 /* forget this pointer (for sanity check) */
4295 batch
->memcg
= NULL
;
4300 * called after __delete_from_swap_cache() and drop "page" account.
4301 * memcg information is recorded to swap_cgroup of "ent"
4304 mem_cgroup_uncharge_swapcache(struct page
*page
, swp_entry_t ent
, bool swapout
)
4306 struct mem_cgroup
*memcg
;
4307 int ctype
= MEM_CGROUP_CHARGE_TYPE_SWAPOUT
;
4309 if (!swapout
) /* this was a swap cache but the swap is unused ! */
4310 ctype
= MEM_CGROUP_CHARGE_TYPE_DROP
;
4312 memcg
= __mem_cgroup_uncharge_common(page
, ctype
, false);
4315 * record memcg information, if swapout && memcg != NULL,
4316 * css_get() was called in uncharge().
4318 if (do_swap_account
&& swapout
&& memcg
)
4319 swap_cgroup_record(ent
, css_id(&memcg
->css
));
4323 #ifdef CONFIG_MEMCG_SWAP
4325 * called from swap_entry_free(). remove record in swap_cgroup and
4326 * uncharge "memsw" account.
4328 void mem_cgroup_uncharge_swap(swp_entry_t ent
)
4330 struct mem_cgroup
*memcg
;
4333 if (!do_swap_account
)
4336 id
= swap_cgroup_record(ent
, 0);
4338 memcg
= mem_cgroup_lookup(id
);
4341 * We uncharge this because swap is freed.
4342 * This memcg can be obsolete one. We avoid calling css_tryget
4344 if (!mem_cgroup_is_root(memcg
))
4345 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
4346 mem_cgroup_swap_statistics(memcg
, false);
4347 css_put(&memcg
->css
);
4353 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4354 * @entry: swap entry to be moved
4355 * @from: mem_cgroup which the entry is moved from
4356 * @to: mem_cgroup which the entry is moved to
4358 * It succeeds only when the swap_cgroup's record for this entry is the same
4359 * as the mem_cgroup's id of @from.
4361 * Returns 0 on success, -EINVAL on failure.
4363 * The caller must have charged to @to, IOW, called res_counter_charge() about
4364 * both res and memsw, and called css_get().
4366 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
4367 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4369 unsigned short old_id
, new_id
;
4371 old_id
= css_id(&from
->css
);
4372 new_id
= css_id(&to
->css
);
4374 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
4375 mem_cgroup_swap_statistics(from
, false);
4376 mem_cgroup_swap_statistics(to
, true);
4378 * This function is only called from task migration context now.
4379 * It postpones res_counter and refcount handling till the end
4380 * of task migration(mem_cgroup_clear_mc()) for performance
4381 * improvement. But we cannot postpone css_get(to) because if
4382 * the process that has been moved to @to does swap-in, the
4383 * refcount of @to might be decreased to 0.
4385 * We are in attach() phase, so the cgroup is guaranteed to be
4386 * alive, so we can just call css_get().
4394 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
4395 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4402 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4405 void mem_cgroup_prepare_migration(struct page
*page
, struct page
*newpage
,
4406 struct mem_cgroup
**memcgp
)
4408 struct mem_cgroup
*memcg
= NULL
;
4409 unsigned int nr_pages
= 1;
4410 struct page_cgroup
*pc
;
4411 enum charge_type ctype
;
4415 if (mem_cgroup_disabled())
4418 if (PageTransHuge(page
))
4419 nr_pages
<<= compound_order(page
);
4421 pc
= lookup_page_cgroup(page
);
4422 lock_page_cgroup(pc
);
4423 if (PageCgroupUsed(pc
)) {
4424 memcg
= pc
->mem_cgroup
;
4425 css_get(&memcg
->css
);
4427 * At migrating an anonymous page, its mapcount goes down
4428 * to 0 and uncharge() will be called. But, even if it's fully
4429 * unmapped, migration may fail and this page has to be
4430 * charged again. We set MIGRATION flag here and delay uncharge
4431 * until end_migration() is called
4433 * Corner Case Thinking
4435 * When the old page was mapped as Anon and it's unmap-and-freed
4436 * while migration was ongoing.
4437 * If unmap finds the old page, uncharge() of it will be delayed
4438 * until end_migration(). If unmap finds a new page, it's
4439 * uncharged when it make mapcount to be 1->0. If unmap code
4440 * finds swap_migration_entry, the new page will not be mapped
4441 * and end_migration() will find it(mapcount==0).
4444 * When the old page was mapped but migraion fails, the kernel
4445 * remaps it. A charge for it is kept by MIGRATION flag even
4446 * if mapcount goes down to 0. We can do remap successfully
4447 * without charging it again.
4450 * The "old" page is under lock_page() until the end of
4451 * migration, so, the old page itself will not be swapped-out.
4452 * If the new page is swapped out before end_migraton, our
4453 * hook to usual swap-out path will catch the event.
4456 SetPageCgroupMigration(pc
);
4458 unlock_page_cgroup(pc
);
4460 * If the page is not charged at this point,
4468 * We charge new page before it's used/mapped. So, even if unlock_page()
4469 * is called before end_migration, we can catch all events on this new
4470 * page. In the case new page is migrated but not remapped, new page's
4471 * mapcount will be finally 0 and we call uncharge in end_migration().
4474 ctype
= MEM_CGROUP_CHARGE_TYPE_ANON
;
4476 ctype
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4478 * The page is committed to the memcg, but it's not actually
4479 * charged to the res_counter since we plan on replacing the
4480 * old one and only one page is going to be left afterwards.
4482 __mem_cgroup_commit_charge(memcg
, newpage
, nr_pages
, ctype
, false);
4485 /* remove redundant charge if migration failed*/
4486 void mem_cgroup_end_migration(struct mem_cgroup
*memcg
,
4487 struct page
*oldpage
, struct page
*newpage
, bool migration_ok
)
4489 struct page
*used
, *unused
;
4490 struct page_cgroup
*pc
;
4496 if (!migration_ok
) {
4503 anon
= PageAnon(used
);
4504 __mem_cgroup_uncharge_common(unused
,
4505 anon
? MEM_CGROUP_CHARGE_TYPE_ANON
4506 : MEM_CGROUP_CHARGE_TYPE_CACHE
,
4508 css_put(&memcg
->css
);
4510 * We disallowed uncharge of pages under migration because mapcount
4511 * of the page goes down to zero, temporarly.
4512 * Clear the flag and check the page should be charged.
4514 pc
= lookup_page_cgroup(oldpage
);
4515 lock_page_cgroup(pc
);
4516 ClearPageCgroupMigration(pc
);
4517 unlock_page_cgroup(pc
);
4520 * If a page is a file cache, radix-tree replacement is very atomic
4521 * and we can skip this check. When it was an Anon page, its mapcount
4522 * goes down to 0. But because we added MIGRATION flage, it's not
4523 * uncharged yet. There are several case but page->mapcount check
4524 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4525 * check. (see prepare_charge() also)
4528 mem_cgroup_uncharge_page(used
);
4532 * At replace page cache, newpage is not under any memcg but it's on
4533 * LRU. So, this function doesn't touch res_counter but handles LRU
4534 * in correct way. Both pages are locked so we cannot race with uncharge.
4536 void mem_cgroup_replace_page_cache(struct page
*oldpage
,
4537 struct page
*newpage
)
4539 struct mem_cgroup
*memcg
= NULL
;
4540 struct page_cgroup
*pc
;
4541 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4543 if (mem_cgroup_disabled())
4546 pc
= lookup_page_cgroup(oldpage
);
4547 /* fix accounting on old pages */
4548 lock_page_cgroup(pc
);
4549 if (PageCgroupUsed(pc
)) {
4550 memcg
= pc
->mem_cgroup
;
4551 mem_cgroup_charge_statistics(memcg
, oldpage
, false, -1);
4552 ClearPageCgroupUsed(pc
);
4554 unlock_page_cgroup(pc
);
4557 * When called from shmem_replace_page(), in some cases the
4558 * oldpage has already been charged, and in some cases not.
4563 * Even if newpage->mapping was NULL before starting replacement,
4564 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4565 * LRU while we overwrite pc->mem_cgroup.
4567 __mem_cgroup_commit_charge(memcg
, newpage
, 1, type
, true);
4570 #ifdef CONFIG_DEBUG_VM
4571 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
4573 struct page_cgroup
*pc
;
4575 pc
= lookup_page_cgroup(page
);
4577 * Can be NULL while feeding pages into the page allocator for
4578 * the first time, i.e. during boot or memory hotplug;
4579 * or when mem_cgroup_disabled().
4581 if (likely(pc
) && PageCgroupUsed(pc
))
4586 bool mem_cgroup_bad_page_check(struct page
*page
)
4588 if (mem_cgroup_disabled())
4591 return lookup_page_cgroup_used(page
) != NULL
;
4594 void mem_cgroup_print_bad_page(struct page
*page
)
4596 struct page_cgroup
*pc
;
4598 pc
= lookup_page_cgroup_used(page
);
4600 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4601 pc
, pc
->flags
, pc
->mem_cgroup
);
4606 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
4607 unsigned long long val
)
4610 u64 memswlimit
, memlimit
;
4612 int children
= mem_cgroup_count_children(memcg
);
4613 u64 curusage
, oldusage
;
4617 * For keeping hierarchical_reclaim simple, how long we should retry
4618 * is depends on callers. We set our retry-count to be function
4619 * of # of children which we should visit in this loop.
4621 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
4623 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4626 while (retry_count
) {
4627 if (signal_pending(current
)) {
4632 * Rather than hide all in some function, I do this in
4633 * open coded manner. You see what this really does.
4634 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4636 mutex_lock(&set_limit_mutex
);
4637 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4638 if (memswlimit
< val
) {
4640 mutex_unlock(&set_limit_mutex
);
4644 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4648 ret
= res_counter_set_limit(&memcg
->res
, val
);
4650 if (memswlimit
== val
)
4651 memcg
->memsw_is_minimum
= true;
4653 memcg
->memsw_is_minimum
= false;
4655 mutex_unlock(&set_limit_mutex
);
4660 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4661 MEM_CGROUP_RECLAIM_SHRINK
);
4662 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4663 /* Usage is reduced ? */
4664 if (curusage
>= oldusage
)
4667 oldusage
= curusage
;
4669 if (!ret
&& enlarge
)
4670 memcg_oom_recover(memcg
);
4675 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
4676 unsigned long long val
)
4679 u64 memlimit
, memswlimit
, oldusage
, curusage
;
4680 int children
= mem_cgroup_count_children(memcg
);
4684 /* see mem_cgroup_resize_res_limit */
4685 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
4686 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4687 while (retry_count
) {
4688 if (signal_pending(current
)) {
4693 * Rather than hide all in some function, I do this in
4694 * open coded manner. You see what this really does.
4695 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4697 mutex_lock(&set_limit_mutex
);
4698 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4699 if (memlimit
> val
) {
4701 mutex_unlock(&set_limit_mutex
);
4704 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4705 if (memswlimit
< val
)
4707 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
4709 if (memlimit
== val
)
4710 memcg
->memsw_is_minimum
= true;
4712 memcg
->memsw_is_minimum
= false;
4714 mutex_unlock(&set_limit_mutex
);
4719 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4720 MEM_CGROUP_RECLAIM_NOSWAP
|
4721 MEM_CGROUP_RECLAIM_SHRINK
);
4722 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4723 /* Usage is reduced ? */
4724 if (curusage
>= oldusage
)
4727 oldusage
= curusage
;
4729 if (!ret
&& enlarge
)
4730 memcg_oom_recover(memcg
);
4734 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
4736 unsigned long *total_scanned
)
4738 unsigned long nr_reclaimed
= 0;
4739 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
4740 unsigned long reclaimed
;
4742 struct mem_cgroup_tree_per_zone
*mctz
;
4743 unsigned long long excess
;
4744 unsigned long nr_scanned
;
4749 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
4751 * This loop can run a while, specially if mem_cgroup's continuously
4752 * keep exceeding their soft limit and putting the system under
4759 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
4764 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
4765 gfp_mask
, &nr_scanned
);
4766 nr_reclaimed
+= reclaimed
;
4767 *total_scanned
+= nr_scanned
;
4768 spin_lock(&mctz
->lock
);
4771 * If we failed to reclaim anything from this memory cgroup
4772 * it is time to move on to the next cgroup
4778 * Loop until we find yet another one.
4780 * By the time we get the soft_limit lock
4781 * again, someone might have aded the
4782 * group back on the RB tree. Iterate to
4783 * make sure we get a different mem.
4784 * mem_cgroup_largest_soft_limit_node returns
4785 * NULL if no other cgroup is present on
4789 __mem_cgroup_largest_soft_limit_node(mctz
);
4791 css_put(&next_mz
->memcg
->css
);
4792 else /* next_mz == NULL or other memcg */
4796 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
4797 excess
= res_counter_soft_limit_excess(&mz
->memcg
->res
);
4799 * One school of thought says that we should not add
4800 * back the node to the tree if reclaim returns 0.
4801 * But our reclaim could return 0, simply because due
4802 * to priority we are exposing a smaller subset of
4803 * memory to reclaim from. Consider this as a longer
4806 /* If excess == 0, no tree ops */
4807 __mem_cgroup_insert_exceeded(mz
->memcg
, mz
, mctz
, excess
);
4808 spin_unlock(&mctz
->lock
);
4809 css_put(&mz
->memcg
->css
);
4812 * Could not reclaim anything and there are no more
4813 * mem cgroups to try or we seem to be looping without
4814 * reclaiming anything.
4816 if (!nr_reclaimed
&&
4818 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
4820 } while (!nr_reclaimed
);
4822 css_put(&next_mz
->memcg
->css
);
4823 return nr_reclaimed
;
4827 * mem_cgroup_force_empty_list - clears LRU of a group
4828 * @memcg: group to clear
4831 * @lru: lru to to clear
4833 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4834 * reclaim the pages page themselves - pages are moved to the parent (or root)
4837 static void mem_cgroup_force_empty_list(struct mem_cgroup
*memcg
,
4838 int node
, int zid
, enum lru_list lru
)
4840 struct lruvec
*lruvec
;
4841 unsigned long flags
;
4842 struct list_head
*list
;
4846 zone
= &NODE_DATA(node
)->node_zones
[zid
];
4847 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
4848 list
= &lruvec
->lists
[lru
];
4852 struct page_cgroup
*pc
;
4855 spin_lock_irqsave(&zone
->lru_lock
, flags
);
4856 if (list_empty(list
)) {
4857 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4860 page
= list_entry(list
->prev
, struct page
, lru
);
4862 list_move(&page
->lru
, list
);
4864 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4867 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4869 pc
= lookup_page_cgroup(page
);
4871 if (mem_cgroup_move_parent(page
, pc
, memcg
)) {
4872 /* found lock contention or "pc" is obsolete. */
4877 } while (!list_empty(list
));
4881 * make mem_cgroup's charge to be 0 if there is no task by moving
4882 * all the charges and pages to the parent.
4883 * This enables deleting this mem_cgroup.
4885 * Caller is responsible for holding css reference on the memcg.
4887 static void mem_cgroup_reparent_charges(struct mem_cgroup
*memcg
)
4893 /* This is for making all *used* pages to be on LRU. */
4894 lru_add_drain_all();
4895 drain_all_stock_sync(memcg
);
4896 mem_cgroup_start_move(memcg
);
4897 for_each_node_state(node
, N_MEMORY
) {
4898 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
4901 mem_cgroup_force_empty_list(memcg
,
4906 mem_cgroup_end_move(memcg
);
4907 memcg_oom_recover(memcg
);
4911 * Kernel memory may not necessarily be trackable to a specific
4912 * process. So they are not migrated, and therefore we can't
4913 * expect their value to drop to 0 here.
4914 * Having res filled up with kmem only is enough.
4916 * This is a safety check because mem_cgroup_force_empty_list
4917 * could have raced with mem_cgroup_replace_page_cache callers
4918 * so the lru seemed empty but the page could have been added
4919 * right after the check. RES_USAGE should be safe as we always
4920 * charge before adding to the LRU.
4922 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
) -
4923 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
);
4924 } while (usage
> 0);
4928 * This mainly exists for tests during the setting of set of use_hierarchy.
4929 * Since this is the very setting we are changing, the current hierarchy value
4932 static inline bool __memcg_has_children(struct mem_cgroup
*memcg
)
4934 struct cgroup_subsys_state
*pos
;
4936 /* bounce at first found */
4937 css_for_each_child(pos
, &memcg
->css
)
4943 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4944 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4945 * from mem_cgroup_count_children(), in the sense that we don't really care how
4946 * many children we have; we only need to know if we have any. It also counts
4947 * any memcg without hierarchy as infertile.
4949 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
4951 return memcg
->use_hierarchy
&& __memcg_has_children(memcg
);
4955 * Reclaims as many pages from the given memcg as possible and moves
4956 * the rest to the parent.
4958 * Caller is responsible for holding css reference for memcg.
4960 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
4962 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
4963 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
4965 /* returns EBUSY if there is a task or if we come here twice. */
4966 if (cgroup_task_count(cgrp
) || !list_empty(&cgrp
->children
))
4969 /* we call try-to-free pages for make this cgroup empty */
4970 lru_add_drain_all();
4971 /* try to free all pages in this cgroup */
4972 while (nr_retries
&& res_counter_read_u64(&memcg
->res
, RES_USAGE
) > 0) {
4975 if (signal_pending(current
))
4978 progress
= try_to_free_mem_cgroup_pages(memcg
, GFP_KERNEL
,
4982 /* maybe some writeback is necessary */
4983 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4988 mem_cgroup_reparent_charges(memcg
);
4993 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state
*css
,
4996 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4999 if (mem_cgroup_is_root(memcg
))
5001 css_get(&memcg
->css
);
5002 ret
= mem_cgroup_force_empty(memcg
);
5003 css_put(&memcg
->css
);
5009 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
5012 return mem_cgroup_from_css(css
)->use_hierarchy
;
5015 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
5016 struct cftype
*cft
, u64 val
)
5019 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5020 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5022 mutex_lock(&memcg_create_mutex
);
5024 if (memcg
->use_hierarchy
== val
)
5028 * If parent's use_hierarchy is set, we can't make any modifications
5029 * in the child subtrees. If it is unset, then the change can
5030 * occur, provided the current cgroup has no children.
5032 * For the root cgroup, parent_mem is NULL, we allow value to be
5033 * set if there are no children.
5035 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
5036 (val
== 1 || val
== 0)) {
5037 if (!__memcg_has_children(memcg
))
5038 memcg
->use_hierarchy
= val
;
5045 mutex_unlock(&memcg_create_mutex
);
5051 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup
*memcg
,
5052 enum mem_cgroup_stat_index idx
)
5054 struct mem_cgroup
*iter
;
5057 /* Per-cpu values can be negative, use a signed accumulator */
5058 for_each_mem_cgroup_tree(iter
, memcg
)
5059 val
+= mem_cgroup_read_stat(iter
, idx
);
5061 if (val
< 0) /* race ? */
5066 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
5070 if (!mem_cgroup_is_root(memcg
)) {
5072 return res_counter_read_u64(&memcg
->res
, RES_USAGE
);
5074 return res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
5078 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5079 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5081 val
= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
5082 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_RSS
);
5085 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
5087 return val
<< PAGE_SHIFT
;
5090 static ssize_t
mem_cgroup_read(struct cgroup_subsys_state
*css
,
5091 struct cftype
*cft
, struct file
*file
,
5092 char __user
*buf
, size_t nbytes
, loff_t
*ppos
)
5094 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5100 type
= MEMFILE_TYPE(cft
->private);
5101 name
= MEMFILE_ATTR(cft
->private);
5105 if (name
== RES_USAGE
)
5106 val
= mem_cgroup_usage(memcg
, false);
5108 val
= res_counter_read_u64(&memcg
->res
, name
);
5111 if (name
== RES_USAGE
)
5112 val
= mem_cgroup_usage(memcg
, true);
5114 val
= res_counter_read_u64(&memcg
->memsw
, name
);
5117 val
= res_counter_read_u64(&memcg
->kmem
, name
);
5123 len
= scnprintf(str
, sizeof(str
), "%llu\n", (unsigned long long)val
);
5124 return simple_read_from_buffer(buf
, nbytes
, ppos
, str
, len
);
5127 static int memcg_update_kmem_limit(struct cgroup_subsys_state
*css
, u64 val
)
5130 #ifdef CONFIG_MEMCG_KMEM
5131 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5133 * For simplicity, we won't allow this to be disabled. It also can't
5134 * be changed if the cgroup has children already, or if tasks had
5137 * If tasks join before we set the limit, a person looking at
5138 * kmem.usage_in_bytes will have no way to determine when it took
5139 * place, which makes the value quite meaningless.
5141 * After it first became limited, changes in the value of the limit are
5142 * of course permitted.
5144 mutex_lock(&memcg_create_mutex
);
5145 mutex_lock(&set_limit_mutex
);
5146 if (!memcg
->kmem_account_flags
&& val
!= RESOURCE_MAX
) {
5147 if (cgroup_task_count(css
->cgroup
) || memcg_has_children(memcg
)) {
5151 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
5154 ret
= memcg_update_cache_sizes(memcg
);
5156 res_counter_set_limit(&memcg
->kmem
, RESOURCE_MAX
);
5159 static_key_slow_inc(&memcg_kmem_enabled_key
);
5161 * setting the active bit after the inc will guarantee no one
5162 * starts accounting before all call sites are patched
5164 memcg_kmem_set_active(memcg
);
5166 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
5168 mutex_unlock(&set_limit_mutex
);
5169 mutex_unlock(&memcg_create_mutex
);
5174 #ifdef CONFIG_MEMCG_KMEM
5175 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
5178 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
5182 memcg
->kmem_account_flags
= parent
->kmem_account_flags
;
5184 * When that happen, we need to disable the static branch only on those
5185 * memcgs that enabled it. To achieve this, we would be forced to
5186 * complicate the code by keeping track of which memcgs were the ones
5187 * that actually enabled limits, and which ones got it from its
5190 * It is a lot simpler just to do static_key_slow_inc() on every child
5191 * that is accounted.
5193 if (!memcg_kmem_is_active(memcg
))
5197 * __mem_cgroup_free() will issue static_key_slow_dec() because this
5198 * memcg is active already. If the later initialization fails then the
5199 * cgroup core triggers the cleanup so we do not have to do it here.
5201 static_key_slow_inc(&memcg_kmem_enabled_key
);
5203 mutex_lock(&set_limit_mutex
);
5204 memcg_stop_kmem_account();
5205 ret
= memcg_update_cache_sizes(memcg
);
5206 memcg_resume_kmem_account();
5207 mutex_unlock(&set_limit_mutex
);
5211 #endif /* CONFIG_MEMCG_KMEM */
5214 * The user of this function is...
5217 static int mem_cgroup_write(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
5220 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5223 unsigned long long val
;
5226 type
= MEMFILE_TYPE(cft
->private);
5227 name
= MEMFILE_ATTR(cft
->private);
5231 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
5235 /* This function does all necessary parse...reuse it */
5236 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5240 ret
= mem_cgroup_resize_limit(memcg
, val
);
5241 else if (type
== _MEMSWAP
)
5242 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
5243 else if (type
== _KMEM
)
5244 ret
= memcg_update_kmem_limit(css
, val
);
5248 case RES_SOFT_LIMIT
:
5249 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5253 * For memsw, soft limits are hard to implement in terms
5254 * of semantics, for now, we support soft limits for
5255 * control without swap
5258 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
5263 ret
= -EINVAL
; /* should be BUG() ? */
5269 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
5270 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
5272 unsigned long long min_limit
, min_memsw_limit
, tmp
;
5274 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5275 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5276 if (!memcg
->use_hierarchy
)
5279 while (css_parent(&memcg
->css
)) {
5280 memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5281 if (!memcg
->use_hierarchy
)
5283 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5284 min_limit
= min(min_limit
, tmp
);
5285 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5286 min_memsw_limit
= min(min_memsw_limit
, tmp
);
5289 *mem_limit
= min_limit
;
5290 *memsw_limit
= min_memsw_limit
;
5293 static int mem_cgroup_reset(struct cgroup_subsys_state
*css
, unsigned int event
)
5295 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5299 type
= MEMFILE_TYPE(event
);
5300 name
= MEMFILE_ATTR(event
);
5305 res_counter_reset_max(&memcg
->res
);
5306 else if (type
== _MEMSWAP
)
5307 res_counter_reset_max(&memcg
->memsw
);
5308 else if (type
== _KMEM
)
5309 res_counter_reset_max(&memcg
->kmem
);
5315 res_counter_reset_failcnt(&memcg
->res
);
5316 else if (type
== _MEMSWAP
)
5317 res_counter_reset_failcnt(&memcg
->memsw
);
5318 else if (type
== _KMEM
)
5319 res_counter_reset_failcnt(&memcg
->kmem
);
5328 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
5331 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
5335 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5336 struct cftype
*cft
, u64 val
)
5338 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5340 if (val
>= (1 << NR_MOVE_TYPE
))
5344 * No kind of locking is needed in here, because ->can_attach() will
5345 * check this value once in the beginning of the process, and then carry
5346 * on with stale data. This means that changes to this value will only
5347 * affect task migrations starting after the change.
5349 memcg
->move_charge_at_immigrate
= val
;
5353 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5354 struct cftype
*cft
, u64 val
)
5361 static int memcg_numa_stat_show(struct cgroup_subsys_state
*css
,
5362 struct cftype
*cft
, struct seq_file
*m
)
5365 unsigned long total_nr
, file_nr
, anon_nr
, unevictable_nr
;
5366 unsigned long node_nr
;
5367 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5369 total_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL
);
5370 seq_printf(m
, "total=%lu", total_nr
);
5371 for_each_node_state(nid
, N_MEMORY
) {
5372 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL
);
5373 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5377 file_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_FILE
);
5378 seq_printf(m
, "file=%lu", file_nr
);
5379 for_each_node_state(nid
, N_MEMORY
) {
5380 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5382 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5386 anon_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_ANON
);
5387 seq_printf(m
, "anon=%lu", anon_nr
);
5388 for_each_node_state(nid
, N_MEMORY
) {
5389 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5391 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5395 unevictable_nr
= mem_cgroup_nr_lru_pages(memcg
, BIT(LRU_UNEVICTABLE
));
5396 seq_printf(m
, "unevictable=%lu", unevictable_nr
);
5397 for_each_node_state(nid
, N_MEMORY
) {
5398 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5399 BIT(LRU_UNEVICTABLE
));
5400 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5405 #endif /* CONFIG_NUMA */
5407 static inline void mem_cgroup_lru_names_not_uptodate(void)
5409 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
5412 static int memcg_stat_show(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
5415 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5416 struct mem_cgroup
*mi
;
5419 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5420 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5422 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
5423 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
5426 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
5427 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
5428 mem_cgroup_read_events(memcg
, i
));
5430 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
5431 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
5432 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
5434 /* Hierarchical information */
5436 unsigned long long limit
, memsw_limit
;
5437 memcg_get_hierarchical_limit(memcg
, &limit
, &memsw_limit
);
5438 seq_printf(m
, "hierarchical_memory_limit %llu\n", limit
);
5439 if (do_swap_account
)
5440 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
5444 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5447 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5449 for_each_mem_cgroup_tree(mi
, memcg
)
5450 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
5451 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
5454 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
5455 unsigned long long val
= 0;
5457 for_each_mem_cgroup_tree(mi
, memcg
)
5458 val
+= mem_cgroup_read_events(mi
, i
);
5459 seq_printf(m
, "total_%s %llu\n",
5460 mem_cgroup_events_names
[i
], val
);
5463 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5464 unsigned long long val
= 0;
5466 for_each_mem_cgroup_tree(mi
, memcg
)
5467 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
5468 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
5471 #ifdef CONFIG_DEBUG_VM
5474 struct mem_cgroup_per_zone
*mz
;
5475 struct zone_reclaim_stat
*rstat
;
5476 unsigned long recent_rotated
[2] = {0, 0};
5477 unsigned long recent_scanned
[2] = {0, 0};
5479 for_each_online_node(nid
)
5480 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
5481 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
5482 rstat
= &mz
->lruvec
.reclaim_stat
;
5484 recent_rotated
[0] += rstat
->recent_rotated
[0];
5485 recent_rotated
[1] += rstat
->recent_rotated
[1];
5486 recent_scanned
[0] += rstat
->recent_scanned
[0];
5487 recent_scanned
[1] += rstat
->recent_scanned
[1];
5489 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
5490 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
5491 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
5492 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
5499 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
5502 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5504 return mem_cgroup_swappiness(memcg
);
5507 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
5508 struct cftype
*cft
, u64 val
)
5510 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5511 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5513 if (val
> 100 || !parent
)
5516 mutex_lock(&memcg_create_mutex
);
5518 /* If under hierarchy, only empty-root can set this value */
5519 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5520 mutex_unlock(&memcg_create_mutex
);
5524 memcg
->swappiness
= val
;
5526 mutex_unlock(&memcg_create_mutex
);
5531 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
5533 struct mem_cgroup_threshold_ary
*t
;
5539 t
= rcu_dereference(memcg
->thresholds
.primary
);
5541 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
5546 usage
= mem_cgroup_usage(memcg
, swap
);
5549 * current_threshold points to threshold just below or equal to usage.
5550 * If it's not true, a threshold was crossed after last
5551 * call of __mem_cgroup_threshold().
5553 i
= t
->current_threshold
;
5556 * Iterate backward over array of thresholds starting from
5557 * current_threshold and check if a threshold is crossed.
5558 * If none of thresholds below usage is crossed, we read
5559 * only one element of the array here.
5561 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
5562 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5564 /* i = current_threshold + 1 */
5568 * Iterate forward over array of thresholds starting from
5569 * current_threshold+1 and check if a threshold is crossed.
5570 * If none of thresholds above usage is crossed, we read
5571 * only one element of the array here.
5573 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
5574 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5576 /* Update current_threshold */
5577 t
->current_threshold
= i
- 1;
5582 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
5585 __mem_cgroup_threshold(memcg
, false);
5586 if (do_swap_account
)
5587 __mem_cgroup_threshold(memcg
, true);
5589 memcg
= parent_mem_cgroup(memcg
);
5593 static int compare_thresholds(const void *a
, const void *b
)
5595 const struct mem_cgroup_threshold
*_a
= a
;
5596 const struct mem_cgroup_threshold
*_b
= b
;
5598 return _a
->threshold
- _b
->threshold
;
5601 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
5603 struct mem_cgroup_eventfd_list
*ev
;
5605 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
5606 eventfd_signal(ev
->eventfd
, 1);
5610 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
5612 struct mem_cgroup
*iter
;
5614 for_each_mem_cgroup_tree(iter
, memcg
)
5615 mem_cgroup_oom_notify_cb(iter
);
5618 static int mem_cgroup_usage_register_event(struct cgroup_subsys_state
*css
,
5619 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5621 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5622 struct mem_cgroup_thresholds
*thresholds
;
5623 struct mem_cgroup_threshold_ary
*new;
5624 enum res_type type
= MEMFILE_TYPE(cft
->private);
5625 u64 threshold
, usage
;
5628 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
5632 mutex_lock(&memcg
->thresholds_lock
);
5635 thresholds
= &memcg
->thresholds
;
5636 else if (type
== _MEMSWAP
)
5637 thresholds
= &memcg
->memsw_thresholds
;
5641 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5643 /* Check if a threshold crossed before adding a new one */
5644 if (thresholds
->primary
)
5645 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5647 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
5649 /* Allocate memory for new array of thresholds */
5650 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
5658 /* Copy thresholds (if any) to new array */
5659 if (thresholds
->primary
) {
5660 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
5661 sizeof(struct mem_cgroup_threshold
));
5664 /* Add new threshold */
5665 new->entries
[size
- 1].eventfd
= eventfd
;
5666 new->entries
[size
- 1].threshold
= threshold
;
5668 /* Sort thresholds. Registering of new threshold isn't time-critical */
5669 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
5670 compare_thresholds
, NULL
);
5672 /* Find current threshold */
5673 new->current_threshold
= -1;
5674 for (i
= 0; i
< size
; i
++) {
5675 if (new->entries
[i
].threshold
<= usage
) {
5677 * new->current_threshold will not be used until
5678 * rcu_assign_pointer(), so it's safe to increment
5681 ++new->current_threshold
;
5686 /* Free old spare buffer and save old primary buffer as spare */
5687 kfree(thresholds
->spare
);
5688 thresholds
->spare
= thresholds
->primary
;
5690 rcu_assign_pointer(thresholds
->primary
, new);
5692 /* To be sure that nobody uses thresholds */
5696 mutex_unlock(&memcg
->thresholds_lock
);
5701 static void mem_cgroup_usage_unregister_event(struct cgroup_subsys_state
*css
,
5702 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5704 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5705 struct mem_cgroup_thresholds
*thresholds
;
5706 struct mem_cgroup_threshold_ary
*new;
5707 enum res_type type
= MEMFILE_TYPE(cft
->private);
5711 mutex_lock(&memcg
->thresholds_lock
);
5713 thresholds
= &memcg
->thresholds
;
5714 else if (type
== _MEMSWAP
)
5715 thresholds
= &memcg
->memsw_thresholds
;
5719 if (!thresholds
->primary
)
5722 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5724 /* Check if a threshold crossed before removing */
5725 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5727 /* Calculate new number of threshold */
5729 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
5730 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
5734 new = thresholds
->spare
;
5736 /* Set thresholds array to NULL if we don't have thresholds */
5745 /* Copy thresholds and find current threshold */
5746 new->current_threshold
= -1;
5747 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
5748 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
5751 new->entries
[j
] = thresholds
->primary
->entries
[i
];
5752 if (new->entries
[j
].threshold
<= usage
) {
5754 * new->current_threshold will not be used
5755 * until rcu_assign_pointer(), so it's safe to increment
5758 ++new->current_threshold
;
5764 /* Swap primary and spare array */
5765 thresholds
->spare
= thresholds
->primary
;
5766 /* If all events are unregistered, free the spare array */
5768 kfree(thresholds
->spare
);
5769 thresholds
->spare
= NULL
;
5772 rcu_assign_pointer(thresholds
->primary
, new);
5774 /* To be sure that nobody uses thresholds */
5777 mutex_unlock(&memcg
->thresholds_lock
);
5780 static int mem_cgroup_oom_register_event(struct cgroup_subsys_state
*css
,
5781 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5783 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5784 struct mem_cgroup_eventfd_list
*event
;
5785 enum res_type type
= MEMFILE_TYPE(cft
->private);
5787 BUG_ON(type
!= _OOM_TYPE
);
5788 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
5792 spin_lock(&memcg_oom_lock
);
5794 event
->eventfd
= eventfd
;
5795 list_add(&event
->list
, &memcg
->oom_notify
);
5797 /* already in OOM ? */
5798 if (atomic_read(&memcg
->under_oom
))
5799 eventfd_signal(eventfd
, 1);
5800 spin_unlock(&memcg_oom_lock
);
5805 static void mem_cgroup_oom_unregister_event(struct cgroup_subsys_state
*css
,
5806 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5808 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5809 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
5810 enum res_type type
= MEMFILE_TYPE(cft
->private);
5812 BUG_ON(type
!= _OOM_TYPE
);
5814 spin_lock(&memcg_oom_lock
);
5816 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
5817 if (ev
->eventfd
== eventfd
) {
5818 list_del(&ev
->list
);
5823 spin_unlock(&memcg_oom_lock
);
5826 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state
*css
,
5827 struct cftype
*cft
, struct cgroup_map_cb
*cb
)
5829 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5831 cb
->fill(cb
, "oom_kill_disable", memcg
->oom_kill_disable
);
5833 if (atomic_read(&memcg
->under_oom
))
5834 cb
->fill(cb
, "under_oom", 1);
5836 cb
->fill(cb
, "under_oom", 0);
5840 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
5841 struct cftype
*cft
, u64 val
)
5843 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5844 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5846 /* cannot set to root cgroup and only 0 and 1 are allowed */
5847 if (!parent
|| !((val
== 0) || (val
== 1)))
5850 mutex_lock(&memcg_create_mutex
);
5851 /* oom-kill-disable is a flag for subhierarchy. */
5852 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5853 mutex_unlock(&memcg_create_mutex
);
5856 memcg
->oom_kill_disable
= val
;
5858 memcg_oom_recover(memcg
);
5859 mutex_unlock(&memcg_create_mutex
);
5863 #ifdef CONFIG_MEMCG_KMEM
5864 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5868 memcg
->kmemcg_id
= -1;
5869 ret
= memcg_propagate_kmem(memcg
);
5873 return mem_cgroup_sockets_init(memcg
, ss
);
5876 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5878 mem_cgroup_sockets_destroy(memcg
);
5881 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5883 if (!memcg_kmem_is_active(memcg
))
5887 * kmem charges can outlive the cgroup. In the case of slab
5888 * pages, for instance, a page contain objects from various
5889 * processes. As we prevent from taking a reference for every
5890 * such allocation we have to be careful when doing uncharge
5891 * (see memcg_uncharge_kmem) and here during offlining.
5893 * The idea is that that only the _last_ uncharge which sees
5894 * the dead memcg will drop the last reference. An additional
5895 * reference is taken here before the group is marked dead
5896 * which is then paired with css_put during uncharge resp. here.
5898 * Although this might sound strange as this path is called from
5899 * css_offline() when the referencemight have dropped down to 0
5900 * and shouldn't be incremented anymore (css_tryget would fail)
5901 * we do not have other options because of the kmem allocations
5904 css_get(&memcg
->css
);
5906 memcg_kmem_mark_dead(memcg
);
5908 if (res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0)
5911 if (memcg_kmem_test_and_clear_dead(memcg
))
5912 css_put(&memcg
->css
);
5915 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5920 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5924 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5929 static struct cftype mem_cgroup_files
[] = {
5931 .name
= "usage_in_bytes",
5932 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
5933 .read
= mem_cgroup_read
,
5934 .register_event
= mem_cgroup_usage_register_event
,
5935 .unregister_event
= mem_cgroup_usage_unregister_event
,
5938 .name
= "max_usage_in_bytes",
5939 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
5940 .trigger
= mem_cgroup_reset
,
5941 .read
= mem_cgroup_read
,
5944 .name
= "limit_in_bytes",
5945 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
5946 .write_string
= mem_cgroup_write
,
5947 .read
= mem_cgroup_read
,
5950 .name
= "soft_limit_in_bytes",
5951 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
5952 .write_string
= mem_cgroup_write
,
5953 .read
= mem_cgroup_read
,
5957 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
5958 .trigger
= mem_cgroup_reset
,
5959 .read
= mem_cgroup_read
,
5963 .read_seq_string
= memcg_stat_show
,
5966 .name
= "force_empty",
5967 .trigger
= mem_cgroup_force_empty_write
,
5970 .name
= "use_hierarchy",
5971 .flags
= CFTYPE_INSANE
,
5972 .write_u64
= mem_cgroup_hierarchy_write
,
5973 .read_u64
= mem_cgroup_hierarchy_read
,
5976 .name
= "swappiness",
5977 .read_u64
= mem_cgroup_swappiness_read
,
5978 .write_u64
= mem_cgroup_swappiness_write
,
5981 .name
= "move_charge_at_immigrate",
5982 .read_u64
= mem_cgroup_move_charge_read
,
5983 .write_u64
= mem_cgroup_move_charge_write
,
5986 .name
= "oom_control",
5987 .read_map
= mem_cgroup_oom_control_read
,
5988 .write_u64
= mem_cgroup_oom_control_write
,
5989 .register_event
= mem_cgroup_oom_register_event
,
5990 .unregister_event
= mem_cgroup_oom_unregister_event
,
5991 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
5994 .name
= "pressure_level",
5995 .register_event
= vmpressure_register_event
,
5996 .unregister_event
= vmpressure_unregister_event
,
6000 .name
= "numa_stat",
6001 .read_seq_string
= memcg_numa_stat_show
,
6004 #ifdef CONFIG_MEMCG_KMEM
6006 .name
= "kmem.limit_in_bytes",
6007 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
6008 .write_string
= mem_cgroup_write
,
6009 .read
= mem_cgroup_read
,
6012 .name
= "kmem.usage_in_bytes",
6013 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
6014 .read
= mem_cgroup_read
,
6017 .name
= "kmem.failcnt",
6018 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
6019 .trigger
= mem_cgroup_reset
,
6020 .read
= mem_cgroup_read
,
6023 .name
= "kmem.max_usage_in_bytes",
6024 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
6025 .trigger
= mem_cgroup_reset
,
6026 .read
= mem_cgroup_read
,
6028 #ifdef CONFIG_SLABINFO
6030 .name
= "kmem.slabinfo",
6031 .read_seq_string
= mem_cgroup_slabinfo_read
,
6035 { }, /* terminate */
6038 #ifdef CONFIG_MEMCG_SWAP
6039 static struct cftype memsw_cgroup_files
[] = {
6041 .name
= "memsw.usage_in_bytes",
6042 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
6043 .read
= mem_cgroup_read
,
6044 .register_event
= mem_cgroup_usage_register_event
,
6045 .unregister_event
= mem_cgroup_usage_unregister_event
,
6048 .name
= "memsw.max_usage_in_bytes",
6049 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
6050 .trigger
= mem_cgroup_reset
,
6051 .read
= mem_cgroup_read
,
6054 .name
= "memsw.limit_in_bytes",
6055 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
6056 .write_string
= mem_cgroup_write
,
6057 .read
= mem_cgroup_read
,
6060 .name
= "memsw.failcnt",
6061 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
6062 .trigger
= mem_cgroup_reset
,
6063 .read
= mem_cgroup_read
,
6065 { }, /* terminate */
6068 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6070 struct mem_cgroup_per_node
*pn
;
6071 struct mem_cgroup_per_zone
*mz
;
6072 int zone
, tmp
= node
;
6074 * This routine is called against possible nodes.
6075 * But it's BUG to call kmalloc() against offline node.
6077 * TODO: this routine can waste much memory for nodes which will
6078 * never be onlined. It's better to use memory hotplug callback
6081 if (!node_state(node
, N_NORMAL_MEMORY
))
6083 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
6087 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6088 mz
= &pn
->zoneinfo
[zone
];
6089 lruvec_init(&mz
->lruvec
);
6090 mz
->usage_in_excess
= 0;
6091 mz
->on_tree
= false;
6094 memcg
->nodeinfo
[node
] = pn
;
6098 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6100 kfree(memcg
->nodeinfo
[node
]);
6103 static struct mem_cgroup
*mem_cgroup_alloc(void)
6105 struct mem_cgroup
*memcg
;
6106 size_t size
= memcg_size();
6108 /* Can be very big if nr_node_ids is very big */
6109 if (size
< PAGE_SIZE
)
6110 memcg
= kzalloc(size
, GFP_KERNEL
);
6112 memcg
= vzalloc(size
);
6117 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
6120 spin_lock_init(&memcg
->pcp_counter_lock
);
6124 if (size
< PAGE_SIZE
)
6132 * At destroying mem_cgroup, references from swap_cgroup can remain.
6133 * (scanning all at force_empty is too costly...)
6135 * Instead of clearing all references at force_empty, we remember
6136 * the number of reference from swap_cgroup and free mem_cgroup when
6137 * it goes down to 0.
6139 * Removal of cgroup itself succeeds regardless of refs from swap.
6142 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
6145 size_t size
= memcg_size();
6147 mem_cgroup_remove_from_trees(memcg
);
6148 free_css_id(&mem_cgroup_subsys
, &memcg
->css
);
6151 free_mem_cgroup_per_zone_info(memcg
, node
);
6153 free_percpu(memcg
->stat
);
6156 * We need to make sure that (at least for now), the jump label
6157 * destruction code runs outside of the cgroup lock. This is because
6158 * get_online_cpus(), which is called from the static_branch update,
6159 * can't be called inside the cgroup_lock. cpusets are the ones
6160 * enforcing this dependency, so if they ever change, we might as well.
6162 * schedule_work() will guarantee this happens. Be careful if you need
6163 * to move this code around, and make sure it is outside
6166 disarm_static_keys(memcg
);
6167 if (size
< PAGE_SIZE
)
6174 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6176 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
6178 if (!memcg
->res
.parent
)
6180 return mem_cgroup_from_res_counter(memcg
->res
.parent
, res
);
6182 EXPORT_SYMBOL(parent_mem_cgroup
);
6184 static void __init
mem_cgroup_soft_limit_tree_init(void)
6186 struct mem_cgroup_tree_per_node
*rtpn
;
6187 struct mem_cgroup_tree_per_zone
*rtpz
;
6188 int tmp
, node
, zone
;
6190 for_each_node(node
) {
6192 if (!node_state(node
, N_NORMAL_MEMORY
))
6194 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
6197 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
6199 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6200 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
6201 rtpz
->rb_root
= RB_ROOT
;
6202 spin_lock_init(&rtpz
->lock
);
6207 static struct cgroup_subsys_state
* __ref
6208 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
6210 struct mem_cgroup
*memcg
;
6211 long error
= -ENOMEM
;
6214 memcg
= mem_cgroup_alloc();
6216 return ERR_PTR(error
);
6219 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
6223 if (parent_css
== NULL
) {
6224 root_mem_cgroup
= memcg
;
6225 res_counter_init(&memcg
->res
, NULL
);
6226 res_counter_init(&memcg
->memsw
, NULL
);
6227 res_counter_init(&memcg
->kmem
, NULL
);
6230 memcg
->last_scanned_node
= MAX_NUMNODES
;
6231 INIT_LIST_HEAD(&memcg
->oom_notify
);
6232 memcg
->move_charge_at_immigrate
= 0;
6233 mutex_init(&memcg
->thresholds_lock
);
6234 spin_lock_init(&memcg
->move_lock
);
6235 vmpressure_init(&memcg
->vmpressure
);
6240 __mem_cgroup_free(memcg
);
6241 return ERR_PTR(error
);
6245 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
6247 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6248 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(css
));
6254 mutex_lock(&memcg_create_mutex
);
6256 memcg
->use_hierarchy
= parent
->use_hierarchy
;
6257 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
6258 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
6260 if (parent
->use_hierarchy
) {
6261 res_counter_init(&memcg
->res
, &parent
->res
);
6262 res_counter_init(&memcg
->memsw
, &parent
->memsw
);
6263 res_counter_init(&memcg
->kmem
, &parent
->kmem
);
6266 * No need to take a reference to the parent because cgroup
6267 * core guarantees its existence.
6270 res_counter_init(&memcg
->res
, NULL
);
6271 res_counter_init(&memcg
->memsw
, NULL
);
6272 res_counter_init(&memcg
->kmem
, NULL
);
6274 * Deeper hierachy with use_hierarchy == false doesn't make
6275 * much sense so let cgroup subsystem know about this
6276 * unfortunate state in our controller.
6278 if (parent
!= root_mem_cgroup
)
6279 mem_cgroup_subsys
.broken_hierarchy
= true;
6282 error
= memcg_init_kmem(memcg
, &mem_cgroup_subsys
);
6283 mutex_unlock(&memcg_create_mutex
);
6288 * Announce all parents that a group from their hierarchy is gone.
6290 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup
*memcg
)
6292 struct mem_cgroup
*parent
= memcg
;
6294 while ((parent
= parent_mem_cgroup(parent
)))
6295 mem_cgroup_iter_invalidate(parent
);
6298 * if the root memcg is not hierarchical we have to check it
6301 if (!root_mem_cgroup
->use_hierarchy
)
6302 mem_cgroup_iter_invalidate(root_mem_cgroup
);
6305 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
6307 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6309 kmem_cgroup_css_offline(memcg
);
6311 mem_cgroup_invalidate_reclaim_iterators(memcg
);
6312 mem_cgroup_reparent_charges(memcg
);
6313 mem_cgroup_destroy_all_caches(memcg
);
6316 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
6318 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6320 memcg_destroy_kmem(memcg
);
6321 __mem_cgroup_free(memcg
);
6325 /* Handlers for move charge at task migration. */
6326 #define PRECHARGE_COUNT_AT_ONCE 256
6327 static int mem_cgroup_do_precharge(unsigned long count
)
6330 int batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6331 struct mem_cgroup
*memcg
= mc
.to
;
6333 if (mem_cgroup_is_root(memcg
)) {
6334 mc
.precharge
+= count
;
6335 /* we don't need css_get for root */
6338 /* try to charge at once */
6340 struct res_counter
*dummy
;
6342 * "memcg" cannot be under rmdir() because we've already checked
6343 * by cgroup_lock_live_cgroup() that it is not removed and we
6344 * are still under the same cgroup_mutex. So we can postpone
6347 if (res_counter_charge(&memcg
->res
, PAGE_SIZE
* count
, &dummy
))
6349 if (do_swap_account
&& res_counter_charge(&memcg
->memsw
,
6350 PAGE_SIZE
* count
, &dummy
)) {
6351 res_counter_uncharge(&memcg
->res
, PAGE_SIZE
* count
);
6354 mc
.precharge
+= count
;
6358 /* fall back to one by one charge */
6360 if (signal_pending(current
)) {
6364 if (!batch_count
--) {
6365 batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6368 ret
= __mem_cgroup_try_charge(NULL
,
6369 GFP_KERNEL
, 1, &memcg
, false);
6371 /* mem_cgroup_clear_mc() will do uncharge later */
6379 * get_mctgt_type - get target type of moving charge
6380 * @vma: the vma the pte to be checked belongs
6381 * @addr: the address corresponding to the pte to be checked
6382 * @ptent: the pte to be checked
6383 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6386 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6387 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6388 * move charge. if @target is not NULL, the page is stored in target->page
6389 * with extra refcnt got(Callers should handle it).
6390 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6391 * target for charge migration. if @target is not NULL, the entry is stored
6394 * Called with pte lock held.
6401 enum mc_target_type
{
6407 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
6408 unsigned long addr
, pte_t ptent
)
6410 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
6412 if (!page
|| !page_mapped(page
))
6414 if (PageAnon(page
)) {
6415 /* we don't move shared anon */
6418 } else if (!move_file())
6419 /* we ignore mapcount for file pages */
6421 if (!get_page_unless_zero(page
))
6428 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6429 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6431 struct page
*page
= NULL
;
6432 swp_entry_t ent
= pte_to_swp_entry(ptent
);
6434 if (!move_anon() || non_swap_entry(ent
))
6437 * Because lookup_swap_cache() updates some statistics counter,
6438 * we call find_get_page() with swapper_space directly.
6440 page
= find_get_page(swap_address_space(ent
), ent
.val
);
6441 if (do_swap_account
)
6442 entry
->val
= ent
.val
;
6447 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6448 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6454 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
6455 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6457 struct page
*page
= NULL
;
6458 struct address_space
*mapping
;
6461 if (!vma
->vm_file
) /* anonymous vma */
6466 mapping
= vma
->vm_file
->f_mapping
;
6467 if (pte_none(ptent
))
6468 pgoff
= linear_page_index(vma
, addr
);
6469 else /* pte_file(ptent) is true */
6470 pgoff
= pte_to_pgoff(ptent
);
6472 /* page is moved even if it's not RSS of this task(page-faulted). */
6473 page
= find_get_page(mapping
, pgoff
);
6476 /* shmem/tmpfs may report page out on swap: account for that too. */
6477 if (radix_tree_exceptional_entry(page
)) {
6478 swp_entry_t swap
= radix_to_swp_entry(page
);
6479 if (do_swap_account
)
6481 page
= find_get_page(swap_address_space(swap
), swap
.val
);
6487 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
6488 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
6490 struct page
*page
= NULL
;
6491 struct page_cgroup
*pc
;
6492 enum mc_target_type ret
= MC_TARGET_NONE
;
6493 swp_entry_t ent
= { .val
= 0 };
6495 if (pte_present(ptent
))
6496 page
= mc_handle_present_pte(vma
, addr
, ptent
);
6497 else if (is_swap_pte(ptent
))
6498 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
6499 else if (pte_none(ptent
) || pte_file(ptent
))
6500 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
6502 if (!page
&& !ent
.val
)
6505 pc
= lookup_page_cgroup(page
);
6507 * Do only loose check w/o page_cgroup lock.
6508 * mem_cgroup_move_account() checks the pc is valid or not under
6511 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6512 ret
= MC_TARGET_PAGE
;
6514 target
->page
= page
;
6516 if (!ret
|| !target
)
6519 /* There is a swap entry and a page doesn't exist or isn't charged */
6520 if (ent
.val
&& !ret
&&
6521 css_id(&mc
.from
->css
) == lookup_swap_cgroup_id(ent
)) {
6522 ret
= MC_TARGET_SWAP
;
6529 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6531 * We don't consider swapping or file mapped pages because THP does not
6532 * support them for now.
6533 * Caller should make sure that pmd_trans_huge(pmd) is true.
6535 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6536 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6538 struct page
*page
= NULL
;
6539 struct page_cgroup
*pc
;
6540 enum mc_target_type ret
= MC_TARGET_NONE
;
6542 page
= pmd_page(pmd
);
6543 VM_BUG_ON(!page
|| !PageHead(page
));
6546 pc
= lookup_page_cgroup(page
);
6547 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6548 ret
= MC_TARGET_PAGE
;
6551 target
->page
= page
;
6557 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6558 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6560 return MC_TARGET_NONE
;
6564 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
6565 unsigned long addr
, unsigned long end
,
6566 struct mm_walk
*walk
)
6568 struct vm_area_struct
*vma
= walk
->private;
6572 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6573 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
6574 mc
.precharge
+= HPAGE_PMD_NR
;
6575 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6579 if (pmd_trans_unstable(pmd
))
6581 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6582 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
6583 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
6584 mc
.precharge
++; /* increment precharge temporarily */
6585 pte_unmap_unlock(pte
- 1, ptl
);
6591 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
6593 unsigned long precharge
;
6594 struct vm_area_struct
*vma
;
6596 down_read(&mm
->mmap_sem
);
6597 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6598 struct mm_walk mem_cgroup_count_precharge_walk
= {
6599 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
6603 if (is_vm_hugetlb_page(vma
))
6605 walk_page_range(vma
->vm_start
, vma
->vm_end
,
6606 &mem_cgroup_count_precharge_walk
);
6608 up_read(&mm
->mmap_sem
);
6610 precharge
= mc
.precharge
;
6616 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
6618 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
6620 VM_BUG_ON(mc
.moving_task
);
6621 mc
.moving_task
= current
;
6622 return mem_cgroup_do_precharge(precharge
);
6625 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6626 static void __mem_cgroup_clear_mc(void)
6628 struct mem_cgroup
*from
= mc
.from
;
6629 struct mem_cgroup
*to
= mc
.to
;
6632 /* we must uncharge all the leftover precharges from mc.to */
6634 __mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
6638 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6639 * we must uncharge here.
6641 if (mc
.moved_charge
) {
6642 __mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
6643 mc
.moved_charge
= 0;
6645 /* we must fixup refcnts and charges */
6646 if (mc
.moved_swap
) {
6647 /* uncharge swap account from the old cgroup */
6648 if (!mem_cgroup_is_root(mc
.from
))
6649 res_counter_uncharge(&mc
.from
->memsw
,
6650 PAGE_SIZE
* mc
.moved_swap
);
6652 for (i
= 0; i
< mc
.moved_swap
; i
++)
6653 css_put(&mc
.from
->css
);
6655 if (!mem_cgroup_is_root(mc
.to
)) {
6657 * we charged both to->res and to->memsw, so we should
6660 res_counter_uncharge(&mc
.to
->res
,
6661 PAGE_SIZE
* mc
.moved_swap
);
6663 /* we've already done css_get(mc.to) */
6666 memcg_oom_recover(from
);
6667 memcg_oom_recover(to
);
6668 wake_up_all(&mc
.waitq
);
6671 static void mem_cgroup_clear_mc(void)
6673 struct mem_cgroup
*from
= mc
.from
;
6676 * we must clear moving_task before waking up waiters at the end of
6679 mc
.moving_task
= NULL
;
6680 __mem_cgroup_clear_mc();
6681 spin_lock(&mc
.lock
);
6684 spin_unlock(&mc
.lock
);
6685 mem_cgroup_end_move(from
);
6688 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6689 struct cgroup_taskset
*tset
)
6691 struct task_struct
*p
= cgroup_taskset_first(tset
);
6693 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6694 unsigned long move_charge_at_immigrate
;
6697 * We are now commited to this value whatever it is. Changes in this
6698 * tunable will only affect upcoming migrations, not the current one.
6699 * So we need to save it, and keep it going.
6701 move_charge_at_immigrate
= memcg
->move_charge_at_immigrate
;
6702 if (move_charge_at_immigrate
) {
6703 struct mm_struct
*mm
;
6704 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
6706 VM_BUG_ON(from
== memcg
);
6708 mm
= get_task_mm(p
);
6711 /* We move charges only when we move a owner of the mm */
6712 if (mm
->owner
== p
) {
6715 VM_BUG_ON(mc
.precharge
);
6716 VM_BUG_ON(mc
.moved_charge
);
6717 VM_BUG_ON(mc
.moved_swap
);
6718 mem_cgroup_start_move(from
);
6719 spin_lock(&mc
.lock
);
6722 mc
.immigrate_flags
= move_charge_at_immigrate
;
6723 spin_unlock(&mc
.lock
);
6724 /* We set mc.moving_task later */
6726 ret
= mem_cgroup_precharge_mc(mm
);
6728 mem_cgroup_clear_mc();
6735 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6736 struct cgroup_taskset
*tset
)
6738 mem_cgroup_clear_mc();
6741 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6742 unsigned long addr
, unsigned long end
,
6743 struct mm_walk
*walk
)
6746 struct vm_area_struct
*vma
= walk
->private;
6749 enum mc_target_type target_type
;
6750 union mc_target target
;
6752 struct page_cgroup
*pc
;
6755 * We don't take compound_lock() here but no race with splitting thp
6757 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6758 * under splitting, which means there's no concurrent thp split,
6759 * - if another thread runs into split_huge_page() just after we
6760 * entered this if-block, the thread must wait for page table lock
6761 * to be unlocked in __split_huge_page_splitting(), where the main
6762 * part of thp split is not executed yet.
6764 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6765 if (mc
.precharge
< HPAGE_PMD_NR
) {
6766 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6769 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6770 if (target_type
== MC_TARGET_PAGE
) {
6772 if (!isolate_lru_page(page
)) {
6773 pc
= lookup_page_cgroup(page
);
6774 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
6775 pc
, mc
.from
, mc
.to
)) {
6776 mc
.precharge
-= HPAGE_PMD_NR
;
6777 mc
.moved_charge
+= HPAGE_PMD_NR
;
6779 putback_lru_page(page
);
6783 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6787 if (pmd_trans_unstable(pmd
))
6790 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6791 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6792 pte_t ptent
= *(pte
++);
6798 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6799 case MC_TARGET_PAGE
:
6801 if (isolate_lru_page(page
))
6803 pc
= lookup_page_cgroup(page
);
6804 if (!mem_cgroup_move_account(page
, 1, pc
,
6807 /* we uncharge from mc.from later. */
6810 putback_lru_page(page
);
6811 put
: /* get_mctgt_type() gets the page */
6814 case MC_TARGET_SWAP
:
6816 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6818 /* we fixup refcnts and charges later. */
6826 pte_unmap_unlock(pte
- 1, ptl
);
6831 * We have consumed all precharges we got in can_attach().
6832 * We try charge one by one, but don't do any additional
6833 * charges to mc.to if we have failed in charge once in attach()
6836 ret
= mem_cgroup_do_precharge(1);
6844 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
6846 struct vm_area_struct
*vma
;
6848 lru_add_drain_all();
6850 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
6852 * Someone who are holding the mmap_sem might be waiting in
6853 * waitq. So we cancel all extra charges, wake up all waiters,
6854 * and retry. Because we cancel precharges, we might not be able
6855 * to move enough charges, but moving charge is a best-effort
6856 * feature anyway, so it wouldn't be a big problem.
6858 __mem_cgroup_clear_mc();
6862 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6864 struct mm_walk mem_cgroup_move_charge_walk
= {
6865 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6869 if (is_vm_hugetlb_page(vma
))
6871 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
6872 &mem_cgroup_move_charge_walk
);
6875 * means we have consumed all precharges and failed in
6876 * doing additional charge. Just abandon here.
6880 up_read(&mm
->mmap_sem
);
6883 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6884 struct cgroup_taskset
*tset
)
6886 struct task_struct
*p
= cgroup_taskset_first(tset
);
6887 struct mm_struct
*mm
= get_task_mm(p
);
6891 mem_cgroup_move_charge(mm
);
6895 mem_cgroup_clear_mc();
6897 #else /* !CONFIG_MMU */
6898 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6899 struct cgroup_taskset
*tset
)
6903 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6904 struct cgroup_taskset
*tset
)
6907 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6908 struct cgroup_taskset
*tset
)
6914 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6915 * to verify sane_behavior flag on each mount attempt.
6917 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
6920 * use_hierarchy is forced with sane_behavior. cgroup core
6921 * guarantees that @root doesn't have any children, so turning it
6922 * on for the root memcg is enough.
6924 if (cgroup_sane_behavior(root_css
->cgroup
))
6925 mem_cgroup_from_css(root_css
)->use_hierarchy
= true;
6928 struct cgroup_subsys mem_cgroup_subsys
= {
6930 .subsys_id
= mem_cgroup_subsys_id
,
6931 .css_alloc
= mem_cgroup_css_alloc
,
6932 .css_online
= mem_cgroup_css_online
,
6933 .css_offline
= mem_cgroup_css_offline
,
6934 .css_free
= mem_cgroup_css_free
,
6935 .can_attach
= mem_cgroup_can_attach
,
6936 .cancel_attach
= mem_cgroup_cancel_attach
,
6937 .attach
= mem_cgroup_move_task
,
6938 .bind
= mem_cgroup_bind
,
6939 .base_cftypes
= mem_cgroup_files
,
6944 #ifdef CONFIG_MEMCG_SWAP
6945 static int __init
enable_swap_account(char *s
)
6947 /* consider enabled if no parameter or 1 is given */
6948 if (!strcmp(s
, "1"))
6949 really_do_swap_account
= 1;
6950 else if (!strcmp(s
, "0"))
6951 really_do_swap_account
= 0;
6954 __setup("swapaccount=", enable_swap_account
);
6956 static void __init
memsw_file_init(void)
6958 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys
, memsw_cgroup_files
));
6961 static void __init
enable_swap_cgroup(void)
6963 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6964 do_swap_account
= 1;
6970 static void __init
enable_swap_cgroup(void)
6976 * subsys_initcall() for memory controller.
6978 * Some parts like hotcpu_notifier() have to be initialized from this context
6979 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6980 * everything that doesn't depend on a specific mem_cgroup structure should
6981 * be initialized from here.
6983 static int __init
mem_cgroup_init(void)
6985 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
6986 enable_swap_cgroup();
6987 mem_cgroup_soft_limit_tree_init();
6991 subsys_initcall(mem_cgroup_init
);